UFDC Home  myUFDC Home  Help 



Full Text  
xml version 1.0 encoding UTF8 REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchemainstance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd INGEST IEID E20101108_AAAABL INGEST_TIME 20101108T16:33:09Z PACKAGE UFE0021707_00001 AGREEMENT_INFO ACCOUNT UF PROJECT UFDC FILES FILE SIZE 1962 DFID F20101108_AABFBN ORIGIN DEPOSITOR PATH ramu_p_Page_049.txt GLOBAL false PRESERVATION BIT MESSAGE_DIGEST ALGORITHM MD5 adc3f13a94739efc1e2e0f1170f0faf5 SHA1 8ce2b217b002ad08b373e7b5740175864c9de371 22346 F20101108_AABEWH ramu_p_Page_050.QC.jpg 7ef73dec1a09baf2768b365259791ff6 2cdebbe2ba165cce1bec000a4d9d11adc5b4da86 46342 F20101108_AABFAZ ramu_p_Page_068.jpg 81593737ad15e42d4ea4981d8bf7b441 dda24f484375a01a33eb2fafa8661d9757c46b27 6431 F20101108_AABFZJ ramu_p_Page_024thm.jpg 6b420ed6d921541b31aff5908dbc8d49 ef23a26e5f65445ac78b78ea6862d12a53c333b7 1332 F20101108_AABFYV ramu_p_Page_002thm.jpg 305a0f3d9bcd285313b383e5bc1aefa2 5ea258fceeef48716f7474d251d8ad7f00336f8e 1138 F20101108_AABFCC ramu_p_Page_099.txt 8e8498753c8bd1ca9ae827d4fb96d936 efa8cbe99d3a1cf14dab5a13e808475f00282245 1053954 F20101108_AABFBO ramu_p_Page_075.tif 029433658b9a79a866359c9dc3d47f18 0f040b8083e14ef5aee1177b1538b0387af3b4a8 1609 F20101108_AABEWI ramu_p_Page_088.txt 3fb115b503dd78d054c993b983645520 26491f0cf35d23d466437bd1b6aa9c612a9aef9c 6987 F20101108_AABFZK ramu_p_Page_026thm.jpg 334c2c73c819c30051c2198f111905c3 4417d39e82293a2b1a8fd7bb395e15d80c9ea826 1372 F20101108_AABFYW ramu_p_Page_003thm.jpg 4528dcb5cae36bfc56f665b3f86f14cb fa9f07ecc3e385dab53fca6a94c2d3aa6e9f1df1 4932 F20101108_AABFCD ramu_p_Page_083.pro 31769b7b17dd2d95cf7f6ccb2d7ca4d6 b08302510e006000f699c58354d0676b76c91e41 29383 F20101108_AABFBP ramu_p_Page_114.jpg 74085807c51d21a3031bd745585fb56e 7c432fc91ebb0d51a21c66b3008a566ce0cb2bcf 514248 F20101108_AABEWJ ramu_p_Page_066.jp2 4cacf4d35e2ebd466fd4a20e99bb6f1c 4d1acaf3265f078fbb7ae79c8c3fbf9dccf89379 26255 F20101108_AABFZL ramu_p_Page_027.QC.jpg 6dc38ed14fab2d73e285795164d72242 09b99246a66ef910d4cc4c60f0c5be7dbccb2ce7 20476 F20101108_AABFYX ramu_p_Page_006.QC.jpg 6fdc1a9b61fda765db157ed884736d53 07889db43c784ae60e3c88ee0c13183ae1bd6c80 8423998 F20101108_AABFCE ramu_p_Page_025.tif 73392c89d4d3c7f2efc2204903f1e1ba 93efe10fbbd9dcffe9ce11c52f92accecdc3d51d 9754 F20101108_AABFBQ ramu_p_Page_114.QC.jpg 8b6a5f8140712479c97dd1b30371912f 7ea653e74cff057f1cd179f733afd6b4b07c5c20 12225 F20101108_AABEWK ramu_p_Page_091.QC.jpg faeee95062ee17a22cd216c6b3fcaf87 a01888bc863005c5c8a190f2031e9ff9485d4858 7013 F20101108_AABFZM ramu_p_Page_027thm.jpg f83edcef8bbce0b3e8bd3a13ae03161d fa88d665bc773a76c035080b5bb92d6d10776b57 5713 F20101108_AABFYY ramu_p_Page_008.QC.jpg 4b9fb00b0fbf66089d305334935a1be4 20d71f3aebf56dabcf8e0ef37fcac3bcaa8352b4 103706 F20101108_AABFCF ramu_p_Page_100.jp2 e41c22dd24af9f0abb43f7674d6c6bc3 7a56072429942d210f445f74ec450d8430a5baad 6955 F20101108_AABFBR ramu_p_Page_054thm.jpg 69a5c55f6a589182fbea75b3ec7e05e4 56fd3bcc757050d97c0bad2de8e5337de7bab781 1816 F20101108_AABEWL ramu_p_Page_042.txt 03b7441ac9f87d4c12a3dd20c965a72f b23e71d33a136e29b8ebb51b30334eda4b61c74d 7027 F20101108_AABFZN ramu_p_Page_028thm.jpg 12d7cca5b10c0d9a818923081e9b80c7 74c33b9c638e8166ed5354fdcde69af4e802a2cc 26002 F20101108_AABFYZ ramu_p_Page_009.QC.jpg ea20afc0e519f4f781dc83486d5d0f05 c1f4157b8c7f4acc8b01a3395622eab3e744f691 F20101108_AABFCG ramu_p_Page_077.tif aa9a82e84f626b7df578bb3e923ab073 874fe3c7d6981d346962cd913f587690dcbc2565 2132 F20101108_AABEXA ramu_p_Page_018.txt aff3e41b2b2973bc5c8718cdb01ef3bc e36ef460cbbc2d80c279734023842a1ff8350032 65369 F20101108_AABFBS ramu_p_Page_040.jpg 524bf9dba5d9d899b920fc355f9b702b 73a379dd6d8550be9c7444f1e3154d0b73a57b1b 266236 F20101108_AABEWM ramu_p_Page_083.jp2 e729a91a5033bc27f6f2b0a367498485 a59c2a9772490f0d4d1573486fba9275bfa63549 25123 F20101108_AABFZO ramu_p_Page_029.QC.jpg a9c28b1d1bcd262fba730b379ef52c7b 9ee5c3f96f72125737c36b973024682aa568c594 51177 F20101108_AABFCH ramu_p_Page_126.jp2 93a04f43e959cf63df172dac38b929bf 56f5490ab3cfbac8af54abc3c4722223fabf1034 14116 F20101108_AABEXB ramu_p_Page_092.jpg 60a1937c25ccf80d9dc1e9c70d9ebd5b 6904697b423605974759ebd345b8b797a0525af6 4130 F20101108_AABFBT ramu_p_Page_080thm.jpg 97a1368817f04d9e938e5a8125d2740b eca721fec3e029a38f123ab902a3944a7212f7eb F20101108_AABEWN ramu_p_Page_102.tif 57e9a0a3ae7263c1d6f7cc7bf07aafb0 a3bfa338de640e8d7fa9c592427a45c1fdbe0f90 6995 F20101108_AABFZP ramu_p_Page_029thm.jpg 0853193cc19a0351c664709ebda783ea 925cba0ba294a862241e74772fedafa25e52cb8a 2320 F20101108_AABFCI ramu_p_Page_124.txt daa1f175302f6236802a31abfef4ea9b 320244bd079ce3e0d84ad8d3268142a3621c1783 12654 F20101108_AABEXC ramu_p_Page_088.QC.jpg 8c364461ae454ff2c5f7771b8378816a a4d2b26415d15e8ecfe5223b4a49d26c3894479b 25271604 F20101108_AABFBU ramu_p_Page_066.tif 6ad3dfa4a98a7a3d9498f6c915be4085 3fbf48be92b25a04c735bfe6f1ff101ff31678fd 49420 F20101108_AABEWO ramu_p_Page_024.pro e288651988490f4cc61c516b94a004d7 e3ac7a20c29a9a92e7846153b19e8e2f0efad40d 23737 F20101108_AABFZQ ramu_p_Page_030.QC.jpg e53c2b7c1b85a25b06f166d27b01f7a5 f34176231541b48240ccb5c38f0d51da658afc99 50691 F20101108_AABFCJ ramu_p_Page_105.pro 69e45364fb413d74364b8c766073c8b3 62b1cfcd4cc095ca833217432d85104b9db3b0e6 8739 F20101108_AABEXD ramu_p_Page_066.QC.jpg 442342aa75c26a26a891f32649c5fa72 80004e6b19df4167c83809f5865c829265102dc6 543 F20101108_AABFBV ramu_p_Page_109.txt 106730eab3fe0372fdfb1c2dce41a5c3 4b518103b6b50401d84a6e4532a491fa3d82cd79 8540 F20101108_AABEWP ramu_p_Page_008.pro 7c91c1ab83393edbc76f1a202bb4996d 57d6c6dcfa8351f350d430962a74935ba2bd63c9 6646 F20101108_AABFZR ramu_p_Page_030thm.jpg 3224e883a1b940462423ed6e403b1da3 4a12b9aa3fcb5c271c5bd3dd02446577880769e0 493696 F20101108_AABEXE ramu_p_Page_047.jp2 e029c60d837e915327379824854715c2 22920fc1d37dc168a84419dff226f5ca1633cb9a 110827 F20101108_AABFBW ramu_p_Page_035.jp2 63af7cc708b5085784643ea0a7b1f5a3 b9fd8b158eec25c1cf703ed9bc4c0f457663d5eb 24877 F20101108_AABEWQ ramu_p_Page_026.QC.jpg a2017bdf313065a30ef34e0e6e5e8f53 1e64ce0d8372dc62ba63835781e9ca18e8636925 21590 F20101108_AABFZS ramu_p_Page_031.QC.jpg 9a2650c55fdb554694e205496a3ecf89 df5aa1e1768dfc90b5b8e803a10a89ba85668108 3811 F20101108_AABFCK ramu_p_Page_011thm.jpg a86774f48a2142692fb0e190908909b1 0c7db9b6431dcc74e60eff362cf89f65e307b5c9 18137 F20101108_AABEXF ramu_p_Page_017.jp2 72289b1794584fe95d030dc9032908cc 85a009c954300c9fe5be88696f332c877de4f609 6919 F20101108_AABFBX ramu_p_Page_071thm.jpg d5a118326c81e6986922050b3fb9ddc5 f6e960ecf123c9035d9d97f4109103dc242973ff 778 F20101108_AABEWR ramu_p_Page_102.txt b813322869431f324f9039671e848af2 27e49c0d513ee3865f391c26b0ccbebbfcbcb960 24136 F20101108_AABFZT ramu_p_Page_032.QC.jpg 43842f95da55b3bdc53186da81ba8370 315c6548cb76d3194fb37c671e9e727923b8e43b F20101108_AABFDA ramu_p_Page_105.tif b2d49c21b318885516dd36b07f46a34d 7c9dc3bf09004145925c7f36af1ba94ece8f9bad 53029 F20101108_AABFCL ramu_p_Page_019.pro 2fe407f81fe03f304adeed93707664bd 87835739e54ca83d2ee83cca4a7f230de023cbce 11430 F20101108_AABFBY ramu_p_Page_087.jp2 f4ac62352bc7f44ce1d0dc3e4517b5da c02b62a681d963eb67e2770f9531a25f6d9c2395 5199 F20101108_AABEWS ramu_p_Page_096thm.jpg 932ef027b38a8d02269548e74d4dcf12 b9e81e30d474d58843dcca6442fc3211ff80329a 6560 F20101108_AABFZU ramu_p_Page_032thm.jpg 43c5fb831424eadcdab27d29cb784dcf e9702d157b5a6925f36158b135f16d738c454d2a 95091 F20101108_AABFDB ramu_p_Page_023.jp2 671a8a440b7dd42669a4489f574ca680 04ee6ff674cf1c75324eb10626e01ea4885655b2 360 F20101108_AABFCM ramu_p_Page_066.txt 4155c668403a113e5f188c4a58b28b35 a7bd9a75dfde10866d7227cd050e816efadeb1bc 56782 F20101108_AABEXG ramu_p_Page_027.pro b63223e70957c268bb70c9d1cd6ea591 61a1b9d2554833af6a03d29372d88de62d19ace5 111850 F20101108_AABFBZ ramu_p_Page_019.jp2 14e913c3452e5b463c219ecd38266655 d8e6d3ba11285df319165162527c5fd5bd9e60f0 736895 F20101108_AABEWT ramu_p_Page_091.jp2 455e34a518be21de04977261a25e30df 4fe4f86f74f05bb8b1676f9aaff1a3c0e6c7f6e3 20505 F20101108_AABFZV ramu_p_Page_033.QC.jpg 6bb723551bb5d82e472bad9bb6feae35 310ba8473ba046a0e5a858a961b73ddbc91322fa 1098 F20101108_AABFDC ramu_p_Page_114.txt 051d65b75a15a53c45f255555d872b51 0b1cdf3783587cc71b59d26d7bb6591e981de038 44569 F20101108_AABFCN ramu_p_Page_034.pro 978bd74e64e873f3667aced9fcdf279e 0ab46af5f6899d1f2848e058f7223f80a5b8e1fd 1064 F20101108_AABEXH ramu_p_Page_113.txt c964ebdabb12c5f090bc16986a67fdd7 2e8d60b7b7a80254841f60c57e59f5502fe3b4d2 F20101108_AABEWU ramu_p_Page_045.tif d927e43a1442e40dd9acda459448877d 97e67dde7a78b826e844c5614ae2e41e7f9a5c75 5505 F20101108_AABFZW ramu_p_Page_033thm.jpg b95a308a8a4ca5d34db56e25b5d53ec6 5d63784f636c7e9d020dff2575ba17b3c9e3e175 25265604 F20101108_AABFDD ramu_p_Page_112.tif 60f815a20c0a4d107730489157178ce2 16824a5cb0b254a57abe80fb42cba0177b8984a4 4398 F20101108_AABFCO ramu_p_Page_068thm.jpg ff6b140e38dfb10f3c803373d0432609 2ddd980b38d33f48dbe0c129cb1e1b0b29fb2a5d 25694 F20101108_AABEXI ramu_p_Page_028.QC.jpg b6d85ca151b04d89e1edd82a9e4064e5 5cae9ca793a6d674d9d42695fd8efd9d4f5de301 6326 F20101108_AABEWV ramu_p_Page_036thm.jpg 0ec2376a75cd0a071dccdd0bbddb9055 bcf71e92f1958b9ff353609763d27177f1adbb1f 6080 F20101108_AABFZX ramu_p_Page_034thm.jpg 7fdd4937f468855e9fd798997c52f301 3573fd2adc03d98c38e1346d90ffa4571b46dfe3 20680 F20101108_AABFDE ramu_p_Page_004.QC.jpg 2efe976eeddb18fb0e8d764c0cfc61e0 9ac6f56020fda15511bf2b41111f41337b606d48 20901 F20101108_AABFCP ramu_p_Page_074.QC.jpg ffc7430bc8cb2e331e4bb22489c11f8f 43bd50b3fabfac03491ec8b8477f653edc7d77bd 23297 F20101108_AABEXJ ramu_p_Page_007.QC.jpg 460b552eeab919ac153165168062a81c 42c80024ebb1e0740e8874ff755afc9e91715cef 77965 F20101108_AABEWW ramu_p_Page_016.jpg ed4ffdc0748bfd74459b7f3e169c2bb9 01ecdfa48bc3450acd70a9546f8b75cf1e508657 24641 F20101108_AABFZY ramu_p_Page_035.QC.jpg e77f5262c499221d39e7550b18700cdc 5b863e759f7146f7c16ace1af6a05f55d2fa09e0 72494 F20101108_AABFDF ramu_p_Page_005.pro 500683e9cc6b39a9ae5ff3f747f18da7 139266ab5e112cd9aea8ed50b7949eaff7da53ea 2253 F20101108_AABFCQ ramu_p_Page_044.txt 5e49c2050134a292a05b4d42538083bc 4a1d283aef7f6d403ca5623df48516e86f8fc50b 112131 F20101108_AABEXK ramu_p_Page_058.jp2 57192d7bc6fbf85ee678e41292574afd 36f74656df3bf8cfd65b9e08c0332bc2a19a44da 7224 F20101108_AABEWX ramu_p_Page_061thm.jpg d4edac7e90d6ae8d497e23169732e583 8aafbc3f1a850d7e09bbd0b5bce34d2f5eda2475 6689 F20101108_AABFZZ ramu_p_Page_035thm.jpg 34532fd2155d6141370277756addc2aa 522948c5d22709aad40bc804f293d3d1fb0e1354 F20101108_AABFDG ramu_p_Page_101.tif 4b460b5d7861ed9c72de7293ac51f443 839b6aae18841d0ae53d119605c7dc5fc998df4a 69476 F20101108_AABEYA ramu_p_Page_032.jpg c8c405312b29b1b11cd71bf18d8c2f61 6b3577e7efad08a82b3d616b0059e2497cc6e127 6327 F20101108_AABFCR ramu_p_Page_101thm.jpg d8ce2f8f133df256e353ff31a3c66e2e 14e1287894ddb2bedb11f89a8c5355a0dfb5fcc9 22409 F20101108_AABEXL ramu_p_Page_005.QC.jpg 7991a9ce962fb0ffef00aa342cd5ecbf 0769ea4746b9353b1fa4e44f0fe26d902dc31cf6 1051977 F20101108_AABEWY ramu_p_Page_005.jp2 2201f37d77b5ae9530666bc448ce5cbf 452ccb81d28f992d1db1af5157ff24722edd8425 5459 F20101108_AABFDH ramu_p_Page_006thm.jpg 14e785b8e62b50676844fcd40edfbedf 60164aff300248444e55b78ff9805d393c7d89e2 20205 F20101108_AABFCS ramu_p_Page_095.pro c311d17c9b595a836421af4af14d5247 4a216e01cc4a037ea7e34875d5d4b3893ed43e4b F20101108_AABEXM ramu_p_Page_109.tif f39cc8cb33a80ea383ab444bd7be22ce 9bc019c63fd4bccf87df265f5bad0f12d5eacc0c F20101108_AABEWZ ramu_p_Page_017.tif da8aff9f7de7cda89f0634832ee11b32 487bf16c9024a64ce0f0f7bb7b5533db2b05cac6 6713 F20101108_AABFDI ramu_p_Page_079thm.jpg 870b38a6df646b0d439b4636f1920bec d7f609e51f7f8b8dfbae7263b433390d1c4a4cfa 85734 F20101108_AABEYB ramu_p_Page_005.jpg 361d26f08ec7025768862a11386d63a8 fff7716df94cc01ef0b2714be0bf009b2626bf4d 8142 F20101108_AABFCT ramu_p_Page_081.QC.jpg 5d2cad97fb12c382fd4b0aa9e7566574 931f3593ef2d734e5b64607958fb58a10245a2f5 2011 F20101108_AABEXN ramu_p_Page_040.txt f38d2bddead8f3407659222ea113e9f1 93378f20854aa4b59923395c96065f70b9aaa96e 700899 F20101108_AABFDJ ramu_p_Page_065.jp2 ac401741513fab12f6fcc5049ca21696 5af1502053278f08c83246dddb599bcd7eea58c5 53819 F20101108_AABEYC ramu_p_Page_028.pro 7db71757626dbcb9669ad5baecc66ecf 726e33e507342aeb1be042f23b0da38f8d23d518 25987 F20101108_AABFCU ramu_p_Page_123.QC.jpg 6d773e751be84b9abd93cab1854fbfa8 71e6e2dad585b45e8adbf25f6a2fdb1217205c9f 25424 F20101108_AABEXO ramu_p_Page_089.jpg 32335a48e62b35728bbaeda993414313 fbdc666af73eff550c9284fe7f6399fab40877bb 43524 F20101108_AABFDK ramu_p_Page_106.pro fb8d6db4b4dd54271f7ab079541a9b37 e0cbf69172998df6ba440a524d994827e05332a8 57576 F20101108_AABEYD ramu_p_Page_061.pro 6515133ccb2b803444b46e2a2abc54c2 f1723619ed32a80a4879a555309d213ee7d4ec14 15845 F20101108_AABFCV ramu_p_Page_113.pro 4563d0f0d98ffd9fb8f97a27b580f6f0 57d04024ba6df44a7faf4c5de357ab661dcd3478 F20101108_AABEXP ramu_p_Page_030.txt 48c0491d7301c4ced68667064a951ff6 db28ec20633917995ea59bb52da602143d8d97b7 F20101108_AABEYE ramu_p_Page_051.tif 2fb50ed6ea3a6f2a10627c52dfac78ce 8ed3b8030dd0fa2590ec9411bb1f1fc981d48800 703846 F20101108_AABFCW ramu_p_Page_109.jp2 3090766756c80e29bfb66fe3cf7dd1de 4f0e32e3bd887ae461452067a84bde0ab3cf0a37 13463 F20101108_AABEXQ ramu_p_Page_099.QC.jpg 4a09a3ef1ec11e15c4610f095854d702 fcbe713814917257555b577924f101e730020176 8647 F20101108_AABFEA ramu_p_Page_070.QC.jpg 797f628816c9004e8f3559138817d1d2 6f271e5d76e9637b3ead20432e5665862efe7cfd 69474 F20101108_AABFDL ramu_p_Page_042.jpg c084c6f0b94518a063b3395054d430eb 8cfec70feb01fd665a3d151c61d3ad2095a262d5 F20101108_AABEYF ramu_p_Page_106.tif c56f8648f66b337af1a311a0a612282d 901edf87293249284e98876724c604fcabae3138 1374 F20101108_AABFCX ramu_p_Page_003.pro a87ddd6fdc076b26bde7bf382a150b8f e379b9d2bf3f324ff910e172d7fc79a4860d1a08 F20101108_AABEXR ramu_p_Page_011.tif 1202d4006a7c47d3b30a9e5ccec20f48 d1b482a6e2ebbb706e1639d65dc36596277a42c7 117761 F20101108_AABFEB ramu_p_Page_028.jp2 dab02a12ad7b5c5669cfe70f1abfd63b a4de3338e97fdf32af31a3d5d3b1ffe0ac9f6fe0 72197 F20101108_AABFDM ramu_p_Page_021.jpg c082489d9081abd5c13f35f5aa879fcd 697ed5e0050cda9d6b9e18dd19fed88a764df0b2 3348 F20101108_AABEYG ramu_p_Page_092.pro 5d321ee63f839b4a8d7d2e0b5ce20be1 f83d57319bb649da62d9464e29fd1e7dd9b3f785 64744 F20101108_AABFCY ramu_p_Page_106.jpg 82f83bcbd3e1801506a3356f3185420c 93bf8a3c1a156d869f901d00e30d2bf8656380bd 38652 F20101108_AABEXS ramu_p_Page_125.jp2 0b12b3cfcb3bb98e505a8139d9533cf8 7e6fcdc7c55cdf1d10feb4195ff4d7efc2b889ac 18000 F20101108_AABFEC ramu_p_Page_107.jpg a8f09950dc29f85d58a69798eeb66556 5ef153589aaf6322450d74c0f8d023615bf75095 34987 F20101108_AABFDN ramu_p_Page_116.jpg b79d3d2c692e874d8136958efd2ee0e8 c39a47aa32896dfd4a6a85e37d273ec8efca9b62 38758 F20101108_AABEYH ramu_p_Page_103.pro 254eaf506dfdc760b9ca349515951e3b dd080f56a6e3a14328111555efbf5c201b59e467 2802 F20101108_AABFCZ ramu_p_Page_070thm.jpg 6de49b49b7cc49ba92a95d36cebcc62f 403b71515fa27a51fb7a19a55f90d32d8bf36490 1899 F20101108_AABEXT ramu_p_Page_103.txt 64968f20b7ddbdb8bf8257eaf8c4bf07 aa1cc134353de9d77e28ade61640ef8b3617ec65 6941 F20101108_AABFED ramu_p_Page_108thm.jpg 89768fe65ee03392f14384dcd3432716 44abc99b309787f52c3f2733f45293c2a77310f4 F20101108_AABFDO ramu_p_Page_058.tif 2e8d61489affbc0f4082dac5bbbe6a39 1da77da5a65c2e681042a7bc99c422c7188fc973 6918 F20101108_AABEYI ramu_p_Page_060thm.jpg b359b90919d3a2f54f6db1b4542c80a8 339aff6ac2a7d3ee6d7cef46d0848da7419d7779 111852 F20101108_AABEXU ramu_p_Page_062.jp2 6915baad842e93211c668a6f1cd23f62 970c74288414dc406f13f8ebfe237e3c4a39c78b 73583 F20101108_AABFDP ramu_p_Page_093.jp2 589781ea6aef078243b511bad6d939e3 fbeec646ec82fdca277f45112e1072c856b2582a 1971 F20101108_AABEYJ ramu_p_Page_100.txt 7a58ff3b69766fd2426804d4e85b2763 e609d0d13b403dd625ca7c65bcbca2c7c0ab284f F20101108_AABEXV ramu_p_Page_009.tif 9b7badba3edb8aec6d198da8dff2227b b9f76ce3d471e979b9706b4996b141b4e8d73c2d 27597 F20101108_AABFEE ramu_p_Page_097.pro b2deef5b3c76c5b1179d8a9f499fd990 12f6cfff64d0cfca451336a110dadd6fbbee6d8c 14053 F20101108_AABFDQ ramu_p_Page_068.QC.jpg e220f0490ed7911c1db14be817a8dd4e 2b5078301cd84005a37632ceb1c36240a5351829 F20101108_AABEYK ramu_p_Page_115.tif 902cf97c01bb1b55dd78f7e3f226adb9 a1a27c2622cc11f55c666b3806b3592cc29fa9ce F20101108_AABEXW ramu_p_Page_070.tif 5916d2b5234e4bac5dbda59bf539f518 565ecd639f03cc6cfb3f4bbb75e9b1b5a7001713 F20101108_AABFEF ramu_p_Page_053.tif 8502b5e13aca35327bac436b0ec160a0 5928eb0a903948a0badf902918b5622da0e7bc64 F20101108_AABEZA ramu_p_Page_019.tif 76350e194425aded87323adbbc75acf1 3623816a22fd546f5c305e6fa754abafde8fdc3c 1859 F20101108_AABFDR ramu_p_Page_023.txt 37416935f46673d028a52539025875b7 40fb332d34dca08583d06ca8fba7c7277515442e 2355 F20101108_AABEYL ramu_p_Page_006.txt 77168c52aa5a0fd94f9535c22937050d 05c836d2fccb1ee737d2391c70be841545b85bf2 115963 F20101108_AABEXX ramu_p_Page_029.jp2 d819fe6f751b95aca99daeba2a1bed90 4b8e70d53ce1d7416b5e6915e18df6f9ebef23d8 F20101108_AABFEG ramu_p_Page_003.tif 776bf67aa01c1374c87fc0b2612fb513 9d6d4138b966d16b86df893211f02dcabcd22f46 35105 F20101108_AABEZB ramu_p_Page_059.pro 559e34a93024fdf85bbad83a290e7b24 0cd4a35ec00b410b4a1fc8d7bf656636a769335a 1346 F20101108_AABFDS ramu_p_Page_098.txt 7297efde899ae1e6735052fc4d7d8b6d f9f262289a473744484e159a12e4af810153bb69 16435 F20101108_AABEYM ramu_p_Page_119.jp2 fc76243f0ba49448efb308f161820e02 92ff5eaa9dd25be43c6dbd0915f7d8c7dbf05dc8 49584 F20101108_AABEXY ramu_p_Page_021.pro cbc6f9a9f8208ffd22b1b575855ef906 2406bf30ece9e469b8a16dad09ebfd2b173c4910 6185 F20101108_AABFEH ramu_p_Page_119.pro 4d6170723d58def8a87083d6998c0b01 460aeaf336e5c18c207317baa27ee73f4509de40 12233 F20101108_AABFDT ramu_p_Page_011.QC.jpg 92f7037fb0f27c15de0900115d0b8dc8 9fd1a17d43f4ae47302b3973da51d7301175e4f5 7096 F20101108_AABEYN ramu_p_Page_121thm.jpg ab98c5af3b1b2cf71fcac4844bfc89a0 6955a29b1dc00f0d6fb6867ac405737eed8b0f2a 5762 F20101108_AABEXZ ramu_p_Page_017.QC.jpg 7ee5304aa356d2c78841d191a8626156 9b267da6805d5d882c3032b6b4ade3e9e8063495 38837 F20101108_AABFEI ramu_p_Page_047.jpg e2f08ed39af3daef8174932915afdd79 653ffcf7445801fe2fb4d008670fcc8a150a2f85 65747 F20101108_AABEZC ramu_p_Page_052.jpg 1531844c0d0226a0d344fab2b4e609ef c1b6291d2c411d55402fbd74d22569168b7177bc 50350 F20101108_AABFDU ramu_p_Page_077.pro 801381de4fe94063772a344cdc18f9fe 5e3fba4d72e4a522eb22cbe1c2245005c937e073 F20101108_AABEYO ramu_p_Page_085.tif ab3cbdc0eccebfcd0cf4d1f357abf355 0b3d6285f93657a884738b58557a4f71cdca8ebf 36312 F20101108_AABFEJ ramu_p_Page_102.jpg 9943a4828580756babbd0ada0efda088 c03c1f07e512060a7580f99cd106664a3f9ad0a0 120857 F20101108_AABEZD ramu_p_Page_061.jp2 4cf65e888d04f847df6e141057ba5f82 fc372971691a2e7cd58278f9b3cd09dfe2e081d1 4155 F20101108_AABFDV ramu_p_Page_087.pro af46961208d26b6e8c81fcb96c2b7619 935b0ab3791da107507131f8ee0c472906743301 6015 F20101108_AABEYP ramu_p_Page_007thm.jpg c1f6e1b5f2be6e36ca019dde644be7d8 bc42a5ccaf7894861f93a646b276564fe1fd35c6 7006 F20101108_AABFEK ramu_p_Page_014thm.jpg d0c647cac12e8aff03006f41578719e1 57647ac59ce62ff781183ecbb48beaa15a0075cc 96818 F20101108_AABEZE ramu_p_Page_049.jp2 1c01baac262414977141be62a3b715cf 10762e2fedd9dcd7375695128506b43415129299 8336 F20101108_AABFDW ramu_p_Page_089.QC.jpg 300ef5e3d90430864f2b448e8807cd94 daebe3772fbe2ed54688d4c60ca26639282877ee 22797 F20101108_AABEYQ ramu_p_Page_104.QC.jpg 7926f2b97a5a3e17e0869c1f2dc3f012 db5e795414c57bc29b0f5c65a2a0688e2041e2b8 24047 F20101108_AABFEL ramu_p_Page_021.QC.jpg abe99acd8ade555c740cbd7120119816 89c2818803a8cf3477a463ca87e0faef365b6a22 24950 F20101108_AABEZF ramu_p_Page_041.QC.jpg 3a8cea43c27c15584a3027768cd7e156 db632384ee5da681f3b7c7154005456aaff1d921 55752 F20101108_AABFDX ramu_p_Page_044.pro 89a8d111a6462b434fff4006a84c6737 6cd5131e22daf9db39ed22e022090c834e047779 593407 F20101108_AABEYR ramu_p_Page_116.jp2 d8224e319904c5f12c95e613a53cb346 4e261cc0ed2862fac0b2d7842f14a889f6100d2e 60676 F20101108_AABFFA ramu_p_Page_072.pro 9f1b4a51e337784d86634285241926b3 0a372be3bd56ccecf0c0f867027e5b53ce8694a7 25048 F20101108_AABEZG ramu_p_Page_072.QC.jpg 3e8b92980a8e2d8009312a63c6090c03 ab82a4f9c6ab3c450ef8b427cc95f016d44af220 23259 F20101108_AABFDY ramu_p_Page_083.jpg 75db753764084caaf0beade5ed1fa84d 8c220701f92fe60002a9ea4c947c02d90a412fd7 22817 F20101108_AABEYS ramu_p_Page_056.QC.jpg 6d54d422e56d91f65aef65a128c54e8a 09016a085b64b162bd7c453a64b2c0f2a8705cf2 2211 F20101108_AABFFB ramu_p_Page_094.txt f16e5c735201678cb2ad1db8bf78a1c0 eef278d246b3df6defa7476a832d9b637ec60225 2038 F20101108_AABFEM ramu_p_Page_076.txt 7260e1f5e49a31a108726c348959dd5c fbbd421a5d6031b423e4fe21102492b8424b1b0c 2643 F20101108_AABEZH ramu_p_Page_009.txt 63b4a68a759eea2c97b531cf23116852 ebcebd1391b1513138290edc40e008d3d52add9e F20101108_AABFDZ ramu_p_Page_022.tif 2c7415ace502f04d5336f95cc622600c 277012be99880cb015f5e779fe88afe0c65e5e30 56694 F20101108_AABEYT ramu_p_Page_007.pro d73bf73c9cc4b010b106a4d6ef1986f1 15cad7eece8afade1a3b1e054f00541b4a57769f 2108 F20101108_AABFFC ramu_p_Page_008thm.jpg 9a2bd211ecd4eb900ed96b0ad6e94376 7a9276adc2a49c9c8396086b5485593a08fe6e0c F20101108_AABFEN ramu_p_Page_033.tif 7af67d7fdc5fad2cc16f2e2382d8147d 82be6b69c33792d7140f689bb19ffa8603ad30be 50607 F20101108_AABEZI ramu_p_Page_096.jpg d366931ab9d396860a4c6b05365e4471 124f0219fcb4aa7b73765ae5b87865481546116a F20101108_AABEYU ramu_p_Page_120.tif 4262faaad1a724ca97800b76151d810c f757bc7797268d8c93edf4b94d9825888cae24e5 6186 F20101108_AABFFD ramu_p_Page_023thm.jpg 1b594f2e04266f88bcfbb42897c49fe1 95375fe563e039cfc2f4f49f08a4134072d472f7 39290 F20101108_AABFEO ramu_p_Page_086.jpg 137ea3e278cfa452cc666bcc7f3008d2 083656ea23bf5f14687a634a12abe7d4db1d3d73 24853 F20101108_AABEZJ ramu_p_Page_090.jpg 7a64b344f60946389e2d8b323d0b799e bfa11e4c10d5b2f1c06c59f513f5db4848b498e9 58479 F20101108_AABEYV ramu_p_Page_123.pro 6f070d8b1596cd2c869599c55a000a91 80db7f5387a2f1c078995a96ce1272a15e2c4240 70513 F20101108_AABFFE ramu_p_Page_104.jpg 7ebf586ade7d4ddd4870e011fd3b50b4 26f8f806d1959e7f9ad1da3f8fcd444a273d538e 99417 F20101108_AABFEP ramu_p_Page_050.jp2 ca614c523c53f28f67ae8a2640f19360 d7aaf08f8507aae47727ff422d1bfc0d65503ffd 51073 F20101108_AABEZK ramu_p_Page_079.pro 2110aa0500818a5f0af418a39eb157d4 4a3303f975ccf5ecec0d7effaeee15f43251f7e2 2095 F20101108_AABEYW ramu_p_Page_024.txt d9d6555858f910b52647950f65361251 122e610101d319ff8f33dee592ec02863b658251 24415 F20101108_AABFFF ramu_p_Page_053.QC.jpg a3a4258967af01ee0b3f5194b8996d2a dfd2c1e8337daa602ee4cb9f4f48c442403a118c 1051961 F20101108_AABFEQ ramu_p_Page_009.jp2 7aaa0c03e32ef158d9ad7bd6ee019f00 e07c0d909b161672d2d472e66e0c74ad1ab35ac1 13222 F20101108_AABEZL ramu_p_Page_069.QC.jpg 3f082d2e5bcbafa15ab62e16c06db9a2 588bd4efcc9856a921be722bd8ba6e009fb03d9f 22607 F20101108_AABEYX ramu_p_Page_040.QC.jpg 048951ba6ade64674bc717e5d5e7fa6c ef567d6fce5d9d01ab5911d4cb3e3f79e8464f26 46900 F20101108_AABFFG ramu_p_Page_050.pro 646d90a5c565747c04c63c60d06bed04 ce1ae4d7919f488430bf4e598e36c200d5282de3 50714 F20101108_AABFER ramu_p_Page_037.pro f6d1b7e50d8bb9a2a04e900ff3371009 f541a98b6a3b4e609415116269410de64d5ad14f 634 F20101108_AABEZM ramu_p_Page_125.txt 20128324c79e4c0657a5aebc2cc2e95d adf9bd50071647bde5e11e07743649102fb814ec F20101108_AABEYY ramu_p_Page_123.tif 6273e7730b68ce4f6000d44e09a0868c 3413274a5d5ec135b40337f2fcd910210302200e F20101108_AABFFH ramu_p_Page_010.tif 0ee813baa61e7c8f80047f065a8f1b0a 0d690fbff9e811903b4faaa3846f74cd2d5097e9 6190 F20101108_AABFES ramu_p_Page_049thm.jpg 3b2cb44ae3734b789283348cb12f9a95 accd3a24c32c111e278fa1b7c16615282a1ca8e5 F20101108_AABEZN ramu_p_Page_088.tif b792eb5869da6eb1c189b71bae434b37 b3357a803201ce0b39c7d9101e46e924752f6548 25513 F20101108_AABEYZ ramu_p_Page_016.QC.jpg 349da5b95ce04661c4e408da047fe88e 17a3000a082f7f01ce0be2d867614ac40db7830b 1908 F20101108_AABFFI ramu_p_Page_034.txt 571ce7a572cf6c9b901f81c18f750819 3cc986763d175a41ae329e89b5f09ec9299533df 71266 F20101108_AABFET ramu_p_Page_079.jpg 2c6307568b36ae24a1ac4814bd0d8ae6 1cf413686e7a3f0b5ddee3c8f48e785e6581765e 6493 F20101108_AABEZO ramu_p_Page_003.jp2 42be902cee246a966a7445eae1ad02e3 89d09d48315b410d6989dd8896b2b8485c0a20d5 2204 F20101108_AABFFJ ramu_p_Page_019.txt ae430a98e046973780e182d8914491df 8606b99e553481cd4483387f74b6c726566573fe 3177 F20101108_AABFEU ramu_p_Page_115thm.jpg f38e0a4311b8cf4247c4232c42bec6e1 ce8dd5b770f26633a4db34a025c68d88e0e182d5 757 F20101108_AABEZP ramu_p_Page_045.txt 5eab1d9ca3a9d4b04a0e49fde83c5815 29d50a9c986ef3d765f5dcfda2cf5351f9cb3623 24197 F20101108_AABFFK ramu_p_Page_073.QC.jpg fd81174c1bda8056782a841868ee5e3a 19c44114df26c81a17f45ad08d263db45fd6db0e 52343 F20101108_AABFEV ramu_p_Page_054.pro f597c1ae5fc9c1008dda2e4637ff24e3 55f517bbdf7cb0d8e6ff4c462d796ee85b0edc77 22767 F20101108_AABEZQ ramu_p_Page_023.QC.jpg 6ac1336775f2640ee7c1c85e64647ad5 954f361aab8e2212c42287520d66af06862cbb90 F20101108_AABFFL ramu_p_Page_064.tif 312e570b32e0fc2415d430d4e916b5dc 39dec70fddbc275d87bc0ee2d6305799015c0559 721 F20101108_AABFEW ramu_p_Page_070.txt 9ff48ea083d2f72a3fb276c9cfae5d67 79c7be504c395b7412b0e43fb26ee168c8f3d2dd F20101108_AABEZR ramu_p_Page_021.tif 03ffea76b475b9c1eec0731b6e7f54d5 a5d8b9853d78cde25c6383d411b2e47caa98d20b 49728 F20101108_AABFGA ramu_p_Page_094.pro 8add447af289d867578fc38de1ca125a 9f62dca0a482e07d0421fe8ebdbb5724130daf1e 71153 F20101108_AABFFM ramu_p_Page_022.jpg 0cf08747cca571e3e3ae15711c28eb20 2aca02104fe64d70a21debf955a760e5be9c568e 6304 F20101108_AABFEX ramu_p_Page_100thm.jpg 57a9bb65913ab0d5240f5f8fd851e8fe f75e6432594e7221481771727e49149f313ab73d F20101108_AABEZS ramu_p_Page_005.tif b3f297b41edecc4769c0102878e16c86 43c4cd0d3dc77e3fa5fef6267fe2b2a63b37ec75 16156 F20101108_AABFGB ramu_p_Page_114.pro 85435353409382ecdd37735570319c28 5244c7a6a43f1b03d80e5de788eeb02a1888ae77 25124 F20101108_AABFEY ramu_p_Page_063.jpg 9814df141b84024bf010a8eb6ea9a9a9 6e08f5c109319f2bb7920b3ea054c8e0b0713ae2 19549 F20101108_AABEZT ramu_p_Page_012.jpg ec7f4785fb63c5503df20b78289bdbe4 18adf04d8b48e81b9e737f0a54ae2aba593b2cd5 2008 F20101108_AABFGC ramu_p_Page_031.txt f1752fb4534a09e66a74e9da99beb39e 752ebf7c85a6ba13cec64788e29891e8a9d23d5c 19242 F20101108_AABFFN ramu_p_Page_068.pro d4dee9936704e71b399da97218b044fa 6159809e301ff8948bfb1908c97678cd10082285 F20101108_AABFEZ ramu_p_Page_037.tif 60a91bc0aadcb1e8ef35bbd5fc5cf7fc facaa59edcc2a0ae71a39831fb3878550dc2f94c 16050 F20101108_AABEZU ramu_p_Page_125.pro 83a070babdb79abf4d4e4736ec4cbd52 40da8f733d05fd7818fcdf0b65d3dbcc30a02350 75822 F20101108_AABFGD ramu_p_Page_108.jpg fc72ae4f7d22e2a76ef334a10230b5b4 ba8e23b5a0ca0a088d200abb74915c3b70df489d 2125 F20101108_AABFFO ramu_p_Page_054.txt a45e54408a2ba88461ffff77deb6ceca 26d1aa27157e84f38fdcb4e06663cfa7f1549db6 6545 F20101108_AABEZV ramu_p_Page_078thm.jpg a3efe280e19ca2fa536c017913c8e700 39872f7c577a667fc68354dfa95f6202e24b6b9f 64006 F20101108_AABFGE ramu_p_Page_023.jpg 11d669b646e5b81e4712bb304fe5d4b2 6e545cceea02fe77d5efacbad618e50cd6832be7 913 F20101108_AABFFP ramu_p_Page_126.txt 02dbef41d3a55236094d5036f689df82 b88815d098155ceab5d42942fd289d038a78de50 51429 F20101108_AABEZW ramu_p_Page_108.pro 1b79f62445df7a237d77e1c326d562e7 0b3fe0f46f7c06dbb534851bf9a3024c4d772c65 2573 F20101108_AABFGF ramu_p_Page_111thm.jpg 9909a8bcd61eea48ab8fadffe9dc78cd 643e55bb7533e6dc7f547bb9293f455b2cfe7bd4 41284 F20101108_AABFFQ ramu_p_Page_095.jpg 5a3a07942c0ab9f70050f25aea7ebcd5 fcb04282e626a2142322e04552ee7c045aff9dbb 1986 F20101108_AABEZX ramu_p_Page_025thm.jpg 19d0270296fe80b9fc603f84ed03e343 a704d654c059df6bbb443282a314992c5560e3a8 F20101108_AABFGG ramu_p_Page_126.tif f4ce1911ad0e6a6434fc275eab2930ca 8f1cde8d3a99c29ad1685bb231df6cb3385ba878 20924 F20101108_AABFFR ramu_p_Page_107.jp2 d7d108a067dcc918841f59aef1905755 a45070d9e4d1b9b115f07d07757770e946200243 78386 F20101108_AABEZY ramu_p_Page_044.jpg e77626bdc0ff724b1b3c0c59a9159fcf 5fec8858aa432adeb6183d3f0f8e04827d4cf7fc 22809 F20101108_AABFGH ramu_p_Page_052.QC.jpg 994405579d524ed13e5a2d508835dbaa b46d05e9f971d4f0f743bd5e9a3b6614d6b22886 F20101108_AABFFS ramu_p_Page_114.tif f1253ba928c5508bf870cfacad7cba41 8c8ada306a974c393e09989b809589621c66a9de F20101108_AABEZZ ramu_p_Page_036.tif c66f1cad3801707b5cd80ddfb3da71b0 3dc29288a5ce04d95c649f7caf0cdb6a4ce0c849 22068 F20101108_AABFGI ramu_p_Page_089.pro 0b418b35624bf0f869acc17163be5158 aeead736d823cb7377fe14dd2bc64229a801bf6d F20101108_AABFFT ramu_p_Page_032.tif 0bda6cb9613616b3521a8ef25f2a7473 c5ad31cbe0f03c348567a447ffabe31fd9289f99 1956 F20101108_AABFGJ ramu_p_Page_013.txt 407afbdce488050dd1076e1c5d6a9eed e03a40a984c1b45dc24576844ec23c958d9a55b8 2171 F20101108_AABFFU ramu_p_Page_022.txt ba6bb59726fd3c0be2c0bcb3dfab3c25 4cad8cf5fad31e623e75173d6ce81852cab8b3dc 5829 F20101108_AABFGK ramu_p_Page_025.QC.jpg 8d0c840812c3ba725f93dd553cedbf11 135fc566ec0fc9161ff1be47a3e16d836998af05 5080 F20101108_AABFFV ramu_p_Page_087.QC.jpg 1415b17c9b9bb61ac0f67a0a033fd224 356b9719c6064bc09ddecfecbe79c5bc7872210f 1149 F20101108_AABFGL ramu_p_Page_112.txt 180260f7c634d63bb48eb4d770dffe86 605c117e8f34fb057cff535ac024caa1f5b15bbc 359796 F20101108_AABFFW ramu_p_Page_084.jp2 c4ba5ffd2cf7ceb598d69338f371e2e7 423d904efda77d1da00ce760083fe57ef99ec18e F20101108_AABFGM ramu_p_Page_038.tif c801585eced60f74191810c9f005366c 5df4272f72d107b4f6be7be1f1c29f730d888c7d F20101108_AABFFX ramu_p_Page_004.tif 7a455fd9abdcce873b5b08728862ea95 a332c7a75ad2b40b8be011ddc4c7848f6c788265 6940 F20101108_AABFHA ramu_p_Page_057thm.jpg 9b4dd5d43a62542b5f86df8efbbd64f4 ba8ac8827d0708aae2ee661081e8229cfc2415d0 56770 F20101108_AABFGN ramu_p_Page_057.pro ac75e16c4ee903a26570171460c05a1e 404e4aae8443c9f794badf18d1d8335b2aec04f7 41285 F20101108_AABFFY ramu_p_Page_045.jp2 e05598ed261476e25ed3247c5af8d7ee 5644e1c17f4185ac8cee3f0b478c9ee9e90a0a0d 1051982 F20101108_AABFHB ramu_p_Page_067.jp2 e73fe9cc9013e4205d67eda032c9f857 7d173c9ebbcd2f5ef00551f2003d0e166d14a975 1884 F20101108_AABFFZ ramu_p_Page_106.txt 0bd8b44c55b661d68d9ece389a4a446e 8001444b38e5c068c815dc6649b74bf830960767 46814 F20101108_AABFHC ramu_p_Page_039.pro 0de20b723f3a53dd8f0eee3547f4f197 a518c359af4f85edad6f74b90dbb88d788a18a73 2688 F20101108_AABFGO ramu_p_Page_083thm.jpg 6bec13579dfc8c01e958e63677aa2eb5 beec091b96483e96102c94d8abd650424c7abb88 F20101108_AABFHD ramu_p_Page_054.tif b808e2ba17265b9d050b331600fa4928 e1614bd06f629e789b10e6037f31ac462eed097d 68404 F20101108_AABFGP ramu_p_Page_036.jpg ca1253005c392bee3f9092fa5173389f 471e13b777aee26cd260e9d9746db0a052758ff4 70013 F20101108_AABFHE ramu_p_Page_078.jpg c5804b042668231b913377b05dade596 0c220e141465406634f94fe20fb6d49eb9beab8c 8745 F20101108_AABFGQ ramu_p_Page_070.pro 2d0653c71700c72a065152b5a7fc201d e51df3195d026c51540f9372089de2499dde514d 2321 F20101108_AABFHF ramu_p_Page_118thm.jpg 43e2cf6de55c435fbafbd063004a2df4 d57bcf1675a3dc54cd0370664d77c099da0656d8 F20101108_AABFGR ramu_p_Page_108.tif bd3071ddb818f7610f2da78fd7f7d43a 9323103db45469e940e8c0f1454176c6469d2d59 50442 F20101108_AABFHG ramu_p_Page_032.pro 1fce835252944586dd38bdaf55122c3c a09d52c713545b618e8a87e091c377acea579108 6412 F20101108_AABFGS ramu_p_Page_022thm.jpg 3c49ecf612ef90494aef6350f677fb4a 5d0e44e00e17f39080ee9cecab4f98c391b9a530 64578 F20101108_AABFHH ramu_p_Page_034.jpg fdfe9619e7689238515818ff17978cd0 01dab952488ebe74352b44b0871c487dfbb273cf 6985 F20101108_AABFGT ramu_p_Page_020thm.jpg f0103680d49e3999d06712e944fee103 cc49fe13b9be7dcff7466f543c78d2c04768fd5a 23684 F20101108_AABFHI ramu_p_Page_022.QC.jpg 2b7898d3fac3d46841ad69c38e8bd5bf a85eebca54e61c9eb0edc4de1c1c3753f6063c70 53262 F20101108_AABFGU ramu_p_Page_073.pro d09463b1fd3e90d176dce1760fbd1570 3f604400027dd2df47dffd4aab1ea44212b330c7 21121 F20101108_AABFHJ ramu_p_Page_106.QC.jpg e604b7083ab896d0e6f65e8771050064 a397202640c87430b11a6a1b8944a598dcc66fd0 F20101108_AABFGV ramu_p_Page_055.tif 77c420127708a9ca7a9d5b2e62ac0d5c c52d287129fddcb24ce81cdc0f6cb98e77fd87cd 17703 F20101108_AABFHK ramu_p_Page_120.jpg 32f66245eef868e5cdb0e98c74bebc58 d8c8b6980b8ae1c27aabdaf5177f099d79cbb371 3518 F20101108_AABFGW ramu_p_Page_003.QC.jpg 1b6c803faf96b23a20cc1484a276b735 175e706838d74dc8a9340908ae59cfb29ac8da83 82331 F20101108_AABFHL ramu_p_Page_059.jp2 0210822ba81510426dd15c14f24622f2 60a1b1bca96dada9a0e5bc99f5a76759e962f377 1206 F20101108_AABFGX ramu_p_Page_110.txt eaaf88db08ee50ed291186194e34c785 95008036e69a69ad46967d99617f3fb2764a9e14 7457 F20101108_AABFIA ramu_p_Page_001.QC.jpg 0f80ed73f41ca20d22d585f981d9116c fa37966959074eeccd505c8bc6c682574bb8433a 23480 F20101108_AABFHM ramu_p_Page_078.QC.jpg 8e5d57e3f9f4ce26fb44871323efa654 28fa7c5a6908aba4440ce72a8aefa95dc9a33ad7 32128 F20101108_AABFGY ramu_p_Page_113.jpg 21923d108a078c212c074155936fc273 6e4493ddf617b1d2c82e571823f0721b99ee646a 112551 F20101108_AABFIB ramu_p_Page_054.jp2 12c74cc8e4c3181759ec324c3e542bc7 a9e01158a0ed66254a4dd269d3de51f04cdee3b9 27265 F20101108_AABFHN ramu_p_Page_038.QC.jpg 7522b214105d34fdc3585e36220c0d04 64333a2ca041e2b1bbd4322ad75ffde9c935732e 6753 F20101108_AABFGZ ramu_p_Page_037thm.jpg b12bd891b081616211350c4146ffe74e b8d04634d6f5b45bd6f061d4d7a495dcc1892493 1600 F20101108_AABFIC ramu_p_Page_096.txt cdb5a6dac936e4858f04f8cfce72ec64 3ea28d84f14c075bd232ff96e375cdc807388471 102292 F20101108_AABFHO ramu_p_Page_094.jp2 844e310e6c3666d24f3e57c300a975cc 0ee10cfa52e991aa6705efc10e3a12f78953da2f 22815 F20101108_AABFID ramu_p_Page_100.QC.jpg 259a52ff7635f168daa9eaf95199b9c6 e7d8a60b6abc0db9485f4bb98c162fe34a2526ba 98400 F20101108_AABFIE ramu_p_Page_052.jp2 65c1f15b627d91c8028ef6defcd42d42 9fc70bb851449e61f75c8bb884118572a845f0f6 75294 F20101108_AABFHP ramu_p_Page_054.jpg e45f09741bff532a82d963fee317c0ce da9a120f456326a4f092cc4d680c45e8227cd981 F20101108_AABFIF ramu_p_Page_113.tif 5a40ad762758bab0c645dcae9e8c32ba ec743b6ce933c53730515e632320a74fd58123e2 761662 F20101108_AABFHQ ramu_p_Page_069.jp2 4c4af278b8ea22f37471f580035c8f1c 2a5a6b73cc24244fda2d71cb74c80fb4d583df55 F20101108_AABFIG ramu_p_Page_100.tif fe310ed45b78aedbfb1c517eabaaca05 5018b438a8d1f8462aef48c36a2fa673ad02e7fd 50991 F20101108_AABFHR ramu_p_Page_104.pro 841a650aebc58599a1fa092dc31942e2 d7399604beb0d05a64dc42d74f6114bb82e08c05 18953 F20101108_AABFIH ramu_p_Page_008.jpg 3a6aea8710a4ea87f28d54cec90bbbf5 a34699f08bf1714176fd72568a69b9743a93293e 3578 F20101108_AABFHS ramu_p_Page_081.pro 3c42ecc72849059099bdd347161e7abb cd0b4adecde3c91e22a4001e984b09591e6e433b 42241 F20101108_AABFII ramu_p_Page_109.jpg 280dbc7fbaa4970a4b4114dae74476a1 084609331069ba6475f250391f5c7a907dff3730 12325 F20101108_AABFHT ramu_p_Page_126.QC.jpg 5fb98d289a9329202dcf2a9356b98ccf 4496fe23a4f809299d2c5da6476703dd12e14d52 49787 F20101108_AABFIJ ramu_p_Page_018.pro 4a16b858a07df505d3aed983db346c1b 74f84f2578cdb47dbc7bc91728539a03311f5d39 77770 F20101108_AABFHU ramu_p_Page_006.jpg d2abb8cb6ecad92d6a5f493bf8d128a6 797f1089a8794cc1b7d1fdb9a67474bb370d9c00 F20101108_AABFIK ramu_p_Page_089.tif a1a8c55a141ac658f906b5bf3c24f571 7f53bff58e7f42ecac3e7cf34349b396acffbd28 F20101108_AABFHV ramu_p_Page_033.txt 35fbe8e827e92336d0abfef97ecd88d0 22822c36c862b0ef0af0605483cb2d5991f2426b 70755 F20101108_AABFIL ramu_p_Page_055.jpg 7b9be2684fe0f385f556e2bd0d5ce297 5e9931d020d8e5fd9664a346512e3de58d27b93b 107854 F20101108_AABFHW ramu_p_Page_077.jp2 e2bb7d56cb9995fc36f560344e477204 077177c1fa4fd52d00556d8d84ee7c2bb661a560 6930 F20101108_AABFJA ramu_p_Page_053thm.jpg 1fd03854b26697915964064355accb87 9563ddcbb726685c375793f98c01c34ccc9c3278 2030 F20101108_AABFIM ramu_p_Page_077.txt 6b1e1870ae23567c9bdba2fd320080bc 5e908107ef1d1152eed7f7e860b617004643e38c 5682 F20101108_AABFHX ramu_p_Page_090.pro fa79f490d226b7440af5ecaf4d60d51a 019b5ef4137ef8ddcec4086c5b7e5a42db962f34 60097 F20101108_AABFIN ramu_p_Page_103.jpg c98a98bfcbfd2864b9ba83f59f01ee6c c8a1605de660f19df0f8edcf9e72e517bdcf19e2 84417 F20101108_AABFHY ramu_p_Page_123.jpg 961abaa4c0356d8d683ad4bb52034634 0be705da64f232386d3bba0adcf209317abfbc78 41656 F20101108_AABFJB ramu_p_Page_004.pro 8ee7c3091266fd86ae32886cdaffdcab 31963b8cdc69929bb439a2bd38ef2954ca8f13a0 86227 F20101108_AABFIO ramu_p_Page_103.jp2 ffffa5cb453691f7944df1135e090431 89ba6ac7a16e2998b44962c29d41b4db7d78a02d 96230 F20101108_AABFHZ ramu_p_Page_106.jp2 ae91491b1e2704446980cbaadf9e5c90 d8968dfc2b88870e80073fe1f871e96324100529 72690 F20101108_AABFJC ramu_p_Page_018.jpg 3512ed27d17c785f76554998190d7efa c79d216f32459171e6da2d583480d4f0cd0461f3 31500 F20101108_AABFIP ramu_p_Page_045.jpg 61e0968d47b52bb509c42854f373b59b cf297525cde67ab39063379f6fcd8e609017cded 6360 F20101108_AABFJD ramu_p_Page_031thm.jpg 616820b10291344d6a69eeb5711bb324 ba71aa31469098c93dcb3dd76b126f55b6c85c6d 9991 F20101108_AABFJE ramu_p_Page_002.jpg 188f6ce36bf3b289ecbe947f1a2a3e47 a612c0b6faf2e177088637365c12336436ee83cd 25678 F20101108_AABFIQ ramu_p_Page_001.jp2 b698da28f255fa0249bd680502e48e4a 867ebcf1fe3782d224136b2c40f41af41acc2d42 F20101108_AABFJF ramu_p_Page_031.tif 65e2002d6aea4d401f34d3d792f56540 2a46566b152425981888c7fc9ff34a953522f077 40359 F20101108_AABFIR ramu_p_Page_033.pro 6dd637753c6f31b835a30720b7d3ca5f ed011b65961bc6b8d1c2065ca78422c1d7886c91 2082 F20101108_AABFJG ramu_p_Page_029.txt c9054cd2ac7d2803484837f897645d4c 3694cfcf26a0669439263129ea63094704ac28e0 106896 F20101108_AABFIS ramu_p_Page_104.jp2 1ba440775ad9482563bc9957258ca54c faa12ffff3fec2ec1985955c1377f1ab27df9427 F20101108_AABFJH ramu_p_Page_090.tif 106d405661bc2a31eb2ebf6d8422be78 2101faafa6729efd80acabf607aa87c9b072a161 6121 F20101108_AABFIT ramu_p_Page_107.QC.jpg a2f8c0f419eaf4daaf111e3ffea554d3 82326d343e850e914928ee952359b25ed2e2d748 21671 F20101108_AABFJI ramu_p_Page_049.QC.jpg 6dffa150dd272a55c89d09ccbd5a0264 070806f31efdb512828ce26fc18790def581d9ff 23634 F20101108_AABFIU ramu_p_Page_018.QC.jpg 231055728594dba91dec42b9873caeb9 674a8d8b66efc33db7d525065bf78ca4120ddfba 33235 F20101108_AABFJJ ramu_p_Page_093.pro 7eed6423730b078cf92f50b8132ae542 e13d0c2494d2526ffcfb1e6fb97578d28437c5c7 535 F20101108_AABFIV ramu_p_Page_119.txt 7695c4416b3e034273c653a119b3df91 57bf3c78ff77f72a7a64c55dcdc2fae270d34fd2 80477 F20101108_AABFJK ramu_p_Page_061.jpg 965993d86049ab1051a325b9c5420088 41ea353eb07d84836e955f18ec5bbb9795054cbc 5442 F20101108_AABFIW ramu_p_Page_002.jp2 29cefca4f3b0e106f2716fa7a4b466fc 09ca983698b7d0d2e323907ce78a67a6c1e13843 54895 F20101108_AABFJL ramu_p_Page_006.pro 6238497e3bea43c2f615653fd6811514 ff4ead8250d8a0df7a92113dfa07ea45d9109733 68104 F20101108_AABFIX ramu_p_Page_031.jpg 85e8b1e095a51b2d41af5cbc400ab088 ace41523a14615902903b33115e5abe443e4d4df 6352 F20101108_AABFKA ramu_p_Page_105thm.jpg 5f5d7a844198c23675cf35587c3d8422 330dfe0269d72ff700b0f6185ee58090eda2b6b0 22031 F20101108_AABFJM ramu_p_Page_126.pro af36241fd7312d1bcef2be3837beb3ab f56c4ddab47370e3246b9ca05b6e7cc79e760bff 2270 F20101108_AABFIY ramu_p_Page_043.txt 8f175c795d03e6a9416ce20abc544dca abf6450c3bd9b01f93c18c82cc06095f70650770 15554 F20101108_AABFKB ramu_p_Page_064.jpg b0fff72414c66a3d88a5d07b18a0006b bb797859a14b2bf88e69804066975c82ef8b1136 1122993 F20101108_AABFJN ramu_p.pdf 7c86797fc3bad4d9f5b4cc35e216de05 b9917b32a8b27eb6aa8c1e342cd8788deb684b9a F20101108_AABFIZ ramu_p_Page_068.tif 2fe1a810a920d9574265c5ed237e9679 07073595ee09a4d01b6753c403a665716fe8ae52 F20101108_AABFKC ramu_p_Page_007.tif 2b89301b4cdb93630e0589d392a7bb07 876f7ad06a5f6b0bd547b4edb9a6ebc4522223ff F20101108_AABFJO ramu_p_Page_010.jp2 8bbe94c367d3c3531c0de0353ce93aeb edc4a96a81a5d7843160d11510c0695f78d6d0cc 2262 F20101108_AABFKD ramu_p_Page_038.txt 3ac6b9fa0d983bb6e69285b6015cafef 86cd6751b1490ba95afe5d708c43fdf18bcabf97 2144 F20101108_AABFJP ramu_p_Page_073.txt 536a1823a9526509ab280a3988c9080b 7874b660d58d5e7131b395268dff1a7000c94f13 6012 F20101108_AABFKE ramu_p_Page_082.pro b4471a5e13c95bdc548980e50e47622f 52706f96ac3d243ccb1c3847917583e5e54a780f 372389 F20101108_AABFJQ ramu_p_Page_085.jp2 0648771ba7d52ed0ede3785aa9455788 dd514b3df2ae25f6628d62fdf1e68dc73ae15b09 491 F20101108_AABFKF ramu_p_Page_001.txt 97c4d264ba7647b359affc97feff075d 195ed4aeb9b9f910fba3036ffa5b791fae66af71 92363 F20101108_AABFKG ramu_p_Page_122.jpg 5f504a2a8cc4c1e5b099c2e93cbd6b14 059d222b383ce83c8d42f4ee3297ad388e1125c8 2060 F20101108_AABFJR ramu_p_Page_052.txt 7ec5b26a6324fd7dc6a1f9f5c3b90fae d8bf14a60e71ec5efac50149457c52042f68996f 53537 F20101108_AABFKH ramu_p_Page_051.jpg 30fb1c42bf623d7e17851124c82c29af f075ba5f8017312346c79800dd845c5a6f4a2918 88052 F20101108_AABFJS ramu_p_Page_033.jp2 c103574fcce39ea5ffbf038703cd1eef 4a64baf8625814ec94a32ba4b4909f4bd7437b01 F20101108_AABFKI ramu_p_Page_012.tif cc9a5b968664123b10c3307de580cb19 e0b7414dd697309ea9ffd276df12cf835dd054ad 56749 F20101108_AABFJT ramu_p_Page_020.pro a567cfffaae136c18ef69066efb88d7d d8d2289d02849044da7a9c314943652eb2a32289 101445 F20101108_AABFKJ ramu_p_Page_039.jp2 8aa8414fdb5367ee96ecf75c6cb010b3 5f8ebf8a61b5efcbac58a69c9910bc0a673034e9 6348 F20101108_AABFJU ramu_p_Page_042thm.jpg 4277aa61c11ea13a8d95a6025180ab6f fa9c96ec6da5c7946917d283e466ab587c836f00 17584 F20101108_AABFKK ramu_p_Page_116.pro 618698313b05f0c773973790d4a4f5cc 127ad28e085d34db2f91fcee5c02f6f1753137ea 3425 F20101108_AABFJV ramu_p_Page_114thm.jpg d7443293d5d0de7240128a90365980ef 29a80fdffae8daceb61e3e5992dded386ae0ee80 1953 F20101108_AABFKL ramu_p_Page_050.txt 4e9ba96967bb41a7b4c02b1f3c790b5b 17eeb16ee071e4e7369e6d462cdabdc9c5a64bbf 49716 F20101108_AABFJW ramu_p_Page_030.pro 706bcb97c75db351ac72cab7e88e387d 3c915948ce82de40b805009665aa0564cad02eae 84651 F20101108_AABFLA ramu_p_Page_010.jpg be7f091236faa36772ce8fa749484dd3 fb2641b39fddd82797dd9153e165d5fb899b620d 96264 F20101108_AABFKM ramu_p_Page_040.jp2 8712f7259632f8d5445a7adf1e628138 02101cd7d5e5c50d6e123c9d5806bc84ee28560c 2235 F20101108_AABFJX ramu_p_Page_026.txt 3196213811c07b6b65923db440007d1b 543dcbe0d91aeda977671273f0cd26e57ccb4d94 50904 F20101108_AABFLB ramu_p_Page_078.pro 05f3d0a07b091f4660ecb717ef3492e5 f32efaa44916081c77a0fee66b1d3c8d420dbca9 2114 F20101108_AABFKN ramu_p_Page_064thm.jpg a87fc6f8a76f19b16e2a22aa97b6695c 66de1743ef364e1fe367747132e4dae492254d5e 511 F20101108_AABFJY ramu_p_Page_111.txt daf322f49063475649fabaef8025985e 12e9ec7e8626565075837da891299d41698158bd 2283 F20101108_AABFLC ramu_p_Page_123.txt b56377c0503887e3a5b95eabda2f1b40 1110580ea9e2f40bca75f6a850608cbfea7c9aa8 666828 F20101108_AABFKO ramu_p_Page_110.jp2 89fe2df59e9d83a4632579731bd3643d 4c4d08d70b7c228ed6ae72fb14743daa5604cddb F20101108_AABFJZ ramu_p_Page_001.tif ffae0a1c72ba3f44ed7e74753f35d634 a6ec47ab1847a62973317ff726a8e0b08c783ad2 9148 F20101108_AABFLD ramu_p_Page_109.pro 41e2e08c4a383bba1f63b2b76eb0361e 7e2f2ff2bc883dc59ce928cbef8e680183bc7520 8372 F20101108_AABFKP ramu_p_Page_063.QC.jpg e88b21165b6d18f7cfeb10fc98b0a899 105b578499bc60fa27ec755c9205a8aaec10a61e 6503 F20101108_AABFLE ramu_p_Page_056thm.jpg 49afc375032920e1ae7d972781ab8663 1b8a1b4b8ca2f26c76bd6cb5deda4a9e625d8136 300 F20101108_AABFKQ ramu_p_Page_092.txt e600357479366bfa7db9a1ca87f2b100 b9292fc7639bb8f955d2cfed7eb9e36c0e5072c5 102 F20101108_AABFLF ramu_p_Page_048.txt 9246b2e14e2bf216582ab46a331931c3 d41bc09213dfbb79f689fd62d80711ef6c8bebd8 2251 F20101108_AABFKR ramu_p_Page_012thm.jpg b7e247f63ccd7bb5c3715148638bc28e 289e9a19eeb4c8577803959e0e87c64b8831d9b9 37743 F20101108_AABFLG ramu_p_Page_126.jpg 1f86d83797a8cd8403747c2aa7570a46 f201280d094f32f47adf75603d43bb84bcf5dea9 F20101108_AABFLH ramu_p_Page_057.tif a02f661e167c1bdef7f9b4035271f333 a89a2971cd43d746411355129275cf576fdba934 1557 F20101108_AABFKS ramu_p_Page_091.txt d381b6afd19f08809e5c755106c4d985 d6637869d20ced433b725f97cdaad15fce615e55 F20101108_AABFLI ramu_p_Page_080.tif c519cc6e53ca76e8992a872fb9131f02 b666feedd119fdff94562db44d31f80886d120ae 47004 F20101108_AABFKT ramu_p_Page_102.jp2 beda44e0a1e0926dcb31722b36f705da 2cb78fd57b5c136dabb4f9492add5b486bd8ea60 2102 F20101108_AABFLJ ramu_p_Page_062.txt 330c3e6e33b44a2090c1c6c183569904 ef3a401a46bef1d0e91a980f81866b6c5863fced 108279 F20101108_AABFKU ramu_p_Page_018.jp2 3be4de4764f3ee07d78f93a32210412f 786b4d294e0ddfd5fc8d2b632c00d65621ade6b0 17172 F20101108_AABFLK ramu_p_Page_117.pro 2b1ffb13e6612e32d89a9c4a05395b47 fcafa3bfc90598210d08d1d1457572bc8062ff2a 7917 F20101108_AABFKV ramu_p_Page_107.pro 235d50c67089ce5c656fd1c59674b43c ea6239cbf29bc035f0c7c31a394089a66d4d4e25 2085 F20101108_AABFLL ramu_p_Page_105.txt 43452fb2f821b85b07859f3a7c97be66 64eff090b161581d0b3d3cedb668f699377ab7fc 50998 F20101108_AABFKW ramu_p_Page_041.pro dad1e8e239d86464afa23502fc5b3327 171cdd2c5c295d82b2fc0399da91ffa71792fdb4 6604 F20101108_AABFLM ramu_p_Page_118.pro 4ccf6d0cfd7365e575c6296d1dee2e74 e536f67d325159e05292a67c6c291c0bb17bf603 112053 F20101108_AABFKX ramu_p_Page_108.jp2 fd2e3cde6d30fe3e14987c74a28d2f9f c9267d2a17eaffda941d6d88b6d8405c15a7a1bb 809 F20101108_AABFMA ramu_p_Page_011.txt 8837b26083a81613d60787666700d9af 3440d6d43764a87dfb9161401b2ee73c692131d5 114159 F20101108_AABFLN ramu_p_Page_071.jp2 6a63fbc3a1dc50b7e2f9966f99c61415 a96b5a530bd5f52e501f95814788da29fdef5ca7 45702 F20101108_AABFKY ramu_p_Page_110.jpg 9af0c1e5c98d0872bf87b199f72d57fe 95d807422910d78c33d5ea555e83234eef30cf65 22029 F20101108_AABFMB ramu_p_Page_034.QC.jpg 9847556ab15d87f30f65e3977a931a24 4d0059590fd89c867007618e61e28176134bb6aa F20101108_AABFLO ramu_p_Page_111.tif fb2e61e80262ca48db879faeb50a3d12 c10bd01a32f7fc038faa39e664aa00e43b6aa927 27683 F20101108_AABFKZ ramu_p_Page_115.jpg 461c81ee563e3488f81e606558138162 2777223ebf04a57e8ea9ffddf0cce4671f291473 102703 F20101108_AABFMC ramu_p_Page_036.jp2 cb424e626fdb3b726f56002658e0c5d3 683fb10daccc77aa496e058ef894d83cd3887584 44300 F20101108_AABFLP ramu_p_Page_023.pro 35d43cb1064633bd71def902caeab31b e8ec124ff93b4ba48a75813e1230e39c9ab7c3cb 269 F20101108_AABFMD ramu_p_Page_017.txt 84bc6e5320caaac76ad31c2fb3dcf1d4 48b797cb7f27625aba3b48ac3f32711fdf9464ec 545275 F20101108_AABFLQ ramu_p_Page_112.jp2 6934d4905db9569e273bc8cb6a90f425 c677c6fee5b2bc9793dea52298e4af1f2c8d0990 63923 F20101108_AABFME ramu_p_Page_010.pro 464e15b2d1daae780455ad639b898981 11db20998a35d4ce7cabc0fb78b06ee3eceacc1c F20101108_AABFLR ramu_p_Page_084.tif 8defc7f9517ca10bc2db8048ea8e9183 bda7d7ff0b092ff278ee5b7965d5fbe4cd5866c3 3980 F20101108_AABFMF ramu_p_Page_099thm.jpg 68bb219b5ffa299ec563bdc728c39a7f 6244f0d4e6a7b77988f40177baf94880dfd13dee 55336 F20101108_AABFLS ramu_p_Page_059.jpg b870508eb18a03adb65b323a82fa169e 4ed383093bfa93694c89b4d2b97982f7508e68be 3557 F20101108_AABFMG ramu_p_Page_116thm.jpg d45965eb777ff0757497838311481013 7b277a3599cff1c7c41bd2863a9588629697adbc 70113 F20101108_AABFMH ramu_p_Page_075.jpg 9d2194e161fad44aef5c72291e0e67a3 bab5336eb78bb8b2393ff1a4b1b0a4e0a55647ea 18553 F20101108_AABFLT ramu_p_Page_119.jpg 82950bf442f1ee1c52227e5f5912e68c 0efe3e94ab22dbae29bce7e830eeeba93b9bf0dd 107494 F20101108_AABFMI ramu_p_Page_078.jp2 a4a60100357bfd83b88548d91a27a303 d0e9fcb747e4ad3dcc9f8355e9a90ba1357bdaea 25137 F20101108_AABFLU ramu_p_Page_043.QC.jpg a89152c8c31e012c9a8f9e1f441c1bf0 7a16b06736d3bbf4cbe2685d45a4ee25ced43d18 25695 F20101108_AABFMJ ramu_p_Page_014.QC.jpg 99a1c5509adb911f24745c9b98c2c5f5 3716c31917de7102c4b029db260bca5a8a4df0eb 27207 F20101108_AABFLV ramu_p_Page_122.QC.jpg bda8d8c975220578dd66e120176bbd74 63f380545921cd9235db44b8a41cff966fc5519b 49241 F20101108_AABFMK ramu_p_Page_055.pro a2ad2ee748d1e21c66dab963f2b069b6 a2e65e0c1043efc50604191e41a5f62fd5a08021 F20101108_AABFLW ramu_p_Page_093.tif e953cfdecaeebd45a22e177be8abb198 6ed15597a18fbc87326abf64ef4c6826e3de3655 117988 F20101108_AABFML ramu_p_Page_014.jp2 697131ace0c1fbc62eb111b7439b2a79 f1eeb446d3972588f2547036d1e8f86bfc67b238 24248 F20101108_AABFLX ramu_p_Page_019.QC.jpg 1d68c83db26de2e1eafa417ff14a6e3f ff5ad22992d05f78e8055d9254bfab42ec7feff7 78808 F20101108_AABFNA ramu_p_Page_014.jpg a360e11a3f34338f8994a9182fbb27fa baaa975e79af76a2a274b27bdaf4e0b464dcba91 F20101108_AABFMM ramu_p_Page_061.tif d20e0349689ed876bfb828fcb3958cc8 218620b85f4143df5d2122920eafd8cecb82494d 495 F20101108_AABFLY ramu_p_Page_069.txt 4f7373523f26b0dfb24c6acb6c45cfad f4917954a5760fb5b88bfb567e11d13bf17114b1 77119 F20101108_AABFNB ramu_p_Page_015.jpg 97fcf446607646418669e680baae5164 9e44876283b9a9e9027fc61cff67810c03fafbd4 5726 F20101108_AABFMN ramu_p_Page_005thm.jpg 79b85e568c6b8b4540efaaf599d4ef43 07a76404a117a1bdc4e5344bffe0923e258f8dba 16889 F20101108_AABFNC ramu_p_Page_017.jpg 43179c4fa4905be701ffb1bf8a48b0fa 12b80ee9e7c2a7187d8e14ec7729062e6ed07ef3 5577 F20101108_AABFMO ramu_p_Page_004thm.jpg 7605fb9a47e9e6c24460197582161227 1ccbefb46c7dc0bb956562234d6ebe3501552d26 78752 F20101108_AABFLZ ramu_p_Page_057.jpg 6f6021ce3384c2167c2421c45db75d4e df62b5d7f7ea77cc173952df327019f224a00688 80693 F20101108_AABFND ramu_p_Page_020.jpg aaa0e9ada0981d5c186c6cd4a2b0ed84 19cd7f0adc0bca6679c7eebb9491617db2f9f043 4302 F20101108_AABFMP ramu_p_Page_110thm.jpg 6a1ca328a5114d5ab587276dfd363246 2d0a74f2f474e481feb963b3cf90423e81481233 70234 F20101108_AABFNE ramu_p_Page_024.jpg 2ae48b705928c9a770a0ec9d37efb8fa 0717fa34bb8c0531c0444de2ba3dec315db75ac3 434 F20101108_AABFMQ ramu_p_Page_084.txt 167cea6d42e5213df3fd91e59706b3fe 9c496997519fa1c4a6905e7f13adba8bcd0d6731 18285 F20101108_AABFNF ramu_p_Page_025.jpg 2879f64d5cd7dd45dbcfbc65652052e3 e0e47e047fbd0ef25b29c0387b6f0a6072e83371 187087 F20101108_AABFMR UFE0021707_00001.xml FULL 2ce65870bba148b25f3baf4afa7691d8 41fe3359905f50f8146b84499ec47b238604240f 77042 F20101108_AABFNG ramu_p_Page_026.jpg 16290d140f4f0d05032c5f57734b95a7 ef8525e67331a17434129be62ae661022cdd39dc 80781 F20101108_AABFNH ramu_p_Page_027.jpg 5983d64f4e3b9cb24a16fe4cd9728d3d 5ed43b0feb63a983f5c7c6b1d62ac1c8cb8a1e3d 78531 F20101108_AABFNI ramu_p_Page_028.jpg 8a0c50a993e59ff82712453c5819885a b85723ca67521c7c36fb1c9501bbe766e6300356 76594 F20101108_AABFNJ ramu_p_Page_029.jpg a3517bbb0534b1cf33ed27c9a52e9a20 d86729f2523e51025cd1d8947fc8014152a18cce 23629 F20101108_AABFMU ramu_p_Page_001.jpg a6dfaac858dfc96a537f95581c353e91 a053f4a204bdc8a2d1e9932f6de8fbd9c4958db9 71990 F20101108_AABFNK ramu_p_Page_030.jpg a456b18239b334c98d3f8f8db438c2d6 86f2d680bf3cd31f1806037a8e4b75c669de2c5b 10688 F20101108_AABFMV ramu_p_Page_003.jpg 75d23efd080e9dea091e3113c7b61932 e3c523057a8eb14dafdcb45cf8c4d311a746b060 59250 F20101108_AABFNL ramu_p_Page_033.jpg 71bb639e9f936004190debbd9b1e23a8 e19684d9fb830f56654552018517664915ec5dbf 63168 F20101108_AABFMW ramu_p_Page_004.jpg fd5698b8fe23f1b31887f2c6488cd587 17b0d7367555d8818a8dff970974478083a8bcf9 73540 F20101108_AABFNM ramu_p_Page_037.jpg cb5f75f75e2cab5c0a710423baebf3a1 5b1df4abe489ee402f05be13ebb3c421af529435 78020 F20101108_AABFMX ramu_p_Page_007.jpg 8f6569e29220b71167ae4e9731699990 7d27214e9d21ac653948fec3e0fd4964804aad60 27373 F20101108_AABFOA ramu_p_Page_066.jpg ff9a194c83819025d98197a049d2552f 496da29c59ef7386a0ff60d0654809d36f30861b 77785 F20101108_AABFNN ramu_p_Page_038.jpg 267a12f244fd08c31297e51dbb358670 7bba1dfb77eb5577f11ce10910fc54f4efae8a45 91805 F20101108_AABFMY ramu_p_Page_009.jpg f8b1b11fc2eba6442880eb1a03d256cc 93026a3d1d5eee822a72335c8744bbb090bfd57c 35841 F20101108_AABFOB ramu_p_Page_067.jpg 5ec4b6ed30f42bbb9985d26753aaea88 d9645492c2cebaebae281f8a0ec5507deaa4e37d 70485 F20101108_AABFNO ramu_p_Page_039.jpg fc551388d23dd3760d464fbb4371a7c9 d28d190b2e5a9ca49ba6c90d155904d764b151b0 35438 F20101108_AABFMZ ramu_p_Page_011.jpg a4288527ca24d70d2e271e3a107a8cbd d3601c97d72c0de0e8c50bd7662b6fd39c1ec580 49469 F20101108_AABFOC ramu_p_Page_069.jpg 43fb11b617f6adff434eb4644fba2c12 93741bbb54bc09a73ee4224ce0862d96db186942 75569 F20101108_AABFNP ramu_p_Page_041.jpg 6175eb9b2d5f33235e7a3319ee6f74bb b39b4869c4944f9ed21236bbdde3c519a4164e9c 26400 F20101108_AABFOD ramu_p_Page_070.jpg 9caf2227e81be23e26512a40cec2d3de 0e4576a2f993ec1be36ea2ebad622aef17a61464 76037 F20101108_AABFNQ ramu_p_Page_043.jpg 1608619aa62b42ea004743ebc0120afb eb0247bdef7456b9259ee5795ddb2e063d84aba9 77224 F20101108_AABFOE ramu_p_Page_071.jpg 76192dd74f45112518f8ee104bb87bd3 d86820c090a40b01619744beaa2ecffbf962b51b 25353 F20101108_AABFNR ramu_p_Page_046.jpg 4ce3c9fc2458ddb8257a3cab9b36a6b1 0f2b6e749c71c4a9b71aed29ad9bc40ff81b8d12 81538 F20101108_AABFOF ramu_p_Page_072.jpg 4e557d4c121985a925c88a4021e29142 c1f82934a46a227a9c5081c7d7dd2b70484bf37a 16961 F20101108_AABFNS ramu_p_Page_048.jpg d2b32a1e03bab563b93b614deb3fba84 0c48a47149dbe727ead40bc16e9d75ceca33c398 72162 F20101108_AABFOG ramu_p_Page_073.jpg 4366e697c0728f86b3d3178469c32fcb 4e6ccc3ce854f5ff16d056dc327aa1b8b0262504 64985 F20101108_AABFNT ramu_p_Page_049.jpg 1cde77f49d7f893d7e1b007c21fbee46 90f8dc1b6d9480e82ac092cf84208e94a4852247 61062 F20101108_AABFOH ramu_p_Page_074.jpg b40163c5302f167919a8b84135c89627 dac1ee701cfe416ad2ba2e211b837fcde4244fb7 66318 F20101108_AABFNU ramu_p_Page_050.jpg 07d6e367656580b5c0a3fc70dfbb789a da934ac3f54e108bb4cb73c8cb5e3b5350606627 65981 F20101108_AABFOI ramu_p_Page_076.jpg 9f5da3c94d08ceda13d7251c2f4e3739 8ffbf39402e53e4355dc0d31fa4b07eadd99840f 72383 F20101108_AABFOJ ramu_p_Page_077.jpg c544bb690dc6e69a98e454e7072a1257 36d119b32939c3bb66899a617644d513b4f00cfd 74383 F20101108_AABFNV ramu_p_Page_053.jpg 34595a99c40c7914b9b383da22d96c44 0e5d93072824774266bb850018dcc104f01d184e 47167 F20101108_AABFOK ramu_p_Page_080.jpg be8aaa2c9f2bbcdf67da02573222138d 4b80193a67611f4bf8b118a81e9a5eddaa551fac 70881 F20101108_AABFNW ramu_p_Page_056.jpg 083a448a08ac604dac88f63f88e4ee8f cedbd9a13e2d8aa9442bb1805b814828361a4e23 24973 F20101108_AABFOL ramu_p_Page_081.jpg 1a73bbb07f99a5688de6e79843eb4dee dd4775661535ec015ba7f96bfd343ef82753d75e 75753 F20101108_AABFNX ramu_p_Page_058.jpg d94b1c559481df46159d67331c85ecf9 0809722a3e9856dfadfa006571a92c611fcb50f3 86803 F20101108_AABFPA ramu_p_Page_121.jpg ed547a6ad70bc4d82ace0f5c17cc00e5 c24a3e2aed745d478d83129725efd92c0297f31e 24462 F20101108_AABFOM ramu_p_Page_082.jpg 61df6a7efabf289f7b820e3ddb0946d0 cf581a266f61e5b6a6596b4e0957a159574d3786 74146 F20101108_AABFNY ramu_p_Page_060.jpg 03c82675c2da2387a69dfb080408307e 02916a1d2a9c24c6058e2e66d843a084ec9d84ab 86056 F20101108_AABFPB ramu_p_Page_124.jpg 35f11ee5abc0c304bf5d1ff571ee8271 365a18cb52da67c931ae68b70d01da4067b09e3d 26189 F20101108_AABFON ramu_p_Page_084.jpg c22cd117345e670e96f6877b14f69fc2 a9697e0fd63357e333e723e2b98d683d1fe343a3 75475 F20101108_AABFNZ ramu_p_Page_062.jpg 79c30b1e80f01df00640152427d54a63 a5aec7c64d4de3621d6d58ec16b94d4970c2cdba 29528 F20101108_AABFPC ramu_p_Page_125.jpg a8623776ced41833717c149e1ccb891a fed11a03acead41c7c80aa06f21ccd0c597b813b 26612 F20101108_AABFOO ramu_p_Page_085.jpg 76a737d0fad32204d7b97300bb8fafb1 f2b14acd73a38559294db97f8209616770a12588 90269 F20101108_AABFPD ramu_p_Page_004.jp2 a232980980b8b8e2f2e06a77d174269a fd6d4e59e5d05844f798b31d6f3129e38486ddeb 40954 F20101108_AABFOP ramu_p_Page_088.jpg c4e374d32e5b1a9c9ff437f195a5ff0f c83f156f8223ba1b4b2de941fe9b8b10e547f2db 1051981 F20101108_AABFPE ramu_p_Page_006.jp2 d532cda91602db2245b5f495df97ab39 e82a11611d04fcd1ad683ecdbcf6f0ec7775c419 38687 F20101108_AABFOQ ramu_p_Page_091.jpg c7815cba37714a7f2eb8909074115c6e adf5b894ce44523285191b05de3d76d19b827712 1051986 F20101108_AABFPF ramu_p_Page_007.jp2 ceb1f0429440e73cd7a7095d2d3508d6 d023a4e043a3795c476ca0da51c0df4a80f75d13 51401 F20101108_AABFOR ramu_p_Page_093.jpg c16ff63e28fd24ed91990cd4b78d8e77 4c78ce36d00ac1e7adf74d5004d3b1dc4712e04e 361607 F20101108_AABFPG ramu_p_Page_008.jp2 e05d65f634a411eb0e2d78096ceb99e8 8190ec15fa38f77aeb8b99bf537ad538e563a6f7 44575 F20101108_AABFOS ramu_p_Page_097.jpg 23b8f75bb035c1d6fdd1e23aa4509f71 0bd49eca351742b392456cd352eccd63e97a40ac 47594 F20101108_AABFPH ramu_p_Page_011.jp2 f4739bd2ce8f2d788bca0d1166748bd0 5eeaf74595764c4ea0a25da131c6ffdd226782b7 41622 F20101108_AABFOT ramu_p_Page_098.jpg fa09ba36762c3410268d3c797fe9176c 705d1184bb3d722eb652e4507be014a08b5840b7 21936 F20101108_AABFPI ramu_p_Page_012.jp2 ed24fa1320e972050dbbb637f907b3ef b89f49906efd588d6ffc2d69999bd8207b223728 43158 F20101108_AABFOU ramu_p_Page_099.jpg a4168ce49c74c3c1ad09eee7e0e40a7f 06170ce8b29c4bf329ebcb50fa28e4fa2b468d0e 98949 F20101108_AABFPJ ramu_p_Page_013.jp2 85c0ca2ad77acbd8d22d4ec4650b0995 8cb0de8f8875ab8592820e6e81d1771e4894fe33 70458 F20101108_AABFOV ramu_p_Page_100.jpg 7f2db5b2546da58f9c97819f24816b5a 547593ed00bf2f8fdcea0c1e32dc26c751deb0d3 117092 F20101108_AABFPK ramu_p_Page_015.jp2 4617600b84da008ccde07495972f96d2 a3c7beabdfc2be4b00e73fe4bb90ada29cd4ac67 119009 F20101108_AABFPL ramu_p_Page_016.jp2 ec52b3543ef70418da851db3433869dc 961861255f8559e0eba5074e27a0ad149f6c0c4d 73185 F20101108_AABFOW ramu_p_Page_105.jpg 44b9f9f7f5241f1dd663ed2aa67c1952 755dc4f82a08d9f5fb4b183e700721254b2420ea 112792 F20101108_AABFQA ramu_p_Page_043.jp2 aa1444f54bdc1542240ce3c25b274e94 c4f67c451c4361b0baeec3655cba71f6a2a30713 121546 F20101108_AABFPM ramu_p_Page_020.jp2 4438799ee5c52f89f230bfc9a6e4cb53 ba69f36e46e7ffb259f24b6d5a9eab0b337d4ee6 24304 F20101108_AABFOX ramu_p_Page_111.jpg 8b34314ac3c4643fd9a4e1278a976ad5 be66978a10206814623277bf9868b736d0fbc439 117547 F20101108_AABFQB ramu_p_Page_044.jp2 855e30cedeb32805aabf5fb787288199 ad065d7f5632e48058033ee8724ca5f044f3bb56 106614 F20101108_AABFPN ramu_p_Page_021.jp2 1f3ae5720893e7a334f082822dcf5e43 a5ebbd891a302b3b8ffc608728d7cd26dfd6a78b 32417 F20101108_AABFOY ramu_p_Page_117.jpg 7ad3e142eb7179f6c032c5cb14ae92b5 ea7d82778f86c751fcfbaf430831f1150b3e434c 351047 F20101108_AABFQC ramu_p_Page_046.jp2 29a7dccad94d138f8dee411b1db1b72b 70f679a5763765c6c93bfa3b604f67904a196dcd 106057 F20101108_AABFPO ramu_p_Page_024.jp2 a1ee74e4e08d35e98ec0b35d9a201e04 648915f898167d5b187d2bcd324d25c9ec7bc16b 19447 F20101108_AABFOZ ramu_p_Page_118.jpg 7c409fefa8a1efdfc756706744bb3f39 20d957ee93f1b94ce4939f47a074b01ab988bdf6 14461 F20101108_AABFQD ramu_p_Page_048.jp2 f15c52dbd4def6112546fecf2fc3e568 f5cf2a70e0856e90e4d5f1b1bac02eb8500e86da 206031 F20101108_AABFPP ramu_p_Page_025.jp2 a8dff70c610a3b5e7052da1c0b6fb3eb 231b3b5e8048ae5d0d57455d2169473c7338db81 111597 F20101108_AABFQE ramu_p_Page_053.jp2 a4ae3e6442bb0e78ac9a8797dcba4a14 7e75e1cf73babf98e21bc28bcf9c9ec37572d981 116634 F20101108_AABFPQ ramu_p_Page_026.jp2 cde3f091f0ff61816be02ff1bcccbe92 a754e1e1a5a8f4ff8517077ed9efd9edb8b4b49b 105790 F20101108_AABFQF ramu_p_Page_055.jp2 b62b030ea81aa79ab0f3b43835cb64b0 682e74d83bdcf601c58a212c0acef34a6dc31bdf 121715 F20101108_AABFPR ramu_p_Page_027.jp2 6d126967cea5c02ea6ce0ffae508c858 d01df7bc5aac54b1053edb01ee5a498246a4f866 103998 F20101108_AABFQG ramu_p_Page_056.jp2 bd8089d4ef6602b0c24cdff5550abaf1 c9d5eb8bfd140f3f6126204219a9432f4a2ee845 108292 F20101108_AABFPS ramu_p_Page_030.jp2 3fafc213ffc6a1d7ec3cca532b39f4d4 0a15239b79b63a2810bf31dac32db4fa3acb4c55 120179 F20101108_AABFQH ramu_p_Page_057.jp2 f86ca0b85ab2fe118a2ffa7234a8fee2 d9e976535b00fd52adc0e667e6b998901b8d8797 98314 F20101108_AABFPT ramu_p_Page_031.jp2 824d99fa40106bf51112c36ac43296ad 49633ccd04c7ffebde6ed8548ec0a84cdb8d73f7 110627 F20101108_AABFQI ramu_p_Page_060.jp2 60b15f8d339480ff52ee45f0b4d11d94 b1625464bcd7428c5e9a6f76ae4b652cba574714 105647 F20101108_AABFPU ramu_p_Page_032.jp2 71ff7d559e6e129737bf2ca58a19583c c07ba0bcccb847274ff001c08cdd4caaca7e716d 32193 F20101108_AABFQJ ramu_p_Page_063.jp2 9206e237fa797fc496d2c4e8e25ddad8 a24ff6e8d9681cd3c6206b731fc73ae0e548dc60 94631 F20101108_AABFPV ramu_p_Page_034.jp2 f0dd728f09c1c15d89577dd5789ddc7c a420f59226bab3deba421ffe69b1a8b4cd567b2b 15071 F20101108_AABFQK ramu_p_Page_064.jp2 8422a17999c6efd97e3e864d577c864c 1e7167624ebfbba9426c19cb47fcbe4c529c2df1 111174 F20101108_AABFPW ramu_p_Page_037.jp2 9eff443df340e979030bcaec96152982 51df76455516886c033a197c17733f3339bd2492 748230 F20101108_AABFQL ramu_p_Page_068.jp2 0226644b7570e870f0fcda31f07fcff3 dfc572e461f6c21d0d1110ed882cc686803896f5 307760 F20101108_AABFQM ramu_p_Page_070.jp2 beead18d426010a51912a73f21ef3f37 30fc10e25c011f3ad74797eb9b0c6a8f10026ec8 115525 F20101108_AABFPX ramu_p_Page_038.jp2 d496745239d8c40972a068edc9aef259 0920cbe1de88991c66485f31a9363a803bc21a14 11310 F20101108_AABFRA ramu_p_Page_092.jp2 82ed7c056dfe5bc3790715ab56053c66 c73c8c9d585638336945fd0540578c5ad81b8ab7 123585 F20101108_AABFQN ramu_p_Page_072.jp2 20ca6e49cbb2105661ab899941771c4a 30dd57560dc5992e198719c8f04b500e862f4221 112272 F20101108_AABFPY ramu_p_Page_041.jp2 6e22f911e8d44e2d504b7c4be030a865 88afaedf9ea89f638786ace4feeabaa703ac5659 512453 F20101108_AABFRB ramu_p_Page_095.jp2 be371cdd619a48dbcb24182463aafb14 43f3b8a7fb02fb0cae2a28d0c60042bc4cd65341 110093 F20101108_AABFQO ramu_p_Page_073.jp2 b71dc199929a40e0f9d32b79ca615a4c b905bea86f96649970717702b2a47a6e7c227672 101251 F20101108_AABFPZ ramu_p_Page_042.jp2 2c02cfcd7a7434eab1ea0fdf3e0f0ad7 7c614c2e9c5a50eb1fefa5b6724393f0affccedd 74734 F20101108_AABFRC ramu_p_Page_096.jp2 c5a4f5043d3b3a3a5d5d7142f1a6c904 105fb228cc1352acde084e24881b3f2b3280edd0 84749 F20101108_AABFQP ramu_p_Page_074.jp2 be26d0e242e8bcea242f756a3b166eca fb0272a44b797cc47d874b45ff184c7c3a294613 62480 F20101108_AABFRD ramu_p_Page_097.jp2 60f61a52dc6030b47faf0e2d25bc9d87 00bab36b7bdea79552895fb95ad6f91d229a66f1 99345 F20101108_AABFQQ ramu_p_Page_075.jp2 d891f772a1dc1da63a826a5811dd9995 909cae297dd8108204fed5c071535aa8e94cbad7 57575 F20101108_AABFRE ramu_p_Page_098.jp2 2347eb18035dc8b41cc5cf54a9a0e192 c4d92b1bf06cb3c10aca4577f3b062ad8c36f5db 100220 F20101108_AABFQR ramu_p_Page_076.jp2 c5fde8bded1bd4cd13148e89a4f1510d 4f46f70028eb86b764e7c32a3784a1281657e79a 59335 F20101108_AABFRF ramu_p_Page_099.jp2 a7675554ff6a17a0f3e18685e91f4036 e3596d4a2bb6be3fbc315b4d0de86894df985bcb 107069 F20101108_AABFQS ramu_p_Page_079.jp2 8f907f627f17ec29e93358b3609cf923 7a7a8ce988ccb29d0948eed2d27d314e5bb57018 101872 F20101108_AABFRG ramu_p_Page_101.jp2 5e61de442274f395f68c664578429029 2c432f70c1ccbf6448011f041122bd9052a750db 63444 F20101108_AABFQT ramu_p_Page_080.jp2 70b6c5f856ecb6c5e155ca1fbf0d9da4 5915094d34b89a7995babb0dae96cbea2845497a 107489 F20101108_AABFRH ramu_p_Page_105.jp2 e45ba8953f17e7a3be7137a8763f5eba c9096befe5367a28c4d347f5e2866a639400a96d 313165 F20101108_AABFQU ramu_p_Page_081.jp2 47c3a48ea416876096d7bcec37a35d13 4767b65038c99ab65c3783e28d6098b4a55e05a5 27187 F20101108_AABFRI ramu_p_Page_111.jp2 a0da3f6b5ad118f44fa5c21dcd6246d5 72722c136b2cf4bfb8bc2e2c7def4830b7fe39e9 318161 F20101108_AABFQV ramu_p_Page_082.jp2 11102b6e10ec634b1ff021e9fcdabbe2 008a7474b04c1da5714828719bb34844b73364c1 553123 F20101108_AABFRJ ramu_p_Page_113.jp2 409712d5058080885de610667217c8cb d3415b3034816c4ce872f50d89fe0c159749092f 714313 F20101108_AABFQW ramu_p_Page_086.jp2 bd32161bc366251ae0c34e531a0e956c fbf47e3c55ebc1cc8c025306c24eefd5d3e23d8a 534296 F20101108_AABFRK ramu_p_Page_114.jp2 715e9d58be2ff656df4fe660349411ca 5bd5fa044d28f54c75ca986514ce3d023ea1da7e 406234 F20101108_AABFRL ramu_p_Page_115.jp2 7914ab43772a8030ba8a20e97da89220 dcad103953d4d5128e1dcadd0c5289d81c968d62 736640 F20101108_AABFQX ramu_p_Page_088.jp2 b90b5e7508ce35fa4628992c37579f03 39a18d45c4f09c617445e8a4c5212fb77390250e F20101108_AABFSA ramu_p_Page_023.tif f7e2080110cf1be3aca7637b33241ca3 ee4c8b67b27beaee9befc09c1a7505620726e3bf 541568 F20101108_AABFRM ramu_p_Page_117.jp2 76acf7dfe392a4fe35def5acd910648a 2edaad9f42e4e2ecc33cc2e2e7f5191b3536758b F20101108_AABFSB ramu_p_Page_024.tif 89782af3161eb4d13d633136ffad614f 36e25e988a471d414b039d6cce0a8225a14761c7 18592 F20101108_AABFRN ramu_p_Page_118.jp2 af223d08433517496af9b1a637adb61d 40c5d5ddd4ddfe3593905a508f6ccf92a457b55c 552317 F20101108_AABFQY ramu_p_Page_089.jp2 567028cafdef85e3347b3252cdb1b5c2 44dee1e52258b9df38e6ab36dfd34a17b2689c88 F20101108_AABFSC ramu_p_Page_026.tif 7a3adf6a3f3a8d59469db12eb1c6fbc3 0071ac011255b5b2c39b8cc383a416dbc02a75cb 15281 F20101108_AABFRO ramu_p_Page_120.jp2 6d5f8a25c26e916d95e6323de731b4c7 374798eb307a88a8528f87d9ba8ac2c024538940 261639 F20101108_AABFQZ ramu_p_Page_090.jp2 595968f2fee1232df97b89b9c9b9d6cf f75746f7e80ed70ce45fb2cc09abcd45aef65f97 F20101108_AABFSD ramu_p_Page_027.tif 2b7eae1dcb6498b5121b36e8d6bdd725 306b5a6cd04de6970597b012051fc4827deed310 137835 F20101108_AABFRP ramu_p_Page_121.jp2 80193c7826dfcce58ae60e46b50dd77e bdcb2ca9bb679bd99843a1241d03e99a428cd2e0 F20101108_AABFSE ramu_p_Page_028.tif 83d6d210f768a113959ecbeef2be5068 8aec3c5f675fc04603d82b1b48c29388cf35f030 127992 F20101108_AABFRQ ramu_p_Page_123.jp2 b730790aa42285179b8aa503946f6124 fa026f7b0de739cae89f9f884829fd073f8b8c04 F20101108_AABFSF ramu_p_Page_029.tif 66fb4bf1afd5bb0afffcbea84d1c6b5c 45c6eebc9d165497839c20b6b7852a99631d2709 130399 F20101108_AABFRR ramu_p_Page_124.jp2 59462ce0d11ca16e71572e53ca8fa6d4 8910c82e44e767180fb617d681e24d8df7cb5db8 F20101108_AABFSG ramu_p_Page_030.tif e84ef904b01962eb6d3b5156a9ef959d 4544921d411aa28747fef94093cb32486386a2ac F20101108_AABFRS ramu_p_Page_002.tif c42d569954734393f104d202e1e6e2f8 33699104157604586795c17e2f2c44eb286bd75e F20101108_AABFSH ramu_p_Page_034.tif b9e1554ba02fa25a56d1b25a90963cde e1f563eb10d4e2b163e6a7c997a3f52bf356cedc F20101108_AABFRT ramu_p_Page_006.tif 0da11d0fd434cb5b52ec0b3fd6b7279c 46e97204a16ec733c736627792b13f85fb76b235 F20101108_AABFSI ramu_p_Page_035.tif 381ad0f8130971a20b7098bef2a763b1 b32f73b3c0a689d0555c4765a84e9f027b42ae49 F20101108_AABFRU ramu_p_Page_008.tif 5ee1e7a7ecf0a30b268f26e358910015 e6c11e88fdfae809c25941b16e85fc473db50a6f F20101108_AABFSJ ramu_p_Page_039.tif c7d8df2a196098e9885e3f172b95335c 815735f8ebf9ce7209c96aee8a3aa6c4e67e5c95 F20101108_AABFRV ramu_p_Page_013.tif 78d92fcc5a6e3081362dae988f7d8bf5 8d97f9403e9d0b6a6c88c5ea57580e6b7feb356a F20101108_AABFSK ramu_p_Page_040.tif 9d8169f5b2c75836b662d2771f944e8b 704a90de1ae16a1214fb6e11cb3b4c037ffd3544 F20101108_AABFRW ramu_p_Page_014.tif fb1d5029b96171918f6181f5e61ccad4 be7a97e17bb8554dad2ec2b28c064b422d8533fa F20101108_AABFSL ramu_p_Page_041.tif e621c7ddcb8d2d2c0735fcb47007fdb3 968708e5258f539217f0637f8c8d5cfeedf624d4 F20101108_AABFRX ramu_p_Page_015.tif 2359ecf5b1f614aabb5a8f8f78625abf 3cbfe8626142fc1ec1ff6b4c32bcb8f9b9c24bc2 F20101108_AABFSM ramu_p_Page_042.tif 1186b2cddea8943b6ee4bd9086cfb92c e9a967d5dea1cf7083a9951d2f438ad188476951 F20101108_AABFRY ramu_p_Page_016.tif da737eab8eb1a29068c1ebbf8b447ed5 41ec480e1183393eddeb7d14f2534509f4d955ee F20101108_AABFTA ramu_p_Page_067.tif 5d4057b2cabf9afa6ba9bd4d38ce2644 8d93fe1c5c5d42b977ea41212130a2022dcd5a07 F20101108_AABFSN ramu_p_Page_043.tif 6f2fc506da5e03cd254382f639f70eec 10925ed7e290cf23a4845322b1e389baa2702e3e F20101108_AABFTB ramu_p_Page_071.tif 16a2fcde58f600d038bcd8a3d2cf4b64 a0c6909f875f40fdc280bcc6e4885a0850c19fef F20101108_AABFSO ramu_p_Page_044.tif 985305ab2445ec3eb8bf10e8ba5ad179 75995b167c533e09c033c91c0866db1d285041ad F20101108_AABFRZ ramu_p_Page_018.tif d64be8ecc79445b8eac4df4423c31c72 eb8e34d73e69490f5470ae33d879b03da841da33 F20101108_AABFTC ramu_p_Page_072.tif ecd84db1fe10b8e84d98e70c2c051507 8de2340acb3ec861a9a4791af2c4e14c476d52ad F20101108_AABFSP ramu_p_Page_047.tif ffc0aa2f91b427e487ea068949f8e509 749f967da5e81d1d673766cdc3b4745a0bb152a1 F20101108_AABFTD ramu_p_Page_073.tif 3f0af01920c11ad269e331f83440830e ec02e4ed2f6f37bb6898c25c9bc1eaf2ac3be1a6 F20101108_AABFSQ ramu_p_Page_048.tif a46df1429b058ddf966616dfdd2ab500 0e573dfc2eea584b59e541c43f31802ab018a0b3 F20101108_AABFTE ramu_p_Page_074.tif 9bdca58648a607d83815d6abc72bd223 f16b38b34a10ecfeedb2b18eba49e0afa28c17cb F20101108_AABFSR ramu_p_Page_049.tif 07d34b65e72db29f19729257e7804f1b 2e81a3564ee44de1955f163d4199eb45f8c0c826 F20101108_AABFTF ramu_p_Page_076.tif b24e1ebb855c7aefb0441020068da659 dbb99ff54bf4cd0353a82fe7e1f91c0c2ba8e99d F20101108_AABFSS ramu_p_Page_050.tif 0ea96e2dbb8d3366cdff665414b6d63c 13b14cf8329be6b7f590ef6d3a73e225ea208a92 F20101108_AABFTG ramu_p_Page_078.tif ed7c0ad188402cf58df2b38610125adc 3e924ed634fe185f91522d6ef2339ec578c37401 F20101108_AABFST ramu_p_Page_052.tif 0ae289d5478e1c0fd0ad003de3d576b0 647b4ddf4441aacf0597f1b727a5ff88cef4bc72 F20101108_AABFTH ramu_p_Page_079.tif f850ed11896e855ee89ac82c7cc761c7 a7a4645fb55b10dac584b537ceebea4df8a7784b F20101108_AABFSU ramu_p_Page_056.tif cd20f5b132f20c10722605157b7fc298 923939fef5f654ce32acf341d52b7bc0bf5e945a F20101108_AABFTI ramu_p_Page_081.tif fdb7d1daa759f852410ab6bda9e1fb29 59db1a014744fa70fb78063c330cd5491e30c211 F20101108_AABFSV ramu_p_Page_059.tif 783237eaab7ef514fb22cf98b46787b4 fb0b8dbf0cf6f4d074fe29bdd2cb1fe0ffc2194d F20101108_AABFTJ ramu_p_Page_083.tif 2f357049dd148875c78ac88510427c4f 80025925bc72de9b0db4773193ca0761d081ccbe F20101108_AABFSW ramu_p_Page_060.tif 10c306eadee5f28dd7f5a5b4b1fc6409 6b6625ececb868a0b1523418b515c1f7d64ad666 F20101108_AABFTK ramu_p_Page_086.tif 24aaba83caad889df0235f30a06fa6f2 828b7ddfeece132ddda65366143acfbba1ac2a5b F20101108_AABFSX ramu_p_Page_062.tif 53b0d2428fa265990c81b10200e12c0e 2f7f8c75f6a2be2d11a5868da2ca41000516a42a F20101108_AABFTL ramu_p_Page_087.tif ffdedf48cf7c3b77229ebfef7bf9199f 3902aaf9fdb2434c6f00ee7c4e2d581c0a488c73 F20101108_AABFSY ramu_p_Page_063.tif eeeb104600a80cfcd716f01d9935bb68 6f87750ac5f46c1b2f6ebaf50d331d653059cf04 F20101108_AABFUA ramu_p_Page_122.tif 4e854ff6d828a71a3b493c951fdf8519 a69bc421a9e24d4f918556447f7fa1a4da0dc42d F20101108_AABFTM ramu_p_Page_091.tif 776e587b3fc8da728df4492c15db8680 05d2659f83ea46041e7ddec263b08c66e5c55268 F20101108_AABFSZ ramu_p_Page_065.tif 94ac64280369d3b95256de04ccac9aed 35f7bf70f39069ca9c210e98e70e27c9ddba4cff F20101108_AABFUB ramu_p_Page_124.tif 15089ad740b2d6053b42a442d3ad8ee8 abe2d88bb317197b365edfd6bb9fbc098a20a473 F20101108_AABFTN ramu_p_Page_092.tif 9333a9c335155dddb1007a340e6cbad5 deb3efccdc5fcfbcae4c7b27ea6fcd68c7251c7a F20101108_AABFUC ramu_p_Page_125.tif b8e39d8a0d6f46b66f8be771fd3ffb4c 7a361e8aabb728a46ddc5dbd543902897313a465 F20101108_AABFTO ramu_p_Page_094.tif c6a9d1f7e14f60fc339591fcb4011379 bb43ff464c5806009fba1e0e2a2ec92dd6238940 8474 F20101108_AABFUD ramu_p_Page_001.pro eb446392edc5cbac1b3c7fa0c896ba44 037492395bb3f67e4e8ea6452ebbd81ff0812477 F20101108_AABFTP ramu_p_Page_095.tif 510c58284ad38b91f3cc74b1ee3513e4 56a0a73590e00200a9fad5e5b19157cc99c90087 880 F20101108_AABFUE ramu_p_Page_002.pro 32b115ccd2e494d192d79fab5d907bc8 c3c4b7f7d0636add72f0477983306f5035eaec25 F20101108_AABFTQ ramu_p_Page_096.tif 3d9af07048c3d58fbb92b1ad785dbd06 76ffdfc7220cbd06ff2c15b3bde2b550f4ce16c6 65294 F20101108_AABFUF ramu_p_Page_009.pro d24e3224874e282005a1fdcde5d6aa16 ed682b771defe20316e9ec2096e90f06b83c138a F20101108_AABFTR ramu_p_Page_097.tif 70dbe523af38c388113b47ebc44c60fd beef08628886a9f95cbf83f6aff095efc4c3d1e2 23242 F20101108_AABGAA ramu_p_Page_036.QC.jpg 002c726e11a1894b706531d31170fb5c 469ed82cc728290d24993ce2699c9f27ddd1a448 18666 F20101108_AABFUG ramu_p_Page_011.pro 4af4b9f73015666e5850c27f8d2efa58 e4314af0395a4014d5f17d70a529ccf0c86b1876 F20101108_AABFTS ramu_p_Page_098.tif 13ba1065bb4390f299406e354d68d3ae b2c798571cbe498cd77f64253f1bf99c14452063 24598 F20101108_AABGAB ramu_p_Page_037.QC.jpg 422386119bf977cc08a1827aff1d011d 6a629da5ee706a1d08e05e2243545d8ec4025123 7676 F20101108_AABFUH ramu_p_Page_012.pro 8a30613a6126595d0bc63e50f27b253a 998fcdb0d9e5e3c08c5321fd221b18be2f5c9cae F20101108_AABFTT ramu_p_Page_103.tif ebd1d759ea2c3c78643fbe059928adc8 107e688bc319af4203abc050b2efce7e25237838 7025 F20101108_AABGAC ramu_p_Page_038thm.jpg 3f0be825cce2f71b91fde7891fc73000 9a70e991df8ec8d9d1a97798e22c65f6b4d47fee 44622 F20101108_AABFUI ramu_p_Page_013.pro cc9ae22428e05bec52e6662262b31b52 d9c594387a0f12825ed97234d3589cfff0d2dce4 F20101108_AABFTU ramu_p_Page_107.tif a8adabc7ae5a42d70e9074a0f61ff756 cfdc760719174f04ba216d5ae6189b203f64afc1 23071 F20101108_AABGAD ramu_p_Page_039.QC.jpg f0e3acc668e80c6e984389ac839a34ce 4328f26aa3f779ef9d0e06937ca21024ad21a613 55519 F20101108_AABFUJ ramu_p_Page_014.pro 7aa0c3ad78790082953ccee6a0243998 d2a7b8ca9a6698d0cf1d8a599899d3358c062eaf F20101108_AABFTV ramu_p_Page_110.tif ba2594480a4082e218f9c32790c737ad e69b15076988b599cfba42ea51e3db31166dd5c9 6469 F20101108_AABGAE ramu_p_Page_039thm.jpg 6dd0590ed00365f3ea92a70e9b20cae7 4a272c7c906c9932769938d047254ba3f0a069cb 55018 F20101108_AABFUK ramu_p_Page_015.pro 9e4e46ef7930f2576edc0fd7d5ab0506 9cd71eda8fcc1d0a0cf2b0efb79b94d8dc9d37f3 F20101108_AABFTW ramu_p_Page_116.tif bc570f4144b1a3a0ed4ee7a4e4eb5dcb cb088caea97f3a95133ee6eb3f029dddde0eb4ca 6369 F20101108_AABGAF ramu_p_Page_040thm.jpg c48df7c358c381ebbec356693cb2f973 98088089f01e769019aecb84a50dc3895b5e7c56 6707 F20101108_AABFUL ramu_p_Page_017.pro 9c45fde39ffae58633d97a3d68a58285 04d0e4a43a9fc3a958d06d71ce7ef824fe835039 F20101108_AABFTX ramu_p_Page_118.tif 3b88e09947006cd0c4e77154b3dd13b4 9d3143edcfe4482670642e42ff3046e9d00cae44 6939 F20101108_AABGAG ramu_p_Page_041thm.jpg e67c861da6f21836b40ce68f6a39845a be075784c84d18e8771d4acf568572ec29e66a51 34299 F20101108_AABFVA ramu_p_Page_051.pro 691452b170219376cd12ad285404aeb2 d5750126d1d85efaf0dfce1eb336b09f42c18464 3666 F20101108_AABFUM ramu_p_Page_025.pro 4d73f01e04e53ab73fc530f0c8bb5008 8a08734169c03215e640f9b288225dc9c7071f20 F20101108_AABFTY ramu_p_Page_119.tif 4327efef9c358d8af2d83ba67998d701 6b33a0826efee89f985ec7dd84354a2477cfce10 46953 F20101108_AABFVB ramu_p_Page_052.pro 1842f036bf663c824dff1c0f7dfc12aa bf919a9923874425ba5768df6d349c86c483d933 54314 F20101108_AABFUN ramu_p_Page_026.pro 24a84c3ebd12315980003f93195aa04a 3c44b682de4a44e99ac52413f2bb744a4f8feeec F20101108_AABFTZ ramu_p_Page_121.tif 26d6c10fb812301d9d2ddb17780cf12f b126038436bd338f77890ab2c709a53a83f95a3d 23353 F20101108_AABGAH ramu_p_Page_042.QC.jpg 5eb211386f11d2d15f1a1790408d0113 08d313423f1b2458a4dfbb5b905273dacda5138a 6823 F20101108_AABGAI ramu_p_Page_043thm.jpg 76353309e35c62aff8baf88a20ab46f9 899de2a920f6ba1e572363e6f6793d73fd14af93 47830 F20101108_AABFVC ramu_p_Page_056.pro ccac46f626c24aba0d65f13a50501f38 4a96521b651eef633ffcdaae4c48f1d664e35fae 52819 F20101108_AABFUO ramu_p_Page_029.pro bdbddb09dbc4334a8d948d1a5fa2f136 6bf4cfc09bbc4a73ec6471ac958d5ec71add7661 26152 F20101108_AABGAJ ramu_p_Page_044.QC.jpg 97bd75b22c93d2f409ecb7a157ca97c7 6785d53ee8f9ad591ddba2ba507d4649ca15ba9d 51906 F20101108_AABFVD ramu_p_Page_058.pro 9bc48bd7098b3b6825a5997779948908 e04950a1f92ef8c5121cb8d5524c383fabcb3f31 45679 F20101108_AABFUP ramu_p_Page_031.pro b9f9bd7b296f127ccbe0f6e60f7171f5 38aef3e4bf3863e6127c234f5e65229e63029ba4 6979 F20101108_AABGAK ramu_p_Page_044thm.jpg 7c8f658315ba6afb473fbf2ba07905a7 f71a080233ee9de15ceda69e316311aa0c2d7650 52757 F20101108_AABFVE ramu_p_Page_060.pro 0516a62763edad07890ee126df0b685d 5aadd6da4c9ed7e00c5e54fb2cdd3b6fc38012f5 52110 F20101108_AABFUQ ramu_p_Page_035.pro 56ecb9cf1a6724c1d138a825026cab5a 6ddbd451c1ffefda7bf69fbd72704261c76d189c 25962 F20101108_AABGBA ramu_p_Page_057.QC.jpg 30284224dbae2e7ab315d0df76715977 3147214bb65d455b8cc9f17e6592368d6df24ae0 10265 F20101108_AABGAL ramu_p_Page_045.QC.jpg d1ed3ab8e5a0284f70840a64c590db51 2375fb5ef9f4b3a68cd04aaac6cc48f4ab010fb7 53520 F20101108_AABFVF ramu_p_Page_062.pro c0fcf7225f19c385cc20fa845e931a4d 16a797540931d27a7ea2467af07f51dd280f1792 49495 F20101108_AABFUR ramu_p_Page_036.pro 9a31bcebb0ecefb5b321704aec21db8a b8f86236f16f9bc398968d4d2d48cda2a996a321 25612 F20101108_AABGBB ramu_p_Page_058.QC.jpg 80406dc5c2a8e88c070d6793f3ad98ec 36b9d9ad10642926bbdb26087d51367f5de62a16 3131 F20101108_AABGAM ramu_p_Page_045thm.jpg 16ca0a3bf724041390de8e961f8bc940 5f8473a8611292018b4ee0c431292cbb5bf9d27d 13430 F20101108_AABFVG ramu_p_Page_063.pro a3dc0442f18f8372f2dc21a86222ae18 eeb7864e387dc9c4fb9636e80d28dc721494612a 55896 F20101108_AABFUS ramu_p_Page_038.pro eae0e3721108875accdf7b09816741df f047e585b78ea3b11eeab30ded9d9b7fa1d66473 6903 F20101108_AABGBC ramu_p_Page_058thm.jpg 8c324d81ee060a0215946c1aeffaa055 26734e339952e67e70afd027be627b175b529a76 7976 F20101108_AABGAN ramu_p_Page_046.QC.jpg d495623979cf28397aae2a0a3b50817c 0688fc030d006988b65ece7149bc0ef6da505f21 4010 F20101108_AABFVH ramu_p_Page_064.pro 055e78fb399aac03bb52d657eace1dff c314a1089e9a651a95751999410c95d1ba8cf377 44198 F20101108_AABFUT ramu_p_Page_040.pro 19a3f7f5d4a9f6bdc3ea7a2c3f630795 c1e63ed5bb3d825fe72b5c5471e97bb7fdb25912 5586 F20101108_AABGBD ramu_p_Page_059thm.jpg 650089348a54515b8c133be3fa2d3d2b bd7fd4290de186aa6227ca48d3ac12dcd5f90421 2362 F20101108_AABGAO ramu_p_Page_046thm.jpg 05285df83fab9a7a78fbfde3aad34394 f431bc990d14b7b9c5961cdeb220b7ebee129e39 10479 F20101108_AABFVI ramu_p_Page_065.pro e117de3c2944968b87d08e55b3c14ee5 b2d886cd0bbc395bc1b8e6a193a6beeebcbe5ab1 53620 F20101108_AABFUU ramu_p_Page_043.pro 5d8e72d80a566ce991d8e4c11b83cc7d 750e767cc3aed50b2952b17f1aa1f35017507630 24639 F20101108_AABGBE ramu_p_Page_060.QC.jpg b7d678ccf097910ebe8c3f027abe0115 09dcebe3063c3eddf89e62f52d4991e9f3bec027 11880 F20101108_AABGAP ramu_p_Page_047.QC.jpg 4ad798384aa7f9009e99f588c10d544c 326110e7d6847ae8b7fe9018c9296415d4824c77 5438 F20101108_AABFVJ ramu_p_Page_066.pro ea7001a6a383cd9d1e872ea87dced882 bf720d0df4f3eafc97ec886515babf37a0e56214 18286 F20101108_AABFUV ramu_p_Page_045.pro 354a55f090b2ed548c3c2a69d3d1e7c0 d9e1d25cb608caa614580e9525f175f0074fbe70 26782 F20101108_AABGBF ramu_p_Page_061.QC.jpg d815e6f6f541ca6e05b754e0ce0adb21 bde556623d0ccae6531f4b8169dec08760a31f57 3755 F20101108_AABGAQ ramu_p_Page_047thm.jpg 1aa592f1e5691d2b47616c365c051d40 1715091ee4cb335c0749747643fc337678d49b98 3063 F20101108_AABFVK ramu_p_Page_067.pro 559b870ed841a4d8ec34141421d98a23 8a63c3e61ad1fc2a7a9d3d17cad7bc53f567dcb7 11336 F20101108_AABFUW ramu_p_Page_046.pro 5175b08e7dd74abc84d195e36287a5e9 54555131a226601e7a1509aeac79b5eb23a2d99f 24210 F20101108_AABGBG ramu_p_Page_062.QC.jpg 0b45dc9d4fad9c019b0c8c6322fab22f 947231049c1f3a86bd2313e3ff68bef320c811fc 5566 F20101108_AABGAR ramu_p_Page_048.QC.jpg da4f51756117a7a7185fac0d254d4876 899df0edc5c33a0e482b04f584b6209eb8311e9d F20101108_AABFVL ramu_p_Page_069.pro 2c8987d54da95a4e5a42a36d76ef2063 ee0c9cc1145e6d2ede3139f70b9d5dee4b445043 8806 F20101108_AABFUX ramu_p_Page_047.pro 930af12c54a64c98ddd9c9ba95016746 5e6954e96aa6684f699dc0fe384cb467f9e759de 6916 F20101108_AABGBH ramu_p_Page_062thm.jpg 61295fd27326c6b844b346c24c5afa69 c4b2c0c0779721f8c33599a7a6f45b6cc38a7660 2032 F20101108_AABGAS ramu_p_Page_048thm.jpg 34e01bfb523bf6dff9e029e01964f450 3d1cd77fb49e8b198a4de6c09844bd92eea8af06 54635 F20101108_AABFVM ramu_p_Page_071.pro 754f4a3bfd03b30ef1989e2a9b69b753 9ae7f4c96e30f5b41a9e9c825222fd7a8e1f7650 1504 F20101108_AABFUY ramu_p_Page_048.pro cd4bf94d5635d072be79d01a1ae794f4 dba1dffe4d8796486151f7c4184c1da078ecb6e5 15726 F20101108_AABFWA ramu_p_Page_110.pro 1f67c5c4af22646f58c87e190bcfe66f 80e3ed514d24dbebe8fe614ec2912ab9a17f4323 6275 F20101108_AABGAT ramu_p_Page_050thm.jpg 838f18062445b7ae1793991857b7c2e3 c4be8619694c118fd48143f543920c9c5cbac6d1 38523 F20101108_AABFVN ramu_p_Page_074.pro 3a4af26ceb094501e9ef31b3fb0cdb29 18923f47e9e093d78b8ed8f2d352537c5fcdb92b 44814 F20101108_AABFUZ ramu_p_Page_049.pro 1b19d774307f8b8b8c6918e7198b44cf e9a30214662425d332c0773128c92924edddf4bc 11204 F20101108_AABFWB ramu_p_Page_111.pro d125622539152632127c205833b26662 8589dc4e76f354c90a842b85f30c503da2d46728 2586 F20101108_AABGBI ramu_p_Page_063thm.jpg fa1387f9dcffa415bb62f717e76db7bf f936891c0c60bbd51dba8a025e8b6aa1a785eaab 17173 F20101108_AABGAU ramu_p_Page_051.QC.jpg 0c8498db0f91d3934c93b9af1c0cda93 9e27616429c8594b7efa45cc540d2d5d39b837ac 45776 F20101108_AABFVO ramu_p_Page_075.pro e57c5f9f21cb9737fea8d9336f25c5df 2cf80d759754e668cc426477ba1e05fb49ffc8e8 16739 F20101108_AABFWC ramu_p_Page_112.pro 851ec97205d848cc00b41b0d53a5ad3a 360a55338b220d0789777a53da36d77c48f442d4 5228 F20101108_AABGBJ ramu_p_Page_064.QC.jpg 0ef03dfc3a41d2650a9338c5ec15e7e7 62f8fc498c1e2c2e5fe4f9ceeb62bd4ce207b92b 5353 F20101108_AABGAV ramu_p_Page_051thm.jpg 4bb61e0859b69d7efebad3728b0a16dc 587da0ae6c0b43da8e109d0346e559d0deb86242 45521 F20101108_AABFVP ramu_p_Page_076.pro 58b0948cc342b0e8e9f0abe578bdc642 c63355ac64c9304fdbef7b540904f536e6441e67 15314 F20101108_AABFWD ramu_p_Page_115.pro e4b3c404ce8e8050838d4097bd933809 d1a93342f949f5a8ac8df2fad8074be442247f4a 13308 F20101108_AABGBK ramu_p_Page_065.QC.jpg 5b018b25aeba94c698ef43c6bcd9b273 10f66f6fe31a8beddcc6b186f5cea4279782a31b 6458 F20101108_AABGAW ramu_p_Page_052thm.jpg 46fe2087e982fa7feed6564d826b4bd2 b742e63590b1371bfd3eb18ffd90ce547e91a7e9 25587 F20101108_AABFVQ ramu_p_Page_080.pro 82c79989422830049a0ed28421d49107 cceab87f7e17210c3eebc3cfa979935a1d82c32c 5049 F20101108_AABFWE ramu_p_Page_120.pro b0b391723f467a6ab708a01d6a5abd86 6efd436d9981fcb4ac4e8634414d3a08e44012e6 3350 F20101108_AABGCA ramu_p_Page_082thm.jpg 66d05b3be69372928f1f3f981ab963df 8be9d02ac59efe477a0404ae46e58afb142b92e9 4368 F20101108_AABGBL ramu_p_Page_065thm.jpg 6e322401c5a47427563e4dd5fcb463f4 22354ac9327a01efc90960df4e41b7ce988311be 24902 F20101108_AABGAX ramu_p_Page_054.QC.jpg 14088da4fc0748e2ea5b11a81df43d6b e2400182f578cad1a8672fda8d45a4570540f5e6 4939 F20101108_AABFVR ramu_p_Page_084.pro f23976adca78513e8322c80bd980529b e4caa5e9d490af2eeb2ae59c40329be339d07254 62670 F20101108_AABFWF ramu_p_Page_121.pro 2acff4d69322ae064e4b591793d5eeff ad0329a300a0abae644035e7ebb5d7df02c424c2 8326 F20101108_AABGCB ramu_p_Page_084.QC.jpg 2394f5766f5fbbd535e0fad10769dbfe 9d0697057feb7ff1ebf1bf6e6cabfe32d9c5390b 3071 F20101108_AABGBM ramu_p_Page_066thm.jpg 20d81db7da700abf1efbe21da7b3e9b2 4fe0f88dde3b0bd28680b11f2bc42829008eeba3 22960 F20101108_AABGAY ramu_p_Page_055.QC.jpg 05519a89327681818d4918101487391e 7c174ff755d06616deb775af1b4caaf262b2b7cb 6439 F20101108_AABFVS ramu_p_Page_085.pro 2aab20bbbd0ef3cd1ab2415c0cdd5992 2ee642fd88e4c6e33af45c6c0a3ca0f32f3e0561 64587 F20101108_AABFWG ramu_p_Page_122.pro bde239ad5e882c4c947fc14c70823f1e a93da736e2a379ee7bacc07e08afa1b488963cf3 2993 F20101108_AABGCC ramu_p_Page_084thm.jpg f63e2a02b60ed9c52fbcd0a50cdc48c5 b6d54432fabb3e8c2ca4c6219c9d7d9e9e1d2191 9598 F20101108_AABGBN ramu_p_Page_067.QC.jpg 73f050ff2ce818d78d3f9310f7c19461 8eddae4f3ec5f0979206b33d6f94dfaa770e0e18 6354 F20101108_AABGAZ ramu_p_Page_055thm.jpg 4b17621f2b19b8ae0122e23ea7248d96 ad5b39cc335a20a92f6cbcaa7ee1bb5056d0f9d9 26863 F20101108_AABFVT ramu_p_Page_086.pro 76dbfd2954393b0a711298c9fef06d54 d023207f92db6366f5fc3656941a5c94cabe090e 59567 F20101108_AABFWH ramu_p_Page_124.pro 2bd5f1d24b1e111243a56b835f20b542 52494001d622cfc5c6fce1b9d14a0a76e3538ea3 2797 F20101108_AABGCD ramu_p_Page_085thm.jpg 0131ae16cb54f5dd88d8370562f0f42b 395de4f5449374bf53639a89c88ddc05d59013ea 2987 F20101108_AABGBO ramu_p_Page_067thm.jpg 9a5957c253c15b1142b18db38259db66 94e40ba8d5cdfc0af80bb6ed67741b8f3b672222 90 F20101108_AABFWI ramu_p_Page_002.txt 5ea0521ccd841c621143ecc45713b122 2027c5f0227f4c8f92c05dbf2d3a81058f60a652 27758 F20101108_AABFVU ramu_p_Page_088.pro 83f910afeae8e0581b431b424e3ad68b 33e9a997e8c6ff169a8e82779fccc910ae0e9b46 12791 F20101108_AABGCE ramu_p_Page_086.QC.jpg 3f1d48923c321b2a62ac372ac5596087 2c10e22b5f65c51b3399e20253e5b89e6e67cd31 4008 F20101108_AABGBP ramu_p_Page_069thm.jpg 6c0c049c2593d8d0a3e9277887434e3a 8e974089584b4eb5ab127ca4db7cd3a1320dcc2c 111 F20101108_AABFWJ ramu_p_Page_003.txt 9ca314dd1111fa2643b579d469bb4a50 e38e7d16b4f2caccd623c1bd47185ded93b1291a 26153 F20101108_AABFVV ramu_p_Page_091.pro cec2bd4e9a9a90da8fda1f1a47268dba af8d3ee129bcf8e6d5e318207e33b391549e618f 3943 F20101108_AABGCF ramu_p_Page_086thm.jpg 381df977fc8572e9f29d168bbcf1f148 a76348bb450fc6da786bf6500f11e24c05df09fb 6509 F20101108_AABGBQ ramu_p_Page_072thm.jpg 7b79549340a3477f03e5cc0c38fbd9ab d107ddc4213d139d2e83185fec213bbe55de4e4d 1702 F20101108_AABFWK ramu_p_Page_004.txt d5cd8ff6723bae41e19f1556160c08dc 118100363dc46b3d87cb246e01c86af15fce0b7a 23118 F20101108_AABFVW ramu_p_Page_098.pro ec36a1dcf84e331f24569af7d3bb783d 6c2cd0e7f02476c8eadb910d91e06bb0d33b8dcd 1903 F20101108_AABGCG ramu_p_Page_087thm.jpg fb9344c0e54d3e5b8e0e8ec4e00eaea2 dacf46e098f184866b0f684b17f0f840fa2a80c7 6610 F20101108_AABGBR ramu_p_Page_073thm.jpg 6e678dd87e577d79fe6ebcbcd7693a7f f2647acb355371284c39a11418eb74aee72d5629 3179 F20101108_AABFWL ramu_p_Page_005.txt 9c9dee7d24860f22166450f281537826 3ece6dd5d3851401dcb3b9eae0640ee80b1aca4a 47604 F20101108_AABFVX ramu_p_Page_100.pro 137ed54d990359f0439197fee93cc888 7431dadbfdc25af4dc557e733464e32b74fd3550 3966 F20101108_AABGCH ramu_p_Page_088thm.jpg 2d42079a74bd8116078d6db783e13ab9 a482db9aabb4ca1b76b05b8931b616362e8d2548 21641 F20101108_AABGBS ramu_p_Page_076.QC.jpg 357a8f2e5d063d0a2616cc196b22784d 1bf54b4064d3a99c321316f115a3fcf91e9f5856 F20101108_AABFXA ramu_p_Page_041.txt 74aa2afbc9d9a731b49a238ddd4ace9d 6b7d5b8c4559a49668181d75a882aa0cbeb43735 2714 F20101108_AABFWM ramu_p_Page_010.txt cdf4731ff36583e0540f7ae0e665ffdf ba17c96f1e850a953f376f614851a06a8cb81b6f 47633 F20101108_AABFVY ramu_p_Page_101.pro b1ff856c575c9420178a381ce4de6bb3 d89587b4e617837f6cc8298418dd317027209623 2810 F20101108_AABGCI ramu_p_Page_089thm.jpg 1eb14e0ee34e03f6fb995c38e1505eed 34f0a1a8fb6f40fbf6a742dbbcf89c8365f38eec F20101108_AABGBT ramu_p_Page_076thm.jpg 6277d73dc1f73e538599a825203daa25 bce04b75099b156dae3ffb0dc30372e25fd54111 513 F20101108_AABFXB ramu_p_Page_046.txt d0ebf742dd0f0316fd5eec121c1fd188 1b05a86341aeb475e18fd7e4942e053934d771e8 337 F20101108_AABFWN ramu_p_Page_012.txt ccd909aafd11ed060e6e518dcff0693a 96f3c23ec305199b5f4e8d00e3c61b3613209dcf 18610 F20101108_AABFVZ ramu_p_Page_102.pro 1a0aa581bc3ca679b91c6445dd0bcb02 c6d709f34ccd12f12057d7f792250f79de6a6ebb 22946 F20101108_AABGBU ramu_p_Page_077.QC.jpg 6a07c71c450dd58178e49247f377b36c ee5e30596a2b716dd16eea36dc3d3b3f1d6acd0d 362 F20101108_AABFXC ramu_p_Page_047.txt ddf78a4826a79e70006a4abb95814530 6495f4b41f35e03e7de0017b0c501cf79f1379e6 2255 F20101108_AABFWO ramu_p_Page_014.txt e82fb1fe78985fcdd39209ca5dc72813 25111b14de286a58cf1d4ceadf92f181542f251b 7924 F20101108_AABGCJ ramu_p_Page_090.QC.jpg fa13b079d14b9df7498d62970c7d8a04 231db18bfe6763061b07c844556a9182a9d717a8 6489 F20101108_AABGBV ramu_p_Page_077thm.jpg b18cc7870306583663193bd1737c04b2 7c8c3e99d6fc0cd79c94a22765248816f19611f2 1756 F20101108_AABFXD ramu_p_Page_051.txt 840e32119341fc9f03a027e2164ac6e9 c3a7cc90ca4d10f3c6e25631b7c7c15794685fce 2160 F20101108_AABFWP ramu_p_Page_015.txt da4e064211f133e4e0e0e9866b1037e1 e598332f58c5ffc0dfde4b360a8dc38ba32bf020 2674 F20101108_AABGCK ramu_p_Page_090thm.jpg 041fba12f1fa97baae33622b848c3648 5f1bf67502941a4b37c163bc0b17c98940e0f807 24137 F20101108_AABGBW ramu_p_Page_079.QC.jpg b003aab6a9a8414c8ed2c303db457960 7e7d50f31cc24c25cbb498d9ad322832a64f4b07 2061 F20101108_AABFXE ramu_p_Page_053.txt 6c0b795e4657635667e52a4ae8da86e3 bc11ce02beadba9fba048cc8fba5c9dde1ca4c83 2178 F20101108_AABFWQ ramu_p_Page_016.txt c7bf6d979e7c0a032d25e5780f7105f2 2989fdf6b9817211a153d82fa25b1342d6b39d1d 3273 F20101108_AABGDA ramu_p_Page_102thm.jpg 693679c8a161aed299b3312fb45ca06c c9dbc43abf2b9c2246c61be1cb2032d6f83789c1 4381 F20101108_AABGCL ramu_p_Page_092.QC.jpg 8cf540d4e163b99999399993b5ac6674 4dfdbccb4a12c70c894b32c3259f7bf9f3d088dd 14281 F20101108_AABGBX ramu_p_Page_080.QC.jpg c6e6a27a07d85c1ae4c4b2cedc6630be 20751f77ff376e002424ff20ab002b5a637fed46 1992 F20101108_AABFXF ramu_p_Page_055.txt 8f1cf79658fc0abf763b35d6b78b4f59 05f15b4783aa9909b399cab725334bd4da5767f4 2218 F20101108_AABFWR ramu_p_Page_020.txt 9e85c391bd39daeed77ae4102ddb20ed 8843d9d4a8d6dcde1ae3ba8b8f4dbd9750a4ac77 18960 F20101108_AABGDB ramu_p_Page_103.QC.jpg 6ba5745ac18a9a0f78fec15eabffabda 01346e8407cc743f7e701c60c0f637f8d025e37b 1743 F20101108_AABGCM ramu_p_Page_092thm.jpg 562ee0d84d0c09cde5bd4f3629f3f13e 1855e2075ecf3cf4ee194e095cd8dc7792e8fa73 3095 F20101108_AABGBY ramu_p_Page_081thm.jpg 4fda36317fcab94310c9ed022f16413f 6116c2cd29f18eb62a123e81b0ab987aa4166ed5 1974 F20101108_AABFXG ramu_p_Page_056.txt 47378caf27758a51e15b9ace621eca6e 08dc1c63d7a91d585ff125cf8be4172c02b3d291 1969 F20101108_AABFWS ramu_p_Page_021.txt 383db3f0bcfc3fdda836f67434b57d85 f5cf27d2ad38e83f7413c94defcb437a20caeeb6 5517 F20101108_AABGDC ramu_p_Page_103thm.jpg c1c433896d90afbd09c14cf527f1c28f f490e6ab29244eb640aea10c8bbb5e579ad0e088 16968 F20101108_AABGCN ramu_p_Page_093.QC.jpg 5f558914e5c7a36c00939f5a5dff2c98 b4fa1bda994d37887544a8d44ecd896670be1d59 8301 F20101108_AABGBZ ramu_p_Page_082.QC.jpg 124c0840aee57837121545bad6512513 c5bdefadbf88fa19ccdcee38317a6f4329fd827b 2258 F20101108_AABFXH ramu_p_Page_057.txt dac4f5851f1e86f94365df3e5ca73712 a8e5b0961a8ef0d835f6a58a3f54ae00692a3ec0 321 F20101108_AABFWT ramu_p_Page_025.txt 26f8f8875576ae2ef5b520000eaa09ff 508ccf6d6e205786f314e27c79e90878090085db 25599 F20101108_AABFAA ramu_p_Page_099.pro 1bee3a2e52c73923e1da1a7ec0bf94c3 b6d6ea5f62426aaa228c050870d3a2c8505dafa2 6346 F20101108_AABGDD ramu_p_Page_104thm.jpg ef569fd84931d5eaf565706b4229f8e5 152739a4211379fecce1cb6372cfddb0caebbaf6 4693 F20101108_AABGCO ramu_p_Page_093thm.jpg c3d93b20c7fa531457ef49de4c707ebd f159528cff18e57704d4d72713e1ed6f596a9895 2039 F20101108_AABFXI ramu_p_Page_058.txt ddc28328f9572fda6e35688a1715b644 8c9d19aef53b76cdba407da6e760593bdd658938 2217 F20101108_AABFWU ramu_p_Page_027.txt a638f050ab709ae9ffde667557640c30 a1d75c27262e64a5576c32bcd66fe6ef2fee5075 6583 F20101108_AABFAB ramu_p_Page_075thm.jpg 6fbfdd0f68b7f15e2ad6048cf352d6fd 43144a8e2385e2ae038cebd1824d3170834ae91b 6009 F20101108_AABGDE ramu_p_Page_106thm.jpg 286f2d2e3dad68cb23813bd33d2dd051 c20ee9ca0053d23a1858f295aa4ca271702b6f74 22842 F20101108_AABGCP ramu_p_Page_094.QC.jpg 4d20ae479201d108933d23d066265c40 d61c7a7edda9e6c22d63c5cd8b2dc9798fd1b529 1541 F20101108_AABFXJ ramu_p_Page_059.txt 25049fedcb1419d8dfd9a1dd60ab686d 7068a851a06477f3fba47c600d3d170321b178ea 2116 F20101108_AABFWV ramu_p_Page_028.txt e0601173a3a7668c87dbed1b3a52a1e6 c26f2dfd514cdc8b194134c419331cae8256d341 8698 F20101108_AABFAC ramu_p_Page_085.QC.jpg 68f7f72dd0207077a8f4cf9b2df84467 676dedd9403b57d7c40be736ee9960bdb2155ec3 F20101108_AABGDF ramu_p_Page_107thm.jpg 90788ae539e5b1f8fe5f7cf9c50ddec8 43f1a5641e37a94f82f63dab3538fc72a5b70ce7 6484 F20101108_AABGCQ ramu_p_Page_094thm.jpg 49bf79d20aa562be8e8b361de006c9d7 4b75304e767def85e43d73883d1f96f6b38cf395 2107 F20101108_AABFXK ramu_p_Page_060.txt 33a87aed2bd28c0990c4c941462983ae 96661719a59f9fcf1d5a49bcac54445cc87ce60c 2045 F20101108_AABFWW ramu_p_Page_032.txt 70838f4835aaf8fa7d009d51913f3ca5 078d6fb464d38bf14b14d171f6b6bc68eb98f2d0 26391 F20101108_AABFAD ramu_p_Page_020.QC.jpg 15f946daf9ed83138b38a456dd3d0fee de72bf9f4110237ccd112237b191794f86a637c4 24879 F20101108_AABGDG ramu_p_Page_108.QC.jpg 75fa2e88cab89e198a5aa00b5ac69391 086cd7184d8385cf14b168ae9a03a51021f510b4 12087 F20101108_AABGCR ramu_p_Page_095.QC.jpg 1656527e0b09058b27cfaa41adfeae8e 108d0358e7397919a4b2d0b1ea448b0549868131 F20101108_AABFXL ramu_p_Page_061.txt 20dec850ddf13f676138e0d9b5870826 63585d73fd7b787a6d708765da207a6cc00aeed6 2091 F20101108_AABFWX ramu_p_Page_035.txt 87f7a898517f217f21c926be5a5e4399 46e02475b2ea819b9d02e5f2292741f47bd3bf84 23006 F20101108_AABFAE ramu_p_Page_105.QC.jpg e92065a00bbc7ab543105b79f98bb8df d2258f5cfeb159afdb8dc31c15b2cb4f23b2c5b2 13299 F20101108_AABGDH ramu_p_Page_109.QC.jpg 6a91c2c506161dd953abadc24186938d a1c07d2fc61b192c624b3e2a9f7f88392486870a 3791 F20101108_AABGCS ramu_p_Page_095thm.jpg 9ddf11cf09e9732e9e1d929a7e3ea3e4 2e766126303ce3cc48044743dd8232484b7b3f3c 541 F20101108_AABFXM ramu_p_Page_063.txt 79dcbabde4a90dd035cb32252d70d866 3f4c38675a460ff9e6d45f9890c00dddfffe87f4 2005 F20101108_AABFWY ramu_p_Page_037.txt 8f5c8411af801203f92f3c0144fc4d78 b792d7c223a0234b64301853f169e2c516dba66b 74243 F20101108_AABFAF ramu_p_Page_035.jpg 71d1f190b7eed6e66ba4e18ddae480bf 3a01a0acda794c54e93903eba5ef63a96037fa2c 277 F20101108_AABFYA ramu_p_Page_083.txt 9fbe32fa43e8ceab599b5e1d0b0e9ff6 a7d2d04d65937d0b254e388a63dc931516f39877 4127 F20101108_AABGDI ramu_p_Page_109thm.jpg faedb877e50f7d8baafd522145c87801 280b442c6095610826a0272fc71f4b220bb47f1e 17384 F20101108_AABGCT ramu_p_Page_096.QC.jpg c19b4eced92d21e30f3b02aa7986e102 13b77fce2d70d0f995dd8208ebb453764aa2ef91 194 F20101108_AABFXN ramu_p_Page_064.txt ba22cd09259cf609af10eea6571c4ec7 f267d7687b660ca03fe2f7e9269800faa86993a7 1907 F20101108_AABFWZ ramu_p_Page_039.txt 2487cd538f83f08ad433b687c084c62d 27347d0bd5c24f59a82f271a3876bf77d8d885b7 F20101108_AABFAG ramu_p_Page_099.tif d1830f8375b383e78ac705296972781c 481a0129e5ad0ad68ef4fb13e8c33367dd789ca5 583 F20101108_AABFYB ramu_p_Page_085.txt 0530605d220cf7aef2540a3267eb1400 071389827f87aaf784619e72119e0ab4ed8f8c45 13609 F20101108_AABGDJ ramu_p_Page_110.QC.jpg e18c921e143f8af71b3d03ea484dcb68 891f3d13ebf11481d3ea106f35de78c4eca0b77a 15502 F20101108_AABGCU ramu_p_Page_097.QC.jpg e59d10cef7504b0e51185bd3856f49d6 9e92f297a28c810f479c99e300aeb212b15a82a1 727 F20101108_AABFXO ramu_p_Page_065.txt 1fd79680e55a05be680a59a2de157d53 c4577fe9a1bdcf2366e844bb137f96ca61264b31 73150 F20101108_AABFAH ramu_p_Page_019.jpg 2f11a043c248b9c5d796206d1362c114 b7628988f386254410db782c5ccc8d13a25d2f99 1524 F20101108_AABFYC ramu_p_Page_086.txt 110c8c3678886e00b6f9d7911385a312 b73711440fb588a70d88f588d7846e620f91521d 4414 F20101108_AABGCV ramu_p_Page_097thm.jpg 208642f848f3d594c64e1fc529a1804d 14d9ecfa954879255e06479313754adf36f7fcab 313 F20101108_AABFXP ramu_p_Page_067.txt c4419646d5bcbacb60c0b34f25a38971 2f18cc91fc8d12c651fd9289275e64ba6a78fa13 347 F20101108_AABFYD ramu_p_Page_087.txt 37b833119a876b3791def189f636bd6f b10eeb052858b95e4447d98ca6c6d80b57e8a5de 7058 F20101108_AABGDK ramu_p_Page_111.QC.jpg a1b3643b6fd43b3e4ff75d4f79921ea8 c6d68691defb59de6b1411e46259869264a29d4f 13990 F20101108_AABGCW ramu_p_Page_098.QC.jpg 7e70203218605353d64b2fbb760fe979 47c6dbd40e61004871bf081fe7f00be32f376f34 1160 F20101108_AABFXQ ramu_p_Page_068.txt 89d69e6f7257d19d0bba1198f43134d5 2b96f9265d5d67e8fe8862e153f338db9331eb8d F20101108_AABFAI ramu_p_Page_117.tif 01482e401d88e8cc334df6f10d6d5389 47bd03018fa8cc8819155edfaa8792d9f57fd5f9 1279 F20101108_AABFYE ramu_p_Page_089.txt ccf50e99b92b11bbd1ad1f7eb61dc7f5 10c73f55adc4dd79c4ac9254efd6af3b607a2cbf 24968 F20101108_AABGEA ramu_p_Page_124.QC.jpg 24d847c0917f4a11181c477c24ed095d 945154ee15971e545e4a862fd810c37914e48ccf 10562 F20101108_AABGDL ramu_p_Page_112.QC.jpg 18fe682c42429dfc7bb4ef483e20c54a ea7e1161c32af133fe0649a061263ce9660ad1d6 4229 F20101108_AABGCX ramu_p_Page_098thm.jpg e4ed2d3f79eea553f440f70bf67ae511 5377e7202e4591d6d279c3e258aa9d34a2093141 2247 F20101108_AABFXR ramu_p_Page_071.txt de78428561d5b803bc7e6a5ea82f315d 4977a40f2bcb4a7c40541ff7cd512b020e9459c6 F20101108_AABFAJ ramu_p_Page_021thm.jpg f18b1fe9c111bdeb22bd76eae7c33bf6 67ff97bf4a86cb568134f845d2cbf14fee60838c 338 F20101108_AABFYF ramu_p_Page_090.txt d0f62c6abac3b8ed943f6e4802e3ea79 cfd7059ef0c15eab4acc2a67e0d986711921df1b 6951 F20101108_AABGEB ramu_p_Page_124thm.jpg f98c26306b44c2b631bb587d9e6df502 7069419949cf6771a4cc645ab3b87e25a0a4f3e2 3424 F20101108_AABGDM ramu_p_Page_112thm.jpg d3b75df1987f50658522cdc551c2f417 730f1b9355740194d489870c43972d97f35ab70c 23584 F20101108_AABGCY ramu_p_Page_101.QC.jpg f61429b7bab3021a2c82a57b480c1900 dea95db24dedc28b71b264445bec393e3f48e23e 2494 F20101108_AABFXS ramu_p_Page_072.txt 3a7dbc22718963470080818c770fc6ae 5658ce255b46e13f8c3bf9bf73740d26b9a50ad5 F20101108_AABFAK ramu_p_Page_020.tif e833989b63efd14785cab8f1bbbad38c 125e17900df40e98edeb0758a9bde6cff13f9c39 1401 F20101108_AABFYG ramu_p_Page_093.txt 7f25f6afde84c2531eecd6185ed139c4 02346681b2e9edadb8687a01de0658b4145428e2 8740 F20101108_AABGEC ramu_p_Page_125.QC.jpg bdfbb0b1f73d95b1ce41681c4c859152 d8142e1179184d8c07c11e89e145a89d0ed2274b 10666 F20101108_AABGDN ramu_p_Page_113.QC.jpg 457e0068a897f314d8b3f0798b0b3537 4dadf7bcad480b77fabe4b8da61f21c8c5b5cae6 10596 F20101108_AABGCZ ramu_p_Page_102.QC.jpg cfcc0493b21c66c1a5c5ba7ae6676f96 24c9beca5728bf16524ea95cbb0dac7512de4632 1681 F20101108_AABFXT ramu_p_Page_074.txt 558a8d152f2611909dec12d9bdb9d73d 3febf4334dd79cd384d6508cbaa420e6d3e770c8 32091 F20101108_AABFBA ramu_p_Page_112.jpg 9ded7d6cbeb147cdf0dfb17ed6e224d7 a94e8e4b5e2984b8fff6efbbdb2382a42ae86ec7 F20101108_AABFAL ramu_p_Page_104.tif d783acef10cd6e94013eef674f9bdfa9 05cf343f1ecd32ecf566702d5fe6927851a3e76f 1066 F20101108_AABFYH ramu_p_Page_095.txt 224e0d50328043fdd88c2af032db23d5 298194c7d6f0fc7fa0ab79122ddb2aa071a2ecae 3673 F20101108_AABGED ramu_p_Page_126thm.jpg 420ea4e36064fdeb3196bdf3c7bea472 928b26ad88eb221abcb8fa946b28f6b824e7eb06 3484 F20101108_AABGDO ramu_p_Page_113thm.jpg 830b95076e5b1625442dd36fccfb82a7 ee2801a4b326416ce5b4f5dcc32075a1d1bc6f4b 1839 F20101108_AABFXU ramu_p_Page_075.txt fe3ce3dcec6a48c856942e0c19ab75b3 e9e90fb692124f931db22c0176c901f97670de25 1990 F20101108_AABFBB ramu_p_Page_017thm.jpg fc39ec75860a15d7c43ce096b7d7dcd5 3e0c14fe3679b62793a65514768faa93ee060c2e 75142 F20101108_AABFAM ramu_p_Page_051.jp2 d1dc9d8af7966388586d37ac5967a2e0 e8404515ebdf2f81482934eedd8aa85c0b00cf0f 1228 F20101108_AABFYI ramu_p_Page_097.txt 96b9839ba3b7e8f3e75e9f872fac2e4b 2f7e8d3ffa91d1e7891629606093b0be7cf1e621 144459 F20101108_AABGEE UFE0021707_00001.mets 140fb6ab518a0d16a917d13e8f62b8c6 7556733c54a6c96930b84e167805b0f0f40b6960 9679 F20101108_AABGDP ramu_p_Page_115.QC.jpg c4f79c103ca2426c54d4e3b6c912b283 03e7c501905790d7d45924c1d971395e447bdf8e F20101108_AABFXV ramu_p_Page_078.txt 3a1e749e0856085e4be9b6a8e17b0aca 3ef7fbbe21510d932bd021fb001b214f891fee83 6832 F20101108_AABFBC ramu_p_Page_019thm.jpg 860dd27ac2fb0903eae8499131e88697 283e232b8037ec84db2506843e9555aa750336d2 50215 F20101108_AABFAN ramu_p_Page_022.pro 40e247f1a99ad9775cc78f08cc8e9534 1f232ca1fddb2c51ed5a2798f6ff50e631bb9669 F20101108_AABFYJ ramu_p_Page_101.txt 49fa69d6fa74123de2fca093eec519ec 456b6763e976a506df6555d6a7d5e232834c27f6 11191 F20101108_AABGDQ ramu_p_Page_116.QC.jpg b6377bad6fbd0bf58d1cb70af1368c4d fa5f29e1084d31e626b3f129bc9af944b30fa7d1 2051 F20101108_AABFXW ramu_p_Page_079.txt 046d96e1939fb4d7c6838466c3f20d0b 17588fcbde0e856dddc82960bf18dab7b31cdec3 F20101108_AABFBD ramu_p_Page_036.txt 53f026124c2b6cb1af13f7dade30c2d2 a45cbafba49409e3d7a763022253a0aa00cdd226 6911 F20101108_AABFAO ramu_p_Page_016thm.jpg c647c73101334170b0071f0cbf2d9857 4322cf42a8d03cc00e165da5e14c2c1d541aab1c 2093 F20101108_AABFYK ramu_p_Page_104.txt 1f2106a0f3d2e0611b4048b0952bbc68 f537c1cbd300ba25af1efdd5bbed76042942a6be 3448 F20101108_AABGDR ramu_p_Page_117thm.jpg c167ff0d15d8dd428d0de6ad9ff74df2 851d2a692534f3dc62790bb57193d9f5024a4d8c 1081 F20101108_AABFXX ramu_p_Page_080.txt 1008e54ae7f5aa04c037ad8481948294 76e9bf27e8c7c608da440f4d1dcb9c6d4c51ae8b 2294 F20101108_AABFBE ramu_p_Page_001thm.jpg 88cb9057f5489f7dcc612911a3d03dd8 c72e1b2f7db0632efb9d4f887cc0d56b43ad0d5d 3757 F20101108_AABFAP ramu_p_Page_091thm.jpg b74802e32f63645a52d1c8b6c076ee0f 8614d77695c8d6597cd5c455b96aaf7fcc0e338b 6717 F20101108_AABFZA ramu_p_Page_009thm.jpg d63295f50309466eaeff893f78417216 195867a04ba3c2b8dd0680cd00bc881e3b2a8e27 318 F20101108_AABFYL ramu_p_Page_107.txt 834bd950869a33b6b837a8bc2b85faf9 a423eab5aac581c6c034133d3344494e316993ef 6592 F20101108_AABGDS ramu_p_Page_118.QC.jpg 7c49c6892728184474c06ee8e7e7c40e 9e8e99a3a6cd733ab75684080764ce53d2e6145f 215 F20101108_AABFXY ramu_p_Page_081.txt 73590659450d36c5cdb47003173f7987 01bb642c2527076f31f4c00eb28d132d33e1d6df F20101108_AABFBF ramu_p_Page_069.tif 4e9ff1412432ba299bee6e59f1d5e4d2 983ace712f4d677a62ebf67a4aea0ce7fb85aa9e 52284 F20101108_AABFAQ ramu_p_Page_053.pro d0840552edcebbb99c073b97ee6bd340 39130d43aa314a2d8a74f978a86438aca3298081 2133 F20101108_AABFYM ramu_p_Page_108.txt 499694fc70a97b4b5bfb65b8803473f7 a589b83cc21d350a0ce4068eccbef4734dfbf2db 6074 F20101108_AABGDT ramu_p_Page_119.QC.jpg 357886fa7750520ed8b6f5d03f8e43f6 c16d56bf67b41ea340699728e06ea253e47b51d7 437 F20101108_AABFXZ ramu_p_Page_082.txt a52777265084e82aa3e873a789748a57 846a14ea48746200408414ac4a1478a5bed640fb 377 F20101108_AABFBG ramu_p_Page_008.txt fb4d17d5f013234ab310823fc5363535 39ffa1ff7abb1932a8536e69cd663dfea78d4df1 F20101108_AABFAR ramu_p_Page_046.tif 2a1263f1f2a19bde0c2f52ce8d63d5a2 03e473cf56377c689e2a49479394e7889bbab281 24572 F20101108_AABFZB ramu_p_Page_010.QC.jpg d70e842b34ef193b24d135f7da35987e 4897ed3efbae21ce2f3f5897fdee3c2f851ba398 891 F20101108_AABFYN ramu_p_Page_115.txt 13add380290f7a2385a1cbef77f7b9d7 f075da2ab03b41301c71263c9e48f86a45ddecc5 2345 F20101108_AABGDU ramu_p_Page_119thm.jpg 1e6d1bef4339ec6c54c505f9c4405dc9 39ee5292bfce5c034e6dbf6e76d92b701ffe94ce 2267 F20101108_AABFBH ramu_p_Page_007.txt 8250e1922f7d742de01ff9ceb2b76280 e47a1d53f207067fd39399e077524f6171137f98 67512 F20101108_AABFAS ramu_p_Page_101.jpg d6d24a0bc9251f7b5ddfd2c614f1983d 764e91877f4d453503262d93deb103d46f5697f3 6021 F20101108_AABFZC ramu_p_Page_010thm.jpg a8751fe0b71aae405b7618d22a08029d 83ce5291f38360140e78ecda5d774abb52514144 1084 F20101108_AABFYO ramu_p_Page_116.txt 76cd0dd61cf9be83f39611742ec3a446 2e9e32121bca3a7d2fe9e7b61918cec8278e721b 5819 F20101108_AABGDV ramu_p_Page_120.QC.jpg 282edebae80ab2e2cc08a7d98269e257 fc332da2822b456f35c00033553c8fc4460c34db 23246 F20101108_AABFBI ramu_p_Page_075.QC.jpg 6d7294805173b628bdd6753434a80aec 219edbdc2d93fc69d9492bd043cb05c08cff1849 14662 F20101108_AABFAT ramu_p_Page_087.jpg 2cf9016d3919f0749f9884cab1487786 ff6ad9117a68743e5f96d5e9825aa87ced77e6de 6542 F20101108_AABFZD ramu_p_Page_012.QC.jpg 228a30a764956a42078c132719bfcec1 78003526f1d483e739338e2fb8bdda58062746d7 1173 F20101108_AABFYP ramu_p_Page_117.txt 3d1c93a1a311c0cfbcac37984b070eee 407b29fbfe826280a9015f1b7a94279f3f5ce81e 2265 F20101108_AABGDW ramu_p_Page_120thm.jpg c9f2f004354e9d078b469dfc85a29dd3 311230c4594c8be00800dc27b9e07c2fc1de82bf F20101108_AABEWC ramu_p_Page_082.tif f182532ce3c48a000919e39bf9eb96a3 cd3cf20be3ca90424bd72381232c7418088db45a 10760 F20101108_AABFAU ramu_p_Page_117.QC.jpg 9125264314f665a85dc472caa9c4511e 9c4855e75dda6ef5af151d0e74c5c58e3123f399 21393 F20101108_AABFZE ramu_p_Page_013.QC.jpg 4c16d4196980ac4cd5830c7da950cc2e b9947da6952154893157d14006c85e4320634bf0 489 F20101108_AABFYQ ramu_p_Page_118.txt 076612903c8bc1fe6fbb28b4ffc74886 70e67a3504c699024f2e7884211444dbcd4e2af0 26581 F20101108_AABGDX ramu_p_Page_121.QC.jpg 2f8ad2901c47028225016f49fe1a7e99 f99e9acc3016083c91f0534d084f863f7a6007b7 55470 F20101108_AABFBJ ramu_p_Page_016.pro 34bbd1a4ec944809ccc4200a673d6566 8a621ad771f69039d3f43beb417098bdbde31284 25195 F20101108_AABEWD ramu_p_Page_015.QC.jpg 0e2ffa2ce92fd1fd94a8114c510a947e 846e261d9f640168d084a8cf4a4471d002256307 70340 F20101108_AABFAV ramu_p_Page_094.jpg b1c749763af9cdac9432bc1571230ba3 e140595cf377a194d0f52a872f01873caaf44b72 5862 F20101108_AABFZF ramu_p_Page_013thm.jpg 399472b389eee6f849027e12abe732c5 527ab6574a9c571b3fde6b6f66a0365579550792 425 F20101108_AABFYR ramu_p_Page_120.txt 8158ab5e833f0434ad6ee0b4c7a3d1fc f28f40dc9cf7de3a51d20a46c2dd4f4f7cbf8265 7313 F20101108_AABGDY ramu_p_Page_122thm.jpg f45ef611f2b91e535a17e6959c7744e9 742d508369a118b72a7988d51c674131106fb71d 106686 F20101108_AABFBK ramu_p_Page_022.jp2 79c2575755ac3505b244a218e3874618 e528a4e7e18de68ee2d346005312f170bbe82a23 139314 F20101108_AABEWE ramu_p_Page_122.jp2 7aa54980dde7fb3ded1393ff427b02b2 698bffb3baec9a6898999a022dc100ed5a303744 67113 F20101108_AABFAW ramu_p_Page_013.jpg 882bbd0a5d230f6ff356fa1c7cd3a4f0 116cd882dcdbd0991844c1a4b8925e74604dfe47 6654 F20101108_AABFZG ramu_p_Page_015thm.jpg d7671b483b96318ce45052ae4018d752 452fe70b5976c64787cb96eb6001e652485ba0b5 2424 F20101108_AABFYS ramu_p_Page_121.txt 130867ab90b290b774b852245c0af9ce ac05d0f92a48567943ed7c8e125076b71a73db50 6966 F20101108_AABGDZ ramu_p_Page_123thm.jpg db10ecee6b299cb123f5785347f7ed84 922917d9bf7f938c01b81e4f9f7ddd41f27c2611 45499 F20101108_AABFBL ramu_p_Page_065.jpg cc372830b528146a81847b92206d57dc 0885f8e0800720eb77042db56f57e58b5340df37 2618 F20101108_AABEWF ramu_p_Page_125thm.jpg 84ca0dd24825f0ef6d04d298e33f086e 39e7b59ec08d957b107ce686c1b5d3545bb9611d 5711 F20101108_AABFAX ramu_p_Page_074thm.jpg 5e0e03b26a13bcaa72b4377b1f237cc6 a2bef6a410f8153dca28c691d6645ee3af1093f5 6655 F20101108_AABFZH ramu_p_Page_018thm.jpg d7d22528e8ce09616154f78221433345 65a21f1b87810fe6de0b53cf5314637d472fc715 2513 F20101108_AABFYT ramu_p_Page_122.txt 7949c40ec45d39c98b4efd75b8d6fac8 d682727383c919c70034e066eff56df436f8a25c 18461 F20101108_AABFCA ramu_p_Page_059.QC.jpg 135aeb237a1febb0af43b986d44ee7f8 0cc9a22b5f1bb70ebf5fa651e09bfc3fabd9252c 31720 F20101108_AABFBM ramu_p_Page_096.pro f7dfbddee606c824b9ec2291aed378d6 51eac7ea1f012798d16c8eb85cfcb69d29c68319 45844 F20101108_AABEWG ramu_p_Page_042.pro 30921a8597cc64b7fc4e880214b14024 8ddb02cbf63c28fe0faf84bdd3eca20ebf6d7170 24775 F20101108_AABFAY ramu_p_Page_071.QC.jpg 5357c33e3235e31eb1c6c99035abeb8c cd5a252333afe7536289341defef62ac2eece588 23522 F20101108_AABFZI ramu_p_Page_024.QC.jpg 6f38fcc9c56a719152c2c10c0652371e 3586ae572a895639853f636463992d25a7fa9b2a 3247 F20101108_AABFYU ramu_p_Page_002.QC.jpg bb495f83a8c4187b72fa866525b17a36 0d47f10b6ce566d5ff2d81b5c4f53bac2a10ca58 7347 F20101108_AABFCB ramu_p_Page_083.QC.jpg e24def5ccef7bea56eb1772bad44f8ba 1fd7c909d9797eeb033d07d7407aeae1abb39ca9 MULTIPLE TAIL MODELS INCLUDING INVERSE MEASURES FOR STRUCTURAL DESIGN UNDER UNCERTAINTIES By PALANIAPPAN RAMU A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2007 O 2007 Palaniappan Ramu Dedicated to my parents Meenakshi and Ramu ACKNOWLEDGMENTS This dissertation would not have been possible if not for the help, support and motivation by my teachers, family and friends. I am extremely thankful to my advisors Dr. Nam Ho Kim and Dr. Raphael T. Haftka for, providing me this wonderful opportunity to pursue a Ph.D, their constant encouragement, patience and excellent guidance. I learnt a lot both in academics and personal life from them. They have been more than mere advisors friends, philosophers and mentors. I have always admired their indepth understanding and knowledge and feel that I am fortunate to have worked under their guidance. I would like to thank Dr. Peter Ifju, Dr. Tony Schmitz and Dr. Stanislay Uryasev for agreeing to serve on my committee, reviewing my dissertation and for providing constructive criticism that helped to enhance the quality of this work. I would like to thank Dr.Choi, Dr.Missoum, Dr.Qu, and Dr.Youn, for collaborating with me. I also thank the staff of the Mechanical and Aerospace Engineering department for their help with administrative support Thanks are due to my former and present colleagues in the Structural and Multidisciplinary Optimization group. I thank all my friends, especially Gators for Asha group members, for making my life outside work wonderful and enj oyable. I greatly appreciate the financial support that I received from the Institute for Future Space Transport Last, but not the least, I would not have completed this work if not for the unconditional love, emotional understanding and support of my family. To you, I dedicate this work! TABLE OF CONTENTS page ACKNOWLEDGMENTS .............. ...............4..... LIST OF TABLES ............ ...... ._._ ...............7.... LIST OF FIGURES .............. ...............9..... NOMENCLATURE ........._..... ...............11.____........ AB S TRAC T ............._. .......... ..............._ 13... CHAPTER 1 INTRODUCTION ................. ...............14.......... ...... 2 RELIABILITY BASED DESIGN OPTIMIZATION ................. ............... ......... ...18 Introducti on ................. ...............18................. Standard Formulations ................. ...............18................. MomentBased Methods .............. ...............21.... Monte Carlo Simulation .............. ...............22.... 3 INVERSE RELIABILITY MEASURES .............. ...............26.... Literature Review .............. ...............26.... Birger Safety Factor............... ...............31. Probabilistic Sufficiency Factor .............. ...............31.... Probabilistic Performance Measure ................. ...............33........... .... Inverse Measure Calculation ................. .. ........ ............3 Simulation Approach Monte Carlo Simulation ................. ...............35............. Analytical Approach Momentbased Methods .............. ...............37.... Reliability based Design Optimization with Inverse Measures ................. ............... ....39 Beam Design Example ............... ...............39.... Design for Stress Constraint ................. ...............40................ Comparison of Inverse Measures ................ ............ ......... ....... ........ ... ...... ..........41 Use of PSF in Estimating the Required Change in Weight to Achieve a Safe Design ..........42 System Reliability Estimation Using PSF and MCS ................ .............. ......... .....43 Design for System Reliability by MCS and PSF ................. ...............44............. 4 TAIL MODELLING AND RELIABILITY ANALYSIS ................... ...............4 Tail Equivalence and Generalized Extreme Value Theory .............. ..... ............... 4 Generalized Pareto Distribution .............. ...............50.... Threshold Selection .............. ...............52.... Parameter Estimation ................ .............54.................. Maximum Likelihood Estimation............... ...............5 Least Square Regression............... .. ... ...... ................5 Accuracy and Convergence Studies for the Quantile Estimation............... ...............5 Cantilever Beam Example .............. ...............57.... Tuned MassDamper Example .............. ...............59.... Alternate Tail Extrapolation Schemes ............ .....__ ...............61.. 5 MULTIPLE TAIL MODELS .............. ...............71.... M multiple Tail M odel ............ ..... .._ ...............71... Numerical Examples.................... .............7 Standard Statistical Distributions .............. ...............72.... Engineering Exam ples .............. ...... ....... ...... ...._ .. .. .. ..........7 Application of multiple tail models for reliability estimation of a cantilever beam........76 Application of multiple tail models for reliability estimation of a composite panel.......77 CONCLUSION S............... ..............9 APPENDIX A PSF ESTIMATION FOR A SYSTEM RELIABILITY ....._____ .........__ ...............94 B GENERALIZED EXTREME VALUE THEORY ....__ ......_____ ...........___.........9 C MAXIMUM LIKELIHOOD ESTIMATION ....__ ......_____ .......___ ............9 D ERROR ANALYSIS FOR GPD .............. ...............100.... E ACCURACY AND CONVERGENCE STUDIES FOR NORMAL AND LOGNORMAL DISTRIBUTIONS ................. ...............103................ F ACCURACY AND CONVERGENCE STUDIES FOR THE CANTILEVER BEAM......108 Stress Failure M ode ................ ...............108................ Deflection Failure Mode ................. ...............108................ G STATISTICAL DISTRIBUTIONS ERROR PLOTS ................. .............................111 REFERENCES .............. ...............121.... BIOGRAPHICAL SKETCH ................. ...............126......... ...... LIST OF TABLES Table page 31 Different approaches to prescribe the probabilistic constraint .............. ....................3 32 Random variables for beam problem ................. ...............40............... 33 Comparison of optimum designs for the stress constraint ................. .......................41 34 Comparison of inverse measures for w=2.4526 and t3.8884 (stress constraint) .............42 35 Comparison of MPP obtained from calculation of PSF. ......____ ... ....._ ..............42 36 Design for System Reliability by PSF, Qu and Haftka (2004)** ........... .....................45 52 Summary of the performance of individual methods and MTM for different di stributi ons ................. ...............75........... .... 53 Summary of the error to range ratio for different distributions ................. ................ ..75 54 PSF reciprocal estimates and standard deviation at different reliability indices ..............77 55 Mechanical properties of the composite laminates ................. ............... ......... ...80 56 Coefficient of variation for random material properties ................ ............... ...._...80 57 Mean of random variables .......... ...._._.....__ ................ ........._.__.....80 58 Deterministic optima found by Qu et al (2003)............... ...............80. 59 Composite laminate. S, reciprocal estimates and standard deviation at different reliability indices............... ...............80 A1 Distribution of random variables. Multiple failure modes example ............... ............._....95 A2 Contribution of failure modes to the system failure ................. ............... ........ ...95 A3 P SF at different target failure prob abiliti es .....__.....___ ..........._ .........9 D1 Modelling error(F,(z)F, (z)*....... ) ..... .. ......... ........ ...._ .... ........102 D2 Sampling error (F ~(z)* F ~(zS) )........_........._ ......... ...._..... ...._........102 E1 Accuracy of the tail modeling approach to estimate quantiles. Normal distribution....... 104 E2 Accuracy of the tail modeling approach to estimate quantiles. LogNormal di stributi on ................. ...............105................ E3 Accuracy of the tail modeling approach to estimate quantiles. Higher thresholds. LogNormal distribution .............. ...............106.... LIST OF FIGURES figure page 21 Reliability analysis and MPP. Subscript refers to iteration number ................. ...............25 31 Schematic probability density of the safety factor S. ............. ...............46..... 32 Schematic probability density of the limit state function. ................ .......................46 33 Illustration of the calculation of PPM with Monte Carlo Simulation for the linear performance function ................. ...............47................. 34 Inverse reliability analysis and MPP for target failure probability 0.00135 ( P = 3)........47 35 Cantilever beam subj ected to horizontal and vertical loads. ...........__... ...._._...........47 41 Tail modeling of F(u) using the threshold u. The region of y>0 is failure ........................64 42 Convergence of PSF at different thresholds. A) MLE B)Regression. Cantilever beam system failure mode. 500 Samples. 100 repetitions............... ..............6 43 Tuned vibration absorber ........._.__ ......._._ ...............66... 44 Normalized amplitude vs rl and rz............... ...............66... 45 Contour of the normalized amplitude ................. ......... ...............67. ... 46 Convergence of PSF at different thresholds. Tuned Mass Damper ................. ...............68 47 Tuned Mass Damper. Comparison of tail models using regression and ML method...._...69 48 Transformation of the CDF of PSF reciprocal (S,). A) CDF of S,. B) Inverse Standard normal cumulative distribution function applied to the CDF. C ) Logarithmic transformation applied to the reliability index. .............. ....................7 51 Lognormal distribution. Classical tail modeling techniques MLE and Regression........8 1 52 Lognormal distribution. Linear fit to tail (LT) and Quadratic fit to half of the data (QH) ........._.__...... ._ __ ...............82.... 53 Lognormal distribution. Quadratic fit to the tail (QT) ................. ......... ................83 54 Lognormal distribution. Multiple tail models ......___ .... ... ._ ......___........8 55 Lognormal distribution. Multiple tail modelsextrapolation region ................. ...............85 56 Normal distribution. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices ................. ...............86........... .... 57 Normal distribution. Boxplot of r ratio. ............. ...............87..... 58 Cantilever Beam. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices............... ...............88 59 Normal QQ plot of the residuals for the normal distribution ................. .....................89 510 Cantilever beam. Boxplot of rl ratio ................. ...............90........... .. 511 Geometry and loading for the composite laminate. Xhoop direction, Yaxial direction .............. ...............90.... 512 Composite laminate of cryogenic tank. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices... ........... ...............91...... 513 Composite laminate of cryogenic tank. Boxplot of rl ratio. ................ ..................9 A1 Two variable system. Algebraic multiple limit state functions example...........................95 F1 Convergence of PSF at different thresholds. A) MLE B) Regression. Cantilever beam stress failure mode. 500 Samples. 100 repetitions .............. .....................0 F2 Convergence of PSF at different thresholds. A) MLE B) Regression. Cantilever beam deflection failure mode. 500 Samples. 100 repetition ................. .........._.. .....110 G1 Lognormal distribution ............ ...... .__ .............. 112.. G2 Gamma Distribution ........._.___..... .___ ...............113... G3 Extreme Value type 1 Distribution ................. ......... ...............114 ... G4 Uniform Distribution. ................ ...............115................ G5 Rayleigh Distribution. ........._.___..... .__. ...............116... G6 Exponential Distribution. ........._.___..... .__. ...............117... G7 Lognormal Distribution. ............ ...... ...............118.. G8 Gamma Distribution ........._.___..... .___ ...............118... G9 Extreme ValueType 1 Distribution. ............ .....__ ...............119 G10 Uniform Distribution. ........._.___......_. ...............119.... G11 Rayleigh Distribution. ........._.___..... .__. ...............120... G12 Exponential Distribution. ........._.___..... .__. ...............120... NOMENCLATURE A Cross sectional area C Capacity C Scaling factor CDF Cumulative Density Function DUU Design Under Uncertainties FORM First Order Reliability Method G Limit state function GPD Generalized pareto distribution M Number of failure modes MCS Monte Carlo Simulation MPP Most Probable Point MPPIR MPP Inverse Reliability MPTP Minimum Performance Target Point MSE Mean square error PMA Performance Measure Approach PPM Probabilistic Performance Measure PSF Probabilistic Sufficiency Factor R Response RBDO Reliability based Design Optimization RIA Reliability Index Approach RSA Response Surface Approximation U Variables in standard normal space u Threshold value t Thickness w Width X Load in x direction Y Load in Y direction /7 Reliability Index (D Standard Normal Cumulative Density Function cr Scale parameter 5 Shape parameter z Exceedance data Fu(z) Conditional CDF F~; (z) Conditional CDF approximated by GPD Pftarget Target probability of failure Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy MULTIPLE TAIL MODELS INCLUDING INVERSE MEASURES FOR STRUCTURAL DESIGN UNDER UNCERTAINTIES By Palaniappan Ramu December 2007 Chair: Nam Ho Kim Cochair: Raphael T. Haftka Major: Aerospace Engineering Samplingbased reliability estimation with expensive computer models may be computationally prohibitive due to a large number of required simulations. One way to alleviate the computational expense is to extrapolate reliability estimates from observed levels to unobserved levels. Classical tail modeling techniques provide a class of models to enable this extrapolation using asymptotic theory by approximating the tail region of the cumulative distribution function (CDF). This work proposes three alternate tail extrapolation techniques including inverse measures that can complement classical tail modeling. The proposed approach, multiple tail models, applies the two classical and three alternate extrapolation techniques simultaneously to estimate inverse measures at the extrapolation regions and use the median as the best estimate. It is observed that the range of the five estimates can be used as a good approximation of the error associated with the median estimate. Accuracy and computational efficiency are competing factors in selecting sample size. Yet, as our numerical studies reveal, the accuracy lost to the reduction of computational power is very small in the proposed method. The method is demonstrated on standard statistical distributions and complex engineering examples. CHAPTER 1 INTRODUCTION Uncertainty is an acknowledged phenomenon in the process of structural design. In an optimization framework design under uncertainties refers to a safe design that not only needs to be optimal but should also warrant survivability against uncertainties. Traditionally safety factors were used to account for the uncertainties. However, use of safety factors does not usually lead to minimum cost designs for a given level of safety because different structural members or different failure modes require different safety factors. Alternately, probabilistic approaches offer techniques to characterize the uncertainties using a statistical method and have the potential to provide safer designs at a given cost. However, the probabilistic approaches require solving an expensive, complex optimization problem that needs robust formulations and efficient computational techniques for stable and accelerated convergence. In addition, they also require statistical information which may be expensive to obtain. Structural reliability analysis requires the assessment of the performance function which dictates the behavior of the structure. The performance function is called the limit state function which is typically expressed as the difference between the capacity (e.g, yield strength, allowable vibration level) and the response of the system (e.g, stress, actual vibration). Approaches available for reliability assessment and analysis can be widely classified as analytical and simulation approaches. Analytical approaches are simple to implement but are mostly limited to single failure modes, whereas simulation methods like Monte Carlo simulation (MCS) are computationally intensive but can handle multiple failure modes. Moreover, they can handle any type of limit state functions unlike analytical approaches which are mostly appropriate for linear limit state functions. Most real life applications exhibit multiple failure modes and the limit state function is not available explicitly in closed form. Since there is no information on nonlinearity of the limit state function, MCS is the obvious choice in such situations. In reliabilitybased design optimization, a widely used probabilistic optimization approach (Rackwitz, 2000) in structural engineering, reliability analysis is an iterative process and using crude MCS is computationally prohibitive. Researchers develop variants of MCS or other approximation methods like response surface that replaces the reliability analysis and obviate the need to repeatedly access the expensive computer models In reliability based design context, high reliability, typical of aerospace applications translates to small probability in the tails of the statistical distributions. Reliability analysis when dealing with high reliability (or low failure probability) designs is mostly dependent on how the tails of the random variables behave. In few cases, the safety levels can vary by an order of magnitude with slight modifications in the tails of the response variables (Caers and Maes, 1998). Therefore, the tails need to be modeled accurately. Limitations in computational power prevent us in employing direct simulation methods to model the tails. One way to alleviate the computational expense is to extrapolate into high reliability levels with limited data at lower reliability levels. Statistical techniques from extreme value theory referred to as classical tail modeling techniques here are available to perform this extrapolation. They are based on the concept of approximating the tail portion of the cumulative distribution function (CDF) with a generalized pareto distribution (GPD). In structural engineering, reliability is measured by quantities like probability of failure or reliability index. Recently, alternate safety measures like the inverse measures in performance space have cornered enough interest because of their multifaceted advantages (Ramu et al, 2006). The inverse measures in performance space ties the concepts of safety factor and target failure probability. The inverse measures translate the deficiency in failure probability to deficiency in performance measure and hence provide a more quantitative measure of the resources needed to satisfy safety requirements. Among the several advantages they exhibit, inverse measures like probabilistic sufficiency factor (PSF) help in stable and accelerated convergence in optimization, better response surface approximations compared to surfaces fit to other reliability measures. Since inverse measure exhibit several advantages, its usefulness in tail modeling is explored in this work. Here we develop 3 alternate tail extrapolation schemes including inverse measures that can complement the classical tail modeling techniques. The alternate extrapolation schemes are based on variable transformations and approximating the relationship between the inverse measure and transformed variable. The primary motivation of this work is to develop an efficient reliability estimation procedure employing tail modeling techniques including inverse measures that can estimate reliability measures corresponding to high reliability with samples that are only sufficient to assess reliability measure corresponding to low reliability levels. This is a tradeoff between accuracy and computational power. Yet, as the studies reveal, the accuracy lost to the reduction of computational power is very reasonable. The reduction in computational cost is about a minimum of 3 orders of magnitude for the same level of accuracy. Goel et al., (2006) developed a method to extend the utility of an ensemble of surrogates. When faced with multiple surrogates, they explored the possibility of using the best surrogate or a weighted average surrogate model instead of individual surrogate models. In a similar fashion, in order to take advantage of both the classical tail modeling techniques and alternate extrapolation schemes and still come up with the best prediction, we propose to apply all the techniques simultaneously and use the median of the five estimates as the best estimate. Here, we call using all the techniques simultaneously as multiple tail models (MTM). In addition to arriving at the estimate, the range of the five techniques can be used to approximate the error associated with the median estimate. Moreover, the MTM approach can be used to replace the reliability analysis in reliability based design framework. CHAPTER 2 RELIABILITY BASED DESIGN OPTIMIZATION Introduction Optimization is the process of minimizing a cost which is a function of design variables (and other variables) subj ect to constraints that prescribe the design requirement. Often times, the variables involved are uncertain and probabilistic approaches are called for to account for the uncertainty as an alternate to the traditional safety factor approach, to obtain better designs. Several probabilistic approaches has been proposed in the last decade like Robust Design Optimization (Phadke, 1989, Gu, 2000), Reliabilitybased Design Optimization(Frangopal, 1985, Tu, 1999), Fuzzy optimization (Sakawa 1993, Maglaras et al, 1997), Reliability design using evidence theory (Soundappan et al., 2004). Reliabilitybased Design Optimization (RBDO) is widely used because it allows the designer to prescribe the level of reliability required. This section presents a review of the standard reliabilitybased design formulations and reliability estimation methods. Standard Formulations Primarily, reliabilitybased design consists of minimizing a cost function while satisfying reliability constraints. The reliability constraints are based on the failure probability corresponding to each failure mode or a single failure mode describing the system failure. The estimation of failure probability is usually performed by reliability analysis. In structural engineering, the system performance criterion is described by the limit state function which is typically expressed as the difference between the capacity of the system and the response of the system, which is expressed as: G(x) = G,(x) G(x) (21) where, G is the limit state function and Gc, Gr are the capacity and the response respectively. All the three quantities are functions of a vector x which consists of design (and random) variables and other parameters. The system is considered failed if G < 0 and safe if G > 0. If x is random, the design is considered a failure if the probability of failure is unacceptably high. Often times, RBDO is formulated as a double loop (nested loop) problem. The outer loop performs the optimization with respect to design variables, while the inner loop performs the reliability analysis estimating the failure probability. This is mathematically expressed as: min cost(x) subj ect to Pf = P(G" < 0) <; P~,, ,, j = 1, 2...nna(22 where the cost is a function of design variables, nm is the number of failure modes, G is the limit state function. Pyre,,wer is the prescribed failure probability levels to which the designer intends to design the structure. Here, estimation of probabilistic constraint requires reliability analysis. Often times the magnitude of failure probability is low and hence engineers prefer to work with reliability index as the alternate measure of reliability. The reliability index and failure probability are related as: Pf = CI(f ) (23) where, # is the standard normal cumulative density function (CDF). Since the inverse of the standard normal CDF is required to estimate the reliability index, it is an inverse measure. Probabilistic constraints in Eq.(22) can be written as: P" >Pt,,,t"(24) Employing Eq. (24) for prescribing the probabilistic constraint is called the Reliability Index Approach (RIA). In RIA, the failure reliability measure is usually calculated via FORM which is an iterative process and so computationally expensive and sometimes features convergence problems (Tu et al., 1999) In order to reduce the computational cost of double loop, various techniques has been proposed which can be classified into two categories as: (i) techniques that improve the efficiency of uncertainty analysis like fast probability integration (Wu 1994) and twopoint adaptive non linear approximations (Grandhi and Wang 1998) (ii) techniques that modify the formulation of the probabilistic constraints, for instance, using inverse reliability measures in the probabilistic constraint. Sometimes, researchers formulate a single loop RBDO that avoids nested loops. The idea of single loop formulations rests on the basis of formulating the probabilistic constraint as deterministic constraints by either approximating the Karush KuhnTucker conditions at the MPP or approximating the relationship between probabilistic design and deterministic design safety factors. A detailed survey of both single and double loop methods in RBDO is presented in the literature survey of Chapter 3. Kharmanda et al (2002, 2004, and 2007) developed RBDO solution procedures relative to two views points: reliability and optimization. From an optimization view point, Kharmanda et al (2002) developed a hybrid method based on simultaneous application of the reliability and the optimization problem that reduced the computational time. The hybrid method, compared to classical RBDO, minimizes a new form of the obj ective function which is expressed as the product of the original obj ective function and the reliability index in the hybrid space. Since the minimization of the obj ective function is carried out in both the deterministic variables and random variable space, it is referred to as hybrid design space. The reliability index in the hybrid space is the distance between the optimum point and the design point in the hybrid space However, the hybrid RBDO problem was more complex than that of the deterministic design and may not lead to local optima In order to address both the challenges, Kharmanda et al (2004) propose an optimum safety factor approach that computes safety factors satisfying a reliability level without demanding additional computing cost for the reliability evaluation. Kharmanda et al (2007) developed the hybrid design spaces and optimal safety factor equations for three distributions namely, the normal, lognormal and uniform. The optimal safety factors are computed by sensitivity analysis of the limit state function with respect to the random variables. They report that the optimal safety factor approach has several advantages like smaller number of optimization variables, good convergence stability, lower computing time and satisfaction of required reliability levels. In order to estimate the reliability measure for reliability analysis, one need to use analytical or simulation approaches which are described below. MomentBased Methods In standard reliability techniques, an arbitrary random vector x is mapped to an independent standard normal vector U (variables with normal distribution of zero mean and unit variance). This transformation is known as Rosenblatt transformation (Rosenblatt, 1952, Melchers, 1999, pp. 118120). The limit state function in the standard normal space can be obtained as G(U) = G(T(X)) where T is the transformation. If the limit state function in the standard normal space is affine, i.e., if G(U) = aTU + y then failure probability can be exactly calculated as P, = O I This is the basis for moment based methods. Momentbased methods provide for less expensive calculation of the probability of failure compared to simulation methods, although they are limited to a single failure mode. The First Order Reliability Method (FORM) is the most widely used momentbased technique. FORM is based on the idea of the linear approximation of the limit state function and is accurate as long as the curvature of the limit state function is not too high. In the standard normal space the point on the first order limit state function at which the distance from the origin is minimum is the MPP. When the limit state has a significant curvature, second order methods can be used. The Second Order Reliability Method (SORM) approximates the measure of reliability more accurately by considering the effect of the curvature of the limit state function (Melchers, 1999, pp 127130). Figure 21 illustrates the concept of the reliability index and MPP search for a two variable case in the standard normal space. In reliability analysis, concerns are first focused on the G(U) = 0 curve. Next, the minimum distance to the origin is sought. The corresponding point is the MPP because it contributes most to failure. This process can be mathematically expressed as: To find /7 = U1 Minimize IT (25) Subject to G(U)= 0 Where U* is the MPP. The calculation of the failure probability is based on linearizing the limit function at the MPP. It is to be noted that reliability index is the distance to the limit state function from the origin in the standard normal space. That is, it is an inverse measure in the input space (standard normal). Monte Carlo Simulation Variants of Monte Carlo methods were originally practiced under more generic names such as statistical sampling. The name and the method were made famous by early pioneers such as Stainslaw Ulam and John von Neumann and is a reference to the famous casino in Monaco. The methods use of randomness and repetitive nature is similar to the casino' s activities. It is discussed in Metropolis (1987) that the method earns its name from the fact that Stainslaw Ulam had an uncle who would borrow money from relative just to go to Monte Carlo. Monte Carlo integration is essentially numerical quadrature using random numbers. Monte Carlo integration methods are algorithms for the approximate evaluation of definite integrals, usually multidimensional ones. The usual algorithms evaluate the integrand at a regular grid. Monte Carlo methods, however, randomly choose the points at which the integrand is evaluated. Monte Carlo Methods are based on the analogy between volume and probability. Monte Carlo calculates the volume of a set by interpreting the volume as probability. Simply put, it means, 22 sampling from all the possible outcomes and considering the fraction of random draws that in a given set as an estimate of the set' s volume. The law of large numbers ensures that this estimate converges to the correct value as the number of draws increases. The central limit theorem provides information about the likely magnitude of the error in the estimate after finite number of draws. For example, consider estimating the integral of a function f over a unit interval. The a =f (x)dx (26) integral can be expressed as an expectation E;[ f(Ul)], with Uuniformly distributed between 0 and 1. The Monte Carlo estimate for the integral can be obtained as: & = f(U,)(27) If fis continuous and integrable over [0, 1], then, by the strong law of large numbers, the estimate in Eq. (27) converges to the actual value with probability 1 as n+ 00. In fact, if fis square integrable and we set "f = ( f(x) a2 )d (28) then the error & a in the Monte Carlo estmate is approximately normally distributed with zero mean and standard deviation of of /~ the quality of this approximation improving as n increases. The parameter of in Eq.(28) will not be known in a real setting but can be approximated by the sample standard deviation. Thus, we not only obtain the estimate but also the error contained in the estimate. The form of the standard error is the main feature of the MCS. Reducing this error to half require about increasing the number of points by four; increasing the precision by one decimal point requires 100 times as many points. Yet, the advantage of MCS lies in the fact that its O(n /) convergence rate is not restricted to integrals over the unit interval and can be extended to any dimension. The standard error will still have the same form irrespective of dimension. This advantage is unlike other numerical methods. Thus, Monte Carlo simulation is attractive to solve integrals in higher dimension. In structural reliability estimation, MCS involves sampling each random variable to give a sample value. Then, the limit state function is computed at every realization. If the limit state is violated, the structure has failed. The experiment is repeated many times with a randomly chosen vector. If Ntrials are conducted, the estimate Pf of the probability of failure is given approximately as: n(G(x,)~ < ) Pf = N (29) Where n(G(x) <; 0) denotes the number of trials n for which the structure failed. The number of trails Nis related to the desired accuracy of failure probability. MCS need at least 100*N samples when the failure probability is of the order of l over N. One can find the standard deviation of failure probability in Eq. (29) using the following expression: Pf (1 Pf ) 0= (210) There are many variants to the abovementioned crude MCS. Researchers use several variance reduction techniques like importance sampling (Glasserman, 2004) to make better use of MCS and extract more information from the same level of simulation. Importance sampling requires one to find a sampling PDF which uses the information of greatest probability density in the failure zone. This is conceptually similar to MPP which is essentially a point that contributes most to failure. An alternate formulation of the MCS is used by Smarslok et al (2006) to exploit the advantage of separable Monte Carlo simulations. This integration method takes advantage of a special form of the limit state function in which the limit state is composed of difference between capacity and response. ff2 (GU)=0(linear) Failure Region G(U)=0.2 G(U)<0 Figure 21. Reliability analysis and MPP. Sub script refers to iteration number .CHAPTER 3 INVERSE RELIABILITY IVEASURES In probabilistic approaches the difference between the computed probability of failure or reliability index and their target values does not provide the designer with easy estimates of the change in the design cost needed to achieve these target values. Alternately, inverse reliability measures are capable of providing this information. Several inverse reliability measures (e.g., probabilistic performance measure and probabilistic sufficiency factor) that are essentially equivalent have been introduced in recent years. The different names for essentially the same measure reflect the fact that different researchers focused on different advantages of inverse measures. This chapter reviews the various inverse measures and describes their advantages compared to the direct measures of safety such as probability of failure and reliability index. Methods to compute the inverse measures are also described. Reliabilitybased design optimization with inverse measure is demonstrated with a beam design example. Literature Review In deterministic approaches, a safety factor is defined as the ratio of the capacity to resistance of a system, typically calculated at the mean values. However, in probabilistic approaches, as both capacity and response are random, the safety factor is also a random number. One safety measure that combines the advantages of the safety factor and probability of failure was proposed by Birger (1970, as reported by Elishakoff, 2001). Birger relates the safety factor and the fractile of the safety factor distribution corresponding to a target probability of failure. It belongs to a class of inverse reliability measures, which carry that name because they require use of the inverse of the cumulative distribution function. Several researchers developed equivalent inverse reliability methods (Lee and Kwak 1987; Kiureghian et al. 1994; Li and Foschi 1998; Tu et al. 1999; Lee et al. 2002; Qu and Haftka 2003; Du et al. 2003) that are closely related to the Birger measure. These measures quantify the level of safety in terms of the change in structural response needed to meet the target probability of failure. Lee and Kwak (1987) used the inverse formulation in RBDO and showed that it is preferable for design when the probability of failure is very low in some region of the design space so that the safety index approaches infinity. Kiureghian et al. (1994) addressed the inverse reliability problem of seeking to determine one unknown parameter in the limit state function such that a prescribed first order reliability index is attained. To solve the inverse reliability problem, they proposed an iterative algorithm based on the HasoferLindRackwitzFiessler algorithm. Li and Foschi (1998) employed the inverse reliability strategy in earthquake and offshore applications to solve for multiple design parameters. They show that it is an efficient method to estimate design parameters corresponding to target reliabilities. KirjnerNeto et al. (1998) reformulated the standard RBDO formulation similar to Lee and Kwak (1987) except that they use an inequality constraint. They developed a semiinfinite optimization algorithm to solve the reformulated problem. This formulation does not require second order derivatives of the limit state functions and obviates the need for repeated reliability indices computation. However, they found that the approach can result in conservative design. Royset et al. (2001) extended the reformulation technique discussed in KirjnerNeto et al. (1998) for reliabilitybased design of series structural systems. The required reliability and optimization calculations are completely decoupled in this approach. Tu et al. (1999) dubbed the inverse measure approach the Performance Measure Approach (PMA) and called the inverse measure probability performance measure (PPM). Lee et al. (2002) adopted the same procedure as Tu et al. (1999) and named it target performance based approach, calling the inverse measure the target performance. They compared the reliability index approach and inverse measure based approach and found that the latter was superior in both computational efficiency and numerical stability. Youn et al. (2003) showed that PMA allows faster and more stable RBDO compared to the traditional Reliability Index Approach (RIA). When accurate statistical information is not available for input data, it is not appropriate to use probabilistic method for stochastic structural analysis and design optimization and researchers resort to possibility based design optimization (PBDO) methods in such situations. Mourelatos and Zhou, 2005, discuss a PMA based PBDO. They formulate the inverse possibility analysis problem for the possibilistic constraint. Unlike FORM based RBDO which is based on linear approximation, the PBDO is exact. To perform the inverse possibility analysis, Du et al (2006b) proposes a maximal possibility search (MPS) method with interpolation to address the challenges of accuracy and computational expense exhibited by traditional methods like the multilevel cut and alpha level optimization method respectively. They report that the MPS evaluates possibility constraint efficiently and accurately for nonlinear structural applications. Often times industry design problems deal with uncertainties with sufficient data and uncertainties with insufficient data simultaneously. To address such situations Du et al., (2006a) extend the possibilitybased design optimization (PBDO) method. They propose a two step approach to generate membership function from the available data. Initially, they develop a temporary PDF using the available data and then generate a membership function from the temporary PDF. They report that PMA based PBDO combined with MPS method can adequately address design problems with mixed input variables and demonstrates that it evaluates possibility constraint efficiently, stably and accurately for nonlinear structural applications. Qu and Haftka (2003, 2004) called the inverse measure probability sufficiency factor (PSF) and explored its use for RBDO with multiple failure modes through Monte Carlo Simulation (MCS) and response surface approximation (RSA). They showed that PSF leads to more accurate RSA compared to RSA fitted to failure probability, provides more effective RBDO, and that it permits estimation of the necessary change in the design cost needed to meet the target reliability. Moreover, PSF enables performing RBDO in variablefidelity fashion and sequential deterministic optimization fashion to reduce the computational cost (Qu and Haftka, 2004a and 2004b). An initial study of using PSF to convert RBDO to sequential deterministic optimization was performed to solve problems with reliability constraints on individual failure modes (Qu and Haftka, 2003). An improved version for system reliability problems with multiple failure modes was developed for reliabilitybased global optimization of stiffened panels (Qu and Haftka, 2004b). Du et al. (2003, 2004) employed PMA to formulate RBDO, but used percentile levels of reliability (1 minus failure probability) in the probabilistic constraint and called the inverse measure as percentile performance. Traditionally, design for robustness involves minimizing the mean and standard deviation of the performance. Here, Du et al. (2003) proposed to replace the standard deviation by percentile performance difference, which is the difference between the percentile performance corresponding to the left tail of a CDF and the right tail of that CDF. They demonstrated increased computational efficiency and more accurate evaluation of the variation of the obj ective performance. In an effort to address reliabilitybased designs when both random and interval variables are present, Du and Sudijianto (2003) proposed the use of percentile performance with worstcase combination of the interval variables for efficient RBDO solutions. Du and Chen (2004) developed the sequential optimization and reliability assessment (SORA) to improve the efficiency of the probabilistic optimization. The method is a serial single loop strategy, which employs percentile performance and the key is to establish equivalent deterministic constraints from probabilistic constraints. This method is based on evaluating the constraint at the most probable point of the inverse measure in Section IV below) based on the reliability information from the previous cycle. This is referred to as "design shift" (Chiralaksanakul and Mahadevan 2004; Youn et al. 2004). They show that the design quickly improves in each cycle and is computationally efficient. The sequential optimization and reliability assessment, however, is not guaranteed to lead to an optimal design. Singlelevel (or unilevel) techniques that are equivalent to the standard RBDO formulation are based on replacing the RIA or PMA inner loop by the corresponding KarushKuhnTucker conditions. Here again, Agarwal et al. (2004) showed that the PMA approach is more efficient than the unilevel RIA approach due to Kuschel and Rackwitz (2000). Since most commercial optimizers are numerically unreliable when applied to problems accompanied by many equality constraints, Agarwal et al (2007) use homotopy methods for constraint relaxation and to obtain a relaxed feasible design. They solve a series of optimization problem as the relaxed optimization problem is transformed via a homotopy to the original problem. They show that it is easier to solve the relaxed problem and make a gradual progress towards the solution than solve the original problem directly. The several inverse measures discussed above are all based on the common idea of using the inverse of the cumulative distribution function. The numerous names for the inverse measures contemplate that they were developed by different researchers for different applications. Since these inverse measures come under various names, it is easy to fail to notice the commonality among them. Birger Safety Factor The safety factor, S is defined as the ratio of the capacity of the system Gc (e.g., allowable strength) to the response Grwith a safe design satisfying Gr < Gc To account for uncertainties, the design safety factor is greater than one. For example, a load safety factor of 1.5 is mandated by FAA in aircraft applications. To address the probabilistic interpretation of the safety factor, Birger (1970) proposed to consider its cumulative distribution function (CDF) Fs : G~ F, (s) = Prob( c s) (31) Note that unlike the deterministic safety factor, which is normally calculated for the mean value of the random variables, Gc/Gr in Eq.(31) is a random function. Given a target probability,Pftarget, Birger suggested a safety factor s*(which we call here the Birger safety factor) defined in the following equation F,(s*) = Prob( ` That is, the Birger safety factor is found by setting the value of cumulative distribution function (CDF) of the safety factor equal to the target probability. That is, we seek to find the value of the safety factor that makes the CDF of the safety factor equal to the target failure probability. This requires the inverse transformation of the CDF, hence the terminology of mnverse measure. Probabilistic Sufficiency Factor Qu and Haftka (2003, 2004) developed a similar measure to the Birger safety factor, calling it first the probabilistic safety factor and then the probabilistic sufficiency factor (PSF). They obtained the PSF by Monte Carlo simulation and found that the response surface for PSF was more accurate than the response surface fitted to failure probability. Later, they found the reference to Birger' s work in Elishakoff' s review (2001) of safety factors and their relations to probabilities. It is desirable to avoid the term safety factor for this entity because the common use of the term is mostly deterministic and independent of the target safety level. Therefore, while noting the identity of the Birger safety factor and the probabilistic sufficiency factor, we will use the latter term in the following. Failure happens when the actual safety factor S is less than one. The basic design condition that the probability of failure should be smaller than the target probability for a safe design may then be written as: P, = Prob(S < 1)= F, (1) < Pftarget (33) Using inverse transformation, Eq.(33) can be expressed as: 1 < F/ (Pftarget )= s* (3 4) The concept of PSF is illustrated in Figure. 31. The design requirement Pftarge is known and the corresponding area under the probability density function of the safety factor is the shaded region in Figure.31.The upper bound of the abscissa s* is the value of the PSF. The region to the left of the vertical line S = 1 represents failure. To satisfy the basic design condition s* should be greater than or equal to one. In order to achieve this, it is possible to either increase Gc or decreaseGr The PSF s*, represents the factor that has to multiply the response Gr or divide the capacity Gc, so that the safety factor be raised to 1. For example, a PSF of 0.8 means that Gr has to be multiplied by 0.8 or Gc be divided by 0.8 so that the safety factor ratio increases to one. In other words, this means that Gr has to be decreased by 20 % (10.8) or Gc has to be increased by 25% ((1/0.8)1) in order to achieve the target failure probability. It can be observed that PSF is a safety factor with respect to the target failure probability and is automatically normalized in the course of its formulation. PSF is useful in estimating the resources needed to achieve the required target probability of failure. For example, in a stressdominated linear problem, if the target probability of failure is 10' and a current design yields a probability of failure of 103, One cannot easily estimate the change in the weight required to achieve the target failure probability. Instead, if the failure probability corresponds to a PSF of 0.8, this indicates that maximum stresses must be lowered by 20% to meet the target. This permits the designers to readily estimate the weight required to reduce stresses to a given level. Probabilistic Performance Measure In probabilistic approaches, instead of the safety factor it is customary to use a performance function or a limit state function to define the failure (or success) of a system. For example, the limit state function is expressed as in Eq(21). In terms of safety factor S, another form of the limit state function is: Gc (X) G(X) = 1= S12 > (35) Gr (X) Here, G (X) and G(X) are the ordinary and normalized limit state functions, respectively. Failure happens when G (X) or G(X) is less than zero, so the probability of failure Pf is: P, = Prob(G (X) < 0) (3 6) P, = Prob(G(X) < 0) (3 7) Using Eq. (36), Eq. (37) can be rewritten as: P, = Prob(G (X) < 0) = FG (0) < Pftaerge (38) P, = Prob(G(X) <;0) = FO (0) < Pftarge (3 9) where FG and FO are the CDF of G (X) and G(X), respectively. Inverse transformation allows us to write Eqs. (38) and (39) as: 0 < FGl(Pftage ) = g* (310) 0 < Fii (Pftage ) = g (311) Here, g* and g are the ordinary and normalized Probabilistic Performance Measure (PPM, Tu et al, 1999), respectively. PPM can be defined as the solution to Eqs. (38) and (39). That is, the value in (*) (instead of zero) which makes the inequality an equality. Hence PPM is the value of the limit state function that makes its CDF equals the target failure probability. Figure 32 illustrates the concept of PPM. The shaded area corresponds to the target failure probability. The area to the left of the line G = 0 indicates failure. g* is the factor that has to be subtracted from Eq. (36) in order to make the vertical line g* move further to the right of the G = 0 line and hence facilitate a safe design. For example, a PPM of 0.8 means that the design is not safe enough, and 0.8 has to be subtracted from G (X) in order to achieve the target probability of failure. A PPM value of 0.3 means that we have a safety margin of 0.3 to reduce while improving the cost function to meet the target failure probability. Considering g as the solution for Eq.(31 1), it can be rewritten in terms of the safety factor as: Prob(G(X)= S1< g*)= Pftarget (312) Comparing Eqs. (15), (23) and (24), we can observe a relationship between s* andg*. PSF (s*) is related to the normalized PPM (g *) as: s*= g*+1 (313) This simple relationship between PPM and PSF shows that they are closely related, and the difference is only in the way the limit state function is written. If the limit state function is expressed as the difference between capacity and response as in Eq.(36), failure probability formulation allows us to define PPM. Alternatively, if the limit state function is expressed in terms of the safety factor as in Eq. (37), the corresponding failure probability formulation allows us to define PSF. PSF can be viewed as PPM derived from the normalized form of Eq. (36). The performancemeasure approach (PMA) notation may appeal because of its generality, while the PSF notation has the advantage of being automatically scaled and being expressed in terms that are familiar to designers who use safety factors. The PPM can be viewed as the distance from the G = 0 to target failure probability line, in the performance function space. This is analogical to reliability index being a measure of distance in the input variable space. The maj or difference is the measurement of distance in different spaces, the performance function (or output) space and the input space. Inverse Measure Calculation Simulation Approach Monte Carlo Simulation Conceptually, the simplest approach to evaluate PSF or PPM is by Monte Carlo simulation (MCS), which involves the generation of random sample points according to the statistical distribution of the variables. The sample points that violate the safety criteria in Eq. (36) are considered failed. Figure.33 illustrates the concept of MCS. A twovariable problem with a linear limit state function is considered. The straight lines are the contour lines of the limit state function and sample points generated by MCS are represented by small circles, with the numbered circles representing failed samples. The zero value of the limit state function divides the distribution space into a safe region and a failure region. The dashed lines represent failed conditions and the continuous lines represent safe conditions. The failure probability can be estimated using Eq.(29) In Figure. 33, the number of sample points that lie in the failure region above the G = 0 curve is 12. If the total number of samples is 100,000, the failure probability is estimated at 1.2x104. For a fixed number of samples, the accuracy of MCS deteriorates with the decrease in failure probability. For example, with only 12 failure points out of the 100,000 samples, the standard deviation of the probability estimate is 3.5x10 more than a quarter of the estimate. When the probability of failure is significantly smaller than one over the number of sample points, its calculated value by MCS is likely to be zero. PPM is estimated by MC S as the net'' smallest limit state function among the N sampled functions, where n = Nx Pftarget For example, considering the example illustrated in Figure. 33, if the target failure probability is 104, to satisfy the target probability of failure, no more than 10 samples out of the 100,000 should fail. So, the focus is on the two extra samples that failed. PPM is equal to the value of the highest limit state function among the n (in this case, n = 10) lowest limit state functions. The numbered small circles are the sample points that failed. Of these, the three highest limit states are shown by the dashed lines. The tenth smallest limit state corresponds to the sample numbered 8 and has a limit state value of 0.4, which is the value of PPM. Mathematically this is expressed as: g* = nth minz(G (Xi)) (314) i= 1 where, net'' min is the n"i' smallest limit state function. So, the calculation of PPM in MCS requires only sorting the lowest limit state function in the MCS sample. Similarly, PSF can be computed as the nth smallest factor among the N sampled safety factors and is mathematically expressed as * s* =nri'' min (S(Xi)) (315) i=1 ** A more accurate estimate of PPM or PSF will be obtained from the average of the nt and (n+1)t smallest values. So in the case of Figure. 4, PPM is more accurately estimated as 0.35. Finally, probabilities calculated through Monte Carlo Simulation (MCS) using a small sample size are computed as zero in the region where the probability of failure is lower than one over the number of samples used in MCS. In that region, no useful gradient information is available to the optimization routine. On the other hand, PSF value varies in this region and thus provides guidance to optimization. MCS generates numerical noise due to limited sample size. Noise in failure probability may cause RBDO to converge to a spurious minimum. In order to filter out the noise, response surface approximations are fitted to failure probability to create a socalled design response surface. It is difficult to construct a highly accurate design response surface because of the huge variation and uneven accuracy of failure probability. To overcome these difficulties, Qu and Haftka (2003, 2004) discuss the usage of PSF to improve the accuracy of the design response surface. They showed that design response surface based on PSF is more accurate compared to design response surface based on failure probability and this accelerates the convergence of RBDO. For complex problems, response surface approximations (RSA) can also be used to approximate the structural response in order to reduce the computational cost. Qu and Haftka (2004b) employ PSF with MCS based on RSA to design stiffened panels under system reliability constramnt. Analytical Approach Momentbased Methods Inverse reliability measures can be estimated using moment based methods like FORM. The FORM estimate is good if the limit state is linear but when the limit state has a significant curvature, second order methods can be used. The Second Order Reliability Method (SORM) approximates the measure of reliability more accurately by considering the effect of the curvature of the limit state function (Melchers, 1999, pp 127130). Figure 34. illustrates the concept of inverse reliability analysis and MPP search. The circles represent the p curves with the target p curve represented by a dashed circle. Here, among the different values of limit state functions that pass through the Prose, curve, the one with minimum value is sought. The value of this minimal limit state function is the MPP as shown by Tu et al. 1999. The point on the target circle with the minimal limit state function is sought. This point is also an MPP and in order to avoid confusion between the usual MPP and MPP in inverse reliability analysis, Du et al. (2003) coined the term most probable point of inverse reliability (MPPIR) and Lee et al. (2002) called it the minimum performance target point (MPTP). Du et al. (2003) developed the sequential optimization and reliability analysis method in which they show that evaluating the probabilistic constraint at the design point is equivalent to evaluating the deterministic constraint at the most probable point of the inverse reliability. This facilitates in converting the probabilistic constraint to an equivalent deterministic constraint. That is, the deterministic optimization is performed using a constraint limit which is determined based on the inverse MPP obtained in the previous iteration. Kiureghian et al. (1994) proposed an extension of the HasoferLindRackwitzFiessler algorithm that uses a merit function and search direction to find the MPTP. In Figure 34, the value of the minimal limit state function or the PPM is 0.2. This process can be expressed as: Minimize :G(U) Subject to : U = J = Freze (316 In reliability analysis the MPP is on the G (U) = 0 failure surface. In inverse reliability analysis, the MPP search is on the Ptarse curve. One of the main advantages when inverse measures are used in the FORM perspective is that, the formulation of the optimization problem is as expressed in Eq. (316) where the constraint has a simple form (circle) compared to the cost function. The cost function has in general a complicated expression due to the nonlinear transformation between physical random variable X to the standard random variable U. It is well known that it is easy to satisfy a simple constraint compared to a complicated constraint irrespective of the cost function. On comparing the formulation in Eq.(25) with the formulation in Eq.(316), it is evident that it is easier to solve the latter because of the simpler constraint. Reliability Based Design Optimization with Inverse Measures The probabilistic constraint in RBDO can be prescribed by several methods like the Reliability Index Approach (RIA), the Performance Measure Approach (PMA), the Probability Sufficiency Factor approach (PSF), see Table31. Table 31. Different approaches to prescribe the probabilistic constraint Method RIA PMA PSF Probabilistic constraint 13 > twer g* > 0 s* > 1 Quantity to be computed Reliability index( P) PPM ( g*) PSF (s*) In RIA, P can be computed as the product of reliability analysis as discussed in the previous section. The PPM or PSF can be computed through inverse reliability analysis or as a byproduct of reliability analysis using MCS. To date, most researchers have used RIA to prescribe the probabilistic constraint. However, the advantages of the inverse measures, illustrated in the next section, have led to its growing popularity. Beam Design Example The cantilever beam shown in Figure. 35, taken from Wu et al. (2001), is a commonly used demonstration example for RBDO methods. The length L of the beam is 100". The width and thickness is represented by w and t. It is subj ected to end transverse loads X and Y in orthogonal directions as shown in the Figure 35 1. The obj ective of the design is to minimize the i For this example, the random variables are shown in bold face weight or equivalently the cross sectional area: A =wt subj ect to two reliability constraints, which require the safety indices for strength and deflection constraints to be larger than three. The first two failure modes are expressed as two limit state functions: Stress limit G = R a=l( R +, (317) 4L 3 iY X (8 Tip displacement limit: Gd = Do D = Do + 38 Ewt t' w where R is the yield strength, E is the elastic modulus, Do the displacement limit, and w and t are the design parameters. R, X, Y and E are uncorrelated random variables and their means and standard deviations are defined in Table 32. Table 32. Random variables for beam problem Random X Y R E Variables (lb) (lb) (psi) (psi) Di stributi on Normal Normal Normal Normal (p, a)(500,100) (1000,100)1b (40000,2000) (29x 106,1.45x106) Design for Stress Constraint The design with strength reliability constraint is solved first, followed by the design with a system reliability constraint. The results for the strength constraint are presented in Table 33. For the strength constraint, the limit state is a linear combination of normally distributed random variables, and FORM gives accurate results for this case. The MCS is performed with 100,000 samples. The standard deviation in the estimated failure probability is calculated by MCS as: Pf (1 Pf ) Op (319) In this case, the failure probability of 0.0013 calculated from 100,000 samples has a standard deviation of 1.14x104. It is seen from Table 33 that the designs obtained from RIA, PMA and PSF match well. Since the stress in Eq. (317)is a linear function of random variables, the RIA and PMA are exact. The more conservative design from PSF is due to limited sampling of MCS. Table 33. Comparison of optimum designs for the stress constraint Minimize obj ective function A =wt such that p 2 3 Method Reliability Inverse Reliability Analysis Exact Analysis FORM M~CS (PSF) Optimnum FORM (PMA) (Qu and Haftka,2003) (Wu et al. (RIA) 2001) Optima w 2.4460 2.4460 2.4526 2.4484 t 3.8922 3.8920 3.8884 3.8884 Obj ective Function 9.5202 9.5202 9.5367 9.5204 Reliability Index 3.00 3.00 3.0162 3.00 Failure Probability 0.00135 0.00135 0.00128 0.00135 Comparison of Inverse Measures The relation between PSF and PPM in Eq. (313) is only approximate when PPM is calculated by FORM and PSF by MCS. PSF suffers from sampling error and PPM from error due to linearization. For the linear stress constraint, and with a large Monte Carlo sample, the difference is small, as seen in Table 34. It may be expected that the MPTP (minimum performance target point) should also be close to the point used to calculate PSF. This result may be useful, because when a response surface is used for an approximation of the response, it is useful to center it near the MPTP. In order to check the accuracy of the MPTP estimation, the MPP of PPM and the point corresponding to the PSF are compared and the results are tabulated in Table 35. The coordinates corresponding to the PSF, computed by a millionsample MCS, deviate considerably from the MPTP. Since the accuracy of the points computed by MCS depends on the number of samples used, an increased number of samples lead to more accurate results, albeit at increased computational cost. Alternative approaches to obtain the points with better accuracy without increasing the number of samples are to average the coordinates computed by repeated MCS simulations with fewer numbers of samples. Alternatively, we can average a number of points that are nearly as critical as the PSF point. That is, instead of using only the X1 corresponding to the S, in Eq (315), we use also the points corresponding to S S,, S ,in computing the PSF, where 2p is the total number of points that is averaged around the PSF. It can be observed from Table 35 that averaging 11 points around the PSF matches well with the MPTP, reducing a Euclidean distance of about 0.831 for the raw PSF to 0.277 with 11point average. The values of X, Y and R presented in Table 3 5 are in the standard normal space Table 34. Comparison of inverse measures for w=2.4526 and t3.8884 (stress constraint) Method FORM MC S (1 x107samples) Pc 0.001238 0.001241 Inverse Measure PPM: 0.00258 PSF: 1.002619 Table 35. Comparison of MPP obtained from calculation of PSF for w=2.4526 and t3.8884 (stress constraint) Coordinates X Y R MPTP 2.1147 1.3370 1.6480 PSF 1)106 Samples 2.6641 1.2146 1.0355 2)Average:10 runs of each 105samples 2.1888 1.8867 1.1097 3)Average:11 points around PSF of 106 Samples 2.0666 1.5734 1.5128 Use of PSF in Estimating the Required Change in Weight to Achieve a Safe Design The relation between the stresses, displacement and weight for this problem is presented to demonstrate the utility of PSF in estimating the required resources to achieve a safe design. Consider a design with the dimensions of Ao= wo to with a PSF of so*" less than one. The structure can be made safer by scaling both w and t by the same factor c. This will change the stress and displacement expressed in Eq.(317) and Eq. (318) by a factor of c3 and c4, respectively, and the area by a factor of c2. If Stress is most critical, it will scale as c3 and the PSF will vary with the area, A as: 1.5 s* = so a (320) Equation (320) indicates that a one percent increase in area will increase the PSF by 1.5 percent. Since nonuniform scaling of width and thickness may be more efficient than uniform scaling, this is a conservative estimate. Thus, for example, considering a design with a PSF of 0.97, the safety factor deficiency is 3% and the structure can be made safer with a weight increase of less than two percent, as shown by Qu and Haftka (2003). For a critical displacement state, s* will be proportional to A2 and a 3% deficit in PSF can be corrected in under 1.5% weight increase. While for more complex structures we do not have analytical expressions for the dependence of the displacements or the stresses on the design variables, designers can usually estimate the weight needed to reduce stresses or displacements by a given amount. System Reliability Estimation Using PSF and MCS System reliability arises when the failure of a system is defined by multiple failure modes. Here, the discussion is limited to series systems. The estimation of system failure probability involves estimating the failure due to individual modes and failure due to the interaction between the modes. This is mathematically expressed for a system with n failure modes as: Psv = P(F, ) +P(F, )+...+ P(F ,) P(F, nF,) P(F; nr,,) ... P(F,_ z Fn ) (3 21) +P(; nr, nr, ,)+...+ P(F,_, Fng,_ Fn)...Higher order terms It is easy to employ MCS to estimate system failure, as the methodology is a direct extension of failure estimation for single modes. In the context of employing analytical methods like FORM, estimation of failure regions bounded by single modes are easy to estimate but estimating the probability content of odd shaped domains created because of the intersection of two or more modes is challenging. Techniques are available to estimate the probability content at the intersection regions but require evaluating multiple integrals (Melchers, pp: 386400). Since the number of multiple integrals that has to be evaluated depends on the number of variables, the task becomes arduous when more variables are involved. Instead, an alternate technique is to develop upper and lower bounds on the failure probability of the structural system. Considering the terms in the first and the third rows (positive contributions) in Eq. (321) permits us to estimate the upper bound. When the terms in first and second rows (negative contributions) are considered, the lower bounds can be estimated. Improved bounds for the system reliability are available in the literature. Owing to these challenges, MCS is mostly used when system reliability is to be estimated. As discussed earlier, MCS is computationally prohibitive when the failure probability to be estimated is low, because of the number of samples required. PSF can be used to estimate system reliability. The method is an extension of the PSF estimation for single failure mode. When M\~ failure modes are considered, the most critical safety factor is calculated for each sample. Then the sorting of the n*h minimum safety factor can proceed as in Eq.(315). This process is explained in the flow chart in Figure 36. The estimation of PSF for a system reliability case is demonstrated with an example in Appendix A Design for System Reliability by MCS and PSF Monte Carlo simulation is a good method to use for system reliability analysis with multiple failure modes. For the cantilever beam discussed in the earlier section, the allowable deflection was chosen to be 2.25" in order to have competing constraints (Wu et al., 2001). The results are presented in Table 36. It can be observed that the contribution of the stress mode to failure probability dominates the contribution due to displacement and the interaction between the modes. The details of the design process are provided in Qu and Haftka (2004). They demonstrated the advantages of using PSF as an inverse safety measure over probability of failure or the safety index as a normal safety measure. They showed that the design response 44 surface (DRS) of the PSF was much more accurate than the DRS for the probability of failure. For a set of test points the error in the probability of failure was 39. 11% from the DRS to the PSF, 96.49% for the DRS to the safety index, and 334.78% for the DRS to the probability of failure . Table 36. Design for System Reliability by PSF, Qu and Haftka (2004)** Optima Obj ective Pf Safety Ppl Pz Ppl n Pf PSF Function Index s* w = 2.6041 9.5691 0.001289 3.01379 0.001133 0.000208 0.000052 1.0006 t= 3.6746 **100,000 samples, qf~Pf, Pf, On Pfz Failure probabilities due to stress constraint, displacement constraint, and intersection between the modes, respectively. PDF of 3 Figure 31. Schematic probability density of the safety factor S. The PSF is the value of the safety factor corresponding to the target probability of failure PDF of G p~gc~ g~ G (x) =G, (x) Gr i) Figure 32. Schematic probability density of the limit state function. The Probabilistic Performance Measure is the value of the limit state function corresponding to the target probability of failure. G(X) < Failure region O O Figure 33. Illustration of the calculation of PPM with Monte Carlo Simulation for the linear performance function Figure 34. Inverse reliability analysis and MPP for target failure probability 0.00135 (/7 Figure 35. Cantilever beam subj ected to horizontal and vertical loads Y' Figure 36. Flowchart for estimating system PSF CHAPTER 4 TAIL MODELLING AND RELIABILITY ANALYSIS Low failure probability problems (extreme value) require one to have sufficient data in the tails of the distribution which represent the extremes. But this is seldom possible and instead researchers use tail modeling techniques based on extreme value theory to predict the probability of extreme events. The theory comprises a principle for model extrapolation based on the implementation of mathematical limits as finite level approximations. Since several advantages are reported by working with inverse measures, it is logical to justify an attempt to perform tail modeling in the performance space along with inverse measures to estimate quantities at unobserved levels. This section discusses the tail modeling technique and how to apply it to find inverse measures. Tail Equivalence and Generalized Extreme Value Theory Tail modeling is based on the concept of approximating the tail distribution of the variable of interest by associating it with the tail of already known distribution and using the parameters and characteristics therein for further analysis. This criterion is adequately defined by tail equivalence. That is, two CDFs F(x) and Q(x) can be called tail equivalent (Maes and Breitung, 1993) at the right tails when: .1 F(x) him = 1 (41) 1Q(x) This reveals that the error in approximating the small probability of 1 F(x) by 1 Q(x) tends to zero as x 4 00 The extremal theorem by Fisher and Tippet (1928) states that, if2I,, is the maximum of a sequence of 'n' independent random variables having a common distribution function and if there exists sequences of constants (a,, > 0) and {b,,) such that Pr (M, ,, V/' a, U 4Hn)a (42) for a nondegenerate distribution function H, then H belongs to one of the extreme value distribution families namely, the Gumbel, Frechet and Weibull families. It is to be noted that regardless of the distribution for the population, the three types of the extreme value distributions are the only possible limits for the distribution of the normalized maxima. In this respect, this theorem is to extreme value theory what central limit theorem is to central statistics. The central limit theorem states that the approximating distribution of the sample mean is normal regardless of the parent distribution. The three families of distributions can be combined into a single family of models having distribution function of the form H(z) = exp 1+(r (43) such that (zl :+ 5(z pu) / a > 0), where the parameters satisfy co < pu < c, a > 0 and co < 5 < CO This is the generalized extreme value (GEV) family of distributions. Here, pu is the location parameter, a is the scale parameter, 5 is the shape parameter. A more detailed discussion of the generalized extreme value distributions is presented in Appendix B Generalized Pareto Distribution In engineering applications, rather than maxima, the interest is to address the excesses over threshold. In these situations, the generalized pareto distribution (GPD) arises as the limiting distribution. The concept of GPD is presented in Figure 41. Let y be a model output which is random and u be a large threshold of y. The observations of y that exceed u are called exceedance. The conditional distribution Fu(z) of the exceedance given that the data y is greater than the threshold u, is modeled fairly well by the GPD. Here, z = y u Let approximation of Fi (z) using GPD be denoted by F,,(z). 5 and ry are the shape and scale parameters respectively. For a large enough u, the distribution function of (y u) conditional on y > u, is approximately written as (Coles, 2001): 1 1+ if 5 +0 i(z)= = (44) 1 exp if 5=0 In Eql (44), (Aj+ = max(0, A) and z > 0.5 plays a key role in assessing the weight of the tail. Eq (44) can be justified as a limiting distribution as u increases (Coles, 2001, pp:7576). Tails can be classified based on 5 as: 5 > 0, heavy tail (pareto type tail) 5 = 0, medium tail (exponential type tail) 5 <0, light tail (Betatype tails) There are several parameter estimation methods to estimate the parameters 5 and ry .It is to be noted that conditional excess CDF F (z) is related to the CDF of interest Fly) through the following expression: Fly) Flu) ,(z) = (45) 1 Flu) From Eq (45), the CDF of y can be expressed as: Fly) = (1 F(u))Fu(z) +Flu) (46) When Eq (44) is substituted for F (z) in the above expression, Eq (46) becomes: F(y)=1 O (1 F(u) ( u (47) For simplicity of presentation, only the case of5 + 0 is considered here. Once we obtain estimates of the parameters as 5 and ry using some parameter estimation method like maximum likelihood estimation method, least square regression that are discussed later in the chapter, it is possible to estimate the pth quantile by inverting Eq.(47) : yp = F (p)= u+ 1 (48)p 1 l F(u)j 48 In order to extend the concept to structural applications, if the output y is replaced by the limit state in Eq (35) then the interest is to model the lower tail rather than the upper tail. However, modeling the upper tail of y is equivalent to modeling the lower tail of y. Therefore, an equivalent way of expressing Eq.(35) is: G, (x)= 1 S,(x) (49) where S, is the reciprocal of the safety factor. Failure occurs when G1 > 0. S, at any required target reliability index can be estimated using Eq(48). If Pf target refers to the target failure probability that we wish to design the structure, PSF can be obtained as: S = F 1P ,,)=* u+ Pf target 1(410) r f taget ( 1 Flu) Coles (2001) opines that it is easy to obj ect to the procedure of tail modeling on the grounds that even with the support of asymptotic argument, there is an implicit assumption that the underlying stochastic mechanism of the process being modeled is sufficiently smooth to enable extrapolation. However no credible alternative has been proposed to date. Threshold Selection The theory of excess over threshold is referred to as peaks over threshold method also. Performance of this approach is based on the choice of the threshold value u. The basic theory identified by Pickands (1975) states that the pareto distribution will be a practical family for statistical estimation of the tails, provided that the threshold is taken sufficiently high. The CDF is composed of 3 portions, the lower tail, upper tail and the central portion. In theory, the threshold should be selected where the actual upper tail starts. But there is no straightforward globally accepted method to select threshold. Selection of threshold is a tradeoff between bias and variance. If the threshold selected is low, then some data points belong to the central part of the distribution and do not provide a good approximation to the tails. On the other hand, if the threshold selected is too high, the number of data available for the tail approximation is much less and this might lead to excessive scatter in the Einal estimate. The proper selection of threshold is very important because it has important repercussions on the estimated value of the shape factor (Caers and Maes, 1998, McNeil and Saladin, 1997) and hence on the Einal estimates such as the quantile, extreme values etc. There are many exploratory techniques like the mean excess plot which help in selecting the threshold. But in a simulation study, it is impractical to perform interactive data analysis required by the exploratory techniques to choose the threshold. The mean excess plot is the sample mean excess function defined as the sum of the excesses over the threshold divided by the number of the data points which exceed the threshold plotted with respect to the threshold. A reasonable choice of the optimal threshold is where this plot becomes linear. For further reading the reader is referred to Coles (2001). Boos (1984) suggests that the ratio of Nex (number of tail data) over N (total number data) should be 0.02 (50 suggests using Nex = 1.5J Caers and Maes (1998) propose to use a Einite sample mean square error (MSE) as a criterion for estimating the threshold. They use the threshold value that minimizes the MSE. In a similar fashion Beirlant et al (1996) Eind an optimal threshold by minimizing an approximate expression for asymptotic mean square error. The other methods include plotting the quantile, shape or scale factor or any quantity of interest with respect to different thresholds and look for a stability in the curve (Bassi et al, Coles pp:8486) Parameter Estimation Generally one is interested to generate values (PDF) from a particular distribution given a set of parameter values B which is expressed as: f (x / ) (411) where, x is a data vector [xl ... xm] of frequency. However, here we are faced with an inverse problem. Given the set of observed data and a model of interest, one is required to find one PDF that among all probability distribution that is possible for the model is most likely to have produced the data. The parameters can be estimated by several methods. The methods used in the literature include, maximum likelihood estimation (MLE), method of moments, probability weighted moments, elemental percentile method, the least square regression method. Generally, maximum likelihood estimation method is widely used and accepted by researchers though it has limitations. That is as long as 5> 0.5 the asymptotic properties (Appendix B) of the maximum likelihood estimators are preserved. When 1 < < 0.5 MLE are obtainable but do not have the standard asymptotic properties. MLE is not obtainable when 5 <1 (Coles, 2001, pp5455). The likelihood estimation approach for the GPD parameters is presented in the Appendix B. However, accurate estimation of the parameters is a huge research area in itself. Beirlant et al (2004), Castillo et al (2005) discuss these methods in detail. Hasking and Wallis (1987) report a comparison between the maximum likelihood estimation method, method of moments and the probability weighted moments. They conclude that method of moments and the probability weighted moments are more reliable compared to the MLE method unless the sample size is greater than 500. This section discusses the maximum likelihood method and least square regression approach in detail Maximum Likelihood Estimation MLE is a popular statistical method that is used to make inferences about the parameters of the underlying probability distribution of a given data set. There is nothing visual about the MLE method but it is very powerful and is precise for large samples. The likelihood of a set of data is the probability of obtaining that particular set of data, given the chosen probability distribution model. ML estimation starts with writing the likelihood function which contains the unknown distribution parameters. The values of these parameters that maximize the likelihood function are called the maximum likelihood estimators. In order to estimate the parameters, the likelihood function has to be constructed by reversing the roles of the data and parameter in Eq (411i), i.e., L(6 / x) = f (xl / ) (412) Thus L represent the likelihood of the parameter B given the observed data x and is a function of 0 For a model with k parameters, the likelihood function takes the shape of kdim geometrical surface sitting above a kdim hyperplane spanned by the parameter vector (Myung, 2003). Both fand L are PDFs but the maj or difference lies in the fact that they are defined in two different axes and are not directly comparable to each other. PDF fis a function of data given a particular set of parameter values defined on the data scale. The likelihood function L is the function of parameter given a particular set of observed data defined on the parameter scale. ML approach is totally analytical in concept. MLE's generally have very desirable large sample properties (Appendix C) However, MLE cannot be adopted in all situations. There are some limitations when small numbers of failures (less than 5, and sometimes less than 10 is small), MLE's can be heavily biased and the large sample asymptotic properties do not apply Least Square Regression The method of least squares assumes that the bestfit curve of a given type is the curve that has the minimal sum of the deviations squared leastt square error) from a given set of data. The parameters are obtained by solving the following minimization problem The GPD CDF can be obtained by using Eq. (44). The empirical CDF are plotting positions which are computed as: F:= ,i1.N(414) N+1 where N is the total number of samples, Nuis the index corresponding to the threshold u and P is the plotting position. Since it is a minimum finding, it can be seen as an optimization technique similar to maximum likelihood estimate method where we try to maximize the likelihood function. Least square regression requires no or minimal distributional assumptions. Unlike MLE, there is no basis for testing hypotheses or constructing confidence intervals. Accuracy and Convergence Studies for the Quantile Estimation In order to understand the capability of GPD to model the tails of the distribution, an error analysis in terms of modeling and sampling error was performed on the probability content estimation using tail modeling approach. A detailed discussion of this is presented in Appendix D. Based on the results, it can be concluded that the sampling error is high compared to the modeling error. If the sampling error is accounted for, tail modeling has a fairly good capacity to estimate extreme probabilities with good accuracy. In structural reliability studies, one is required to model the tails of the response which might follow some standard statistical distributions. Normal and Lognormal distributions are widely used to model physical quantities in structural reliability studies. Therefore, an investigation was done to study the accuracy and convergence capabilities of tail modeling approach to model the tails of normal and lognormal distribution which is discussed in detail in Appendix E. Then, the approach is demonstrated on a cantilever beam example and a tuned mass damper example to estimate the failure probability. The simulation studies (Appendix E) help us to conclude that based on the target quantile value we wish to estimate, the threshold and number of exceedance data need to be chosen. In the context of structural analysis, the interest is to address the response function. The failure probability is based on the response. If we wish to design a system with a low failure probability of the order of le4 or higher, then we are required to estimate quantiles of 0.9999 or higher. So, if the response follows a distribution close to normal, and we wish to estimate a quantile of 0.9999, a quantile of 0.95 and 100 exceedance data would enable us to estimate the quantile under 3% error. On the other hand, if the response follows more of a lognormal distribution (heavy tailed), then one is required to choose a higher threshold of the order of 0.99 or 0.995 with 100 or more exceedance data to be guaranteed to have a bias error of less than one percent and a RMSE error of less than 5 percent. However, it is not possible to come up with a table that explains the minimum exceedance data required for a level of quantile and required accuracy level because, in practical situations the distribution of the response is not known. Moreover, the challenge is to develop methodologies that can estimate the inverse measure corresponding to high reliability, with limited samples. Cantilever Beam Example The beam example that was considered in Chapter 3 is treated here also. Though simulating the beam example is not very expensive, we pretend that we have limited computer resources and are restricted to use a maximum of 500 simulations. We are interested in estimating the PSF for low failure probabilities with only 500 response data. Three different cases are discussed. Case 1 discusses the estimation of PSF for the strength failure mode using the tail modeling approach at different threshold values. Similarly the PSF estimation for the deflection failure mode is investigated in Case 2 and the system failure case where both the failure modes are considered simultaneously is discussed in Case 3. Case 1 and Case 2 are presented in Appendix F. For all the cases, the convergence of PSF with respect to different thresholds and accuracy of PSF at different number of samples are investigated. Two methods, namely the maximum likelihood method and the least square regression methods are used to estimate the parameters. The results from these methods are compared. In the system failure mode, both the failure modes are considered simultaneously. The safety factor for each failure mode is evaluated. For each realization, the critical of the two safety factor is considered the system safety factor. Once, the system safety factor is obtained for 100 simulations, the tail modeling approach is carried out and the corresponding system PSF is estimated. Here the design variables are w=2.6041; t3.6746 and the allowable deflection Do=2.145. This combination of the variables allows equal contribution of the modes to the total failure. The contribution of the modes are: P, = 0.00099; Pf, = 0.00117 ; nf ~f = 0.00016 . The convergence plots of PSF for the system failure case is presented in Figure 42. Figure 42 shows that the PSF estimated through parameters estimated by the regression method is unstable in contrast to the PSF estimated through parameters from the ML method. This behavior was observed in the other two cases too. The PSF estimate at a threshold of 0.9 is 0.99 with a standard deviation of 0.003. The corresponding failure probability estimate is 0.0017 with a standard deviation of 0.0002 which is about 10% of the estimate. The number of samples used in the above studies is 500. The PSF is expected to converge to the actual value when the number of samples is increased. Tuned MassDamper Example The tail modeling approach to estimate the failure probability is demonstrated here with the help of a tuned massdamper example. The tuned massdamper problem in Figure 43 is taken from Chen et al (1999). The problem involves the failure probability estimation of a damped single degree of freedom system with dynamic vibration absorber. Figure 43 illustrates the tuned damper system consisting of the single degree of freedom system and a dynamic vibration absorber to reduce the vibrations. The original system is externally excited by a harmonic force. The absorber serves to reduce the vibration. The amplitude of vibration depends on R = the mass ratio of the absorber to the original system r, the damping ratio of the original system S=" ratio of the natural frequency of the original system to the excitation frequency r2= n2, ratio of the natural frequency of the absorber to the excitation frequency The amplitude of the original system y is normalized by the amplitude of its quasi static response and is a function of four variables expressed as: y = (415) 1 1 1 1 1 1 r, a r2 6 2 6 1 2 This example treats r, and r2 as random variables. They follow a normal distribution N(1,0.025) andR = 0.01, r=0.01. The normalized amplitude of the original system is plotted in Figure 44. There are two peaks where the normalized amplitude reached undesirable vibration levels. The corresponding contour plot is presented in Figure 45. The failure region is denoted by the red band. This failure region is an island. That is, the failure region has safe regions on either side of it. This introduces additional challenges of not being able to use analytical approaches like FORM because the failure region is not a half plane. The obj ective of the problem is to reduce the risk of the normalized amplitude being larger than a certain value. The limit state for this case is expressed as: %(6,6= MG,)70(416) where yo can be considered as the allowable level of vibration. When the limit state in Eq. (416) is greater than 0, failure is said to occur. Increasing or decreasing yo will help in decreasing or increasing the failure probability respectively. yo =27 is considered for the discussion. The corresponding failure probability with le7 sample MCS is estimated to be 0.01031. The tail modeling approach with 500 samples and 100 repetitions are used to study the convergence and accuracy estimates of PSF. The plot for PSF at various thresholds is presented in Figure 46. From the plots, it seems that regression behaves better compared to the ML method. There is a discrepancy in the plot corresponding to the ML method. The PSF estimated at 0.96 threshold is more accurate than the PSF estimate at a threshold of 0.98. In order to understand the behavior of the tail of the tuned mass damper, a log plot of the CDF with le5 samples is examined. The plot is presented in Figure 47. It is evident from the plot that there are two curvatures in the tail which are difficult to model. The GPD has to capture this shape with less exceedance data. This is the reason for the discrepancy in the ML plot, the shape of the tail at the area of interest (Sr=1) modeled by ML with a threshold of 0.96 was better than the tail model with a threshold of 0.98. In order to further explore the tail models from each method, the tail model from each method is superimposed on each other and the comparative plots are presented in Figure 47. It is clearly observed that the tail modeled based on ML approach denoted by the dashed line attempts to capture the second curvature and in the process introduces error in the PSF value corresponding to a failure probability level of 0.01. On the other hand, the model obtained based on regression parameters represented by the solid line, approximates the tail in a linear fashion and is accurate compared to the tail based on ML method for a failure probability of 0.01. However, the ML method is expected to perform better when the failure probability to be estimated is low. When the failure probability is low, it becomes necessary to model the second curvature adequately to estimate the PSF with reasonable accuracy. The ML method can perform better than the regression approach in modeling the second curvature. Such types of CDF with double curvature might not be encountered in structural applications often and solving this problem with less number of samples is very challenging. With more insight on the problem, it is possible to select the suitable parameter estimation method depending on the requirements. Here, if one needs to estimate PSF corresponding to failure probabilities lesser than le3, then the second curvature has to be modeled well and hence ML method is a good choice. Else, for higher failure probabilities, regression can perform better. Alternate Tail Extrapolation Schemes In this section, simple tail extrapolation techniques are developed to estimate the PSF for low target failure probability using MCS that is sufficient only to estimate the PSF for substantially higher failure probability (lower target reliability index). Failure probability can be transformed to reliability index by using Eq. (23). The same transformation is applied here to the CDF of the reciprocal of inverse measure. Since we are modeling the upper tail, reciprocal of the inverse measure is of interest here. The tail portion of the resulting transformed CDF (relationship between inverse measure and reliability index) is approximated by a linear polynomial in order to take advantage of fact that normally distributed inverse measure is linearly related to the reliability index. Since this approximation will not suffice if the inverse measure follows distributions other than normal, a logarithmic transformation is then applied to the reliability index which tends to linearize the tail of the transformed CDF. This tail is approximated using a quadratic polynomial. Since both these techniques use only data from the tail, another technique that approximates the relationship between inverse measure and reliability index from the mean to the maximum data using a quadratic polynomial is developed. The three extrapolation techniques discussed are described with the help of a figure. A data set of size 500 with a mean of 10 and variance 9 following a lognormal distribution is used to illustrate the three techniques. Figure 48, A) shows the general relationship between PSF reciprocal and 1Pf. Applying the transformation modifies the CDF as in B). If the PSF reciprocal follows a normal distribution, this curve will be linear. Therefore, the first technique approximates the tail of the curve in B) by a linear polynomial. In order to take advantage of the remaining data, half of the curve in B) is approximated by a quadratic polynomial. It is sufficient to consider only half of the data because of the symmetry in reliability index. The third technique applies a logarithmic transformation to the reliability index and approximates the tail of the curve shown in C) by a quadratic polynomial. In all the above three techniques, once the coefficients of the polynomial are obtained, the PSF reciprocal corresponding to any higher reliability index can be found. It is to be noted that with 500 samples, it is possible to estimate only PSF reciprocal at reliability levels less than 3. Classical tail models discussed in the previous chapter and the alternate extrapolation techniques discussed in this chapter allow predicting PSF reciprocal at much higher reliability indices without additional reliability analysis. The alternative extrapolation techniques and classical tail modeling techniques are conceptually the same. The maj or difference in perceiving the two classes is that the classical tail modeling techniques model the CDF of PSF reciprocal, whereas the extrapolation schemes approximates the trend of PSF reciprocal in terms of reliability index. Figure 41. Tail modeling of F(u) using the threshold u. The region of y>0 is failure 1  Tail part ~ Safe region Failed region 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 Fu I.L. 1 1.( 0 0.! Figure 42 Convergence of PSF at different thresholds. A) MLE B)Regression. Cantilever beam system failure mode. 500 Samples. 100 repetitions Nex 1 090 90 80 70 60 50 40 30 20 10 I . 32~~~~~   LCI +UCI 31C1   M ean M Median 98  97 r u.. 0.i 0.i 0e 0'J.8 m, On2 Absorber k2 Mn, mni Original y I Isystem F = cos(mt) clkl Figure 43. Tuned vibration absorber 9095 1 1 05 1 71 1.1 0 9 r2 Figure 44. Normalized amplitude vs rl and r2 ~i~f~f~i_~ I : I i I1 0. ' 0.9 0.95 1 B~eta Figure 45. Contour of the normalized amplitude 67 ~_ Nex 1. O0 90 80 70 60 50 40 30 20 10 Actual PSF SLCI   UC I BMean xMedian 1.4 1.2 L. 0.8 ~~ ~ 0.7 ' r ;c~~  .8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 Fu 00 90 80 70 60 50 40 30 20 10 ( Actual PSF LC I tUCI Mean  edian    t=*~ C~ 1.86 1.72 1.58 1.44 1.16 0.88 0.74 0.~8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 Fu Figure 46. Convergence of PSF at different thresholds. Tuned Mass Damper. A) MLE. B)Regression. 500 samples. 100 repetitions. LCI Lower Confidence Interval. UCI Upper Confidence Interval 185 Samples i I ,, MLE I .. Reg h: 500 Samples .1 I 10 1 V 102 L. 103 0.5 1 1.5 2 2.5 3 3.5 Figure 47 Tuned Mass Damper. Comparison of tail models using regression and ML method 69 102 2 2 100 0.2 0.4 0.6 0.8 1 30 0.2 0.4 0.6 0.8 1 S Sr A 'B 3 0.4 0.5 0.6 0.7 0.8 0.9 1 Sr C Figure 48. Transformation, V of the CDF of PSF reciroclV L (Sr).l A) CDF. of S. B) Inverse Standard normal cumulative distribution function applied to the CDF. C ) Logarithmic transformation applied to the reliability index. 70 CHAPTER 5 MULTIPLE TAIL MODELS Multiple Tail Model A total of Hyve methods to model the tails of data were discussed in the previous two sections. However, in practical applications where there is no information on the form of the output distribution, there is no way to explore which technique performs better. A single technique that performs well for a particular example might not perform the best for another example. In fact, a single method that performs the best at a particular reliability index might not perform the best at a different reliability index for the same example, as will be shown in the numerical examples. In addition, there are no known straight forward ways to quantify the error in the estimates. In order to take advantage of all the techniques and still come up with the best prediction, we propose to apply the Hyve techniques simultaneously and use the median of the Hyve estimates. It is observed that the median is a robust estimate than the mean. This method is demonstrated first on seven different statistical distributions followed by two engineering examples. In all examples, it is observed that the median of the five estimates is at least the second best compared to estimates from individual techniques. Moreover, the range of the Hyve estimates closely approximates the error in the median. Thus using multiple tail models not only ensures to choose a good estimate and buys insurance against bad predictions but also provides an error of the estimate. Figures 5.1 to 5.5 show the various tail modeling techniques for a lognormal distribution with a data mean of 10 and standard deviation of 3. In each case, the le5 sample curve and 500 sample curve are superimposed to highlight the extrapolation. The Sr is computed as discussed in the implementation in next section. Figure 5.4 shows the multiple tail model approach (MTM). In order to understand to clearly see the range of the five methods, the region beyond the threshold is zoomed and presented in Figure 5.5. The range of the five methods at three different reliability indices is marked by double headed arrows. Mean of these ranges computed over the entire extrapolation zone are capable of representing the error associated with the median estimate from MTM.A representative actual error is presented in the plot. The QH is the median and overlaps with the MTM curve. Numerical Examples Standard Statistical Distributions The standard statistical distributions that are discussed in this section are presented in Table 51. A data set with mean 10 and standard deviation 3 is considered. In order to compare the performance of the proposed method on different distributions, a common target failure probability is used to find the capacity which is in turn is used to normalize the parameters. Implementation of this exercise is as follows: Pf smse = 0.00135 ( smsget =3), X=10, COV = 0.3, 0=COV x X 1. Use x and CO~to estimate parameters (ao and bo) for each distribution from Table 51. These parameters can be used to generate random numbers. 2. Find capacity C as (14~,ef quantile. Inverse cumulative distribution functions (icdf) can be used for this. In the case of single parameter distributions, add shift factor to the C. 3. Find normalized mean using x= We are interested in the safety factor reciprocal S, following different statistical distributions. The normalized mean is the mean of the Sy 4. Use x and CO~to find new parameters (a andb) for all the distributions. 5. Generate N = 500 Latin hypercube samples in [0, 1] 6. Find sy at each sample using (a, b) and inverse icdf 7. Use Eq (414) to estimate plotting positions 8. Plotting Sy and P provides the empirical CDF ofSy It is to be noted that the ordinate is (1Pf). That is, we are interested in the upper tail. As Pf decreases, 1 Pf increases, reliability index increases, S decreases and sy increase. 9. A threshold of 0.9 is considered for classical tail modeling techniques and the parameters of GPD that approximates the conditional CDF beyond the threshold are estimated using Maximum Likelihood and least square regression approaches. 10. The transformations discussed in the previous section can be applied to the CDF of S, and s, at higher reliability indices can be estimated directly from the approximated relationships. 11. There are totally five estimates (steps 9 and 10) of sy denoted by Sind in the extrapolated regions. Record the median and range of f md .as 5m and 5Ra respectively. 12. Exact values of Sy can be obtained using icdf functions. 13. Compute error in the estimate of individual methods eind as abs (sy Smd ). 14. Compute error in multiple tail model (MTM) estimate eMTM~n RS abs(si. 'm) 15. Compute the ratio of mean of eMTM to mean of 5Ra over the entire extrapolation zone and denote it as rl 16. Repeat steps 5 to 15, 1000 times for each distribution. 17. Record the median, mean of eind and eMTM OVer 1000 repetition as '"ooomind > 1000pind and oomMTM, >100MTM 18. Record the median and mean of the ratio in step 15 over 1000 repetition as 'ooom, For each distribution, boxplots of eind and eMTM are preSented. 'ooomind and 'ooomMTM COmpared at three reliability indices between 3 and 4.2. The results for the normal distribution are presented in Figure 56. The obj ective of this plot is to see how well the 5m estimate compares to the Sind For 1000 repetitions, only the median and the mean can be compared because the range is volatile. It was observed that the median was a robust estimate compared to the mean. It is evident that the median is less sensitive to the outliers compared to the mean. At P =3, the regression (Rg) performs the best followed by the linear tail (LT). The MTM performs as good as the second best. At p =3.6, the LT performs the best followed by the regression (Rg) and MTM performs as well as LT. At P =4.2, the LT performs the best followed by the quadratic tail (QT) and MTM's performance is as good as LT. It is to be noted that the individual techniques that provided the best estimate differed at different reliability indices. However, the MTM estimate was consistent and it performed at least as good as the second best estimate from the individual techniques. This approach serves as an insurance against bad predictions, if we were to use a single technique. Box plots of rl ratio are presented to see how well the range of the methods compares to the mean of the eMTM Over the entire extrapolation region. The results for the normal distribution are presented in Figure 57. It is clearly observed that the ratio of the mean of the error to the mean of the range in the entire extrapolation region is around 1. If the ratio is 1, then the range captures the error in the median perfectly. Here, the range provides an estimate of eMTM that might slight over or under estimate the actual error. The boxplots for the remaining distributions are presented in the Appendix G. Table 52 provides a summary of the performance of the individual methods and the MTM for all the distributions. Table 52 shows that not a particular technique was the best for all the distributions and the technique that provided the best estimate for a particular distribution varied as the reliability indices changed. However, the MTM estimate was mostly close to the 2nd best available estimate. The summary of rl ratio is presented in Table53. It is observed that other than uniform distribution, the ratio varies between 0.15 to Table 51. 7 standard statistical distributions and their parameters Distribution Parameters a b Normal ;nr 5(' LogNormal n( 05b Gamma 2 Extreme Type _0.577 i 1 b o Uniform 1 2 xa x+a 2 2 Single parameter distributions Exponential Y x Rayleigh a 74 0.35. This means that half of the range over estimates the error by a factor of 1.5 to 3. In order to provide an insight what number should one expect for the r ratio, a simple exercise is performed. Standard normal random numbers of size 5x7 (represents 5 estimates at 7 different reliability indices) are generated and rl is computed over le5 repetition. The resultant number is 0.18. Table 52. Summary of the performance of individual methods and MTM for different di stributi ons Rel Index 3 3.6 4.2 Performance 1st best 2nd best MTM 1st best 2nd best MTM 1st best 2nd best MTM Distributions Normal Rg LT 2nd LT QH 1st LT QT 1s Lognormal Rg LT 1st QH Rg 3rd QH QT 3rd Gamma Rg LT 2nd LT Rg 1st QT LT 2nd Ev QH Rg 1st QH QT 1st QH QT 2nd Uniform QT Rg 1st Rg ML 2nd Rg ML 2nd Rayleigh Rg LT 2nd LT Rg 1st LT QT 1st Exponential LT Rg 2nd Rg QT 1st Rg QH 3rd Table 53. Summary of the Error to range ratio for different distributions Distribution looom Normal 0.279 Lognormal 0.349 Gamma 0.181 Ev 0.183 Uniform 0.003 Rayleigh 0.148 Exponential 0.241 Since the concept of regression is used to estimate the parameters in all methods other than the ML approach, the residuals are monitored for each method to see whether they follow the normal distribution as per classical statistical theory. It is expected that the type of distribution of the residuals can provide a good idea about the quality of the fit. A similar study was made for the normal distribution case and the normal QQ plots for the residuals are presented in Figure 58. It is observed that the residuals of none of the methods follow a normal distribution. Investigating the corresponding histograms reveal that they follow a log normal distribution. In these situations, other regression schemes like weighted approaches can be adopted depending on the knowledge of the histograms and QQ plots of the residual which is scope for future work. Engineering Examples Application of multiple tail models for reliability estimation of a cantilever beam The cantilevered beam example treated in Chapter 3 is considered here. The obj ective of the original problem is to minimize the weight or, equivalently, the cross sectional area, A =w. t subject to two reliability constraints, which require the reliability indices for strength and deflection constraints to be larger than three. Eqs (317) and (318) are rewritten as 600 600 ,Y+ X sR R 4L Y (X D Ew t t'w2 Gd = 1 1 (5 2) Do Do It is to be noted that the performance functions are expressed in a fashion such that failure occurs when G, or Gd is greater than 1. Here, we consider a system failure case. That is, both the failure modes are considered simultaneously. The optimal design variables w=2.6041, t3.6746 is taken from Qu and Haftka (2003) for a system reliability case. The corresponding reliability index is 3. The allowable deflection Do is taken to be 2. 145. For each sample the critical safety ratio reciprocal (maximum of the two) is computed. Since we are dealing with S,, it is to be noted that n=1Perazer. The conditional CDF of Sy can be approximated by classical techniques and the relationship between Sy and reliability index be approximated by three different alternate techniques. Finally, using the MTM approach will give a better prediction compared to any individual method. This procedure is repeated 1000 times and the results are presented in Figure 59 It is observed that the LT performs the best at all three reliability indices followed by the QH as the second best. MTM consistently performed close to the second best estimate. Figure 510 presents the plot of rl ratio. The ratio is 0.23 and thus strengthens our conclusion that the half of the range approximates the error in the median by a factor between 1.5 and 3, which is close to 2 here. The main advantages of the techniques discussed are that no statistical information is required and huge reduction in computational power. In order to understand the sacrifice in accuracy for the reduction in computational power a MCS study is performed. 100 repetitions of Sy estimates with 5E5 samples and the corresponding standard deviation are computed and presented in Table 54. At a reliability index of 4.2, the standard deviation in Sy estimate is 0.04 which is the same error from MTM using 500 samples. Therefore, for a same level of accuracy, the reduction in computational power is about 3 orders of magnitude (5E5 to 5E2). Table 54. PSF reciprocal estimates and standard deviation at different reliability indices Rel Index 3 3.2 3.4 3.6 3.8 4 4.2 S, 1.012 1.032 1.05 1.07 1.09 1.12 1.13 SD 0.003 0.004 0.01 0.01 0.01 0.02 0.04 Application of multiple tail models for reliability estimation of a composite panel The design of the wall of a hydrogen tank which is a composite laminate operating in cryogenic temperatures addressed by Qu et al., (2003) is considered as a second engineering example. In order to reduce matrix cracking and hence hydrogen leakage in cryogenic environments, the laminates need to have smaller angle between the plies. However, these types of laminates do not carry the biaxial hoop and axial stresses due to pressure. Qu et al., (2003) investigated options for minimizing the increase in thickness required to carry both mechanical 77 and thermal residual strains. The geometry and loading conditions of the problem is presented in Figure 51 1. The laminate is subj ect to mechanical (Nx and N,) load and thermal loading due to operating temperature 423oF where the stress free temperature is 300oF. The obj ective of the actual problem was to optimize the weight of laminate with two ply angles [Oz,,102 The ply angles and ply thickness (t, and t2 ) are the design variables. The material used is IM600/133 graphite epoxy of ply thickness 0.005 inch. Qu et al., (2003) performed the deterministic design optimization using continuous design variables. In order to account for the uncertainties and make the deterministic optimal design comparable to probabilistic optimal design, they used a safety factor of 1.4. For further details, the reader is referred to Qu et al., (2003). The mechanical properties are presented in Table 55. The deterministic optimization problem is formulated as Minimize h =4(tz + t2 Such that t,,, t2 >0.005 (53) El < SFE 1" U, ef S;F 2 2U SFZ ~ 5;121 1 I: All material properties are assumed to be random, uncorrelated and follow normal distributions. The coefficient of variation is presented in Table 56. Here, thickness is in inches and h is the laminate thickness. Superscripts U and L represent upper and lower limits of the associated quantities. El, E2, and 712 are the ply strains along fiber direction, transverse to fiber direction, and shear strain, respectively. E2, G127 1 and a2 are functions of temperature. Qu et al (2001) show that at 20.2K (423 o F) the thickness of the optimum laminates obtained by using temperature dependent material properties is 80% less than that using constant material properties at 298 K (77o F ). Qu et al (2003) observed that e2U is the active constraint. For a feasible design, 21 different temperatures uniformly distributed between 20oK and 300oK is 78 considered and the strain constraints are applied to these temperatures. The mean is computed at a particular temperature and random numbers generated based on coefficient of variation. The mean for other parameters are presented in Table 57. The deterministic optimal design obtained by Qu et al (2003) is presented in Table 58. The transverse strain on the first ply (direction 2) is the critical strain. The limit state is defined as the difference between the critical strain and the upper bound of the allowable strain. G; = Ez  E"r (54) Failure is said to occur if Eq. (54)> 0. In order to extend this example for demonstrating the multiple tail models approach, the second optimal design in Table 58 is considered. The loads are increased by a factor of 1.1 (Nx 5280 lb/inch and Ny 2640 lb/inch) so that the failure probability is of the order of l e3. 500 random critical strain values are generated. Similar to the previous examples, 5 different techniques are used to fit the tail of the response distribution. The results are presented in Figures 512 and 513. As observed in other examples the MTM compared at least close to the second best in estimates from individual methods. However, the ratio of mean error to mean range is 0.22 which means that half of the range overestimates the error by a factor of around 2.which falls within the range of 1.5 and 3 that was observed earlier in section. The le7 sample based PSF reciprocal estimation and the corresponding standard deviation is presented in Table 59. At reliability index of 4.2, it is seen that the standard deviation is 0.0063 and the multiple tail model provides an estimate with an error of 0.05 which is off by a factor of 8. The number of samples used for this estimation is 500, which is about 4 orders of magnitude less than le7 Due to the prohibitive computational power required, an exercise similar to the previous example is Table 57: Mean of random variables not performed here. Instead the estimate from 500 samples and le7 samples are compared directly . Table 55. Mechanical properties of the composite laminates Elastic properties E,,E ,G,, and ,, Coefficients of thermal a, and a, expansion Stressfree temperature Failure strains Safety factor 6 ,6 ,6f,ef nd7{ S, Table 56: Coefficient of variation for random material properties 0.035 0.035 0.03 0.06 0.09 21.5x100 0.359 300 0.0109 0.0103 0.013 0.0154 0.0138 Table 58: Or (deg) Deterministic optima found by Qu et al (2003) 6~ (deg) tl (in) t2 (in) h (in) 0.00 27.04 25.16 28.16 27.04 27.31 0.005 0.01 0.005 0.02 0.015 0.020 Table 59: Composite laminate. S, reciprocal estimates and standard deviation at different reliability indices Rel Index 3 3.2 3.4 3.6 3.8 4 4.2 1E+07 1.0043 1.0304 1.0579 1.0862 1.1157 1.1480 1.1824 SD 0.0007 0.0009 0.0015 0.0021 0.0025 0.0046 0.0063 101 10~ 13 104 104 0.5 1 1.5 Figure 51.Lognormal distribution. Classical tail modeling techniques MLE and Regression 3.5 2.5 . Th reshold c~z I1015 Samples 1.5 LT 1 500 Samples 0.5 0.4 0.6 0.8 11.2 1.4 Figure 52. Lognormal distribution. Linear fit to tail (LT) and Quadratic fit to half of the data (QH) 0.5 S0.5  1.5 0.5 0.6 0.7 0.8 0.9 1 1.1 1 Figure 53. Lognormal distribution. Quadratic fit to the tail (QT) Th reshold 10e5 Samples ,QT 500 Samples .2 1.3 1.4 1.5  Reg eLT 1 J IQH 0.5 T 500 Sample 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 Figure 54.Lognormal distribution. Multiple tail models Adual Ermrr r. *Threshold 165% Samples tRange 2 / IML LT 1.t ,P I*QH 1 * MTMd 0.6 0 7 0.8 0.9 1 1.1 1 2 1.3 1 4 1.5 1.6 Figure 55 Lognormal distribution. Multiple tail modelsextrapolation region + ML Rg OT LT H looom = 0.016 1000 Id=008 MTM i 1 ML Rg OT LT H looom = 0.022 1000 Id=005 MTM + ML Rg OT LT Q looom = 0.028 100 n = 0.032 MTM  1 = 4.2 0.074 0.034 0.028 0.074 0.051 0.033 S= 3 0.006 0.023 0.016 0.017 0.006 0.029 0.018 0.020 p = 3.6 0.031 0.031 0.031 0.041 looom = 0.018 1001n=0.01B ind oom. = 0.047 1001n=0.045 ind looom = 0.093 1000md007 FIind .8 0.041 0.048 0.022 0.028 0.026 0.033 0.2 0.15 0.3 S0.2 0.1 0 0.15 0.05 0.05 0 0.12 0.1 0.08 S0.06 0.04 0.02 O 0.06 S0.04 0.02 0 0.05 0 MTM MTM MTM Med= Median. ML=Maximum Likelihood, Rg Regression, QT Quadratic fit to the Tail data between Sy and In( P ), LT Linear fit to the Tail data between S~ and a QH Quadratic fit to Half of the data between S~ and a Dashed line inside the box depicts the mean and the solid line, the median. Figure 56. Normal distribution. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices oom =0.279 1000=0.5 0.55 025 0.55 0.45 0.45 0.35 S0.3 0.25 0.2 0.15 0.1 0.05t MTM Figure 57: Normal distribution. Boxplot of I ratio. MTM MTM MTM Med= Median. ML=maximum likelihood, Rg Regression, QT Quadratic fit to the Tail data between Sy and In( P ), LT Linear fit to the Tail data between S~ and a QH Quadratic fit to Half of the data between S~ and a Dashed line inside the box depicts the mean and the solid line, the median. Figure 58. Cantilever Beam. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices ML Rg QT LT H looom = 0.017 1000 Id=001 MTM 002 I + . ML Rg OT LT H looom = 0.028 100 n = 0.034 MTM 0.8+ 0.6 0.4 0.0 ML Rg QlT LT O looom = 0.040 100 n = 0.051 MTM 0.25 0.2 S0.15 e 0.1I 0.05 1 = 3 0.037 0.020 0.015 0.016 0.052 0.026 0.018 0.021 S= 4.2 0.135 0.051 0.435 0.060 13 = 3.6 10m.d= 0.040 0.078 0.034 0.022 0.027 100 In=0.054 0.144 0.041 0.025 0.032 ind 0.6 = 0.021 :0.024 1000m. 1000 n FIind 0.3 0.25 0.2 e= 0.15 0 5 0 looom = 0.087 100 In=0.101 ind 0.028 0.039 0.033 0.048 0.5 0.4 0.3 0.15 0.1 0.08 E 0.06 e0.04 0.05 O 0.02 0 QQ Plot of Sample Data versus Standard Normal Regression 0 3 7  m 0 4 .,.+ 4 + +  '0 7 + m~ +' O + 0 9 + QQ Plot of Sample Data versus Standard Norrnal QT 0 2 0 3 m 0 4 ~,+ 4. 4.4 '0 7 m .+ E + 0 9 25 2 1 5 1 0.5 0 0 5 1 1 5 2 2.5 2 5 2 1 5 1 0 5 0 0.5 1 1 5 2 2 5 Standard Norrnal Quantiles Standard Normal Quantiles QQ Plot of Sample Data versus Standard Norrnal QQ Plot of Sarnple Data versus Standard Norrnal QH LT 02 ,., 02 0 3 I 0 3 0~ 4 4.~,+ 4 +1 0 4 4. 4.+ + + 2 5 1 . O 5 1 15 2 2.25 15 05 O.5 1 1 52 2 Stndr Nora Quanile Stndr Nom lQun Figure 59. NomlQ lto hersdasfrteomldsrbto looom = 0.231 1000 =0.7 3.5 2.5 MTM Figure 510. Cantilever beam. Boxplot of I ratio. Ny ~J;, Figure 511. Geometry and loading for the composite laminate. Xhoop direction, Yaxial direction + + i ML Rg Q~T LTQH loo = 0.026 MTM MTM $ + j+ ML Rg QT LT QH 1000 m =008 0100 T = 0.0 61 MTM + 0 MTM ML Rg QTLT QH 100m = 0.081 1000 MM .0 MTM + MTM S= 3 0.034 0.044 S= 3.6 0.111 0.073 0.040 0.032 1.668 0.094 0.049 0.039 1000 1= 4.2 looom = 0.166 0.210 0.121 1001n=0.547 227.298 0.156 2 looom = 0.027 1001n=0.032 ind 0.25 0.2 0.15 0.05 0.3 oom. = 0.080 1000nd016 Cind 010 0.8 0.042 0.085 0.026 0.021 0.032 0.025 0.058 0.049 0.067 0.057 0.6 S0.4 0.2 0 0.6 '0.4 0.2 0 0.5 1.2 1 0.8 0.4 0.2 0 ML=Maximum Likelihood, Rg Regression, QT Quadratic fit to the Tail data between SI and In( 4), LT Linear fit to the Tail data between S, and a QH Quadratic fit to half of the data between S, and a .Dashed line inside the box depicts the mean and the solid line, the median Figure 512. Composite laminate of cryogenic tank. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices... looom = 0.220 1000 =0.5 2.5 2 F 1.5 0.5 MTM Figure 513. Composite laminate of cryogenic tank. Boxplot of rl ratio. CHAPTER 6 CONCLUSIONS High reliability estimations with expensive computer models pose a challenge because of the high computational expense. Extrapolating into higher reliability indices with information from lower reliability indices can be one solution to this problem. In order to do this, the tails of the distribution are to be modeled well. Inverse reliability measures feature several advantages and is widely used by researchers in structural reliability. This work explores the usefulness of inverse measures in tail modeling. Two classical tail modeling techniques were discussed.Three extrapolation techniques that can complement the classical tail modeling techniques were developed. This work proposes to use a multiple tail model approach in which all the five techniques are applied simultaneously to estimate the inverse measure, and the median is taken to be the best estimate. It is shown that the range of the five methods can be utilized to approximate the error associated with the estimate. The proposed method was demonstrated on seven statistical distributions and engineering examples. It was shown that the sacrifice in accuracy compared to the crude MCS is relatively small compared to the immense reduction in computational power. APPENDIX A PSF ESTIMATION FOR A SYSTEM RELIABILITY The estimation of PSF for a system reliability case is demonstrated with an example. This example is artificial and is an illustration of a case with multiple limit state functions. The example is presented in Figure Al.The limit state functions are: G, =( y 3 X (A l) (x + y5)23 (x y 12)21 (2 G2 =+ 0.8 (2 G3=y 20 (A3) The limit states are functions of two random variables x and y. The distributions of the random variables are presented in Table A1. The example has been created in such a fashion that each limit state and its intersection with other limit states contributes significantly to the total failure. It should be noted that non linear functions are used to emphasize the advantage of using MCS and a PSFbased approach to estimate failure probability. Using FORM in these circumstances might introduce huge error due to linearizing the limit state function at the most probable point of failure. The contributions of failure modes to the total failure are listed in Table A2. The system PSF can be estimated as described in the flowchart in Figure A1. The limit states are functions of two random variables x and y. The PSF estimated for different target failure probabilities are listed in Table A3. For a target failure probability of 0.5 57 which is the actual failure probability as listed in Table A3, the PSF is estimated as 1. When a lower target failure probability of 0.5 is used, the PSF is estimated to be 0.844 indicating that the system is unsafe and when a larger target failure probability of 0.6 is used, the PSF is found to be 1.12 showing that the system is over safe. This shows that PSF based approach can adequately address system reliability case accounting for the interactions between the modes. Table A1. Distribution of random variables. Multiple failure modes example Random Variable x y Distribution Normal (4,0.5) Normal (2,3) Table Limit states A2. Contribution of failure modes to the system failure Failure % Contribution probability to total failure 1 2 3 1 & 2 1 & 3 2 & 3 1,2 & 3 Total 0.432 0.468 0.336 0.386 0.253 0.249 0.209 0.557 77.55 84.02 60.32 69.29 45.42 44.71 37.52 100 Table A3. Target Pf 0.557 0.5 0.6 PSF at different target failure probabilities PSF 0.9996 0.8444 1.1191 n Failure i *' :' ,"':'" \ :," i~ ~ ~ ~r ;I~ .' .'c: 'r.CI' "r i I ' ... :~; I~ FlllrlP x 'r. 2 3 4 5 6 '7 Figure A1. Two variable system. Algebraic multiple limit state functions example APPENDIX B GENERALIZED EXTREME VALUE THEORY Let X1, X2,...be a sequence of independent random variables with a common distribution function F, and let Mnn = maxt (X,..., Xn) Pr(Mn < z} = Pr(X1 < z,...Xn <; z) (B 1) = PrfX, = (F` is not known, so it does not help. Here we look for approximate families of models for Fn which can be estimated based on extreme data only. This is done as follows: Mn n n(B2) for sequences of constants (an > 0) and {bn) Appropriate choices of constants stabilize the location and scale of Mn as n increases. Essentially we seek limit distributions for Mn with appropriate choices of constants. The all possible range of limit distributions for Mn is given by the following theorem, the extremal type theorem, a product of FisherTippet (1928) work. If there exists sequences of constants {an > 0) n b) uhta Pr ((M~ b)l /~ a< z)~ H(z) as na a (B3) for a non degenerate distribution function H, then H belong to one of the following families: 1. H(z) = exp exp ,JI c < z < cO (B4) 2. H(z)= ex z b a iz> b (B5) 3. H(z) = a (B6) 2b~ Here, a and b are the scale and location parameters. a is the shape parameter. These three classes of distributions are called the extreme value distributions with types I, II and III, widely known as Gumbel, Frechet and Weibull families respectively. The above theorem implies that the normalized maxima have a limiting distribution that must be one of the three types of extreme value distribution. The three families of distributions can be combined into a single family of models having distribution function of the form H(z)= exp [ 1+( z ~ug (B7) such that (zl :+ 5(z u) / a > 0), where the parameters satisfy co < pu < co, a > 0 and co < 5 < CO This is the generalized extreme value (GEV) family of distributions. Here, pu is the location parameter, a is the scale parameter, 5is the shape parameter. Type II and III families correspond to 5 > 0 and 5 < 0 cases in (9). 5 = 0 can be interpreted as the Gumbel family with distribution function: H~z)=exp xp (B8) APPENDIX C MAXIMUM LIKELIHOOD ESTIMATION The PDF of the Pareto Distribution is given by: 1 r~l~rz 1+i o f (x) = O (C1) e O 5= 0 Here, the z represents the exceedances. 5 and a are the shape and scale parameters respectively. The likelihood function for a sample of zl...zn of iid GP random variables is given by: n~= ls i 1+~)5~ L (zi,5 o ) = ::i i=1 5= (C2) The log likelihood function is expressed as: nlogE( 51+1 log= 1+ zi 1+ 5zi >0o log L(f ) = (C3) n logs a zi 5 = 0 Oi=1 The maximization of (C3) is best made by reparametrization: (,, r>() (, 5) with r = yielding h = log L (r, 5) = n log f1+ jlog (1 zi)Si (C4) = n log5+ n log r l1+ lg1ri (CS) The max of (CS) is found as the minimum of the negative of(C5). The MLE are given by differentiating (CS) and equating to zero. dhn+ 2 Ilog( + rzi) =0 (C6) ++1z Ci 0= (C7) (C6) yields i = 1 log(1+ fzi) and (C7)can be solved for Z Numerical techniques are used for i=1 the maximization of h in (CS) using algorithms that are available in standard software packages. In most situations estimation is done using iid samples. In such cases, the interest is to determine the behavior of the estimator as a sample size 3 0 This is referred to as asymptotic behavior. Under certain regularity conditions like: the first and second derivatives of the log likelihood function must be defined., the fisher information matrix must not be equal to zero, MLEs exhibit following properties Asymptotically unbiased (i.e.,) its bias tends to zero as the number of samples increases to infinity. MLE is asymptotically guassian. That is, as sample increases the distribution of MLE tends to normal distribution. Here, the covariance matrix can be obtained as the inverse of the fisher information matrix. APPENDIX D ERROR ANALYSIS FOR GPD This appendix describes the error analysis for the probability content estimated using tail modeling. When the distribution of exceedance is approximated using GPD, there are two sources of error: 1. Modelling error 2. Sampling component of the total error The modelling error refers to the capacity of the GPD model to accurately represent the distribution of exceedance data or the tail of the CDF. If the exact CDF is available, the CDF estimated through GPD can be compared and the resulting error is the modeling error. Depending on the empirical CDF, the GP fit will vary. This introduces sampling error. The sampling error comes from the fact that the GPD parameters vary depending on the exceedance. The input variables generally come from sampling distributions like the LHS or any other suitable DOE. Depending on these sampling points, the threshold differs. Hence, the exceedance is affected and therefore the estimated parameters and eventually the fitted GPD. If we have the privilege to generate large number of samples and use a fraction of it as exceedance data to fit a GPD, the estimated parameters in the process can be considered as 'optimal' parameters since they are based on a large sample size. These parameters can be used to estimate probability contents at certain levels and can be compared with the probability content estimated by a GPD with small sample size. This difference will dictate the sampling error contribution to the total error. In an attempt to study the modeling and sampling error, an example problem is treated here. The example under investigation is Y =exp(x) and x~N(0, 1). The modeling error is studied comparing the estimated probability content using a GPD with a large sample size and the exact probability content generated by analytical expression of the lognormal distribution. The parameters for the GPD are obtained by MLE method. Let the conditional CDF P (Y > ut +: / Y > ur) be represented by F4, (z) In this example, it is possible to estimate this quantity exactly. Using extreme value theory for exceedances, F (z) can be approximated by a GP distribution denoted by F ,(z) Let the estimated CDF be denoted by F ,(z)$ Let the GPD which is based on the optimal parameters be denoted by F ~(z)* . le6 samples based on LHS design are generated. The Y's are sorted as: Y(, < Y(2) .. (le6) This is called the order statistics of Y. Choosing the appropriate 'u', the threshold value is very critical. u defines the beginning of the upper tail. The F, is chosen to be 0.90 and the corresponding value of Y is chosen as u. Once u is chosen, the exceedance data is formed and the ML estimates are obtained for the parameters. These estimates are used to estimate the probability content at different levels of Y. The results are presented in Table D1. Five repetitions of le6 samples are used for each level. The mean and standard deviation of the CDF estimates based on GPD are presented. The error refers to the total error in percentage. The sampling error component of the total error is extracted using estimated optimal parameters. The optimal parameters were obtained based on a million samples. Here, 3 percent of the data was considered as exceedance. The results are presented in Table D 2. From Tables D1 and D2, it can be clearly seen that the sampling error is high compared to the modeling error. So, if we can account for the sampling errors, tail modeling has a fairly good capacity to model the tails of the distribution to estimate probability contents. TableD1. Modelling error (F,(z) ,,(z) ). The standard deviation is very low that it is not presented. Exact F, (z) 0.08283 0.05376 0.03659 0.02583 0.01879 0.01400 0.01065 0.00824 0.00648 GPD F,,(z)*A 0.07988 0.05299 0.03647 0.02587 0.01884 0.01403 0.01065 0.00822 0.00645 Error% 3.56 1.43 0.33 0.16 0.28 0.19 0.00 0.24 0.49 TableD2: Sampling error( (F,,(z)* ,;,(z) ) Error 9.04 9.39 9.61 9.15 8.20 7.17 6.53 6.69 7.95 Y GPD le6 samples ,,(z)*A 4 0.07988 5 0.05299 6 0.03647 7 0.02587 8 0.01884 9 0.01403 10 0.01065 11 0.00822 12 0.00645 GPD le2 samples ~(z)S 0.08710 0.05797 0.03997 0.02824 0.02039 0.01503 0.01135 0.00877 0.00696 APPENDIX E ACCURACY AND CONVERGENCE STUDIES FOR NORMAL AND LOGNORMAL DISTRIBUTIONS The quantile of a probability distribution can be estimated using Eq.(48). The quantile2 can be related to the inverse measure in the structural reliability context where the probability content is expressed as the target failure probability. Due to the various advantages of using inverse measures in RBDO, we are interested in estimating them using tail modeling approach. Here, we study the accuracy of the tail model to capture the quantile in the tails of the normal distribution. In the following simulation study a y=x example where x ~N (0, 1) is treated. The error metrics that are monitored are: a) Bias error: Bias(x )=Ex x] (El) 1Ns x~ xP (E2) Nsim = where, x : Estimated quantile x, : Actual quantile Nsim : Number of simulations b) RMSE error: d~ x)2_NS 2 (E3) Here, we set out to study the effects of the threshold value, the number of exceedance data on the accuracy of the quantile estimate. For each threshold value, four different exceedances are tested. Since the exceedances vary, the number of samples used in each case varies. The results are presented in Table E1.The overall capacity of the tail modeling approach to capture the quantiles in the tails of the normal distribution seems to be pretty good. For all the four threshold 2 A p quantile of a random variable X is any value x such that Pr(XI x,) = p values, the % Bias error for the 25 exceedance data seem to be very less compared to higher exceedance data. But, it is to be noted that the corresponding %RMSE errors are high emphasizing the fact that the estimates from low number of exceedance data could be highly unreliable. 25 exceedance data is considered low to estimate quantiles. 50 to 100 data is considered realistic and 200 ideal. It can be clearly observed that the accuracy of the % Bias increases as the threshold increases for the Table E1. Accuracy of the tail modeling approach to estimate quantiles. Normal distribution y = x. x~N(0, 1). 500 simulations. ML estimated parameters Fu Nex NS % Bias % RMSE 0.999 0.9999 0.99999 0.999 0.9999 0.99999 0.8 25 125 8 9 5 48 98 216 50 250 6 10 14 22 34 46 100 500 3 7 11 11 16 22 200 1000 2.5 7.3 12.5 5.1 8.5 12.0 0.85 25 166 6 7 2 37 81 211 50 333 4 8 11 17 28 44 100 666 2 6 10 8 13 18 200 1333 2.1 6.1 10.6 4.1 6.9 9.8 0.9 25 250 3 2 3 25 51 107 50 500 4 7 10 11 18 25 100 1000 2.2 5.7 9.4 5.4 9.3 13.1 200 2000 0.8 3.5 6.8 2.6 4.8 7.1 0.95 25 500 3 3 0 12 26 57 50 1000 2 4 5 6 11 17 100 2000 0.8 2.5 4.5 2.8 5.5 8.5 200 4000 0.4 2.4 5.2 1.3 2.6 4.0 Fu Threshold in the CDF. Nex Number of exceedance data. NS Number of total samples used. 0.999, 0.9999, 0.99999 > Quantiles required to be estimated. For a threshold of 0.85 and Nex=100, the 2,6,10 % Bias error in quantile translates to 27.6, 139.9 and 524.12% error in probability content same number of exceedance data. The threshold values, number of exceedance data is dependent on the level of target quantile we wish to estimate. It can be observed that for a quantile of 0.999, reasonably low quantiles such as 0.8 perform well enough with 50 to 100 exceedance data, if 5% bias error is allowable. For a 0.9999 quantile, a threshold of 0.9 with 100 exceedance data 104 performs fairly well. A 0.95 threshold is capable of estimating a 0.99999 quantile well with about 100 data. But it can be observed that there is oscillation in the error values as the 100 exceedance data seem to perform better than the 200 exceedance data in % bias error. A higher threshold value is expected to rectify this oscillation. Since normal distribution is the most encountered distribution in structural analysis, it is good news that the tail modeling approach can capture the quantiles in the tail of the normal distribution fairly well. The next distribution that is treated here is a log normal distribution which has a heavy tail and is generally used to model loads in structural analysis. The simulation procedure is same as above. The results are provided in Table E2. Comparison of Tables E1 and E2 allows us to conclude Table E2. Accuracy of the tail modeling approach to estimate quantiles. LogNormal distribution y = x. xlogN(0,1). 500 simulations. ML estimated parameters Fu Nex NS % Bias % RMSE 0.999 0.9999 0.99999 0.999 0.9999 0.99999 0.8 25 125 18 110 527 228 1214 8065 50 250 8 42 126 77 199 517 100 500 7 29 79 37 91 229 200 1000 3.8 17.4 44.0 16.6 36.4 75.1 0.85 25 166 18 170 1517 256 2848 40477 50 333 4 37 152 67 270 1410 100 666 6 26 70 28 68 159 200 1333 1.8 12.0 32.9 12.3 26.1 51.1 0.9 25 250 5 71 439 101 632 5683 50 500 6 38 136 44 152 590 100 1000 1.1 14.2 43.7 16.8 40.3 92.4 200 2000 1.5 10.5 29.4 8.6 18.6 36.4 0.95 25 500 2 38 211 41 196 1363 50 1000 1 20 82 19 71 358 100 2000 1.4 12.4 39.0 9.3 24.2 56.2 200 4000 0.1 5.6 19.6 4.2 10.2 20.8 Fu Threshold in the CDF. Nex Number of exceedance data. NS Number of total samples used.0.999, 0.9999, 0.99999 > Quantiles required to be estimated. For a 0.95 threshold and Nex=100, the 1.4, 12.4 and 39% Bias error in quantile translates to 4.5, 37.5 and 78.3% error in probability content that estimating quantiles in lognormal distribution is more challenging than normal distribution. For a 0.999 quantile, 0.8 and 0.85 thresholds perform fairly well with 200 samples. Above a 0.9 threshold, 100 samples seem to render very good results. There is considerable difficulty in estimating 0.9999 quantile with even a 100 exceedance data and a 0.95 threshold value. Needless to mention, trying to estimate 0.99999 quantile deteriorates the accuracy of the estimation only. In order to test whether higher threshold values enable us to estimate the extreme quantiles accurately, simulation is conducted for threshold values ranging from 0.96 to 0.99 and results are presented in Table E3. Table E3. Accuracy of the tail modeling approach to estimate quantiles. Higher thresholds. LogNormal distribution y = x. xlogN(0,1). 500 simulations. ML estimated parameters Fu Nex NS % Bias % RMSE 0.9999 0.99999 0.9999 0.99999 0.96 25 625 50 254 171 958 50 1250 16 61 46 160 100 2500 9 30 18 42 200 5000 4 16 8 16 0.97 25 834 20 135 102 693 50 1667 12 49 35 98 100 3334 7 29 16 43 200 6667 3 13 6 12 0.98 25 1250 17 126 74 563 50 2500 6 29 19 49 100 5000 4 20 10 22 200 10000 1 8 4 8 0.99 25 2500 5 65 30 225 50 5000 2 19 10 25 100 10000 0.9 10 5 12 200 20000 0.2 6 2 5 Fu Threshold in the CDF. Nex Number of exceedance data. NS Number of total samples used. 0.9999, 0.99999 > Quantiles required to be estimated Results from Table 43 clearly show that as the threshold is increased, the extreme quantiles are also captured fairly well. With a 0.99 threshold and 100 exceedance data, we can achieve less 106 than one percent error in bias for estimating 0.9999 quantile. This requires 10000 samples. Similarly the accuracy in estimating 0.99999 quantile is also better with 0.99 threshold and 200 exceedance data. A more accurate estimate can be expected if we raise the threshold to 0.999 or higher. APPENDIX F ACCURACY AND CONVERGENCE STUDIES FOR THE CANTILEVER BEAM Stress Failure Mode The stress failure mode for the cantilever beam discussed in Chapter 3 is considered here. The optimal design variables w=2.4526; t3.8884 for a target failure probability of 0.00125 (le7 Samples) are adopted from (Qu and Haftka, 2004). The corresponding value of PSF for this design is 1. For the optimal combination of the design variables, 500 samples based on the distribution of the random variables are generated and the PSF is estimated for different thresholds as discussed in Chapter 3. This procedure is repeated 100 times and the mean, 5%, 95% confidence intervals are estimated. The results are presented in Figure E1. It can be observed from the figure that the PSF estimated by regression approach is very unstable whereas the median of PSF estimated by the ML approach is relatively stable. There is instability towards higher threshold like 0.98 but the corresponding exceedance data are very less, which explains the instability. The estimated mean of the PSF in the ML approach is close to 0.99 which is about 1% bias error in estimating the actual value of the PSF. The standard deviation of the PSF estimate at a threshold of 0.9 with ML approach is about 0.003 which is 0.3% of the estimate. The corresponding estimate of failure probability is 0.001 1 with a standard deviation of 0.0001 which is about 10% of the estimate. Deflection Failure Mode The allowable deflection is chosen to be Do=2.25. Similar to the stress failure case, the PSF is estimated at different thresholds for the optimal combination of the design variables. The results are presented in Figure E2. At a threshold of 0.9 and ML approach, the PSF is estimated to be 0.99 with a standard deviation of 0.004. The corresponding failure probability estimate is 0.0013 with a standard deviation of 0.0001 which is about 8% of the estimate itself. ;F~ C 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 Fu Nex 50 40 30 ,400 90 80 70 60 20 10 0 MM~an xMedian  i ii 1.0 1.06 1.024 1.012 0.976 0.9 0.8 Nex Figure F1. Convergence of PSF at different thresholds. A) MLE B) Regression. Cantilever beam stress failure mode. 500 Samples. 100 repetitions 109 1.0 0 90 80 70 60 50 40 30 20 10       M e a n 1edian ~~i . 1.0461 1.032 1.018 1.004 L. 0.976 0.962 0.948 0.9341 0'~.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 Fu 1.1 0 90 80 70 60 50 40 30 20 10 4LCI UUCI Mean xMedian L + i : LI. S1.15 .5 1 .O 0.95 0. 0.2 084 .86 0.8 0.9 0.92 0.94 0.96 0.98 Fu B Figure F2. Convergence of PSF at different thresholds. A) MLE B) Regression. Cantilever beam deflection failure mode. 500 Samples. 100 repetition 110 APPENDIX G STATISTICAL DISTRIBUTIONS ERROR PLOTS Boxplot comparison of eind and eMTM at different reliability indices. ML=Maximum Likelihood, Rg Regression, QT Quadratic fit to the Tail data between S, and In(P ), LT  Linear fit to the Tail data between S, and P QH Quadratic fit to half of the data between S, and ; .Dashed line inside the box depicts the mean and the solid line, the median. 1000 repetition 100 p= 3 looom = 0.029 0.006 0.034 0.012 0.021 1001n= 0.030 0.008 0.047 0.015 0.028 1000 1= 3.6 looom = 0.083 0 .025 0.034 0.056 0.019 10001nd= 0.082 0.027 0.058 0.052 0.032 1000 1= 4.2 looom = 0.177 0.081 0.079 0.145 0.027 100Pin=0.175 0 .080 0.000 0.135 0.041 + + ML Rg OT LT H oom. = 0.008 100pIn = 0.016 MTM I~ I MTM ML Rg OT LT OH oom. = 0.032 100Pin = 0.039 MTM MTM ML Rg Q1T LT Q1H 100 m = 0.090 MTM + I MTM 0.5 0.4 o 0.3 0.2 0.1 0 0.25 0.2 S0.15 0.1 0.05 0 0.3 0.2 0.1 0 0.15 S0.1 0.05 0 Figure G1 Lognormal distribution 1.2 1 0.B ei 0.5 0.4 0.2 0 0.4 0.2 e 0. 0.1 0 ML RgI QT LT 1H lom. = 0.009 100PIn = 0.015 MTM I MTM ML RgI QT LT H 1000 nd=006 m.~~ =.1 MTM ML RgI QT LT QH oom. = 0.047 In MTM 100 p= 3 looom = 0.025 0.005 0.030 0.008 0.024 100Pn= 0.026 0.006 0.040 0.013 0.029 1000 1 = 3.5 oom. = 0.068 0.027 0.029 0.016 0.037 1001n=0.06B 0.028 0.052 0.01B 0.046 ind 1000 = 4.2 oom. = 0.137 0.078 0.040 0.042 0.053 1000n=0.134 0.076 0.066 0.039 0.066 0.8 0.6 i' 0.4 0.2 0 0.3 E 0.2 0.1 0.4 0.3 dC0.2 0.1 0 0.25 0.2 0.15 e 01 0.05 0 0.2 1 m 0.15 S0.1 0.05 0.5 0.15 Figure G2: Gamma Distribution I I~I i~ i I i ML RgI QT LT 1H looom = 0.076 100 n = 0.132 MTM MTM + II ML RgI Q;T LT H monom = 0.291 1000 Id=032 MTM MTM + ML RgI QT LT Q~H looom = 0.720 100 n = 0.692 MTM MTM lomi = 0.249 0.163 0.356 0.510 0.076 1000 = 0.255 0.159 0.435 0.544 0.106 ind p = 3.6 13 = 4.2 oom. = 0.652 0.615 0.602 1.020 0.270 lomi = 1.203 1.245 0.B69 1.654 0.679 looo = 0.611 0.606 0.737 1.068 0.256 1000 = 1.10B 1.232 1.063 1.717 0.619 ind ind 4 3 r2 e1 2.5 2 S1.5 S1 2 1.5 o0.5 0 1.5 I 1 o 0.5 0 0.5 0 1 0.8 S0.6 0.2 Figure G3. Extreme Value type 1 Distribution ML Rg QT LT QH looom = 0.000 100 Ind = 0.000 x 10 MT MTM ML Rg Q~T LTQH looom = 0.001 4 100 n = 0.001 x 10 MTM MTM ML Rg Q~T LT Q~H looo = 0.001 4 100 n = 0.001 x 10 T 0  MTM 1000 1= 3 1001 = 3.6 looom = 0.000 0.000 0.000 0.029 0.037 oomind = 0.001 0.000 0.005 0.058 0.118 loo = 0.001 0.000 0.000 0.029 0.037 100in= 0.001 0.000 0.005 0.058 0.118 x 10' ,x10 3 8 1000 p = 4.2 oom. = 0.001 0.000 0.014 0.087 0.245 100 d=0.001s 0.000 0.014 0.087 0.245 xn X10. 2 10 5 0 1l e4 2 0 1 o Figure G4. Uniform Distribution. ML RgI QT LT lH oom. = 0.015 In MTM i~ 1 B~ B ML RgI QT LTQH oom. = 0.016 100 n = 0.024 MTM ML RgI QT LT Q~H looom = 0.018 1000 n=0.3 MTM=.3 loom = 0.022 0 .005 0.028 0.015 0.027 ind looo = 0.023 0.006 0.036 0.018 0.031 I . , 1000 = 3.6 loom = 0.059 0.027 0.036 0.016 0.049 ind looop= 0.057 0.027 0.051 0.021 0.055 1000 1= 4.2 looom = 0.117 0.072 0.037 0.015 0.078 Ind0 looop, = 0.110 0.072 0.062 0.022 0.088 i i 0.3 0.25 ei 0 1 0.05 0 0.15 0.05 0 0.5 0.4 0.1 0.15 0.05 0.1 0.08 E 0.05 o 0.04 0.02 0 0.25 0.2 ~0.15 e) 0.05 MTM MTM Figure G5. Rayleigh Distribution. B j ML RgI Q;T LTQH oom. = 0.025 1000 Id=003 MTM 003 I . MTM ML RgI Q;T LTQH looom = 0.040 1000 Id=002 MTM MTM +i t +3 ML Rg Q LT Q oom. = 0.109 100 n = 0.120 MTM + MTM 1000 1= 4.2 looom. d = 0.209 0 .072 0.107 0.229 0.081 1000 =0.241 0 .087 0.132 0.214 0.106 100p = 3 looom = 0.040 0.025 0.056 0.023 0 .042 100 d=0.044 0.026 0.076 0.025 0 .052 1000 1 = 3.6 looom.n = 0.104 0.043 0.054 0 .098 0.061 1000 =0.115 0.049 0.093 0 .090 0.07B 2.5 2 S1.5 1 0.5 0.8 0.6 S0.4 0.2 0 0.6 0.5 0.4 d'0.2 0.0 0.3 0.25 0.2 S0.15 o 0.05 0 = 0.6 0.4 0.0 0.5 0.4 S0.3 0.2 0.1 0 Figure G6. Exponential Distribution. Boxplot of r ratio. Different distributions oom =0.349 1000 = 0.340 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 MTM Figure G7. Lognormal Distribution. 1000m .8 looop = 0.220 0.6 0.5 0.4 0.3 0.2 0.1 MTM Figure G8. Gamma Distribution. oom =0.183 looo = 0.185 0.55 0.5 0.45~ 0.4 0.35 S0.3 0.25 0.1 0.05 MTM Figure G9. Extreme ValueType 1 Distribution oom =0.003 x 10 3 looo 0.003 6 5 4 3 2 0 MTM Figure G10. Uniform Distribution. 119 oom =0.146 10004= 0.185 0.4 0.2  MTM Figure G11. Rayleigh Distribution. 1000m .4 1000 0.276 0.7 0.6 0.5 0.4 0.3 0.2 0.1 MTM Figure G12. Exponential Distribution. 120 REFERENCES Agarwal, H.; Lee, J.C.; Watson, L.T.; Renaud, J.E. 2004: A unilevel method for reliability based de sign opti mi zati on, Proceedings of the 45th A1AA ASME 'ASCE AHS/ ASC Structures, Structural Dynamics, and Material Conference, Palm Springs, CA, April 1922, AIAA Paper 20042029. Agarwal H., Shawn E.G., Renaud J.E., Perez V.M., Watson. L.T. 2007: Homotopy Methods for Constraint Relaxation in Unilevel Reliabilitybased Design Optimization, Submitted to the ASM\E Journal of2~echanical Design Bassi, F.; Embrechts, P.; Kafetzaki, M. 1998: Risk Management and Quantile Estimation In: A Practical Guide to Heavy Tails, Edited by Adler, R. J., et al., Boston, Birkhaeuser, pp. 111130 Beirlant, J.; Goegebeur, Y.; Segers, J.; Teugels, J. 2004: Statistics ofExtrenzes: Theory and Applications, John Wiley & Sons Beirlant, J.; Vynckier, P.; Teugels, J.L. 1996: Excess Functions and Estimation of Extreme Value Index. Bernoulli, 2, pp:293318 Boos, D.D. 1984: Using Extreme Value Theory to Estimate Large Percentiles, Technonsetrics, 26, pp:3339 Caers, J.; Maes, M.A. 1998: Identifying Tails, Bounds and Endpoints of Random Variables. Structural Safety, Vol 20, pp 123 Castillo, E.; Hadi, A.S.; Balakrishnan, N.; Sarabia, J.M. 2005: Extreme Vahue and Related Models n ithr Applications in Engineering and Science, Hoboken, New Jersey:Wiley Interscience Chen, S.; Nikolaidis, E.; Cudney, H. H. 1999: Comparison of Probabilistic and Fuzzy Set Methods for Designing under Uncertainty. Proceedings, AIAA ASME 'ASCE AHS ASC Structures, Structural dynamics, andMaterials Conference and Exhibit, 28602874. Chiralaksanakul, A.; Mahadevan, S. 2005: Multidisciplinary Design Optimization Under Uncertainty, in press, Optimization and Engineering. Coles, S. 2001: An Introduction to Sta~tistical M'odeling of Extrenze Values, London, England: SpringerVerlag Du, X.; Chen, W. 2004: Sequential Optimization and Reliability Assessment Method for Efficient Probabilistic Design," Journal of2~echanical Design, Vol 126, No. 2, pp. 225 233. Du, X.; Sudjianto, A.; and Chen, W. 2004: An Integrated Framework for Optimization under Uncertainty Using Inverse Reliability Strategy, ASM~E Journal of2~echanical Design, Vol. 126, No.4, pp. 561764. Elishakoff, I. 2001: hIterrelation between Safety Factors and Reliability, NASA report CR2001 211309 Elishakoff, I. 2004: Safety Factors and Reliability: Friends or Foes? Kluwer Academic Publishers, Dordrecht, The Netherlands Frangopol, D.M. 1985: Structural Optimization Using Reliability Concepts. Journal of Structural Engineering. Vol. 111, no. 11, pp. 22882301 Frangopol, D. M.; Maute, K. 2003: Lifecycle Reliabilitybased Optimization of Civil and Aerospace Structures, Computers and Structures, Vol. 81, No. 7, pp. 397410. Glasserman, P. 2004: M~onte Calrlo M~ethods in Financial Engmneering. Springer Verlag Goel, T., Haftka, R.T., Shyy, W., Queipo, N.V.(2006) "Ensemble of Multiple Surrogates", Structural and M~ultidisciplinary Optimization, 33 (3) Gu, X.; Renaud, J.E.; Batill, S.M.; Brach, R.M.; Budhiraja, A.S. 2000: Worst Case Propagated Uncertainty of Multidisciplinary Systems in Robust Design Optimization. Structural and M\~ultidisciplinary Optimization, 20(3), 190213. Gurdal, Z.; Haftka, R.T.; Haj ela, P. 1988: Design and Optimization ofLamninated Composite Structures. New York: Wiley Hasofer, A.M. 1996: Parametric Estimation of Failure Probabilities. In: Mathematical Models for Structural Reliability Analysis. Edited by Casicati, F.; Roberts, B. Boca Raton, FL: CRC Press Hosking, J.R.M.; Wallis, J.R. 1987: Parameter and Quantile Estimation for the Generalized Pareto Distribution. Technometrics, Vol 29, no.3., pp. 339349 Kharmanda. G and Olhoff N, 2007: Extension of Optimum Safety Factor Method to Nonlinear Reliabilitybased Design Optimization, Journal of Structural and M~ultidisciplinaryy Optimization, Vol. 34, No.5, pp. 367380. Kharmanda. G., Olhoff N.,ElHami .A. 2004: Optimum Values of Structural Safety Factors for a Predefined Reliability Level with Extension to Multiple Limit States, Journal ofStructural and M\~ultidisciplinary Optimization, Vol. 26, No.6, pp. 421434. Kharmanda. G., Mohamed A., Lemaire M. 2002: Efficient Reliabilitybased Design Optimization Using Hybrid Space with Application to Finite Element Analysis, Journal of Structural and M~ultidisciplinary Optimization, Vol. 24, No.3, pp. 233245. KirjnerNeto, C.; Polak, E.; Kiureghian, A. D. 1998: An Outer Approximations Approach to Reliabilitybased Optimal Design of Structures, Journal of Optimization Thzeory and Applications, Vol. 98, No. 1, pp 116. Kiureghian, A. D.; Zhang, Y.; Li, C. C. 1994: Inverse Reliability Problem, Journal of Engineering M~echanics, Vol. 120, No. 5, pp. 1 1541 159. Kuschel, N.; and Rackwitz, R. 2000: A new approach for structural optimization of series systems, Applications of Statistics and Probability/,Vol.2, No 8, pp987994 Lee, J. O.; Yang, Y. S.; Ruy, W. S. 2002: A Comparative Study on Reliabilityindex and Target performancebased Probabilistic Structural Design Optimization, Computers and' Structures, Vol. 80, No. 34, pp. 257269. Li, H.; Foschi, O. 1998: An Inverse Reliability Measure and Its Application, Structural Safety, Vol. 20, No. 3, pp. 257270. Liu. D., Choi. K. K., Youn .B. D., Gorsich. D. 2006: Possibilitybased Design Optimization Method for Design Problems with Both Statistical and Fuzzy Input Data, Journal of2~echanical Design, Vol 128, pp. 928935. Liu. D., Choi. K. K., Youn .B. D. 2006: An Inverse Possibility Analysis Method for Possibility Based Design Optimization, A1AA Journal, Vol 44, No. 11, pp. 26822689 Madsen, H. O.; Krenk, S.; Lind, N. C. 1986: M~ethods of Structural Safety, NJ: PrenticeHall Maes, M.A.; Breitung, K. 1993: ReliabilityBased Tail Estimation. Proceedings, IUTAM\~ Symposium on Probabilistic Structural M\~echanics (Advances in Structural Reliability Methods), San Antonio, Texas, 335346 Maglaras, G.; Nikolaidis, E.; Haftka, R.T.; Cudney, H.H. 1997: AnalyticalExperimental Comparison of Probabilistic methods and Fuzzy Set based Methods for Designing under Uncertainty, Structural Optimization, 13(2/3), pp. 6980. Melchers, R. E. 1999: Structural Reliability Analysis and Pred'iction, New York: Wiley Metropolis, N. 1987: The Beginning of Monte Carlo Method, Los Alamnos Science, Special Issue dedicated to Stanislaw Ulam: 125130 Moore, G.E. 1965: Cramming More Components onto Integrated Circuits, Electronics, 38, (8). Mourelatos. Z.P., and Zhou, J. 2005: Reliability Estimation and Design with Insufficient Data Based on Possibility Theory, AIAA Journal, Vol 43, No. 8, pp 16961705 Myung J.I. 2003: Tutorial on Maximum Likelihood Estimation, Journal of2athematical Psychology, 47, 90100 Phadke M. S. 1989: Quality Engineering Using Robust Design, Englewood Cliffs, NJ: Prentice Hall Pickands, J. 1975: Statistical Inference Using Extreme Order Statistics. Annals ofStatistics. Vol 3,119131 Qu, X.; Haftka, R. T. 2004: Reliabilitybased Design Optimization Using Probabilistic Sufficiency Factor, Journal ofStructural and M~ultidisciplinary Optimization, Vol. 27, No.5, pp. 314325. Qu, X.; Haftka, R. T., Venkataraman, S., Johnson, T.F.2003: Deterministic and ReliabilityBased Optimization of Composite Laminates for Cryogenic Environments, AIAA Journal, Vol 41, No. 10, pp 20292036 Qu, X.; Singer, T.; Haftka, R. T. 2004:Reliabilitybased Global Optimization of Stiffened Panels Using Probabilistic Sufficiency Factor, Proceedings of the 45th AIAA/ASM\~E/ASCE/AHS/ASC Structures, Structural Dynamics, and Material Conference, Palm Springs, CA, April 19 22,AIAA Paper 20041898. Qu, .X.,Venkataraman, S., Haftka, R T., Johnson, T.F. 2001: Reliability, Weight and Cost Tradeoffs in the Design of Composite Laminates for Cryogenic Environments, AIAA Paper 20011327, April 2001 Rackwitz, R.2000: Reliability Analysis Past, Present and Future. 8th ASCE Speciality Conference on Probabilsitic M~echanics and Structural Reliability, 2426 July, Notre Dame, INT, USA. Ramu, P.; Qu, X.; Youn, B.; Haftka, R.T.; Choi, K.K. 2006: Safety Factors and Inverse Reliability Measures, Int. J. Reliability and Safety, Vol. 1, Nos. 1/2 Rosenblatt, M. 1952: Remarks on a Multivariate Transformation, The Annals of2athematical Statistics, 23(3), 470 472 Royset, J.O.; Kiureghian, A. D.; Polak, E. 2001: Reliabilitybased Optimal Structural Design by the Decoupling Approach," Reliability Engineering and System SafetyVol 73, No. 3, 2001, pp.213221 Sakawa, M. 1993: Fuzzy sets and Interactive M~ultiobjective Optimization. Plenum press Smarslok, B. P.; Haftka, R. T.; Kim, N.H. 2006: Taking Advantage of Separable Limit States in Sampling Procedures. 47th A1AA/ASM\~E/ASCE/AHS/ASC Structures, Structural Dynamnics, and Materials Conference, AIAA 20061632, Newport, RI, May, Soundappan, P.; Nikolaidis, E.; Haftka, R.T.; Grandhi, R.; Canfield, R. 2004: Comparison of evidence theory and Bayesian Theory for Uncertainty Modeling, Reliability Engineering & System Safety, 85 (13): 295311 Tu, J.; Choi, K. K.; Park, Y. H. 1999: A New Study on Reliability Based Design Optimization, Journal of2\~echanical Design, ASME, Vol 121, No. 4, 557564 Wu, Y T.; Shin Y.; Sues, R.; Cesare, M. 2001: Safety Factor Based Approach for Probability based Design Optimization, Proceedings of 42nd AIAA/ASM\~E/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, Seattle, WA, AIAA Paper 20011522. Youn B. D.; Choi K. K.; Du. L. 2005: Enriched Performance Measure Approach (PMA +) for Reliabilitybased Design Optimization," AIAA Journal, Vol. 43, No. 4, pp. 874884. Youn, B. D.; Choi, K. K.; Park, Y. H. 2003: Hybrid Analysis Method for Reliabilitybased Design Optimization, Journal of2\~echanicalDesign, ASM~E, Vol. 125, No. 2, pp. 221232. BIOGRAPHICAL SKETCH Palani Ramu was born in Chennai, India in 1978. He earned his BEng in Mechanical from Madurai Kamaraj University, India in 1999. He worked at the Indian Institute of Technology, Bombay and InfoTech Pratt & Whitney Center of Excellence as Research Assistant and Design Engineer respectively till 2002. In 2003, he started his graduate studies at the University of Florida, Gainesville and worked towards a Ph.D in Aerospace Engineering. During his graduate studies, he was a visiting researcher at the Leonardo Da Vinci University, Paris, France and was awarded a summer fellowship at the Center for Space Nuclear Research Idaho National Labs, Idaho. His areas of interest include application of probability and statistical methods to design aerospace structures, specifically highly reliable designs. 126 PAGE 1 1 MULTIPLE TAIL MODELS INCLUDING IN VERSE MEASURES FOR STRUCTURAL DESIGN UNDER UNCERTAINTIES By PALANIAPPAN RAMU A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2007 PAGE 2 2 2007 Palaniappan Ramu PAGE 3 3 Dedicated to my parents Meenakshi and Ramu PAGE 4 4 ACKNOWLEDGMENTS This dissertation would not have been possibl e if not for the help, support and motivation by my teachers, family and friends. I am extrem ely thankful to my advisors Dr. Nam Ho Kim and Dr. Raphael T. Haftka for, providing me th is wonderful opportunity to pursue a Ph.D, their constant encouragement, patience and excellent guidance. I learnt a lot both in academics and personal life from them. They have been more than mere advisors friends, philosophers and mentors. I have always admired their indepth understanding and knowledge and feel that I am fortunate to have worked under their guidance. I would like to thank Dr. Pe ter Ifju, Dr. Tony Schmitz and Dr. Stanislav Uryasev for agreeing to serve on my committee, reviewing my dissertation and for providing constructive criticism that helped to enhance the quality of this work. I would like to thank Dr.Choi, Dr.Missoum, Dr.Qu, and Dr.Youn, fo r collaborating with me. I al so thank the staff of the Mechanical and Aerospace Engineering departme nt for their help with administrative support Thanks are due to my former and present coll eagues in the Structural and Multidisciplinary Optimization group. I thank all my friends, espe cially Gators for Asha group members, for making my life outside work wonderful and enjo yable. I greatly appreciate the financial support that I received from the Instit ute for Future Space Transport Last, but not the least, I would not have comp leted this work if not for the unconditional love, emotional understanding and support of my family. To you, I dedicate this work! PAGE 5 5 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF TABLES................................................................................................................. ..........7 LIST OF FIGURES................................................................................................................ .........9 NOMENCLATURE................................................................................................................... ...11 ABSTRACT....................................................................................................................... ............13 CHAPTER 1 INTRODUCTION..................................................................................................................14 2 RELIABILITY BASED DESIGN OPTIMIZATION............................................................18 Introduction................................................................................................................... ..........18 Standard Formulations.......................................................................................................... ..18 MomentBased Methods.................................................................................................21 Monte Carlo Simulation..................................................................................................22 3 INVERSE RELIABILITY MEASURES...............................................................................26 Literature Review.............................................................................................................. .....26 Birger Safety Factor........................................................................................................... .....31 Probabilistic Sufficiency Factor.............................................................................................31 Probabilistic Performance Measure........................................................................................33 Inverse Measure Calculation..................................................................................................35 Simulation ApproachMonte Carlo Simulation..............................................................35 Analytical ApproachMomentbased Methods..............................................................37 Reliability based Design Optimization with Inverse Measures...........................................39 Beam Design Example...........................................................................................................39 Design for Stress Constraint............................................................................................40 Comparison of Inverse Measures....................................................................................41 Use of PSF in Estimating the Required Change in Weight to Achieve a Safe Design..........42 System Reliability Estimation Using PSF and MCS..............................................................43 Design for System Reliability by MCS and PSF....................................................................44 4 TAIL MODELLING AND RELIABILITY ANALYSIS......................................................49 Tail Equivalence and Generalized Extreme Value Theory....................................................49 Generalized Pareto Distribution.............................................................................................50 Threshold Selection............................................................................................................ ....52 Parameter Es timation........................................................................................................... ...54 PAGE 6 6 Maximum Likelihood Estimation....................................................................................55 Least Square Regression..................................................................................................56 Accuracy and Convergence Studies for the Quantile Estimation...........................................56 Cantilever Beam Example......................................................................................................57 Tuned MassDamper Example...............................................................................................59 Alternate Tail Extrapolation Schemes....................................................................................61 5 MULTIPLE TAIL MODELS.................................................................................................71 Multiple Tail Model............................................................................................................ ....71 Numerical Examples............................................................................................................. ..72 Standard Statistical Distributions....................................................................................72 Engineering Examples.....................................................................................................76 Application of multiple tail models for reli ability estimation of a cantilever beam........76 Application of multiple tail models for reli ability estimation of a composite panel.......77 CONCLUSIONS.................................................................................................................... ........93 APPENDIX A PSF ESTIMATION FOR A SYSTEM RELIABILITY.........................................................94 B GENERALIZED EXTREME VALUE THEORY.................................................................96 C MAXIMUM LIKELIHOOD ESTIMATION.........................................................................98 D ERROR ANALYSIS FOR GPD..........................................................................................100 E ACCURACY AND CONVERGENCE STUDIES FOR NORMAL AND LOGNORMAL DISTRIBUTIONS......................................................................................103 F ACCURACY AND CONVERGENCE STUD IES FOR THE CANTILEVER BEAM......108 Stress Failure Mode............................................................................................................ ..108 Deflection Failure Mode.......................................................................................................108 G STATISTICAL DISTRIBU TIONS ERROR PLOTS..........................................................111 REFERENCES..................................................................................................................... .......121 BIOGRAPHICAL SKETCH.......................................................................................................126 PAGE 7 7 LIST OF TABLES Table page 31 Different approaches to pres cribe the probabilis tic constraint..........................................39 32 Random variables for beam problem.................................................................................40 33 Comparison of optimum desi gns for the stress constraint.................................................41 34 Comparison of inverse measures for w =2.4526 and t =3.8884 (stress constraint).............42 35 Comparison of MPP obtaine d from calculation of PSF.....................................................42 36 Design for System Reliability by PSF, Qu and Haftka (2004)**......................................45 52 Summary of the performance of i ndividual methods and MTM for different distributions.................................................................................................................. ......75 53 Summary of the error to range ratio for different distributions.........................................75 54 PSF reciprocal estimates and standard de viation at different reliability indices...............77 55 Mechanical properties of the composite laminates............................................................80 56 Coefficient of variation fo r random material properties....................................................80 57 Mean of random variables.................................................................................................80 58 Deterministic optima found by Qu et al (2003).................................................................80 59 Composite laminate. rS reciprocal estimates and sta ndard deviation at different reliability indices............................................................................................................ ....80 A1 Distribution of random variables. Multiple failure modes example..................................95 A2 Contribution of failure modes to the system failure..........................................................95 A3 PSF at different target failure probabilities........................................................................95 D1 Modelling error (()gFz * () Fz)..................................................................................102 D2 Sampling error (* () Fz$ () Fz)................................................................................102 E1 Accuracy of the tail modeling approach to estimate quantiles. Normal distribution.......104 PAGE 8 8 E2 Accuracy of the tail modeling appr oach to estimate quantiles. LogNormal distribution................................................................................................................... ....105 E3 Accuracy of the tail modeling approach to estimate quantiles. Higher thresholds. LogNormal distribution...................................................................................................106 PAGE 9 9 LIST OF FIGURES Figure page 21 Reliability analysis and MPP. S ubscript refers to iteration number..................................25 31 Schematic probability density of the safety factor S.........................................................46 32 Schematic probability density of the limit state function..................................................46 33 Illustration of the calculation of PPM with Monte Carlo Simulation for the linear performance function.........................................................................................................47 34 Inverse reliability analysis and M PP for target failure probability 0.00135 ( = 3)........47 35 Cantilever beam subjected to horizontal and vertical loads...............................................47 41 Tail modeling of F (u) using the threshold u. The region of y >0 is failure........................64 42 Convergence of PSF at different threshol ds. A) MLE B)Regressi on. Cantilever beam system failure mode. 500 Samples. 100 repetitions...........................................................65 43 Tuned vibration absorber...................................................................................................66 44 Normalized amplitude vs r1 and r2.....................................................................................66 45 Contour of the normalized amplitude................................................................................67 46 Convergence of PSF at different thresholds. Tuned Mass Damper...................................68 47 Tuned Mass Damper. Comparison of tail models using regression and ML method........69 48 Transformation of the CDF of PSF reciprocal ( Sr). A) CDF of Sr. B) Inverse Standard normal cumulative distributi on function applied to the CDF. C ) Logarithmic transformation applied to the reliability index..............................................70 51 Lognormal distribution. Classical tail modeling techniques MLE and Regression........81 52 Lognormal distribution. Linear fit to tail (LT) and Quadra tic fit to half of the data (QH)........................................................................................................................... ........82 53 Lognormal distribution. Quad ratic fit to the tail (QT).......................................................83 54 Lognormal distribution. Multiple tail models....................................................................84 55 Lognormal distribution. Multiple tail modelsextrapolation region..................................85 56 Normal distribution. 1000 repe titions. Boxplot comparison of eind and eMTM at different reliability indices.................................................................................................86 PAGE 10 10 57 Normal distribution. Boxplot of ratio.............................................................................87 58 Cantilever Beam. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices............................................................................................................ ....88 59 Normal QQ plot of the residua ls for the normal distribution.............................................89 510 Cantilever beam. Boxplot of ratio...................................................................................90 511 Geometry and loading for the compos ite laminate. Xhoop direction, Yaxial direction...................................................................................................................... .......90 512 Composite laminate of cryogenic tank 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices............................................................................91 513 Composite laminate of cryogenic tank. Boxplot of ratio................................................92 A1 Two variable system. Algebraic mu ltiple limit state functions example...........................95 F1 Convergence of PSF at different thres holds. A) MLE B) Regression. Cantilever beam stress failure mode. 500 Samples. 100 repetitions.................................................109 F2 Convergence of PSF at different thres holds. A) MLE B) Regression. Cantilever beam deflection failure mode. 500 Samples. 100 repetition............................................110 G1 Lognormal distribution....................................................................................................112 G2 Gamma Distribution.........................................................................................................113 G3 Extreme Value type 1 Distribution..................................................................................114 G4 Uniform Distribution.......................................................................................................115 G5 Rayleigh Distribution...................................................................................................... .116 G6 Exponential Distribution..................................................................................................117 G7 Lognormal Distribution...................................................................................................118 G8 Gamma Distribution.........................................................................................................118 G9 Extreme ValueType 1 Distribution.................................................................................119 G10 Uniform Distribution...................................................................................................... .119 G11 Rayleigh Distribution..................................................................................................... ..120 G12 Exponential Distribution..................................................................................................120 PAGE 11 11 NOMENCLATURE A Cross sectional area C Capacity C Scaling factor CDF Cumulative Density Function DUU Design Under Uncertainties FORM First Order Reliability Method G Limit state function GPD Generalized pareto distribution M Number of failure modes MCS Monte Carlo Simulation MPP Most Probable Point MPPIR MPP Inverse Reliability MPTP Minimum Performance Target Point MSE Mean square error PMA Performance Measure Approach PPM Probabilistic Performance Measure PSF Probabilistic Sufficiency Factor R Response RBDO Reliability based Design Optimization RIA Reliability Index Approach RSA Response Surface Approximation U Variables in standard normal space u Threshold value t Thickness PAGE 12 12 w Width X Load in x direction Y Load in Y direction Reliability Index Standard Normal Cumulative Density Function Scale parameter Shape parameter z Exceedance data Fu(z) Conditional CDF Fz Conditional CDF approximated by GPD target fP Target probability of failure PAGE 13 13 Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy MULTIPLE TAIL MODELS INCLUDING IN VERSE MEASURES FOR STRUCTURAL DESIGN UNDER UNCERTAINTIES By Palaniappan Ramu December 2007 Chair: Nam Ho Kim Cochair: Raphael T. Haftka Major: Aerospace Engineering Samplingbased reliability estimation with expensive computer models may be computationally prohibitive due to a large numbe r of required simulations. One way to alleviate the computational expense is to extrapolate re liability estimates from observed levels to unobserved levels. Classical tail modeling techniques provide a cla ss of models to enable this extrapolation using asymptotic theory by approximating the tail region of the cumulative distribution function (CDF). This work proposes three alternate tail ex trapolation techniques including inverse measures that can complement classical tail modeling. The proposed approach, multiple tail models, applies the two classical and three alternate extrapolation techniques simultaneously to estimate inverse measures at th e extrapolation regions and use the median as the best estimate. It is observed that the rang e of the five estimates can be used as a good approximation of the error associated with the median estimate. Accuracy and computational efficiency are competing factors in selecting samp le size. Yet, as our num erical studies reveal, the accuracy lost to the reduction of computationa l power is very small in the proposed method. The method is demonstrated on standard statis tical distributions a nd complex engineering examples. PAGE 14 14 CHAPTER 1 INTRODUCTION Uncertainty is an acknowledged phenomenon in the process of structural design. In an optimization framework design under uncertainties re fers to a safe design that not only needs to be optimal but should also warrant survivability against uncertainties. Traditionally safety factors were used to account for the uncertainties. Howeve r, use of safety factors does not usually lead to minimum cost designs for a given level of sa fety because different structural members or different failure modes require different safety factors. Alternately, probabilistic approaches offer techniques to characterize the unc ertainties using a statistical me thod and have the potential to provide safer designs at a given cost. However, the probabilis tic approaches require solving an expensive, complex optimization problem that needs robust formulations and efficient computational techniques for stab le and accelerated convergence. In addition, they also require statistical information which may be expensive to obtain. Structural reliability analysis requires the assessment of the performance function which dictates the behavior of the stru cture. The performance function is called the limit state function which is typically expressed as the difference betw een the capacity (e.g, yield strength, allowable vibration level) and the response of the system (e.g, stress, actual vibration). Approaches available for reliability assessment and analysis can be widely classified as analytical and simulation approaches. Analytical approaches are simple to impl ement but are mostly limited to single failure modes, whereas simulation me thods like Monte Carlo simulation (MCS) are computationally intensive but can handle multiple failure modes. Moreover, they can handle any type of limit state functions unlik e analytical approaches which ar e mostly appropriate for linear limit state functions. Most real life applications exhibit multiple failure modes and the limit state function is not available explicit ly in closed form. Since there is no information on nonlinearity PAGE 15 15 of the limit state function, MCS is the obvious ch oice in such situations In reliabilitybased design optimization, a widely used probabilistic optimization approach (Rackwitz, 2000) in structural engineering, reliability analysis is an iterative process and using crude MCS is computationally prohibitive. Researchers develo p variants of MCS or other approximation methods like response surface that replaces the re liability analysis and obviate the need to repeatedly access the expensive computer models In reliability based design context, high re liability, typical of aerospace applications translates to small probability in the tails of the statistical distributions. Reliability analysis when dealing with high reliability (or low failure proba bility) designs is mostly dependent on how the tails of the random variables behave. In few cases the safety levels can vary by an order of magnitude with slight modifications in the ta ils of the response va riables (Caers and Maes, 1998). Therefore, the tails need to be modeled accurately. Limitations in computational power prevent us in employing direct simulation methods to model the tails. On e way to alleviate the computational expense is to extrapolate into high reliability levels with limited data at lower reliability levels. Statistical techniques from ex treme value theory referred to as classical tail modeling techniques here are avai lable to perform this extrapolation. They are based on the concept of approximating the tail portion of the cumulative dist ribution function (CDF) with a generalized pareto distribution (GPD). In structural engineering, reliability is measur ed by quantities like probability of failure or reliability index. Recently, alte rnate safety measures like the inverse measures in performance space have cornered enough interest because of their multifaceted advantages (Ramu et al, 2006). The inverse measures in performance space tie s the concepts of safety factor and target failure probability. The inverse measures transl ate the deficiency in failure probability to PAGE 16 16 deficiency in performance measure and hence provide a more quantitative measure of the resources needed to satisfy safety requiremen ts. Among the several advantages they exhibit, inverse measures like probabilistic sufficiency factor (PSF) help in stable and accelerated convergence in optimization, better response surfa ce approximations compared to surfaces fit to other reliability measures. Since inverse measure e xhibit several advantages, its usefulness in tail modeling is explored in this work. Here we develop 3 alternate tail extrapolation schemes including inverse measures that can complement the classical tail modeling techniques. The alternate extrapolation schemes are based on va riable transformations and approximating the relationship between the inverse m easure and transformed variable. The primary motivation of this work is to develop an efficient reliability estimation procedure employing tail modeling techniques in cluding inverse measures that can estimate reliability measures corresponding to high reliabili ty with samples that are only sufficient to assess reliability measure corresponding to low re liability levels. This is a tradeoff between accuracy and computational power. Yet, as the st udies reveal, the accuracy lost to the reduction of computational power is very reasonable. Th e reduction in computational cost is about a minimum of 3 orders of magnitude for the same level of accuracy. Goel et al., (2006) developed a method to extend the utility of an ensemble of surrogates. When faced with multiple surrogates, they explored the possibility of using the best surrogate or a weighted average surrogate model instead of in dividual surrogate models. In a similar fashion, in order to take advantage of both the cl assical tail modeling t echniques and alternate extrapolation schemes and still come up with the best prediction, we pr opose to apply all the techniques simultaneously and use th e median of the five estimates as the best estimate. Here, we call using all the techniques simultaneously as multiple tail models (MTM). In addition to PAGE 17 17 arriving at the estimate, the range of the five techniques can be used to approximate the error associated with the median estimate. Moreover, the MTM approach can be used to replace the reliability analysis in reliab ility based design framework. PAGE 18 18 CHAPTER 2 RELIABILITY BASED DE SIGN OPTIMIZATION Introduction Optimization is the process of minimizing a co st which is a functi on of design variables (and other variables) subject to constraints that prescribe the design requirement. Often times, the variables involved are uncertain and probabilistic approaches are called for to account for the uncertainty as an alternate to the traditional sa fety factor approach, to obtain better designs. Several probabilistic approaches has been pr oposed in the last d ecade like Robust Design Optimization (Phadke, 1989, Gu, 2000), Reliability based Design Optimization(Frangopal, 1985, Tu, 1999), Fuzzy optimization (Sakawa 1993, Maglar as et al, 1997), Reliability design using evidence theory (Soundappan et al., 2004). Reliab ilitybased Design Optimization (RBDO) is widely used because it allows the designer to pr escribe the level of reliability required. This section presents a review of the standard reliabilitybased design formulations and reliability estimation methods. Standard Formulations Primarily, reliabilitybased design consists of minimizing a cost function while satisfying reliability constraints. The reliability cons traints are based on the failure probability corresponding to each failure mode or a single failure mode describing the system failure. The estimation of failure probability is usually performed by reliability analysis. In structural engineering, the system performance criterion is described by the limit state function which is typically expressed as the di fference between the capacity of the system and the response of the system which is expressed as: ()()()crGGG xxx (21) where, G is the limit state function and GC, Gr are the capacity and the response respectively. All the three quantities are functions of a vector x which consists of design (and random) variables PAGE 19 19 and other parameters. The syst em is considered failed if G < 0 and safe if G > 0. If x is random, the design is considered a failure if the pr obability of failure is unacceptably high. Often times, RBDO is formulated as a double loop (nested loop) problem. The outer loop performs the optimization with respect to design variables, wh ile the inner loop performs the reliability analysis estimating the failure proba bility. This is mathematically expressed as: argmin cost(x) subject to (0) 1,2...jjj fftetPPGPjnm (22) where the cost is a func tion of design variables, nm is the number of failure modes, G is the limit state function. f targetP is the prescribed failure probability levels to which the designer intends to design the structure. Here, estimation of probabilis tic constraint requires reliability analysis. Often times the magnitude of failure probabil ity is low and hence engineers prefer to work with reliability index as the alternate measur e of reliability. The reli ability index and failure probability are related as: ()fP (23) where, is the standard normal cumulative density function (CDF). Sin ce the inverse of the standard normal CDF is required to estimate the reliability index, it is an inverse measure. Probabilistic constraints in Eq.(22) can be written as: j j target (24) Employing Eq. (24) for prescribing the probabilistic cons traint is called the Reliability Index Approach (RIA). In RIA, the failure reliability measure is usually calculated via FORM which is an iterative process and so computationally expensive and sometimes features convergence problems (Tu et al., 1999) In order to reduce the computational cost of double loop, various techniques has been proposed which can be classi fied into two categories as: (i) techniques that improve the efficiency of uncertainty analysis like fast probability integration (Wu 1994) and twopoint adaptive non linear approximations (Grandhi and Wang 1998) (ii) techniques that PAGE 20 20 modify the formulation of the probabilistic constr aints, for instance, using inverse reliability measures in the probabilistic constraint. Someti mes, researchers formulate a single loop RBDO that avoids nested loops. The id ea of single loop formulations re sts on the basis of formulating the probabilistic constraint as deterministic co nstraints by either approximating the KarushKuhnTucker conditions at the MPP or approxi mating the relationship between probabilistic design and deterministic design safety factors. A detailed survey of both single and double loop methods in RBDO is presented in the literature survey of Chapter 3. Kharmanda et al (2002, 2004, a nd 2007) developed RBDO solution procedures relative to two views points: reliability and op timization. From an optimization view point, Kharmanda et al (2002) developed a hybrid method based on simultane ous application of the reliability and the optimization problem that reduced the computational time. The hybrid method, compared to classical RBDO, minimizes a new form of the objective function which is expressed as the product of the original objective function and the reliability index in the hybrid space. Since the minimization of the objective function is carried out in both the deterministic variables and random variable space, it is referred to as hybrid design space. The reliability index in the hybrid space is the distance between the optimum point and the design point in the hybrid space However, the hybrid RBDO problem was more comple x than that of the deterministic design and may not lead to local optima In order to addr ess both the challenges, Kharmanda et al (2004) propose an optimum safety factor approach that computes safety factors satisfying a reliability level without demanding additional computing cost for the reliability evaluation. Kharmanda et al (2007) developed the hybrid design spaces an d optimal safety factor equations for three distributions namely, the normal, lognormal a nd uniform. The optimal safety factors are computed by sensitivity analysis of the limit stat e function with respect to the random variables. PAGE 21 21 They report that the optimal safety factor appr oach has several advantages like smaller number of optimization variables, good c onvergence stability, lower compu ting time and satisfaction of required reliability levels. In order to estimate th e reliability measure for reliability analysis, one need to use analytical or simulation approaches which are described below. MomentBased Methods In standard reliability technique s, an arbitrary random vector x is mapped to an independent standard normal vector U (variables with normal distri bution of zero mean and unit variance). This transformation is known as Ro senblatt transformation (Rosenblatt, 1952, Melchers, 1999, pp.118120). The limit state func tion in the standard normal space can be obtained as G (U) = G (T(X)) where T is the transformation. If the limit state function in the standard normal space is affine, i.e., if (U)UTG then failure probability can be exactly calculated as fP This is the basis fo r moment based methods. Momentbased methods provide for less expensiv e calculation of the pr obability of failure compared to simulation methods, although they ar e limited to a single failure mode. The First Order Reliability Method (FORM) is the most wi dely used momentbased technique. FORM is based on the idea of the linear approximation of the limit state function and is accurate as long as the curvature of the limit state function is not to o high. In the standard normal space the point on the first order limit state function at which the di stance from the origin is minimum is the MPP. When the limit state has a significant curvat ure, second order methods can be used. The Second Order Reliability Method (SORM) appr oximates the measure of reliability more accurately by considering the eff ect of the curvature of the lim it state function (Melchers, 1999, pp 127130). PAGE 22 22 Figure 21 illustrates the concep t of the reliability index and MPP search for a two variable case in the standard normal space. In reliability analysis, concerns are first focused on the ()0 G U curve. Next, the minimum distance to the origin is sought. The corresponding point is the MPP because it contributes most to failure. Th is process can be mathematically expressed as: To find *U Minimize Subject to ()TUU GU (25) Where U is the MPP. The calculation of the failure probability is based on linearizing the limit function at the MPP. It is to be noted that re liability index is the distance to the limit state function from the origin in the standard normal space. That is, it is an inverse measure in the input space (standard normal). Monte Carlo Simulation Variants of Monte Carlo methods were originally practiced u nder more generic names such as statistical sampling. The na me and the method were made famous by early pioneers such as Stainslaw Ulam and John von Neumann and is a re ference to the famous casino in Monaco. The methods use of randomness and repetitive nature is similar to the casi nos activities. It is discussed in Metropolis (1987) that the method earns its name from the fact that Stainslaw Ulam had an uncle who would borrow money from relative just to go to Monte Carlo. Monte Carlo integration is e ssentially numerical quadrature using random numbers. Monte Carlo integration methods are algorithms for the approximate evaluation of definite integrals, usually multidimensional ones. The usual algorithms evaluate the integrand at a regular grid. Monte Carlo methods, however, randomly choose the points at which the integrand is evaluated. Monte Carlo Methods are based on the analogy be tween volume and probability. Monte Carlo calculates the volume of a set by interpreting th e volume as probability. Simply put, it means, PAGE 23 23 sampling from all the possible outcomes and cons idering the fraction of random draws that in a given set as an estimate of the sets volume. Th e law of large numbers ensures that this estimate converges to the correct value as the number of draws incr eases. The central limit theorem provides information about the likely magnitude of the error in the estimate after finite number of draws. For example, consider estimating the integral of a function f over a unit interval. The () f xdx (26) integral can be expres sed as an expectation () E fU, with U uniformly distributed between 0 and 1. The Monte Carlo estimate for the integral can be obtained as: ()n i i f U n (27) If f is continuous and integrable over [0, 1], th en, by the strong law of large numbers, the estimate in Eq. (27) converges to the actual va lue with probability 1 as n In fact, if f is square integrable and we set (())f f xdx (28) then the error in the Monte Carlo estimate is approxi mately normally distributed with zero mean and standard deviation of /fn, the quality of this approximation improving as n increases. The parameter f in Eq.(28) will not be known in a real setting but can be approximated by the sample standard deviation. Thus, we not only obtain the estimate but also the error contained in the estimate. The form of the standard error is th e main feature of the MCS. Reducing this error to half require about increasing the num ber of points by four; increasing the precision by one decimal point requires 100 times as many points. Yet, the advantage of MCS lies in the fact that its /() On convergence rate is not restricted to integrals PAGE 24 24 over the unit interval and can be extended to any dimension. The standard error will still have the same form irrespective of dimension. This a dvantage is unlike other numerical methods. Thus, Monte Carlo simulation is attractive to solve integrals in higher dimension. In structural reliability estimation, MCS i nvolves sampling each random variable to give a sample value. Then, the limit state function is com puted at every realization. If the limit state is violated, the structure has failed. The experiment is repeated many times with a randomly chosen vector. If N trials are conducted, the estimate f Pof the probability of failure is given approximately as: (())i fnGx P N (29) Where (()) nGx denotes the number of trials n for whic h the structure failed. The number of trails N is related to the desired accuracy of failure probability. MCS need at least 100*N samples when the failure probabi lity is of the order of 1over N. One can find the standard deviation of failure probability in Eq. (29) using the following expression: () f fPP N (210) There are many variants to the abovementi oned crude MCS. Researchers use several variance reduction techniques li ke importance sampling (Glasserman, 2004) to make better use of MCS and extract more information from the same level of simulation. Importance sampling requires one to find a sampling PDF which uses the information of greatest probability density in the failure zone. This is conceptually similar to MPP which is essentially a point that contributes most to failure. An alternate formulation of the MCS is used by Smarslok et al (2006) to exploit the advantage of separable Monte Carlo simulatio ns. This integration method takes advantage of a special form of the limit state function in wh ich the limit state is composed of difference between capacity and response. PAGE 25 25 Figure 21. Reliability anal ysis and MPP. Subscript re fers to iteration number PAGE 26 26 .CHAPTER 3 INVERSE RELIABILITY MEASURES In probabilistic approaches the difference be tween the computed probability of failure or reliability index and their target values does not provide the designer with easy estimates of the change in the design cost needed to achieve these target values. Alternately, inverse reliability measures are capable of providing this informati on. Several inverse reliability measures (e.g., probabilistic performance measure and probabilistic sufficiency factor) that are essentially equivalent have been introduced in recent years. The different names for essentially the same measure reflect the fact that different research ers focused on different advantages of inverse measures. This chapter reviews the various inve rse measures and describes their advantages compared to the direct measures of safety such as probability of failure and reliability index. Methods to compute the inverse measures ar e also described. Reliabilitybased design optimization with inverse measure is dem onstrated with a beam design example. Literature Review In deterministic approaches, a safety factor is defined as the ratio of the capacity to resistance of a system, typically calculated at the mean values. However, in probabilistic approaches, as both capacity and response are random the safety factor is also a random number. One safety measure that combines the advantages of the safety factor and probability of failure was proposed by Birger (1970, as reported by Elisha koff, 2001). Birger relate s the safety factor and the fractile of the safety f actor distribution corres ponding to a target proba bility of failure. It belongs to a class of inverse reli ability measures, which carry that name because they require use of the inverse of the cumulative distribution f unction. Several researcher s developed equivalent inverse reliability methods (Lee and Kwak 1987; Kiureghian et al 1994; Li and Foschi 1998; Tu et al. 1999; Lee et al. 2002; Qu and Haftka 2003; Du et al. 2003) that are closely related to the PAGE 27 27 Birger measure. These measures quantify the level of safety in terms of the change in structural response needed to meet the ta rget probability of failure. Lee and Kwak (1987) used the inverse formulation in RBDO and showed that it is preferable for design when the probability of failure is very lo w in some region of the design space so that the safety index approaches infinity. Kiureghian et al. (1994) addressed the inverse reliability problem of seeking to determine one unknown para meter in the limit state function such that a prescribed first order reliability index is attained To solve the inverse reliability problem, they proposed an iterative algorithm based on the Haso ferLindRackwitzFiessler algorithm. Li and Foschi (1998) employed the inverse reliability st rategy in earthquake and offshore applications to solve for multiple design parameters. They show that it is an efficient method to estimate design parameters corresponding to targ et reliabilities. Kirj nerNeto et al. (1998) reformulated the standard RBDO formulation similar to Lee and Kwa k (1987) except that they use an inequality constraint. They developed a semiinfinite op timization algorithm to solve the reformulated problem. This formulation does not require seco nd order derivatives of the limit state functions and obviates the need for repeated reliability i ndices computation. However, they found that the approach can result in conserva tive design. Royset et al. (2 001) extended the reformulation technique discussed in KirjnerNet o et al. (1998) for reliabilityba sed design of series structural systems. The required reliability and optimization calculations are completely decoupled in this approach. Tu et al. (1999) dubbed the invers e measure approach the Performance Measure Approach (PMA) and called the inverse measure probability performance measure (PPM). Lee et al. (2002) adopted the same procedure as Tu et al. (1999) and named it target performance based approach, calling the inverse measure th e target performance. They compared the reliability index approach and inverse measure based approach and found that the latter was PAGE 28 28 superior in both computational efficiency and num erical stability. Youn et al. (2003) showed that PMA allows faster and more stable RBDO co mpared to the traditional Reliability Index Approach (RIA). When accurate statistical information is not avai lable for input data, it is not appropriate to use probabilistic method for stochastic stru ctural analysis and design optimization and researchers resort to possibility based design optimization (PBDO) methods in such situations. Mourelatos and Zhou, 2005, discus s a PMA based PBDO. They formulate the inverse possibility analysis problem for the possibilistic constrai nt. Unlike FORM based RBDO which is based on linear approximation, the PBDO is exact. To perform the inverse possibility analysis, Du et al (2006b) proposes a maximal possibility search (MPS) method with interpolation to address the challenges of accuracy and computational expe nse exhibited by traditional methods like the multilevel cut and alpha level optimization method respectively. They report that the MPS evaluates possibility constraint efficiently and accurately for nonlinear st ructural applications. Often times industry design problems deal with uncertainties with sufficient data and uncertainties with insufficient data simultaneously To address such situations Du et al., (2006a) extend the possibilitybased design optimizati on (PBDO) method. They propose a two step approach to generate membership function from the available data. Initially, they develop a temporary PDF using the available data and th en generate a membership function from the temporary PDF. They report that PMA based PB DO combined with MPS method can adequately address design problems with mixed input variables and demonstrates that it evaluates possibility constraint efficiently, stably and accurately for nonlinear structural applications. Qu and Haftka (2003, 2004) called the invers e measure probability sufficiency factor (PSF) and explored its use for RBDO with multiple failure modes through Monte Carlo PAGE 29 29 Simulation (MCS) and response surface approximation (RSA). They showed that PSF leads to more accurate RSA compared to RSA fitted to failure probability, provides more effective RBDO, and that it permits estimation of the necessary change in the design cost needed to meet the target reliability. Moreover, PSF enables performing RBDO in variablefidelity fashion and sequential deterministic optimization fashion to reduce the computational cost (Qu and Haftka, 2004a and 2004b). An initial study of using PSF to convert RBDO to sequential deterministic optimization was performed to solve problems with reliability constraint s on individual failure modes (Qu and Haftka, 2003). An improved vers ion for system reliab ility problems with multiple failure modes was developed for reliab ilitybased global optimization of stiffened panels (Qu and Haftka, 2004b). Du et al. (2003, 2004) employed PMA to formul ate RBDO, but used pe rcentile levels of reliability (1 minus failure proba bility) in the probabilistic co nstraint and called the inverse measure as percentile performance. Traditiona lly, design for robustness involves minimizing the mean and standard deviation of the performance. Here, Du et al. (2003) proposed to replace the standard deviation by percentile performance di fference, which is the difference between the percentile performance corresponding to the left tail of a CDF a nd the right tail of that CDF. They demonstrated increased computational e fficiency and more accurate evaluation of the variation of the objective performa nce. In an effort to address reliabilitybased designs when both random and interval variable s are present, Du and Sudijian to (2003) proposed the use of percentile performance with worstcase combination of the interval variables for efficient RBDO solutions. Du and Chen (2004) developed the sequentia l optimization and reliability assessment (SORA) to improve the efficiency of the probabili stic optimization. The me thod is a serial single PAGE 30 30 loop strategy, which employs percentile performa nce and the key is to establish equivalent deterministic constraints from probabilistic cons traints. This method is based on evaluating the constraint at the most probable point of the i nverse measure in Section IV below) based on the reliability information from the previous cycl e. This is referred to as design shift (Chiralaksanakul and Mahadeva n 2004; Youn et al. 2004). They show that the design quickly improves in each cycle and is computationall y efficient. The sequential optimization and reliability assessment, however, is not guaranteed to lead to an optimal design. Singlelevel (or unilevel) techniques that are equivalent to the standard RBDO formulation are based on replacing the RIA or PMA inner loop by the corresponding KarushKuhnTucker conditions. Here again, Agarwal et al. (2004) showed that the PMA approach is more efficient than the unilevel RIA approach due to Kuschel and Rackw itz (2000). Since most commercial optimizers are numerically unreliable when applied to pr oblems accompanied by many equality constraints, Agarwal et al (2007) use homot opy methods for constraint rela xation and to obtain a relaxed feasible design. They solve a series of optimi zation problem as the relaxed optimization problem is transformed via a homotopy to the original problem. They show that it is easier to solve the relaxed problem and make a gradual progress to wards the solution than solve the original problem directly. The several inverse measures discussed above are all based on the common idea of using the inverse of the cumulative distribution f unction. The numerous names for the inverse measures contemplate that they were devel oped by different researchers for different applications. Since these inverse measures come under various names, it is easy to fail to notice the commonality among them. PAGE 31 31 Birger Safety Factor The safety factor, S is defined as the ratio of the capacity of the system cG (e.g., allowable strength) to the responserGwith a safe design satisfyingrcGG To account for uncertainties, the design safety factor is greater th an one. For example, a load safety factor of 1.5 is mandated by FAA in aircraft applications. To address the probabilistic interpretation of the safety factor, Birger (1970) proposed to cons ider its cumulative dist ribution function (CDF)SF: ()Prob()c S rG Fss G (31) Note that unlike the deterministic safety factor, which is normally calculated for the mean value of the random variables, crGG in Eq.(31) is a random function. Given a target probability, f targetP Birger suggested a safety factor s*(which we call here the Birger safety factor) defined in the following equation Prob()Prob()c Sftarget rG F(s*)s*Ss*P G (32) That is, the Birger safety f actor is found by setting the valu e of cumulative distribution function (CDF) of the safety factor equal to the target probability That is, we seek to find the value of the safety factor that makes the CDF of the safety factor equal to the target failure probability. This requires the inverse transformation of the CDF, hence the terminology of inverse measure. Probabilistic Sufficiency Factor Qu and Haftka (2003, 2004) developed a sim ilar measure to the Birger safety factor, calling it first the probabilistic sa fety factor and then the probab ilistic sufficiency factor (PSF). They obtained the PSF by Monte Carlo simulatio n and found that the re sponse surface for PSF was more accurate than the re sponse surface fitted to failure probability. Later, they found the PAGE 32 32 reference to Birgers work in E lishakoffs review (2001) of safety factors and their relations to probabilities. It is desirable to avoid the term safety factor fo r this entity because the common use of the term is mostly deterministic and independent of the target safety level. Therefore, while noting the identity of the Birger safety f actor and the probabilistic sufficiency factor, we will use the latter term in the following. Failure happens when the actual safety factorS is less than one. The basic design condition that the probability of failure should be smaller than the target probability for a safe design may then be written as: Prob(1)(1) f SftargetPSFP (33) Using inverse transformation, Eq.(33) can be expressed as: 11()SftargetFPs* (34) The concept of PSF is illustrated in Figure. 31. The design requirement f targetP is known and the corresponding area under the probability density function of the safety factor is the shaded region in Figure.31.The upper bound of the abscissa s* is the value of the PSF. The region to the left of the vertical line 1 S represents failure. To sa tisfy the basic design condition s* should be greater than or equal to one. In order to achieve this, it is possible to either increase cG or decreaserG The PSF s*, represents the factor that has to multiply the response rG or divide the capacitycG so that the safety factor be raised to 1. For example, a PSF of 0.8 means that rG has to be multiplied by 0.8 or cG be divided by 0.8 so that the safety factor ratio increases to one. In other words, this means that rG has to be decreased by 20 % (10.8) or cG has to be increased by 25% ((1/0 .8)1) in order to achieve the target failure probability. It can be observed that PSF is a safety f actor with respec t to the target failure probability and is automatically norma lized in the course of its formulation. PAGE 33 33 PSF is useful in estimating the resources needed to achieve the required target probability of failure. For example, in a stressdominated linear problem, if the target probability of failure is 105 and a current design yields a probability of failure of 103, one cannot easily estimate the change in the weight required to achieve the ta rget failure probability. Instead, if the failure probability corresponds to a PSF of 0.8, this indica tes that maximum stresses must be lowered by 20% to meet the target. This permits the designers to readily estimate the weight required to reduce stresses to a given level. Probabilistic Performance Measure In probabilistic approaches, instead of th e safety factor it is customary to use a performance function or a limit stat e function to define the failure (or success) of a system. For example, the limit state function is expressed as in Eq(21). In terms of safety factor S another form of the limit state function is: ()110c rG GS G X X X (35) Here,() G X and () GXare the ordinary and normalized limit state functions, respectively. Failure happens when() G X or () GX is less than zero, so the probability of failure f P is: Prob(()0)fPG X (36) Prob(()0)fPG X (37) Using Eq. (36), Eq. (37) can be rewritten as: Prob(()0)(0) f GftargetPGFPX (38) Prob(()0)(0) f ftarget GPGFPX (39) where GF and GF are the CDF of () GX and () GX, respectively. Inverse transformation allows us to write Eqs. (38) and (39) as: PAGE 34 34 10()GftargetFPg* (310) 10()ftarget GFPg* (311) Here, g* and g* are the ordinary and normalized Probabilistic Performance Measure (PPM, Tu et al, 1999), respectively. PPM can be defined as the solution to Eqs. (38) and (39). That is, the value in () (instead of zero) wh ich makes the inequality an equality. Hence PPM is the value of the limit state function that makes its CDF equa ls the target failure probability. Figure 32 illustrates the concept of PPM. The shaded area co rresponds to the target failure probability. The area to the left of the line 0 G indicates failure. g* is the factor that has to be subtracted from Eq. (36) in order to make the vertical line g* move further to the right of the 0 G line and hence facilitate a safe design. For example, a PPM of 0.8 means that the de sign is not safe enough, and 0.8 has to be subtracted from () GXin order to achieve the target probability of failure. A PPM value of 0.3 means that we have a safety margin of 0.3 to reduce while improving the cost function to meet the target failure probability. Considering g* as the solution for Eq.(311), it can be rewritten in terms of the safety factor as: Prob(()1) f targetGSg*PX (312) Comparing Eqs. (15), (23) and (24), we can observe a relationship between s* and g* PSF (s*) is related to the normalized PPM ( g* ) as: 1 s*=g* (313) This simple relationship between PPM and PSF s hows that they are closely related, and the difference is only in the way the limit state func tion is written. If the limit state function is expressed as the difference between capacity and response as in Eq.(36), failure probability PAGE 35 35 formulation allows us to define PPM. Alternativ ely, if the limit state function is expressed in terms of the safety factor as in Eq. (37), the corresponding failure pr obability formulation allows us to define PSF. PSF can be viewed as PPM derived from the normalized form of Eq. (36). The performancemeasure approach (PMA) notation may appeal because of its generality, while the PSF notation has the advantage of being automatica lly scaled and being e xpressed in terms that are familiar to designers who use safety factors. The PPM can be viewed as the distance from the G = 0 to target failure probability line, in the performance function space. This is analogi cal to reliability index being a measure of distance in the input variable space. The major difference is the measurement of distance in different spaces, the performance functi on (or output) space a nd the input space. Inverse Measure Calculation Simulation ApproachMo nte Carlo Simulation Conceptually, the simplest approach to eval uate PSF or PPM is by Monte Carlo simulation (MCS), which involves the generation of rando m sample points according to the statistical distribution of the variables. The sample poi nts that violate the safety criteria in Eq. (36) are considered failed. Figure.33 illustrates the co ncept of MCS. A twovariable problem with a linear limit state function is considered. The straig ht lines are the contour lines of the limit state function and sample points generated by MCS are represented by small circles, with the numbered circles representing failed samples. Th e zero value of the limit state function divides the distribution space into a safe region and a fa ilure region. The dashed lines represent failed conditions and the continuous lin es represent safe conditions. The failure probability can be estimated using Eq.(29) In Figure. 33, the number of sample points that lie in the failure region above the 0 G curve is 12. If the total number of PAGE 36 36 samples is 100,000, the failure probability is estimated at 1.2x104. For a fixed number of samples, the accuracy of MCS deteriorates wi th the decrease in failure probability. For example, with only 12 failure points out of the 100,000 samples, the standard deviation of the probability estimate is 3.5x105, more than a quarter of the estimate. When the probability of failure is significantly smaller than one over th e number of sample points, its calculated value by MCS is likely to be zero. PPM is estimated by MCS as the thn smallest limit state function among the N sampled functions, where f targetnNP. For example, considering the example illustrated in Figure. 33, if the target failure probability is 104, to satisfy the target probability of failure, no more than 10 samples out of the 100,000 should fail. So, the focu s is on the two extra samples that failed. PPM is equal to the value of the highest limit state function among the n (in this case, n = 10) lowest limit state functions. The numbered small circles are the sample points that failed. Of these, the three highest limit states are shown by the dashed lines. The tenth smallest limit state corresponds to the sample numbered 8 and has a lim it state value of 0.4, which is the value of PPM. Mathematically this is expressed as: i i = 1(())N thg*nminG X (314) where, thn min is the thn smallest limit state function. So, the calculation of PPM in MCS requires only sorting the lowest limit state func tion in the MCS sample. Similarly, PSF can be computed as the nth smallest factor among the N sampled safety factors and is mathematically expressed as **: i i = 1(())N ths*nminS X (315) ** A more accurate estimate of PPM or PSF will be obtained from the average of the nth and (n+1)th smallest values. So in the case of Figure. 4, PPM is more accurately estimated as 0.35. PAGE 37 37 Finally, probabilities calculated through M onte Carlo Simulation (MCS) using a small sample size are computed as zero in the region wher e the probability of failure is lower than one over the number of samples used in MCS. In that region, no useful gradient information is available to the optimization routin e. On the other hand, PSF value va ries in this region and thus provides guidance to optimization. MCS generates numerical noise due to limited sample size. Noise in failure probability may cause RBDO to converge to a spurious minimu m. In order to filter out the noise, response surface approximations are fitted to failure prob ability to create a socalled design response surface. It is difficult to constr uct a highly accurate design res ponse surface because of the huge variation and uneven accuracy of failure probability. To overcome these difficulties, Qu and Haftka (2003, 2004) discuss the usage of PSF to improve the accuracy of the design response surface. They showed that design response surfac e based on PSF is more accurate compared to design response surface based on failure probabi lity and this accelerates the convergence of RBDO. For complex problems, response surface a pproximations (RSA) can also be used to approximate the structural response in order to reduce the computationa l cost. Qu and Haftka (2004b) employ PSF with MCS based on RSA to desi gn stiffened panels unde r system reliability constraint. Analytical ApproachMomentbased Methods Inverse reliability measures can be estimated using moment based methods like FORM. The FORM estimate is good if the limit state is linear but when the limit state has a significant curvature, second order methods can be used. The Second Order Reliability Method (SORM) approximates the measure of reliability more accurately by considering the effect of the curvature of the limit state function (Melchers, 1999, pp 127130). PAGE 38 38 Figure 34. illustrates the c oncept of inverse reliability analysis and MPP search. The circles represent the curves with the target curve represented by a dashed circle. Here, among the different values of limit state functions that pass through the target curve, the one with minimum value is sought. The value of th is minimal limit state function is the MPP as shown by Tu et al. 1999. The point on the target circle with the minima l limit state function is sought. This point is also an MPP and in orde r to avoid confusion be tween the usual MPP and MPP in inverse reliability analys is, Du et al. (2003) coined th e term most probable point of inverse reliability (MPPIR) and Lee et al. (2002) called it the mi nimum performance target point (MPTP). Du et al. (2003) devel oped the sequential optimization a nd reliability analysis method in which they show that evaluati ng the probabilistic constraint at the design point is equivalent to evaluating the deterministic constraint at the most probable point of the i nverse reliability. This facilitates in converting the probabi listic constraint to an equivale nt deterministic constraint. That is, the deterministic optimization is performed using a constraint limit which is determined based on the inverse MPP obtained in th e previous iteration. Kiureghi an et al. (1994) proposed an extension of the HasoferLindR ackwitzFiessler algorithm that us es a merit function and search direction to find the MPTP. In Fi gure 34, the value of the mini mal limit state function or the PPM is 0.2. This process can be expressed as: Minimize :() Subject to :UT targetGU UU (316) In reliability analysis the MPP is on the ()0 G Ufailure surface. In inverse reliability analysis, the MPP search is on the target curve. One of the main advantages when inverse measures are used in the FORM perspective is that, the formulation of the optimization problem is as expressed in Eq. (316) where the constraint has a simple form (circle) compared to the cost function. The cost function has in general a complicated expression due to the nonlinear PAGE 39 39 transformation between physical random variable X to the standard random variable U. It is well known that it is easy to satisfy a simple cons traint compared to a complicated constraint irrespective of the cost function. On comparing the formulation in Eq.(25) with the formulation in Eq.(316), it is evident that it is easier to solv e the latter because of the simpler constraint. Reliability Based Design Optimi zation with Inverse Measures The probabilistic constraint in RBDO can be prescribed by several methods like the Reliability Index Approach (RIA ), the Performance Measure Approach (PMA), the Probability Sufficiency Factor appro ach (PSF), see Table31. Table 31. Different approaches to prescribe the probabilistic constraint Method RIA PMA PSF Probabilistic constraint target 0 g* 1s* Quantity to be computed Reliability index ( ) PPM (g*) PSF (s*) In RIA, can be computed as the pr oduct of reliability analysis as discussed in the previous secti on. The PPM or PSF can be computed through inverse reliability analysis or as a byproduct of re liability analysis using MCS. To date, most researchers have used RIA to prescribe the probabilistic constraint. However, the advantages of the inverse measures, illustrated in the next section, have led to its growing popularity. Beam Design Example The cantilever beam shown in Figure. 35, taken from Wu et al (2001), is a commonly used demonstration example for RBDO methods. The length L of the beam is 100". The width and thickness is represented by w and t. It is subjected to end transverse loads X and Y in orthogonal directions as shown in the Figure 351. The objective of the design is to minimize the 1 For this example, the random variables are shown in bold face PAGE 40 40 weight or equivalently the cross sectional area: A=wt subject to two reliability constraints, which require the safety indices for stre ngth and deflection constraints to be larger than three. The first two failure modes are expresse d as two limit state functions: Stress limit: sGRRYX wtwt (317) Tip displacement limit: dOOLYX GDDD Ewt tw (318) where R is the yield strength, E is the elastic modulus, D0 the displacement limit, and w and t are the design parameters. R, X, Y and E are uncorrelated random vari ables and their means and standard deviations are defined in Table 32. Table 32. Random variables for beam problem Random Variables X (lb) Y (lb) R (psi) E (psi) Distribution (, ) Normal (500,100) Normal (1000,100)lb Normal (40000,2000) Normal (296,1.456) Design for Stress Constraint The design with strength reliability constraint is solved first, followed by the design with a system reliability constraint. The results for the strength constraint are presented in Table 33. For the strength constraint, the limit state is a linear combination of normally distributed random variables, and FORM gives accurate results fo r this case. The MCS is performed with 100,000 samples. The standard deviation in the estima ted failure probability is calculated by MCS as: p1 f fP(P) N (319) In this case, the failure probability of 0.0013 calculated from 100,000 samples has a standard deviation of 1.144. It is seen from Table 33 that the designs obtai ned from RIA, PMA and PSF match well. Since the stress in Eq. (317)is a linear functi on of random variables, PAGE 41 41 the RIA and PMA are exact. The more conservati ve design from PSF is due to limited sampling of MCS. Table 33. Comparison of optimum designs for the stress constraint Minimize objective function A=wt such that 3 Inverse Reliability Analysis Method Reliability Analysis FORM (RIA) FORM (PMA) MCS (PSF) (Qu and Haftka,2003) Exact Optimum (Wu et al. 2001) w 2.4460 2.4460 2.4526 2.4484 Optima t 3.8922 3.8920 3.8884 3.8884 Objective Function 9.5202 9.5202 9.5367 9.5204 Reliability Index 3.00 3.00 3.0162 3.00 Failure Probability 0.00135 0.00135 0.00128 0.00135 Comparison of Inverse Measures The relation between PSF and PPM in Eq. (313) is only approximate when PPM is calculated by FORM and PSF by MCS. PSF suffers from sa mpling error and PPM from error due to linearization. For the linear stress constraint, and with a large Monte Carlo sample, the difference is small, as seen in Table 34. It may be expe cted that the MPTP (minimum performance target point) should also be close to the point used to calculate PSF. This resu lt may be useful, because when a response surface is used fo r an approximation of the response, it is useful to center it near the MPTP. In order to check th e accuracy of the MPTP estimation, the MPP of PPM and the point corresponding to the PSF are compared and the results are tabulated in Table 35. The coordinates corresponding to the PSF, computed by a millionsample MCS, deviate considerably from the MPTP. Since the accuracy of the poi nts computed by MCS depends on the number of samples used, an increased number of samples lead to more accurate results albeit at increased computational cost. Alternative approaches to obtain the points with better accuracy without increasing the number of samples are to averag e the coordinates computed by repeated MCS simulations with fewer numbers of samples. A lternatively, we can average a number of points PAGE 42 42 that are nearly as critical as the PSF point. That is, instead of using only the iX corresponding to the nS in Eq (315), we use also th e points corresponding to ...,npnpnpSSS in computing the PSF, where 2p is the total number of points that is averaged around the PSF. It can be observed from Table 35 that averaging 11 points around the PSF matches well with the MPTP, reducing a Euclidean distance of about 0.831 for the raw PSF to 0.277 with 11point average. The values of X, Y and R presented in Table 35 are in the standard normal space Table 34. Comparison of inverse measures for w =2.4526 and t =3.8884 (stress constraint) Method FORM MCS (17samples) Pf 0.001238 0.001241 Inverse Measure PPM: 0.00258 PSF: 1.002619 Table 35. Comparison of MPP obt ained from calculation of PSF for w =2.4526 and t =3.8884 (stress constraint) Use of PSF in Estimating the Required Chan ge in Weight to Achieve a Safe Design The relation between the stresses displacement and weight for this problem is presented to demonstrate the utility of PSF in estimating the required resources to ach ieve a safe design. Consider a design with the dimensions of A0= w0 t0 with a PSF of 0s* less than one. The structure can be made safer by scaling both w and t by the same factor c This will change the stress and displacement expressed in Eq.(317) and Eq. (318) by a factor of c3 and c4, respectively, and the area by a factor of c2. If stress is most critical, it will scale as c3 and the PSF will vary with the area, A as: Coordinates X Y R MPTP 2.1147 1.3370 1.6480 PSF 1)106 samples 2.6641 1.2146 1.0355 2)Average:10 runs of each 105samples 2.1888 1.8867 1.1097 3)Average:11 point s around PSF of 106 Samples 2.0666 1.5734 1.5128 PAGE 43 43 15 0 0A s*s A (320) Equation (320) indicates that a one percent in crease in area will in crease the PSF by 1.5 percent. Since nonuniform scaling of width and thickness may be more efficient than uniform scaling, this is a conservative estimate. Thus, for example, considering a design with a PSF of 0.97, the safety factor deficiency is 3% and the structure can be made safer with a weight increase of less than two percent, as shown by Qu and Haftka (2003). For a critical displacement state, s* will be proportional to 2 A and a 3% deficit in PSF can be corrected in under 1.5% weight increase. While for more complex structur es we do not have analytical expressions for the dependence of the displacements or the st resses on the design variables, designers can usually estimate the weight needed to redu ce stresses or displacements by a given amount. System Reliability Estimation Using PSF and MCS System reliability arises when the failure of a system is defined by multiple failure modes. Here, the discussion is limited to series systems. The estimation of system failure probability involves estimating the failure due to individual modes and failure due to th e interaction between the modes. This is mathematically expressed for a system with n failure modes as: 12 12131 12321()()...() ()()...() +()...()...Higher order termssysn nn nnnPPFPFPF PFFPFFPFF PFFFPFFF (321) It is easy to employ MCS to estimate system failu re, as the methodology is a direct extension of failure estimation for single modes. In the co ntext of employing analytical methods like FORM, estimation of failure regions bounded by single m odes are easy to estimate but estimating the probability content of odd shaped domains created because of the intersection of two or more modes is challenging. Technique s are available to estimate th e probability content at the intersection regions but require evaluating multip le integrals (Melchers, pp: 386400). Since the PAGE 44 44 number of multiple integrals that has to be ev aluated depends on the number of variables, the task becomes arduous when more variables are i nvolved. Instead, an alternate technique is to develop upper and lower bounds on the failure probabi lity of the structural system. Considering the terms in the first and the third rows (positive contributions) in Eq. (321) permits us to estimate the upper bound. When the terms in firs t and second rows (negative contributions) are considered, the lower bounds can be estimated. Improved bounds for the sy stem reliability are available in the literature. Owing to these challenges, MCS is mostly used when system reliability is to be estimated. As discussed earlier, MCS is co mputationally prohibitive when the failure probability to be estimated is low, because of the number of sample s required. PSF can be used to estimate system reliability. The method is an extension of th e PSF estimation for single failure mode. When M failure modes are considered, the most critical sa fety factor is calculated for each sample. Then the sorting of the thnminimum safety factor can proceed as in Eq.(315). This process is explained in the flow chart in Figure 36. The es timation of PSF for a system reliability case is demonstrated with an example in Appendix A Design for System Reliability by MCS and PSF Monte Carlo simulation is a good method to use for system reliability analysis with multiple failure modes. For the cantilever beam di scussed in the earlier section, the allowable deflection was chosen to be 2.25" in or der to have competing constraints (Wu et al ., 2001). The results are presented in Table 36. It can be obs erved that the contribution of the stress mode to failure probability dominates the contribution du e to displacement and the interaction between the modes. The details of the design proce ss are provided in Qu a nd Haftka (2004). They demonstrated the advantages of using PSF as an inverse safety measure over probability of failure or the safety index as a normal safety measure. They showed that the design response PAGE 45 45 surface (DRS) of the PSF was much more accurate than the DRS fo r the probability of failure. For a set of test points the erro r in the probability of failure was 39.11% from the DRS to the PSF, 96.49% for the DRS to the safety index, and 334.78% for the DRS to the probability of failure. Table 36. Design for System Reliabi lity by PSF, Qu and Haftka (2004)** Optima Objective Function Pf Safety Index Pf1 Pf2 Pf1 Pf2 PSF s w= 2.6041 t = 3.6746 9.5691 0.001289 3.01379 0.0011330.000208 0.000052 1.0006 ** 100,000 samples, Pf1, Pf2, Pf1 Pf2 Failure probabilities due to stress constraint, displacement constraint, and intersection between the modes, respectively. PAGE 46 46 Figure 31. Schematic probability density of the safety factor S. The PSF is the value of the safety factor corresponding to th e target probability of failure Figure 32. Schematic probability density of the limit state function. The Probabilistic Performance Measure is the value of the limit state f unction corresponding to the target probability of failure. PAGE 47 47 Figure 33. Illustration of the calculation of PPM with Monte Carlo Simulation for the linear performance function Figure 34. Inverse reliability analysis a nd MPP for target failure probability 0.00135 ( = 3) Figure 35. Cantilever beam subjected to horizontal and vertical loads PAGE 48 48 Figure 36. Flowchart fo r estimating system PSF 1...i x iN () 1...(#of modes)j iSxjM (min(()))j criSsortSx ftarget 1min() ; N th syscr iSnSnNP PAGE 49 49 CHAPTER 4 TAIL MODELLING AND RELIABILITY ANALYSIS Low failure probability problems (extreme value) require one to have sufficient data in the tails of the distribution which represent the extr emes. But this is seldom possible and instead researchers use tail modeling techniques based on extreme value theory to predict the probability of extreme events. The theory comprises a pr inciple for model extrapolation based on the implementation of mathematical limits as finite level approximations. Since several advantages are reported by working with inverse measures, it is logical to justify an attempt to perform tail modeling in the performance space along with in verse measures to estimate quantities at unobserved levels. This section di scusses the tail mode ling technique and how to apply it to find inverse measures. Tail Equivalence and Generalized Extreme Value Theory Tail modeling is based on the concept of approxi mating the tail distribution of the variable of interest by associating it w ith the tail of already known dist ribution and using the parameters and characteristics therein for further analysis This criterion is ade quately defined by tail equivalence. That is, two CDFs () Fx and () Qxcan be called tail equiva lent (Maes and Breitung, 1993) at the right tails when: 1() lim1 1()xFx Qx (41) This reveals that the error in approximating the small probability of 1() Fxby 1() Qxtends to zero as x The extremal theorem by Fisher and Tippet (1928) states that, ifn M is the maximum of a sequence of n independent random variables having a common distri bution function and if there exists sequences of constants 0na and nbsuch that PAGE 50 50 Pr()/() as nnnMbazHzn (42) for a nondegenerate distribution function H then H belongs to one of the extreme value distribution families namely, the Gumbel, Frechet a nd Weibull families. It is to be noted that regardless of the distribution for the population, the three types of the extreme value distributions are the only possible limits for th e distribution of the normalized maxima. In this respect, this theorem is to extreme value theory what central limit theorem is to central statistics. The central limit theorem states that the approximating distri bution of the sample mean is normal regardless of the parent distribution. The three families of distributions can be comb ined into a single family of models having distribution function of the form 1/()exp1 z Hz (43) such that :1()/0 zz, where the parameters satisfy 0 and This is the generalized extreme valu e (GEV) family of distributions. Here, is the location parameter, is the scale parameter, is the shape parameter. A more detailed discussion of the generalized extreme value distributions is pres ented in Appendix B Generalized Pareto Distribution In engineering applications, rather than maxima the interest is to address the excesses over threshold. In these situations, the generalized pareto distribution (GPD) arises as the limiting distribution. The concept of GPD is presented in Figure 41. Let y be a model output which is random and u be a large threshold of y The observations of y that exceed u are called exceedance. The conditional distribution ()uFz of the exceedance given that the data y is greater than the threshold u, is modeled fairly well by the GPD. Here,zyu Let approximation of PAGE 51 51 ()uFz using GPD be denoted by ()Fz. and are the shape and scale parameters respectively. For a large enough u, the distributio n function of () y u conditional on y > u, is approximately written as (Coles, 2001): 1 ,11 if 0 1exp if =0z Fz z (44) In Eq (44), max(0,) AAand0 z plays a key role in a ssessing the weight of the tail. Eq (44) can be justified as a limiting distribution as u increases (Coles, 2001, pp:7576). Tails can be classified based on as: 0, heavy tail (pareto type tail) 0, medium tail (exponential type tail) 0, light tail (Betatype tails) There are several parameter estimation methods to estimate the parameters and .It is to be noted that conditional excess CDF ()uFzis related to the CDF of interest () Fy through the following expression: ()() () 1()uFyFu Fz Fu (45) From Eq (45), the CDF of y can be expressed as: ()(1())()()uFyFuFzFu (46) When Eq (44) is substituted for()uFzin the above expression, Eq (46) becomes: 1()11()1() FyFuyu (47) For simplicity of presentation, only the case of0 is considered here. Once we obtain estimates of the parameters as and using some parameter estimation method like maximum PAGE 52 52 likelihood estimation method, least square regression th at are discussed later in the chapter, it is possible to estimate the pth quantile by inverting Eq.(47) : 11() ()1 1()pFp yFpu Fu (48) In order to extend the concept to st ructural applicati ons, if the output y is replaced by the limit state in Eq (35) then the interest is to model th e lower tail rather than the upper tail. However, modeling the upper tail of y is equivalent to modeling the lower tail of y Therefore, an equivalent way of expressing Eq.(35) is: 1()1()rGxSx (49) where rS is the reciprocal of the safety factor. Failure occurs when G1 > 0. rS at any required target reliability index can be estimated using Eq(48). If target fP refers to the target failure probability that we wish to design th e structure, PSF can be obtained as: target 1 target 11 1()f rfP SFPu Fu (410) Coles (2001) opines that it is easy to object to the pro cedure of tail modeling on the grounds that even with the support of asymptotic argument, there is an implicit assumption that the underlying stochastic mechanism of the pro cess being modeled is sufficiently smooth to enable extrapolation. However no credible alternative has been proposed to date. Threshold Selection The theory of excess over threshold is refe rred to as peaks over threshold method also. Performance of this approach is base d on the choice of the threshold value u. The basic theory identified by Pickands (1975) states that the pa reto distribution will be a practical family for statistical estimation of the tails, provided that the threshold is taken sufficiently high. The CDF is composed of 3 portions, the lower tail, uppe r tail and the ce ntral portion. In theory, the PAGE 53 53 threshold should be selected where the actual up per tail starts. But there is no straightforward globally accepted method to select threshold. Selection of threshold is a tradeoff between bias and variance. If the threshold selected is low, then some data points belong to the central part of the distribution and do not provide a good approximation to the tails. On the other hand if the threshold selected is too high, the number of data available for the tail approximation is much less and this might lead to excessive scatter in the final estimate. The proper selecti on of threshold is very important because it has important repercussions on the estimated value of the shape factor (Caers and Maes, 1998, McNeil and Saladin, 1997) and henc e on the final estimates such as the quantile, extreme values etc. There are many exploratory techniques like the mean excess plot which help in selecting the threshold. But in a simulation study, it is im practical to perform interactive data analysis required by the exploratory techniques to choose the threshold. The mean excess plot is the sample mean excess function defined as the su m of the excesses over th e threshold divided by the number of the data points which exceed the thre shold plotted with respect to the threshold. A reasonable choice of the optimal threshold is wher e this plot becomes linear. For further reading the reader is referre d to Coles (2001). Boos (1984) suggests that the ratio of Nex (number of tail data) over N (total number data) should be 0.02 (50< N <500) and the ratio should be 0.1 for 500< N <1000. Hasofer (1996) suggests using 1.5 NexN Caers and Maes (1998) propose to use a finite sample mean square error (MSE) as a criterion for estimating the threshold. They use the threshold value that minimizes the MSE. In a similar fashion Beirla nt et al (1996) find an optimal threshold by minimizing an approximate expression for asympt otic mean square error. The other methods PAGE 54 54 include plotting the quantile, shape or scale factor or any quantity of interest with respect to different thresholds and look for a stability in the curve (Bassi et al, Coles pp:8486) Parameter Estimation Generally one is interested to generate values (PDF) from a particul ar distribution given a set of parameter values which is expressed as: (/) fx (411) where, x is a data vector [ x1xm] of frequency. However, here we are faced with an inverse problem. Given the set of observed data and a mode l of interest, one is re quired to find one PDF that among all probability distri bution that is possibl e for the model is most likely to have produced the data. The parameters can be estima ted by several methods. The methods used in the literature include, maximum likelihood estima tion (MLE), method of moments, probability weighted moments, elemental percentile method, the least square regression method. Generally, maximum likelihood estimation method is widely used and accepted by researchers though it has limitations. That is as long as0.5 the asymptotic properties (Appendix B) of the maximum likelihood estimators are preserved. When 10.5 MLE are obtainable but do not have the standard asymptotic properties MLE is not obtainable when 1 (Coles, 2001, pp5455). The likelihood estimation approach for the GPD para meters is presented in the Appendix B. However, accurate estimation of the parameters is a huge research area in its elf. Beirlant et al (2004), Castillo et al (2005) discuss these methods in detail. Hasking and Wallis (1987) report a comparison between the maximum likelihood esti mation method, method of moments and the probability weighted moments. They conclude that method of moments and the probability weighted moments are more reliable compared to the MLE method unless the sample size is greater than 500. This secti on discusses the maximum like lihood method and least square regression approach in detail PAGE 55 55 Maximum Likelihood Estimation MLE is a popular statistical method that is used to make inferences about the parameters of the underlying probability distributi on of a given data set. There is nothing visual about the MLE method but it is very powerful and is precise for large samples. The likelihood of a set of data is the probability of obtaining that particular set of data, given the chosen probability distribution model. ML estimation starts with writing th e likelihood function which contains the unknown distribution parameters. The values of these pa rameters that maximize the likelihood function are called the maximum likelihood estimators. In order to estimate the parameters, the lik elihood function has to be constructed by reversing the roles of the data and parameter in Eq (411), i.e., (/)(/) Lxfx (412) Thus L represent the likeli hood of the parameter given the observed data x and is a function of For a model with k parameters, the likelihood function takes the shape of k dim geometrical surface sitting above a k dim hyperplane spanned by the parameter vector (Myung, 2003). Both f and L are PDFs but the major difference lies in the fact that they are defined in two different axes and are not directly comparable to each other. PDF f is a function of data given a particular set of parameter values define d on the data scale. The likelihood function L is the function of parameter given a par ticular set of observed data de fined on the parameter scale. ML approach is totally analyt ical in concept. MLE's genera lly have very desirable large sample properties (Appendix C) However, MLE ca nnot be adopted in all situations. There are some limitations when small numbers of failure s (less than 5, and sometimes less than 10 is small), MLE's can be heavily biased and the la rge sample asymptotic properties do not apply PAGE 56 56 Least Square Regression The method of least squares assumes that the best fit curve of a given ty pe is the curve that has the minimal sum of the deviations squared ( least square error ) from a given set of data. The parameters are obtained by solving the following minimization problem 2 ()uN i iN M inFzP (413) The GPD CDF can be obtained by using Eq. (44). The empirical CDF are plotting positions which are computed as: ,i=1... 1ii PN N (414) where N is the total number of samples, Nu is the index corresponding to the threshold u and P is the plotting position. Since it is a minimum finding, it can be seen as an optimization technique similar to maximum likelihood estimate met hod where we try to maximize the likelihood function. Least square regression requires no or minimal distributional assumptions. Unlike MLE, there is no basis for testing hypothese s or constructing c onfidence intervals. Accuracy and Convergence Studies for the Quantile Estimation In order to understand the capabil ity of GPD to model the tails of the distribution, an error analysis in terms of modeling and sampling error was performed on the probability content estimation using tail modeling approach. A detailed discussion of this is presented in Appendix D. Based on the results, it can be concluded that the sampling error is high compared to the modeling error. If the sampling e rror is accounted for, tail modeli ng has a fairly good capacity to estimate extreme probabilities with good accuracy. In structural reliability studies, one is required to model the tails of the response which might follo w some standard statistical distributions. Normal and Lognorma l distributions are widely used to model physical quantities in structural reliability studies Therefore, an investigation wa s done to study the accuracy and PAGE 57 57 convergence capabilities of tail modeling approach to model the tails of normal and lognormal distribution which is discussed in detail in Appendix E. Then, the approach is demonstrated on a cantilever beam example and a tuned mass damper example to estimate the failure probability. The simulation studies (Appendix E) help us to conclude that based on the ta rget quantile value we wish to estimate, the threshold and num ber of exceedance data need to be chosen. In the context of structural analys is, the interest is to address the response function. The failure probability is based on the response. If we wish to design a system with a low failure probability of the order of 1e4 or higher, then we are required to estimate quantiles of 0.9999 or higher. So, if the response follows a distribution close to normal, and we wish to estimate a quantile of 0.9999, a quantile of 0.95 and 100 exceedance data would enable us to estimate the quantile under 3% error. On the other hand, if the res ponse follows more of a lognormal distribution (heavy tailed), then one is required to choose a higher threshold of the order of 0.99 or 0.995 with 100 or more exceedance data to be guaranteed to have a bias error of less than one percent and a RMSE error of less than 5 percent. However, it is not possible to come up with a table that explains the minimum exceedance data required for a level of quantile and required accuracy level because, in practical situations the distri bution of the response is not known. Moreover, the challenge is to develop methodol ogies that can estimate the inverse measure corresponding to high reliability, with limited samples. Cantilever Beam Example The beam example that was considered in Chapter 3 is treated here also. Though simulating the beam example is not very expensiv e, we pretend that we have limited computer resources and are restricted to use a maximu m of 500 simulations. We are interested in estimating the PSF for low failure probabilities with only 500 response data. Three different cases are discussed. Case 1 disc usses the estimation of PSF for th e strength failure mode using PAGE 58 58 the tail modeling approach at different threshold values. Similarly the PSF estimation for the deflection failure mode is investigated in Case 2 and the system failure case where both the failure modes are considered simultaneously is discussed in Case 3. Case 1 and Case 2 are presented in Appendix F. For all the cases, th e convergence of PSF with respect to different thresholds and accuracy of PSF at different num ber of samples are investigated. Two methods, namely the maximum likelihood method and the l east square regression methods are used to estimate the parameters. The results from these methods are compared. In the system failure mode, both the failure modes are considered simultaneously. The safety factor for each failure mode is evaluated. For each realization, the critical of the two safety factor is considered the system safety factor. Once, the system safety factor is obtained for 100 simulations, the tail modeling approach is car ried out and the corresponding system PSF is estimated. Here the design variables are w =2.6041; t =3.6746 and the allowable deflection D0=2.145. This combination of the variables allows equal contribution of the modes to the total failure. The contribution of the modes are:10.00099fP ;20.00117fP ;120.00016ffPP The convergence plots of PSF for the system failure case is presented in Figure 42. Figure 42 shows that the PSF estimated through parameters estimated by the regre ssion method is unstable in contrast to the PSF estimated through paramete rs from the ML method. This behavior was observed in the other two cases too. The PSF es timate at a threshold of 0.9 is 0.99 with a standard deviation of 0.003. Th e corresponding failure probability estimate is 0.0017 with a standard deviation of 0.0002 which is about 10% of the estimate. The number of samples used in the above studies is 500. The PSF is expected to conve rge to the actual va lue when the number of samples is increased. PAGE 59 59 Tuned MassDamper Example The tail modeling approach to estimate the fa ilure probability is demonstrated here with the help of a tuned massdamper example. Th e tuned massdamper problem in Figure 43 is taken from Chen et al (1999). The problem i nvolves the failure probab ility estimation of a damped single degree of freedom system with dynamic vibration absorber. Figure 43 illustrates the tuned damper system consisting of the single degree of freedom system and a dynamic vibration absorber to reduce the vibrations. Th e original system is externally excited by a harmonic force. The absorber serves to reduce th e vibration. The amplitude of vibration depends on m R M the mass ratio of the absorber to the original system the damping ratio of the original system 1 1 nr ratio of the natural frequency of the orig inal system to the excitation frequency 2 2 nr ratio of the natural frequency of th e absorber to the excitation frequency The amplitude of the original system y is normalized by the amplitude of its quasi static response and is a function of four variables expressed as: 2 2 2 2 222 2 222 112121121 1 111111 14 r y R rrrrrrrr (415) This example treats12 and rr as random variables. They follow a normal distribution N(1,0.025) and0.01, =0.01 R The normalized amplitude of the original system is plotted in PAGE 60 60 Figure 44. There are two peaks where the norm alized amplitude reached undesirable vibration levels. The corresponding contour plot is presen ted in Figure 45. The failure region is denoted by the red band. This failure region is an island. That is, the failure region has safe regions on either side of it. This introduces additional ch allenges of not being able to use analytical approaches like FORM because the failure regi on is not a half plane. The objective of the problem is to reduce the risk of the normalized amplitude being larger than a certain value. The limit state for this case is expressed as: 12120(,)(,) grryrry (416) where 0 y can be considered as the allowable level of vibration. When the limit state in Eq. (416) is greater than 0, failure is said to occur. Increasing or decreasing 0 y will help in decreasing or increasing the fa ilure probability respectively. 0 y =27 is considered for the discussion. The corresponding failu re probability with 1e7 samp le MCS is estimated to be 0.01031. The tail modeling approach with 500 samp les and 100 repetitions are used to study the convergence and accuracy estimates of PSF. The plot for PSF at various thre sholds is presented in Figure 46. From the plots, it seems that regression behaves better compared to the ML method. There is a discrepancy in the plot co rresponding to the ML method. The PSF estimated at 0.96 threshold is more accurate than the PSF estimate at a threshold of 0.98. In order to understand the behavior of the tail of the tune d mass damper, a log plot of the CDF with 1e5 samples is examined. The plot is presented in Figure 47. It is evident from the plot that there are two curvatures in the tail which are difficult to model. The GPD has to capture this shape with less exceedance data. This is the reason for the disc repancy in the ML plot, the shape of the tail at the area of interest (Sr=1) modeled by ML with a threshold of 0.96 was better than the tail model with a threshold of 0.98. PAGE 61 61 In order to further explore the tail models from each method, the tail model from each method is superimposed on each other and the compara tive plots are presented in Figure 47. It is clearly observed that the tail modeled based on ML approach denoted by the dashed line attempts to capture the second curvature and in the process introduces error in the PSF value corresponding to a failure probability level of 0. 01. On the other hand, the model obtained based on regression parameters represented by the solid line, approximates the tail in a linear fashion and is accurate compared to the tail based on ML method for a failure probability of 0.01. However, the ML method is expected to perform better when the failure probability to be estimated is low. When the failure probability is low, it becomes necessary to model the second curvature adequately to estimate the PSF with reasonable accuracy. The ML method can perform better than the regression approach in modeling the second curvature. Such types of CDF with double curvature might not be encountered in stru ctural applications often and solving this problem with less number of samples is very cha llenging. With more insight on the problem, it is possible to select the suitable parameter estimation method depe nding on the requirements. Here, if one needs to estimate PSF corresponding to fail ure probabilities lesser than 1e3, then the second curvature has to be modeled well and he nce ML method is a good choice. Else, for higher failure probabilities, regr ession can perform better. Alternate Tail Extrapolation Schemes In this section, simple tail extrapolation t echniques are developed to estimate the PSF for low target failure probability using MCS that is sufficient only to estimate the PSF for substantially higher failure probability (lower targ et reliability index). Failure probability can be transformed to reliabili ty index by using Eq. (23). The same transformation is applied here to the CDF of the reciprocal of inverse measure. Si nce we are modeling the uppe r tail, reciprocal of the inverse measure is of interest here. The tail portion of the resulting transformed CDF PAGE 62 62 (relationship between inverse measure and reliability index) is approximated by a linear polynomial in order to take adva ntage of fact that normally di stributed inverse measure is linearly related to the reliability index. Since this approximation will not suffice if the inverse measure follows distributions other than normal, a logarithmic transformation is then applied to the reliability index which tends to linearize th e tail of the transformed CDF. This tail is approximated using a quadratic polynomial. Since both these techniques use only data from the tail, another technique that appr oximates the relations hip between inverse measure and reliability index from the mean to the maximum data usin g a quadratic polynomial is developed. The three extrapolation techniques discusse d are described with the help of a figure. A data set of size 500 with a mean of 10 and variance 9 following a lognor mal distribution is used to illustrate the three techniques. Figure 48, A) shows the general re lationship between PSF reciprocal and 1Pf. Applying the transformation modifies the CDF as in B). If the PSF reciprocal follows a normal distribution, this curve will be linear. Therefore, the first technique approximates the tail of the curve in B) by a linear polynomial. In order to ta ke advantage of the rema ining data, half of the curve in B) is approximated by a quadratic polynomial. It is sufficient to consider only half of the data because of the symmetry in reliability in dex. The third technique applies a logarithmic transformation to the reliability index and approxi mates the tail of the curve shown in C) by a quadratic polynomial. In all the above three techniques, once the coefficients of the polynomial are obtained, the PSF reciprocal corresponding to any higher reliabil ity index can be found. It is to be noted that with 500 samples, it is possi ble to estimate only PSF reci procal at reliability levels less than 3. Classical tail models discussed in the previous chapter and the alternate PAGE 63 63 extrapolation techniques discusse d in this chapter al low predicting PSF recipr ocal at much higher reliability indices without a dditional reliability analysis. The alternative extrapolation techniques and classical tail mo deling techniques are conceptually the same. The major difference in perc eiving the two classes is that the classical tail modeling techniques model the CDF of PSF reci procal, whereas the extrapolation schemes approximates the trend of PSF reciprocal in terms of reliability index. PAGE 64 64 0 Tail part () Fy () FuuyzFailed region Safe region 1fP Figure 41. Tail modeling of F (u) using the threshold u. The region of y >0 is failure PAGE 65 65 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 0.95 0.96 0.97 0.98 0.99 1 1.01 1.02 1.03 1.04 1.05 Fu1/PSF LCI UCI Mean Median 10 20 30 40 50 60 70 80 90 100 NexA 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 0.9 0.935 0.97 1.005 1.04 1.075 1.11 1.145 1.18 1.215 1.25 Fu1/PSF LCI UCI Mean Median 10 20 30 40 50 60 70 80 90 100 NexB Figure 42 Convergence of PSF at different thre sholds. A) MLE B)Regre ssion. Cantilever beam system failure mode. 500 Samples. 100 repetitions PAGE 66 66 Figure 43. Tuned vibration absorber Figure 44. Normalized amplitude vs r1 and r2 r1 r2 n2, m n1, M Absorber Original system c1k1 y cos() F t k2 PAGE 67 67 Figure 45. Contour of the normalized amplitude PAGE 68 68 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 Fu1/PSF Actual PSF LCI UCI Mean Median 0 10 20 30 40 50 60 70 80 90 100 NexA 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 0.6 0.74 0.88 1.02 1.16 1.3 1.44 1.58 1.72 1.86 2 Fu1/PSF Actual PSF LCI UCI Mean Median 0 10 20 30 40 50 60 70 80 90 100 NexB Figure 46. Convergence of PSF at differen t thresholds. Tuned Mass Damper. A) MLE. B)Regression. 500 samples. 100 repetitions. LCI Lower Confidence Interval. UCIUpper Confidence Interval PAGE 69 69 Figure 47 Tuned Mass Damper. Comparison of ta il models using regression and ML method PAGE 70 70 0 0.2 0.4 0.6 0.8 1 10 10 10 100 PfSrA 0 0.2 0.4 0.6 0.8 1 0 1 2 3 = (1 Pf )SrB 0.4 0.5 0.6 0.7 0.8 0.9 1 0 1 2 log( )SrC Figure 48. Transformation of the CDF of PSF reciprocal ( Sr). A) CDF of Sr. B) Inverse Standard normal cumulative distribution function applied to the CDF. C ) Logarithmic transformation applied to the reliability index. PAGE 71 71 CHAPTER 5 MULTIPLE TAIL MODELS Multiple Tail Model A total of five methods to model the tails of data were discussed in the previous two sections. However, in practical applications wh ere there is no information on the form of the output distribution, there is no way to explor e which technique performs better. A single technique that performs well for a particular example might not perfor m the best for another example. In fact, a single method that performs the best at a part icular reliability index might not perform the best at a different reliability index for the same example, as will be shown in the numerical examples. In addition, there are no known straight forw ard ways to quantify the error in the estimates. In order to take advantage of all the techniqu es and still come up with the best prediction, we propose to apply the five techniques simultaneous ly and use the median of the five estimates. It is observed that the median is a robust esti mate than the mean. This method is demonstrated first on seven different statistic al distributions followed by tw o engineering examples. In all examples, it is observed that the median of the five estimates is at least the second best compared to estimates from individual techniques. More over, the range of the five estimates closely approximates the error in the median. Thus usi ng multiple tail models not only ensures to choose a good estimate and buys insurance against bad pr edictions but also provides an error of the estimate. Figures 5.1 to 5.5 show the vari ous tail modeling tec hniques for a lognormal distribution with a data mean of 10 and standard deviation of 3. In each case, the 1e5 sample curve and 500 sample curve are superimposed to highlight the extrapolation. The Sr is computed as discussed in the implementation in next section. Figure 5.4 shows the multiple tail model approach (MTM). In order to under stand to clearly see the range of the five methods, the region beyond the threshold is zoomed and presented in Figure 5.5. The range of the five methods at PAGE 72 72 three different reliability indices is marked by double headed arrows. Mean of these ranges computed over the entire extrapolation zone are capable of representing the error associated with the median estimate from MTM.A representative act ual error is presented in the plot. The QH is the median and overlaps with the MTM curve. Numerical Examples Standard Statistical Distributions The standard statistical distri butions that are discussed in this section are presented in Table 51. A data set with mean 10 and standard deviation 3 is considered. In order to compare the performance of the proposed method on diff erent distributions, a common target failure probability is used to find the capacity which is in turn is used to normalize the parameters. Implementation of this exercise is as follows: targettarget0.00135 (=3), =10, 0.3, =fPxCOVCOVx 1. Use x and COV to estimate parameters (0aand 0b) for each distribution from Table 51 These parameters can be used to generate random numbers. 2. Find capacity C as target1th fP quantile. Inverse cumulative di stribution functions (icdf) can be used for this. In the case of single parame ter distributions, add shif t factor to the C. 3. Find normalized mean using x x C We are interested in the safety factor reciprocal rS following different statistical distributions. The normalized mean is the mean of the rS. 4. Use x and COV to find new parameters (a andb) for all the distributions. 5. Generate N = 500 Latin hype rcube samples in [0,1] 6. Find rSat each sample using (a b) and inverse icdf 7. Use Eq (414) to estimate plotting positions 8. Plotting rSand P provides the empirical CDF ofrS. It is to be noted that the ordinate is (1Pf). That is, we are intere sted in the upper tail. As Pf decreases, 1Pf increases, reliability index increases, S decreases and rS increase. 9. A threshold of 0.9 is considered for classical tail modeling technique s and the parameters of GPD that approximates the conditional CDF beyond the threshold are estimated using Maximum Likelihood and least s quare regression approaches. 10. The transformations discussed in the previous section can be applied to the CDF of rSand rSat higher reliability indices can be es timated directly from the approximated relationships. PAGE 73 73 11. There are totally five estimates (steps 9 and 10) of rSdenoted by ind rSin the extrapolated regions. Record the median and range ofind rS.as 5m and 5Ra respectively. 12. Exact values of rScan be obtained using icdf functions. 13. Compute error in the estimate of individual methods eind as abs (rS ind rS). 14. Compute error in multiple tail model (MTM) estimate eMTM as abs(rS 5m) 15. Compute the ratio of mean of eMTM to mean of 5Ra over the entire ex trapolation zone and denote it as 16. Repeat steps 5 to 15, 1000 times for each distribution. 17. Record the median, mean of eind and eMTM over 1000 repetition as 1000 indm, 1000 indand 1000 MTMm,1000 MTM 18. Record the median and mean of the ra tio in step 15 over 1000 repetition as 1000 m For each distribution, boxplots of eind and eMTM are presented. 1000 indmand 1000 MTMmcompared at three reliability indices between 3 and 4. 2. The results for the normal distribution are presented in Figure 56. The objective of this plot is to see how well the 5m estimate compares to the ind rS. For 1000 repetitions, only the median and the mean can be compared because the range is volatile. It was observed that the median was a robust estimate compared to the mean. It is evident that the median is less sensitive to the outliers compared to the mean. At =3, the regression (Rg) performs the best followed by the linear tail (LT). The MTM performs as good as the second best. At =3.6, the LT performs the best followed by the regression (Rg) and MTM performs as well as LT. At =4.2, the LT performs the best followed by the quadratic tail (QT) and MTMs performance is as good as LT. It is to be noted that the individual techniques that provided the best estimate differed at di fferent reliability indices. However, the MTM estimate was consistent and it pe rformed at least as good as the second best estimate from the individual techniques. This approach serves as an insurance against bad predictions, if we were to use a single technique. PAGE 74 74 Box plots of ratio are presented to see how well the range of the methods compares to the mean of the eMTM over the entire extrapolation region. The results for the normal distribution are presented in Figure 57. It is clearly observed that the ratio of the mean of the error to the mean of the range in the entire extrapolation region is around 1. If the ratio is 1, then the range captures the error in the median perfectly. Here, the range provides an estimate of eMTM that might slight over or under estimate the actual er ror. The boxplots for the remaining distributions are presented in the Appendix G. Table 52 pr ovides a summary of the performance of the individual methods and the MTM fo r all the distributions. Table 52 shows that not a particular technique was the best for all the distributions and the technique that provided the best estimate for a particular distribution va ried as the reliability indices changed. However, the MTM estimate was mostly close to the 2nd best available estimate. The summary of ratio is presented in Table53. It is observed that other than unifo rm distribution, the ratio varies between 0.15 to Table 51. 7 standard statistical distributions and their parameters Distribution Parameters a b Normal x LogNormal 2ln0.5 x b 2ln1x Gamma 2 x a Extreme Type 1 0.577x b 6 Uniform 12 2x 12 2x Single parameter distributions Exponential x x Rayleigh 2 x b 2 2 PAGE 75 75 0.35. This means that half of the range over estimates the error by a factor of 1.5 to 3. In order to provide an insight what number should one expect for the ratio, a simple exercise is performed. Standard normal random numbers of size 5 (rep resents 5 estimates at 7 different reliability indices) are generated and is computed over 1e5 repetition. The resultant number is 0.18. Table 52. Summary of the performance of individual methods and MTM for different distributions Rel Index 3 3.6 4.2 Performance 1st best 2nd best MTM1st best2nd bestMTM1st best 2nd bestMTM Distributions Normal Rg LT 2nd LT QH 1st LT QT 1st Lognormal Rg LT 1st QH Rg 3rd QH QT 3rd Gamma Rg LT 2nd LT Rg 1st QT LT 2nd Ev QH Rg 1st QH QT 1st QH QT 2nd Uniform QT Rg 1st Rg ML 2nd Rg ML 2nd Rayleigh Rg LT 2nd LT Rg 1st LT QT 1st Exponential LT Rg 2nd Rg QT 1st Rg QH 3rd Table 53. Summary of the Error to ra nge ratio for different distributions Distribution 1000 m Normal 0.279 Lognormal 0.349 Gamma 0.181 Ev 0.183 Uniform 0.003 Rayleigh 0.148 Exponential 0.241 Since the concept of regression is used to es timate the parameters in all methods other than the ML approach, the residuals are monito red for each method to see whether they follow the normal distribution as per classical statistical theory. It is expected that the type of distribution of the residuals can provide a good idea about the quali ty of the fit. A similar study was made for the normal dist ribution case and the normal QQ plots for the residuals are presented in Figure 58. It is observed that the residuals of none of the methods follow a normal distribution. Investigating th e corresponding histograms reveal that they follow a log normal PAGE 76 76 distribution. In these situati ons, other regression schemes like weighted approaches can be adopted depending on the knowledge of the histogra ms and QQ plots of the residual which is scope for future work. Engineering Examples Application of multiple tail models for re liability estimation of a cantilever beam The cantilevered beam example treated in Chap ter 3 is considered here. The objective of the original problem is to minimize the weight or, equivalently, the cross sectional area, A=w.t subject to two reliability constraints, which require the reliability indices for strength and deflection constraints to be larger than three. Eqs (317) and (318) are rewritten as sYX S wtwt G RR (51) d OOLYX Ewt tw D G DD (52) It is to be noted that the pe rformance functions are expressed in a fashion such that failure occurs when s Gor dGis greater than 1. Here, we consider a system failure case. That is, both the failure modes are considered simulta neously. The optimal design variables w=2.6041, t=3.6746 is taken from Qu and Haftka (2003) for a system reliability case. The corresponding reliability index is 3. The allowable deflection Do is taken to be 2.145. For each sample the critical safety ratio reciprocal (maximum of the two) is computed. Since we are dealing with rS, it is to be noted that n=1Pftarget. The conditional CDF of rScan be approximated by classical techniques and the relationship between rSand reliability index be approxima ted by three different alternate techniques. Finally, using the MTM approach wi ll give a better prediction compared to any individual method. This procedur e is repeated 1000 times and the results are presented in Figure PAGE 77 77 59 It is observed that the LT pe rforms the best at all three reliability indices followed by the QH as the second best. MTM consistently perf ormed close to the second best estimate. Figure 510 presents the plot of ratio. The ratio is 0.23 and thus strengthens our conclusion that the half of the range approximates the error in the median by a factor between 1.5 and 3, which is close to 2 here. The main advantages of the techniques discussed are that no statistical information is required and huge re duction in computational power. In order to understand the sacrifice in accuracy for the re duction in computational power a MCS study is performed. 100 repetitions of rSestimates with 5E5 samples a nd the corresponding standard deviation are computed and presented in Table 54. At a reliability index of 4.2, the standard deviation in rS estimate is 0.04 which is the same error from MTM using 500 samples. Therefore, for a same level of accuracy, the redu ction in computational pow er is about 3 orders of magnitude (5E5 to 5E2). Table 54. PSF reciprocal estimates and standa rd deviation at differe nt reliability indices Rel Index 3 3.2 3.4 3.6 3.8 4 4.2 rS 1.012 1.032 1.05 1.071.091.121.13 SD 0.003 0.004 0.01 0.010.010.020.04 Application of multiple tail models for re liability estimation of a composite panel The design of the wall of a hydrogen tank whic h is a composite laminate operating in cryogenic temperatures addressed by Qu et al., (2003) is consid ered as a second engineering example. In order to reduce matrix crack ing and hence hydrogen leakage in cryogenic environments, the laminates need to have smalle r angle between the plies. However, these types of laminates do not carry the bi axial hoop and axial stresses due to pressure. Qu et al., (2003) investigated options for minimizing the increase in thickness required to carry both mechanical PAGE 78 78 and thermal residual strains. The geometry and lo ading conditions of the problem is presented in Figure 511. The laminate is subject to mechanical (Nx and Ny) load and thermal loading due to operating temperature 423oF where the stress free temperature is 300oF. The objective of the actual problem was to optimize the weight of laminate with two ply angles 12, The ply angles and ply thickness (12 and tt) are the design variables. The material used is IM600/133 graphite epoxy of ply thickness 0.005 inch. Qu et al., (2003) performed the deterministic design optimization using continuous design variables. In order to account fo r the uncertainties and make the deterministic optimal design comparable to probabilistic optimal design, they used a safety factor of 1.4. For further details, the reader is referred to Qu et al., (2003). The mechanical properties are presen ted in Table 55. The deterministic optimization problem is formulated as 12 12 1112221212Minimize 4() Such that ,0.005 L ULUU FFFhtt tt SSS (53) All material properties are assumed to be random, uncorrelated and follow normal distributions. The coefficient of va riation is presented in Table 56. Here, thickness is in inches and h is the laminate thickness. Superscripts U and L represent upper and lower limits of the associated quantities. 1212,, and are the ply strains along fiber di rection, transverse to fiber direction, and shear strain, respectively. 21212,, and EG are functions of temperature. Qu et al (2001) show that at 20.2K (423 F) the thickness of the optim um laminates obtained by using temperature dependent material properties is 80% less than th at using constant material properties at 298 K (77 F ). Qu et al (2003) observed that 2 U is the active constraint. For a feasible design, 21 different temperat ures uniformly distributed between 20oK and 300oK is PAGE 79 79 considered and the strain constrai nts are applied to these temperat ures. The mean is computed at a particular temperature and random numbers gene rated based on coefficient of variation. The mean for other parameters are presented in Tabl e 57. The deterministic optimal design obtained by Qu et al (2003) is presented in Table 58. The transverse strain on the first ply (direction 2) is the critical st rain. The limit state is defined as the difference between the critical strain and the uppe r bound of the allowable strain. 21 UG (54) Failure is said to occur if Eq. (54)> 0. In order to extend this example for demonstrating the multiple tail models approach, the second optimal design in Table 58 is considered. The loads are increased by a factor of 1.1 (Nx 5280 lb/inch and Ny 2640 lb/inch) so that the failure probability is of the order of 1e3. 500 random critical strain values are genera ted. Similar to the previous examples, 5 different techniques are used to f it the tail of the response distri bution. The results are presented in Figures 512 and 513. As observed in other ex amples the MTM compared at least close to the second best in estimates from individual methods However, the ratio of mean error to mean range is 0.22 which means that half of the ra nge overestimates the error by a factor of around 2.which falls within the range of 1.5 and 3 that was observed earlier in section. The 1e7 sample based PSF reciprocal estimation and the correspo nding standard deviation is presented in Table 59. At reliability index of 4.2, it is seen that the standard deviation is 0.0063 and the multiple tail model provides an estimate with an error of 0.05 which is off by a factor of 8. The number of samples used for this estimation is 500, which is about 4 orders of magnitude less than 1e7 Due to the prohibitive computational power required, an exercise similar to the previous example is PAGE 80 80 not performed here. Instead the estimate from 500 samples and 1e7 samples are compared directly. Table 55. Mechanical properties of the composite laminates Elastic properties 121212,, and EEG Coefficients of thermal expansion 12 and Stressfree temperature zeroT Failure strains 112212,,, and L ULUU Safety factor F S Table 56: Coefficient of varia tion for random material properties 121212,,,EEG 12, zeroT 11, L U 2212,, L UU 0.035 0.035 0.03 0.06 0.09 Table 57: Mean of random variables 1E 12 zeroT 1 L 1 U 2 L 2 U 12 U 21.5x106 0.359 300 0.0109 0.0103 0.013 0.0154 0.0138 Table 58: Deterministic optim a found by Qu et al (2003) 1 (deg) 2 (deg) t1 (in) t2 (in) h (in) 0.00 28.16 0.005 0.02 0.1 27.04 27.04 0.01 0.015 0.1 25.16 27.31 0.005 0.020 0.1 Table 59: Composite laminate. rSreciprocal estimates and stan dard deviation at different reliability indices Rel Index 3 3.2 3.43.63.844.2 1E+07 1.0043 1.0304 1.0579 1.0862 1.1157 1.1480 1.1824 SD 0.0007 0.0009 0.0015 0.0021 0.0025 0.0046 0.0063 PAGE 81 81 Figure 51.Lognormal distribution. Classical ta il modeling techniques MLE and Regression PAGE 82 82 Figure 52. Lognormal distri bution. Linear fit to ta il (LT) and Quadratic fit to half of the data (QH) PAGE 83 83 Figure 53. Lognormal dist ribution. Quadratic fi t to the tail (QT) PAGE 84 84 Figure 54.Lognormal distribution. Multiple tail models PAGE 85 85 Figure 55 Lognormal distribution. Multip le tail modelsextrapolation region PAGE 86 86 Med= Median. ML=Maximum Likelihood, RgRegression, QT Quadratic fit to the Tail data between rSand ln( ), LT Linear fit to the Tail data between rSand QH Quadratic fit to Ha lf of the data between rS and Dashed line inside the box depicts th e mean and the solid line, the median. Figure 56. Normal distribution. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices PAGE 87 87 MTM 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 1000m = 0.279 1000= 0.254 Figure 57: Normal dist ribution. Boxplot of ratio. PAGE 88 88 Med= Median. ML=maximum likelihood, RgRegression QT Quadratic fit to the Tail data between rSand ln( ), LT Linear fit to the Tail data between rSand QH Quadratic fit to Ha lf of the data between rS and Dashed line inside the box depicts th e mean and the solid line, the median. Figure 58. Cantilever Beam. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices PAGE 89 89 Figure 59. Normal QQ plot of the residuals for the normal distribution PAGE 90 90 MTM 0 0.5 1 1.5 2 2.5 3 3.5 1000m = 0.231 1000= 0.374 Figure 510. Cantilever beam. Boxplot of ratio. NX 2 X Y 1 2 1 NY NX 2 X Y 1 2 1 NY Figure 511. Geometry and loading for the co mposite laminate. Xhoop direction, Yaxial direction PAGE 91 91 ML=Maximum Likelihood, RgRegression, QT Quadratic fit to the Tail data between rSand ln( ), LT Linear fit to the Tail data between rSand QH Quadratic fit to half of the data between rS and .Dashed line inside the bo x depicts the mean and th e solid line, the median Figure 512. Composite laminate of cryogenic tank. 1000 repetitions. Boxplot comparison of eind and eMTM at different reliability indices... PAGE 92 92 MTM 0 0.5 1 1.5 2 2.5 3 1000m = 0.220 1000= 0.356 Figure 513. Composite laminate of cryogenic tank. Boxplot of ratio. PAGE 93 93 CHAPTER 6 CONCLUSIONS High reliability estimations with expensive co mputer models pose a challenge because of the high computational expense. Extrapolating in to higher reliability indices with information from lower reliability indices can be one solution to this problem. In order to do this, the tails of the distribution are to be modele d well. Inverse reliability measur es feature several advantages and is widely used by researchers in structural reliability. This work explores the usefulness of inverse measures in tail modeling. Two classical tail modeling techniques were discussed.Three extrapolation techniques that can complement the classical tail modeling techniques were developed. This work proposes to use a multiple tail model approach in which all the five techniques are applied simultaneously to estimate the inverse measure, and the median is taken to be the best estimate. It is shown that the range of the five methods can be utilized to approximate the error associated with the es timate. The proposed method was demonstrated on seven statistical distributions and engineering examples. It was shown that the sacrifice in accuracy compared to the crude MCS is relative ly small compared to the immense reduction in computational power. PAGE 94 94 APPENDIX A PSF ESTIMATION FOR A SYSTEM RELIABILITY The estimation of PSF for a system reliability case is demonstrated with an example. This example is artificial and is an illustration of a case with multiple limit state functions. The example is presented in Figure A1 .The limit state functions are: 32 115 40 xx Gy (A1) 22 2(5)(12) 0.8 30140 xyxy G (A2) 380 20Gy x (A3) The limit states are functions of two random variables x and y. The distributions of the random variables are presented in Table A1. The exampl e has been created in such a fashion that each limit state and its intersection with other limit stat es contributes significantly to the total failure. It should be noted that non lin ear functions are used to empha size the advantage of using MCS and a PSFbased approach to estimate failure probability. Using FORM in these circumstances might introduce huge error due to linearizing the limit st ate function at the most probable point of failure. The contributions of failure modes to the total failure are listed in Table A2. The system PSF can be estimated as described in th e flowchart in Figure A1. The limit states are functions of two random variab les x and y. The PSF estimated for different target failure probabilities are listed in Table A3. For a target failure probabil ity of 0.557 which is the actual failure probability as listed in Table A3, the PSF is estimated as 1. When a lower target failure probability of 0.5 is used, the PSF is estimated to be 0.844 indicating that the system is unsafe and when a larger target failure probability of 0.6 is used, the PSF is found to be 1.12 showing that the system is over safe. This shows that PS F based approach can adequately address system reliability case accounting for the interactions between the modes. PAGE 95 95 Table A1. Distribution of random vari ables. Multiple failure modes example Table A2. Contribution of failure modes to the system failure Limit states Failure probability % Contribution to total failure 1 0.432 77.55 2 0.468 84.02 3 0.336 60.32 1 & 2 0.386 69.29 1 & 3 0.253 45.42 2 & 3 0.249 44.71 1,2 & 3 0.209 37.52 Total 0.557 100 Table A3. PSF at different target failure probabilities Target Pf PSF 0.557 0.9996 0.5 0.8444 0.6 1.1191 Figure A1. Two variable system. Algebr aic multiple limit state functions example Random Variable x y Distribution Normal (4,0.5) Normal (2,3) PAGE 96 96 APPENDIX B GENERALIZED EXTREME VALUE THEORY Let X1, X2,be a sequence of independent random variables with a common distribution function F, and let 1 1 1max,..., PrPr,... = Pr...Pr = ()nn nn n nMXX MzXzXz X zXz Fz (B1) Generally, the distribution function of n M can be derived exactly for all values of n. But here F is not known, so it does not help. Here we look for approximate families of models for nFwhich can be estimated based on extreme data only. This is done as follows: nn n n M b M a (B2) for sequences of constants 0na and nb. Appropriate choices of constants stabilize the location and scale of *n M as n increases. Essentially we seek limit distributions for *n M with appropriate choices of constants. The all possible range of limit distributions for *n M is given by the following theorem, the extremal type theo rem, a product of FisherTippet (1928) work. If there exists sequences of constants 0na and nbsuch that Pr()/() as nnnMbazHzn (B3) for a non degenerate distribution function H, then H belong to one of the following families: 1. ()expexp, zb Hzz a (B4) 2. 0 zb () exp, Hz zb zb a (B5) PAGE 97 97 3. exp, () 1 zb, zb zb Hz a (B6) Here, a and b are the scale and location parameters. is the shape parameter. These three classes of distributions are called the extreme value di stributions with types I, II and III, widely known as Gumbel, Frechet and Weibull families respectively. The a bove theorem implies that the normalized maxima have a limiting distribution that must be one of the three types of extreme value distribution. The three families of distributions can be combined into a single family of models having distribution function of the form 1/()exp1 z Hz (B7) such that :1()/0 zz, where the parameters satisfy 0 and This is the generalized extreme valu e (GEV) family of distributions. Here, is the location parameter, is the scale parameter, is the shape parameter. Type II and III families correspond to 0 and 0 cases in (9). 0 can be interpreted as the Gumbel family with distribution function: ()expexp z Hz (B8) PAGE 98 98 APPENDIX C MAXIMUM LIKELIHOOD ESTIMATION The PDF of the Pareto Di stribution is given by: 1 11 1 0 () 1 = 0i zz fx e (C1) Here, the z represents the exceedances. and are the shape and scale parameters respectively. The likelihood functi on for a sample of z1 zn of iid GP random variables is given by: 1 1 1 1 11 1 0 ,, 1 = 0n i n i i n n iz Lz e (C2) The log likelihood function is expressed as: 1 11 log1log1 1+0 log(,) 1 log = 0n i i i n i iz nz L nz (C3) The maximization of (C3) is best made by reparametrization: ,, with yielding 11 log,log1log1n i ihLnz (C4) 11 loglog1log1n i innz (C5) The max of (C5) is found as the minimum of the negative of(C5). The MLE are given by differentiating (C5) and equating to zero. PAGE 99 99 2 11 log(1)0 n i ihn z (C6) 111 10 1n i i ihn z z (C7) (C6) yields 11 log(1)n i iz n and (C7)can be solved for Numerical techniques are used for the maximization of h in (C5) using algorithms that are availa ble in standard software packages. In most situations estimation is done using iid samples. In such cases, the interst is to determine the behavior of the estimator as a sample size This is referred to as asymptotic behavior. Under certain regularit y conditions like: the first and second derivatives of the log likelihood function must be defined., the fisher information matrix must not be equal to zero, MLEs exhibit following properties Asymptotically unbiased (i.e.,) its bias tends to zero as the number of samples increases to infinity. MLE is asymptotically guassian. That is, as sample increases the distribution of MLE tends to normal distribution. Here, the covarian ce matrix can be obtai ned as the inverse of the fisher information matrix. PAGE 100 100 APPENDIX D ERROR ANALYSIS FOR GPD This appendix describes the error analysis for the probability content estimated using tail modeling. When the distribution of exceedance is approximated using GPD, there are two sources of error: 1. Modelling error 2. Sampling component of the total error The modelling error refers to the capacity of the GPD model to accurately represent the distribution of exceedance data or the tail of th e CDF. If the exact CDF is available, the CDF estimated through GPD can be compared and th e resulting error is the modeling error. Depending on the empirical CDF, the GP fit will vary. This introduces sampling error. The sampling error comes from the fact that the GPD parameters vary depending on the exceedance. The input variables generally come from samp ling distributions like the LHS or any other suitable DOE. Depending on these sampling points, the threshold differs. Hence, the exceedance is affected and therefore the estimated para meters and eventually the fitted GPD. If we have the privilege to generate large number of samples and use a fraction of it as exceedance data to fit a GPD, the estimated para meters in the process can be considered as optimal parameters since they are based on a large sample size. These parameters can be used to estimate probability contents at certain leve ls and can be compared with the probability content estimated by a GPD with small sample si ze. This difference will dictate the sampling error contribution to the total error. In an attempt to study the modeling and samp ling error, an example problem is treated here. The example under investigation is Y =exp( x) and x~N(0,1). The modeling error is studied comparing the estimated probability content using a GPD with a large sample size and the exact PAGE 101 101 probability content generated by analytical expression of the lognormal distribution. The parameters for the GPD are obtained by MLE method. Let the conditional CDF/ PYuzYu be represented by()uFz In this example, it is possible to estimate this quantity exactly. Using extreme value theory for exceedances, ()uFz can be approximated by a GP distribution denoted by, () Fz. Let the estimated CDF be denoted by$ () Fz. Let the GPD which is based on the optimal parameters be denoted by* () Fz. 1e6 samples based on LHS design are generated. The Ys are sorted as:(1)(2)(16)....eYYY This is called the order statistics of Y. Choosing the appropriate u, the threshold value is very critical. u defines the beginning of the upper tail. The uF is chosen to be 0.90 and the corresponding value of Y is chosen as u. Once u is chosen, the exceedance data is formed and the ML estimates are obtained for the parameters. These estimates are used to estimate the probability content at different levels of Y. The results are presented in Table D1. Five repetitions of 1e6 samples are used for each level. The mean and standard deviation of the CDF estimates based on GPD are presented. The error refers to the total error in percentage. The sampling error component of the total error is ex tracted using estimated optimal parameters. The optimal parameters were obtained based on a million samples. Here, 3 percent of the data was considered as exceedance. The results are presen ted in Table D2. From Tables D1 and D2, it can be clearly seen that the sampling error is high compared to the modeling error. So, if we can account for the sampling errors, tail modeling has a fairly good capacity to model the tails of the distribution to estimate probability contents. PAGE 102 102 TableD1. Modelling error (()gFz * () Fz). The standard deviation is very low that it is not presented. TableD2: Sampling error (* () Fz$ () Fz) Y Exact g F z GPD* () F z Error% 4 0.08283 0.07988 3.56 5 0.05376 0.05299 1.43 6 0.03659 0.03647 0.33 7 0.02583 0.02587 0.16 8 0.01879 0.01884 0.28 9 0.01400 0.01403 0.19 10 0.01065 0.01065 0.00 11 0.00824 0.00822 0.24 12 0.00648 0.00645 0.49 Y GPD1e6 samples* () F z GPD1e2 samples$ () Fz Error % 4 0.07988 0.08710 9.04 5 0.05299 0.05797 9.39 6 0.03647 0.03997 9.61 7 0.02587 0.02824 9.15 8 0.01884 0.02039 8.20 9 0.01403 0.01503 7.17 10 0.01065 0.01135 6.53 11 0.00822 0.00877 6.69 12 0.00645 0.00696 7.95 PAGE 103 103 APPENDIX E ACCURACY AND CONVERGENCE STUD IES FOR NORMAL AND LOGNORMAL DISTRIBUTIONS The quantile of a probability distribution can be estimated using Eq.(48). The quantile2 can be related to the inverse m easure in the structural reliability context where the probability content is expressed as the targ et failure probability. Due to th e various advantages of using inverse measures in RBDO, we are interested in estimating them using tail modeling approach. Here, we study the accuracy of the tail model to capture the quantile in the tails of the normal distribution. In the following simulation study a y=x example where x ~ N (0, 1) is treated. The error metrics that are monitored are: a) Bias error: Bias()=E p pp x xx (E1) 11Ns p p i x x Nsim (E2) where, : Estimated quantile : Actual quantile : Number of simulationsp px x Nsim b) RMSE error: 22 11 ()()Ns pppp ixxxx Ns (E3) Here, we set out to study the effects of the threshold value, the number of exceedance data on the accuracy of the quantile estimate. For each threshold value, four different exceedances are tested. Since the exceedances vary, the number of samples used in each case varies. The results are presented in Table E1.The overall capacity of the tail modeling approach to capture the quantiles in the tails of the norma l distribution seems to be pretty good. For all the four threshold 2 A p quantile of a random variable X is any value x such that Pr(X xp) = p PAGE 104 104 values, the % Bias error for the 25 exceedance data seem to be very less compared to higher exceedance data. But, it is to be noted that the corresponding %RMSE errors are high emphasizing the fact that the estimates from low number of exceedance data could be highly unreliable. 25 exceedance data is considered low to estimate quantiles. 50 to 100 data is considered realistic and 200 ideal It can be clearly observed th at the accuracy of the % Bias increases as the threshold increases for the Table E1. Accuracy of the tail modeling approa ch to estimate quantiles. Normal distribution y = x. x~N(0,1). 500 simulations. ML estimated parameters % Bias % RMSE Fu Nex NS 0.999 0.9999 0.99999 0.999 0.9999 0.99999 25 125 8 9 5 48 98 216 50 250 6 10 14 22 34 46 100 500 3 7 11 11 16 22 0.8 200 1000 2.5 7.3 12.5 5.1 8.5 12.0 25 166 6 7 2 37 81 211 50 333 4 8 11 17 28 44 100 666 2 6 10 8 13 18 0.85 200 1333 2.1 6.1 10.6 4.1 6.9 9.8 25 250 3 2 3 25 51 107 50 500 4 7 10 11 18 25 100 1000 2.2 5.7 9.4 5.4 9.3 13.1 0.9 200 2000 0.8 3.5 6.8 2.6 4.8 7.1 25 500 3 3 0 12 26 57 50 1000 2 4 5 6 11 17 100 2000 0.8 2.5 4.5 2.8 5.5 8.5 0.95 200 4000 0.4 2.4 5.2 1.3 2.6 4.0 Fu Threshold in the CDF. Nex Nu mber of exceedance data. NSNumber of total samples used. 0.999, 0.9999, 0.99999 Quantiles required to be estimated. For a threshold of 0.85 and Nex=100, the 2,6,10 % Bias erro r in quantile translates to 27.6, 139.9 and 524.12% error in probability content same number of exceedance data. The threshold va lues, number of exceedance data is dependent on the level of target quantile we wish to estimate It can be observed that for a quantile of 0.999, reasonably low quantiles such as 0.8 perform we ll enough with 50 to 100 exceedance data, if 5% bias error is allowable. For a 0.9999 quantile, a threshold of 0.9 with 100 exceedance data PAGE 105 105 performs fairly well. A 0.95 threshold is cap able of estimating a 0.99999 quantile well with about 100 data. But it can be observed that there is oscillation in the error values as the 100 exceedance data seem to perform better than the 200 exceedance data in % bias error. A higher threshold value is expected to rectify this oscillation. Since normal distribution is the most encountered distribution in structural analysis, it is good news that the tail modeling approach can capture the quantiles in the tail of the normal distribution fairly well. The next distribution that is treated he re is a log normal distribution which has a heavy tail and is generally used to model loads in structural anal ysis. The simulation procedure is same as above. The results are provided in Table E2. Comparison of Tables E1 a nd E2 allows us to conclude Table E2. Accuracy of the tail modeling approa ch to estimate quantiles. LogNormal distribution y = x. x~logN(0,1). 500 simulations. ML estimated parameters % Bias % RMSE Fu Nex NS 0.999 0.9999 0.99999 0.999 0.9999 0.99999 25 125 18 110 527 228 1214 8065 50 250 8 42 126 77 199 517 100 500 7 29 79 37 91 229 0.8 200 1000 3.8 17.4 44.0 16.6 36.4 75.1 25 166 18 170 1517 256 2848 40477 50 333 4 37 152 67 270 1410 100 666 6 26 70 28 68 159 0.85 200 1333 1.8 12.0 32.9 12.3 26.1 51.1 25 250 5 71 439 101 632 5683 50 500 6 38 136 44 152 590 100 1000 1.1 14.2 43.7 16.8 40.3 92.4 0.9 200 2000 1.5 10.5 29.4 8.6 18.6 36.4 25 500 2 38 211 41 196 1363 50 1000 1 20 82 19 71 358 100 2000 1.4 12.4 39.0 9.3 24.2 56.2 0.95 200 4000 0.1 5.6 19.6 4.2 10.2 20.8 Fu Threshold in the CDF. Nex Nu mber of exceedance data. NSNumber of total samples used.0.999, 0.9999, 0.99999 Quantiles required to be estimated. For a 0.95 threshold and Nex=10 0, the 1.4, 12.4 and 39% Bias error in qu antile translates to 4.5, 37.5 and 78.3% error in probability content PAGE 106 106 that estimating quantiles in lognor mal distribution is more challe nging than normal distribution. For a 0.999 quantile, 0.8 and 0.85 thresholds perf orm fairly well with 200 samples. Above a 0.9 threshold, 100 samples seem to render very good results. There is considerable difficulty in estimating 0.9999 quantile with even a 100 exceedan ce data and a 0.95 threshold value. Needless to mention, trying to estimate 0.99999 quantile de teriorates the accuracy of the estimation only. In order to test whether higher threshold values enable us to estimate the extreme quantiles accurately, simulation is conducted for threshold va lues ranging from 0.96 to 0.99 and results are presented in Table E3. Table E3. Accuracy of the tail modeling appr oach to estimate quantiles. Higher thresholds. LogNormal distribution y = x. x~logN(0,1). 500 simulations. ML estimated parameters % Bias % RMSE Fu Nex NS 0.99990.999990.99990.99999 25 625 50 254 171 958 50 1250 16 61 46 160 100 2500 9 30 18 42 0.96 200 5000 4 16 8 16 25 834 20 135 102 693 50 1667 12 49 35 98 100 3334 7 29 16 43 0.97 200 6667 3 13 6 12 25 1250 17 126 74 563 50 2500 6 29 19 49 100 5000 4 20 10 22 0.98 200 10000 1 8 4 8 25 2500 5 65 30 225 50 5000 2 19 10 25 100 10000 0.9 10 5 12 0.99 200 20000 0.2 6 2 5 Fu Threshold in the CDF. Nex Nu mber of exceedance data. NSNumber of total samples used. 0.9999, 0.99999 Quantiles required to be estimated Results from Table 43 clearly show that as the threshold is increased, the extreme quantiles are also captured fairly well. With a 0.99 threshol d and 100 exceedance data, we can achieve less PAGE 107 107 than one percent error in bias for estimating 0.9999 quantil e. This requires 10000 samples. Similarly the accuracy in estimating 0.99999 quantil e is also better with 0.99 threshold and 200 exceedance data. A more accurate estimate can be e xpected if we raise the threshold to 0.999 or higher. PAGE 108 108 APPENDIX F ACCURACY AND CONVERGENCE STUDIES FOR THE CANTILEVER BEAM Stress Failure Mode The stress failure mode for the cantilever beam discussed in Chapter 3 is considered here. The optimal design variables w=2.4526; t=3.8884 for a target failure probability of 0.00125 (1e7 Samples) are adopted from (Qu and Haftka, 2004). The corresponding value of PSF for this design is 1. For the optimal combination of the design variables, 500 samples based on the distribution of the random vari ables are generated and the PSF is estimated for different thresholds as discussed in Chapter 3. This pr ocedure is repeated 100 times and the mean, 5%, 95% confidence intervals are estimated. The resu lts are presented in Figure E1. It can be observed from the figure that the PSF estimated by regression approach is very unstable whereas the median of PSF estimated by the ML approach is relatively stable. There is instability towards higher threshold like 0.98 but the corresponding ex ceedance data are very less, which explains the instability. The estimated mean of the PSF in the ML approach is close to 0.99 which is about 1% bias error in estimating th e actual value of the PSF. The standard deviation of the PSF estimate at a threshold of 0.9 with ML approach is about 0.003 which is 0.3% of the estimate. The corresponding estimate of failure probability is 0.0011 with a standa rd deviation of 0.0001 which is about 10% of the estimate. Deflection Failure Mode The allowable deflection is chosen to be D0=2.25. Similar to the stress failure case, the PSF is estimated at different thresholds for the optimal combination of the design variables. The results are presented in Figure E2. At a threshold of 0.9 and ML approach, the PSF is estimated to be 0.99 with a standard deviation of 0.004. The corresponding failure probability estimate is 0.0013 with a standard deviation of 0.0001 wh ich is about 8% of the estimate itself. PAGE 109 109 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 0.94 0.952 0.964 0.976 0.988 1 1.012 1.024 1.036 1.048 1.06 Fu1/PSF LCI UCI Mean Median 0 10 20 30 40 50 60 70 80 90 100 NexA 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 0.9 0.935 0.97 1.005 1.04 1.075 1.11 1.145 1.18 1.215 1.25 Fu1/PSF LCI UCI Mean Median 0 10 20 30 40 50 60 70 80 90 100 NexB Figure F1. Convergence of PSF at different thresholds. A) MLE B) Regression. Cantilever beam stress failure mode. 500 Samples. 100 repetitions PAGE 110 110 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 0.92 0.934 0.948 0.962 0.976 0.99 1.004 1.018 1.032 1.046 1.06 Fu1/PSF LCI UCI Mean Median 10 20 30 40 50 60 70 80 90 100 NexA 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 0.96 0.98 0.9 0.95 1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 Fu1/PSF LCI UCI Mean Median 10 20 30 40 50 60 70 80 90 100 NexB Figure F2. Convergence of PSF at different thresholds. A) MLE B) Regression. Cantilever beam deflection failure mode. 500 Samples. 100 repetition PAGE 111 111 APPENDIX G STATISTICAL DISTRIBUTIONS ERROR PLOTS Boxplot comparison of eind and eMTM at different reliability indices. ML=Maximum Likelihood, RgRegression, QT Quadra tic fit to the Tail data between rSand ln( ), LT Linear fit to the Tail data between rSand QH Quadratic fit to half of the data between rS and .Dashed line inside the box depicts the m ean and the solid line, the median. 1000 repetition PAGE 112 112 Figure G1 Lognormal distribution PAGE 113 113 Figure G2: Gamm a Distribution PAGE 114 114 Figure G3. Extreme Value type 1 Distribution PAGE 115 115 Figure G4. Uniform Distribution. PAGE 116 116 Figure G5. Rayleigh Distribution. PAGE 117 117 Figure G6. Exponential Distribution. PAGE 118 118 Boxplot of ratio. Different distributions Figure G7. Lognorma l Distribution. MTM 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1000m = 0.181 1000= 0.220 Figure G8. Gamma Distribution. MTM 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1000m = 0.349 1000= 0.340 PAGE 119 119 MTM 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 1000m = 0.183 1000= 0.185 Figure G9. Extreme Valu eType 1 Distribution MTM 0 1 2 3 4 5 6 x 10 1000m = 0.003 1000= 0.003 Figure G10. Uniform Distribution. PAGE 120 120 MTM 0 0.2 0.4 0.6 0.8 1 1000m = 0.146 1000= 0.185 Figure G11. Rayleigh Distribution. MTM 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 1000m = 0.241 1000= 0.276 Figure G12. Exponential Distribution. PAGE 121 121 REFERENCES Agarwal, H.; Lee, J.C.; Watson, L.T.; Renaud, J.E. 2004: A unilevel met hod for reliability based design optimization, Proceedings of the 45th AIAA/ASME/ASC E/AHS/ASC Structures, Structural Dynamics, and Material Conference, Palm Springs, CA, April 1922, AIAA Paper 20042029. Agarwal H., Shawn E.G., Renaud J.E., Perez V. M., Watson. L.T. 2007: Homotopy Methods for Constraint Relaxation in Unilevel Reliabilitybased Design Optimization, Submitted to the ASME Journal of Mechanical Design Bassi, F.; Embrechts, P.; Kafetzaki, M. 1998: Ri sk Management and Quantile Estimation In: A Practical Guide to Heavy Tails, Edited by Adler, R. J., et al., Boston, Birkhaeuser, pp. 111130 Beirlant, J.; Goegebeur, Y.; Segers, J.; Teugels, J. 2004: Statistics of Extremes: Theory and Applications, John Wiley & Sons Beirlant, J.; Vynckier, P.; Teuge ls, J.L. 1996: Excess Functions and Estimation of Extreme Value Index. Bernoulli, 2, pp:293318 Boos, D.D. 1984: Using Extreme Value Theory to Estimate Large Percentiles, Technometrics, 26, pp:3339 Caers, J.; Maes, M.A. 1998: Identifying Tails, Bounds and Endpoints of Random Variables. Structural Safety, Vol 20, pp 123 Castillo, E.; Hadi, A.S.; Balakr ishnan, N.; Sarabia, J.M. 2005: Extreme Value and Related Models with Applications in Engineering and Science, Hoboken, New Jersey:Wiley Interscience Chen, S.; Nikolaidis, E.; Cudney, H. H. 1999: Comparison of Probabilistic and Fuzzy Set Methods for Designing under Uncertainty. Proceedings, AIAA/ASME/ASCE/AHS/ASC Structures, Structural dynamics, and Materials Conference and Exhibit, 28602874. Chiralaksanakul, A.; Mahadevan, S. 2005: Mu ltidisciplinary Design Optimization Under Uncertainty, in press, Optimization and Engineering. Coles, S. 2001: An Introduction to Statistical Modeling of Extreme Values, London, England: SpringerVerlag Du, X.; Chen, W. 2004: Sequential Optimizatio n and Reliability Assessment Method for Efficient Probabilistic Design, Journal of Mechanical Design, Vol 126, No. 2, pp. 225 233. Du, X.; Sudjianto, A.; and Chen, W. 2004: An Integrated Framework for Optimization under Uncertainty Using Inverse Reliability Strategy, ASME Journal of Mechanical Design, Vol.126, No.4, pp. 561764. Elishakoff, I. 2001: Interrelation between Saf ety Factors and Reliability, NASA report CR2001211309 PAGE 122 122 Elishakoff, I. 2004: Safety Factors and Reliabi lity: Friends or Foes? Kluwer Academic Publishers, Dordrecht, The Netherlands Frangopol, D.M. 1985: Structural Optimi zation Using Reliability Concepts. Journal of Structural Engineering. Vol. 111, no. 11, pp. 22882301 Frangopol, D. M.; Maute, K. 2003: Lifecycle Reliabilitybased Optimization of Civil and Aerospace Structures, Computers and Structures, Vol. 81, No. 7, pp. 397410. Glasserman, P. 2004: Monte Carlo Methods in Financial Engineering. Springer Verlag Goel, T., Haftka, R.T., Shyy, W., Queipo, N.V.(2006) Ensemble of Multiple Surrogates, Structural and Multidisciplinary Optimization, 33(3) Gu, X.; Renaud, J.E.; Batill, S.M.; Brach, R.M. ; Budhiraja, A.S. 2000: Worst Case Propagated Uncertainty of Multidisciplinary Systems in Robust Design Optimization. Structural and Multidisciplinary Optimization, 20(3), 190213. Gurdal, Z.; Haftka, R.T.; Hajela, P. 1988: Design and Optimization of Laminated Composite Structures. New York: Wiley Hasofer, A.M. 1996: Parametric Esti mation of Failure Probabilities. In: Mathematical Models for Structural Reliability Analysis. Edited by Casicati, F.; Roberts, B. Boca Raton, FL: CRC Press Hosking, J.R.M.; Wallis, J.R. 1987: Paramete r and Quantile Estimation for the Generalized Pareto Distribution. Technometrics, Vol 29, no.3., pp. 339349 Kharmanda. G and Olhoff N, 2007: Extension of Optimum Safety Factor Method to Nonlinear Reliabilitybased Design Optimization, Journal of Structural and Multidisciplinary Optimization, Vol. 34, No.5, pp. 367380. Kharmanda. G., Olhoff N.,ElHami .A. 2004: Optimum Values of Structural Safety Factors for a Predefined Reliability Level with Extension to Multiple Limit States, Journal of Structural and Multidisciplinary Optimization, Vol. 26, No.6, pp. 421434. Kharmanda. G., Mohamed A., Lemaire M. 2002: Efficient Reliabilitybased Design Optimization Using Hybrid Space with App lication to Finite Element Analysis, Journal of Structural and Multidisciplinary Optimization, Vol. 24, No.3, pp. 233245. KirjnerNeto, C.; Polak, E.; Kiureghian, A. D. 1998: An Outer Approximations Approach to Reliabilitybased Optimal Design of Structures, Journal of Optimization Theory and Applications, Vol. 98, No. 1, pp 116. Kiureghian, A. D.; Zhang, Y.; Li, C. C. 1994: Inverse Reliability Problem, Journal of Engineering Mechanics, Vol. 120, No. 5, pp. 11541159. PAGE 123 123 Kuschel, N.; and Rackwitz, R. 2000: A new appr oach for structural opt imization of series systems, Applications of Statistics and Probability,Vol.2, No 8, pp987994 Lee, J. O.; Yang, Y. S.; Ruy, W. S. 2002: A Comparative Study on Reliabilityindex and Targetperformancebased Probabilistic St ructural Design Optimization, Computers and Structures, Vol. 80, No. 34, pp. 257269. Li, H.; Foschi, O. 1998: An Inverse Relia bility Measure and Its Application, Structural Safety, Vol. 20, No. 3, pp. 257270. Liu. D., Choi. K. K., Youn .B. D., Gorsich. D. 2006: Possibilitybased Design Optimization Method for Design Problems with Both Statistical and Fuzzy Input Data, Journal of Mechanical Design, Vol 128, pp. 928935. Liu. D., Choi. K. K., Youn .B. D. 2006: An Inve rse Possibility Analysis Method for PossibilityBased Design Optimization, AIAA Journal, Vol 44, No. 11, pp. 26822689 Madsen, H. O.; Krenk, S.; Lind, N. C. 1986: Methods of Structural Safety, NJ: PrenticeHall Maes, M.A.; Breitung, K. 1993: Reliability Based Tail Estimation. Proceedings, IUTAM Symposium on Probabilistic Structural Mechanics (Advances in Structural Reliability Methods), San Antonio, Texas, 335346 Maglaras, G.; Nikolaidis, E.; Haftka, R.T.; Cudney, H.H. 1997: AnalyticalExperimental Comparison of Probabilistic methods and Fu zzy Set based Methods for Designing under Uncertainty, Structural Optimization, 13(2/3), pp. 6980. Melchers, R. E. 1999: Structural Reliability Analysis and Prediction, New York: Wiley Metropolis, N. 1987: The Beginni ng of Monte Carlo Method, Los Alamos Science, Special Issue dedicated to Stanislaw Ulam: 125130 Moore, G.E. 1965: Cramming More Co mponents onto Integrated Circuits, Electronics, 38, (8). Mourelatos. Z.P., and Zhou, J. 2005: Reliability Estimation and Design with Insufficient Data Based on Possibility Theory, AIAA Journal, Vol 43, No. 8, pp 16961705 Myung J.I. 2003: Tutorial on Maximum Likelihood Estimation, Journal of Mathematical Psychology, 47, 90100 Phadke M.S. 1989: Quality Engineering Using Robust Design, Englewood Cliffs, NJ: Prentice Hall Pickands, J. 1975: Statistical Infere nce Using Extreme Order Statistics. Annals of Statistics. Vol 3,119131 PAGE 124 124 Qu, X.; Haftka, R. T. 2004: Reliabilitybase d Design Optimization Using Probabilistic Sufficiency Factor, Journal of Structural and Multidisciplinary Optimization, Vol. 27, No.5, pp. 314325. Qu, X.; Haftka, R. T., Venkataraman, S., Johnson, T.F.2003: Deterministic and ReliabilityBased Optimization of Composite Lamina tes for Cryogenic Environments, AIAA Journal, Vol 41, No.10, pp 20292036 Qu, X.; Singer, T.; Haftka, R. T. 2004:Reliabilityb ased Global Optimization of Stiffened Panels Using Probabilistic Sufficiency Factor, Proceedings of the 45th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynam ics, and Material Conference, Palm Springs, CA, April 1922,AIAA Paper 20041898. Qu,.X.,Venkataraman, S., Haftka, R T., Johnson, T.F. 2001: Reliability, Weight and Cost Tradeoffs in the Design of Composite Laminates for Cryogenic Environments, AIAA Paper 20011327, April 2001 Rackwitz, R.2000: Reliability Analysis Past, Present and Future. 8th ASCE Speciality Conference on Probabilsitic Mechan ics and Structural Reliability, 2426 July, Notre Dame, IN, USA. Ramu, P.; Qu, X.; Youn, B.; Haftka, R.T.; Choi, K.K. 2006: Safety Factors and Inverse Reliability Measures, Int. J. Reliability and Safety, Vol. 1, Nos. 1/2 Rosenblatt, M. 1952: Remarks on a Multivariate Transformation, The Annals of Mathematical Statistics, 23(3), 470 472 Royset, J.O.; Kiureghian, A. D.; Polak, E. 2001: Reliabilitybased Optimal Structural Design by the Decoupling Approach, Reliability Engineering and System Safety, Vol 73, No. 3, 2001, pp.213221 Sakawa, M. 1993: Fuzzy sets and Interactive Multiobjective Optimization. Plenum press Smarslok, B. P.; Haftka, R. T.; Kim, N.H. 2006: Taking Advantage of Separable Limit States in Sampling Procedures. 47th AIAA/ASME/ASCE/AHS/ASC Stru ctures, Structural Dynamics, and Materials Conference, AIAA 20061632, Newp ort, RI, May, Soundappan, P.; Nikolaidis, E.; Haftka, R.T.; Grandhi, R.; Canfield, R. 2004: Comparison of evidence theory and Bayesian Theory for Uncertainty Modeling, Reliability Engineering & System Safety, 85 (13): 295311 Tu, J.; Choi, K. K.; Park, Y. H. 1999: A New Study on Reliability Based Design Optimization, Journal of Mechanical Design, ASME, Vol 121, No. 4, 557564 PAGE 125 125 Wu, Y T.; Shin Y.; Sues, R.; Cesare, M. 2001: Sa fety Factor Based Approach for Probability based Design Optimization, Proceedings of 42nd AIAA/ ASME/ ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, Seattle, WA, AIAA Paper 20011522. Youn B. D.; Choi K. K.; Du. L. 2005: Enriched Performance Measure Approach (PMA +) for Reliabilitybased Design Optimization, AIAA Journal, Vol. 43, No. 4, pp. 874884. Youn, B. D.; Choi, K. K.; Park, Y. H. 2003: Hybrid Analysis Method for Reliabilitybased Design Optimization, Journal of Mechanical Design, ASME, Vol. 125, No. 2, pp. 221232. PAGE 126 126 BIOGRAPHICAL SKETCH Palani Ramu was born in Chennai, India in 1978. He earned his BEng in Mechanical from Madurai Kamaraj University, India in 1999. He wo rked at the Indian Institute of Technology, Bombay and InfoTech Pratt & Whitney Center of Excellence as Research Assistant and Design Engineer respectively till 2002. In 2003, he starte d his graduate studies at the University of Florida, Gainesville and worked towards a Ph.D in Aerospace Engineering. During his graduate studies, he was a visiting researcher at the Leon ardo Da Vinci University, Paris, France and was awarded a summer fellowship at the Center fo r Space Nuclear Research Idaho National Labs, Idaho. His areas of interest include applicati on of probability and statistical methods to design aerospace structures, specifical ly highly reliable designs. 