<%BANNER%>

Scalable Query Processing in Service-Oriented Sensor Networks

Permanent Link: http://ufdc.ufl.edu/UFE0019614/00001

Material Information

Title: Scalable Query Processing in Service-Oriented Sensor Networks
Physical Description: 1 online resource (124 p.)
Language: english
Creator: Bose, Raja
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: databases, networks, phenomena, processing, query, sensors, soa
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: The widespread availability of sensor devices and the rapid increase in their deployment, everywhere from industrial plants to private homes has put sensor network research in the spotlight for the past several years. Moreover, the requirement for rich highly configurable sensor network applications has led to the emergence of Service Oriented Sensor Networks (SOSNs), which imports the concept of Service Oriented Architecture (SOA) into the sensor network domain. It represents each of its sensors as a service object in a service framework that allows their dynamic discovery and composition into applications. Representing sensors as composable services and utilizing their associated knowledge can lead to significant enhancements in information processing capabilities of a sensor network, allowing it to operate on sophisticated data types and events beyond the primitive data types typically originating from individual hardware sensors. This dissertation describes the research and development of Sensable, a scalable query processing middleware which extends the capabilities of service-oriented sensor networks as follows: (1) Provides adaptive sensor-aware query processing to minimize overall power consumption in sensor networks by utilizing knowledge associated with sensor services to identify and minimize sensing operations which cause significant power drain; (2) Enhances Smart Space capabilities to sense virtual types of data which cannot be directly sourced from physical sensors due to their inherent sophistication or higher Quality of Service (QoS) requirements and; (3) Enables Smart Spaces to monitor and track phenomena event clouds whose shape, size and motion cannot be modeled precisely. Sensable attempts to bring query processing using service-oriented sensor networks into the realm of immediate utility by providing query scalability based on actual network infrastructure using current and emerging technologies and advanced querying capabilities encompassing abstract data types fused from multiple sensors, the concept of data quality in face of failure and detection and tracking of events susceptible to uncertainty in face of noise.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Raja Bose.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.
Local: Adviser: Helal, Abdelsalam A.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2010-05-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0019614:00001

Permanent Link: http://ufdc.ufl.edu/UFE0019614/00001

Material Information

Title: Scalable Query Processing in Service-Oriented Sensor Networks
Physical Description: 1 online resource (124 p.)
Language: english
Creator: Bose, Raja
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: databases, networks, phenomena, processing, query, sensors, soa
Computer and Information Science and Engineering -- Dissertations, Academic -- UF
Genre: Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: The widespread availability of sensor devices and the rapid increase in their deployment, everywhere from industrial plants to private homes has put sensor network research in the spotlight for the past several years. Moreover, the requirement for rich highly configurable sensor network applications has led to the emergence of Service Oriented Sensor Networks (SOSNs), which imports the concept of Service Oriented Architecture (SOA) into the sensor network domain. It represents each of its sensors as a service object in a service framework that allows their dynamic discovery and composition into applications. Representing sensors as composable services and utilizing their associated knowledge can lead to significant enhancements in information processing capabilities of a sensor network, allowing it to operate on sophisticated data types and events beyond the primitive data types typically originating from individual hardware sensors. This dissertation describes the research and development of Sensable, a scalable query processing middleware which extends the capabilities of service-oriented sensor networks as follows: (1) Provides adaptive sensor-aware query processing to minimize overall power consumption in sensor networks by utilizing knowledge associated with sensor services to identify and minimize sensing operations which cause significant power drain; (2) Enhances Smart Space capabilities to sense virtual types of data which cannot be directly sourced from physical sensors due to their inherent sophistication or higher Quality of Service (QoS) requirements and; (3) Enables Smart Spaces to monitor and track phenomena event clouds whose shape, size and motion cannot be modeled precisely. Sensable attempts to bring query processing using service-oriented sensor networks into the realm of immediate utility by providing query scalability based on actual network infrastructure using current and emerging technologies and advanced querying capabilities encompassing abstract data types fused from multiple sensors, the concept of data quality in face of failure and detection and tracking of events susceptible to uncertainty in face of noise.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Raja Bose.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.
Local: Adviser: Helal, Abdelsalam A.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2010-05-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0019614:00001


This item has the following downloads:


Full Text

PAGE 1

1 SCALABLE QUERY PROCESSING IN SERVICE ORIENTED SENSOR NETWORKS By RAJA BOSE 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 2009

PAGE 2

2 2009 Raja Bose

PAGE 3

3 To my grandfather Samarendra Kumar Mitra, who designed and built Indias first computer

PAGE 4

4 ACKNOWLEDGMENTS I thank my advisor Dr. Abdelsalam (Sumi) Helal for his constant help, guidance and support. I thank all the members of the Mobile and Pervasive Computing Laboratory, both past and present who were involved in design and development of the Atlas Platform and associated middleware. I especially thank Hen I Yang for all his help during development of the Virtual Sensors Framework and its experimen tal evaluation. I thank Steven Vanderploeg, James Russo and Steven Pickles for their role in designing, building and testing the Atlas h ardware. I express my utmost gratitude to Chao Chen for his constructive suggestions and debugging help.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS .................................................................................................................... 4 LIST OF TABLES ................................................................................................................................ 8 LIST OF FIGURES .............................................................................................................................. 9 ABSTRACT ........................................................................................................................................ 11 CHAPTER 1 INTRODUCTION ....................................................................................................................... 13 2 RELATED WORK ..................................................................................................................... 18 Energy -Efficient Q uery Processing ............................................................................................ 18 Virtual Sensors ............................................................................................................................ 20 Phenomena Detection and Tracking .......................................................................................... 21 3 OVERVIEW OF THE ATLAS PLATFORM ........................................................................... 24 Hardware ...................................................................................................................................... 24 Firmware ...................................................................................................................................... 25 Management and Service Layer ................................................................................................. 25 Enabling Features for a Query Processing System ................................................................... 26 4 THE SENSABLE QUERY PROCESSING MIDDLEWARE .................................................. 28 Query Processing Engine ............................................................................................................ 28 Sensor Service Data Handler ...................................................................................................... 29 Onboard Query Processor ........................................................................................................... 29 5 SENSOR -AWARE ADAPTIVE QUERY PROCESSING ...................................................... 31 Motivation .................................................................................................................................... 31 Re -examining Sensor Sampling Strategies ................................................................................ 33 Sensor Querying Strategies ........................................................................................................ 36 Definitions ............................................................................................................................ 36 Query Dissemination ........................................................................................................... 38 Push Strategy ........................................................................................................................ 39 Selective Pull Strategy ......................................................................................................... 41 Hybrid P ull -Push Strategy ................................................................................................... 45 Choosing the Best Query Plan ............................................................................................ 46 Monitoring Plan Performance ............................................................................................. 48 Fault Tolerance ............................................................................................................................ 49

PAGE 6

6 Experimental Per formance Analysis .......................................................................................... 51 Method of Experimentation ................................................................................................ 51 Results and Analysis ............................................................................................................ 52 6 VIRTUAL SENSORS FRAMEWORK .................................................................................... 60 The Concept and Classification of Virtual Sensors .................................................................. 62 Singleton Virtual Sensor ..................................................................................................... 62 Basic Virtual Sensor ............................................................................................................ 62 Derived Virt ual Sensor ........................................................................................................ 63 System Framework for Virtual Sensors ..................................................................................... 64 The Knowledge Base ........................................................................................................... 64 Framework Controller ......................................................................................................... 65 On Demand Cr eation of Virtual Sensors ................................................................................... 66 Virtual Sensor Composition Graph .................................................................................... 66 Smart Space Ambience Sensor Example ........................................................................... 67 Detecting the Enhanced Sentience of a Smart Space ................................................................ 69 Operations of a Basic Virtual Sensor ......................................................................................... 71 Basic Virtual Sensor Aggregation Process ........................................................................ 74 Fault Tolerance of Basic Virtual Sen sors ........................................................................... 74 Monitoring Data Quality of Basic Virtual Sensors ........................................................... 78 Operations of a Derived Virtual Sensor ..................................................................................... 79 Derived Virtual Sensor Aggregation .................................................................................. 80 Fault Tolerance of Derived Virtual Sensors ...................................................................... 80 Monitoring Data Quality of Derived Virtual Sensors ....................................................... 82 Experimental Performance Analysis .......................................................................................... 82 Fault Tolerance and Data Quality Monitoring ................................................................... 82 Latency of Data Arrival and Energy Consumption ........................................................... 86 7 PHENOMENA DETECTION AND TRACKING ................................................................... 90 Phenomena Clouds ...................................................................................................................... 91 Major Challenges ................................................................................................................. 92 Representation ...................................................................................................................... 92 Detection and Tracking ............................................................................................................... 93 Classification of Sensors ..................................................................................................... 94 Keeping Tabs on the Neighborhood ................................................................................... 95 Transition Rules ................................................................................................................... 96 Initial Selection of Candidate Sensors ................................................................................ 97 Monitoring for Initial Occurrences ..................................................................................... 98 Notification of Initial Occurrence ....................................................................................... 99 Growth of Phenomenon Cloud ........................................................................................... 99 Shrinking of Phenomenon Cloud ...................................................................................... 100 Handling Failures ............................................................................................................... 101 Real Time Moni toring by Applications ........................................................................... 102 A Practical Application of Phenomena Detection and Tracking ........................................... 103

PAGE 7

7 Experimental Analysis and Performance Evaluation .............................................................. 107 Experiment I: Effectiveness of Detection Strategy ......................................................... 108 Experimental setup ..................................................................................................... 108 Results and analysis ................................................................................................... 109 Experiment II: Resource and Power Consumption ......................................................... 112 Experimental setup ..................................................................................................... 114 Results and analysis ................................................................................................... 114 8 CONCLUSIONS AND FUTURE WORK .............................................................................. 119 LIST OF REFERENCES ................................................................................................................. 121 BIOGRAPHICAL SKETCH ........................................................................................................... 124

PAGE 8

8 LIST OF TABLES Table page 5 1 Atlas ZigBee node hardware specifications ......................................................................... 52 7 1 Actions taken by a sensor node with respect to its neighbors which are not idle .............. 96 7 2 Energy consumption specifications for Atlas ZigBee nodes ............................................. 113

PAGE 9

9 LIST OF FIGURES Figure page 1 1 Sensable middleware components ........................................................................................ 14 3 1 Layers of an Atlas nod e ......................................................................................................... 24 3 2 Block diagram of node firmware .......................................................................................... 25 4 1 Basic architecture of Sensable ............................................................................................... 28 5 1 Network cost versus sensor sampling cost in MICA2 and RCB ......................................... 32 5 2 Query dissemination and evaluation ..................................................................................... 36 5 3 Comparing energy consumption of different query plans ................................................... 53 5 4 Comparing latency of response for different query plans ................................................... 54 5 5 Effect of selectivity on energy consumption of query plans ............................................... 55 5 6 Effect of selectivity on latency of query plans ..................................................................... 57 5 7 Effect of number of sensors on energy consumption of query plans .................................. 58 5 8 Effect of number of sensors on latency of query plans ........................................................ 59 6 1 Virtual Sensors Framework architecture .............................................................................. 64 6 2 Sensor composition graph...................................................................................................... 67 6 3 Sensing ambience using virtual sensors ................................................................................ 68 6 4 Initial sensing capability of the smart space ......................................................................... 70 6 5 Effect of introducing humidity sensors into the smart space .............................................. 71 6 6 Logical versus network view of an aggregation tree ........................................................... 72 6 7 Relative error % vs. VSQI ..................................................................................................... 81 6 8 Number of sensor failures vs. VSQI and relative error ....................................................... 83 6 9 Effect of sensor failure pattern on VSQI .............................................................................. 85 6 10 Comparing latency in arrival of final output ........................................................................ 86 6 11 Energy consumption of virtual sensor per epoch ................................................................. 88

PAGE 10

10 7 1 Dissection of a phenomenon cloud ....................................................................................... 93 7 2 Classification of participating sensors .................................................................................. 94 7 3 Detection and tracking of a phenomenon cloud ................................................................... 98 7 4 Ratio of total active sensors to cloud size in a rectangular sensor grid .............................. 99 7 5 Gator Tech Smart House...................................................................................................... 103 7 6 Smart Floor tile with force sensor and Atlas Platform node ............................................. 104 7 7 Ripple effect of a foot step on the Smart Floor .................................................................. 105 7 8 Walking motion as a phenomenon ...................................................................................... 106 7 9 Effect of varying n with pT=0.4 and m=150 .................................................................... 107 7 10 Determining the optimal value of n .................................................................................. 108 7 11 Effect of varying pT with n=3 and m=150 ....................................................................... 109 7 12 Determining the optimal value of pT ................................................................................ 111 7 13 Effect of varying m with n=3 and pT=0.4 ........................................................................ 112 7 14 Determining optimal value of m ....................................................................................... 113 7 15 Number of update messages sent to the query processor .................................................. 113 7 16 Average number of active sensors required ....................................................................... 115 7 17 Number of network messages exchanged........................................................................... 116 7 18 Average energy consumption per node ............................................................................... 118

PAGE 11

11 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 SCALABLE QUERY PROCESSING IN SERVICE ORIENTED SENSOR NETWORKS By Raja Bose May 2009 Chair: Abdelsalam (Sumi) Helal Major: Computer Engineering The widespread availability of sensor devices and the rapid increase in their deployment, everywhere from industrial plants to private homes has put sensor network research in the spotlight for the past several years. Moreover, the requirement for rich highly configurable sensor network applications has led to the emergence of Service Oriented Sensor Networks (SOSNs), which imports the concept of Service Oriented Architecture (SOA) into the sensor network domain. It represents e ach of its sensors as a service object in a service framework that allows their dynamic discovery and composition into applications. Representing sensors as composable services and utilizing their associated knowledge can lead to significa nt enhancement s i n information processing capabilities of a sensor network, allow ing it to operate on sophisticated data types and events beyond the primitive data types typically originating from individual hardware sensors. This dissertation describes the research and development of Sensable a scalable query processing middleware which extends the capabilities of service oriented sensor networks as follows: (1) Provides adaptive sensor aware query processing to minimize overall power consumption in sensor networks by uti lizing knowledge associated with sensor services to identify and minimize sensing operations which cause significant power drain ; (2) Enhances Smart Space capabilities to sense virtual types of data which cannot be directly

PAGE 12

12 sourced from physical sensors du e to their inherent sophistication or higher Quality of Service (QoS) requirements and; (3) Enables Smart Spaces to monitor and track phenomena event clouds whose shape, size and motion cannot be modeled precisely. Sensable attempts to bring query processi ng using service -oriented sensor networks into the realm of immediate utility by providing query scalability based on actual network infrastructure using current and emerging technologies and advanced querying capabilities encompassing abstract data types fused from multiple sensors, the concept of data quality in face of failure and detection and tracking of events susceptible to uncertainty in face of noise.

PAGE 13

13 CHAPTER 1 INTRODUCTION T he requirement for rich highly -configurable sensor network applications led to the emergence of Service Oriented Sensor Networks (SOSN ). A SOSN imports the concept of Service Oriented Architecture (SOA) into the sensor network domain. It represents each of its sensors as a service object in a service framework that allows their dynamic discovery and composition into applications [9]. This provides a loose ly -coupled model where there is a strict separation between application logic and device -specific operational logic. In such a system, s ensors are not integrated using hard -coded routines inside applications but instead are discovered integrated and accessed on -demand. Moreover such loose -coupling also enable s mu ltiple applications to simultaneously share the same set of devices deployed in the space without interfering with each others operations. The representation of a sensor as a service immediately opens up a number of possibilities for enhancing the informa tion processing capabilities of the sensor network. Each sensor is now associated with some knowledge about the device which can be exploited to provide more power -efficient query processing in the network. Representing individual sensors as services in a n SOA can also provide the ability to dynamically compose multiple data sources either to derive more sophisticated data types which cannot be obtained directly from physical sensors or to provide higher standards of Quality of Service (QoS) and availability guarantees. Furthermore instead of restricting query processing to well -defined data and events, the sensor network can be made capable of monitoring and tracking phenomena events whose behavior cannot be modeled (shape, size and motion) accurately over time W e describe Sensable a scalable query processing middleware which utilizes the SOA characteristics of a SOSN to provide information processing capabilities beyond those available

PAGE 14

14 in traditional query processors which only deal with primitive data directly obtained from low level physical sensors. Th e middleware is designed and built on top of the Atlas Platform [ 18], which is a plug and -play service -oriented sensor platform developed at the Mobile and Pervasive Computing Laboratory a t the University of Florida. The middleware is composed of three distinct components (Figure 1 1) : 1 Adaptive sensor aware in network query processing for minimizing power consumption in sensor networks utilizing low power communications such as ZigBee 2 Vi rtual Sensors f ramework for processing queries involving sophisticated derived data types and queries requiring QoS guarantees. 3 Phenomena Detection and Tracking for monitoring and tracking events which cannot be defined in precise terms or have their beha vior (shape, size and motion) modeled accurately over time. Figure 1 1. Sensable middleware c omponents The rest of this dissertation is organized as follows. In Chapter 2, we present related work. In Chapter 3, we introduce basic concepts of service oriented sensor networks and give an overview of the Atlas Platform and its hardware and software components, which provide the Atlas ServiceOriented Sensor Network Sensor Aware Query Processing Virtual Sensors Phenomena Detection & Tracking

PAGE 15

15 basic impl ementation layers for Sensable In Chapter 4, we describe the overall system architecture of the Sensable query processing middleware. In Chapter 5, we describe sensor aware adaptive query processing, which utilizes knowledge provided by a sensors servic e representation to execute distributed, in-network queries involving multiple sensors in a power -efficient manner. To date, sensor network research has assumed that the cost of transmitting a sensor reading over the network is much higher than the cost of sampling a sensor. However, this assumption is no longer always valid, due to availability of new generation sensor platform hardware which utilizes industry standard mesh -networking protocols such as ZigBee on top of relatively high speed yet low power wireless radios In fact, we have experimentally verified that the energy consumed for acquiring a sample from a sensor can be significantly higher than the energy consumed for transmitting its reading over the network. Hence, new querying strategies need to be formulated which optimize the order of sampling sensors across the network in such a manner that sensors with expensive acquisition costs are not sampled unless absolutely required. We propose distributed pull -push querying mechanisms whi ch optimiz e the query plan by adapting to variable costs of acquiring readings from different sensors across the network. The goal of these mechanisms is to minimize the energy consumption of nodes executing a query while ensuring that the latency of query response does not exceed user -specified bounds. We also analyze the performance of various plan options on different hardware configurations, based on their energy consumption and latency, through experiments using real -world data. In Chapter 6, we describe the Vir tual Sensors framework. Contemporary sensor network query proces sing systems are typically equipped to only process queries involving primitive data types which can be directly sourced from hardware sensors. The Virtual Sensors framework on

PAGE 16

16 the other hand enables the query processor to handle queries involving more abstract and derived types of data fused from multiple sensors. The framework leverages SOA concepts and plug and -play features of the Atlas Platform to allow the sensor network to automatically detect new derived sensing capabilities, whenever new physical sensors are introduced into the space. It provides capabilities for on demand creation of virtual sensors and their lifecycle management and also allows applications to monitor sensor data qual ity and provides fault -tolerance mechanisms which utilize approximation algorithms to maximize the availability of sensing resources. We also address the issues of scalability and latency and propose distributed in network algorithms for the creation and execution of virtual sensors. We show through experiments that the proposed mechanisms provide excellent fault tolerance and data quality monitoring capabilities and result in significantly lower latency and energy consumption as compared to the centralized stream -based approaches. In Chapter 7, we describe phenomena detection and tracking which enables the query processor to detect and track events which cannot be precisely described or modeled. The phenomena detection and tracking component extends the information processing capability of the sensor network to execute queries which involve real time detection and tracking of groups of clustered events called phenomena clouds. Phenomena clouds are characterized by non deterministic, dynamic variations over time, of their shape, size and direction of motion along multiple axes. This makes it difficult to apply contemporary techniques proposed for tracking moving objects using wireless sensor network s (WSNs). In the past, the utility of phenomena detection and tracking has been limited to applications such as tracking oil spills and gas clouds. However, through our collective experience over the years in a completely different deployment domain (Smart Spaces), we have discovered great utility and value in applying this concept to

PAGE 17

17 accurately and efficiently observe other types of phenomena. In this dissertation, we describe distributed sensor network algorithms for in -situ detection a nd tracking of phenomena clouds, which utilize localized in -network processing to simultaneously detect and track multiple phenomena clouds in a sensor space. Our algorithms not only ensure low processing and networking overhead at the centralized query processor but also m inimize the number of sensors which are actively involved in the detection and tracking processes at any given time We validate our approach using both r eal life smart home applications as well as simulation experiments, which analyze the effectiveness an d efficiency of our detection and tracking algorithms. We further show that our distributed algorithms also result in significant savings in resource usage and energy consumption.

PAGE 18

18 CHAPTER 2 RELATED WORK In this section, we look at past research work done in the area of sensor network query processing systems. We organize related work into three categories corresponding to each of the three main features of Sensable : (1) E nergy-efficient query processing; (2) Virtual Sensors and; (3) Phenomena Detection and Tracking. Energ y -Efficient Query Processing There is a large body of available research proposing techniques for querying sensor data. These approaches can be broadly di vided into stream -based and acquisition -based approaches. The stream -based approaches view data originating from sensors as a series of data streams. They assume a priori existence of sensor data and do not factor in sensor acquisition costs. The acquisition -based techniques on the other hand closely look at ways to sample sensors and transmitting their information so that the total energy con sumption is minimized. O ur sen sor aware approach is more closely related to acquisition -based techniques rather than streaming methods hence, the related work covered in this sub-section reflects that. The related work is classified into two broad categories: acquisition -based techniqu es for minimizing the cost of sampling sensors and push -pull mechanisms for query execution in sensor networks. In TinyDB, Madden et al [23] proposed algorithms for minimizing energy consumption by optimizing the order of sampling sensors and the predica tes being applied on them using a series parallel graph. However, their approach was only proposed for ordering sensors connected to the same node rather than all sensors participating in a query. This is probably due to the fact that when TinyDB was desi gned, network costs were much more significant than sensor sampling costs. A model -based approach (BB Q) was proposed by Deshpande et al [8 ] utilizing a time -varying multivariate Gaussian model for processing sensor

PAGE 19

19 queries. BBQ pulls readings whenever th e model needs to be updated and a lso utilizes correlation among various phenomena to select sensors which are less expensive to sample. Both the above mentioned strategies only look at the tradeoffs based on the cost of sampling one sensor versus another a nd do not consider the tradeoff between network costs and sensor sampling costs. Hartl et al [11] describe d an inference -b ased technique where only a subset of nodes are activated and the readings from their sensors are used to approximate readings from o ther sensors using Bayesian inference. Their mechanism was proposed specifically for obtaining the global average of reading values and is subject to prediction errors due to incomplete data modeling or significant fluctuation of readings. However, the com bination of our approach and model -based techniques such as those listed above can potentially provide a more optimal solution in the future. The push -pull approach for query execution has been applied by a number of query processing strategies for sensor networks. However, as we shall see, these mechanisms mainly seek to minimize the network cost, based on the assumption that it always significantly outweighs the s ensor sampling cost. Trigoni et al [32] propose d a hybrid push pull mechanism where sensors have their readings pushed to intermediate nodes (called view nodes), from where they are pulled by the query processor based on query requirements. However, their approach only seeks to reduce the network cost and late ncy of answering queries and does not factor in the cost of sampling sensors (as evidenced by the fact that it uses the push a pproach to sample sensors). Shenker et al [28 ] propose d a structured hybrid push-pull mechanism where data is pushed from sensors onto intermediate nodes and stored u sing geographic hash tables. The sink nodes apply the same hash function to determine where specific data is stored and use the pull approach to retrieve data. In contrast, Liu et al [21] use d an unstructured pushpull approach

PAGE 20

20 (called Comb -Needle) where queries are disseminated along the horizontal lines of a sensor grid (visualized as a comb with horizontal teeth) and data is independently pushed along the vertical lines of a sensor grid (visualized as vertical needles). However, all these hybrid push -pu ll mechanisms mentioned above require nodes to proactively push sensor readings to intermediate nodes in the network and hence, are fundamentally different from the push -pull approach proposed in this dissertation. In the context of actual sensor sampling all of th e above methods can be classified as mechani sms which use the push strategy which do es not place a premium on sensor sampling cost. This is essentially due to the fact that they assume that network cost always outweighs sensing cost. In contrast our sensor aware query processing makes no such assumption and tries to avoid sampling sensors which do not contribute to the query result at a given epoch thereby, addressing the issue where sensing cost can outweigh network cost by a large margin and b ecome the major contributor to total energy consumption. Virtual Sensors Sensor fusion and virtual sensors have always been actively researched topic s in the sensor network community. However, the majority of publications in this area have tended to focus on developing algorithms for sensor fusion and application specific virtual sensors [10, 14, 26, 3 1 ]. However, there are some groups of researchers who have looked at virtual sensors from a systems perspective and developed mechanisms for the creation and deployment of virtual sensors. Kabadayi et al. [ 16] describe d a middleware for virtual sensors where heterogeneous sensor readings are aggregated to generate high er abstractions of data. However, they only look ed at aggregation aspects of virtual sensors w ithout addressing reliability and availability issues. Furthermore, their proposed middleware seems to be too rudimentary and can easily be superseded by native mechanisms available in service oriented architecture (SOA) frameworks

PAGE 21

21 such as OSGi [19], which form the basis of SOSNs. Lewis [20] defined virtual sensor as a combination of a physical transducer along with data signal processing (DSP) elements for reliable data estimation. Both Kabadayi and Lewis have considered different aspects of virtual sensor s but none of them have propose d practical distributed mechanisms for execution, data quality monitoring and fault tolerance. The Global Sensor Network (GSN) [ 1 ] is a project at EPFL, Switzerland which aims at providing a unified framework to users for ac cessing real and virtual sensors. Their work comes closest to our concept of a virtual sensor framework which enables on demand creation and execution of software sensors. However, GSN follows a stream based centralized approach for implementing virtual sensors where physical sensors are required to stream up all their readings to the central host where virtual sensors are executed. This not only leads to increased overhead at the central host but also leads to high latency and increased power consumption. In contrast, even though we also have a centralized service layer in our architecture, we take a distributed acquisition -based approach for the internal operations of a virtual sensor and propose mechanisms which reduce latency and improve fault tolerance of virtual sensors and lower power consumption of the sensor nodes, without sacrificing the advantages of providing a user with a unified framework for accessing virtual sensors. Unlike the systems described above, the Virtual Sensors framework described i n this dissertation provides full -featured capabilities to query multiple types of virtual data with quality monitoring and fault tolerance mechanisms. Furthermore, it also enables the ondemand creation and distributed in -network execution of virtual sens ors to minimize energy consumption and latency of response. Phenomena Detection and Tracking Nile -PDT [2, 3 ] is a Phenomena Detection and Tracking (PDT) framework running on top of the centralized Nile data stream management system, developed by the Indian a Center for

PAGE 22

22 Database Systems (ICDS) at Purdue University Nile -PDT wa s designed for det ecting and tracking phenomenon clouds such as gas clouds, oil spills and chemic al waste spillage. Nile -PDT uses two custom database operators, namely, SN Scan and SN Join to perform phenomenon detection and tracking. The SN Scan operator scans all the sensors in the network and chooses candidate sensors which have a high probability of detecting the phenomenon. The SN Join operator then evaluates each of these c andidate sensors and checks if they join with other candidates a certain number of times and hence, are detecting a phenomenon event. NilePDT use s feedback control to continuously tune the SN -Scan and SN -J oin parameters to maxi mize efficiency of the detec tion process The main drawback of the Nile -PDT approach is that it takes a streaming database view of the process. It does not consider any mechanisms for controlling the flow of data at the source sensors themselves or address power consumption and netwo rk bandwidth issues inside the sensor network. Furthermore, it requires all sensors to pump readings to the SN -Scan operator to allow it to choose phenomenon candidates, which can lead to potentially massive scalabi lity issues. Omotayo et al. [ 27] have de scribed a data harvesting framework for tracing phenomena. They propose d algorithm s for maintaining a data farm on the nodes by maximum utilization of their on board non -volatile storage, to enable backtracking to deter mine the cause of a phenomenon. McErl ean et al. [ 25] propose d distributed event detection and tracking algorithms for moving objects using WSNs. Their approach involves distributed collaboration between sensors to detect an event. Each sensor which detects an event notifies neighboring sensor s about its occurrence and the central base station is only notified if a certain number of sensors agree that the object is moving, that is, the event is propagated a certain number of times. However, this system assumes the prior availability of optimal ad -hoc routing mechanisms and is primarily

PAGE 23

23 designed for detecting individual discrete objects with well defined shape and size, as opposed to phenomenon clouds whose shape and size typically cannot be defined in exact terms. Chintalapudi and Govindan [ 7 ] d escribe d algorithms for detecting sensors lying closest to the edges of a phenomenon cloud. They utilize d image processing and classifier techniques to determine if a sensor lies at the edge of the phenomenon or not. However, their approach not only made s implifying assumptions regarding the shape of the edges (such as whether it is a line or an ellipse), but also le d to detection of a high number of false positives, since the extreme fringes of a phenomenon are more susceptible to sensor errors and rapid f luctuations. In contrast to the above approaches, the phenomena detection and tracking algorithms described in this dissertation do not make assumptions regarding the shape, size or motion of phenomena cloud s and execute in a localized, in -network fashion which m ake them more scalable in terms of processing, networking and energy resources as compared to contemporary stream -based techniques.

PAGE 24

24 CHAPTER 3 OVERVIEW OF THE ATLA S PLATFORM In this section we provide a brief overview of the Atlas Platform, which is a service oriented plug and -play sensor and actuator platform developed at the University of Florida Mobile and Pervasive Computing Laboratory. Atlas provides the basic building blocks for building a SOSN a nd in its most basic incarnation consists of a set of hardware nodes and a collection of management components running inside an OSGi framework which forms the core service layer. Hardware The Atlas platform hardware consists of a set of nodes which provi de a physical interface to sensors and actuators. Each Atlas node is a modular hardware device composed of stackable, swappable layers, with each layer providing specific functionality. A ba sic Atlas node consists of 3 layers: Processing Layer, Communication Layer, and Device Connection Layer (Figure 3 1). Figure 3 1. Layers of an Atlas node The Processing Layer is based around the Atmel ATmega128L microcontroller. The ATmega128L is an 8MHz chip that in cludes 128KB FLASH memory and an 8 -channe l 10-bit Communication Layer Processing Layer Device Connection Layer

PAGE 25

25 Analog to Digital C onverter (ADC) which can operate at voltages between 2.7 and 5.5V. The Communication Layer handles the transfer of data and control -messages over the network. Currently Atlas has 4 communication options 802.11b WiFi, wired 10BaseT Ethernet, ZigBee and USB. The Device Connection Layer is used to connect various sensors and actuators to the node. Atlas has numerous available options for connecting multiple analog and digital sensors and actuators to a single node. Figure 3 2. Bl ock diagram of node firmware Firmware The firmware is responsible for performing four fundamental tasks, namely, controlling sensor and actuator operations, performing network communications, handling commands and control instructions and executing proc esses and filters pushed on the node by the service layer. Each of these tasks is handled by separate components (Figure 3 2), which abstract away the low level deta ils of their operation from other firmware components which may be interested in using thei r functionality. Management and Service Layer The Atlas management components run inside an OSGi service framework hosted on a central host. The host computer can be anything from a Single Board Computer (SBC) running sensor Command Handler Onboard Processing Engine Process Scheduler Process Execution Environment Device Controller Communications Module sensor Process Table Schedule Table

PAGE 26

26 Linux to a full fledged standard desk top PC. OSGi is a Java -based service oriented architecture (SOA) based framework that provides a runtime environment for dynamic, transient service modules known as bundles. It provides functionalities such as life cycle management as well as service regis tration and discovery that are crucial for scalable composition and maintenance of applications. Each sensor or actuator is represented by Atlas as an individual OSGi bundle in the service framework. Applications are able to dynamically discover and access sensors via their respective services using standard OSGi mechanisms. The core management components of Atlas running inside the OSGi framework consist of the Network Manager, Configuration Manager and the Bundle Repository. The Network Manager handles th e joining and departure of nodes in the network. The Configuration Manager manages the configuration settings of each node and enables remote configuration. The Bundle Repository stores and manages all the supported sensor and actuator service representati ons. Features of these three common services are also accessible to the user through an intuitive web interface. Enabling Features for a Query Processing System The Atlas Platform provides the following enabling features for building a query processing sys tem for SOSNs: 1 Each sensor is represented as a service thereby abstracting away the low level operational details of the hardware from the query processor. 2 Atlas enables the query processor to access heterogeneous sensors by using generic high level methods without having to adapt to changing sensor or node hardware specifications. 3 The plug andplay capability of Atlas allows the query processor to be notified in case a new sensor enters the network or an old sensor suddenly goes offline. This enable s the query processor to immediately take appropriate actions for queries involving the offline sensor. 4 Atlas allows different segments of the sensor network to run using diverse networking protocols, yet it provides a single unified network view to the query processor by abstracting away the routing and network management details.

PAGE 27

27 5 Atlas provides distributed computing capabilities where, each Atlas node has an on-board processing engine which can execute processes pushed by the service and application lay ers. The query processor can push in range filters using Java method calls which are automatically translated into node instructions and automatically routed to the appropriate node. 6 Queries and applications do not have to be written as firmware code and pre loaded on the nodes and can be pushed dynamically through the service layer. Hence, deployment of new applications or modification of existing ones do not require recompilation and updating of node firmware

PAGE 28

28 CHAPTER 4 THE SENSABLE QUERY PROCESSING MI DDLEWARE In this section, we give an overview of the system architecture of the Sensable query processing middleware The Sensable system is built on top of the Atlas Platform and consists of three major componen ts distributed am ong var ious layers (Figure 4 1) Figure 4 1. Basic architecture of Sensable Query Processing Engine T h e query processing engine runs as a service inside the OSGi framework. It resides in the service layer and is responsible for the overall management of all query operations in the system. It accepts queries from applications, generates optimized query plans and schedules their execution. It interacts with the sensors through their service objects and uses standard OSGi mechanisms to determine when a new sensor comes online or an existing sensor goes offline. The query plan determines whether a query needs to be executed entirely on the nodes or the processing engine or if some sub-queries have to be pushed on to the nodes while the rest are s1 s3 s2 Sensor Services Sensors Onboard Query Processor Physical Layer Service Layer Data Handler Data Handler Data HandlerNode Layer Application Application Application Application Layer Onboard Query Processor sensor history sensor history Atlas Management Components Query Processing Engine Knowledge BaseAtlas Nodes

PAGE 29

29 executed centrally. The query processin g engine also acts as the central clock source and can perform time synchronization between the nodes and the service laye r to ensure that the nodes are in step with each other and the service layer. The knowledge base inside the query processor stores inf ormation such as phenomenon definitions, static optimization parameters and virtual sensor model definitions. Sensor Service Data Handler The second Sensable component also resides in the service layer inside each sensor bundle. It consists of a data han dler which listens for sensor readings based on epochs of queries running for that sensor. The Sensable query processor does not store the history of each sensor in a central location. Instead the sensor history is stored in a distributed manner at the Atlas node connected to that sensor. Onboard Query Processor The Onboard Processor is responsible for scheduling and executing processes which have been pushed on the node. It exists as part of the nodes firmware and is responsible for executing queries pushed on to the node. It works in tandem with other node level components and performs tasks such as sampling sensors, applying filters on their data and transmitting readings back to the service layer. It also performs additional tasks such as ensuring t hat the nodes operations are time -synchronized with the service layer and maintaining the nodes portion of the distributed history of readings originating from sensors connected to it. The query processing engine can specify processes for individual sens ors for tasks such as filtering out sensor data using range filters and triggering an action, in case a filter condition is satisfied. For pushing a process on a particular sensor, the pushProcess() method of that sensors OSGi service, is called. The sens or service then encodes the process specifications into a process packet and forwards it to the node controlling that sensor. The nodes command handler receives the message and

PAGE 30

30 detects that it contains a process packet. It pushes the packet into the Proce ss Table and notifies the Process Scheduler which decodes it and based on the processs sampling and transmission rates, reserves time slots for it in the schedule table. The Process Execution Environment looks up the Schedule Table and executes the variou s processes whose references are stored in it, based on the time slots allotted to them.

PAGE 31

31 CHAPTER 5 SENSOR -AWARE ADAPTIVE QUERY PROCESSING The main focus of research on query processing in sensor networks till date has been the minimization of en ergy consumption of the sensor network. Current research on acquisition based query processing on sensor networks assumed that the energy cost of transmitting a sensors reading is always significantly higher than the energy cost of sampling that sensor. H owever, this assumption seems to no longer be universally valid, due to rapid advances in radio hardware technology and availability of relatively high speed yet low power networking protocols such as ZigBee/802.15.4. From our experience, we found that the energy consumption of a node can be quite high even when it samples a sensor without transmitting the reading over the network and the contribution of network cost to the total energy consumption of a node can be significantly less than the sensor samplin g cost. To the best of our knowledge, there is no published work which has explicitly looked at this issue and has suggested query processing strategies to deal with it. Motivation First, we explain why the cost of sampling a sensor can become so signific ant that its contribution to the total energy consumption of a node becomes extremely high as compared to network cost. Figure 5 1 shows the comparative percentage contributions of networking and sensing tasks to total energy consumption for the Crossbow M ICA2 mote and the Atmel Zlink RCB sensor platform. The MICA2 was relea sed in 2002 and has been among the first and most widely used sensor platforms in sensor network research. The Atmel Zlink RCB on the other hand, was released relatively recently in 2007 and is among a new generation of sensor platforms being developed by companies such as Atmel and Texas Instruments. The MICA2 hardware is built around an Atmel Atmega128 microcontroller and Chipcon CC1000 radio. It typically runs

PAGE 32

32 the proprietary ad -hoc ro uting protocol used by TinyOS and has a transmission baud rate of 38.4Kbps. The Atmel Zlink RCB is based around the Atmega1281 microcontroller and AT86RF230 802.15.4 radio. It runs a full featured industry standard ZigBee stack and has a comparatively high transmission baud rate of 250Kbps. The energy consumption values for the MICA2 were derived from exp eriments performed by Madden et al [23]. The energy consumption values for the Zlink RCB were calculated based on the sensor configuration used by Madden et al [23] and hardware specifications provided by Atmel. Figure 5 1. Network cost versus s enso r sampling c ost in MICA2 and RCB W e can observe that the contribution of sensing to the total energy consumption for the Zlink RCB is significantly higher (by about 20%) as compared to the MICA2 mote. This can be explained as follows. The newer 802.15.4 radios used by ZigBee have lower current consumpt ion and significantly higher baud rates (250Kbps) as compared to older low -power radios such as the CC1000 (38.4Kbps) Hence, 802.15.4 radios take much less time to transmit a 80.50 95.62 13.50 2.87 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% MICA2 RCB Application Energy Cost Network Energy Cost Sensor Sampling Energy Cost

PAGE 33

33 packet and therefore consume much less energy. On the other hand, the technology used for acquiring readings from a sensor has more or less remained the same. Apart from a few power hungry sensors such as magnetometers or organic byproduct sensors, most sensors have very low power requirements of the order of fractions of a milliAmpe re (mA) In fact, the cost of sampling a sensor is largely due to the cost of operating the microcontro ller and its Analog -to -Digital Converter (ADC), which is of the order of a few milliAmperes (mA) In order to sample an analog sensor (most sensors used today fall in this category), a microcontroller has to operate its ADC to access and convert the sensors output voltage into discrete numbers. This is a relatively power hungry operation and also typically requires some time to complete since the speed of sampling is bounded by delays such as time required to initialize the ADC, waiting for input line voltages to stabilize and taking multiple samples to ensure correctness of reading. These delays are inevitable due to physical limitations of the microcontr oller and sensor and the fact that the sensors themselves are extremely simple electrical devices such as resistors or diodes. Therefore, the energy required to take a sensor reading has remained more or less the same whereas the energy required for transm itting a reading over the network has decreased drastically. With more advances in radio technology we expect this gap to widen further in the future. Hence, the tables are now being turned and one is faced with situations where the cost of sensing can act ually outweigh network cost by a significant margin. Re -examining Sensor Sampling Strategies In light of the above discussion, we feel there is a need to take a fresh look at sensor sampling and acquisition strategies. We no te that even though in case of t he MICA2 the sensing cost has a higher contribution to the total energy cost as compared to network cost, which may not be significant enough to outweigh the total network cost of communicating over a multi hop network. On the other hand, for newer hardwa re such as the Atmel RCB, the difference is

PAGE 34

34 significant enough to warrant a fresh look at query processing strategies for in -network execution of queries. Currently almost all query execution plans use the Push approach. Push requires nodes to autonomousl y sample their sensors and push their readings into the network. Optimizations have been suggested where nodes can avoid sampling their sensors if the result of the query can be deduced from already existing information. These optimizations either involve the use of partial aggregate information or utilize models to determine which sensors to sample. The former approach used in TAG [22] only avoids sampling of sensors directly connected to a node and does not control sensor sampling of other nodes below it in the aggregation tree. The latte r approach used by Deshpande et al in BBQ [8 ], makes the decision of which sensors to sample based on data requirements of the model, rather than e nergy consumption. Deshpande et al do consider the acquisition cost of s ampling a sensor however, they are primarily interested in utilizing this data to find another sensor which is less expensive to sample based on correlations between various phenomena (example, temperature and battery voltage). Therefore, their approach d oes not look at the question of whether to entirely avoid sampling a sensor based on energy considerations; rather it depends on a model to determine which sensors to sample and attempts to find replacement sensors if possible, having lower energy cost of sampling. The Pull approach requires nodes to wait for an explicit command before sampling its sensors. The pull approach has been utilized very frugally in sensor network query processing till now, mainly due to the assumption that network cost is more than the cost of sensing. But both the push and pull approaches have their advantages and drawbacks. The push approach allows nodes to execute their task autonomously and has low latency of query response. However, the push approach also almost always requir e s a node to sample its sensors barring certain

PAGE 35

35 intermediate nodes as mentioned before This approach worked well in the past when the cost of sampling a sensor was a minor contributor to the total energy cost. But this may not work that well now since acq uiring readings from a sensor makes up the largest fraction of a nodes energy consumption by a wide margin. Hence, even if a sensors reading is not transmitted over the network, simply sampling it causes the nodes energy consumption to be quite signific ant. The pull approach on the other hand only requires nodes to sample their sensors when explicitly asked to. However, using pull naively by pulling readings from sensors in parallel will not prove useful in terms of energy consumption due to obvious reas ons. On the other hands, if sensor readings are pulled in an optimal order based on their acquisition costs, this can lead to lower energy costs since sensors will be sampled only when required. Madden et al [23] h ave proposed such a technique for TinyDB, however their mechanism is only meant for sensors connected to the same node and provides a locally optimal ordering of sampling of sensors connected to the same node rather than a network -wide ordering of all sensors involved in the query. The disadvanta ge of the pull approach is that it suffers from higher network cost as it requires the exchange of two messages as opposed to one message in pull. Moreover, its data delivery latency is also higher than push. Hence, we feel that a third plan option namely, a hybrid Pull Push approach is required when data is pushed from some sensors and pulled from others. In this hybrid approach pull mechanisms are utilized to cut down on unnecessary sampling of sensors thereby reducing energy consumption and push mechanis ms are utilized to reduce network traffic and latency of query response. The f ollowing sections provide a detailed discussion of all three approaches and describe how a query optimizer generates and chooses the query execution plan which best meets the goa l of minimizing node energy consumption while ensuring that the latency of query response does not go beyond user -specified bounds.

PAGE 36

36 Sensor Querying Strategies We first begin by identifying the type of queries we are targeting for optimization and define some terms which will be used in subsequent sections. Then we provide a description of how a query is disseminated among the nodes for in-network execution based on network topology. Then we cover three querying strategies namely, Push, Pull and hybrid Pul l -Push. For each of the approaches we also define cost functions which will be used by the query optimizer to identify the best query plan. Then, we describe how the query optimizer generates and chooses the best query plan for minimizing energy consumption while ensuring that the response latency remains within user -specified bounds. Finally, we describe how the cost performance of a plan can be monitored during its execution to ensure its effectiveness. Figure 5 2. Query dissemination and evaluation Def initions We focus on continuous selection queries involving multiple predicates applied on multiple sensors, since these types of queries are directly affected by sensor acquisition costs and the order of sampling sensors. We den ote these predicate queries in conjunctive normal form or CNF (Equation 5 1). If we consider any query expressed as a SQL statement such as SELECT d (C12V C41 ) (C32VC51) C41, C42 query processor coordinator/gateway N1 N2 N6 N4 N3 N5 C11, C12C21C31, C32C51C61a b c e (C11V C21) (C11V C21 V C31) (C51V C32 V C42)

PAGE 37

37 FROM WHERE , then Equation 5 1 corresponds to the where -clause -expression. Q = ) (ij j iC (5 1) Each literal Cij involves a predicate applied on a sensor and is of the form P(s), where P is a predicate and P(s) denotes that predicate P is being applied on sensor s. SelP(s) de notes the selectivity of sensor s with respect to predicate P. We assume SelP [0, 1] and is the probability that P(s) evaluates to False. Hence, the higher the value of SelP(s), the more selective the sensor s is with respect to predicate P. SelP(s) can be calculated by the query processor based on sensor history [5 ] or can be obtained directly from statist ics stored onboard the sensor nodes. If the predicate being referred to in the text is unambiguous, we use Sel(s) instead of SelP(s). We denote the cost of sampling a sensor s as E(s) where E(s) denotes energy consumed by the microcontroller for sampling sensor s and is measured in milliJoules (mJ). Note that the microcontroller may consume different amounts of energy based on the type of sensor being sampled. Assume that the node evaluating Cij only transmits the result when Cij evaluates to True. We define the network cost of transmitting the result of a literal Cij as NwkCost(Cij), where NwkCost(Cij) is the total energy consumed (mJ) by nodes in the network for transmitting the result to its destination. This cost is the sum total of the energy spent by each node along the route (including the source and destination nodes) for receiving and transmitt ing the result. NwkCost can be calculated if the network structure and link quality information is available. In a ZigBee network, this information is readily available from the Coordinator (one of whose roles is to maintain an overall view of the network) and the query processor does not need to expend extra effort towards gathering this information from all the nodes in the network. This information is automatically transmitted to the Coordinator by piggybacking it on status update messages.

PAGE 38

38 Query Dissemi nation Users issue queries to the query processor residing in the service layer. The query processor generates a suitable query plan after utilizing its optimizer and injects it into the network via the coordinator/gateway. The query is disseminated into t he network (Figure 5 3) for execution using a n overlay tree structure similar to what is used in TinyDB [ 23]. We will use this example to illustrate all the three strategies namely, Push, Selective Pull and Pull -Push. In this particular example, the query is given by Equation 5 2. Q =) ( ) ( ) ( ) (61 42 32 51 41 12 31 21 11C C C C C C C C C (5 2) Evaluation of each Cij is assigned to the node connected to the sensor associated with the literal. For example, C11 and C12 are pushed to node N1. Hence, a node may have to perform evaluation of literals belonging to multiple clauses. The cost of evaluating the set of literals assigned to a node depends on the selectivity of each sensor involved and its cost of sampling. Since a node is required to evaluate multiple disjunctions of literal s (often belonging to different clauses), the order of sensor sampling is done as follows. For each group of literals assigned to a node belonging to the same clause, the node samples the sensors serially in the ascending order of their selectivity. This e nsures that the sensors which are most likely to satisfy their predicates and hence, cause the disjunction to evaluate to True are sampled first. The sampling halts as soon as one of the literals evaluates to True. If we arrange the sensors in ascending or der of their selectivity and enumerate them an d their associated predicates, then the expected energy cost of evaluating a group of literals assigned to a node belonging to the same clause is given by Equation 5 3. C ( ) (1 i i is P ) = 1 1 1 1 1) ( ) (i i k k k P is Sel s E (5 3) SelP 0(s0) is constant and set equal to 1. A node evaluates groups of literals in ascending order of their expected energy cost. Note that for the sake of simplicity, this expression assumes

PAGE 39

39 that each group of literals is evaluated independently. In reality however, if a sensor belongs to more than one group of literals, it is only sampled once and its reading shared among all the literal evaluations This reduces energy consumption since the Analog to Digital Convertor is only turned on once for sampling the sensor as opposed to multiple times for each literal Push Strategy The Push approach is the most widely adopted strategy for sampling sensors and evaluating queries. We cover this approach for the sake of reference and for deriving cost functions, since we will be using certain aspects of Push based querying in our proposed query plans. During each epoch of execution, nodes sample their respective sensors and create partial evaluation records which they transmit to their parents. Since each node can be assigned literals belonging to multiple clauses, multiple partial evaluation records can originate from a single node. This entire process of evaluation uses the slotted approach of time scheduling as described in [ 22], where each node divides its epoch into multiple slots and requires its children to respond within a certain sub interval. Clauses are evaluated in full at the intermediate node serving as the root of the sub-tree containing all the nodes involved in the clause. For example, in Figure 5 2 the final evaluation of the clause ) (42 32 51C C C takes place at node d. The conjunction of clauses also gets evaluated in a similar manner to give the final result of the query. The roots of all the evaluation sub trees get selected during the query dissemination phase, based on routing of clauses and literals down the cluster -tree network. If a node detects tha t a clause gets split up among its children, then it knows that it is the root of the evaluation tree for that clause. Similarly if a node detects that two or more clauses get routed to different child nodes then it determines it is a root of the evaluation tree for conjunction of those clauses. The intermediate routing nodes are capable of suppressing the transfer of partial evaluation records if they are able to deduce the

PAGE 40

40 final result based on information contained in them. For example, a clause evaluation root node will not transmit the result if it finds out that the clause evaluated to False, since th at will result in the entire query which is a conjunction of clauses to evaluate to False. The energy cost of evaluating a clause using Push strategy can be calculated by modifying the expression given by Equation 5 3 to factor in the cost of evaluating literals and associated network cost for all nodes involved in the clause (denoted by the set Node ). The expected evaluation cost of a clause )) ( (1 i N i N i Node Ns P using push approach (denoted by EvalCostPush) is given b y Equation 5 4. C ( )) ( (1 i N i N i Node Ns P ) = Node N N i i N i PC NwkCost s SelN)) (s P C( ) ( ) ) ( 1 (i N i N 1 i 1 (5 4) N is the Id of a node belonging to the set Node i s SelN i PN 1 ) (0 )) ( (1 i N i N i Ns P C and partial clause evaluation cost C( ) (1 i N i N is P ) is given by Equation 5 3 The sum inside the box brackets gives the total energy consumed by a node for evaluating the literals and transmitting the result over the network if any of the literals evaluates to True. The final summation is over all members of the Node set. TotalCo stPu sh ( ) (ijj iC ) = FinalAggregationCost + i ij i PushC EvalCost ) ( (5 5) In Equation 5 5, Final AggregationCost is the additional network cost required to aggregate partial results created during evaluation of each clause and aggregation of results from multiple clauses to obtain the final result of the conjunction. The NwkCost term is calculated based on the number of hops required to transmit the partial evaluation record to the next intermediate node which will merge it. Fo r example, referring to Figure 5 2 NwkCost(C11) 2 (the number of hops between nodes N1 and b) whereas NwkCost(C12) 3 (the number of hops between nodes N1 and a). In case the evaluation tree structure is s uch that the final result can be obtained as a side -effect of clause evaluation then Final AggregationCost will be minimal

PAGE 41

41 F inally, the total energy cost for in-network evaluation of a query (Q = ) (ij j iC ) using push strategy is given by Equation 5 5. This total energy cost is used by the query optimizer for comparing different plan options to decide on t he most optimal plan for distributed execution of queries injected by the user Selective Pull Strategy The order of evaluation in the Push strategy is governed by the structure of the evaluation tree which in turn depends on the topology of the network. The nave Pull strategy would involve pulling all the sensor readings in parallel and essentially following the same evaluation order as above, but such a plan would be of no use since it will always have higher energy cost due to extra network traffic and double the latency of the Push strategy. Instead of the nave approach, we propose a selective pull strategy where the sampling of sen sors across the network is ordered according to the expected cost of evaluating of clauses involving them and clauses are evaluated in serial order. Consider the example depicted in Figure 5 2 For the sake of discussion, suppose the optimal order of evaluation obtained by the query optimizer (based on ascending cost of evaluation) is: ) ( ) ( ) ( ) (41 12 61 42 32 51 31 21 11C C C C C C C C C The first step is the query dissemination phase has been described before where the entire query is pushed on the network. It has been suggested that for the pull approach, nodes do not nee d to store query information [ 29]. However, we feel that for continuous queries it is more energy efficient and fault tolerant for nodes to store the query instructions even though they do not execute them without the arrival of an explicit command from a parent node The next step involves evaluation of the first clause in the execution schedule. In the above example node a which is the root of the 1st clauses evaluation tree transmits a pull command to

PAGE 42

42 nodes b and c. These nodes in turn relay the command to nodes N1, N2 and N3. The root does not issue 3 pull commands rather the n umber of pull commands issued is equal to the number of its children it needs to transmit the command to (in this case, 2) After eac h of the nodes receive their respective pull commands, they sample their sensors in the order of their sampling cost and selectivity. One important thing to note here is that a pull command does not reference a particular sensor. In fact, when a node rece ives a pull command it samples sensors associated with all the literals assigned to it. Recall that a node can have multiple sensors spread across multiple clauses. Hence, by adopting this strategy a node does not have to sample its sensors multiple times. This naturally saves energy consumption due to networking but it also saves a significant amount of energy consumed by sensor sampling. This is due to the fact that it is far more energy efficient to initialize the ADC once and sample all the sensors rath er than keep starting and stopping it for each individual sensor sampling. Hence, when a node sends a response to a pull command it transmits multiple partial evaluation records up to the root. Once a node finishes evaluation, it transmits the partial rec ords to the root of the evaluation tree. The partial records get merged on their way up to the root and o nce the root receives all the responses and it is able to determine whether the clause evaluated to True or False. If it evaluated to False, the execution of the query for that epoch is terminated since it results in the entire conjunction expression evaluation to False. If the clause evaluates to True, the root of this clauses evaluation tree (node a in the exampl e) transmits the partial evaluation records of the other clauses to the root of the next clauses evaluation tree (node d in the example). This root in turn examines the partial evaluation records and sends pull commands to only those nodes which have se nsors whose readings are not in any of the partial record. The

PAGE 43

43 process of execution continues in this manner till one of the clauses evaluates to False or all clauses evaluate to True, in which case a response is sent to the query processor. We can observe that the effectiveness of the selective pull strategy depends on the order in which clauses are evaluated across the network. Ordering of sensor samples with the aim of minimizing acquisition costs has been studied before by Madden et al in [ 23 ] but thei r mechanism only provides a locally optimal ordering of sampling of sensors connected to the same node rather than a network -wide ordering of all sensors involved in the query. The cost effectiveness of the selective pull strategy depends not only on the cost of evaluating a clause but also on the cost of transferring control from the root of one clause evaluation tree to the next. The formula for calculating the cost of evaluating a clause using pull is similar to EvalCostPush described before except it c ontains an additional term corresponding to the network cost of the pull command. T he expected evaluation cost of a clause using pull a pproach is given by Equation 5 6. EvalCostPu ll (( )) ( (1 i N i N i Node Ns P ) = Node N Push NEvalCost C NwkCost ) ( (5 6) N is the Id of a node belonging to the set Node and )) ( (1 i N i N i Ns P C The cost of transferring control from the root of one evaluation tree to the next is basically the network transmission cost for transferring the partial records to the other root. Both these pieces of information can be easily obtained by the query processor. Given an ordered pair of two clauses, the service layer can provide information as to the possible number of partial evaluation records that need to be transferred based on wh ich sensors are connected to which nodes. Furthermore, based on network information obtained from the ZigBee coordinator /gateway the query processor can also find out the cost of transmitting those records from one root node to the other.

PAGE 44

44 The total expect ed energy cost of evaluating a query (Q = ) (ij j iC ) using pull strategy (where clauses are numbered in ascending order of execution), is given by Equation 57. TotalCostPu ll ( ) (ij j iC ) = i ij i Pull i i iC EvalCost TrfCost p )) ( () 1 ( (5 7) i iTrfCost ) 1 ( denotes the network energy cost of transferring control from root of evaluation tree of clause numbered i 1 to the root of the evaluation tree of clause numbered i. pi denotes the probability that clause i will be executed, that is, it is the probability that execution control will transition from clause i 1 to i. This probability is dependent on the probability that the previous clause i 1 evaluates to True. p1 = 1 and pi= 1 1'i k kp (for i pk = P ( kj jC = True ). This implies that pk = 1 P ( kj jC = False ) which further implies that pk = 1 P ( j False Ckj ). Hence, pk = 1 kjA ss Sel ) ( where Akj is the set of sensors involved in evaluation of clause kj jC Note that there is an assumption of independence of sensor selectivity in the formula for pk. If possible one can refine this probability by using correlation information about various phenomena such as the approach used by Deshpande et al [8 ]. As discussed previously, determining the optimal order of execution is essential for the selective pull approach to succeed. The goal here is to determine the optimal order of evaluating clauses so that the total energy consumption is minimized. Consider a query Q = i iB We can view the execution space as a complete directed -graph G = (V, E) where clause Bi in the query Q is represented by vertex Vi. The weight of the ed ge E (i, j) going from a vertex Vi to Vj is set equal to (Bj)) EvalCost ( 'Pull j i jTrfCost p where p j, EvalCostPull (Bj) and TrfCosti are defined above. The edge weight gives the expected cost of transitioning from Vi and Vj based on transition probability pi Also, a dummy node D is added to G such that E (D, i) is equal to

PAGE 45

45 EvalCostPull(Bi) and E (i, D) = 0, with i ranging from 1 to |V|. Thus, the problem of optimizing the order of clause evaluation with the goal of minimizing total energy consumption can now be s olved as a Traveling Salesman Problem (TSP) with starting point as node D. There have been numerous solutions proposed for the Traveling Salesman Problem such as dynamic programming for finding optimal solutions and heuristic techniques such as simulated a nnealing and local search for near optimal solutions. Since the dynamic programming solutions have time complexity which is exponential in terms of the number of clauses being evaluated, it is more practical to use one of the heuristic techniques in the qu ery optimizer. Hybrid Pull -Push Strategy Both the push and selective pull strategies have their advantages and disadvantages. The push strategy provides greater autonomy to the nodes and works without external supervision. It also will typically have lower latency of query response as compared to selective pull. The selective pull strategy on the other hand, takes into consideration the fact that sensing costs now significantly dominate the total energy consumption and tries to optimize the order of sensor sampling across the network in order to minimize energy consumption. However, this comes at a cost of greater reliance on the network and higher latency as compared to the push approach, since each node is subject to e xternal supervision from a parent node We feel that one of the best ways to utilize the strong points of both push and pull strategies is to adopt a hybrid pull push approach. The main goal of the hybrid pull -push approach is to minimize the total energy required to execute a query in the ne twork while ensuring that the latency of query response is within bounds specified by the user. In order to achieve this, the plan utilizes the push approach on a number of sensors based on cost and latency considerations and utilizes the selective pull ap proach on the rest of the sensors. After query dissemination, the group of sensors corresponding to a set of clauses having the lowest

PAGE 46

46 evaluation cost is asked to execute the push strategy. Since the query i s a conjunction of clauses therefore, only if all the initially selected clauses evaluate to True, are the other clauses evaluated using the selective pull approach. The construction of a hybrid pull -push plan is described in more detail when we discuss how the query optimizer chooses the best plan of execution. The energy cost of a hybrid pull push plan depends on which of the clauses were evaluated using push approach and which were evaluated using selective pull. The total energy cost is simply the sum of the energy cost of each clause calculated by applying the appropriate cost formula for the push or pull approach. Suppos e the query in question in Q = i n iB1 where the clauses are numbered in the order of execution. Suppose the hybrid plan calls for m clauses to be evaluated using pu sh and the rest evaluated using pull, thereby implying in this case that the first m clauses will be evaluated using push. T he total energy cost of the hyb rid pull push plan is given by Equation 5 8. TotalCostHybrid ( i n iB1 ) = m i i PushB EvalCost1) ( + n m i i Pull i i iB EvalCost TrfCost p1 ) 1 ()) ( ( (5 8) Choosing the Best Query Plan Heidemann et al [12 ] proposed that data dissemination algorithms need to be mapped to application requirements. Through our hands on experience we found that the actual fulfillment of these application requirements eventually depends on the specific sensor hardware that is being targeted. Hence, the query optimizer needs to be flexible enough so that it can generate multiple query plans which exploit specific characteristics of differe nt hardware with the aim of satisfying application requirements without attempting to generate a one -size -fits all solution. The Sensable query optimizer generates all three types of query plans discussed in the preceding sub -sections and chooses the one which best minimizes the energy consumption while meeting

PAGE 47

47 the user requirements on latency of query response. To aid the query optimizer in making a decision a user is required to provide information regarding its tolerance for latency. Instead of requiri ng the user to provide a hard number, the query optimizer asks for tolerances in terms of percentage (denoted by Lmax). The percentage value represents the magnitude by how much the latency can exceed that of the Push approach. If LPush and LPlan respectively denote the latencies for push approach and the plan that was chosen then, max100 L x L L LPush Push Plan must hold true. If we consider a query Q = i iB then the latency for evaluating it using the push approach is given by Equation 5 9. The latency for the selective pull approach is given by Equation 5 10. The latency of a hybrid plan is calculated as the sum of latencies due to its push and pull operations. If we consider the hybrid pull push plan then its query response latency is given by Equation 5 11. LPush= i i Push PushB Latency gation FinalAggre Latency ) ( (5 9) LatencyPush(Bi) denotes the latency in receiving a response for clause Bi if it evaluates to True, and LatencyPushFinalAggregation denotes the latency in generating the final results from results of the clause evaluations. For the sake of simplicity, these individual latency values can be simply calculated as the total number of hops between the nodes evaluating Bi and the root of Bis evaluation tree. Given latency tolerance Lmax, the query optimizer undertakes the following steps to d etermine the optimum query plan. First it generates a query plan using the push strategy and calculates the energy cost (TotalCostPush) and latency ( LPush). Next it generates a query plan using the selective pull approach by solving the optimization problem using simulated annealing and calculates the energy cost (TotalCostPull) and latency (LPull). Finally, it generates a hybrid pull push query plan a s follows. Suppose the selective pull query plan generated for user defined query Q = i iB resulted in a query plan with the following order of evaluation of clauses: B1, B2,

PAGE 48

48 B3, ... Bn. The query optimizer progressively replaces the pull action associated with each clause with push, starting from the first clause. Each time it does this, it recalculates the latency of the plan (LHybrid) and checks if its percentage relative difference with LPush is less than Lmax or not. This process is continued till the relative difference in latency becomes less than Lmax. If the final hybrid plan consists of m push based evaluations followed by (n -m) selective pull base d evaluations then there exists no k < m for which Equation 512 holds true. LPull= i i i i Pull iTrfLatency B Latency p ) ) ( () 1 ( (5 10) LatencyPull(Bi) is calculated as twice the total number of hops between the nodes evaluating Bi and the root of Bis evaluation tree. pi is as defined before. i iTrfLatency ) 1 ( is the latency in transferring con trol from the root of clause (i 1)s evaluation tree to clause is evaluation tree. LHybrid= m i i Push PushB Latency gation FinalAggre Latency1) ( + 1 ) 1 () ) ( (m i i i i Pull iTrfLatency B Latency p (5 11) k i i Push PushB Latency gation FinalAggre Latency1) ( + max 1 ) 1 () ) ( ( L TrfLatency B Latency pk i i i i Pull i (5 12) Finally, the query optimizer compares the energy costs of all the plans which meet the users latency bounds and chooses the one with the lowest cost. Typically one would expect that the hybrid pull -push plan will be always chosen. However, this may not be true in every case since the cost effectiveness of the plan depends heavily on energy consumption characteristics of the sensors and the sensor platform hardware. Hence, the approach outlined above allows the optimizer to fall back on traditional approach es such as Push, if the new proposed strategies do no prove to be more cost -effective. Monitoring Plan Performance Since the query plan is generated based on a snapshot of history its effectiveness in minimizing energy consumption can decrease over time du e to changing conditions in the

PAGE 49

49 deployment environment. In order to monitor the effectiveness of the query plan being executed, each node can keep track of selectivity of its sensors and calculate the expected cost of evaluation of literals assigned to it This information can be periodically transmitted up the evaluation tree as a partial record getting merged along the way, in the same manner as the query evaluation records. This will lead to the root of each clauses evaluation tree to have updated cost estimates for that clause. For selective pull and pull -push approaches, the total cost also depends on the probability of transferring control from one root to another. Hence, each evaluation tree root also factors into the cost calculation, the expected n etwork cost based on the actual transition probability of transferring control to the next root in the execution sequence. Finally, the query processor in the service layer compares the actual updated cost with the estimated cost and if required, regenerat es plans and chooses the best one using more upto -date selectivity data. The new plan can either be disseminated in its entirety into the network or the query processor only disseminates those portions of the new plan which differ from the existing plan. Selectively adding or removing query execution assignments from specific nodes in this manner, without having to reset the entire query execution process, leads to greater energy efficiency and higher system availability. Fault Tolerance In this sub -sectio n we discuss some mechanisms for dealing with node and network malfunctions, since failure is an integral part of any sensor network. The query execution strategies described utilize in -network evaluation trees and hence, are vulnerable to failure of any of the nodes participating in the evaluation process or nodes which are not participating in query evaluation but are nonetheless playing a vital role by linking up different segments of an evaluation tree. The following are possible types of failure that can affect the query execution plans that are executed in network :

PAGE 50

50 1 Node whose sensors are involved in evaluation of a clause fail : In such a case, the ancestor of this node responsible for merging its evaluation record will not receive any messages from it If the node is using the push approach, there are a number of techniques that have been suggested to cope with failure su ch as the use of child caches [ 22] and Bayesian techniques to determine whether the lack of response indicates suppression or node failure [30]. In case the node was using the pull approach then its failure will get detected by the next epoch of execution when a pull command is issued. Th is is due to the fact that ZigBee aims at providing reliable message delivery through acknowledgement mechanisms present in the link layer and above. Hence, a sender is able to determine if its message got delivered or not. The effect of failure of sensors participating in a query depends on the type of query that was issued. For queries targeting specific sensors, the failure of one or more of those sensors may result in the query execution being halted, whereas queries targeting the entire network may not be affected drastically by such failures. 2 Non -parti cipating intermediate node failure : In case an intermediate router node which is not participating in query evaluation fails, then its child nodes simply re associate themselves with a new parent in the network. This is a feature of ZigBees self -healing characteristics which allows the network to cope with loss of a limited number of routers without affecting network connectivity. 3 Root of evaluation tree fails: This is the most serious type of failure t hat can affect query plan execution since the failure of the root will not only cause the evaluation of its clause to fail but might also lead to the entire query execution getting stalled depending upon the location of that node in the network. To handle root failure we utilize a leader election algorithm. Leader election algorithms for ad hoc networks have been previously proposed [24, 33] and we follow a similar philosophy. We propose that for each node which is the root of an evaluation tree or sub tree should inform a small set of its immediate neighbors (which are within 1 -hop broadcast range ) about its status and have them cache all the information necessary to act as a root node. These neighbors are responsible for monitoring whether the root is onl ine or not. Since they are only 1 hop away, the network overhead and network energy cost is not that significant. When one of the nodes discovers that the root in its neighborhood is dead it can take over the role of the root since it already has all the n ecessary information. To avoid multiple nodes from detecting failure at the same time and becoming roots of the same tree, we can have each node monitor the roots health in round robin fashion for a certain number of contiguous epochs. One of the most im portant practical issues which will occur in the above scenarios is the fact that whenever a node rejoins the network it is assigned a different network address by its new parent. Similarly, when the root of an evaluation tree fails and a new root takes ov er, naturally the network address of the new root wont match the address of the original root. This can create significant problems since other nodes in the tree will be unaware of such changes and

PAGE 51

51 will be unable to transmit readings or commands. Fortunat ely, ZigBee provides a solution to deal with such issues in the form of logical addresses. A logical address is a n application -specific address ind ependent of the network address which can be user -defined and used for communication between applications. During query dissemination phase, we can assign each node in the tree a unique logical address which it retains for the duration of the query lifetime. Hence, even if a node rejoins the network by associating with a new parent its logical address still remai ns the same and gets automatically mapped to its new network address. Similarly, when a node takes over as the root of an evaluation tree it can changes its logical address to that of the deceased root. This ensures that it is able to receive evaluation re cords from the other nodes in the tree without having to inform them of the change. Experimental Performance Analysis W e compare and analyze each of the query plan options discussed in this paper, based on two performance metrics: energy consumption and l atency of query response. We studied the effect of varying sensor selectivity and the numbers of sensors operating in push and pull modes on these two metrics. Method of Experimentation Our experiments consisted of simulating the construction and in -network execution of all the three types of query plans discussed in this paper namely, Pus h, Selective Pull and hybrid Push Pull For analyzing the effect of varying the numbers of sensors operating in push and pull modes, the simulator randomly constr ucted a set of 5 selection queries involving multiple predicates. Each query involved the participation of 16 sensors in the form of range filter predicates. The 16 sensors were generated such that the selectivity of their associated range filters followed the Gaussian distribution with mean 0.66 and variance 0.33. F or each iteration, the simulator randomly generated a multi hop network of sensors nodes (each responsible for the

PAGE 52

52 operation of one or more sensors participating in the query) and generated quer y plans as described previously. Then, it simulated the execution of each plan using hardware specifications of the Atlas ZigBee node (Table 5 1) For studying the effect of varying mean sensor selectivity, for each iteration the simulator generated a new group of sensors, such that each new groups selectivity followed the Gaussian distribution with mean equivalent to the sensor selectivity for that iteration and variance equivalent to one third of the selectivity magnitude. Table 5 1. Atlas ZigBee node ha rdware specifications Operation Current (mA) Duration (seconds) Sampling Sensors, Processing or Listening for messages 6.0 2.0 Receive Message @256Kbps 15 0.002 Transmit Message @256 Kbps 15 0.002 Results and Analysis First we look at the effect of varying the numbers of sensors operating in push and pull mode, on energy consumption per epoch (Figure 5 3). We observe that the energy consumption (measured in milliJoules per epoch) increases steadily as the ratio of the number of sensors operating in push mode to the number of sensors operating in pull mode increases If we look at the two extreme cases we observe that the energy consumption for the conventional all push plan is approximately 2.5 times more than the energy consumption of the selective pull plan. This is due to the fact that in case of the all push plan, each sensor is sampled independently of others in a parallel manner hence, regardless of whether the sensors reading satisfies the associated range filter or not, the node incurs the e nergy cost due to sensor sampling. However, in case of the selective pull plan, sensors are sampled in a linear fashion so that if a sensors reading fails to satisfy its range filter, all subsequent sensors which were supposed to be read after it are not sampled. For ZigBee based sensor nodes such as the ones used by Atlas, the sensing cost makes up for an overwhelming majority of the total energy consumption (Figure 5 -

PAGE 53

53 1). Hence, the all push plan has much higher energy costs than the selective pull and an y of the hybrid push -pull plans. We notice that in case of the hybrid pushpull plans, their energy consumption falls somewhere between selective pull and all push plans. The hybrid plans which have more number of sensors operating in push mode have higher energy costs as compared to ones which have comparatively more number of sensors operating in pull mode. Figure 5 3. Comparing energy consumption of different query plans However, if we look at the latency of response for the various query plans, we fin d a different picture altogether (Figure 5 4) We observe that the latency of response decreases as more sensors operate in push mode as compared to pull mode. If we consider the two extreme cases, we find that the average latency in response for the selec tive pull plan is nearly 20 times more than the all push plan. However, in case of the hybrid push-pull plans as the number of sensors which operate in push mode increases the latency comes down drastically. This is due to the fact all sensors operating in push mode are sampled in parallel whereas sensors operating in 0 100 200 300 400 500 600 700 Total Energy Consumption per Epoch Energy Consumption due to Sensing Energy Consumption due to Networking Energy Consumption per Epoch (mJ)Push:Pull Ratio

PAGE 54

54 selective pull mode are sampled in serial order with the control of the execution process passing from one clause to another, which results in higher latency. This clearly brings forth the que stion of a tradeoff where a user has to decide whether energy consumption is of greater concern or the latency of response. In most cases, we feel one of the hybrid pushpull plans will usually satisfy a users requirements by providing more energy efficie ncy as compared to the conventional all push plan at the cost of a small increase in latency of response. Figure 5 4. Comparing latency of response for different query plans Next, we look at the effect of varying sensor selectivity on the energy consumption of different query plans (Figure 5 5). We compare the performance of the all push plan with that of the best hybrid query plan that was generated by the query optimizer (in terms of energy consumption per epoch). We observe the energy consumption for the all push plan does not change significantly with the change in average selectivity of sensors. This is due to the fact that in case of push, the selectivity only affects the fact whether a reading is transmitted over the network or not. Since the network cost in case of ZigBee -based hardware is a tiny fraction of the overall energy cost hence, it does not affect the energy consumption significantly. 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 Push:Pull RatioLatency per Response (seconds)

PAGE 55

55 Figure 5 5. Effect of selectivity on energy consumption of query plans In case of the hybrid push -pull plans we notice that for low sensor selectivity the best plan is usually the one with a high number of sensors operating in push mode and consequently, a very low number of se nsors in pull mode. This can be explained by the fact that since average selectivity is low hence the probability a sensors output satisfies the associated range filter is quite high. Hence, in such a case the selective pull strategy will perform much wor se than the push strategy since most of the time almost all the sensors will have to be sampled. The push strategy will be able to handle this in a much more energy efficient way since it does not require sensors to be sampled in serial order and hence, do es not incur the extra network cost of hopping from one clause to another. However, as the average selectivity increases the hybrid plans with higher number of sensors operating in pull mode prove to be more energy efficient. Furthermore, the energy consum ption of the hybrid plans also drop drastically. This is due to the fact that as selectivity increases the probability that a sensors output fails to satisfy its associated filter becomes higher. This in turn implies that in case of selective pull, the probability that the 0 100 200 300 400 500 600 700 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 AllPush Best Push Pull Plan Energy Consumption per Epoch (mJ)Mean Selectivity Push1:Pull15

PAGE 56

56 subsequent sensors do not have to be sampled also becomes higher. Since sensing cost is a major portion of the total energy cost hence, the energy saved by not sampling these sensors leads to a significant reduction in total energy consumption. We also notice that for selectivity less than 0.3, the all push plan is the best in terms of energy consumption. However, as soon as selectivity exceeds 0.3, the selective pull plan seems to have the lowest energy drain. The reason for such a dramatic shift from all push to selective pull is due to the fact that when selectivity is less than 0.3, the deciding factor is the network energy cost. Whereas, when selectivity exceeds 0.3, the deciding factor becomes t he sensor sampling cost. T he cos t of sampling one sensor has a much larger magnitude than the cost of transmitting one network packet. Hence, for higher values of selectivity, reducing the probability of sampling even one additional sensor leads to a larger impact on total energy consumption which usually makes selective pull the most energy efficient plan in such cases. If we consider the effect of varying selectivity on the latency of query plans (Figure 5 6), we observe that the all push plan is consistently better in terms of lower l atency as compared to the best hybrid push -pull plan generated by the query optimizer. This is due to the fact that selectivity does not affect latency of the all push plan. However, it does negatively affect the latency of any query plan which has a selec tive pull component. Since selectivity determines the number of sensors that will eventually get sampled using a selective pull strategy hence, having low selectivity increases the latency of the plan significantly. As average selectivity increases the latency of the hybrid plan decreases somewhat however we note that it stays within the range of 0.03 to 0.04 seconds. This is due to the fact that any gains made due to not sampling some sensors might get offset by the additional latency brought on by having to hop from one clause to another in a serial fashion. We also observe that for extremely high selectivity, the best hybrid

PAGE 57

57 plan is actually the selective pull plan since it takes advantage of the fact that a sensors output will fail to satisfy its associated range filter most of the time hence, eliminating the need to sample the rest of the sensors. Figure 5 6. Effect of selectivity on latency of query plans Next, we look at the effect of the varying the number of sensors participating in the query, on energy consumption (Figure 57). We observe that the energy cost of the all push query plan increases almost linearly as the number of participating sensors increases. This can be attributed to the fact that since all sensors are sampled in parallel and th e sensing cost is by far the largest contributor to the total energy consumption hence; as the number of sensors increase the total energy consumption goes up by nearly the same factor. However, in case of best push -pull query plans we observe that the inc rease in energy consumption is much more gradual. Moreover, we notice that the relative difference between the energy costs due to the all push plan and the best hybrid plan increases as the number of sensors increase. This is due to the fact that unlike a ll push strategy, in case of selective pull strategy simply increasing the number of sensors 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 AllPush Best Push Pull Plan (in terms of Energy Consumption) Mean SelectivityLatency per Response (seconds) Push1:Pull15 Push1:Pull15 Push1:Pull15 Push1:Pull15 Push1:Pull15 Push1:Pull15

PAGE 58

58 participating in the query does not necessarily imply that those sensors will be sampled every time the query is executed In fact, as explained previously the que stion of how many sensors are sampled during each execution depends on sensor selectivity rather than the total number of sensors participating in the query. H ence, the total energy consumption does not increas e by the same factor and we can say that the h ybrid plans or selective pull plans scale better in terms of number of sensors participating in the query. Figure 5 7. Effect of number of sensors on energy consumption of query plans Finally, we take a look at the effect of varying the number of participating sensors, on the latency of query plans (Figure 5 8). We find that the latency of the all push plan remains more or less uniformly low regardless of the number of participating sen sors. This is due to the fact that all sensors in the all push plan are sampled in parallel and their readings pushed and merged up the network. So the effect on latency due to increasing the number of sensors is minimal. On the other hand, for selective pull and hybrid plans, the latency increases as the number of 0 500 1000 1500 2000 2500 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 AllPush Best Push Pull Plan Energy Consumption per Epoch (mJ)Number of Sensors participating in QueryPush0:Pull4 Push0:Pull8 Push0:Pull12 Push0:Pull20 Push0:Pull24 Push1:Pull27 Push0:Pull32 Push2:Pull34 Push0:Pull40 Push0:Pull44 Push0:Pull48 Push0:Pull52 Push1:Pull55 Push0:Pull60

PAGE 59

59 participating sensors increase. This is due to the fact that in these plans at least some of the sensors are sampled in a serial fashion using selective pull strategy. Hence, as the number of sen sors increases the length of the queue of sensors being sampled increases and so does t he latency of response Figure 5 8. Effect of number of sensors on latency of query plans In conclusion, we can state that in general if the user is only con cerned about conserving energy then the selective pull plan will usually be the plan of choice. On the other hand, i f fast response time and low latency is the primary concern then the all push plan will be the one desired. However, if a user wants to cons erve energy but is willing to tolerate somewhat slower response times then one of the intermediate hybrid push-pull plans might be chosen where the specific ratio of sensors operating in push mode versus pull mode will be determined by the users preferred tradeoff between energy efficiency and latency. In the end, the choice of query plans is determined by application requirements and hardware capabilities of the deployed sensor network. 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 4 8 12 16 20 24 28 32 36 40 44 48 52 56 60 AllPush Best Push Pull Plan (in terms of energy consumption) Push0:Pull4 Push0:Pull8 Push0:Pull12 Push0:Pull16 Push0:Pull20 Push0:Pull24 Push1:Pull27 Push0:Pull32 Push2:Pull34 Push0:Pull40 Push0:Pull44 Push0:Pull48 Push0:Pull52 Push1:Pull55 Push0:Pull60Latency per Response (seconds)Number of Sensors participating in Query

PAGE 60

60 CHAPTER 6 VIRTUAL SENSORS FRAM EWORK Most sensor network database syst ems are geared toward answering user queries involving raw data, directly originating from individual sensors. However, as sensor network applications become more sophisticated, there is a real requirement for mechanisms which enable users to query data ty pes of more abstract nature and over a larger spatial scope with user -specific reliability requirements. We define virtual data as a type of data whose values are a function of a variable number of multiple sensory inputs. This can be due to inherent s ophistication of the data type itself or the spatial scope within which it is being queried. An example of virtual data would be weather. There is no single sensor which can detect weather since it is an amalgamation of different types of environmental inp uts such as temperature, atmospheric pressure and humidity. But if users are only interested in querying weather conditions, it would be inefficient on their part to individually query temperature, pressure and humidity sensors and then aggregate their rea dings every time the query is executed. Furthermore, user applications may not have access to domain specific knowledge required to fuse raw sensor inputs into the desired output. Another example of virtual data is when a user queries for the temperature of a room which has multiple sensors deployed in it. Even though the actual data type is still primitive and can be answered by a single sensor, the query itself involves a specific spatial scope which is larger than each individual sensor. The user is onl y interested in the temperature of the whole room and not that being reported by each indivi dual sensor. In such a case it is also more efficient for the query processing system to automatically abstract away the presence of multiple sensors and behave as if there is one large temperature sensor covering the entire room. Putting the onus of aggregating readings over the spatial domain or multiple inputs on the application not only

PAGE 61

61 makes it in efficient and expensive, but also in feasible due to the dynamic nature of sensor networks where sensors can come online or fail wit hout warning. To address these requirements, w e propose a Virtual S ensors Framework which utilizes SOA mechanisms to provide on demand creation and life -cycle management of virtual sensor s and process queries involving abstract derived types of data. Virtual sensors are software sensors which do not exist physically; however, they interact with applications as service s just like the physical sensor services Moreover, virtual sensors get deployed on an on -demand basis depending on queries currently executing in the network. Based on queries issued by users, the virtual sensor framework utilizes the service -oriented architecture of SOSNs for dynamic composition of physical sensor services i nto virtual sensors. As part of the Virtual Sensors Framework, we developed distributed, in network algorithms for enabling virtual sensors and monitoring their data quality. Furthermore, we also developed distributed fault tolerance mechanism s which compe nsate for failure s of multiple physical member sensors. The rest of this chapter is organized as follows. We begin by introducing the concept of virtual sensors which enable s the querying of virtual data. Then, we classify the different types of virtual sensors accord ing to their roles and discuss the virtual sensor framework model and its components Next we describe the on -demand creation of virtual sensors with the help of an example and show how the virtual sensors framework enables a smart space to detect when its sentience gets enhanced. Then, we provide detailed descriptions of b asic and derived virtual sensor operations including their distributed in -network execution; fault tolerance and data quality monitoring algorithms. Finally, we perform experimental analysis of fault -tolerance and data quality monitoring algorithms an d compare the energy efficiency of the virtual sensors f ramework with that of other co ntemporary virtual sensor systems.

PAGE 62

62 The Concept and Classification of Virtual Sensor s We define a virtual sensor as a software sensor service entity consisting of a group of physical or other virtual sensors along with associated knowledge which enables qu ery processing involving virtual data. We classify virtual sensors into three categories, namely, Singleton virtual sensor Basic virtual sensor and Derived virtual sensor. Each category of virtual sensor is modeled as a set of tuples which represents the knowledge possessed by them. Singleton Virtual Sensor This type of virtual sensor represents a single physical sensor. It contains specific knowledge about the sensor in form of attributes such as location, the phenomena it detects, unit of measurement et c. For our purposes we assume that a physical sensor outputs a numeric value which is then converted using a conversion formula into a real -world measurement. The role of the singleton virtual sensor is to enable the creation and composition of other virtual sensor types. We model a singleton virtual sensor as a 5 tuple (Equation 6 1) S = < I, L T, U, F> (6 1) where, I: Sensor ID L: Location of sensor T: Type of phenomenon detected (temperature, velocity etc.) U: Unit of measurem ent for the phenomena F: Convers ion formula F: X nu meric readings of the physical sensor and Y is the range of real world measurements. Basic Virtual Sensor A basic virtual sensor is composed of a group of singleton virtual sensors of the same type. A basic virtual sensor detects the same type of phenomena which is detected by its member singleton virtual sensors. The role of the basic virtual sensor is to ans wer queries on virtual data which specify the spatial scope and hence, require the abstraction of multiple sensors of the same type deployed in the space into a single sensor entity. For example, a basi c virtual temperature

PAGE 63

63 sensor for a room will be compos ed of multiple singleton virtual sensors of type temperature which are deployed in that room Equation 6 2 describes a basic virtual sensor. B = < I, L T, U, F, S > (6 2) where, S: Set of singleton virtual sensor instances which are members of this basic virtual sensor. F: Aggregation f ormula where F = f ({Rs }); Rs is a reading from singlet on virtual sensor s and range of F is the set of aggregated readings. I, L, T and U have their usual meanings defined above. Derived Virtual Sensor A derived virtual sensor is composed of a group of basic and/or other derived virtual sensors of heterogeneous types The role of a derived virtual sensor is to answer queries on virtual data which cannot be answered by individual singleton or basic virtual sensors. This is due to the fact that the type of data required as output is too abstract to be generated by any of the other individual sensors. Consider an example where the application wants to query the weather in a specific location which con tains a deployment of multiple temperature, atmospheric pressure and hu midity sensors. These physical sensors can be encapsulated as singleton virtual sensors, which in turn can be grouped by type into basic virtual sensors Finally a derived virtual sen sor for sensing weather can be composed out of these basic virtual sensors. A derived virtual sensor is modeled as shown in Equation 6 3. D = < I, L, Y, Z, F, M, R> (6 3) where, Y: Set of basic/derived virtual sensor type s which are required by the derived sensor. Z: Set of basic/derived virtual sensor instances which are members of the derived sensor. F Formula for integrating various sensor readings where F=f( {Rz }), where Rz is a reading from basic/derived virtual sensor z Z. M: F the set of possible responses that can be given by this derived virtual sensor. I and L have their usual meanings An interesting feature of a derived virtual sensor is that a large number of such sensors can be created from a sma ll finite set of physical or singleton sensor services. Simply changing the

PAGE 64

64 logical glue which fuses inputs from such sensors in a domain-specific way can provide a rich set of sensing capabilities to the space without any additional hardware deployment cost. System Framework for Virtual Sensors W e propose a framework model (Figure 6 1) for managing the life cycle of virtual sensors. The framework resides as component of the Sensable query processor in the service layer and consists of the knowledge base, f ramework controller and the virtual sensors currently active in the network. Figure 6 1 Virtual Sensors Framework a rchitecture The Knowledge Base The knowledge base is the central repository of information of the framework model. It is responsible for managing the storage, addition, modification and deletion of information records associated with virtual sensors running inside the framework. It stores the following types of Framework Controller Derived Virtual Sensors Basic Virtual Sensors Singleton Virtual Sensors Physical Sensors Virtual Sensors Framework Knowledge Base

PAGE 65

65 information records which are availa ble for lookup by both the framework controller and active virtual sensors: 1 Sensor Model Definition: These records store the model info rmation for each of the 3 types of virtua l sensors defined before: singleton, basic and derived 2 Virtual Data definition : These records store the definitions of various types of virtual data that can be detected by the sensor network. Examples include basic phenomena such as temperature and humidity and more complex phenomena such as weather, ambience and security. A virtua l data definition is stored as a 2 tuple: V = ; where, N is the identifying name of the phenomenon and D is stored model definition of the derive d / basic virtual sensor which can detect this type of data. Framework Controller The framework contr oller is responsible for the overall operation of the Virtual Sensors Framework. It receives virtual data que ries from the query processor and then determine with the help of the knowledge base, which virtual sensors need to be created. It is also responsi ble for removing virtual sensors which are not being used anymore Furthermore, it also prompts the query processor whenever the sentience of the smart space gets upgraded as a result of introducing a new physical sensor into the space. We assume that a q uery from a user application i s of the form SELECT FROM [] The query processor extracts the type of virtual data being queried, the location where it has to be detected and the quality threshold and passes it on to the framework controller. The framework controller then starts the process of creating and organizing virtual sensors in the framework to enable the sensor network to work towards fulfilling that query. This process is d escribed in further detail in the following sections T he framework controller also keeps track of sensors joining the network and for each new sensor, it creates a singleton virtual sensor. An important point to note here is the fact that basic and derived virtual sensors are created and destroyed based on user requirements or as a

PAGE 66

66 result of quality control, but singleton virtual sensors are created whenever new physical sensors join the sensor network and are destroyed when the physical sensor they are representing fails or goes offline. On -Demand Creation of Virtual Sensors The virtual sen sor framework performs dynamic composition of sensor services in response to a users requirements. It only requires a user to state the type of phenomenon a nd the location where it is to be sensed at. Based on the type of phenomenon that has to be sensed the framework controller looks up the corresponding definition from the knowledge base. This definiti on contains a reference to the sensor model definition of the basic or derived vir tual sensor which can detect that phenomenon The fram ework controller retrieves the sensor m odel Definition from the knowledge base and sets some of the attributes in the model definition such as the sensor ID and sensor locatio n. Based on the model definition, it then creates a new service object representing this virtual sensor. This newly created virtual sensor in turn, uses information from its model definition to create or access other service objects representing virtual se nsors required by it and the process continues. Once all the virtual sensors have been created and organized the user is able to retrieve data from the sensor network. In case a virtual sensor fails to start due to a software or hardware error, it returns an error message which is forwarded back up the hierarchy till it reaches the framework controller. The framework cont roller in turn notifies the query processor that its query cannot b e executed at the present time. Virtual Sensor Composition Graph To fa cilitate the on demand creation of virtual sensors and to enable the smart space to detect when its sensing capabilities are enhanced due to introduction of a new p hysical sensor into the space, the knowledge base utilizes a sensor composition graph with h ash indexed nodes (Figure 6 2) The sensor composition graph is created from the sensor model definitions stored in

PAGE 67

67 the knowledge base. Each node in the graph represents a singleton, basic or derived virtual sensor and contains its sensor model definition The sensor composition graph is a directed graph where an edge from sensor A to sensor B indicates that sensor B is a member of sensor A. For example, there exists an edge from the node representing virtual temperature sensor to the node representing vir tual ambience sensor (Figure 6 2). Each node in the graph is also hash indexed with phenomena types being used as keys. This allows constant time look ups to initiate the virtual sensor creation process depending on the type of basic or derived phenomenon that has to be sensed, without having to traverse the entire graph. We utilize an example scenario to demonstrate the utility of the sensor composition graph in subsequent section s Figure 6 2 Sensor composition g raph Smart Space Ambience Sensor Example We present an example scenario to demonstrate how all the components of the virtual sensor f ramework come together and work in unison to perform automatic onthe -fly service composition to enable the on -demand creation of virtual sensors Virtual Ambience Sensor (derived) Virtual Temperature Sensor (basic) Virtual Humidity Sensor (basic) Temperature Sensor (singleton) Humidity Sensor (singleton) Ambience Humidity Temperature

PAGE 68

68 Consider a sens or network consisting of temperature and humidity sensors deployed in a smart space which is divided into 4 rooms A, B, C and D ( Figure 6 3 ) T he knowledge base contains the sensor model definitions of the following virtual sensors: Singleton Temperature s ensor (denoted by a solid black circle), Singleton Humidity sensor (denoted by a solid black square), Basic Temperature sensor (T), Basic Humidity sensor (H) and Derived Ambience (W). It also contains a user -defined data type definition < Ambience W> where W represents the model definition of the derived a mbience sensor. Figure 6 3 Sensing ambience using virtual s ensors Suppose a user sends the following query to the query processor: SELECT Ambience FROM A. The query processor extracts the vir tual data type, Ambience and t he location, A from the query, which it forwards to the framework controller. The framework controller then looks up the knowledg e base to retrieve the phenomenon definition for Ambience Since the phenomenon name is has h indexed to the virtual sensor which has the capability to sense D C A Physical Temperature Virtual Sensor Physical Humidity Virtual Sensor Framework Controller Virtual Data Type = Ambience, Location = A Derived Ambience Sensor WS 1 Basic Temperature Sensor BTS 1 Basic Humidity Sensor BHS 1 B Derived Ambience VS Model W = Basic Temperature VS Model T = Basic Humidity VS Model H = Knowledge Base W = T = H =

PAGE 69

69 ambience it retrieves the reference for the node representing the derived ambience sensor. From the node i t gets the sensor mo del definition for the derived ambience s ensor. It sets the loc ation attribute in the model definition to A and creates a sensor service object representing a n ambience sensor. From the outbound edges of its node in the sensor composition graph, the ambience sensor knows that it needs to get data from basic temperat ure and humidity virtual sensors located at A. The ambience sensor then retrieves the sensor model definitions of these basic virtual sensors and sets their location attributes to A. It then uses the model definitions to create a new service object rep resenting each basic virtual sensor. Each of these basic virtual sensors then figures out from the sensor composition graph that they need to aggregate readings from singleton virtual sensors of their respective types, which are located inside A. Using mechanisms for service discovery and composition provided by t he underlying service framework the basic virtual sensors are able to access and aggregate data from the singleton temperature and humidity sensors. Once all the member virtual sensors are active and data starts flowing back up the chain, the ambience sensor becomes active Detecting the Enhanced Sentience of a Smar t Space The role of the sensor composition graph is not limited to the creation of virtual sensors. In fact, it plays a very important role in enabling a smart space to automatically detect whenever its sentience is enhanced due to introduction of new phys ical sensors into the space. The framework controller maintains two lists of virtual sensors: (1) Potential List: A list of sensors which can be created and; (2) Active List: A list of active sensors which have been created. The active sensors are created on-demand by traversing the sensor composition graph, as described previously. However, to maintain the potential list essentially requires a reverse lookup of the sensor composition graph. When a new physical sensor connected to an Atlas node is introduc ed into the smart space and powered on, it automatically becomes available as a singleton sensor. The

PAGE 70

70 virtual sensors framework then performs a reverse lookup by traversing all the edges inbound into the node corresponding to the sensor type of the new sensor. This process is recursively done till all back edges have been traversed. Figure 6 4 Initial sensing capability of the smart space The framework controller adds any basic virtual sensors which is part of the traversal path into the potential list. Moreover, it also adds into the potential list, any derived virtual sensors which have all their member sensors in the act ive list or potential list. Consider the ambience sensor example described previously and assume that initially only temperature senso rs are deployed in the space (Figure 6 4 ). In that case the smart spaces initial sensing capability is limited only to temperature. However, as soon as humidity sensors are introduced into the space, the virtual sensors framework is able to detect that no t only can the smart space detect humidity and temperature but can now also detect ambience (Figure 65 ). But one must note that, none of the virtual sensors are actually created till a user application explicitly issues a query involving

PAGE 71

71 one of them. In t his manner, it is able to prompt the user whenever the smart spaces sentience gets enhanced both in terms of new primitive and virtual data types. Figure 6 5. Effect of introducing humidity sensors into the smart space Operations of a Basic Virtual Sens or When the user queries for a basic virtual data type, the service layer looks up a knowledge base in the virtual sensor framework to find out which virtual sensor to create. This is done using sensor definitions stored in the knowledge base and using standard SOA mechanisms to compose the required singleton/physical sensor services. The service composition takes place in the centralized service layer and the nave way would be to stream up all the data from each of the physical sensors for centralized aggregation. However this is clearly not scalable when the number of sensors involved is large. Furthermore, streaming up all the data through a multi -hop mesh network will lead to excessive latency and waste of network bandwidth and node energy. To addres s the issue of scalability, we push down the process of basic virtual sensor aggregation from the service layer on to the sensor nodes so that it runs in a distributed in

PAGE 72

72 network fashion without requiring centralized processing. The aggregation is done usi ng an aggregation tree overlaid over the topology of the network. Such aggregation techniques have been previously proposed in the context of ad -hoc networks [ 22] and our mechanism follows a similar philosophy since Atlas nodes support ZigBee -based mesh networking The basic virtual sensor service object created in the service layer sends a tree -construction packet to the coordinator containing the i d entifier of the basic virtual sensor, its sampling rate in epochs and a list of its member sensor nodes repr esented by their network addresses. Figure 6 6. Logical versus network view of an aggregation tree Before transmitting the packet, t he co ordinator applies its routing algorithm on the list of network addresses to determine the non-duplicate set of child nodes that it needs to propagate the message to. A hash function which uses network addresses as keys can be used for this purpose. This ensures that each recipient node does not receive duplicate copies of the tree -construction packet as a consequence of multiple member nodes sharing the same routing path. Each router node in the network which receives the packet continues this process. Each router is also able to root service layer aggregation tree application layer service layer application layer aggregation tree root coordinator Logical View Network View member router member end device nonmember router cardinality > 1

PAGE 73

73 determine if it is the root of the virtual sensor aggregation tree or not by looking at the set of child nodes that it needs to propagate the packet to. The first router whose set of recipient child nodes has cardinality greater than 1 (Figure 6 6 Network View), becomes the root of the tree. This follows from the fact that the message only trave lled along a single routing path before reaching this node and now for the first time the route gets split into multiple sub -paths. In order to indicate to other nodes that a root has been selected, this node sets a special bit in the message packet before forwarding it further down the network. In case the coordinator is the root node it sets this special bit before forwarding the initial packet. Whenever a router or enddevice finds its address listed in a tree-construction packet, it removes its address from the list and also extracts the basic virtual sensor id along with its associated sampling rate. In this manner the aggregation tree is constructed on top of the ZigBee network. However, one must note that the logical view of the tree may be quite diff erent from its network view as we can see in the example (Figure 6 6 Logical View ). The network view of the tree might contain router nodes whose sensors are not members of the basic virtual sensor and whose only function is to act as a link with differen t segments of the aggregation tree. This is a consequence of self -organizing networks where nodes connect to their nearest routers based on proximity and signal strength. In case of proprietary ad hoc routing protocols it might be possible to tweak the algorithms to force nodes to form a network in a certain way. Semantic Routing Trees (SRTs) [34] used in conjunction with link based parent selection by TinyDB, provide such a functionality. However, this implies that the network has to be re -organized whenever the index attributes change. This is not only expensive but also of limited utility when multiple applications are running simultaneously. Moreover, such techniques require modification of networking layers in an application -specific manner which may not be advisable.

PAGE 74

74 Basic Virtual Sensor Aggregation Process The aggregation mechanism consists of propagation of partial records starting from the leaf nodes all the way to the root of the tree. Every time a participating end -device samples its sensor, it creates and sends a partial aggregation record to its parent. Each router on receiving a partial aggregation record from its child nodes merges them, updates it if necessary and forwards it to its parent. For example, if the aggregation formula being used was the arithmetic mean then a node would send a partial aggregation record , where Sum denotes the partial sum of readings collected and Count denotes the number of sensors sampled for the subtree rooted at this node. The root node on receiving the aggregation record calculates the final output, for example, for arithmetic mean it calculates the output as Sum divided by Count. Fault Tolerance of Basic Virtual Sensors Failure is an integral part of any sensor network especially for large scale deployments. Whenever a member node (and the sensor physically connected to it) fails, the basic virtual sensor loses a data source and this affects its data quality. A basic virtual sensor consisting of N Singleton Vir tual Sensors, outputs a single reading, which is an aggregate of the N readings obtained from them. In case of failure of one or more singleton sensors, it is likely that the basic virtual sensor can still continue to function albeit at the cost of providi ng data of a lower quality. We provide an approximation algorithm which enables basic virtual sensors to be fault tolerant up to a certain number of sensor failures, depending on user preferences. For derivation purposes we assume that the aggregation fun ction of the basic virtual sensor is the arithmetic mean We begin by definin g the following terms : N = Number of s ingl eton virtual s ensors which are initially members of the basic virtual s ensor S = {si}i = 1 to N : Set of the singleton virtual s ensors

PAGE 75

75 SF = Set of dea d singleton virtual s ensors NC = Number of singleton virtual s ensors currently alive DQT = Minimum Data Qual ity Threshold. Tstart = Time when the sensor network was started up Rs i = Reading from sensor s i S Nt = Number of r eadings sampled till time t. B = N x N matrix where Bij = number of instances when |Rs i Rs j the sensitivity of the sensor hardware and is typically chosen between 5 and 20). Bij gives the number of observed instances when sensors si and sj exhibited similar behavior. B is updated whenever the sensors are sampled. Ps i s j is the probability that sensors si and sj behaved in a similar manner and depends on the time t at which it is calculated. Ps i s j = t j iN B Ws i = Probabilistic weight associated with Rs i where, 1 ; if s i is alive Ws i = Ws jPs i s j ; if s i is dead and its readings are approximated using Rs j for some j As i = Set of sensors whose readings are being approximated using readings from sensor s i. Initially for all i = 1 to N, As i is set equal to VSreading denotes a reading from the basic virtual s ensor and is calculated by Equation 6 4. VSreading = S s s S s s si i i j iW R W (6 4) We also make the following assumption. All singleton virtual sensors which are members of a particular Basic Virtual Sensor represent physical sensors having the same characteristics such as sensitivity, er ror and r ange of output values. All s ingleton virtual s ensors are sampled together periodically in a synchronized fashion. All singleton virtual s ensors are alive at startup

PAGE 76

76 and hence, initially Ws i =1 for i = 1 to N. Suppose S is the set of singleton virtual sens ors that fail at time t then the approximation algorithm is given as follows: 1 Set NC = NC |S|. 2 For each sensor sx i ) Calculate Ps x equal to m ax {{0} U {Ps x s j for all sj S As x | Ps x s j > PT}} where PT is the minimum user defined threshold probability. PT ensures that the maximum observed probability of similar behavior is high enough not to be attributed to chance. Note that in case none of the probabilities exceed PT, then Ps x = 0. Suppose Ps x is equal to Ps x s y for some y. This implies that sensor sy x. ii) Set Ws x equal to Ws yPs x and As y equal to As y U {sx}. Hence, from now onwards, readings for sensor sx will be approximated using readings from sensor sy, which implies that Rs x is equal to Rs y. We assume that the equations for calculating Ws x and Rs x are stored in the framework. This ensures that: (a) Subsequent to each sampling, sensor readings for sx are generated using data obtained from sensor sy and; (b) Whenever Ws y chan ges (for example, when sensor sy fails), Ws x also gets updated. 3 Update Ws As x and set SF equal to SF U {sx}. 4 If Ws less than PT S, set Ws equal to 0 and set S equal to S {s}. This approximation algorithm ensures that during computation of VSreading the absence of sensors is compensated by readings obtained from other live sensors thereby increasing availability of sensor data while reducing the loss of data quality Each router in the aggregation tree runs the fault tolerance algorithm described above and maintai ns a history of similar behavio r am ong its children who are members of the basic virtual sensor. For example, if a router has 3 member sensors i, j and k as its children then it maintai ns three variables s(i, j), s(i,k) and s(j,k). If we use Euclidean distance as a measure of similar behavior then, during any sampling epoch, whenever the absolute difference between readingi and readingj is less than sor), s(i, j) is incremented by 1. Similarly s(i, k) and s(j, k) are also updated as required. The larger the value of these similarity statistics, the better is the quality of sensor -based approximation. If a child sensor node fails, its

PAGE 77

77 parent router aut omatically approximates its readings by choosing a live child node which was observed to be behaving most similar to the failed sensor. This choice is made based on the statistics collected till that point of time. Suppose in our example sensor i fails the n another live sensor (call it x) is used to approximate its readings, where x is equal to maxy s(i, y) with y is equal to {j, k}. The variables used for maintaining similarity statistics are periodically flushed to ensure that they do not get skewed by ou tdated information. This sensor based approximation mechanism not only has low processing overhead as compared to model -based approximation techniques but also compensates for the loss of sensors without requiring assistance from the service layer. Approxi mating readings of failed sensors may result in an improvement in the quality of virtual data as compared to having no compensation at all; however, it will still result in a drop in the overa ll data quality. Failure among member nodes of a basic virtual s ensor can be one of three types: 1 End device failure: The router which is the parent of the end-device detects this situation as a consequence of the ZigBee sync mechanism and executes the approximation algorithm described above. 2 Router failure: The parent of the router executes the approximation algorithm while its children on detecting the disconnection, try to re -connect to an alternate router available in the network utilizing the self -healing features of ZigBee. On re -establishing connection, a node se nds a message to the new router announcing its basic virtual sensor id and aggregation formula for merging partial records. This router then adds them as child nodes and takes on the responsibility of collecting similarity statistics and executing the faul t tolerance algorithm if any of them fails. 3 Root of aggregation tree f ails : If the root node of the aggregation tree fails, then a new root needs to be discovered. We propose a root discovery algorithm which guarantees that at most one root gets selected even if several nodes are in contention for that role. Each of the failed roots children initiates the discovery process by first transmitting a packet to its other siblings to determine if they are alive or not. After waiting for a certain period of time it transmits a root -search packet containing its virtual sensor id and number of live siblings. If required, it also contains the sensor ids and network addresses of root nodes of the derived virtual sensors which get inputs from this basic virtual senso r. This packet has its destination address is set equal to 0 which is the network address assigned to a coordinator by default. Each router forwards this message to its parent and caches it. However, if any router

PAGE 78

78 receives messages from all the siblings it considers itself as the new root of the aggregation tree. It suppresses further transmission of the root -search packet and sends a new -root message to each sibling declaring itself as the new root node. Since it is possible that due to time delays and synchronization errors, multiple nodes can declare themselves as the root of the same tree, each router which considers itself the root node also suppresses forwarding of any new root messages. This ensures that only the node with the shortest network path to all the siblings becomes the new root. In case the basic virtual sensor is a member of a derived virtual sensor, the new root node is also responsible for linking up with the root of the derived virtual sensor. Monitoring Data Quality of Basic Virtual Se nsors The fault tolerance mechanism approximates readings of failed sensors by using readings from other live sensors. This naturally affect s the quality of data being output by the virtual sensor. In order to monitor data quality, each reading originating from a basic virtual sensor is associated with a virtual sensor quality indicator (VSQIBVS) value. This value provides a measure of the virtual sensors data quality. VSQIBVS is a function of the number of live sensors contributing readings to the basic v irtual sensor and the number of failed sensors whose readings are being approximated using live sensors by the fault tolerance mechanisms. The formula for computing VSQIBVS is given by Equation 6 5. VSQIBVS = N W t g NFS s s C) ) ( ( s (6 5) where, g (x) = 1 ex/ a and VSQIBVS VSQIBVS has a maximum value of 1, if all member sensors are alive and 0 if all the sensors are dead. N is the number of live physical sensors which were originally members of the basic virtual sensor NC denotes the n umber of physical sensors which are currently alive. SF is the s et of failed sensorss is the time elapsed between the start up and failure of a failed sensor s. Ws is a weight associated with sensor s where, Ws Ws is computed by the fault tolerance mechanism, as the probability that the live sensor being used for approximation behaved similar to the failed sensor. The term g(x) is a time dependent weight function

PAGE 79

79 associated with Ws, where a is a constant greater than 0. The value of a determines how fast g(x) converges to 1 and depends on the sensors and their deployment environment. If Ws is calculated based on observations taken over a long time interval, then that value of Ws will be given more weight than sa y, a value of Ws which has been calculated based on obser vations taken over a short er period of time. The value of VSQIBVS is also computed using the partial aggregation method described before A partial aggregation record for VSQIBVS is given by . QP denotes VSQI information, for the sub tree (call it ST) rooted at this node. The QP term is given by T FS S s s LW t g N s) ( where NL denotes the number of live sensors in ST and g(t), ts and Ws are as defined above. The final VSQIBVS value is calculated at the root as QP/N, where N is as defined above. For practical purposes, the partial aggregation records for sensor readings and VSQIBVS are not transmitted separately but instead are merged into a single record: . Ope rations of a Derived Virtual Sensor The role of a derived virtual sensor is to apply a fusion function on inputs from multiple heterogeneous virtual sensors (whether basic or derived) to produce more sophisticated and abstract types of data. This enables i t to answer queries on data types which cannot be processed by the other virtual and physical sensors such as ambience and complex events like activities of daily living. Intuitively a derived virtual sensor should be residing solely inside the service lay er since it is not directly bound to inputs originating from any physical sensor. However, this will lead to scalability and latency issues due to the centralized service layer. In order to address these issues, we push down the operations of a derived vir tual sensor on to the distributed mesh network of nodes. When a user queries for a derived data type, the service layer looks up the definition of

PAGE 80

80 the associated virtual sensor from the knowledge base. It then composes together the different member virtual sensor services into a derived sensor service and initiates the process of creating an in network aggregation tree. First the derived virtual sensor service obtains the network addresses of the root nodes of each of its member sensors. This is possible due to the fact that the packets containing the output of any virtual sensor will have its source address equal to that of the root node of that sensors aggregation tree. Then a sensor -construction packet containing the list of these network addresses is di spatched to the ZigBee network via the coordinator. In a manner similar to the one described previously for basic virtual sensors each router applies its routing algorithm on the list of network addresses to determine the non-duplicate set of child nodes that it needs to propagate the message to. The first router whos set of recipient child nodes has cardinality greater than 1, becomes the root of the aggregation tree for the derived virtual sensor. This node becomes responsible for applying sensor fusion to the data originating from sensors which make up the derived virtual sensor. Derived Virtual Sensor Aggregation The aggregation of readings in a derived virtual sensor is pretty straightforward, with each member sensor sending its output to the root node which computes the final output and transmits it to the service layer via the coordinator. Unlike a basic virtual sensor which aggregates data from homogeneous sources, a derived virtual sensor has to operate on multiple heterogeneous inputs. Hence, the partial aggregation techniques used for basic virtual sensors are no t utilized for derived virtual sensors. Fault Tolerance of Derived Virtual Sensors By virtue of its definition, a derived virtual sensor cannot be too resilient when it comes to toleratin g failures. We take the position that if a member sensor fails completely then the derived virtual sensor should cease to operate due to lack of one of its inputs. Since we are not focusing

PAGE 81

8 1 on a particular application scenario for derived virtual sensors, we are forced to adopt such a pragmatic stance. It is entirely possible that in some application domains derived virtual sensors may be able to operate without the presence of one or more inputs. However, in the interests of providing a generic framework f or virtual sensors, we decided to adopt an all or -none approach in the context of derived virtual sensors. We hope to rectify this as part of future work. We discussed most of the failures when we covered fault tolerance for basic virtual sensors. Howeve r, if the root of a derived virtual sensors evaluation tree fails we need to take special measures. In this case, t he root of each member virtual sensor sends a root -search message addressed to the coordinator. Any router receiving messages from all the r oot nodes of the members considers itself the new root of the derived virtual sensor and sends a new root message to the sender nodes. It also su ppresses further transmission of root search and new -root messages VSQIDVS = BVS i x DVS j; VSQIDVS (6 6) 11 Figure 6 7. Relative error % vs. VSQI 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 0.00% 0.50% 1.00% 1.50% 2.00% 2.50% VSQI Relative Error %

PAGE 82

82 Monitoring Data Quality of Derived Virtual Sensors A derived virtual sensor may be composed of multiple heterogeneous b asic or derived virtual sensors; h ence, monitoring the performance of a derived virtual sensor requires a slightly different approach than that of a basic virtual sensor. The VSQIBVS formula gives the probability that the approximated output of a basic virtual sensor matches its expected output, if it was fully functional. Following the same trend, we define d VSQI associated with a derived virtual sensor as the product of VSQIs of its member basic and derived virtual sensors (Equation 6 6). Equation 6 6 reflects the fact that in case of a derived virtual sensor there is no concept of compensating for the loss of one of its member sensors. If one of the member basic or derived virtual sensors fails then its parent derived virtual sensors also fail. Experimental Pe rformance Analysis W e analyz e the performance of the virtual sensor mechanisms presented in this paper, using both simulated and real -world data. We conducted two sets of experiments, one for evaluating fault tolerance and data quality mo nitoring and the o ther for analyz ing the performance of our methods as compared to centralized techniques, in terms of latency and energy consumption of the sensor nodes. Fault Tolerance and Data Qualit y Monitoring The goal of the first set of experiments was to measure th e effectiveness of the fault tolerance algorithm and data quality monitoring. We were interested in evaluating whether the sensor -based approximation method was an effective way of compensating for failed sensors and if the virtual sensor quality indicator (VSQI) accurately reflected the effect of the approximation scheme on overall data quality. We used a real -world data set from Intel Research Lab, Berkeley for conducting this set of experiments. This data set consists of temperature, humidity and light level readings collected

PAGE 83

83 from 54 nodes deployed at the Intel Research Lab in Berkeley, California, over a period of one month. Out of the 54 sets of data, only 40 were short listed as the remaining ones had incomplete sets of readings. Each sensor log was also sorted to ensure that all the readings were ascending order of their time stamps. Figure 6 8. Number of sensor failures vs. VSQI and relative error Based on sensor location information provided in the documentation, we created basic virtual sensors composed of 8 physical temperature sensors which used arithmetic mean for aggregation. S ensor failures are simulated at periodic in tervals and the VSQI value and percentage relative error logged for each epoch of sampling. The relative error value was computed as the deviation of the actual output of the virtual sensor (based on aggregation of live and approximated sensor readings) from the expected output (based on aggregation of readings from all live sensors ) Since at the time of logging of the original data sets, all sensors were functioning properly, the induced failure allows us to evaluate how well the virtual sensors compensate for it. 0.00% 0.50% 1.00% 1.50% 2.00% 2.50% 0 0.2 0.4 0.6 0.8 1 0 1 2 3 4 5 6 7 8 Relative Error % Virtual Sensor Quality Indicator (VSQI) Number of Failed Sensors VSQI Relative Error %

PAGE 84

84 To determine the effectiveness of data quality monitoring we explore d the correlation between VSQI and the rela tive error. As described earlier, VSQI is a measure of the quality of a virtual sensor; in terms of the probability that compensated values accurately reflect the readings should all sensors remain functional. Relative error is a different measurement of t he quality of virtual sensor, which focuses on the deviation of the output from the expected value because of the internal compensation within a virtual sensor. Relative error can be heavily dependent on the specific datasets and the configuration of the v irtual sensors. However, basic virtual sensors components deployed in realistic settings are usually spatially correlated, which leads to smaller and bounded relative errors. W e observe that in general, the VSQI and the relative error are inversely proport ional to each other (Figure 6 7 ). As the relative error increases (usually due to more failures and heavier compensation), the VSQI decreases accordingly. From the plot, one also observes that when the relative error is approximately 2.3%, its corresponding VSQI value seems to be quite high. Upon inspection of the simulation logs, it was determined that this anomaly occurred due to temporary malfunctioning of one of the compensating sensors, which resulted in the output of spurious readings for a short dura tion. Next, we studied the effect of sensor failure on VSQI (Figure 6 8). The simulation showed that as more singleton virtual sensors fail the VSQI decreases. However, VSQI only decreases gently as sensors fail, which demonstrates the fault -resiliency of the virtual sensor mechanism. Depending on the specific functioning sensor chosen to compensate for a failed one, the relative error may spike irregularly, but in general, the relative error increases as more sensors fail. One can also observe that, there is a temporary increase in the relative error when the first sensor fails. Upon further inspection of the logs, it was determined that the compensating sensor experienced some temporary malfunction and as a result output spurious readings for a short

PAGE 85

85 dura tion, before resuming normal operation. As mentioned earlier, even under extreme wide spread failures (87.5%), the rel ative error always remains bounded (less than 2.5%), which demonstrates the effectiveness of fault tolerance in virtual sensors. Figure 6 9. Effect of sensor failure pattern on VSQI The results of the simulation also show the effect of failure patterns on VSQI (Figure 6 9) We simulated random singleton virtual sensor failures beginning at time 0, 180, 720, 2470 and 7920, with the stipulat ion that all the sensors will fail within 80 epochs. The purpose wa s to explore whether allowing the sensors to accumulate a long correlation history improves VSQI when sensors start to fail. When the choice of the converging time constant a is kept low (we chose a=4), the weight of approximation plays a more dominating role. We observed that the later the sensor failures are introduced, the sharper the drop in VSQI when sensors do fail. This phenomenon appears to be contradictory to our intuition at firs t, but upon further analysis, it is found that since the sensors that are used to compensate are not always placed at the exact same locality, the probability that two sensors exhibiting comparatively similar behavior always generate readings of equivalent magnitude, decreases as the history accumulates over time. 0 0.2 0.4 0.6 0.8 1 0 1 2 3 4 5 6 7 8 VSQI Number of Failed Sensors D = 260 / F = [181, 260] D = 800 / F = [721, 800] D = 2550 / F = [2471, 2550] D = 8000 / F = [7921, 8000]

PAGE 86

86 Latency of Data Arrival and Energy Consumption The second set of experiments was conducted to evaluate system p erformance in terms of latency in arrival of virtual sensor output and energy consumption of nodes per epoch of sampling. We compared the performance of our distributed virtual sensor mechanisms versus the mechanism where virtual sensors are executed exclusively inside the service layer o f the SOSN and perform centralized aggregation on data streams. The Global Sensor Network (GSN) [ 1] is example of a virtual sensor framework which utilizes this form of centralized streaming approach for implementing virtual sensors. Figure 6 10. Compar ing latency in arrival of final output For distributed tree -based aggregation, we used a simulator to randomly generate aggregation trees overlaid on top of a mesh network. The simulator takes as input the number of sensor nodes which are members of a virt ua l sensor and randomly generates an aggregation tree of nodes for each iteration To simulate the overlaying of the tree on top of a mesh network, it randomly assigns to each node one of the following roles: (1) enddevice which is a member of 0 10 20 30 40 50 60 70 80 90 10 60 110 160 210 260 310 360 410 460 Latency in Arrival of Final Output (seconds) Number of Nodes CentLat DistLat

PAGE 87

87 the virtual sensor; (2) router which is a member of the virtual sensor; and (3) router which is not a member of the virtual sensor (that is, it only links different segments of the aggregation tree). To calculate latency of virtual sensor outp ut, we assume a node takes 0.003 seconds to receive or transmit a packet (based on a packet size of 100 bytes and ZigBee transmission bit rate of 250 Kbps). For computing energy consumption of nodes we obtained power consumption data from Atmel for their Zlink RCB nodes. We use a power c onsumption cost model where each member end device and routers cost is calculated as the sum of processing, sensor sampling and transmission costs. Additionally each member router also incurs the cost of receiving messages. On the other hand, each non -me mber routers cost is simply the cost of receiving and transmitting messages. W e observe that in case of centralized aggregation (denoted by CentLat), the latency increases significantly as the number of sensors increases (Figure 6 10) This is due to t he fact the each sensor reading has to traverse multiple hops all the way up the tree in order to reach the service layer and the aggregation operation cannot take place till readings from all the member sensors have streamed up. One would assume that sinc e messages flow up the tree in parallel, latency of the streaming mechanism will be low. However, this assumption turns out to be invalid in a sensor network which uses multi -hop networking This is due to the fact that in the nave streaming approach, eac h sensor reading packet travels all the way up to the base station and there is no merging of packets at any routing node. Hence, at each hop a packet has to wait till the radio of that router node is free to transmit before it can propagate up the tree. T his wait time can be long if there are a number of child nodes which are waiting to transmit their packet via this router node since the radio of the node can only transmit one packet at a time Therefore with each hop the latency of the packet increases and since every sensor reading packet has to propagate all the way up the tree, an increase in the number of

PAGE 88

88 sensors leads to a drastic increase in latency of response for the virtual sensor The distributed aggregation tree mechanism (denoted by DistLat ) on the other hand exhibits extremely low latencies of approximately 1 second or less, even when the number of sensors is large. This can be explained by the fact that a reading from each sensor node only travels only 1 hop to its parent aggregation node where it is merged with other partial aggregation records. Hence, readings from each sensor do not need to be individually transmitted over multiple hops all the way to the service layer thereby, significantly reducing latency of virtual sensor output. Moreover, due to its lower utilization of network resources, it reduces the processing and networking costs i n intermediate routers, which translates into significant reduction in overall energy consumption. Figure 6 11. Energy consumption of virtual sens or per epoch With regards to total energy consumption we observe that for the centralized method (denoted by CentEnergy), the total energy consumed by nodes for each epoch of sampling dramatically increases as the number of member sensors increases wher eas for the distributed mechanism (denoted by DistEnergy) there is no significant increase in energy consumption 0 5 10 15 20 25 30 10 60 110 160 210 260 310 360 410 460 Energy Consumed per Epoch of Sampling (Joules) Number of Nodes CentEnergy DistEnergy

PAGE 89

89 (Figure 6 11) In fact the total energy requirements per epoch for the distributed mechanism is anywhere from 9098% less than the centralize d method. This implies that the distributed mechanism is not only more scalable but is also more energy efficient as compared to the centralized method.

PAGE 90

90 CHAPTER 7 PHENOMENA DETECTION AND TRACKING Contemporary wireless sensor network (WSN) research done in the area of detection and tracking has primarily concentrated on and proposed distributed algorithms for, observing motion of objects whose shape and size are invariant [ 15, 25]. In contrast, phenomena clouds such as oil spills and gas clouds not only exhibit motion but are also characterized by non deterministic variation of their shape and size, over time. However, the utility of phenomena cloud detection and tracking is not restricted only to application dom ains involving gas clouds or oil spills. In fact, they can also be utilized in situations where the quality of data originating from individual sensors cannot be trusted in isolation. In such cases, the raw sensor data originating from the system is typica lly extremely noisy which makes it very difficult to distinguish actual events from random stimuli. Hence, a quorum of multiple sensors which are located in close proximity to each other is required to reduce the probability of false positives. Through our collective research and systems experience over the years in building Smart Spaces at University of Floridas Mobile and Pervasive Computing Laboratory, we have found great utility in applying the phenomena cloud concept for efficiently and accurately m onitoring various events in the space. Stream -base d mechanisms such as Nile PDT [ 2, 3], which have been proposed for detection and tracking of phenomena clouds do not take into account the cost of acquiring and transmitting sensor readings and typically r equire participation from all sensors in the network. Unfortunately, sensor sampling costs and networking and processing overheads can have a critical effect on the practical viability of the entire smart space and hence there is a need for distributed in -network mechanisms, where execution of the detection and tracking process is localized to the immediate neighborhood of a phenomenon at any given time. Furthermore, due

PAGE 91

91 to primitive processing capabilities of wireless sensor nodes such mechanisms should a void relying on construction of mathematically complex cloud models. Hence, there is a need for localized, in -network algorithms which provide energy-efficient, real time phenomena detection and tracking capabilities without requiring complex cloud models. The remai nder of this chapter is organized as follows. First we define phenomena clouds and their characteristics. We analyze their structure when overlaid on a sensor space to generate a set of parameters to comprehensively describe them without requirin g a formal mathematical model Next, we describe the critical challenges faced during the detection and tracking process. Then we describe energy-efficient, localized, in network algorithms for real -time detection and tracking of phenomena clouds, which do not require customization of the network routing layers. The proposed algorithms work in an autonomous manner without requiring intervention from the centralized query processor residing in the base station and hence, are suitable for disconnected mode of operation, when continuous communication with the base station cannot be maintained. Next, we describe an interesting and practical real -world application of our phenomena detection and tracking algorithm in a smart space. The application described in thi s chapter is presently running in the Gator Tech Smart House, a real -world smart space to solve critical challenges in detecting certain activities Finally we evaluate and analyze the performance of our approach through real life experiments and simulati ons. We al so provide a comparison of the resource usage and energy consumption characteristics of our approach against contemporary real -world stream -based detection and tracking techniques such as Nile -PDT. Phenomena C louds A p henomenon cloud can be defined as a manifestation of a number of simultaneous event s reaching critical mass and spanning a contiguous space. The shape, size and direction of movement of such phenome na cloud s either cannot be modeled accurately or have models which

PAGE 92

92 are usually too complex for real time computing by sensor networks, which largely consist of low -end nodes with limited processing capabilities. Examples of such phenomena include gas clouds, flooding, oil sp ills, fires or even movement of tourists in a museum. A ph enomenon cloud can exhibit non-deterministic behavior over time making it difficult to anticipate its path and motion. Major Challenges M ajor challenges faced during detection and tracking of phenomena clouds are as follows: 1 Initial detection of phenomeno n. Initial occurrences of the phenomenon might be scattered throughout the space. Hence, detection needs to be attempted at multiple locations. 2 Avoiding false positives. The probability of a single sensor outputting in accurate readings at a specific point in time is not low. Hence, it is quite possible for a sensor to temporarily malfunction or be subject to environmental conditions which might cause it to output values which incorrectly indicate the occurrence of a phenomenon. 3 Tracking a phenomenon in re al time. A phenomenon can suddenly grow or shrink in size and also move in multiple direction s simultaneously ; hence, tracking it in real time can become a massiv ely complex task. To enable cost -effective real time tracking, the rate of status update s from the sensor network to the user needs to be kept at a minimum to reduce network cost and processing overhead 4 Operating under harsh conditions resulting in disconnected operation: Hostile phenomena like fires can lead to disruption of communications betwee n the sensor network and the base station hence, the detection and tracking process should be able to operate in an autonomous manner without requiring remote supervision. Representation We represent a phenomenon cloud as a 5 tuple, P = . T he lower and upper bounds of the range of sensor values which constitute a phenomenon are denoted by a and b respectively. For example, a hydrogen gas cloud can have a = 20% volume and b = 100 % volume. pT is the threshold probability, m is the obs ervation count and n is the minimum quorum. We refer to a sensors reading lying in the range [a, b] with probability greater than pT during the last m observations (that is, in a sliding window of size m) as satisfying the Probability -Condition. We defi ne a sensors neighborhood as the set of sensors immediately

PAGE 93

93 surrounding that sensor. We further state that a sensor is said to participate in a phenomenon cloud (or satisfy the Phenomenon-Condition ) given by P = , if it satisfies the Probability -Condition and at least n sensors in its neighborhood also satisfy the Probability Condition. This criterion ensures that a sensor must have a sufficient number of neighboring sensors in agreement with it before it can claim the existence of a phenomenon cloud, thereby reducing the occurrence of false positives. We define Phenomenon-Set to be the set of sensors satisfying the Phenomenon-Condition Figure 7 1. Dissection of a phenomenon cloud We consider a phenomenon cloud to be composed of m ultiple regions ( Figure 7 1 ). Th e innermost region or the Core r egion of the cloud is where the phenomenon is observed to be the strongest. The sensors lying in the Core region satisfy the requirements of the P henomenonCondition and hence, are members of the Phenomenon-Set The Outer r egion denotes the fringes where uncertainty regarding the occurrence of the phenome non is highest. The Middle r egion of the cloud is where the probability that the phenomenon is occurring is somewhere between that of the Core and Outer r egions. The Middle and Outer regions are essentially areas where the occurrence of the phenomena has not yet been fully verified. D etection and Tracking In this section we provide a detailed description of the phenomenon cloud detection and tracking process. We begin by classifying sensors into different categories according to the roles Core Region Middle Region Outer Region

PAGE 94

94 they play in the detection and tracking process. Next, we describe the various resp onsibilities of a sensor node with respect to each of its neighbors, based on the categories they currently fall in. Then, we list a set of rules which govern the transition of sensors from one category to another and form a core part of our detection and tracking strategy. The rest of this section covers the different stages of the detection and tracking process in chronological order (Figure 7 3) and also describes mechanisms for handling node failures. Finally, we discuss how applications can utilize the real time tracking data output by the sensor network. Figure 7 2 Classification of participating sensors Classification of Sensors Figure 7 2 superimposes the phenomenon cloud in Figure 7 1 over a group of sensors. Sensors are classified according to the region where they are located which determines their role in the detection and tracking of a cloud. The different c ategories are Candidate s ensor : A candidate sensor is one which is currently not part o f any phenomenon cloud but is actively monitoring its readings to determine if it will become part of a cloud or not. Candidate sensors make up the M iddle region of a phenomenon cloud. A sensor becomes a candidate if it has been selected as such by the que ry processor as part of the initial detection phase or it has transitioned from the potential candidate stage. The role of a candidate sensor is to receive notifications from its neighboring sensors whenever they satisfy the Probability -Condition. Based on the number of such neighbors it decides whether it is eligible to becom e a tracking sensor or not. Here we also note that in case a candidate sensor has one or more neighbors who are candidate sensors themselves, then it is responsible for notifying them whenever its own readings satisfy the Probability -Condition. Idle Sensor Potential Candidate Candidate Tracking

PAGE 95

95 Potential candidate s ensor : All sensors which are immediate neighbors of candidate sensors but are not candidates themselves are called potential candidate sensors. These sensors also monitor the ir readings to enable a neighboring candidate sensor to check the validity of its observations. The role of a potential candidate sensor is to notify its neighboring candidate sensors whenever its readings satisfy the Probability -Condition. Potential candidate sensors form the fringes of detection and make up the O uter region of the phenomenon cloud. Essentially the set of potential candidate sensors form s a phenomenon front which grows and shrinks dynamically. Tracking s ensor : A tracking s ensor is a sensor which has already detected a phenomenon event and is now actively engaged in tracking it. A candidate sensor becomes a tracking sensor after it satisfies the Phenomenon-Condition. Tracking sensors make up the C ore region. The Phenomenon-S et is the collection of all tracking sensors hence, each cloud that a user observes through the sensor network, consists of sub-sets of tracking sensors from the Phenomenon-Set Idle s ensor : All sensors which do not belong to any of the above three categor ies are called idle sensors. These sensors are not engaged in phenomenon detection or tracking and do not perform any monitoring whatsoever. Typically most sensors in the space will fall in this category since only selected clusters of sensors will be acti vely engaged in the dete ction and tracking of phenomena clouds at any given time. This ensures that the detection and tracking process is executed in a localized manner with minimal expenditure of energy. Keeping Tabs on the Neighborhood Each sensor node k eeps track of which category each of its neighboring sensors falls in. This is done in a peer to -peer fashion, where a sensor transitioning from one category to another, notifies the other sensors in its neighborhood via a 1 -hop broadcast without involvem ent of the centralized query processor. The Atlas nodes support ZigBee communication which natively supports 1 -hop broadcasting. The categories a sensor and its neighbor fall into determine their mutual responsibilities towards each other. Table 7 1 lists the actions a sensor node is required to perform with respect to a neighborhood sensor, based on which category each of them belongs to. For example, a candidate sensor node A has a neighbor B which is also a candidate sensor and another neighbor C which i s a tracking sensor. In that case, A will alert B whenever its readings satisfy the Probability -Condition but A will only alert C whenever its readings do not satisfy the Probability -Condition. Hence, a single sensor node can play different roles with resp ect to different neighbors depending on which categories they fall into. The cells marked

PAGE 96

96 Not Applicable imply that such combinations are not possible according to transit ion rules given in the next sub section. Table 7 1 Actions t aken by a sensor n ode with respect to its n eighbors which are not i dle Sensor node Potential c andidate Candidate Tracking Idle None N ot a pplicable N ot a pplicable Potential c andidate None Send alert whenever readings satisfy Probability Condition Not a pplicable Candidate Receive alerts from neighbor whenever its readings satisfy Probability Condition Send alerts/ receive alerts whenever their readings satisfy Probability Condition Send alert whenever readings do not satisfy Probability Condition Tracking Not a pplicable Send alert whenever readings satisfy Probability Condition Send alerts/ receive alerts whenever their readings do not satisfy Probability Condition Transition Rules A set of rules govern the transition of a sensor from one category to another. These rules are executed in network and control the entire detection and tracking process. R1: Candidate If a sensor satisfies the PhenomenonCondition then it transitions into the tracking category. Once a sensor is in the tracking category, it beco mes a member of the PhenomenonSet R2: Potential Candidate A potential candidate sensor will transition to a candidate sensor if any of its neighbors transitions into a tracking sensor. This rule corresponds to the fact that whenever a phenomenon cloud moves or expands and a new sensor becomes part of its core region, the phenomenon front also moves or expands. R3: Idle An idle sensor transitions into a potential candidate if any of its neighbors becomes a candidate sen sor. R4: Tracking A tracking sensor will transition down to the candidate category if it is unable to satisfy the Phenomenon-Condition In such a case, the sensor will cease to be a member of the Phenomenon-Set

PAGE 97

97 R5: Candidate Candidate A candidate sensor will transition to a potential candidate sensor if none of its neighbors are tracking sensors. R6: Potential Candidate A potential candidate transitions into an idle sensor if all its neighbors are either potential can didates or idle, that is none of its neighbors are in the candidate category. Initial Selection of Candidate Sensors The main goal of this stage is to detect initial occurrences of phenomena clouds. A phenomenon cloud can manifest itself in multiple locati ons simultaneously hence, monitoring one particular location is not adequate. However, in the interest of conserving network resources and power for the entire sensor grid, we cannot require each and every node to monitor its readings. A compromise between the two approaches can be followed where specific sensors are directly chosen to be candidates by the centralized query processor. The criterion for such a selection can be based on the location of nodes or past history of their readings. For example, if we are planning to detect gas leaks in a pipeline, it might be useful to choose the sensors located at the valves and joints to be the initial set of candidate sensors, since the probability of a leak getting started at those locations is higher. We make u se of this criterion in the Smart Floor ap plication described later where sensors located near doorways are selected as initial candidates so that whenever a person enters the room, the system is immediately able to pick up their presence and commence the detection and tracking process to monitor their movement. Another criterion can be the offline use of an available mathematical model of the phenomenon cloud to determine locations where the probability of occurrence is the highest. In case such a criteri on is hard to formulate, alternatively the system can randomly select sensors as candidates such that they are uni formly distributed over the sensor space These sensors and their respective neighborhoods can be viewed as autonomous clusters of early warni ng systems for detecting the sudden manifestation of possibly multiple phenomena clouds.

PAGE 98

98 Figure 7 -3. Detection and tracking of a phenomenon cloud Since sensor deployment patterns tend to be highly application and phenomena -specific, we do not go into details of their deployment. For purposes of discussion for the rest of this chapter we assume that sensor nodes are deployed in such a manner that each sensor has a sufficient number of neighbors to potentially avoid false positives. Monitoring for Initial Occurrences The query processor pushes the ph enomenon cloud parameters on to each of the selected initial candidate sensor nodes in the network At the beginning of every epoch, each potential candidate node monitors its readings and sends a 1 hop broadca st message every time the Probability -Condition is satisfied In order to enable sensor nodes to receive alert broadcasts from multiple neighbors simultaneously during the same epoch, a slotted approach is used to

PAGE 99

99 ensure collision avoidance. Each epoch is sub -divided into multiple sub -epochs and each node only broadcasts alerts during its assigned sub -epoch. The candidate sensor node aggregates alerts received via broadcasts from its neighbors and determines if it satisfies the PhenomenonCondition. A candidate sensor satisfies the Phenomenon-Condition if it s readings satisfy the Probability -Condition and it also receives broadcast alerts from at least n neighbors which have also satisfied the Probability -Condition in the same epoch. Notification of Initial Occurrence If a candidate node has satisfied the Phenomenon-Condition it notifies the query processor residing in the base station that it has detected presence of a phenomenon cloud. The query processor adds the candidate node to the Phenomenon -Set and the candidate sensor transitions to a tracking sensor role using rule R1 Figure 7 4. Ratio of total active sensors to cloud size in a rectangular sensor grid Growth of Phenomenon Cloud The candidate sensor also broadcasts an alert about its change of state. As soon as the neighborhood sensors receive the alert, they transition to the candidate category using rule R2. 0 5 10 15 20 25 30 1 25 81 169 289 441 625 841 1089 1369 1681 2025 2401 2809 3249 3721 Active Sensors : Phenomena Cloud Size Phenomena Cloud Size (Number of Tracking Sensors)Number of Active Sensors : Phenomenon Cloud Size

PAGE 100

100 Each of these sensors in turn, broadcasts its change of state and this causes all their respective neighbors which had been idle till now, to transition in to potential candidate category (rule R3). In this manner, the detection mechanism gets distributed and pro pagated in network, without involvement of the centralized query processor, as the phenomenon cloud grows with time. Each sensor node keeps track of its neighborhood via the broadcast alerts that it receives and determines the actions to be undertaken with respect to a specific neighbor based on which category each neig hbor falls in (Table 7 1 ). We observe that in our distributed innetwork approach, the number of active sensors required at any given time is only slightly more than the number of sensors actually participating in the phenomenon cloud and the ratio of act ive sensors versus phenomena cloud size decreases with increase in cloud size (Figure 7 4) This is due to the fact that the detection and tracking process is executed innetwork in a localized manner to ensure maximum efficiency. Only those sensor nodes w hich are in the immediate vicinity of a phenomenon cloud or are participating in the cloud are actively involved in the detection and tracking process. The propagation of this process in the network is governed solely by the behavior of the phenomenon cloud and handled by the distributed nodes in a cooperative manner using the rules (R1 R6) specified earlier without requiring assistance from the centralized query processor. Shrinking of Phenomenon Cloud The phenomenon cloud is said to shrink when sensors falling in the tracking category determine that they can no longer satisfy the Phenomenon-Condition A fter a sensor transitions into tracking, its neighbors will only send it alerts if their readings fail to satisfy the Probability Condition (Table 7 1) A tracking sensor is no longer participating in the observation of the phenomenon cloud if it determines that less than n of its neighbors currently satisfy the Probability -Condition. In such a case, t he tracking sensor node notifies the query processor

PAGE 101

101 which removes t he tracking sensor from the Phenomenon-Set thereby signifying that the phenomenon cloud has shrunk. The affected sensor also transitions into the candidate category using rule R4 and notifie s the other sensors in its neighborhood of this role change via broadcast. The neighbors which are candidate sensors and meet the conditions of rule R5 then transition into potential candidate sensors. Furthermore, the neighbors of these sensors which mee t the conditions of rule R6 transition into idle sensors. We make a note that if all the phenomena clouds disappear completely then after all the transitions are applied as per the given rules the sensor space will revert back to the initial state where only the initial set of candidate sensors will remain active along with their res pective neighborhoods consisting of potential candidate sensors. Handling Failures The failure of sensors is an inevitable part of deploying sensor networks. In the context of phenomenon detection and tracking, this is especially true for phenomenon clouds which occur in hostile conditions such as wild fires or occur in very large scale sensor deployments In case ZigBee communication is used a sensor node can detect failure a mong its neighbors while notifying them to transition into a different category. ZigBee networks implement ACKs on th e MAC level hence, if any of the neighbors are dead the node will be able to detect that its messages could not be delivered to a particul ar neighbor successfully. Whenever a node detects that one of its neighbors is dead, it updates its total neighbor count (call it N). We note that every node at some point in time, as per the terms of the Phenomenon-Condition and rules R1 through R6, will transition to a different category either due to its ability or inability to satisfy the Phenomenon-Condition or based on the transition of i ts neighbors. Hence, whenever the node to inform its neighbors of this transition, it will be able to detect which of them are dead. A more robust but slightly more expensive way to ensure which nodes in a neighborhood are alive is to

PAGE 102

102 require each node to broadcast alive mess ages at regular intervals if it has not broadcast any alerts for a certain period of time. In case a candidate sensor fails all of its neighbors transition into candidate roles on detecting the failure. Also, if the value of a nodes neighbor count N falls below a certain minimum value N ( -hop broadcast so t h at potential candidates among them can transition into candidate category and their idle neighbors in turn can transition into potential candidate category. This strategy ensure s that the detection of the phenomenon cloud does not get stalled due to failu re of any of the candidate sensors or lack of enough sensors in a candidate nodes neighborhood. Real -Time Monitoring by Applications The centralized query processor maintains the Phenomenon-Set which as described before, is the collection of all sensor n odes which fall in to the tracking category at any given time. T he query processor is able to track phenomenon clouds in real time, with minimum processing and networking overhead of receiving updates from the sensor nodes. The PhenomenonSet is only update d whenever a sensor transitions to or from the tracking category. Hence, the query processor only requires minimal updates to continuously track phenomena clouds. Since the location of each sensor node is known beforehand, an application such as a phenome non cloud visualization tool with a graphical user interface (GUI) can easily reconstruct a view of the various phenomenon clouds in real time using information from the PhenomenonSet in conjunction with sensor location information. By looking at which se nsors enter or leave the Phenomenon-Set the motion of multiple phenomenon clouds can also be tracked over time and more sophisticated analysis and prediction performed at a centralized level. This is extremely use ful for applications such as those which can determine safe passages for rescue workers through multiple occurrences of phenomenon clouds such as gas leaks and wild fires

PAGE 103

103 [4]. As future work, we are concentrating on better organization of the PhenomenonSet to enable efficient updates and look ups of specific clouds and their characteristics. The cardinality of the Phenomenon-Set in combination with sensor location can give applications information about the size of various phenomena clouds. The size of a phenomenon cloud can have different impact s depending on the application context. For example, if an application is concerned with detection of phosphate dust clouds, a PhenomenonSet of low cardinality may not have much significance. However, if an application is tasked with the detection of hydr ogen cyanide (HCN) leakage, then the detection of even a small cloud indicates serious consequences. Figure 7 5. Gator Tech Smart House A P ractical Application of Phenomena Detection and Tracking We describe a real -world application of phenomena detecti on and tracking which is radically different from the gas cloud and oil slick simulations that are usually presented. The application that we desc ribe here involves the Smart Floor [ 6, 17] in the Gator Tech Smart House [13] (Figure 7 5) The Smart Floor deployed in 2005, consists of a grid of force sensors deployed under the raised floor tiles of the house. Each tile has a single sensor c onnected to an Atlas Platform node [18] placed below its center ( Figure 7 6 ), which allows a step anywhere on the tile to be detected. The Smart Floor covers the entire residential area of the 2500 sq. ft. house and

PAGE 104

104 allows it to monitor its residents movement and location without encumbering them with tags or other tracking devices. Figure 7 6 Smart Floor tile with for ce sensor and Atlas Platform node While designing the Smart Floor application, the nave expectation was that when a person steps on a tile, only the sensor underneath that tile outputs a reading of significant magnitude. Unfortunately, based on our experi ence over the years we found that this was clearly not the case. Due to various reasons including seemingly random vibrations, individual sensors sometimes output large readings even when nobody is stepping on them. This results in a very noisy sensory env ironment where one cannot distinguish between a genuine step and random spikes by relying on individual sensors alone. We also observed that when a person steps on a tile not only does this result in that tiles sensor registering a strong reading but som e of its neighboring tiles also output significantly large readings. Hence, the stepping action of a foot on a floor tile causes a ripple effect in the immediate ne ighborhood of the tile (Figure 7 7) as seen in an actual screenshot of this

PAGE 105

105 phenomenon occu rring in the Smart Floor where red dots indicate tiles registering readings of higher magnitude and green indicates tiles with lower yet significant magnitude. We used this observation to describe walking as a phenomenon by defining a step in terms of a phenomenon cloud ( Figure 7 8 ), in order to reduce the number of false positives and provide accurate location information about the homes resident. Moreover, since our approach to phenomenon detection and tracking does not rely on mathematical modeling to track the direction of movement of a phenomenon hence, this makes it extremely suitable for observing phenomena such as walking where it is difficult to accurately model the path that a person will follow at any given time. Figure 7 7. Ripple effect of a foot step on the Smart Floor A step can be described as a phenomenon cloud S = , where a and b denote the lower and upper bounds of a force sensor reading indicating that a foot has stepped on a tile or in its immediate vicinity. This value depends on the particular sensor being used. For example, based on empirical study, we found that for the Interlink force sensors used in the Smart Floor (having an output range of [0, 1023]), a is equal to 150 and b is equal to 600 for

PAGE 106

106 an indivi dual weighing between 110 to 240 pounds. The optimal values of the other parameters were det ermined via experimentation and are described in the following section. Figure 7 8 Walking motion as a p henomenon W alking motion is characterized by stride length (the distance between two footfalls of the same foot), gait velocity (speed at which a person walks) and cadence (the number of steps a person takes per minute) T he observation of these parameters is of paramount importance for monitoring patients suffering from obesity and Parkinsons disease. For example, people with Parkinsons disease have significantly shorter stride length and slower gait velocity as compared to healthy individuals. Similarly, people suffering from morbid obesity typically ha ve comparatively low gait velocity and cadence. Phenomena detection and tracking can be used to monitor all three walking parameters in the privacy of ones home without encumbering the resident in any way. Stride length can be instantaneously calculated a s twice the distance between two sensors which send consecutive update messages. Gait velocity can be calculated over an observation perio d P whose length depends on how long the resident walks in a straight path without turning. If tiles i and j are the first and last tiles stepped on during P, then gait direction of walk Idle Sensor Potential Candidate Candidate Tracking

PAGE 107

107 velocity can be calculated as the difference between distancei and distancej divided by the magnitude of P. Finally, cadence can be calculated as one -half of the number of update messages received by t he monitoring application in one minute. Figure 7 9. Effect of varying n with pT=0.4 and m=150 Experimental Analysis and Performance Evaluation W e evaluate d various aspects of the distributed phenomenon detection and tracking approach using both real -w orld and simulation experiments. The fi rst set of experiments evaluated the effectiveness of our detection strategy in a real worl d sensor deployment and analyzed the effect of the varying phenomena definition parameters pT, m and n Th e second set of experiments used simulation to e valuate the resource usage and energy consumpt ion of our approach and compare it with that of stream -based detection and tracking. We relied on simulation in this case because we wanted to measure resource usage and power consumption in large sensor networks of varying sizes. Hence, it was not practically feasible for us to physically deploy sensors in such large numbers for the purpose of experimentation. 0 100 200 300 400 500 600 700 800 900 1000 0 1 2 3 4 5 6 7 8 False Positives Missed Detected n

PAGE 108

108 Figure 7 10. Determining the optimal value of n Experiment I: E ffectiveness of Detection Strategy In this first set of experiments we study the effectiveness of our phenomenon detection and tracking mechanism in a real -world sensor deployment inside the Gator Tech Smart House. Experimental s etup We chose to evaluate effectiveness by performing ex periments using the Smart Floor in the Gator Tech Smart House, where human footsteps are represented as phenomena clouds and phenomenon detection and tracking is used to monitor the location of the resident in the house. We o bserved the effect of phenomenon definition parameters pT, m and n on t he detection efficiency of our approach We varied the values of each of these parameters and studied their effect by logging the number of false positives, false negatives/misse s and correct detections of a human step. In order to aid our evaluation, we restricted movement to a 100 sq. ft. area in the living room of the smart house and had test subjects walk along a clearly marked path on the floor. Th is allowed us to log the ste ps that a person was taking and collect statistics on correct detections and detection errors. 600 400 200 0 200 400 600 0 1 2 3 4 5 6 7 8 S (Score Function) n

PAGE 109

109 Figure 7 11. Ef fect of varying pT with n=3 and m=150 Results and analysis The experimental results are presented in three graphs ( Figures 7 9, 7 11 and 7 13). First we consider the effect of varying parameter n (Figure 7 9) which determines the minimum quorum of neighboring sensors required to conclusively determine the occurrence of a phenomenon. Since the Smart Floor is deployed as a rectangular grid of se nsors, the value of n varies from 0 to 8. We observe that for n=0 (which corresponds to the nave case) the number of false positives is extremely high since the system is entirely relying on outputs from single sensors to determine the occurrence of a f ootstep event. Hence, even though there are no misses/false negatives and all the actual footsteps are detected, their occurrence is lost in the noise of having an extremely large number of false alerts. As we increase the value of n, the number of false positives comes down sharply since now multiple neighboring sensors need to agree on the occurrence of a phenomenon. We also notice that as n increases, the number of misses also increase, thereby reducing the number of correct detections. This is due t o the fact 0 500 1000 1500 2000 2500 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 False Positives Missed Detected PT

PAGE 110

110 that walking is essentially a transient event where a footstep has to be detected by a specific sensor within a very small time window. Hence, even though we postulated that the action of stepping on a tile causes a ripple effect among neighbori ng tiles, it is not necessary that the number of neighbors experiencing this effect will always meet the minimum quorum requirement (n) within the time window. For large values of n, the number of misses is very high and consequently, the number of cor rect detections becomes very low, since the quorum requirements become too stringent and cannot be satisfied in most or all cases. In order to determine the optimal value of n, we devised a score function (Equation 7 -1) which is a weighted sum of the number of successful detections, number of false positives and number of false negatives. S (D, FP, FN) = D FP 0.5FN (7 1) where D is the number of footsteps detected successfully, FP is the number of false positives and FN is the number of false negatives. In case of the Smart Floor, we decided that the occurrence of a false positive is less desirable than the occurrence of a false negative and we set their respective weights in the equation accordingly. Note that the formula of the score function (Equation 7 1) is specific to the Smart Floor and it is dependent on the type of application and the relative impact of false positives and false negatives on it. We found that for the Smart Floor in the Gator Te ch Smart House, the score is the highest when n is equal to 3 (Figure 7 10). This is supported by the fact that this value of n ensures a reasonably good level of performance, where the number of false positives is comparatively low as compared to the number of correct detections and approximately 77% of all footsteps are successfully detected. W e observe that as threshold probability pT increases the number of false positives decreases since it filters out random spikes (Figure 7 11) Random spikes typica lly result in only

PAGE 111

111 a few readings of significant magnitude within a fixed size sliding window hence, there is a sharp drop in the number of false positives even when we only increase pT from 0.1 to 0.2. However making the probability requirement more strin gent also results in an increase in the number of false negatives/misses. This is due to the fact that we are using a sliding window of fixed size. A s the requirement on the number of readings that have to lie within the phenomena defined bounds [a, b] in creases, the chances of the Probability -Condition getting satisfied decreases. Usi ng the score function defined previously (Equation 7 1 ), we found that setting pT equal to 0.4 (Figure 7 12) results in a reasonably good detection rate with a low number of false positives and misses. Figure 7 12. Determining the optimal value of pT We observe that if the sliding window size is too low, this results in a large number of false positives despite having a high probability threshold (Figure 7 13) This is due to the fact that in case of sensors which have a high sampling rate even random spikes can result in a fairly large contiguous set of significant readings. Since the system essentially uses the threshold frequency (m. pT) to evaluate whether a sensor satis fies the Probability -Condition, for small window sizes 2000 1500 1000 500 0 500 1000 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 S (Score Function) pT

PAGE 112

112 the corresponding threshold frequency is also low even if threshold probability pT is kept high. Hence, there is a high probability that random spikes get mistaken for actual footsteps. Increasing the sliding window size on the other hand, raises the threshold frequency, which as we can observe results in a moderate increase in the number of misses. Using the score function (Equation 7 1), w e found that for the Smart Floor, setting the sliding window si ze m=150 results in reasonably good detection performance without taxing memory resources of individual sensor nodes (Figure 7 14) Figure 7 13. Effect of varying m with n=3 and pT=0.4 Experiment II: Resource and Power Consumption In the second set of experiments we stu died the system resource and energy consumption characteristics of our approach and compare d it with that of a stream -based approach. We use d a simulation based approach where we simulate d the spawning and random movement of multiple phenomenon clouds on a rectangular grid of sensors and conduct ed a compara tive analysis of the resource and energy consumption of each technique. 0 200 400 600 800 1000 1200 25 50 75 100 125 150 175 200 225 250 275 300 325 350 375 400 425 450 475 500 False Positives Missed Detected m

PAGE 113

113 Table 7 2 Energy consumption s pecifications for Atlas ZigBee n odes Operation Current (mA) Durati on (seconds) Sampling Sensors, Processing or Listening for messages 6.0 2.0 Receive Message @256kbps 15 0.002 Transmit Message @256 kbps 15 0.002 Figure 7 14. Determining optimal value of m Figure 7 15. Number of upda te messages sent to the query processor 400 300 200 100 0 100 200 300 400 500 600 25 50 75 100 125 150 175 200 225 250 275 300 325 350 375 400 425 450 475 500 S (Score Function) m 13658 46203 58741 45032 49986 0 1000000 2000000 3000000 4000000 5000000 6000000 50x50 100x100 200x200 300x300 400x400 Total Number of Update Messages Sensor Grid Size DistPDT StreamPDT

PAGE 114

114 Experimental s etup We simulate d the movement of phenomenon clouds over 100 epochs on a rectangular grid of sensors of sizes varying from a 50 x 50 grid to a 400 x 400 grid. Each sensor wa s connected to an Atlas ZigBee node. At the beginning of each simulation, between 2 to 5 phenomeno n clouds were spawned randomly in different areas of the grid. During each epoch the dire ction of motion for each cloud wa s randomly decided and the variation of its motion and size simulated by selecting sub sets of sensors which enter ed or left the Phenomenon-Set Hence, the simulation can be viewed as a random walk of multiple phenomenon clouds over a sensor grid. Our simulation i ntroduced a high level of uncertainty regarding phe nomenon cloud movement and t ested the performance of detection and tracking algorithms to the fullest extent, especially in the presence of multiple clouds. S tatistics such as the number of network messages, processing costs and updates sent to the query processor w e re logged during the simulation process The energy consumption of the nodes wa s calculated as a function of processing costs (including sampling sensors) and network costs (incurred in receiving and transmi tting data over the radio) and wa s based on the Atlas ZigBee node hardware specifications (Table 7 2 ), which is in turn based on the Atmel Zlink R adio C ontrol Board (RCB) design. The d etection and tracking strategies that were simulated in the experiments for purpose s of comparison were: (1) StreamPDT which is the c entr alized stream -based algorithm where phenomena detection and tracking is performed by a centralized query processor such as Nile PDT and; (2) DistPDT t he d istributed acquisition based phenomenon detection and tracking algorithm described in this chapter R esults and a nalysis The simulation results are shown in Figures 7 15, 7 16, 7 17 and 7 18. The numbers shown on each graph correspond to values of data points of their respecti ve plots for DistPDT.

PAGE 115

115 First, we consider the total number of updates rega rding the position of phenomena clouds that were sent to the query processor during the entire course of the simulation. The total number of update messages sent is a reflection of the processing and networking workload put on the centralized query processor by each algorithm and can be considered to be a comparative test of scalability. Figure 7 16. Average number of active sensors required We observe that in the case of Stream PDT the number of update messages increases significantly with the size of the sen sor grid (Figure 7 15) This is due to the fact that stream based algorithms use centralized processing as opposed to in -network processing. Therefore, these mechanisms require inputs from all the sensors in the grid to detect and track phenomenon clouds. The number of update messages sent by DistPDT shows negligible increase as sensor grid size varies. In fact, DistPDT requir es transmission of between 78% to 99% fewer update messages than StreamPDT This can be explained by the fact that in DistPDT, all the processing 1680 4951 6280 5130 5854 0 50000 100000 150000 200000 250000 50x50 100x100 200x200 300x300 400x400 Average Number of Sensors involved in Detection and Tracking Sensor Grid Size StreamPDT DistPDT

PAGE 116

116 is done innetwork and update messages are only sent to the query processor when a phenomenon cloud grows, shrinks or moves that is, when sensors move in or out of the Phenomenon-Set Therefore, regardless of the increase in size of the sensor grid, given similar phenomenon cloud patterns, the number of updates required by the query processor from the DistPDT mechanism does not change significantly, thereby reducing the query processors work load and avoiding potential bottlenecks. Figure 7 17. Number of network messages exchanged This becomes an important issue when we consider smart spaces where base station hardware is not a full -fledged computer or server but instead might be something along the lines of a set -top box with limited process ing and memory resources. In that case the base station will lack the high-performance computing capabilities required to simultaneously process data streams from a large number of sensors. In such circumstances our approach is more suitable since the quer y processor running inside the set -top box is only required to inject queries into the sensor network and receive intermittent updates about the state of the phenomena in the smart 114456.4 348693.8 450815.7 327036.8 377017.6 0 1000000 2000000 3000000 4000000 5000000 6000000 50x50 100x100 200x200 300x300 400x400 Number of Network Messages exchanged Sensor Grid Size StreamPDT DistPDT

PAGE 117

117 space. The actual execution of the detection and tracking process is done i n -network in an autonomous fashion without requiring any form of intervention from the centralized base station. Next we look at the average number of sensors involved in the real -time tracking of phenomenon clou ds during simulation (Figure 7 16) Since St ream PDT require s inputs from all sensors in the grid to be str eamed up to the query processor, the increase in number of sensors actively involved in sampling their readings is equivalent to the increase in sensor grid size. For DistPDT, we observe that th e average number of active sensors is around 20 006000 sensors, anywhere from 3 0% to 98% less than StreamPDT This is due to the fact that, the DistPDT mechanism tries to localize the process of detection and tracking to are as of the grid, where phenomena clouds have been observed or are likely to be present, significantly reduc ing the number of sensors which have to be actively sampled Moreover, due to localization of in network processing among sensor nodes, the number of active sensors does not change s ignificantly with increase in grid size, given similar phenomenon cloud behavior. Distributed, in -network detection and tracking reduces bottlenecks but this comes at the cost of exchanging a higher number of inter -node messages to coordinate efforts amo ng multiple nodes. Unlike Stream PDT where coordination is done in a centralized fashion and data aggregated at a single sink, DistPDT requires each sensor node to co -ordinate its own activities based on the state of its neighboring sensors. This naturally le ads to an increase in the total number of network messages exchanged as compared to centralized mechanisms. However this is true only for small grid sizes, where the percentage of sensors participating in the detection and tracking process will be higher than larger grids (Figure 7 17) For large grids, the cost of additional inter -node messaging incurred by DistPDT is comparatively outweighed by StreamPDTs processing and multi -hop networking costs incurred by non-participating sensor nodes.

PAGE 118

118 If we consider the average energy consumption per node over the entire simulation (Figure 7 18), we observe that DistPDT has 60% to 98% lower average energy consumption than StreamPDT. Hence, DistPDT enables non -participating sensors to conserve their energy and prolong their availability for other tasks. Based on these experiments, we can conclude that DistPDT is much more scalable and energy -efficient than centralized stream -based al gorithms, thereby making it suitable for practical deployment in real -world sma rt spaces such as the Gator Tech Smart House. Figure 7 18. Average energy consumption per node 0.34 0.18 0.08 0.03 0.02 0 0.1 0.2 0.3 0.4 0.5 0.6 50x50 100x100 200x200 300x300 400x400 Average Energy Consumption per Sensor Node (Joules) Sensor Grid Size StreamPDT DistPDT

PAGE 119

119 CHAPTER 8 CONCLUSIONS AND FUTU RE WORK W e described Sensable a scalable query processing middleware for service-oriented sensor networks. Sensable l everages service oriented architecture (SOA) concepts to provide query and event processing capabilities which go beyond the primitive types of data directly available from individual hardware sensors and allow the sensor network to handle queries involvin g abstract data types and difficult to define events and phenomena It also utilizes knowledge associated with devices to minimize energy consumption of sensor networks utilizing new low -power networking technologies such as ZigBee. F uture work for the Se nsable project is as follows For sensor aware query processing, we are planning on conducting experimental evaluations of query plan performance based on actual execution monitored on various sensor platforms connected to real sensors. We are also conside ring more rigorous simulation -based evaluation of plan performances using specifications of other contemporary sensor platform hardware. Furthermore, we are exploring more sophisticated methods for optimizing sensor sampling schedules across the network, e specially approximation techniques based on model -based prediction. For the virtual sensors framework, we are planning on developing fault tolerance mechanisms for derived virtual sensors by u sing the SOA -based virtual sensor concept in conjunction with se rvice selection and replacement mechanisms for ubiquitous service composition in pervasive spaces. This implies that w henever a virtual sensor fails, service selection and replacement mechanisms will be activated to locate another suitable virtual sensor or instantiate a new one. We are also actively pursuing real -world virtual sensor application deployments in the Gator Tech Smart House and are currently concentrating on a pplying virtual sensors for unencumbered monitoring of conditions such as obesity and diabetes of its residents

PAGE 120

120 Finally, in the case of phenomena detection and tracking, we are looking at model assisted detection and tracking where, distribution of detection tasks will be streamlined based on predictions regarding direction of movement of phenomenon clouds. We do not require the existence of formal cloud models for our detection and tracking to work H owever, in case they do exist and are computationally feasible, their input can be utilized towards reducing the number of nodes involved in the grid, further reducing network and processing costs. We are also working on ways to better organize the Phenomenon-Set to enable updating and lookup of specific clouds and their characteristics if possible, in constant time. We are also actively sea rching for real -life datasets from different application domains for further validation and fine tuning of our approach.

PAGE 121

121 LIST OF REFERENCES [1 ] K. Aberer, M. Hauswirth, and A. Salehi A middleware for fast and f le xible sensor network Deployment. In Proceedings of the 32nd International Conference on Very Large Data Bases pages 11991202, Seoul, Korea, 2006. [2 ] M. Ali W. Aref, R. Bose, A. Elmagarmid, A. Helal, I. Kamel and M. Mokbel. NILE PDT: A phenomenon detection and t racki ng f ramework for data str eam management systems. In Proceedings of the 31st International Conference on Very Large Data Bases pages 12951298, Trondheim, Norway, 2005. [3 ] M. Ali M. Mokbel, W. Aref and I. Kamel. Detection and tracking of discrete phenomena in se nsor -network databases. In Proceedings of the 17th International Conference on Scientific and Statistical Database Management pages 163172, Santa Barbara, USA, 2005. [4 ] S. Bhattacharya, N. Atay, G. Alankus, C. Lu, B. Bayazit, and G. C. Roman. Roadmap query for sensor network assisted navigation i n dynamic environments. In Proceedings of the International Conference on Distributed Computing in Sensor Systems pages 17 36, San Francisco, USA, 2006. [5 ] R. Bose and A. Helal. Selectivity -based query plan generation in service -oriented s e nsor networks. In Technical Report Department of Computer and Informati on Science and Engineering, E305 CSE Building, University of Florida, Gainesville, USA, 2007. http://www.c ise.ufl.edu/~rbose/TSH TR.pdf [6 ] R. Bose J. King, H. El Zabadani, S. Pickles, and A. Helal. Building plug and p l ay s mart homes using the Atlas Platform. In Proceedings of the 4th International Conference on Smart Homes and Health Telematics pages 265 272, Belfast, Northern Ireland, 2006. [7 ] K. Chintalapudi and R. Govindan. Localized edge detection in sensor fields. In Proceedings of the 1st IEEE Workshop on Sensor Network Protocols and Applications pages 59 70, Anchorage, USA, 2003. [8 ] A. Deshpande, C. Guestrin, S. Madden, J. Hellerstein, and W. Hong. Model -driven data a cquisition in sensor networks. In Proceedings of the 30th International Conference on Very Large Data Bases pages 588599, Toronto, Canada, 2004. [9 ] S. de Deugd, R. Carroll, K. Kelly, B. Millett, an d J. Ricker. SODA: Servi ce-oriented device architecture. IEEE Pervasive Computing 5(3) : 94 96, 2006. [10] J. Garcia, A. Robertson, J. Ortega, and R. Johansson. Sensor fusion for compliant robot motion control. IEEE Transactions on Robotics 24(2): 430441, 2008. [11] G. Hartl and B. Li. Infer: Bayesian inference approach towards energy-efficient data collection in dense sensor n etworks. In Proceedings of the 25th IEEE International Conference on Distributed Computing Systems pages 371 380, Columbus, USA, 2005.

PAGE 122

122 [12] J Heidemann, F. Silva, and D. Estrin. Matching data dissemination algorithms to application requirements. In Proceedings of the 1st International Conference on Embedded Networked Sensor Systems pages 218229, Los Angeles, USA, 2003. [13] A. Helal, W. Mann, H. El -Zabadani, J. King, Y. Kaddoura, and E. Jansen. The Gator Tech Smart House: A programmable pervasive space. IEEE Computer 38( 3 ): 50-60, 2005. [14] R. Jiang and B. Chen. Fusion of censored d ecisi ons in wireless sensor networks. IEEE Transactions on Wireless C ommunications 4(6) :26682683, 2005. [15] P. Juang, H. Oki, Y. Wang, M. Martonosi, L. Peh, and D. Rubenstein. Energy-efficient c om puting for wildlife tracking: Design t radeoffs and early experiences with ZebraNet. In Proceedings of the 10th International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS-X) pages 96 107, San Jose, USA 2002. [16] S. Kabadayi, A. Pridgen, and C. Julien. Virtual s ensors: Abstrac ting data from physical sensors. In Proceeding s of the IEEE International Symposium on a World of Wireless, Mobile and Multimedia Networks pages 587592, Niagara Falls, USA, 2006. [17] Y. Kaddourah, J. King and A. Helal. Cost precision tradeoffs in unencumbered f loor based indoor location tracking. In Pr oceedings of the 3rd International C onference o n Smart Homes and Health Telematics pages 75 82, Sherbrooke, Canada, 2005. [18] J. King, R. Bose S. Pickles, A. Helal, and H. Yang. Atlas A service -oriented sensor platform. In Proceedings of the 1st IEEE Inter national Workshop on Practical Issues in Building Sensor Network Applications pages 630638, Tampa, USA, 2006. [19] C. Lee, D. Nordstedt and A. Helal. Enabling smart spaces with OSGi. IEEE Pervasive Computing, 2 (3 ): 89 94, 2003. [20] F. Lewis. Wireless sensor networks. Smart Environments: Technologies, Protocols and Applications pages 13 46, John Wiley, 2004. [21] X. Liu, Q. Huang, and Y. Zhang. Combs, needles and haystacks: Balancing push and pull for discovery in large -scale sensor networks. In Proceedings of the 2nd International Conference on Embedded Networked Sensor Systems pages 122 133, Baltimore, USA, 2004. [22] S. Madden, M. Franklin, J. Hellerstein, and W. Hong. TAG: A tiny aggregation service for ad hoc sensor networks. In Proceedings of the 5th Annual Symposium on Operating Systems Design and Implementation, pages 131146, Boston, USA, 2002. [23] S. Madden M. Franklin, J. Hellerstein, and W. Hong. TinyDB: An acquisitional query process ing system for sensor networks. ACM Transactions on Database Systems 30(1) : 122 173, 2005.

PAGE 123

123 [24] N. Malpani, J. Welch, and N. Vaidya. Leader election algorithms for mobile ad hoc networks. In Proceedings of the 4th International Workshop on Discrete Algorithms and Methods for Mobile Computing and Communications pages 96 103, Boston, USA, 2000. [25] D. McErlean and S. Narayanan. Distributed detection and tracking in sensor networks. In Proceedings of the 36th Asilomar Conference on Signals, Systems and Computers pages 11741178, Asilomar, USA, 2002. [26] R. OlfatiSaber and J. Shamma. Consensus filters for sensor networks and distributed sensor fusion. In Proceedings of the 44th IEEE Conference on Decision and Control and the European Control Conference pages 66986703, Seville, Spain, 2005. [27] A. Omotayo, M. Hammad, and K. Barker. Efficient data h arvesting for tracing phenomena in sensor n etw orks. In Proceedings of the 18th International Conference on Scientific and Statistical Database Management pages 59 70, Vienna, Austria, 2006. [28] S. Shenker S. Ratnaswamy, B. Karp, R. Govindan, and D. Estrin. D ata -centric s torage in sensornets. ACM SIGCOMM Computer Communication Review 33(1) : 137-142, 2003. [29] A. Silberstein. Push and p ul l in sensor network query processing. In Proceedings of the Southeast Workshop on Data and Information Management Raleigh, USA, 2006. [30] A. Silberstein, A. Gelfand, K. Munagala, G. Puggioni, and J. Yang. Making sense of suppressions and failures in sensor data: A Bayesian approach. In Proceedings of the 33rd International Conference on Very Large Data Bases pages 842 853, Vienna, Austria, 2007. [31] C. Town and Z. Zhu. Sensor fusion and environmental modeling for multimodal sentient c omputing In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1 2 Minneapolis, USA, 2007. [32] N. Trigoni Y. Yao, A. Demers, J. Gehrke, and R. Rajaraman. Hybrid push -pull query processing for sensor networks. GI Jahrestagung, (2): 370374, 2004. [33] S. Vasudevan, B. DeCleene, N. Immerman, J. Kurose and D. Towsley. Leader election algorithms for wireless ad hoc net works. In Proceedings of the 3rd DARPA Information Survivability Conference and Exposition, pages 261272, Washington DC, USA, 2003. [34] A. Woo and D. Culler. A transmission control scheme for media access in sensor networks. In Proceedings of the 7th Annual I nternational Conference on Mobile Computing and Networking pages 221235, Rome, Italy, 2001.

PAGE 124

124 BIOGRAPHICAL SKETCH Raja Bose was born in Darjeeling in 1981 and raised in New D elhi, India. He has a Bachelor of Science degree (with Hono rs) in s tatistics from Delhi University, India and two Master of Science degrees from the University of Florida: one in applied mathematics and the other in computer e ngineering. Since 2005, h e has been conducting research under the guidance of Dr. Abdelsalam (Sumi) Helal, in the area of service -oriented sensor networks and smart spaces, at the Mobile and Pervasive Computing Laboratory of the University of Florida. He w as awarded the Doctorate in Computer Engineering in May 2009 and joined Nokia Research Center Palo Alto in January 2009, as a member of research s taff specializing in ubiquitous device interoperability.