ENERGY EFFICIENT WIRELESS SENSOR NETWORKS BASED ON

MACHINE LEARNING

MOHAMMAD ABDULAZIZ ALWADI

A Thesis Submitted for the

Degree of Doctor of Philosophy

Faculty of Education Science Technology and Mathematics

October 2015

i

In the name of Allah, the most Merciful, the most compassionate

ii

ح لِّي ص "ر ِّ اْشر ْر لِّ يْدرب ن لِّساني ي * واْحل ي أ مرِّ * ويس ِّ ْل ع قدةً مِّ

(52-52* ي فقه وا ق ْولِّي" )سورة طه

My Lord, 'expand my chest, And ease my task for me. (Grant me self

confidence, contentment, and boldness)

Unloose the knot upon my tongue,

That they may understand my speech. (Quran 20: 25-28)

iii

I dedicate this to My Mother and My Father, My wife EMAN and My Son

OMAR.

v

Abstract

The field of wireless sensor networks have become a focus of intensive research in recent years,

especially for monitoring and characterizing of large physical environments, and for tracking

various environmental or physical conditions such as temperature, pressure, wind and humidity.

Wireless Sensor networks can be used in many applications, such as wildlife monitoring,

military target tracking and surveillance, hazardous environment exploration, and natural

disaster relief. Given the huge amount of sensed data, automatically classifying them becomes

a critical task in many of these applications. Energy efficiency is a key issue in wireless sensor

networks where the energy sources and battery capacity are very limited. To address some of

key WSN challenges, a novel integrated framework for achieving energy efficiency is proposed

consisting of three stages of modelling from data. The first stage is a joint energy efficiency–

event detection model, where a novel sensor node selection technique is designed, that

conserves the energy in the wireless sensor network and at the same time maximizes the event

recognition performance. Here, the scheme utilises, fewer sensor nodes at a time, and placing

unwanted sensor nodes in the sleep mode. For this, a novel objective quantitative metric is

proposed to assess the energy efficiency achieved, namely, the life time extension factor

(LTEF). It was shown with extensive experimental evaluation, that this joint scheme, allows

selection of most significant and influential sensor nodes for participation in different WSN

tasks, and contributes significantly towards energy savings and event detection accuracy. As

the WSN needs to adapt to the state of the environment being monitored dynamically, the

number of sensor nodes participating in the routing tree cannot remain fixed, and need to adapt,

in order to accurately monitor and predict the physical environment, and the second stage in

this framework, is a proposal for adaptive models for sensor selection and classifier learning

for achieving energy efficiency and prediction accuracy, based on performance targets

specified. The third stage is a joint energy efficiency–adaptive routing model, where an

appropriate sensor selection and adaptive routing strategy allows addressing the WSN

challenges corresponding to energy efficiency, prediction accuracy, and MAC layer adaptation.

We show that this joint model, also meets non-functional performance targets, such as missing

or faulty sensors, model building time, needed for adaptation of routing protocol.

vii

Acknowledgments

I would like to express my sincere appreciation to all the staff at the University of Canberra and

Dr. Girija Chetty for all her support and assistance in the preparation of this thesis. My sincere

appreciation to all my Family, Dad, Mum, wife and My Son for supporting me and their

encouragement throughout my life.

ix

Table of Contents

Contents

TABLE OF CONTENTS IX

FORM B XVII

CERTIFICATE OF AUTHORSHIP OF THESIS XVII

KEY TERMS XIX

CHAPTER 1 INTRODUCTION 1

1.1 Introduction 1

1.2 Significance and Motivation 3

1.3 Background 5

1.4 Research Questions 14

1.5 Thesis Contributions 14

1.6 Publications 16

1.7 Organisation of Thesis 17

CHAPTER 2 RELATED WORK AND LITERATURE REVIEW 19

2.1 Machine Learning Based Approaches 19

2.2 Related Work on Machine Learning for WSNs 23

2.2.1 Supervised Machine Learning 23

2.2.2 Unsupervised Machine Learning 27

2.2.3 Reinforcement Machine Learning 28

2.3 Operational Challenges 29

2.3.1 WSN Routing Issues 29

2.3.2 Data Collection and Clustering Issues 34

2.3.3 Event Recognition & Query Processing Issues: 40

2.3.4 Challenges Related to Localisation and Object Targeting 44

2.3.5 Medium Access Control (MAC) Issues: 50

2.4 Non-operational Aspects of WSN 53

2.4.1 Security and Anomaly Intrusion Detection 54

x

2.4.2 Data Integrity, Fault Detection, and QoS Enhancement: 56

2.4.3 Application Specific Unique Challenges 59

2.5 Research Gap in Wireless Sensor Networks Based on Machine Learning/Data Mining

Techniques 62

2.5.1 Better Methods for Selecting Sensors 62

2.5.2 Adaptive and Distributed Machine Learning Approaches For WSNs 62

2.5.3 Managing Resources Using Machine Learning 63

2.5.4 Spatio-Temporal Correlation Detection 63

2.6 Research Plan and Thesis Road Map 63

CHAPTER 3 JOINT SENSOR SELECTION - EVENT DETECTION SCHEME 65

3.1 Introduction 65

3.2 Joint Energy Efficiency - Event Detection Scheme 65

3.2.1 Energy Efficiency with Feature Ranking Algorithm 65

3.2.2 Naïve Bayes Machine Learning Classifier Algorithm 67

3.3 Experimental Validation 68

3.3.1 Experiment 1 (Isolet Data set) 69

3.3.2 Experiment 2 (Ionoshpere dataset) 72

3.3.3 Experiment 3 (forest Cover type data set) 73

3.3.4 Experiment 4 (Forest fires Dataset) 75

3.4 Chapter Summary 77

CHAPTER 4 ADAPTIVE MODELS FOR ENERGY EFFICIENCY 79

4.1 Introduction 79

4.2 Adaptive Classifier Model Based Scheme 79

4.2.1 Data set Description 80

4.2.2 Classification Algorithms 81

4.2.3 Experimental Evaluation 82

4.3 Discussion 84

4.4 Adaptive Classifier Scheme with Gas Sensor Drift Dataset 87

4.4.1 Experimental Validation with Gas Drift Dataset 87

xi

4.4.2 Experimental Validation with Gas Drift Dataset using Ensemble Learning for Weak

Classifiers 89

4.5 Chapter Summary 90

CHAPTER 5 JOINT SENSOR SELECTION- ADAPTIVE ROUTING MODEL 91

5.1 Introduction 91

5.2 Intel Berkeley Lab WSN dataset 91

5.3 Intel Lab data file versus Intel Lab data file restructured for experiments 94

5.4 Sensor Selection and Adaptive Routing Model 98

5.5 Experimental Results and Discussion 99

5.6 Chapter Summary 104

CHAPTER 6 CONCLUSIONS AND FUTURE DIRECTIONS 107

BIBLIOGRAPHY 111

xiii

List of Figures

FIGURE 1 A TYPICAL WIRELESS SENSOR NETWORK [6] 5

FIGURE 2 TAXONOMY OF ENERGY EFFICIENT APPROACHES FOR WIRELESS SENSOR NETWORKS. 7

FIGURE 3 ESTIMATING NODE LOCALIZATION CO-ORDINATES IN WSN USING NEURAL NETWORKS [82] 25

FIGURE 4 SCHEMATIC OF SVM CLASSIFICATION PROCESS [91] 26

FIGURE 5 TWO DIMENSIONAL VISUALIZATION OF PCA PROCESS [103] 28

FIGURE 6 VISUALIZATION OF Q-LEARNING ALGORITHM [108] 29

FIGURE 7 SIMPLIFIED NETWORK ROUTING BASED ON MACHINE LEARNING [46] 31

FIGURE 8 VISUALIZATION OF Q-LEARNING ALGORITHM [118] 35

FIGURE 9 EVENT DETECTION AND QUERY PROCESSING USING MACHINE LEARNING [46] 41

FIGURE 10 HMM AND NAÏVE BAYES EVENT DETECTION AND QUERY PROCESSING [132] 42

FIGURE 11 LOCALIZATION USING BEACON NODES IN WSN [82] 45

FIGURE 12 ALOHA-QIR SCHEME FOR MAC LAYER IN WSN [152] 52

FIGURE 13 ADAPTIVE DECISION TREE BASED MAC PROTOCOL (SAML) [155] 53

FIGURE 14 BASIC CONCEPTS OF ANOMALY INTRUSION DETECTION [54] 54

FIGURE 15 WSN BASED Q-LEARNING FOR OBJECT TRACKING APPLICATION [174] 60

FIGURE 16 BLOCK SCHEMATIC FOR JOINT ENERGY EFFICIENCY - EVENT DETECTION SCHEME 66

FIGURE 17 SENSOR SELECTION AND RANKING ALGORITHM 66

FIGURE 18 EVENT DETECTION ACCURACY VS. LIFE TIME EXTENSION FACTOR(LTEF) (ISOLET 5 DATA SET) 71

FIGURE 19 ACCURACY AND LIFE TIME EXTENSION FACTOR (IONOSPHERE) 73

FIGURE 20 ACCUARCY AND LIFE TIME EXTENSION FACTOR (FOREST COVER TYPE DATA SET) 75

FIGURE 21 ACCURACY AND LIFE TIME EXTENSION FACTOR FOR FOREST FIRES DATA SET 77

FIGURE 22 ADAPTIVE FEATURE SELECTION AND CLASSIFIER MODEL FOR ENERGY EFFICIENCY 80

FIGURE 23 PERFORMANCE OF CLASSIFIERS WITH 10 FOLDS CROSS VALIDATION 82

FIGURE 24 PERFORMANCE OF CLASSIFIERS WITH FULL TRAINING SET 83

FIGURE 25 PERFORMANCE OF CLASSIFIERS WITH FEATURE SELECTION 83

FIGURE 26 PERFORMANCE OF CLASSIFIERS WITH FEATURE SELECTION ON FULL TRAINING SET 84

FIGURE 27 COMPARATIVE CLASSIFIER PERFORMANCE 86

FIGURE 28 GAS DRIFTS SUMMARY OF EXPERIMENTAL RESULTS 88

FIGURE 29 INTEL BERKELEY WIRELESS SENSOR NETWORK DATA SET: LOCATION OF 54 SENSORS IN AN AREA OF 1200

M2 92

FIGURE 30 JOINT SENSOR SELECTION – ADAPTIVE ROUTING MODEL 94

FIGURE 31 INTEL LAB MAIN SOURCE FILE STRUCTURE 94

FIGURE 32 SAMPLE FILES TEMPERATURE READINGS 35, 2700 AND 5400 SAMPLES 96

FIGURE 33 SAMPLE FILES TEMPERATURE READINGS 35, 2700 AND 5400 SAMPLES 97

FIGURE 34 TEMPERATURE SENSOR SELECTION MAP FOR 3 EXPERIMENT SCENARIOS- 1,2 AND 3 98

FIGURE 35 HUMIDITY SENSOR SELECTION MAP FOR 3 EXPERIMENT SCENARIO 1,2 AND 3 99

FIGURE 36 TEMPRATURE EXPERIMENT 1,2 AND 3 RESULTS 101

FIGURE 37 HUMIDITY EXPERIMENT 1,2 AND 3 RESULTS 101

FIGURE 38 TEMPERATURE, ROOT MEAN SQUARE ERROR 102

xiv

FIGURE 39 HUMIDITY, ROOT MEAN SQUARE ERROR 103

FIGURE 40 TIME TAKEN TO BUILD THE MODEL, TEMPERATURE 104

FIGURE 41 TIME TAKEN TO BUILD THE MODEL, HUMIDITY 104

xv

List of Tables

TABLE 1. DATA SETS FOR EXPERIMENTAL VALIDATION 68

TABLE 2 FEATURES SELECTED IN ISOLET 5 69

TABLE 3 NAÏVE BAYES CLASSIFIER PERFORMANCE 70

TABLE 4.RESULTS OF EXPERIMENT 1 71

TABLE 5. EXPERIMENT 1 ACCURACY WITH SENSOR FAILURE PROBABILITY 72

TABLE 6 EXPERIMENT 2 FEATURES SELECTED AND RANKED ON IONOSPHERE DATASET 72

TABLE 7 EXPERIMENT 2 ACCURACY 72

TABLE 8. EXPERIMENT 2 RESULTS 73

TABLE 9 EXPERIMENT 3 FEATURES RANKED AND SELECTED FOR FOREST COVER TYPE DATASET 74

TABLE 10 EXPERIMENT 3 ACCURACY AND LIFE TIME EXTENSION FACTOR 74

TABLE 11. EXPERIMENT 3 RESULTS 75

TABLE 12 SELECTED FEATURES ON FOREST FIRES DATASET 76

TABLE 13 EXPERIMENT 2 ACCUARCY FOREST FIRES DATA SET 76

TABLE 14 EXPERIMENT 4 RESULTS 77

TABLE 15 FOREST COVER TYPE ORIGINAL DATA SET AND SUBSET DATA SET DESCRIPTION 81

TABLE 16 GAS SENSOR ARRAY DRIFT DATA SET DESCRIPTION 87

TABLE 17 PERFORMANCE OF GAS DRIFTS SENSOR DATASET 88

TABLE 18 ENSEMBLE LEARNING ON GAS DRIFT SENSOR ARRAY DATA SET 90

TABLE 19 INTEL LAB DATA SET FILE SCHEMA 92

TABLE 20 TEMPERATURE RESULTS FROM THREE EXPERIMENTS SCENARIOS 100

TABLE 21 HUMIDITY RESULTS FROM THREE EXPERIMENTS SCENARIOS. 100

xvii

Form B

Certificate of Authorship of Thesis

Except where clearly acknowledged in footnotes, quotations and the bibliography, I certify

that I am the sole author of the thesis submitted today entitled

ENERGY EFFICIENT WIRELESS SENSOR NETWORKS BASED ON MACHINE

LEARNING.

(Thesis title)

I further certify that to the best of my knowledge the thesis contains no material previously

published or written by another person except where due reference is made in the text of the

thesis.

The material in the thesis has not been the basis of an award of any other degree or diploma

except where due reference is made in the text of the thesis.

The thesis complies with University requirements for a thesis as set out in Gold Book Part 7:

Examination of Higher Degree by Research Theses Policy, Schedule Two (S2). Refer to

http://www.canberra.edu.au/research-students/goldbook

Signature of Candidate

........................................................................

Signature of chair of the supervisory panel

Date: ……12/7/15……………..

xix

Key Terms

Sensor network

Wireless sensor network

Wired sensor network

Data mining

Classification

Feature selection

Data set

Attributes

Physical environment

Environment Monitoring

Environment characterization

Source node

Sink node

Sensor Failure

Active mode

Sleep mode

Accuracy

Life time extension factor

Energy Efficiency

WEKA data mining software

UCI Repository

Intel lab Wireless sensor network

Mote ID

Root Mean squared error

Feature ranking Algorithm

Feature selection Algorithm

Intelligent monitoring

Intel Berkeley lab

Routing approach

Routing map

Ensemble Learning

Trade off

Simulation tools

Chapter 1

1

Chapter 1 Introduction

1.1 Introduction

The real world physical environment consists of large and diverse information sources, such as

light, temperature, motion, seismic waves, and many others. For a better understanding of the

environment, it is necessary to capture the information from multiple disparate sources, and the

wireless sensor network is an easy to deploy infrastructure allowing capturing of such rich

information.

A wireless sensor network (WSN) consists of spatially distributed autonomous sensors to

monitor the physical environment, and to co-operatively pass their data through the network to a

main node or central location (base station). Modern wireless sensor networks are bi-directional,

allowing transmission of information being monitored from nodes to central node or base station,

as well as enabling control of sensor activity from base station to sensors. The development of

wireless sensor networks was motivated primarily by military applications such as battlefield

surveillance; but today such networks are used in many industrial and consumer applications,

such as industrial process monitoring and control, machine health monitoring, environmental

detection, and habitat monitoring. The WSN is built of "nodes” from a few to several hundreds

or even thousands of nodes (sometimes called as motes), where each node is connected to one

(or sometimes several) sensors. Each such sensor network node has typically several parts: a

radio transceiver with an internal antenna or connection to an external antenna, a microcontroller,

an electronic circuit for interfacing with the sensors and an energy source, usually a battery or an

embedded form of energy harvesting. A sensor node might vary in size from that of a shoebox

down to the size of a grain of dust, although functioning "motes" of genuine microscopic

dimensions have yet to be created. The cost of sensor nodes is similarly variable, ranging from a

few to hundreds of dollars, depending on the complexity of the individual sensor nodes. Size and

cost constraints on sensor nodes result in corresponding constraints on resources such as energy,

memory, computational speed and communications bandwidth. The topology of the WSNs can

vary from a simple star network to an advanced multi-hop wireless mesh network. The

propagation technique between the hops of the network can be determined based on routing or

flooding protocol[1, 2].

Chapter 1

2

A wireless sensor network can be used for various applications; we can summarize some of the

useful applications as the following:

1. Habitat/Area monitoring: Area monitoring is a common application of WSNs. In area

monitoring, the WSN is deployed over a region where some phenomenon is to be

monitored. A military example is the use of sensors to detect enemy intrusion; a civilian

example is the geo-fencing of gas or oil pipelines. When the sensors detect the event

being monitored (heat, pressure), the event is reported to one of the base stations, which

then takes appropriate action (e.g., send a message on Internet or to a satellite).

Similarly, wireless sensor networks can use a range of sensors to detect the presence of

vehicles ranging from motorcycles to trains and cars.

2. Environmental/Earth monitoring: The term Environmental Sensor Networks [3], has

evolved to cover many applications of WSNs to earth science research. This includes

sensing volcanoes oceans, glaciers and forests.

3. Critical Events/Forest fire detection: A network of sensor nodes can be installed in a

forest to detect when a fire has started. The nodes can be equipped with sensors to

measure temperature, humidity and gases which are produced by fire in the trees or

vegetation. Early detection is crucial as it will allow protection of highly valued

resources.

4. Data Logging: Wireless sensor networks are also used to collect data for monitoring

information from the environment. For example, monitoring the temperature in a fridge

to the level of water in over flow tanks in nuclear power plants.

As outlined above, a wide spectrum of applications ranging from habitat monitoring to battlefield

surveillance can be benefited by deploying the wireless sensor network (WSN) technology [1,

2]. Some of the benefits include low cost, easy deployment, high fidelity sensing, self-

organization of WSNs, among several other benefits [2]. However, despite many opportunities

the wireless sensor networks provide, using WSN technology comes with great challenges. These

challenges are associated with characteristics of wireless sensor networks namely:

1. Power consumption constraints for nodes using batteries or energy harvesting.

2. Ability to cope with node failures.

3. Mobility of nodes.

4. Communication failures.

5. Scalability to large scale of deployment.

Chapter 1

3

6. Ability to withstand harsh environmental conditions.

7. Ease of use.

Out of these characteristics, need to operate under severe resource constraints is one of the biggest

challenge with WSNs, which makes efficient design highly necessary.

1.2 Significance and Motivation

A wireless sensor network, or WSN for short, is a large-scale network comprising of wirelessly

interconnected transducer devices called sensor nodes or “mote”. A sensor node, as the name

implies, can have one or more sensor modules, for sensing light, temperature, humidity, pressure,

and sound. In addition, each sensor node can include four other components, namely: memory,

processing, communication, and battery modules. The first use of sensor networks can be traced

back to the cold war era, when a distributed network of radars, and hydrophones were deployed

to monitor the skies and oceans, respectively [4]. Of late, contemporary monitoring networks use

tiny and resource-constrained sensor nodes.

The field of wireless sensor networks (WSN) has become a focus of intensive research in recent

years and various theoretical and practical questions have been addressed. It has drawn a lot of

attention as a result of the possibility of coupling these devices with their surroundings. Well

beyond their direct use, such as surveillance and environmental monitoring, WSNs can help us

pursue one of the ultimate goals in information technology, namely ambient intelligence [5]. The

small size and wireless communication capability of sensor nodes in a WSN provides us with

not only the information about the physical world around us, but also the flexibility to have them

integrated deeply within building material, fabrics, and embedded in inaccessible or hostile

locations in the real world operating scenarios. By using wireless sensor networks we can develop

automated intelligent systems that can co-operate with each other to exchange information

concerning their internal states and the conditions of the physical environment around them, and

provide services to users, and prevent disasters with better efficiency and robustness without any

human intervention [5].

The evolution of sensor networks has extended the computing horizons from desktop computing

to the entire physical environment computing (ambient computing). Due to this, the user-driven

model of traditional computing has shifted to an event-driven model in sensor networks.

Noticeably, the event-driven model entails that the volume of data generated by stimuli of

Chapter 1

4

environmental phenomena exceeds the rate of any user input by multiple folds. The traditional

model for interpreting this large volume of measurements normally involves sending large

sensory data to a base-station for analysis. The collected data could sometimes be locally

processed before being sent in the network, and could involve intermediate sensor nodes for

further processing of the data. Finally, the sensory data is integrated centrally at the base station

to infer the status of the observed environment at the base-station. The base station performs

optimal detection and tracking mechanism based on conventional signal processing methods.

This traditional model, however, suffers from many limitations due to resource-constraints and

the bandwidth limitations. The computational power and speed of base station computers can

create a processing bottleneck and can cause total system failure if the base-station fails. Further,

relaying all sensory data of geographically dispersed sensor nodes to a centralized base-station is

generally ineffective as it requires a significant communication overhead leading to resource

depletion and shortening of the lifetime of the network.

Several research works in the past [9],[10], [11], [12], and [13], tried to address these challenges

using the methods drawn from signal communication theory in telephony/telegraphy, where the

main purpose is the reliable transmission of data in the presence of noisy channels. However,

these approaches did not appear to work well for wireless sensor networks, as the purpose of

WSNs is not just the reliable transmission of data from sender to receiver, but also the detection

of occurrence of catastrophic events from large sets of sensory data, such as earthquakes,

Tsunamis, forest fires, land cover usage etc. Most of the current methods focus on solving the

local short term problem of enhancing the communication capacity between nodes or managing

the resources efficiently for a small WSN, with studies conducted on simulated setups.

Interpreting catastrophic global events from large volumes of data is a challenging task; and

research efforts needs to focus on development of novel approaches to improve the detection

accuracy and detection quality of high level information, where the WSN is deployed, such as,

accurate physical environment event detections, in addition to reduction in amount of data and

energy consumption in the sensor nodes in the network. Approaches to reduce the energy

consumption is one of the most important requirement, as there is no continuous power support

for battery powered sensors in WSNs deployed in the field. The life time of a sensor is very

restricted based on very limited power source. Therefore keeping the energy consumption in the

lowest level is one the key requirement.

Chapter 1

5

1.3 Background

Figure 1 A typical wireless sensor network [6], consisting of a base-station and a collection of

sensor nodes (also called “motes”). Although WSNs are anticipated for dense deployment of

thousands of nodes, some of current deployments range from ten to hundreds of sensor nodes.

Generally, in sensor networks for environmental monitoring and surveillance applications, the

events of interest occur rarely and suddenly. Therefore, the network traffic is typically very low.

However, the traffic flow increases abruptly, when and event of interest occurs leading to large

amounts of sensory data from various sensor nodes being conveyed to the base-station in the

event of a phenomenon of interest, leading to abrupt increase in traffic. To ensure that the event

of phenomenon of interest is captured properly and accurately, sensor nodes are deployed

densely. The densely deployed nodes not only ensure coverage and communication but also

tolerate node failures.

Figure 1 A typical wireless sensor network [6]

A dense sensor node deployment within close proximity of each other can result in an overlap in

coverage and communication. As a result, sensory measurements can contain high correlations

and redundancies. For instance, when the sensing range of two nodes covers the same area, both

sensor nodes will be likely transmitting identical sensory data. Although this ascertains a robust

sensor network tolerant to node failures and noisy sensory measurements, it can cause the sensor

nodes to consume precious battery resources for conveying the redundant data. One effective

approach to control the redundant data being communicated is to adjust the physical location of

Chapter 1

6

sensor nodes so as to minimize the overlap in their sensing ranges. However, adjusting the

location of sensor nodes may not be always possible, especially when using sensor nodes for

applications that require ad hoc and random deployment, e.g. battle field and emergency

applications. Hence, preferably, redundant data needs to be detected and removed.

One effective approach to minimize this energy consumption is the use of an appropriate

communication protocol called “multi hop communication”. With multi hop protocol, the

sensory data is communicated in several “hops” to neighbouring nodes to the base station, instead

of delivering data directly through a maximum range radio link to the base station. This is usually

better in terms of energy consumption. The multi hop communication protocol not only aids in

routing the data through several intermediate nodes, but also doing some node-level processing,

such as removing the data redundancy or for combining the data from other nodes. This

behaviour called in-network processing can contribute significantly to maximizing the longevity

of the WSN by switching off idle nodes. Since the events of interest occur rarely, switching off

battery powered sensor nodes located in inaccessible locations can conserve energy. Secondly,

the processing capabilities of sensor nodes in multi hop path can effectively help reduce the

volume of data transmitted in the network.

In order to predict the energy costs of different algorithms and protocols, and develop energy

efficient techniques, it is important to have an accurate understanding of the amount of energy

consumed at the sensor nodes. There are several sources of power consumption in a typical sensor

node, such as:

1. Sensor start up power,

2. Signal sampling rate,

3. Physical signal-to-electrical conversion,

4. Signal conditioning, and

5. Analogue-to-digital conversion.

In general, the amount of power consumed in the sensors for above mentioned processing stages

is negligible as compared to the energy consumed in the communication of the signal. To manage

the power consumed by a sensor node, for different types of sensors including temperature, photo

resistor, barometric pressure, humidity, passive infrared sensors, sonar rangers, and array sensors,

the processor within these sensor nodes support several operating modes, including active and

sleep modes. In the sleep mode a sensor node completely withholds all its activities and shuts

Chapter 1

7

down almost all of its components. The power consumption for Berkeley motes [7] , an example

wireless sensor node is 8 mill watts in active mode and 75 microwatts in sleep mode (around 10

times less power is consumed). The energy consumption in actual communication of data

between the sensor nodes is much higher than the sensor nodes. In general, there are four modes

of communication modes in a sensor node: transmit, receive, idle, and sleep modes. In transmit

mode, the energy consumption depends on the data rate (40kbps, 38.4kbps, and 250kbps). Other

factors associated with the power consumption and performance of a radio component (wireless

node) includes the type of modulation scheme used, choice of antenna, and duty cycle. The

receive mode of a sensor node also consumes lot of energy, and often has a third operating mode,

called the “idle” mode. The idle mode is different from the sleep mode. In sleep mode all the

radio components within the sensor node are completely shut down (larger energy saving),

whereas in the idle mode a sensor node switches off all its components except the receive radio

antenna.

Figure 2 Taxonomy of energy efficient approaches for wireless sensor networks.

For environment monitoring using wireless sensor networks, many applications are expected to

run continuously, in an unattended manner for several days and months. However, sensor nodes

are constrained by limited resources in terms of energy. And since communication between the

sensors, and, from sensors to central base station is more energy-consuming than the energy

consumption within the individual sensor nodes, appropriate design of sensor data processing

Chapter 1

8

techniques and the collection techniques that can limit the amount of transmitted data continues

to be an important and central issue for diffusion of wireless sensor network technology in real

world application, particularly for civilian operating scenarios, in spite of development of several

protocol driven approaches, such as multi hop communication protocols. Figure 2, shows the

taxonomy of some of earliest attempts on energy efficient methods for WSNs, into information

processing based and sleep-mode based methods. The sleep mode-based approaches conserve

energy by keeping as many nodes as possible for the longest time in the sleep mode. In the

information processing-based approaches, energy saving is achieved by means of reducing the

amount of data communicated in the network, by intermediate node-level processing.

The authors in [8] proposed an information fusion based approach for saving energy by node-

level processing of data to reduce the communication load, and this approach determines the

routing strategy. An approach based on collaborative routing is proposed by authors in [9]. The

authors show that this approach called CRAWL is adaptive to non-uniform distribution of

available energy in sensor networks. Collaborative and non-collaborative algorithms perform

equally when the available energy distribution is uniform, but when the distribution is non-

uniform collaborative algorithms is found to have 20.2% longer network life. For achieving

this, the authors in [9] propose different node scheduling options:

1. Initial network with all surviving nodes.

2. Uneven distribution of surviving sensor nodes.

3. More uniform distribution of surviving sensor nodes.

4. Optimal distribution for the last four surviving nodes for area coverage.

The wireless sensor network is fully effective when all of the sensor nodes are alive and they

cover the entire region of interest. CRAWL algorithm with Collaborative and non-collaborative

scheduling can increase WSN scalability and adaptability and was suggested by authors as the

next generation WSN energy management scheme.

A handoff algorithm for conservation of energy was proposed by authors in [10]. One of the

important characteristic of wireless systems is the signal variation caused by the movement of

the mobile stations. The existing radio link between a base station and the mobile station may

terminate if the radio link between the mobile station and another base station degrades due to

motion of the mobile terminal, and it is necessary to switch, or handoff, the communication link

from one base station to another. This can ensure the signal quality is maintained and the

Chapter 1

9

interference caused to other radio links is minimized leading to energy efficient management of

WSN nodes.

A frequency hopping Spread Spectrum (FHSS) technique was proposed by authors in [11] for

managing the energy in WSN, where the transmitter broadcasts on one frequency for a small

amount of time then switches to another frequency using a known switching algorithm called as

hopping or hopping pattern. The receiver knows the same hopping code so it is able to slide the

code past the incoming signal until it synchronies with the sender. Once they are synchronized,

the transmitter and receiver follow the hopping code to switch frequencies and communicate.

The resulting transmission is spread over a large frequency range and therefore appears as noise

to other receivers unless they know (or can decipher) the hopping code. Four different algorithms

were proposed to decipher the hopping codes:

1. The Brute- force method attempts to decode the signal by using every possible hopping

code.

2. Sequential scanning algorithm: Approach observes once frequency at a time to determine

the hopping sequences and was referred to as sequential scanning.

3. Parallel scanning algorithm: there is a receiver for each possible channel used in the

hopping code.

4. Hybrid algorithm: using concepts from the first three techniques with a set of parallel

receivers that switch through the possible channels.

Each algorithm was analysed theoretically and by simulation, for reduction of power

consumption, the authors in[11] showed that the results were positive in ability to decipher the

hopping codes.

The authors in [12] proposed a WSN scheme with an environment sensing/event detection focus

instead of energy management focus. Wireless sensor networks have a number of strengths such

as distribution, parallelism, redundancy, and comparatively high cost-effectiveness due to lack

of wires. On the other hand, their low cost, need to operate continuously, for a long term and

dependency on batteries, impose severe restrictions on the system. Hence, services provided in

sensor networks need to be lightweight in terms of memory and processing power and should not

require high communication costs. The authors in this work [12] proposed an algorithm in the

context of office monitoring system, which can distinguish abnormal office access pattern from

normal access, using an Adaptive Resonance Theory (ART) based anomaly detection technique.

Chapter 1

10

A resource reservation scheme for managing the energy in WSN was proposed by authors in

[13]. The scheme involving a hand off strategy for small size cells uses transfer probabilities to

predict the destination cell. Here, the reserved resources in each base station are proportional to

the user's transfer probabilities. In order to obtain accurate value of transfer probabilities, they

construct a movement or motion model to study the relationship between the user's initial states

and its transfer probabilities. According to authors, this algorithm turned out to be very easy to

implement and adaptable for different situations. It could offers accurate classification about the

user's random movement in small size cells and improves the efficiency when resources are

limited in wireless systems [13].

An approach, again with focus on environment sensing/event detection focus instead of energy

management focus was proposed by authors in [14]. Event detection is the process of observing

and evaluating an event using multiple sensor nodes without the help of a base station or other

means of central coordination and processing. In this work, authors propose a distributed event

detection approach based on distributed sampling of sensor nodes. It is a self-contained approach,

and it operates without a central component or base station canter, for coordination or processing,

and makes active use of the redundantly placed sensor nodes in the network to improve detection

accuracy.

Schurgers and Srivastava [15] propose an energy efficient routing scheme based on energy

histograms. The scheme involves aggregation of packet streams in a robust way (resulting in

energy reduction of a factor 2 to 3), and shaping of the traffic flow for uniform resource

utilization.

An approach based on opportunistic communication topology control is proposed by authors in

[16] to improve energy efficiency without sacrificing network performance. This technique

involves, wisely choosing a group of nodes to form a connected infrastructure, allowing other

nodes to directly connect to the infrastructure. The nodes that belong to the infrastructure are

called coordinator nodes. a non-coordinator node only turns on when it needs to connect the

infrastructure, and its energy can be significantly saved. A control algorithm for topology control

was developed to minimize the number of coordinator nodes to satisfy given end-to-end network

performance requirements from all the sensor nodes to the single sink [16].

Chapter 1

11

Ali and Uzmi in [17] proposed a scheme based on node address naming for energy efficient

management. The node address naming scheme assigns locally unique addresses without extra

overhead bits, and allows the reduction of address size by a factor of 3.6. Reducing the extra

overhead number of bits from each packet transmission ultimately leads to greater energy

efficiency and increases the lifetime of the network.

A Voronoi diagram based approach was proposed by the authors in [18] for energy management

in WSNs. In case of a network with a high density of sensor nodes, several problems may arise

such as the intersection of sensing area, redundant data, communication interference, and energy

waste. A high density network can introduce a fault-tolerant mechanism, increase precision and

provide multi-resolution data. The authors in this work developed a mechanism to control the

network density based on a criterion to decide which nodes should be turned off or on. Their

solution is based on the Voronoi Diagram, which decomposes the space into regions around each

node, to determine which sensor node should be turned off or on. Given the location of the nodes

and the area to be monitored, each node represents a point, and the desired area to monitor is the

polygon that is defined by the Voronoi diagram [18].

The authors in [19] proposed a power aware routing protocol for energy efficient WSNs, which

involves adapting the routes to available power. This allows a reduction in the total power used

as well as more even power usage across nodes. The authors included three major considerations

in developing this approach: The overall power dissipation, DSAP routing, and Power-DSAP

routing When the power considerations were added to the protocol, the overall power

consumption is much more balanced than without taking power into account.

A highly resilient, multipath routing scheme for energy management was proposed by authors in

[20]. In this work the authors proposed a novel braided multipath route to enable energy efficient

recovery from failure. In this approach, the authors propose localized algorithms to compute

approximations to the idealized disjoint and braided paths. Evaluation of two algorithms was

done using different failure modes: isolated node failures, where each individual node has an

independent probability of failure; and patterned failures, in which all nodes within a certain fixed

radius fail simultaneously. They further evaluate the performance of these approaches across

several parameters: density, probability of isolated failure, spatial separation of source and sink,

and frequency and radius of patterned failures. They found that, for comparable resilience to

patterned failures, braided multipath expends only 33% of the energy of disjoint paths for

Chapter 1

12

alternate path maintenance in some cases, and have a 50% higher resilience to isolated failures

[20].

An approach based on balanced cost cluster-heads selection protocol for reducing the power

consumption in WSNs is proposed in [21]. The protocol named LEACH (Low-Energy Adaptive

Clustering Hierarchy) is completely decentralized, and allows a best distribution of the

transmission energy in the network, and a large stable network lifetime [21].

The authors in [22] introduced a new energy efficient WSN algorithm called e3D (energy-

efficient Distributed Dynamic Diffusion routing algorithm), and compared it to two other similar

algorithms, namely directed, and random clustering communication. The authors take into

account the setup costs and analyse the energy-efficiency and the useful lifetime of the system.

In order to better understand the characteristics of each algorithm and how well e3D really

performs, they also compare e3D with its optimum counterpart and an optimum clustering

algorithm. The benefit of introducing these ideal algorithms is to show the upper bound on

performance at the cost of an astronomical prohibitive synchronization costs. They compare the

algorithms in terms of system lifetime, power dissipation distribution, cost of synchronization,

and simplicity of the algorithm. Their simulation results show that e3D performs comparable to

its optimal counterpart while having significantly less overhead. The proposed algorithm e3D

performed well in terms of achieving its goal to evenly distribute the power dissipation

throughout the network while not creating a very large burden for synchronization purpose [22].

The authors in [23] propose a WSN energy management scheme to conserve energy during

routing. Routing is a main energy demanding operation when nodes become ready for transfer

of data to the sink, an ample amount of research has been conducted to overcome routing energy

issues. However, Quality of Service (QOS) has a very important role especially in critical

applications such as defence, chemical and healthcare, where the accuracy and guaranteed timely

data transfer is an important issue. Hence, besides energy efficiency, QoS based routing is also

required to ensure best use of nodes. In this work, authors tried to focus on operational and

architectural challenges of handling QoS routing traffic in sensor networks and propose a new

protocol for QoS based routing, by applying different techniques simultaneously, and show a

significant improvement towards networks efficiency and QoS [23].

Chapter 1

13

A brief discussion on some of the earlier attempts on addressing the challenges of energy

efficiency done above, show that most of these earlier attempts for designing energy efficient

solutions revolved around classical approaches drawn from telecommunication theory and

communication protocols area. An efficient way to address this challenge, however, is to

combine some of these classical approaches with new developments in soft computing and

evidence based data driven approaches, and exploit the immense amount of data produced by

sensors, exploit the correlation and redundancy between the sensors in the network, and

understand the energy consumption within the sensor nodes and between the sensor nodes, or in

other words, learn the relationships from the data available in the network, and model the spatio-

temporal relationships by means of mathematical models. Learning from data with appropriate

mathematical modelling approaches, can inform the evolution of the measurements taken by

sensors over space and/or time. By building a mathematical model from data or true

measurements, with an appropriate algorithm, it is possible to obtain significant improvements

in prediction of events occurring in the WSN environment, and manage communication capacity

and energy efficiency within the wireless sensor network. These mathematical models that learn

from real data collected by the sensors in the environment, will provide flexible options to the

user, in terms of strategy for WSN setup, optimal selection of number of sensors and base station

nodes in the WSN, and their locations and their grouping into subnets or clusters, choice of

appropriate protocols for transmitting optimal information in the network, and monitor the

overall physical environment accurately. As this is a myriad set of requirements that need to be

satisfied, no single approach can address all these requirements, and though there was some work

done previously in using data driven approaches or machine learning based techniques to address

some of the above mentioned, they were mostly incoherent and often work well in isolation.

These previously proposed techniques mostly focus on the local, within the network protocols

for resource and capacity management and fall short of achieving higher level benefits, in terms

of overall event detection capability and energy efficiency for large dense networks deployed for

real world physical environments. Some of these earlier approaches proposed are reviewed in

detail in Chapter 2.

The construction of mathematical models from data needs to be automatic, since most of the

time, there is little or no information about the variations captured by sensor measurements.

Further, these mathematical models need to be simple and not computation intensive, sensor

nodes have limited computation capacity, and limited energy sources. This calls for some novel

strategies for WSN setup, selection of number of sensors and base station nodes in the WSN,

Chapter 1

14

their locations and their grouping into subnets or clusters, choice of appropriate protocols for

transmitting optimal information in the network, and monitor the overall physical environment

accurately. The hypothesis proposed here is that, it is possible to achieve many of these

objectives coherently, by exploiting the spatial and temporal relationships within the sensor data

that is large and continuously available within the WSN, discover the hidden relationships

between them, and identify the redundancies and correlations, to achieve the most important

objectives of energy efficiency and global event detection capabilities. This will address the

current gap that exists in this area, and is possible with some novel machine learning and data

mining based approaches, which can allow the modelling, prediction and evolution of future

measurements and states of wireless sensor network, and the detect the higher level information

that exists in the physical environment, based on the past measurements or data. For this purpose,

the research questions identified and contributions made for addressing these questions is

presented in next two Sections.

1.4 Research Questions

1. Whether it is possible to develop a strategy for jointly addressing the goals of energy

efficiency and event detection accuracy together?

2. Whether it is possible to develop an adaptive learning strategy to address the dynamic

changing requirements of WSN and address the challenges corresponding to energy

efficiency, event detection accuracy and QoS targets?

3. Whether it is possible to address develop a strategy for jointly addressing the goals of

energy efficiency, prediction accuracy and MAC layer routing issues together?

1.5 Thesis Contributions

To address various WSN challenges, a novel integrated framework for achieving energy

efficiency is proposed and consists of three stages as discussed below:

The first main contribution is the proposal of a joint energy efficiency–event detection

accuracy model, where a novel sensor node selection technique is designed, that

conserves the energy in the wireless sensor network, and at the same time maximizes the

event recognition performance. Here, the scheme utilises, fewer sensor nodes at a time,

Chapter 1

15

and placing unwanted sensor nodes in the sleep mode. For this, a novel objective

quantitative metric is proposed to assess the energy efficiency achieved, namely, the life

time extension factor (LTEF). We show that this joint scheme, allows selection of most

significant and influential sensor nodes for participation in different WSN tasks, and

contributes significantly towards energy savings and event detection accuracy.

As the WSN needs to adapt to the state of the environment being monitored dynamically,

the number of sensor nodes participating in the routing tree cannot remain fixed, and need

to adapt, in order to accurately monitor and predict the physical environment, and the

second contribution of this work is a proposal for adaptive models for sensor selection

and classifier learning for achieving energy efficiency and prediction accuracy, based on

performance targets specified. It turns out that this scheme which involves selection of

an appropriate classifier model, in conjunction with the previous sensor selection

approach, not only results in better prediction accuracy, but also contributes towards

quality of service (QoS) enhancements.

The third and the final contribution is a joint energy efficiency–adaptive routing model,

where an appropriate sensor selection and adaptive routing strategy can address the WSN

challenges corresponding to energy efficiency, prediction accuracy, and MAC layer

adaptation. We show that this joint model, also meet non-functional performance targets,

such as missing or faulty sensors, model building time, needed for adaptation of routing

protocol.

To summarise, this thesis attempts to address some of the important challenges in wireless sensor

networks for physical environment monitoring, such as the energy efficiency, the event

detection/monitoring accuracy, and quality of service aspects, based on evidence-based data

driven machine learning techniques. As can be seen in next Chapter (Chapter 2) on related work

and literature review, to the best of my knowledge, there are not many integrated and joint

approaches investigated in past, that can address multiple objectives of energy efficiency, event

detection accuracy, and quality of service aspects, simultaneously. Some of my attempts to

address these challenges have been published as peer reviewed contributions, and are outlined in

the next Section.

Chapter 1

16

1.6 Publications

The list of peer reviewed publications made during this thesis work is summarised below in

chronological order.

1. Alwadi, M.d. and G. Chetty, A novel feature selection scheme for energy efficient

wireless sensor networks Y.Xiang et al. (Eds): Proceedings ICA3PP 2012, Part II,

Springer LNCS 7440, pp. 264-273, 2012.

2. Moh'd ALWADI, and Girija CHETTY, “Feature Selection and Energy Management

for Wireless Sensor Networks”, IJCSNS International Journal of Computer Science and

Network Security, VOL.12 No.6, June 2012, 46 – 51.

3. Alwadi, Mohammad; Chetty, Girija, “Energy Efficiency Data Mining for Wireless

Sensor Networks Based on Random Forests”, International Journal on Data Mining and

Intelligent Information Technology Applications, 4.1 (Jun 2014): 1-8.

4. Alwadi, Mohammad; Chetty, Girija, “Energy Efficient Data Mining Scheme for Big

Data Biodiversity Environment”, Proceedings 2014 ASE Big Data/Social Comp/Cyber

Security Conference, Stanford University, 27th -31st May 2014. ISBN: 978-1-62561-

000-3. URI: http://www.ase360.org/handle/123456789/100 .

5. Mohammad Alwadi, Girija Chetty, “Energy Efficient Data Mining Scheme for High

Dimensional Data”, Procedia Computer Science 46(2015), 483-490.

doi:10.1016/j.procs.2015.02.047.

6. Alwadi, M. and G. Chetty, " Sensor Selection Scheme in Wireless Sensor

Networks: A New Routing Approach". pp. 73–79, 2015. © CS & IT-CSCP 2015.

7. Mohammad Alwadi and Girija Chetty, “Sensor Selection Scheme in Temperature

Wireless Sensor Network”, International Journal of Wireless and Mobile Networks,

ISSN: 0975-3834. June 2015, Vol 7, No. 3, pp. 47-53. DOI: 10.5121/ijwmn.2015.7304.

Chapter 1

17

8. Mohammad Alwadi and Girija Chetty, “A Novel Sensor Selection Scheme For Energy

Efficient Environment Monitoring of Wireless Sensor Networks”, Journal of Advances

in Computer Networks, ISSN: 1793-8244. (Accepted and in Press).

1.7 Organisation of Thesis

The rest of the thesis is organised as follows. Chapter 2 provides the related work and background

literature review on challenges associated with wireless sensor networks for physical

environment monitoring, and some of earlier research efforts using machine learning techniques,

and contributions made by the research community in addressing these challenges. Chapter 3

presents the first contribution of this thesis, and presents the joint energy efficiency and event

detection model, with discusses the development of an objective measure and sensor selection

scheme to assess the energy efficiency achieved. Chapter 4 discusses the problem of dynamic

behaviour of nature of wireless sensor networks and how the adaptive learning models based on

machine learning approaches can address this problem and maintain the prediction and

monitoring accuracy of the physical environment being monitored. The attempts to address the

challenges corresponding to MAC layer routing adaptation protocol is discussed in Chapter 5,

where a joint sensor selection – adaptive routing model to address the challenges energy

efficiency, prediction accuracy and adaptive routing under dynamic WSN changes is presented.

The thesis concludes with conclusions and further scope of this work in Chapter 6, with some

key references listed in Bibliography Section.

Chapter 2

19

Chapter 2 Related Work and Literature Review

In this Chapter, a review of machine learning approaches proposed in the literature to address the

design challenges in WSNs is presented. As can be seen in this Chapter, a myriad of attempts

have made so far, and many design challenges in wireless sensor networks have been resolved

using several machine learning methods. Utilizing machine learning based algorithms in WSNs

has to consider several constraints, such as limited resources of the network, and application that

requires different events to be monitored, and other operational and non-operational aspects.

2.1 Machine Learning Based Approaches

The recent advancements in machine learning and soft computing techniques allow better

prediction models to be developed based on a set of measurements. The learned model could

be just a simple parametric function, learned from data, a set of input variables - normally

historical measurements or observation, permitting output state or variable to be predicted

accurately.

As discussed before, a Wireless sensor network (WSN) can consist of heterogeneous, multiple

autonomous, tiny, low cost and low power sensor nodes. The purpose of these nodes is to gather

data about the physical environment being monitored, and collaborate with each other to

forward sensed data to centralized controller units called base station nodes or sink nodes for

further processing. The sensor nodes in the WSN could be heterogeneous, that is, they could

be equipped with various types of sensors, including thermal/temperature, acoustic, chemical,

pressure, weather, and optical sensors. Due to this heterogeneity, WSNs have tremendous data

diversity allowing powerful applications to be built, with different characteristics and

requirements. Developing efficient algorithms that are suitable for many different applications

is a challenging task. WSN designers have to address several issues pertaining to collection or

aggregation of data, and reliability of data, in addition to node clustering, energy aware routing,

events scheduling, fault detection and security.

In late 1950s, Machine learning (ML) was initially introduced as a special technique for

artificial intelligence (AI) [41]. It’s focused slowly shifted and evolved more towards

algorithms that are computationally feasible and forceful over the years. Its application grew

Chapter 2

20

extensively in last few years in several areas including bioinformatics, speech recognition,

spam detection, fraud detection and advertising networks. The machine learning tasks involved

were mainly that of classification, regression and density estimation, and involved algorithms

and techniques drawn from many diverse fields including statistics, mathematics, and

neuroscience and computer science. The essence of machine learning can be captured by

following two classical definitions:

o The learning processes for development of computer models that can enhance the

performance of systems and provide solutions to the problem of knowledge acquisition

[41].

o Detecting and describing consistencies and patterns in training data by employing

computational methods that can improve machine performance [42].

Machine learning technology appears very promising as per these definitions, to address

challenges in WSNs, as it allows exploiting historical data to improve the performance of

network on given task, or predict the future performance. For WSNs, using machine learning

technology can be immensely beneficial for a number of reasons, such as:

Better monitoring of dynamic environments that change rapidly over time. For instance,

in soil monitoring scenario, it is possible that the location of sensor nodes may change

due to soil erosion or ocean turbulence, and WSN based on machine learning can allow

automatic adaption and efficient operation in such dynamic environments.

Acquisition of new knowledge from unreachable, dangerous locations in exploratory

applications [43], volcanic eruptions, and early detection of tremors before earth quakes

for example. By detecting anomalies and unexpected behaviour patterns, a WSN that

can learn from data, can provide early warnings to catastrophic events well in advance

for emergency evacuations and calibrate and configure the WSN to collect additional

data from crucial nodes for better tracking of events.

Providing computationally feasible, low-complexity mathematical models for

complicated environments. For these environments, it is difficult to build accurate

Chapter 2

21

mathematical models, and difficult for sensor nodes to compute the algorithm

corresponding to these mathematical models. Under such circumstances, WSN based

on machine learning techniques can provide low complexity approximations for the

system models, allowing its implementation within sensor nodes. Routing problem is a

representative example here [44], [45].

Providing opportunity to extract spatial and temporal correlations between sensor nodes.

There is significant correlation in spatial and temporal dimension in the data that is being

collected in WSN. A WSN based on machine learning can leverage several algorithms

that operate on historical data being captured by sensors, and identify the correlations and

eliminate redundant sensors, localise the sensors at optimal locations or help in failure

recovery mechanism to be invoked in the event of breakdown in the network [46].

Increased automation and novel applications development, such as ubiquitous, ambient

computing systems. WSN based on machine learning can allow increase automation and

new uses by integration with other WSNs leading to fully sensored very large applications

such as Internet of Things technologies, Cyber-physical systems and machine-to-

machine communications. These applications use several different types of WSNs and if

based on machine learning, can support more intelligent decision-making and

autonomous control, with extraction of different levels of abstractions needed to perform

the AI tasks with limited human intervention [47], [48].

However, it is quite possible that WSN based on machine learning techniques may not lead to

any improvements if some of the issues outlined below are not considered during design stage.

As the WSN environment is a resource limited, significant energy is expended on

predicting the hypothesis with accuracy, and for global event detection type scenarios,

energy-efficiency and prediction accuracy is essentially a trade-off, [50].

Since the WSN becomes intelligent by learning from data, there is a need for large data

set. However, just the size of data being big does not ensure better learning or intended

generalization, and it is essential that it is right type of data, not just large data is used for

Chapter 2

22

building the mathematical model. Without right type of data, designer will not have full

control over knowledge discovery process [49], [50].

Recently, an increasing use of machine learning technologies in automation of WSNs operations

is being experienced. The authors in [51], present an excellent survey of machine learning

approaches applied to WSNs for processing the information in the network and improving the

performance. A similar survey, but more focussed towards ad-hoc networks, and how machine

learning techniques have been adopted in ad-hoc networks is presented by authors in [52].

Another seminal work on applications of three popular machine learning algorithms (i.e.,

reinforcement learning, neural networks and decision trees) at all communication layers in the

WSNs is presented in [53]. Some of the work also addressed specific challenges in WSNs, such

as authors in [54], [55], who developed an efficient outlier detection technique based on machine

learning concepts. Authors in [56] proposed an approach based on computational intelligence

technique for addressing challenges corresponding to data aggregation, routing, task scheduling

and optimal deployment and localisation. Computational intelligence techniques are a class of

machine learning techniques that focus on biologically inspired learning approaches such as

neural networks, fuzzy logic and evolutionary algorithms ]57].

Most of the earlier work on using machine learning techniques for WSNs, focussed on

reinforcement learning, neural networks and decision trees which were well established in their

reputation of being efficient at conceptual level and implementation level. Some of the machine

learning algorithms to address functional or operation challenges in WSNs such as routing,

localization, clustering, data aggregation, query processing and medium access control. The

operational or functional issues are those issues which are essential for the basic operation of

WSNs. Then, there are some approaches which have addressed the non-operational on non-

functional in WSNs, such as those that determine the quality or enhance the performance of

functional components, including security, quality of service (QOS) and data integrity.

We present a comprehensive review of some of the related approaches where machine learning

technology has been used for WSNs, which can also act as a design primer and comparative

guide.

Chapter 2

23

2.2 Related Work on Machine Learning for WSNs

In practice, the data science community depicts machine learning techniques as a collection of

algorithms and tools for creation of prediction models. However, machine learning researchers

recognize it as a rich field with very large goals and objectives. Appreciating such large goals

will be useful for designers who wish to apply machine learning to WSNs, which are very

complex in their own way as well. This understanding can provide better insight into

tremendous flexibility and benefit, machine learning algorithms can provide to a wide range of

complex WSN applications. For this, it is necessary to visit some of the theoretical concepts

that form the basis for machine learning technology in the context of WSNs.

Existing machine learning techniques can be categorized into supervised, unsupervised and

reinforcement learning techniques [58]. For supervised learning category, the learning

algorithm is provided with a labelled training data set. The system model is built by using the

labelled training data to make the machine learn the relation between the input, output and

system parameters. On the contrary, no labelled data is provided (there is no output vector) for

unsupervised learning algorithms, For an unsupervised learning algorithm, the relationship is

discovered in an unsupervised manner by clustering several sets of data into different groups

or clusters, and by discovering the similarity between input data samples. The third category is

a reinforcement learning algorithm, where the machine learns interactively, with online

learning from its environment. Finally, another way a machine can learn is a combination of

supervised and unsupervised learning style, and these are called hybrid algorithms or semi-

supervised learning approaches, and they try to inherit the strength of both supervised and

unsupervised learning approaches [59]. Further, a thorough discussion on theoretical concepts

of machine learning is presented in [60].

2.2.1 Supervised Machine Learning

For supervised machine learning, the system model is built with a labelled training set (known

outputs and predefined inputs). The learned relationship between the input, output and system

parameters is learned by the system model. This type of learning approach is extensively used

to solve several challenges in WSNs such as localization and objects targeting [61], [62] [63],

query processing and event detection [64], [65], [66], [67], medium access control [68], [69],

[70], intrusion detection and security [71], [72], [73], [74], data integrity, quality of service

Chapter 2

24

(QoS) and detection of faults [75], [76], [77]. Some well know supervised machine learning

algorithms are discussed next.

2.2.1.1 K-nearest neighbour (k-NN):

For this supervised learning algorithm, a test data sample is classified based on the labels (or

output values) of nearest data samples. By computing an average of readings within its

neighbourhood, the missing or unknown test sample measurement is predicted. Determination

of nearest set of nodes is done by using different methods. One of simplest method to determine

the neighbourhood is by using the Euclidean distance between different sensors [81]. As the

distance measure is computed using few local points, with k normally a small positive integer,

the k-NN approach does not need high computational power. Due to its simplicity, the k-NN

algorithm is suitable for query processing tasks in WSNs [64], [65].

2.2.1.2 Decision Trees

The decision tree classification involves predicting output labels by iterating the input data

using a learning tree [80]. During the iterative process, a comparison of feature properties

relative to decision conditions is done to reach a particular category. A significant amount of

research was done in using decision trees to address different design challenges in WSN, such

as identifying link reliability in WSNs using decision trees. Here use of decision trees provides

a simple method to identify critical features for link reliability, including loss rate, mean time

to failure (MTTF), and mean time to restore (MTTR). However, the limitation of decision trees

is that, it requires linearly separable data [80].

2.2.1.3 Neural Networks

Neural networks are one of the most popular learning algorithms for learning from data and

can be constructed by cascading chains of decision units, often called perception or radial basis

functions [49]. The cascading chains of decision units allow recognitions of non-linear and

complex relationships in data. However, the learning process with multiple cascading chains is

highly computations intensive [81]. An illustrative example of using neural networks for WSNs

is the sensor node localization problem, or determining the node’s geographical position in 3

dimensions. The sensor node’s geographical position has a complex, nonlinear relationship

with propagating angle, and distance measurements of the received signals from anchor nodes

Chapter 2

25

[82]. With supervised training of a neural network with different measurements in WSN as the

inputs, including RSSI (Received Signal Strength Indicator), TOA (Time of Arrival) and

TDOA (Time Difference Of Arrival), the network learns the relationship between RSSI, TOA,

TDOA and Node geometrical position, and can predict/estimate the 3 dimensional node

localisation co-ordinates. Figure 3 shows the schematic of this WSN node localisation

estimation using cascaded layers of neurons (computational units).

Figure 3 Estimating Node Localization co-ordinates in WSN Using Neural Networks [82]

There are several algorithms for training the network of neurons to learn the complex, nonlinear

relationship between the inputs and outputs, including Kohonen’s maps (self organising maps)

and LVQ (learning vector quantisation) [83]. One of the problems with most of the neural

network based estimation techniques is a significant amount of hand crafted feature

engineering required to do precise estimation. However, recently, some of the recent work on

deep learning architectures allows learning directly from high dimensional streaming big data

to learn the relationships between different variables without any feature engineering [84].

2.2.1.4 Support Vector Machines

Support Vector Machines provide alternatives to neural networks, and are preferred options for

solving nonconvex unconstrained optimization problems [79]. In the context of WSN, they

have been used for intrusion detection, or detecting malicious behaviour of sensor nodes,

security [73], [74], [86], [87], [88] and localisation [89], [90] [91]. With SVM, it is possible to

uncover the spatio-temporal correlations in data, as the algorithm involves constructing a set

Chapter 2

26

of hyperplanes (or optimizing a quadratic function with linear constraints) separating WSN

data measurements in feature space, by as wide as possible margins. Figure 4 shows the

schematic of SVM classifies WSN measurements.

Figure 4 Schematic of SVM Classification Process [91]

2.2.1.5 Bayesian Learners:

While most of the machines learning algorithms require large number of training samples to

learn, learning techniques based on Bayesian statistics require lesser training samples [92]. The

learning happens in Bayesian methods by adapting the probability distribution to efficiently learn

the uncertain labels. The important aspect for this learning technique is, it uses the current

knowledge (that the collected data samples (D)) to refine values of prior belief into posterior

belief values (Eq. 3.1).

𝑝(𝜃|𝐷) 𝛼 𝑝(𝜃) ∗ 𝑝(𝐷|𝜃) (3.1)

Where 𝑝(𝜃|𝐷) is the posterior probability of the parameter 𝜃, given observation D. And 𝑝(𝐷|𝜃)

is the prior likelihood of observation D, given the parameter 𝜃. In WSNs, this type of Bayesian

learners are useful for assessing event consistency (𝜃). using incomplete data sets (D) by

investigating prior knowledge about the environment. Several variations of Bayesian learners

allow better learning of relationships, such as Gaussian Mixture Models, Hidden Markov

Models, Conditional Random Fields, Dynamic Bayesian Networks [93].

Chapter 2

27

2.2.2 Unsupervised Machine Learning

For unsupervised learning there are no labels provided or there is no output vector. The sample

set is classified into different groups by investigating the similarity between them with an

unsupervised learning algorithm. This type of learning algorithm finds use in WSN node

clustering or data aggregation at a sink code scenarios [94], [95], [96], [97], [98], 99, [100].

With no labels provided, the unsupervised machine learning algorithm discover the hidden

relationships and is suitable for WSN problems, with complex relationships between variables.

Two most important type of algorithms in this category are K-means clustering [101], and

Principal component analysis [102], 103].

2.2.2.1 K-Means Clustering

This unsupervised learning algorithm classifies data into different clusters or classes and works

in sequential steps involving, random selection of k nodes as initial centroids for different

clusters, use of a distance function to label each node with the closest centroid, iteratively re-

compute the centroids using a predefined threshold value on current node memberships, and

stop the iterations if the convergence condition is met. The K-means clustering algorithm is

widely used in WSN sensor node clustering due to the simplicity and linear in its complexity

[101].

2.2.2.2 Principal Component Analysis

This unsupervised learning algorithm is quite popular in data compression field, and is used for

dimensionality reduction. It is a multivariate method and aims to extract important information

from data in terms of principal components, which is nothing but a set of new orthogonal

variables [102].

It is a multivariate method for data compression and dimensionality reduction that aims to extract

important information from data and present it as a set of new orthogonal variables called

principal components. These principal components are ordered such that the first principal

component is aligned towards highest-variance direction of data, with decreasing variance for

other components in order. This allows, the least variance components to be discarded as they

contain minimum information content, leading to dimensionality reduction. For WSN scenarios,

this can help in reduce the amount of data being transmitted among sensor nodes, by finding a

Chapter 2

28

small set of uncorrelated linear combination of original readings [103]. Further, it can solve the

big data problem into small data by allow selection of only significant principal components and

discarding other lower order insignificant components from the model. The details of PCA

theory, also known as eigenvalue/eigenvector or covariance matrix analysis is discussed

elsewhere [102], [103]. Figure 5 shows a simple two dimensional visualisation of the principal

component analysis (PCA) algorithm in dealing with high dimensional data.

Figure 5 Two dimensional visualization of PCA process [103]

2.2.3 Reinforcement Machine Learning

This type of learning algorithm for WSNs involves learning by interaction with the

environment. Here, a rewards process is involved, and a sensor node learns to take best actions

so that its long-term rewards get maximized with experience. Q-learning is most well-known

reinforcement learning algorithm, useful algorithm for WSN routing problems, where each

node seeks to choose actions that are expected to maximize its long term rewards. [104], [105,

[106], [107], [108]. Here, the sensor node in Q-learning regularly updates the rewards it

achieves based on the action it takes at a given state. The computation of future total reward

(also known as Q-value) of performing an action at a given state is obtained using Eqn. 3.2 as:

𝑄(𝑠𝑡+1, 𝛼𝑡+1) = 𝑄(𝑠𝑡, 𝛼𝑡) + 𝛾(𝑟( (𝑠𝑡, 𝛼𝑡) − 𝑄(𝑠𝑡, 𝛼𝑡)) (3.2)

Chapter 2

29

Where 𝑟( (𝑠𝑡, 𝛼𝑡) denotes the immediate reward of performing an action 𝛼𝑡 at a given state 𝑠𝑡,

and 𝛾 is the learning rate that determines how fast learning occurs (usually set to value between

0 and 1).

Figure 6 shown below illustrates how the sensor node can regularly update its achieved rewards

based on action taken at a given state.

Figure 6 visualization of Q-learning Algorithm [108]

2.3 Operational Challenges

There are several operational or functional challenges in design of WSNs, such as, power and

memory constraints of sensor nodes, topology changes, communication link failures, and

decentralized management. These operational challenges can be addressed by adopting

machine learning paradigms in the ways the WSNs work, so that they can be intelligent, and

can make conscious decisions for achieving energy efficiency, real-time adaptive routing,

query processing, global event detection, localization, node clustering and data

collection/aggregation at sink nodes.

2.3.1 WSN Routing Issues

As the sensor nodes have limited processing capabilities, small memory and low bandwidth,

design a routing protocol for WSNs has to consider various design challenges such as energy

consumption, fault tolerance, scalability and data coverage [46].

Formulation of a routing problem in wireless sensor networks, traditionally is done as a graph

problem, G = (V, E), where V represents the set of all nodes, and E represents the set of

bidirectional communication channels connecting the nodes. With this graph modelling

Chapter 2

30

approach, the routing problem can be described as the process of finding the minimum cost

path from the source vertex to all destination vertices, by using the available graph edges. We

call this path a spanning tree T = (V, E), whose vertices include the source or root node, and

destination nodes or leaf nodes. The solution to such a spanning tree problem with optimal data

aggregation is normally an NP-hard problem, even with the knowledge of full topology [45].

Learning from previous experiences is an important feature of machine learning, and sensor

networks can benefit immensely from machine learning, including selecting optimal routing

actions and adapt to the dynamic environment. Some of the benefits can be summarized as

follows:

Learn the optimal routing paths that can lead to energy efficiency and prolong the

lifetime of dynamically changing WSNs.

Divide the complex routing problem into simpler sub-routing problems, where the

nodes in the sub problem formulate the graph structures, by considering only their local

neighbours, and achieving low efficient and real time routing.

Use relatively simple computational methods and classifiers, and meet Quality of

Service (QoS) requirements in routing problems.

A simple sensor network routing problem using a graph and spanning tree routing algorithm,

is shown in Figure 7.

Figure 7 illustrates a simple sensor network routing problem using a graph, and the traditional

spanning tree routing algorithm, respectively. The network nodes have to exchange their

routing information with each other, to find the optimal routing paths. The illustration of how

machine learning reduces the complexity of a typical routing problem by considering

neighbouring nodes’ information to predict the full path quality. With a routing procedure

backed up with machine learning algorithm, each node will independently decide which

channels to assign, and how to optimise the transmission power. Such an approach will provide

near optimal routing decisions with very low computational complexity.

Chapter 2

31

Figure 7 Simplified Network Routing Based on Machine Learning [46]

A summary of different WSN routing protocols that have used machine learning based

approaches is given below.

2.3.1.1 Distributed Regression Approach:

A general framework for sensors data modelling was proposed by Guestrin et al. in [109]. In

this framework, the network nodes fit a global function to match their own measurements. A

kernel linear regression type of machine learning algorithm is run at the nodes. A set of kernel

functions map the training samples to learn the correlation between different features,

exploiting the fact that the readings of multiple sensors are highly correlated [111], [112]. Due

to the kernel mapping, the communication overhead in detecting the structure in the sensor data

is minimized. This approach contributes to developing a distributed learning framework for

wireless networks based on linear regression methods, the main advantage being good fitting

results, and the small overhead of the learning phase. However, the only disadvantage is that it

cannot learn non-linear and complex functions.

2.3.1.2 SOM (Self Organising Map) based data routing approach:

Using self organised map (SOM) based unsupervised machine learning approach was proposed

by Barbancho et al. in [110], and it involves detecting optimal routing paths as illustrated in

Figure 3.5. This approach is slightly different form the well know Dijkstra’s algorithm which

allows network backbone and shortest paths to be formed from base station to every node in

the network. In this approach, the second layer neurons compete with each other to reserve

high weights in the learning chain, during route learning, and the weights of the winning neuron

and its neighbouring neurons get updated further to match the input patterns. This learning

Chapter 2

32

phase being a highly computational process has to run at central base station node or sink node.

However, the execution phase is not computational and can be made to run on the network

nodes. This algorithm due to its hybrid nature (combination of the Dijkstra’s algorithm and

the SOM, the QoS(quality of service requirements, including latency, throughput, packet error

rate and duty cycle) during the process of updating neuron’s weights are taken into account.

However, some of the drawbacks of this algorithm are the complexity of the algorithm and

computational overhead in the learning phase due to change in network topology and settings.

2.3.1.3 Reinforcement learning based routing enhancement:

A reinforcement learning based algorithms, such as Q-learning algorithm, can enhance routing

protocol to guarantee reliable resource allocation. Sun et al. [105] have shown how a Q-learning

algorithm, called Q-MAP algorithm, can enhance multicast routing in wireless ad hoc

networks, where a node has to send the same messages to several receivers. For a mobile adhoc

network, consisting of heterogeneous nodes, with different nodes having different capabilities,

it is difficult to track the overall, dynamic information about the global state of the network

structure, and the Q-MAP multicast routing algorithm is designed to guarantee reliable resource

allocation for such complex scenario. The Q-MAP algorithm involves two phases, with first

phase as “Join Query Forward” that discovers an optimal route, and updating the Q-values.

In multicast routing, a node sends the same message to several receivers. Sun et al. [65]

demonstrated the use of Q-learning algorithm to enhance multicast routing in wireless ad hoc

networks. Basically, the Q-MAP multicast routing algorithm is designed to guarantee reliable

resource allocation. A mobile adhoc network may consist of heterogeneous nodes, where

different nodes have different capabilities. In addition, it is not feasible to maintain a global,

up-to-date knowledge about the whole network structure. The multicast routes are determined

in two phases. The first phase is “Join Query Forward” that discovers an optimal route, as well

as updates the Q-values (prediction of Q-values) of the Q-learning algorithm. In the second

phase called “Join Reply Forward”, an optimal path is created for facilitating multicast

transmissions. Hence, using a machine learning approach based on Q-learning can reduce the

overhead for route searching for multicast routing in mobile ad-hoc networks. However, this

path may not be energy efficient, and hence Q-MAP needs to be modified to make it an energy

efficient routing strategy.

Chapter 2

33

An alternate routing scheme in WSNs is based on UWB (Ultra Wide Band) communication. A

frequency band of 3.1 to 10.6 GHz (7,500 MHz of spectrum) has been dedicated by FCC

(Federal Communications Commission) towards the use of unlicensed UWB [103]. In UWB

technique, bulky data for short distances is transmitted using a wide spectrum of frequency

bands with relatively low power. An enhanced geographical routing approach with UWB

equipped sensor networks was proposed by Dong et al. [106]]. The authors in [105] used a

reinforcement learning algorithm to enhance geographical routing protocol (RLGR), where an

optimal route is computed by considering sensor node energy and delay as metrics for

formulating the learning reward function. The benefit of using reinforcement learning based

routing protocol is that it does not require information about global network structure to obtain

an optimal routing path. This routing protocol leverages the UWB technology for detecting the

nodes’ location, with UWB devices are placed on cluster heads only. Further, each node

maintains a simple look up table to keep the information about neighbouring nodes, and uses

the location and energy information of neighbouring nodes for network learning. The best

routing actions are learnt by exchanging short “hello” messages between these neighbouring

nodes.

Another enhanced geographic routing scheme based on reinforcement learning algorithm was

proposed by Arroyo-Valles in [108], called “Q-probabilistic Routing” (Q-PR), for WSNs that

can learn from previous routing decisions (for instance, selecting the routing path that has the

highest delivery rate over the past period of time). The difference between this protocol and

one previously discussed, RLGR, is the support for QoS. The Q-PR uses the message priory,

expected delivery rate and the power constraints to determine the optimal route, and uses a

learning model based on reinforcement learning and Bayesian decision models. Here, the

Bayesian method handles the decision to transmit the packets to a set of candidate neighbouring

nodes, by incorporating knowledge about data priority, profile of the nodes, reception energy,

and expected transmission rate. Further, it can discover the next hop online during the message

routing time.

Another enhanced reinforcement learning based WSN scheme was proposed by Forster and

Murphy [107], and it involves exchanging local information in nodes as a feedback response

to other nodes, called as “Feedback Routing for Optimizing Multiple Sinks (FROMS). This

routing algorithm allows efficient routing between multiple sources and multiple sinks, with

initialisation of Q-values based on the hop counts to every node in the network. The hop counts,

Chapter 2

34

which could be short “hello messages” are exchanged between the nodes at earlier stages of

the network deployment, and essentially extends the basic mechanism of RGLR in [106] with

an assumption that there is a direct communication between all the neighbouring nodes.

However, the main shortcoming of reinforcement learning based WSN routing algorithm is

their inability to look ahead or limited recognition of future knowledge. Hence they are unable

to perform in highly dynamic environments, as learning optimal routes in such environments

can take longer times.

2.3.2 Data Collection and Clustering Issues

It is inefficient to transmit all data to the sink directly for large scale energy-constrained sensor

networks, and as proposed by authors in [114], an alternate efficient approach is to pass the

data to an intermediate cluster head (also called as local data collectors), which collects data

from all the sensors within its cluster and forwards it to the sink node or the base station node.

Depending on how the cluster head selection or election is done, it is possible to achieve

significant energy savings. Due to this, several algorithms have been proposed for cluster head

selection/election to maximise the energy efficiency [115], [[116], [117]. A detailed taxonomy

and comparison of different clustering algorithms was done by authors in [118]. Figure 8 shows

an example WSN architecture with different clusters of nodes categorised as working, dead or

head cluster nodes.

Chapter 2

35

Figure 8 visualization of Q-learning Algorithm [118]

As discussed in [118], machine learning based approaches can improve the benefits of

clustering and data collection mechanism between nodes in WSNs in different ways, such as:

Identify non-functional nodes and remove them from routing schemes, using machine

learning algorithms, which can compress data locally at cluster heads, with

dimensionality reduction techniques, that extract similarity and dissimilarity in

different sensors’ readings.

Identify (select or elect) appropriate cluster head that can maximise energy efficiency

and increase lifetime of WSNs with an appropriate feature ranking and feature selection

approaches from machine learning field.

There are several solutions proposed to this end, for selecting the forming different clusters in

a large WSN, and choosing the cluster heads, and assigning nodes to each cluster and devising

a routing scheme from source nodes to sink nodes.

Chapter 2

36

2.3.2.1 Self-Managed Clustering Scheme

Hongmei et al. [119] suggested a scheme based on neural networks for self-managed clusters.

This clustering approach works well for large WSNs with short transmission distances, but for

large distance WSNs, the clustering efficiency is not significant in terms of energy efficiency

and quality of service.

2.3.2.2 LEACH Algorithm

A decision tree based machine learning algorithm was suggested by Ahmed et al. [120] for

solving the cluster head problem. In this work, the authors used several critical features in the

decision tree algorithm for learning the input vector iteratively, including, distance from nodes

to cluster centroids, battery energy level, the mobility degree, and vulnerability indicators.

They did a simulation study and showed its improved performance relatively when compared

to the “Low Energy Adaptive Clustering Hierarchy” or LEACH algorithm proposed by authors

in [127].

2.3.2.3 Gaussian Process Modeling

An approach based on Gaussian modelling of sensor data was proposed by several authors in

[121], [122], [93], and [94]. Gaussian models involve representation using random variables

(stochastic variables) that parameterize mean and covariance functions from the sensor data.

Ertin in [121] proposed an approach based on Gaussian process regression for initializing

probabilistic models. Whereas Kho et al. [122] extended this Gaussian regression approach for

adaptively sample sensor data depending on its importance. Authors in [122], proposed an

approach focussing on energy consumption, which provides a trade-off between optimal

solution and computational cost. In general, with smaller training data sets (less than few

thousand samples), Gaussian models are preferable for prediction of smooth functions [93].

However, they become computational intensive with large scale WSNs, and appropriate

strategies to deal with complexity need to be addressed by WSN designers.

2.3.2.4 CODA Algorithm

Another machine learning based architecture based on self organizing maps (SOM) for data

collection at cluster heads was proposed by Lee et al. in [94]. The SOM approach is an

unsupervised competitive learning technique for mapping high dimensional spaces to lower

Chapter 2

37

dimensions, and in this novel architecture, called, “Cluster based self organization and data

aggregation (CODA), and nodes can classify the collected data using a self-organising

algorithm. For a SOM algorithm, the winning neuron n* , has the weight vector w(t), close to

input vector x(t), at convergence of an optimization algorithm, defined as:

𝑛∗ = 𝑎𝑟𝑔 min𝑛

||𝑥𝑛 (𝑡) − 𝑤𝑛(𝑡)||, 𝑛 = 1, ⋯ , 𝑁 (Eqn.3.3)

Where N represents the number of neurons in the second layer. The updating of winning node

and its neighbours is done as follows:

𝑤𝑛(𝑡 + 1) = 𝑤𝑛(𝑡) + ℎ(𝑡)( 𝑥𝑛(𝑡) − 𝑤𝑛(𝑡)) (Eqn.3.4)

Here 𝑤(𝑡) and 𝑤(𝑡 + 1) represent the values of a neuron at time 𝑡 and 𝑡 + 1, respectively.

Here, ℎ(𝑡) is a Gaussian neighbourhood function defined as:

ℎ(𝑡) = 1

√2𝜋𝜎𝑒𝑥𝑝 (

‖𝑛∗−𝑛‖2

2𝜎2(𝑡)) (Eqn.3.5)

An improvement in energy efficiency and reduction in network traffic was observed by using

the CODA based machine learning approach for WSN.

2.3.2.5 ALVG Algorithm

While there is need for complete knowledge about the network topology in the most of the

methods discussed above, there are some algorithms, which are free from such restrictions. For

instance, one of the algorithm in such category is “Adaptive Learning Vector Quantization”

(ALVG) proposed by authors in [123]. This algorithm uses data correlation and historical

patterns to accurately retrieve compressed versions of reading from sensor nodes. ALVQ

algorithm uses well known learning vector quantization algorithm (LVQ), and uses the past

training samples to predict the code-book. The extension of LVQ to ALVQ enhances the

accuracy of recovering the original data from the compressed data, and reduces the bandwidth

required during transmission. This algorithm though has a capability to represent big size of

data with few vectors [83], it does not use isolated unused nodes in prediction, and hence it is

not robust against outliers.

Chapter 2

38

2.3.2.6 Dimensionality Reduction Techniques

Finally, there are few other algorithms proposed for data collection which reduce the bandwidth

during transmission using dimensionality reduction techniques, such as:

2.3.2.6.1 Compressive Sensing Approach

Compressive Sensing (CS) approach, replaces traditional schemes involving “sampling first

and then compression” to “sample while compressing” scheme. For compressive sensing

scheme, the sparsity feature of the signal is used to recover the original signal from few random

measurements, and is discussed in detail in [128].

2.3.2.6.2 EM Approach

Expectation Maximization (EM) approach, which is basically an iterative algorithm with two

main steps, an expectation (E) step and a maximization (M) step. The formulation of cost

function while setting the current expected value of system parameters happens in E-step, and

recomputation of system parameters, that minimizes the estimation error of the cost function

happens in M-step [129].

2.3.2.6.3 PCA Approach

Principal Component Analysis (PCA) technique, one of the most popular dimensionality

reduction techniques has also found its way in improving WSN performance. A method for

estimating distributed observations using few collected samples, based on PCA was proposed

by Masiero et al. [95], [96]. This technique uses PCA to produce orthogonal components which

is used by compressive sensing scheme to reconstruct original readings. As the PCA technique

here exploits spatial and temporal correlations, this method is independent of routing protocol.

A similar work by Rooshenas et al. [97], has proposed an approach to optimize the direct

transmission of readings to the base station or sink node, based on PCA technique. The use of

PCA here leads to considerable reduction in traffic by extracting fewer packets from combined

nodes’ collected data. The process of data reduction using PCA at intermediate nodes instead

of forwarding them all to destination sink, results in significant reduction in communication in

WSN, and hence makes the network energy efficient. Another approach involving use of PCA

for improving WSN performance was proposed by Macua et al in [98]. This approached uses

Chapter 2

39

a distributed consensus-based method for dimensionality reduction, and uses a combination of

PCA and maximum likelihood measure of the data observed. The two variations of this method

called, consensus based distributed PCA (CB-PCA), extracts the eigenvectors of local

covariance matrices, whereas, the consensus based EM distributed PCA (CB-EM-PCA) uses a

distributed EM algorithm. Both the variations. Use a consensus algorithm proposed in [130],

to predict the probability distribution of the data, and compute the global dominant

eigenvectors based on single hop (local) communication parameters. It is possible to achieve a

trade-off between the dimensionality reduction and the communication costs, by tuning the

consensus round parameter both in CB-DPCA and CB-EM-DPCA algorithms. This implies

increasing the consensus rounds for improvement in algorithm accuracy, but at the cost of

increased computation requirements.

2.3.2.6.4 Distributed PCA

Another recent contribution on using PCA for improving WSN performance is by Fenxiong et

al. [124], who have addressed the problem of data reduction in WSN by transforming the data

from a high dimensional space to a lower one using PCA technique. Here, the data which is

continuously collected over time by each node is sent to its corresponding cluster head, and the

cluster head, the data redundancy is eliminated by compressing the data matrix, by using PCA,

and ignoring the least significant components. By choosing the number of PCA components

appropriately, it is possible to achieve a trade-off between the computational cost and

compression accuracy in WSN.

2.3.2.7 Collaborative Mobile Node Processing

Collaborative mobile node processing with machine learning approach, proposed by authors in

[99], [100], involves use of mobile nodes in the WSN architecture, unlike the fixed location of

WSN layout discussed so far. The use of mobile nodes is particularly needed for collecting

massive data from surveillance camera networks. Here, the power mobile sensor nodes are

deployed along with traditional surveillance cameras, to enhance the intelligence gathering

capabilities of integrated mobile surveillance wireless sensor network systems [99]. Here the

mobile sensor nodes are grouped into several clusters using k-means unsupervised learning

algorithm, with each cluster monitored by single mobile sensor. However, the clustering of

Chapter 2

40

sites with k-means algorithm though simple and straightforward in the implementation with

low complexity, it is sensitive to outliers and selection of initial seed values.

2.3.2.8 Role-free clustering

Role-free clustering approach was proposed by Forster and Murphy in [125], where a Q-

learning technique is used for WSN cluster formulation. In this approach, labelled CLIQUE

method, instead of using an election or selection criteria, it uses a rewards criteria to assign a

node as a cluster-head node. A combination of Q-learning algorithm in combination with

certain dynamic network parameters such as energy levels is used for this method.

2.3.2.9 Decentralised Learning

Reduced data latency using decentralised learning approach is another interesting approach

proposed by Mihaylov et al. [126] to address the problem of data latency that can creep in for

WSNs based on Q-learning with random topology sensor network setups. Here learning

happens in a decentralised manner locally in the cluster head nodes to optimise the data

aggregation, instead of central control/base station node. Due to this, the efficiency of the whole

WSN is improved with smaller learning transmission overheads. Due to the savings in the node

energy budget during data collection process, the lifetime of the network is extended.

2.3.3 Event Recognition & Query Processing Issues:

In addition to routing and node clustering issues mentioned in the previous two sections, the

event recognition and query processing are also important operational requirements of large

scale WSNs. The functionality needed here, is a trustworthy event scheduling and recognition

with minimal human intervention. In general, WSN monitoring can be classified as event-

driven, continuous or query-driven [46]. With machine learning based event monitoring

approach, it is possible to obtain efficient event detection and query processing solutions under

constrained environment with restricted query areas. Figure 9 illustrates different query

processing and event detection operations in WSNs.

Chapter 2

41

Figure 9 Event Detection and Query Processing Using Machine Learning [46]

Adopting machine learning based techniques for these operations can lead to several benefits

including:

Facilitate the development of efficient event detection techniques using learning

algorithms and simple classifiers, particularly with limited availability of storage and

computing resources.

Facilitate the development of effective query processing techniques for WSNs, for

instance, determine the search regions whenever a query is coming from, and localise

the communication efforts there, instead of flooding the whole network.

Several research works have focussed on efficient design of good event detection and query

processing strategies for WSNs. Some of the simpler approaches involve defining a strict

threshold value for phenomenon being sensed and triggering the alarms in case of any violations,

while recent WSN set ups use more complex approaches than using simple threshold values. The

complex, emerging approaches used advanced machine learning based casting of problem for

event detection and query processing. Some of them are discussed next.

2.3.3.1 Bayesian Event Detection Algorithm

Using decentralised Bayesian learning, Krishnamachari and Iyengar [131] investigated the use

of WSNs for detecting environmental phenomenon, and obtained a fault detection accuracy of

Chapter 2

42

up to 95%, with simple threshold criteria. With a decentralised learning approach, they could

isolate the faulty region and focus the query processing in that region leading to better event

recognition accuracy. A follow-up approach was propose by Chen et al [134], which addressed

some of errors in the formulation of distributed learning algorithm problem in [131], leading

to enhanced performance calculations.

2.3.3.2 HMM-Bayes Activity Event Recognition

The work proposed by Zappi et al. [132], involved extension of event recognition and query

processing from WSN area to activity recognition. Here, the authors presented a real-time

approach for activity recognition using WSNs that accurately detects body gesture and motion.

The WSN nodes were initially spread throughout the body, and could detect the organ motion

through accelerometer sensors, measuring three axis measurements (positive, negative and

null). These measurements were then used to build a machine learning model such as Hidden

Markov Model (HMM) to predict the activity at each sensor. The prediction accuracy depends

on selecting appropriate sensors that can provide the most informative description of the

gesture. A final gesture decision is obtained by using a naïve Bayes classifier, which combines

the independent node predictions and maximises the Bayes posterior probability. The

architecture of this system is shown in Figure 10.

Figure 10 HMM and Naïve Bayes Event Detection and Query Processing [132]

Chapter 2

43

2.3.3.3 Neural Networks for Forest Fire Event Recognition

The authors in [135] presented an approach for fire detection and rescue system using WSNs,

where they have shown that better forest fire detection performance can be achieved with use

of WSNs instead of using satellite based solutions, while costing much less. Further, a real time

forest detection scheme based on neural network classifiers was proposed in [66], where, the

distributed processing scheme, with data processing at cluster heads, and important data gets

communicated and collected at the central station for final decision making. The system

however is complex to interpret, specially under real time detection environments, and needs

better strategies for data processing, communication and collection for final decision making

than what has been proposed here.

2.3.3.4 K-Nearest Neighbourhood for Query Processing:

One of the simple but highly effective query processing technique in WSNs is K-nearest

neighbour method for query processing in WSNs. An in-network query processing solution

using the k-nearest neighbour algorithm, called the k-NN boundary tree or KBT algorithm was

proposed by Winter et al. in [64], where, each node, aware of its location can determine its k-

NN search region whenever a query arrives from the application manager. An extension of

KBT query processing approach to 3D space was proposed by Jayaraman et al. [65], called the

“3D-KNN” processing scheme for WSNs, where the query region is restricted to bound at least

k-nearest nodes in 3D space. Further, the SNR (signal to noise ratio) and distance

measurements are used to refine the k-nearest neighbour. One of the main real time constraints

of using such machine learning approaches for query processing, including k-NN based

algorithms is a need for large memory footprint of WSN nodes to store every collected sample,

and high latency, or processing delay in large sensor networks in communication of k-NN

classifier outputs from cluster heads to sink or base station control nodes for final decision

making.

2.3.3.5 Decision Trees for Distributed Event Detection For Disaster Management:

Bahrepour et al. [67] developed a decision tree based event detection and recognition approach

using WSNs for disaster prevention systems. It uses a decentralised mechanism, with its main

application as the fire detection in residential areas. Here the final decision on event detection

is made by using a simple voting scheme from highest reputation nodes.

Chapter 2

44

2.3.3.6 Principal Component Analysis (PCA) for Query Optimization:

An optimized query processing approach using WSN data attributes and PCA was proposed

by Malik et al. [133], where the PCA can dynamically detect dominant principal components

(i.e. important WSN data attributes) from the correlated data set.

The four stage workflow of fundamental steps involved in this algorithm is shown in Fig. 3.8.

In stage 1, an SQL request, containing the human friendly and intelligible attributes is to

DBMS. At the DBMS, this original query is optimized by using only high variance components

of PCA algorithm output extracted from historical data in stage 2. In stage 3 and 4, this

optimized query is transmitted to WSN nodes to extract the data from individual sensor nodes.

The original attributes are then reconstructed from the optimized attributes by reversing the

PCA process. The authors in [133] have shown how this four stage query optimization process

with PCA can result in around 25% energy savings in the WSN nodes at 93% event recognition

accuracy. However, this enhancement does not fully exploit the abundant data effectively, i.e.,

it collected large data at the sensor nodes in the first place, it doesn’t use it fully. So, as such

the process is not cost effective. Therefore, for the applications with high accuracy and

precision requirements, this solution may not be ideal.

2.3.4 Challenges Related to Localisation and Object Targeting

The process of determining the geographic coordinates of network’s nodes is called

localisation, and location awareness of sensor nodes in WSNs is an important capability, since

most of WSN operations are based on the location [136]. Use of GPS hardware in each node

of WSN though can provide location awareness, it not feasible cost wise. Further, GPS services

may not be available in observed remote and certain indoor locations. For such use case

scenarios, relative location measurement may be sufficient, and by using absolute location

measurements for a small group of nodes, relative locations for other nodes can be converted

into absolute location measurements [137].

Moreover, GPS service may not be available in the observed environment (e.g., indoor).

Further, by using proximity based localization, additional measurements relying on distance,

angle or a hybrid of them can be used to enhance the performance of proximity based

localization. These distance measurements can be calculated by different approaches including

RSSI (received strength signal indication), TOA (time of arrival), and TDOA (time difference

Chapter 2

45

of arrival). Also, certain angular measurements can be obtained by using compasses or special

smart antennas [138]. The authors in [82], provide more details about different range based

localization techniques. Sometimes, sensor nodes can encounter changes in their location after

WSN deployment, perhaps, due to the movement of nodes. Use of machine learning techniques

can aid WSN node localisation process in different ways, such as:

Conversion of relative locations of nodes to absolute ones using few anchor beacon or

beacon nodes, eliminating the need for range measurement hardware to obtain distance

estimations.

Machine learning techniques can be used in surveillance and object targeting systems,

to divide the monitored sites into a number of clusters, where each cluster represents

specific location indicator.

Figure 11 Localization Using Beacon Nodes in WSN [82]

Figure 11 shows the layout of beacon nodes (anchor nodes) and unknown node (a node which

cannot determine its location). The beacon nodes can determine their location due to

Chapter 2

46

positioning hardware it consists, and this location serves as a reference point to estimate the

co-ordinates of other unknown nodes.

Some important approaches proposed by researchers for WSN localisation using machine

learning approaches can be described as follows.

2.3.4.1 WSN node localisation using Bayesian approach

WSN node localisation scheme based on Bayesian approach with very few anchor points

(beacon nodes) was proposed by Morelande et al. in [61]. The approach involves extension of

progressive correction technique, proposed in [149], where the predictive samples from

likelihoods get closer to the posterior likelihood. This algorithm works well in localisation in

both small and large WSNs, with few thousands of nodes, as the Bayesian algorithm can

gracefully handle incomplete data sets due to its capability to learn from priors (previous data)

and probabilities.

2.3.4.2 Location Aware Bayesian approach for Activity Recognition

The problem of both WSN sensor and activity localization in smart homes was proposed by Lu

and Fu [62], where the activities of interest including use of phone, listening to musing, using

the refrigerator, studying were detected. The authors reiterated, that in such applications,

designers need to take into consideration both human and environmental constraints, and their

framework named “Ambient Intelligent Compliant Object detects the human interaction with

the home electric devices in a more intelligent manner. This is done using several naive Bayes

classifiers to determine the resident’s current location and evaluate the reliability of the system

by detecting any sensors that didn’t work. This turns out to be a simple and robust mechanism

for localization, though with certain constraints in terms of scope of ambient environment

limited to predefined activities only. If there is a deviation in activities, the location awareness

and the activity detection does work well. To overcome this limitation in this centralized

system, there is a need for less engineering of features, with unsupervised feature learning

techniques, such as those proposed in [49], [150].

Chapter 2

47

2.3.4.3 Neural Network based WSN Localisation Approach

Using different neural networks Shareef et al [63] developed a localisation scheme for WSNs.

By using a combination of MLP (multi-layer perceptron), RBF (radial basis network), and

RNN (recurrent neural network), the authors show that RBF network results in the minimum

error at the cost of high resource requirements, whereas, MLP allows minimization of

computational and memory or storage resources.

A slightly different approach was proposed by Yun et al. [139], where two different processing

modules were used along with RSSI information from anchor/beacon nodes for localisation.

The first processing module uses a combination of fuzzy logic and genetic algorithm system,

whereas for the second processing module, and adaptive neural network that uses the RSSI

measurements from all anchor/beacon nodes as an input vector, to predict the sensor location

is used. A similar approach for WSN localisation with RSSI from anchor nodes as an input to

a set of neural networks was proposed by Chagas et al. in [140]. The advantages of these

multiple NN based localisation algorithms with RSSI information, is their capability to use the

location coordinates in terms of 3D space coordinates (continuous valued vectors). However,

the weakness of neural network based classifiers as compared to Bayesian or statistical

classifiers is their inability to work under uncertainty, as most of the neural networks that have

been used here, are non-probabilistic approaches. Hence prediction estimates cannot exploit

the prior knowledge effectively, leading to increase in localisation errors.

2.3.4.4 Support Vector Machine (SVM) based WSN Localisation Approach

For those scenarios where sensors cannot be equipped with self positioning devices, SVM

based WSN localisation approach was used. To this end, Yang et al. [91] developed a mobile

node localization scheme by employing SVM and connectivity information capabilities. The

algorithm first detects the node movement using the RSSI metric, and SVM estimates the new

location in the second step. A similar approach was proposed by Tran and Nguyen in [90],

called “LSVM” approach for node localization in WSNs. Here LSVM adopts several decision

metrics, including connectivity information and RSSI indicators, and offers a fast and an

effective localisation, it does suffer from sensitivity to outliers in training samples, causing

reduced performance with many outlier samples.

Chapter 2

48

2.3.4.5 Light Weight Support Vector Regression (LWSVR) based Localisation

Kim et al. [89], proposed a light weight implementation of SVR approach, due to problems

with adoption of normal SVR approach, due to limited processing resources in WSN nodes and

high dimensionality of incoming data. In this approach, the original regression problem is

divided into several sub-problems, and algorithm works on several subnetworks with smaller

data processing with each regression algorithm, which they call it as sub-predictors. Then using

a custom ensemble combination technique, the sub-predictor models that were learnt, are

combines together, to predict overall network estimates, with better performance, including

low computational requirements, robustness against noisy data, and convergence to the

preferred solution with low computational requirement.

2.3.4.6 Localisation using Decision Trees

A different application with WSNs, involving acoustic target localisation for WSNs based on

decision tree learning was proposed by Merhi et al. [141], where the exact locations of targets

are determined using time difference of arrival (TDOA) metric and a spatial correlation

decision tree. Also, in this work an EB-MAC protocol (Event Based Medium Access Control)

that allows event-based localization and targeting in acoustic WSNs. This framework was

implemented using MicaZ sensor boards that support ZigBee 802.15.4 specification for

personal area networks. As using GPS functionality in underwater WSN’s applications may

not be feasible, due to limited propagation capability of GPS signal through water [151],

another approach was proposed by Erdal et al [142] for submarine detection in underwater

surveillance systems, a randomly deployed node can find its location in 3D space using beacon

node co-ordinates. Here, a sensor is fixed with a cable to a surface buoy in each monitoring

unit, and data is collected using the buoys and transmitted to central controller and processing

unit. The central unit consists of a decision tree classifier, which can detect any submarines in

the monitored sites.

2.3.4.7 Localisation using Gaussian Processes

For a WSN temperature monitoring system, an optimized solution to sensor placement based

on spatially correlated data was proposed by Krause et al. [143] Here, the authors developed a

lazy learning scheme based on Gaussian process model, which involves storing training

samples, and delay the major processing task until a classification request has arrived. When

Chapter 2

49

choosing optimal locations for sensors, this solution aims to achieve robustness against node

failures and model ambiguity. In another work, Gu and Hu [144] developed an approach based

on spatial Gaussian process regression, for a distributed protocol for collective node motion, A

distributed Gaussian process regression (DGPR) was used to predict optimal location for

mobile nodes’ movements. Further, it uses a sparse version of Gaussian process regression

algorithm to reduce such computational complexity, as compared to traditional Gaussian

process regression (GPR) algorithm, which has a computational complexity of O(N3), where

N is the size of the samples. Using only spatiotemporal information from local neighbours,

each node executes the regression algorithm independently.

2.3.4.8 Localisation using Self Organising Map (SOM):

Paladina et al. [105] proposed the SOM based localisation solution for WSNs consisting of

thousands of nodes. In each WSN node, SOM algorithm is implemented with 2 neurons of the

output layer connected to the 3x3 input layer. The input layer is constructed using spatial

coordinates of 8 anchor nodes surrounding the unknown node. In the output layer, the unknown

node’s 2D spatial co-ordinates evolve, after sufficient training. However, the shortcoming of

this scheme, since it uses its neighbouring nodes, the algorithm expects that the nodes should

be distributed uniformly and equally spaced throughout area that is being monitored. While

most of the traditional methods use absolute locations of a few nodes to find the positions of

the unknown nodes, Giorgetti et al. [146] proposed a localisation algorithm that uses only the

connectivity information and SOM algorithm. Since this method does not require a GPS

enabled device, this method is highly suitable for networks with limited resources. However,

it suffers from latency issues, as this algorithm is implemented in a centralised manner, with

each node transmitting its neighbouring node information to the central control station node

for calculating the adjacency matrix and hence the node’s location. Another algorithm proposed

by Hu and Lee [147], proposed a scheme that does not require anchor nodes, for node

localization service in WSNs. The difference between [147] as compared to [146] is that the

algorithm in [147] is distributed and eliminates the needs for a central unit, and by distributing

the computation tasks to all nodes in the network, eliminates the need for a central unit, and

minimizes the transmission overhead of the algorithm.

Chapter 2

50

2.3.4.9 Reinforcement Learning based Localisation:

A reinforcement learning-based localization scheme for WSNs based on Q-learning was

developed by Li et al. [148], which allows real-time management of the mobile beacons. In

this method called “Dynamic Path determination of Mobile Beacons” (DPMB), the mobile

beacon (MB), is aware of the physical location during its movement, and used to determine the

positions of large number of sensor nodes. Here the different positions of the MB are

determined from different states of Q-learning algorithm, and due to its mobility, the algorithm

can cover all the sensors in the monitored area, with location update message from MB at

different times. This style of mobile beacon functioning can save the resources of the unknown

nodes, as the entire operation is run on mobile devices. However, being centralised, there could

be malfunctioning mobile beacons, and could lead to entire system failure.

2.3.5 Medium Access Control (MAC) Issues:

There are several challenges in the design of MAC protocols for WSNs, such as, the energy

consumption, latency, prediction accuracy etc., in addition to basic operational feature, that a

number of different sensors cooperate to efficiently transfer data [152]. Therefore, the MAC

protocols have to be designed appropriately, to allow efficient data transmission and reception

of the sensor nodes. The authors in [153] have provided a comprehensive survey of MAC

protocols in WSNs. Recently, few machine learning methods have also been proposed for

designing appropriate MAC protocols and enhancing the performance of WSNs. In these

works, machine learning plays a role in a variety of ways, including:

Using the transmission history of the network to adaptively determine the duty cycle of

a node. Here, the assumption is, that the nodes, which are able to predict when the other

nodes’ transmissions will finish, can sleep in the meantime and wake up (to transmit

data) just when the channel is expected to be idle, and no other node is transmitting.

Using the concepts of secured data transmission along with machine learning in

designing the MAC layer protocol. Such a secure MAC layer scheme would be

independent of the proposed application and can learn sporadic attack patterns,

iteratively.

A brief description of how the WSN protocol design issues were addressed by machine

learning, and other related approaches is discussed next.

Chapter 2

51

2.3.5.1 MAC design using Bayesian Statistical Models

A contention-based MAC protocol for managing active and sleep times in WSNs was proposed

by Kim and Park [68]. By using a Bayesian statistical model to learn when the channel can be

allocated, it reduces the need for continuous sensing of medium, and hence save energy. Some

of extensions of this scheme, target the CSMA contention based protocols, and are proposed

as “S-MAC” (Sensor MAC) and “T-MAC” (Timeout MAC”) by authors in [156], [157].

2.3.5.2 MAC design using Neural Network Models

One of the popular medium access protocols in traditional computer networking is TDMA or

time division multiple access protocols, which employ periodic time slots to separate medium

access of different machines, and uses a central server unit to broadcast a transmission schedule

in case of change in topology of the network. This can adopted for WSN scenario, and Shen

and Wang [69] proposed a MAC protocol, which involves broadcasting of the transmission

schedule in TDMA using a fuzzy Hopfield neural network (FHNN) approach. To prevent any

potential transmission collisions and latency issues, the authors propose distribution of time-

slots among different nodes in the network. Another similar approach was proposed by

Kulkarni and Venayagamoorthy [70], which includes security aspects in addition to MAC

issues in WSN protocol. Their CSMA-based MAC approach, can prevent denial-of-service

(DoS) attacks in WSNs, and uses a neural network learning to prevent flooding the WSN with

fake and mendacious data by learning the network properties and variations such as packet

request rate and average packet waiting time. Denial of service attack or DoS attack that

generates large useless data and floods the network, and prevents the delivery of useful data,

and it is much easier to attack WSNs with DoS attacks, as the attacker tries to exploit the

vulnerability of WSNs in terms of limited buffering and storage capacity and limited bandwidth

capabilities. With neural network based MAC protocol, if the neural network exceeds a

predefined threshold level, the MAC layer will be blocked. Further, blocking does not impact

the functioning of the whole network, as the scheme is implemented in a distributed manner,

and only affected site is blocked.

2.3.5.3 MAC design using Reinforcement Learning Models

Use of reinforcement learning based techniques for medium access control (MAC) was

proposed by Liu and Elhanany in [154], called RL-MAC protocol. The adaptive RL-MAC

Chapter 2

52

protocol for WSNs, optimizes the duty cycle of the network node for reduce energy

consumption and increased throughput. RL-MAC works in a similar manner as S-MAC [156]

and T-MAC [157], and synchronises node’s transmission on a common schedule in a frame-

based structure. By using the traffic load and channel bandwidth, the RL-MAC adaptively

determines the slot length, duty cycle and transmission active time. Another proposal by Chu

et al. in [152], proposed a combination of slotted ALOHA and Q-learning algorithm to

introduce a new MAC protocol for WSNs, called ALOHA-QIR, the ALOHA and Q-Learning

based MAC with Informed Receiving. By using the best features of both ALOHA and Q-

Learning, it provides benefits in terms of simple design, low-resource requirements and low-

collision probability. The method works by nodes broadcasting their future transmission

allocation, in their transmission frames, so that nodes can be put in sleep mode. The willingness

to research a slot is represented by Q-value map in each node, where the node with higher Q-

value will attain the right of slot allocation and hence transmission of its own data. An

illustration of steps involved in updating the Q-values over three frames for a node that is

allowed to transmit a maximum of two packets in each frame is shown in Figure 12. The Q-

learning based medium access control can suffer from high collision rates in the initial

exploration phases, though it is appealing due to its distributed mode of operation, a small

storage and computational resource requirement.

Figure 12 ALOHA-QIR Scheme For MAC Layer in WSN [152]

2.3.5.4 MAC design using Adaptive Decision Trees

For modern application scenarios, such as in healthcare and assisted living systems, design of

MAC layer in WSN is quite challenging, particularly to address the dynamic communication

patterns and service requirements over time, and the data in WSNs, need to directly share the

collected data with the users’ mobile phone or smart phone. To this end Sha et al. in [155]

proposed a “Self Adapting MAC layer” (SAML) design, consisting of two components, the

Chapter 2

53

RMA component and the MAC engine component. The RMA or reconfigurable MAC

architecture allows chooses different MAC protocols, and MAC engine, allows learning the

chosen MAC protocol from the current network data. For learning, the MAC engine uses a

decision tree classifier, and uses several features for learning, including, IPI (Interpacket

interval), RSSI (Received Signal Strength Indicator), the application QoS requirements

(reliability, energy usage and latency), statistical parameters (mean and variance), traffic

pattern, and PDR (packet delivery rate). Figure 13 shows the design of SAML protocol.

Figure 13 Adaptive Decision Tree Based MAC Protocol (SAML) [155]

2.4 Non-operational Aspects of WSN

While the operational challenges are directly related to the basic operational or functional

behaviour of the systems with WSN, the non-operational aspects are not related to basic

operational needs of the system, and though non-functional, are highly desirable, performance

enhancing requirements that can used by vendors for differentiating and achieving competitive

edge in the market. Some of the performance enhancing requirements could include updates

and analytics on the environment being monitored by WSNs, QoS (quality of service), security,

and data integrity. Recent advances in machine learning techniques can be harnessed to address

the non-operational aspects, and enhance the WSN performance. Some of the work reported in

this area is discussed next.

Chapter 2

54

2.4.1 Security and Anomaly Intrusion Detection

Due to limited resource requirements, security and intrusion management techniques are

challenging to implement in WSNs [54]. Some of the methods based on machine learning,

proposed for intrusion detection involve introduction of anomaly, or unexpected, misleading

observations to the network, emulating an attack scenario. A brief schematic of general concept

of anomaly detection in monitoring the WSN system is shown in Figure 14.

Figure 14 Basic Concepts of Anomaly Intrusion Detection [54]

Here, the data is classified into two classes corresponding to most observations that may belong

to these two regions, but the measurements that are inconsistent and unusual due to suspected

attacks are considered as anomalies or intrusions. Detection of outliers and misleading

measurements can be done by different machine learning algorithms, including supervised,

unsupervised and reinforcement learning algorithms, and by analysing well known malicious

activities and vulnerabilities, several attacks and intrusions can be detected. Such WSN security

enhancements by adopting machine learning techniques can lead to several benefits, including:

Preventing the transmission of anomalous and suspicious data, by detecting outliers,

save WSN node energy, and significantly expand WSN lifetime.

Eliminating faulty and malicious readings, and avoiding the discovery of unexpected

information impacting on the critical actions, so as to enhance the WSN reliability.

Chapter 2

55

Prevention of malicious attacks and vulnerabilities, by automatic online learning and

prevention of malicious attacks and vulnerabilities.

Some of the approaches based on machine learning, addressing the security issue in WSNs, is

presented next.

2.4.1.1 Outlier detection

An outlier detection scheme based on Bayesian belief networks (BBM) is proposed by

Janakiram et al. [71]. In this scheme, first, the conditional relationships between the nodes’

readings are modelled, since most the nodes’ neighbours have similar readings due to spatial

and temporal correlations. Then, the BBN learns the conditional dependencies in the

observations for detecting the outliers in the collected data.

Another approach based on k-nearest neighbours for outlier detection was developed by Branch

et al. [72]. Here, the anomaly is detected by computing the average value of the k-nearest

neighbour readings, and comparing it with a pre-determined threshold.

2.4.1.2 Anomaly detection:

Kaplantzis et al. [73] proposed a scheme for detecting black hole attacks and selective

forwarding attacks, using routing information bandwidth and hop count to determine the

malicious WSN nodes. In black hole attacks, misleading RREP (Routing Reply) messages are

sent by malicious nodes in response to “Route Request” (messages) from weak and vulnerable

(prone to attack) nodes, indicating incorrectly, that routes to the destinations are found. This

leads to source notes assuming that their packets are being delivered correctly to the

destination, whereas, vulnerable nodes will drop all network’s messages. The selective packet

dropping attack prevention technique based one class SVM (support vector machine) was

proposed by the authors in [73] to address this issue. However, use of traditional SVM is highly

computational intensive, and Rajasegarar et al. [74] proposed a light weight SVM approach for

anomaly detection, called quarter-sphere one class SVM to alleviate this problem. The

approach allows distributed implementation in WSN, and can distinguish anomalies in data

while minimizing communication overheads. Further, Yang et al. in [96], improvised this

Chapter 2

56

algorithm, by having unsupervised clustering technique for learning the anomalies in

distributed nodes, and using the one-class quarter-sphere SVM at the centralised control station

nodes, and show significant improvement in computational complexity. Their approach is

similar to the one proposed in [74]. Another approach using artificial immunity algorithm in

conjunction with SVM for intrusion detection is proposed by Chen et al in [[87]. Artificial

immunity algorithm is a computational intelligence algorithm inspired by biological immunity

systems [164] for problem solving, and involves automatic generation of immune bodies

(antibodies) against the antigen or virus through the cell fission mechanism. For the intrusion

detection scheme, the immunity algorithm was used for pre-processing, the sensor data, which

was fed to SVM after pre-processing for anomaly intrusion detection. Another approach, using

one-class ellipsoid SVM was proposed by Zhang et al. [88], which extracts the temporal and

spatial correlations from the collected readings to train the SVM for developing an outlier

detection technique. The ellipsoid SVM method uses linear optimization instead of quadratic

optimization used for traditional SVM, leading to efficient learning, good performance, and

ability to learn nonlinear and complex problems. However, high computational and large

memory requirements are the main disadvantages, due to scalability problems with large data

sets [85]. An alternative approach using self-organising map (SOM) was proposed by Avram

et al. in [163], who addressed the issue of detecting network attacks in wireless adhoc networks

using an unsupervised learning approach based on SOM, where the weights are learnt from the

statistical analysis of the input data vectors. However, this approach also is not capable enough

to detect malicious attacks in complex data sets from large scale WSNs.

2.4.2 Data Integrity, Fault Detection, and QoS Enhancement:

The state of the art and general QoS requirements in WSNs have been reviewed in [166], and

authors here reiterated that since WSNs suffer from energy and bandwidth constraints, which

can limit the quantity of information that can be transmitted from a source to destination node.

Further, due to random network topologies, and faulty, unreliable data

aggregation/dissemination in WSNs, QoS (Quality of Service) guarantees are necessary. The

QoS enhancements guarantee high priority delivery of real-time events and data, particularly

for complex WSN architectures with multi-hop transmissions of data to the end user, and

distribution of queries from a central system controller to the WSN nodes [165]. Some of recent

efforts discussed next on using machine learning techniques to achieve specific QoS and data

integrity metrics, ascertain several advantages, such as:

Chapter 2

57

Use of machine learning approaches can eliminate the need for flow-aware and stream-

aware management techniques, as they can be trained to recognise different types of

streams automatically.

Machine learning methods can automatically detect the type of network service and the

type of WSN application, and it is possible to meet requirements corresponding to QoS

guarantees, data integrity and detection of faults, while ensuring efficient resource

utilization, mainly bandwidth and power utilization.

Some of the approaches proposed on using machine learning methods for QoS guarantees, data

integrity and fault detection in WSNs are as follows.

2.4.2.1 Using Neural Networks for QoS estimation

Of late, there is a significant interest in estimating and enhancing the WSN performance using

automated approaches. A sensor network dependability metric was proposed by Snow et al.

[75] to represent the availability, reliability, maintainability and survivability of the sensor

network. To estimate the dependency metric, the authors used features performance measures

such as MTBF (Mean Time Between Failure), and MTTR (Mean Time To Repair). Another

approach for modelling dynamic fault detections was proposed by Moustapha and Selmic [76],

where the method models the dynamic behaviour of nodes’ and their effects on other

neighbouring nodes. Further, they used an innovative variation in terms of using the

backpropagation method used for neural network learning for node identification and fault

detection similar to how it was used in [75]. This variation allowed a nonlinear sensor model

to be derived that can adapt to different application with fault detection requirements.

2.4.2.2 Learning Based Quality Estimation Framework

Wang et al. [77] proposed a link quality estimation framework called MetricMap, which

addresses the inadequacies of traditional link quality measurement tools, due to different

operating conditions such as signal variations and interference, leading to inaccurate and

unstable readings across different environments [172]. The proposed MetricMap framework,

for link quality estimation, uses supervised learning techniques, to obtain link quality

Chapter 2

58

indicators. MetricMap is an enhancement over previous MintRoute protocol proposed by

authors in [173], where combination of online and offline learning for decision tree classifiers

was adopted for obtaining link quality indicators. MetricMap builds the classification tree using

several local features, such as RSSI (received signal strength indicator), size of the transmission

buffer, channel load, and forward/backward probabilities. Here, the ratio of the received to the

total transmitted packets is termed as forward probability pf(l), whereas the calculation over the

reverse path is the backward pb(l). Further, as the global features over far away nodes are

communication intensive, local features in the neighbouring nodes are preferred. Experimental

validation of MetricMap framework allowed around three times improvement in data delivery

rate as compared to basic MintRoute method.

2.4.2.3 Use of Multi Output Gaussian Processes for WSN node Accuracy and

Reliability Assessment

A real time algorithm to discover a set of nodes that can handle information processing tasks

corresponding to assessment of accuracy of collected sensor readings, and prediction of

missing readings was proposed by Osborne et al. in [167]. Here, as shown in Eqn 3.6, the

algorithm uses a probabilistic Gaussian process to estimate a reasonable size of training data

by using the priors (historical data/previous experience) and a multivariate Gaussian process

to predict the posterior distribution of an observed environmental variable x (the sea-surface

temperature).

𝑝((𝑥|𝜇, 𝐾, 𝐼)) ≜1

√𝑑𝑒𝑡2𝜋𝐾𝑒𝑥𝑝 (−

1

2(𝑥 − 𝜇)𝑇𝐾−1) (𝑥 − 𝜇) (Eqn 3.6)

where μ, K are the prior mean and covariance of the variable x, respectively, and I denotes the

historical data that is updated online (a sequence of time-stamped samples) to include the new

sequentially collected observations.

2.4.2.4 QoS guarantee based on reinforcement learning

A Q learning based approach for QoS guarantee was proposed by Ouferhat and Mellouk in

[168]. Here, the authors introduced a QoS task scheduler for multimedia sensor networks based

on Q-learning type of reinforcement learning technique, and shown that it is possible to

enhance the network throughput significantly by reducing the transmission delay. Seah et al

[169] on the other hand used WSN coverage as the QoS metric and shown how a Q-learning

Chapter 2

59

method allows efficient monitoring of area of interest in WSN setup. They developed a Q-

learning based distributed learner, that can detect weakly monitored regions, which need to be

scheduled for upgrades in future WSN deployment stages.

Another approach that used the capabilities of Q-learning, considering energy harvesting for

QoS guarantees is proposed by Hsu et al. [170]. The authors, introduced energy harvesting

capabilities, for a QoS-aware WSN power management scheme, and called it - “Reinforcement

Learning based QoS-aware Power Management” (RLPM). The RLPM employs Q-learning

technique to adapt to the dynamic levels of nodes’ energy (In systems with energy harvesting

capabilities). QoS-aware RLPM allows QoS awareness and manages nodes’ duty cycle under

the specified energy restraints. A different approach for QoS guarantee was proposed by Liang

et al. [171] called “Multiagent Reinforcement Learning based multi-hop mesh Cooperative

Communication” (MRL-CC), where MRL-CC is adopted to reliably assess the data in a

cooperative manner. Here, MRL-CC can also examine the impact of traffic load and node

mobility on the performance of the whole network.

2.4.3 Application Specific Unique Challenges

There are some novel application specific challenges, which cannot be categorized into

mainstream machine learning WSN literature, but nevertheless, are unique and provide insight

into how some unforeseen aspects of WSNs were addressed. Some of these are briefly

discussed here.

2.4.3.1 Reinforcement Learning for WSN Resource Management

An algorithm that exploits the local information and constraints imposed on the WSN

application, to optimize various tasks over a period of time, while maximising energy

efficiency was presented by Shah and Kumar in [174]. For this algorithm, termed as DIRL

(Distributed Independent Reinforcement Learning), each WSN node learns the minimum

required resources to perform its scheduled tasks, with rewards assigned by Q-learning method

and finds the optimal parameters of the application equipped with WSN. As an example, for

an object recognition and tracking application shown in Figure 15, the Q-learning based DIRL

algorithm can allow learning of task priorities for a certain task schedule of this application.

The object tracking application, which consists of five different tasks, such as:

Chapter 2

60

Collection of two or more readings into a single reading

Transmission of a message to the next hop

Receipt of incoming messages

Reading of next sample

Placing the node into sleep mode.

These tasks need to be performed in certain priority for maximising the lifetime, and WSN

does not have a predetermined schedule for achieving this performance goal (such as

knowledge of physical proximity of object to a node for enabling the task of reading samples).

Under such circumstances, Q-learning based DIRL task scheduler, can learn from penalties and

rewards assigned for wrong/right decisions during learning stage, and can perform better in

real time based on this knowledge.

Figure 15 WSN Based Q-Learning for Object Tracking Application [174]

2.4.3.2 Decision Tree Based Learning for Animal Behaviour Classification Application

Applications such as habitat and environment monitoring also have used WSNs and used

simple machine learning classifiers to learn the behaviour of herds of animals [175]. Nadimi et

Chapter 2

61

al. [176] utilized a decision tree learner to classify the animal as active or inactive, using

features such as the pitch angle of the neck and movement velocity, from a herd of animals.

This application performed well due to simple implementation and low complexity, with a

decision tree learner and use of few critical features.

2.4.3.3 SOM (Self Organising Map) based Clock Synchronisation

As the modern WSN nodes have to perform several tasks until limited resources, clock

synchronisation between sensor nodes is an important requirement, to maintain consistency in

execution of tasks between the sensor nodes for large scale WSNs. A SOM (self organising

map) based reliable clock synchronisation technique was proposed by Paladina et al. in [177],

where the nodes can predict the near optimal estimation of current time without a need for

central timing device, with restricted storage and computing resources. This method, however,

presumes that the nodes are deployed uniformly over the monitored area, and all the nodes

have same transmission powers, which is not always the case.

2.4.3.4 Neural network based Air Quality Monitoring

A neural network based air quality monitoring approach for measuring pollution levels was

proposed by Postolache et al. in [178]. Here the detection of air quality and gas concentration

was done, by making the neural network learn the readings of inexpensive gas sensor nodes in

the WSN set-up. The implementation was done in a distributed manner by client and server

side scripting on web server and end-user computers.

2.4.3.5 Neural network based Intelligent Lighting Control

A new standard for lighting control for smart buildings based on neural networks was presented

by Gao et al. in [179]. Here, a RBF (Radial Basis Function) neural network was used to extract

a computational entity called “I-Matrix” (Illuminance Matrix), to measure the degree of

illuminance in the lighted area. This is quite a unique application field, and this application

field has several challenges, in terms of converting the detected data from photo sensors to a

quantitative or qualitative feature that can processed by computers, and can impact the

performance of the system significantly. The authors show that their approach based I-Matrix

results in 60% improvement in performance over the standard methods.

Chapter 2

62

2.5 Research Gap in Wireless Sensor Networks Based on Machine

Learning/Data Mining Techniques

As can be seen from the comprehensive previous work presented in this Chapter, a large body

of work exists in using machine learning techniques for addressing various challenges in

WSNs, included operational, non-operational, and application specific challenges, there is still

a research gap, and there is a need for further research efforts as many issues are still open and

need to be solved. Some of the gaps and further research needed are discussed below:

2.5.1 Better Methods for Selecting Sensors

A large number of sensor measurements are needed in practice, to monitor the events and

maintain desired detection accuracy. With the requirement for WSN nodes to operate under

resource constraint, network designers face several design challenges, corresponding to

network management and communication bandwidth. Since around 80% of the energy in the

sensor nodes is consumed for communication activity (sending and receiving data), efficient

data compression and dimensionality reduction techniques are needed to reduce transmission

reduce transmission and hence extend the network lifetime. Most of the previous approaches

discussed here, used PCA (principal component analysis) technique for dimensionality

reduction or data compression. However, PCA is too computationally intensive to be

implemented on WSN nodes, impacting on memory requirements, and causing severe latency

issues ( if implemented on nodes), or extra energy consumption due to the need to transmit the

data for cluster heads or sink codes for extracting features for compression or dimensionality

reduction. Though there is a trade-off between energy consumption and dimensionality

reduction or compression achieved, there is a need for alternate light weight approaches to PCA

and its variants, due to their computational intensive nature and limited resources on WSN

nodes for computing the PCA components.

2.5.2 Adaptive and Distributed Machine Learning Approaches For WSNs

Due to WSN sensors being devices with limited resources, distributed machine learning

techniques are needed for WSNs as compared to centralised learning algorithms. This will

allow less computational power requirements and smaller memory footprint (since they don’t

need to know about the whole network). Further, the algorithms need to be adaptive, allowing

nodes to learn current environment conditions and rapidly adapt their future behaviour and

Chapter 2

63

predictions dynamically. Hence, adaptive and distributed learning algorithms are needed for

reducing the communication overheads and alleviate the computational burden on the nodes.

2.5.3 Managing Resources Using Machine Learning

As discussed in the previous Sections of this Chapter, WSN designer face different types of

challenges, including operational, non-operational or application specific challenges.

Energy efficiency is one of key challenge and energy efficient design goal can be achieved

using improving operational aspects, such as enhanced communication protocols (routing and

MAC protocol design) and by detecting non-operational, energy wasteful activities, such as

listening to neighbouring nodes, transmitting redundant information, by being in active

listening mode all the time. As discussed in the previous sections of this Chapter, while the

first aspect – the design of enhances communication layer protocols based on machine learning

approaches have received significant research attention, with large body of literature available,

the 2nd aspect on design of energy saving approaches has received less attention, and there are

not many approaches available.

2.5.4 Spatio-Temporal Correlation Detection

With several sensor nodes it is quite possible, there is large redundant information being

communicated within the network. This could lead in wastage of energy, and if correlation and

dependencies between the sensors can be detected, both spatially and temporally, and reduced

number of sensors can be used for communication and event detection and monitoring,

significant energy savings can be possible. With most of the earlier approaches examined in

the previous sections in this Chapter, there seems to be not many approaches that exploit the

spatio-temporal correlations for achieving energy efficiency in WSNs.

2.6 Research Plan and Thesis Road Map

To address these research gaps on achieving energy efficient WSN design with machine

learning techniques, in this thesis a novel integrated framework is proposed, which takes into

consideration both operational, non-operational and application-specific challenges to address

the WSN challenges. The integrated framework for energy efficient WSN based on machine

learning, consists of three stages:

Chapter 2

64

Stage I: The Stage1 is based on the proposal of a joint energy efficiency–event

detection model, where we develop a novel sensor node selection scheme that conserves

the energy in the wireless sensor network, and at the same time maximizes the event

recognition performance. Here, the scheme utilises, fewer sensor nodes at a time, and

placing unwanted sensor nodes in the sleep mode. For this, a novel objective

quantitative measure is proposed to assess the energy efficiency achieved, namely, the

life time extension factor (LTEF). We show that this joint scheme, allows selection of

most significant and influential sensor nodes for participation in different WSN tasks,

and contributes significantly towards energy savings and event detection accuracy. The

detailed design and experimental validation for this scheme is presented in Chapter 3.

Stage 2: As the WSN components need to adapt to the state of the WSN environment

being monitored dynamically, the number of sensor nodes participating in the routing

tree cannot remain fixed, and need to adapt, in order to accurately monitor and predict

the physical environment, and the second contribution of this work is on design of

adaptive models for sensor selection and classifier learning which can energy efficiency

and prediction accuracy, based on performance targets specified. It turns out that this

scheme which involves selection of an appropriate classifier model, in conjunction with

the previous sensor selection approach, not only results in better prediction accuracy,

but also contributes towards quality of service (QoS) enhancements. This stage can be

implemented in a decentralised manner in WSN nodes or collectively at the central base

station control code. This module can be implemented in a decentralised manner in

WSN nodes. The detailed design and experimental validation for this scheme is

presented in Chapter 4.

Stage 3: The third and the final contribution is proposal of a joint sensor selection

adaptive routing model, for addressing the dynamic WSN environment, which has a need

to adapt the routing scheme while maintaining the energy efficiency and prediction

accuracy targets. This scheme, also leads in improvement in some non-functional

challenges such as recovery from sensor failure, and model building time, which are

important for maintaining QoS guarantees the detailed design and experimental

validation for this scheme is presented in Chapter 5.

The details of each of these modules are presented in next 3 Chapters of this thesis.

Chapter 3

65

Chapter 3

Joint Sensor Selection - Event Detection Scheme

3.1 Introduction

In this Chapter, the details of joint sensor selection and event detection scheme are presented. In

this scheme, a data driven method was used for learning the most significant sensors. The sensors

are modelled here with the features extracted from the data sets corresponding to different WSN

application scenarios, including acoustic data Isolet, Ionosphere data and Forest cover type data.

In this formulation, minimizing the number of sensors for energy efficient management becomes

equivalent to minimizing the number of features [25]. For minimizing, a feature ranking

approach is used, where the features are ranked according to their significance in the wireless

sensor network. That means we first rank the sensors from the most significant to the least

significant, and then select optimal number of sensors to meet a specified accuracy target[26].

For validating the proposed scheme, we used different publicly available datasets corresponding

to wireless sensor networks in UCI Machine Learning repository [25]. This Chapter will explain

results and studies done on Isolet, Ionosphere, forest cover type and forest fires datasets. Each

data set consists of different number of sensors (features).

3.2 Joint Energy Efficiency - Event Detection Scheme

The block schematic of the joint sensor selection and event detection scheme for the integrated

framework is proposed is as shown in Figure 16.

3.2.1 Energy Efficiency with Feature Ranking Algorithm

The sensor selection algorithm uses the feature selection and ranking technique to determine

most influential sensor by learning the influence of each feature on the event detection

performance, and discards insignificant sensor in the WSN cluster, and keeps the significant

sensor for predicting the application event. For this, a feature selection and ranking algorithm

has been developed which uses the 'independent features' significance testing [175] to extract the

significant sensors in the WSN, and this involves calculation of the significance level of each

sensor from input data measurements, and their ability to distinguish WSN event categories, with

Chapter 3

66

a pre-determined threshold, and sorting them for ranking. Figure 17 shows the implementation

of algorithm for selecting the significant sensors.

Figure 16 Block Schematic for Joint Energy Efficiency - Event Detection Scheme

Figure 17 Sensor Selection and Ranking Algorithm

Chapter 3

67

3.2.3 Naïve Bayes Machine Learning Classifier Algorithm

The Naive Bayes Classifier algorithm is based on the Bayesian theorem and is particularly

suited when the dimensionality of the inputs is high, and number of instances is low. Given a

set of variables, X = {x1, x2, x...,xd}, if we want to construct the posterior probability for the

event Cj among a set of possible outcomes

C = {c1, c2,c...,cd}. In a more familiar nomenclature, X is the predictors and C is the set of

categorical levels present in the dependent variable. Using Bayes' rule:

Equation 1: Bayes's Rule

where p(Cj | x1, x2, x...,xd) is the posterior probability of class membership, i.e., the probability

that X belongs to Cj. Since Naive Bayes assumes that the conditional probabilities of the

independent variables are statistically independent we can decompose the likelihood to a

product of terms:

Equation 2

and rewrite the posterior as:

Equation 3

Using Bayes' rule above, we label a new case X with a class level Cj that achieves the highest

posterior probability.

Although the assumption that the predictor (independent) variables are independent is not

always accurate, it does simplify the classification task dramatically, since it allows the class

conditional densities p(xk | Cj) to be calculated separately for each variable, i.e., it reduces a

multidimensional task to a number of one-dimensional ones. In effect, Naive Bayes reduces a

high-dimensional density estimation task to a one-dimensional kernel density estimation.

Chapter 3

68

Furthermore, the assumption does not seem to greatly affect the posterior probabilities,

especially in regions near decision boundaries, thus, leaving the classification task unaffected.

3.3 Experimental Validation

Four different data sets corresponding to different event recognition application were used for

experimental validation. The data sets used were from publicly available repository.

Table 1. Data sets for experimental validation

The purpose of ISOLET dataset is to predict which letter or name was spoken. As can be seen in

Table 1, the ISOLET is a large data set with 7797 instances and 617 attributes (features). The

data set is divided into number of batches - Isolet 1+2+3+4 and isolet5. In this section Isolet5

part was used consisting of 1559 instances and 617 features.

Ionosphere data set contains radar data, and was collected by system in Goose Bay, Labrador.

The targets were free electrons in the Ionosphere. "Good" radar returns are those showing

evidence of some type of structure in the Ionosphere. "Bad" returns are those that do not let their

signals pass through the Ionosphere [28] . In Ionosphere dataset experiment we used all 34

attributes in addition to the class "good" and "bad".

The Forest Cover type is a huge data set with very large number of attributes (581000 attributes).

This date set used to predict the forest cover type from cartographic variables [25]. In experiment

4 we used all the attributes and instances to find out application’s event detection accuracy.

Forest fires is a regression dataset, and its aim is to predict the burned area due to forest fires.

Several of attributes in forest fires data set could be correlates, thus feature selection and ranking

can reduce the dimensionality of sensors used for detecting the application events [30]. In our

experiments, the features have been minimized to 5 features as some of attributes such as date,

time and month are not the sensor readings, and need not be included in machine learning scheme.

Data set #of instances #of Attributes Missing values? Associated tasks

ISOLET 7797 617 No Classification

Ionosphere 351 34 No Classification

Cover Type 581012 54 No Classification

Forest fires 517 13 N/A Regression

Chapter 3

69

The main aim of our experiments was to show that, to what extent, the number of features

selected may affect the accuracy and the life time extension factor (life time of the sensor network

before the sensor becomes unavailable). In the following experiments, it is shown, that the

accuracy and the life time of a sensor network depends on a variety of factors.

3.3.1 Experiment 1 (Isolet Data set)

The first experiment is on ISOLET dataset. The actual size of data we used consists of 1559

instances with 617 features, whereas the original size of the dataset is 7797 instances and 617

features. After applying our Isolet5 dataset to our feature ranking algorithm, the ranking of the

most significant features are as shown in the Table 2, where hundred features have been ranked

from 1 to 100.

Table 2 Features selected in Isolet 5

As can be seen from this table, The first row represents the first 10 features ranked in order of

significance, from 1 to 10 (455, 453, 454…..462), 2nd row shows the next 10 features ranked in

order from 11 to 20 ( 69,6,101,38. 37…..462), and so on until all features are ranked. This ranking

process, determines which particular sensor is most significant in first batch (1559 instances out

of 7797 instance) of data that has arrived in WSN, and by for determining how many sensors

need to be active to be able to detect the events in WSN, network needs to be trained first and

then used for prediction.

Ranked

Features numbers

1 2 3 4 5 6 7 8 9 10

Most significant 455 453 454 456 457 458 459 460 461 462

69 6 101 38 37 70 39 5 262 261

7 102 40 71 72 103 43 104 8 44

76 73 42 2 41 133 74 75 230 9

106 11 110 108 109 78 77 263 105 45

107 10 12 293 3 46 264 134 229 135

34 111 66 290 98 226 79 47 137 140

227 258 294 231 139 136 225 165 332 166

138 265 130 112 80 486 259 142 48 232

233 141 295 13 81 545 266 167 481 113

.. .. … … … … …. … … …

…. … … … … … … … … …

Least significant 236 467 157 177 329 485 94 147 270 239

Chapter 3

70

This is done by training a machine learning classifier by taking into consideration different

ranked features – first 10 significant features, 20, 30 …features. A simple Naïve Bayes type

machine learning classifier was used, as Bayesian classifiers work well with lesser data, and the

prediction accuracy achieved was noted for deciding the number of sensors that need to be active

in the WSN at a point of time. Table 3 shows the prediction accuracy for Naïve Bayes classifier:

Table 3 Naïve Bayes Classifier Performance

𝐿𝑖𝑓𝑒 𝑡𝑖𝑚𝑒 𝐸𝑥𝑡𝑒𝑛𝑠𝑖𝑜𝑛 𝑓𝑎𝑐𝑡𝑜𝑟(𝐿𝑇𝐸𝐹) = 𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑒𝑎𝑡𝑢𝑟𝑒𝑠

𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑒𝑎𝑡𝑢𝑟𝑒𝑠 𝑢𝑠𝑒𝑑1

The third column in Table 3 is the proposed measure for measuring the energy efficiency

achieved in terms of LTEF metric. As can be seen from Table 3, prediction accuracy improves

with increase in number of features/sensors selected by the classifier. However, this will be at

the cost of the life time extension factor. Life time extension factor (LTEF) is increased if the

number of features/sensors used are lesser, and redundant features are eliminated, and the

increase in LTEF represents increase in energy efficiency. There is a trade-off between energy

efficiency (LTEF) and prediction / event detection accuracy, meeting the performance target for

one at the cost of another. Here, an appropriate feature ranking and selection algorithm can

determine most influential sensors or most significant features, and allow redundant features to

be eliminated. Figure 18 shows a visualisation of trade- off between numbers of features/sensors

vs. event detection accuracy.

1 Life time Extension factor Equation

Features Accuracy Lifetime extension factor

10

9.62%

617/10 = 61.7

20

11.80%

617/20 = 30.85

30

13.79%

20.56

40

14.62%

15.42

50

16.10%

12.34

100

23.92%

6.17

200

41.05%

3.08

Chapter 3

71

Figure 18 Event Detection Accuracy vs. Life time extension Factor(LTEF) (Isolet 5 data set)

Table 4 Results of experiment 1

Number of Features 10 20 30 40 50 100 200

Accuracy 9.62% 11.80% 13.795 14.62% 16.10% 23.92% 41.05%

Life time extension factor 61.7 30.85 20.56 15.42 12.34 6.17 3.08

From Figure 18, it can be seen that the life time extension factor increases with lesser sensors

at the cost of accuracy. And the accuracy of a network event detection performance could be

increased at the cost of decreased life time extension factor. Further, in the event of a sensor

failure or unavailability, it is possible to maintain the accuracy by increasing the number of

features used. To emulate the sensor failure, we assigned a probability that one of the sensor Si

is not available with probability p= 0 , 0.01 , 0.05 , 0.10 , 0.50 [31]. In this experiment, we have

multiplied our Isolet5 data set with all probability values above. We have selected 10 features

and used Naïve Bayes Classifier for finding event detection accuracy, and the results are as

shown in the following table.

0

50

100

150

200

250

1 2 3 4 5 6 7

# o

f fe

atu

res

Experiment 1

features

accuracy

Life timeextensionfactor

Chapter 3

72

Table 5 Experiment 1 Accuracy with Sensor Failure Probability

As shown in Table 5, the system is quite stable with respect to occasional sensor faults. In case

of using 20, 30, 40, 50, 100 and 200 features with sensor failure taken into consideration, the

accuracy achieved was quite stable.

3.3.2 Experiment 2 (Ionoshpere dataset)

This experiment was based on Ionosphere data set. We used all 34 attributes with only two output

state of the application class, either “good" or "bad". After applying ionosphere data set into the

proposed feature ranking algorithm, the ranking of features from most significant to least

significant features are as shown in Table 6, and performance of these ranked features on

detection accuracy and Energy efficiency is shown in Table 6.

Table 6 Experiment 2 Features selected and Ranked on Ionosphere dataset

Table 7 Experiment 2 Accuracy

Features

Accuracy

Without P

Accuracy

P= 0.01

Accuracy

P=0.05

Accuracy

P=0.10

Accuracy

P=0.5

10

9.62%

9.55%

9.42%

9.56%

9.56%

Feature number/

Column number

1 2 3 4 5 6 7 8 9 10

1 2 3 5 7 1 9 31 33 29 21

2 15 23 8 13 25 14 11 12 16 6

3 19 10 18 22 27 4 17 34 28 32

4 20 24 30 26

Features Accuracy Life time extension factor

10

38.74%

34/10 = 3.4

20

35.89%

34/20 = 1.7

30

35.89%

34/30 = 1.1

34

35.89%

34/34 = 1

Chapter 3

73

A comparison of performance of Ionosphere and Isolet datasets in Figure 19 shows that using

more sensors can improve the prediction accuracy, but at a highest cost - in terms of reduced

energy efficiency.

Figure 19 Accuracy and life time extension factor (Ionosphere)

Table 8 Experiment 2 results

Number of Features Accuracy Life time extension factor

10 38.74% 3.4

20 35.89% 1.7

30 35.89% 1.1

34 35.89% 1

3.3.3 Experiment 3 (forest Cover type data set)

The data set used in experiment 3 is Forest Cover Type dataset. This dataset is a large data set

with large number of samples, consisting of 581012 instances and 54 attributes. After applying

feature ranking algorithm to the forest cover type data, features are ranked and selected in the

following table from the most significance to the least significance of relative importance.

0

5

10

15

20

25

30

35

40

1 2 3 4

# o

f fe

atu

res

Experiment 2

features

accuracy

Life time extensionfactor

Chapter 3

74

Table 9 Experiment 3 features ranked and selected for forest cover type dataset

Table 10 Experiment 3 Accuracy and life time Extension factor

The results of this experiment are shown in Table 11, and Figure 20 shows the prediction

accuracy vs. the energy efficiency in terms of LTEF metric. As can be seen from Table 11 and

Figure 20, for same set of features- say 10 sensors, the prediction accuracy achieved is better than

the previous two data sets. This could be due larger data size available for training stage and

ability to learn the model better for Forest cover type data, as compared to Ionosphere and Isolet

type of data. Further, it can be seen that increase in number of sensors used does not improve the

prediction accuracy. That is, as we increase in number of sensors, from 10 to 54, the prediction

accuracy improves from 68% to 68.49%, and impact of this on energy efficiency is worst, as the

LTEF drops from 5.4 to 1.0.

Feature Number 1 2 3 4 5 6 7 8 9 10

1 15 19 28 29 51 1 26 36 37 52

2 24 53 12 25 27 54 44 14 18 43

3 10 6 32 8 40 17 48 38 20 49

4 16 35 42 7 33 5 23 3 13 31

5 30 4 45 2 11 21 41 9 39 22

6 47 46 50 34

Features Accuracy Life time extension factor

10

68.00%

54/10 = 5.4

20

68.16%

54/20 = 2.7

30

68.27%

54/30 = 1.8

40

68.37%

54/40 = 1.3

54

68.49%

54/54 = 1

Chapter 3

75

Figure 20 Accuarcy and Life time extension factor (Forest cover Type data set)

Table 11 Experiment 3 results

Number of Features Accuracy Life time extension factor

10 68.00% 5.4

20 68.16% 2.7

30 68.27% 1.8

40 68.37% 1.3

54 68.49% 1

3.3.4 Experiment 4 (Forest fires Dataset)

For the fourth experiment the forest fires dataset was used. This data set has a size of 517 * 13

(517 samples with 13 features). Features have been reduced to 5 because 8 other attributes such

as date, time and month were not relevant features. After applying feature ranking algorithm, the

following table shows features in the order of their significance, most significant to the least

significant feature.

0

10

20

30

40

50

60

1 2 3 4 5

# o

f Fe

atu

res

Experiment 3

features

accuracy

Life timeextension factor

Chapter 3

76

Table 12 selected features on forest fires Dataset

Table 13 Experiment 2 Accuarcy Forest Fires data set

The relationship between accuracy and lifetime extension factor for the forest fires data set is

similar to experiment number one. That is increasing the number of features increases the

accuracy at the cost of life time extension factor. Further, in the event of sensor failure or

unavailability, it is possible to maintain the specified accuracy by including more sensors for

classifying the area affected by fire. However, for a healthy sensor network, using more features

or sensors is costing more resources and reduces the life time of the sensor network. It would be

energy efficient if lesser number of sensors with more significance can be used. The Accuracy

versus life time extension factor for selected features is shown below. The poor detection

accuracy is due to smaller data size and inability of network to learn the relationship between

input and output with not enough data. Figure 21 shows the performance for dataset 4.

Feature 1 Feature 2 Feature 3 Feature 4 Feature 5

3 4 5 1 2

Features Accuracy Life time extension factor

1 ( 3)

10.77%

5/1= 5

2 (3,4)

11.04%

5/2 = 2.5

3 (3,4,5)

13.37%

5/3= 1.6

4 (3,4,5,1)

13.56%

5/4= 1.25

5 (3,4,5,1,2)

13.75%

5/5= 1

Chapter 3

77

Figure 21 Accuracy and life time extension factor for forest fires data set

Table 14 Experiment 4 results

Number of Features Accuracy Life time extension factor

1 10.77% 5

2 11.04% 2.5

3 13.37% 1.6

4 13.56% 1.25

5 13.75% 1

3.4 Chapter Summary

In this Chapter, a joint sensor selection and event detection model/scheme was proposed for

WSN, based on machine learning approach. As the method is data driven, and tries to learn the

relationship between sensor data and event detection capability, different types of publicly

available datasets from UCI Machine Learning repository were used to test the proposed joint

model/scheme. Here, modelling of sensors in WSN is done with features/attributes of a dataset,

the output classes or variables modelled as the application events, and sensor

samples/measurements modelled with instances of dataset. A feature ranking algorithm was

developed which ranks the features or sensors in order of their significance in being able to

0

1

2

3

4

5

6

1 2 3 4 5

# o

f fe

atu

res

Experiment 4

features

accuracy

Life time extensionfactor

Chapter 3

78

predict the output or WSN state. Also, an objective measure to determine the energy efficiency

achieved was devised with a metric called life time extension factor (LTEF), which needs to

be improved by WSN learning from features and predicting the output with a Bayesian (Naïve

Bayes) machine learning classifier. As reiterated before, due to resource constraints on WSN

and its sensor nodes, there is a need to come up with light weight machine learning approaches,

and the scheme proposed in this Chapter based on a joint sensor selection and event detection

model is one such scheme, that can be implemented in WSN nodes, in both decentralised or

centralised topologies easily. It turns out that this scheme based on a feature ranking technique

and Naïve Bayes classifier, can indeed address the non-operational or non-functional

challenges as well, such as QoS guarantees, as it can take into account sensor failures and

guarantee event detection accuracy under sensor failures. It allows graceful management of

sensor network in the event of sensor failures, by increasing the number of sensors to meet the

specified accuracy requirements. However, the event detection performance is quite low, and

needs improvement as such. It could be possible, that better sensor selection and machine

learning approaches can address this issue. In the next Chapter, we discuss the next stage of

the proposed integrated framework, to address this shortcoming, and extend the joint sensor

selection - event detection model, with adaptive classifier models instead of simple Naïve

Bayes classifier and feature ranking algorithm used here, to address both operational

(functional) and non-operational (non-functional) challenges in WSNs.

79

Chapter 4 Adaptive Models for Energy Efficiency

4.1 Introduction

In this Chapter, we extend the scheme developed in the previous Chapter with adaptive classifier

and adaptive sensor selection models for improving the performance of integrated framework in

addressing the WSN challenges.

As the WSN nodes need to adapt to the state of the WSN environment being monitored

dynamically, the number of sensor nodes participating in the routing tree cannot remain fixed,

and need to adapt, in order to accurately monitor and predict the physical environment, and in

this Chapter, the design of data driven adaptive classifier models for improving prediction

accuracy, based on performance targets specified, is presented. It turns out that this scheme which

involves selection of an appropriate classifier model, in conjunction with the previous sensor

selection approach, not only results in better prediction accuracy, but also contributes towards

quality of service (QoS) enhancements, similar to joint sensor selection and event detection

scheme discussed in previous chapter, where the scheme can detect the sensor failures and

gracefully manage the performance targets. The adaptive classifier model scheme discussed in

this Chapter can be implemented in a decentralised manner in WSN nodes or collectively at the

central base station control code, depending on algorithm complexity, and computational

resources available.

4.2 Adaptive Classifier Model Based Scheme

The block schematic for the adaptive classifier model scheme is shown in Figure 22. Random

forests, random trees and decision tree classifier was compared with baseline Naïve Bayes

classifier to achieve energy efficiency. Depending in the WSN configuration, this scheme can be

implemented in a decentralised, distributed manner at WSN cluster head nodes, or in a centralised

manner at central control station. Along with sensor selection scheme proposed in the previous

Chapter, augmentation with adaptive classifier models allows better energy efficiency and

prediction accuracy, as compared to simple Naïve Bayes classifier.

80

Figure 22 Adaptive Feature Selection and Classifier Model for Energy Efficiency

The experimental validation of the proposed scheme was done on a publicly available UCI

machine learning dataset, shows that the proposed adaptive classifier models, based on random

forests, random trees, perform significantly better than the conventional statistical classifiers,

such as Naïve Bayes, discriminant classifiers and decision trees, and can lead towards energy

efficient, intelligent event detection and monitoring and QoS enhancements in WSNs [32].

4.2.1 Data set Description

Accurate natural resource inventory information is vital to any private, state, or federal land

management agency. Forest cover type dataset provides such important information and is made

available publicly through UCI machine learning repository [25]. The original Cover type data

set is very large, and contains 581012 instances and 54 attributes. There are seven forest cover

type classes (Class 1 to Class 7), such as spruce/fire, lodgepole Pine, Ponderosa Pine,

Cottonwood/Willow, Aspen, Douglas-fir and Krummholz. We used smaller subsets of this data,

with each subset containing around 500 instances from each class (Class 1 to 7), with total

81

number of instances 500 * 7 (3500) instances. Table 15 describes the forest cover type data set

[33].

Table 15 Forest cover type original data set and subset data set Description

4.2.2 Classification Algorithms

For baseline comparison with conventional classification schemes, four different classification

algorithms have been examined in this work, including Naive Bayes, Decision Trees, Random

Forests and Random trees. Naïve Bayes Classifier has been described in the previous Chapter,

and in this Chapter rest of the classifier approaches are discussed.

4.2.2.1 Decision Tree Classifier

Decision Tree classifier, is another statistical classifier, similar to Naïve Bayes classifier, that

builds on decision trees from a set of training data, using the concept of information entropy. The

training data is a set S of already classified samples. Each sample S consists of a p-dimensional

vector X, where the Xj represent attributes or features of the sample, as well as the class in which

Si falls. Details of decision tree algorithm is discussed in [34].

Forest Cover Type original data set

Number of

Attributes

54

Class 1 spruce/fir

Class 2 lodgepole

Pine

Class 3 Ponderosa

Pine

Class 4

Cottonwood/Willow

Class 5 Aspen

Class 6 Douglas-fir

Class 7 Krummholz

Number of

Instances

581012

Forest Cover Type subset data set used for experiments

Number of

Attributes

54

Class 1 spruce/fir

Class 2 lodgepole

Pine

Class 3 Ponderosa

Pine

Class 4

Cottonwood/Willow

Class 5 Aspen

Class 6 Douglas-fir

Class 7 Krummholz

Number of

Instances

3500

82

4.2.2.2 Random forests and random trees

Random forests are based on ensemble learning method for classification (and regression) that

operate by constructing a multitude of decision trees at training time and outputting the class that

is the mode of the classes output by individual trees.

Random tree, on the other hand, involves construction of multiple decision trees randomly. When

constructing each tree, the algorithm picks a “remaining" feature randomly at each node

expansion without any purity function check. A categorical feature (such as gender) is considered

"remaining" if the same categorical feature has not been chosen previously in a particular

decision path starting from the root of tree to the current node. The details of random forests and

random trees are available in [35].

4.2.3 Experimental Evaluation

For all the experiments, 10 folds cross validation was used, with data partitioned into 10 folds,

and 9 out of 10 folds used for training and 1fold for testing with unknown data. Further, for

estimation of performance benchmarks, full training set mode was also used for evaluation. We

also examined the performance with and without feature selection/ranking algorithm to find the

optimal number of sensors needed for energy efficiency and prediction accuracy targets. The

results are as shown in Figure 23, 24, 25 and 26.

Figure 23 Performance of classifiers with 10 folds cross validation

83

Figure 24 Performance of classifiers with full training set

Figure 25 Performance of Classifiers with feature selection

84

Figure 26 Performance of classifiers with feature selection on full training set

4.3 Discussion

The comparative performance evaluation of the adaptive classifier model is shown in Figure 18-

21. As can be seen in these figures, the proposed adaptive classifier model scheme based on

random forest and random tree classifiers perform significantly better than conventional

statistical classifier approaches based on Naïve Bayes and decision trees. With 10 fold cross

validation, it was possible to achieve 86.45% with random forests, and 78.14% with random

trees, as compared to 71.08% with Naïve Bayes, and 86.05% with decision trees. With full

training set mode, which serves as a benchmark mode, random forest results in 99.94% and

random tree results in 100% accuracy. This means, there is a need to use appropriate strategies

for improving generalisation abilities, for the classifier model scheme to perform in test mode as

close as possible to learning or training mode.

For the benchmarking, for full training set mode, we use the entire training data for building the

model with each classifier, and use the same data for testing it. However, when we use k fold

cross validation (k = 10 here), we partition the data into 10 equal sized subsets. For the first fold,

the first nine subsets (90% labelled data) are used for training, and last subset (10% data) is used

for testing. For next fold, the training data consists of subset 2 to 10, and test set consists of subset

1. Likewise for each fold, the training data rotates to next 9 folds, so for each fold, the test data

is unseen 10% data, as compared to 90% of training data. As can be expected, testing with unseen

85

data (i.e. 10 fold cross validation), results in a marginal improvement for proposed random forest

(86.45%)/random tree(78.14%) as compared to conventional Naïve Bayes (71.08%) and decision

tree classifiers(86.05%).

However, the improvement is significantly higher with feature selection algorithm involved,

which is a wrapper type feature selection method used here, unlike previous Chapter, where

significant feature test was the criteria to select the significant features. With 10 fold cross

validation and feature selection, the prediction accuracy achieved is 77.94% (random forest) and

74.20% (random tree), as compared to 66.74% (Naïve Bayes) and 75.85% (decision). For

comparison with how these classifier models fare as compared to benchmark mode, testing with

full training set was done.

With full training set ( testing done on same data as training data), the improvement achieved

was much higher, as is evident from Comparative classifier performance. It must be noted that

use of feature selection method denotes improvement in energy efficiency, as lesser number of

features results in lesser computational power and storage requirements. So, a trade-off between

accuracy and energy efficiency can be achieved with appropriate choice of feature selection and

classification model. As the WSN environment changes dynamically, classifier model is adapted

from a choice of four different classifier models, so as to meet the energy efficiency and

prediction accuracy targets. With a joint and adaptive scheme, with feature selection techniques

and classifier models, it is possible, to monitor the large complex WSN for different event

recognition applications

For the feature selection method used here, we selected 8 features (sensors) using wrapper

method for feature selection. Wrapper method searches for the best subset of features, where the

feature subset assesses the quality of a set of features using a specific classification algorithm by

internal cross validation. Here, the wrapper type feature selection method allows selection of

most significant 8 features, instead of full feature set (54 features), resulting in reduced energy

consumption in terms of sensor computation and storage requirements.

86

Figure 27 Comparative classifier performance

As each feature represents a sensor in WSN, use of reduced features (8) here, implies 8 sensors

in active mode and 46 sensors in sleep mode for classifying the forest cover type environment.

This can lead to increased life for sensors, which we measure with a metric called as life time

extension factor. The life time extension factor can be obtained as ratio of total number of

features to number of features in active mode. In this case, the life time extension factor

achieved is 54/8 = 6.75, that is around 6 times increase in life of sensors or improvement in

energy efficiency. Further, the combination the feature selection and adaptive classifier models,

here can also handle sensor similar to scheme discussed in previous chapter, as the sensor

selection scheme adapts to different set of sensors and a different type of classifiers, to graceful

management of performance targets, including energy efficiency, prediction accuracy, and QoS

guarantees.

However, the weakness of the scheme is in generalisation ability, as benchmark performance

with full training set is higher than 10 fold cross validation mode. This could be due to the

characteristic of data set used or the approach used. So, to ascertain this, we examined the

scheme with adaptive classifier models for a different data set and is discussed in the next

Section.

87

4.4 Adaptive Classifier Scheme with Gas Sensor Drift Dataset

The Gas Sensor Array drift dataset is larger compared to Forest cover type dataset, and consists

of 13,910 measurements from 16 chemical sensors to predict 6 different gases at different

concentration levels. The purpose of this dataset is to provide information about the concentration

level at which the sensors were exposed for each measurement. The data set is divided into 10

batches collected over 36 months , each containing the number of measurements per class and

month indicated in the following table, with details of the data set description provided in [25,

35].

Table 16 Gas sensor Array drift data set description

Number

of

Attributes

Number

of

Instances

Number

of

Classes

Class 1 Class 2 Class 3 Class 4 Class 5 Class 6

129 13910 6 Ethanol Ethylene Ammonia Acetal

Deyhde

Acetone Toulene

4.4.1 Experimental Validation with Gas Drift Dataset

For experimental evaluation with this dataset, the adaptive classifier model was enhanced with

more powerful machine learning classifiers and adaptive feature selection model. The adaptive

classifier model consists of five different classification algorithms was examined, including

Naive Bayes, J48, MLP, Random Forests, Random trees and Random Committee. As can be

seen in the experimental validation with Gas Sensor Array drift dataset, consistent results are

obtained similar to the experiments done with Forest cover type and those done in previous

Chapter. Gas sensor Array drift data set being large, we performed the experiments for 10 batches

and averaged the results over these 10 batches. After applying Naive Bayes, Random forest, J48

(Decision Trees), Random tree and Random committee classification for each batch, the average

of the 10 batches were taken. The details are shown in Figure 28.

88

Table 17 Performance of Gas drifts sensor dataset

Gas drift Sensor data set Naive Bayes Random Forest

J48 Random Tree Random committee

Gas drift/10 folds 89.50%

99.91%

99.54%

100.00%

100.00%

Gas drift/Training set 88%

100%

100%

100%

100%

Gas drift/ folds Feature selection (best first method)

86.95%

99.98%

99.57%

100%

100%

Gas drift/Training set Feature selection best first method

86.53%

99.97%

99.57%

100.00%

100.00%

Gas drift/10 folds Feature selection (Greedy stepwise method)

87.89%

99.97%

99.55%

100.00%

100.00%

Gas drift/Training set Feature selection Greedy stepwise method

86.89%

99.97%

99.54%

100.00%

100.00%

Figure 28 Gas drifts summary of experimental results

89

The performance for the Gas Sensor Array drift dataset was much better as compared to the

forests cover type, particularly with extending the adaptive classifier model with ensemble

learning/random committee classifier, and adaptive feature selection model with best first search

and greedy search method. Naive Bayes classifier results in detection accuracy from 86.89% to

89.50%. Random Forest, J48, Random Tree and Random committee achieved very high accuracy

from 99.57% to 100%. Using different feature selection method instead of wrapper method or

significant feature method, LTEF ( the life time extension factor) has jumped up to achieved up

to 25 times (128/5 = 25.6) for the 10 folds, and the same results for the full training set. With

only 5 features selected instead of 128 numbers of features for this dataset, the energy efficiency

has been improved significantly, at highest prediction accuracy of 100%. The combination of

adaptive classifier model and adaptive feature selection model has resulted in improvement in

generalisation ability, as the 10 fold cross validation performance was 100% and is equal to

performance achieved with full training set. This can also impact on the further QoS

enhancements, in terms of sensor failures, and resource management features.

4.4.2 Experimental Validation with Gas Drift Dataset using Ensemble Learning

for Weak Classifiers

In this set of experiments, ensemble learning method was used to examine whether

performance of weak classifiers can be improved, such as Naïve Bayes and J48 (decision trees).

As can be seen in Table 18 Ensemble Learning on Gas drift sensor Array data set, the

performance of weak classifiers is improved. With 10 fold cross validation mode, for Naïve

Bayes classifier, due to bagging, the classification accuracy improves from 67.14% to 71%, and

accuracy with J48 classifier improves from 81.94% to 88 % due to bagging. The improvement

in performance due to ensemble learning is similar with full training set. This validates that for

bagging method of ensemble learning, the generalisation performance is better, as the

performance is improved for both previously seen data (full training data – a benchmark

performance), and unseen data (10 fold cross-validation). For rest of the experiments, we just

used the benchmark case, i.e. full training set.

90

Table 18 Ensemble Learning on Gas drift sensor Array data set

Ensemble Learning Methods Use cross Validation 10 folds Use Training set

MLP Multilayer Perceptron

97.94%

99.58%

Meta- Bagging- NB

59.25%

59.50%

Meta- Bagging-j48

98.61%

99.72%

Meta-Adaboost-NB

59.16%

59.38%

Meta-Adaboost-J48

99.38%

100%

Meta-stacking-NB

16.66%

16.66%

Meta-Stacking-J48

16.66%

16.66%

4.5 Chapter Summary

In this Chapter the adaptive classifier models were proposed to address the WSN

challenges. The adaptive classifier models scheme performed extremely well, and along

with adaptive feature selection scheme, it could achieve energy efficiency and prediction

accuracy targets, as well as address the QoS and resource management issues. For

experimental validation, two different types of large datasets was used, the forest cover

type dataset and Gas drift type data set to emulate a large physical environment instrumented

with WSN, with each attribute/feature from the data set representing the model of a WSN

node/sensor - set up for monitoring a complex and large physical environment. With Gas sensor

data set, it was showing consistency with findings from Forest Cover Type experiments – a

significant performance improvement with combined adaptive classifier and feature selection

model, with random forests, random tree, and random committees, and with best first and greedy

search feature selection techniques. Further, using a different learning scheme within the adaptive

classifier model - the ensemble learning scheme, it was possible to pull up combined performance

of weak classifiers, such as Naïve Bayes, and J48 decision trees and improve their prediction

accuracy performance metric. This validates the hypothesis that the powerful machine learning

approaches can indeed address different WSN challenges, including, the operational/functional

challenges such as energy efficiency and event detection accuracy, and the non-operational/non-

functional challenges such as failure recovery and resource management. In the next Chapter, the

proposed integrated framework is further extended - with a joint sensor selection - adaptive

routing model/scheme for energy efficiency.

Chapter 5

91

Chapter 5

Joint Sensor Selection- Adaptive Routing Model

5.1 Introduction

In this Chapter, the third stage of integrated framework for addressing WSN challenges is

presented. The third stage, involves the joint sensor selection and adaptive routing scheme, which

can address WSN challenges with missing data or lack of sufficient data due to sensor failures.

The proposed approach involves an adaptive routing scheme to be used for energy efficiency and

works in conjunction with extensions to sensor selection scheme proposed in earlier chapters.

The experimental validation of the proposed scheme for publicly available Intel Berkeley lab

Wireless Sensor Network dataset shows, it is indeed possible to achieve energy efficiency, even

under the missing data or insufficient data scenarios, with an adaptive routing protocol.

Here, the adaptive routing scheme is based on selecting most significant sensors based on Akaike

criterion, for learning the physical environment from sensor measurements. The experimental

validation of this scheme was done with of a publicly available WSN dataset acquired from real

indoor physical environment, the Intel Berkeley Lab [38].

5.2 Intel Berkeley Lab WSN dataset

The publicly available data set used for experimental validation consists of Mica2Dot sensors

with weather boards collected time stamped topology information, along with humidity,

temperature, light and voltage values once every 31 seconds. Data was collected using the

TinyDB in-network query processing system, built on the TinyOS platform [38] . The sensors

were arranged according to the Figure 29. The x and y coordinates of sensors (in meters relative

to the upper right corner of the lab) are given in a separate file. The three columns correspond

to mote id, x location, and y location.

This csv file extracted from the downloaded dataset includes a log of about 2.3 million readings

collected from these sensors. The file is 34MB gzipped, 150MB uncompressed. The schema is

as follows:

Chapter 5

92

Table 19 Intel lab data set file schema

date:

yyyy-mm-

dd

time:

hh:mm:ss.xxx

epoc

h:int

moteid:

int

temperature:

real

humidity:

real

light:

real

voltage:

real

To examine the WSN performance on quantity of data available for learning the relationships

between different variables, for prediction capability, we used three different sample sizes - 35

samples , 2700 samples and 5400 samples, corresponding to temperature and humidity sensor

measurements, which come from a deployment of 54 sensors in the Intel research laboratory

at Berkeley [38]. A picture of the deployment is provided in Intel Berkeley Wireless sensor

network Data set: location of 54 sensors in an area of 1200 m2, where sensor nodes are

identified by numbers ranging from 1 to 54.

Figure 29 Intel Berkeley Wireless sensor network Data set: location of 54 sensors in an area of

1200 m2

Many sensor readings from WSN test bed were missing, due to this being a simple prototype

testbed. This gives us a challenging opportunity and test whether the proposed integrated

machine learning framework can cope with missing and insufficient information. We selected

from this data set few subsets of measurements. The readings were originally sampled every

thirty-one seconds. A pre-processing stage where data was partitioned and normalised was

Chapter 5

93

applied to the data set. Also, for this WSN test bed, all the sensors can play the role of both

sources as well as sink node, and can be configured to be a source node or sink node in a test

session. This is how nodes in most distributed WSNs are set up, and can be configured as

source or sink node, based on the decentralised or centralised topology, and assignment of

different nodes as cluster heads, control nodes, sensing nodes etc. This arrangement allows

different type of routing protocols to be tested as well, under different operational or functional

challenges. Those nodes which actively participate in sensing the environment, whether it is a

source node or sink node, can transmit the data, and consume the power and those which don’t

participate in this activity do not consume any power. This can allow an energy efficient WSN

design; by involving optimum number of sensors to participate in environment sensing and

transmission task, and leaving non-participating sensors in sleep mode (no energy consumption).

This can however, impact on the accuracy of sensing the environment, if number of sensors

participating in routing scheme is not properly chosen. To ensure a trade-off between accuracy

and energy efficiency is achieved, it is essential that a dynamic or adaptive routing scheme is

used, where, the machine learning/data mining technique can use larger training data from

previous/historical data sets to decide the sensors participating in the routing scheme, and meet

the performance targets, in terms of energy efficiency, prediction accuracy and other QoS

metrics. The block schematic of this joint sensor selection and adaptive routing model is shown

in Figure 30 below.

Chapter 5

94

Figure 30 Joint Sensor Selection – Adaptive Routing Model

5.3 Intel Lab data file versus Intel Lab data file restructured for

experiments

The files used for experiment contains approximate readings of 65000 samples for each mote ID

the following diagram shows the process done on the main file to achieve the sensor selection

and routing approach

Figure 31 Intel lab main source file structure

Chapter 5

95

The main original Data set contains huge number of readings about 65000 samples for all 54

sensors. In this research, samples of 35, 2700 and 5400 readings have been taken in 3 separate

files for each temperature and humidity make the total number of files is six for all experiments.

The same set of experiments and samples have been repeated for humidity from the main source

file and as per the following structure for temperature and humidity.

Chapter 5

96

Figure 32 Sample files temperature readings 35, 2700 and 5400 samples

Chapter 5

97

Figure 33 Sample files temperature readings 35, 2700 and 5400 samples

Chapter 5

98

5.4 Sensor Selection and Adaptive Routing Model

The proposed sensor selection and routing approach is based on a feature selection technique

called Akaike criterion [40], [41], that selects the attributes (sensors), by evaluating the worth of

a subset of attributes by considering the individual predictive ability of each feature/sensor along

with the degree of redundancy between them. Subsets of features that are highly correlated with

the class while having low inter correlation are preferred [39, 40].

Further, this feature/sensor selection algorithm identifies locally predictive attributes, and

iteratively adds the attributes with the highest correlation with the class as long as there is not

already an attribute in the subset that has a higher correlation with the attribute in question. Once

the appropriate group of sensors are selected, the prediction of sensor output at sink node or base

station is done by linear regression algorithm, using the Akaike criterion [40], [41], which

involves stepping through the attributes, removing the one with smallest standardized coefficient

until no improvement is observed in the estimate of the error given by Akaike information metric.

Figure 34 shows how the sensor selection evolves as the training data (historic data) used for

predicting the sink sensor output is increased, and ensures the prediction accuracy/error is

maintained at a particular threshold value. Here, prediction error (RMSE) was used as the metric,

in contrast to detection/prediction accuracy used in earlier chapters.

Figure 34 Temperature Sensor selection map for 3 experiment scenarios- 1, 2 and 3

Chapter 5

99

Figure 35 Humidity sensor selection map for 3 experiment scenario 1,2 and 3

5.5 Experimental Results and Discussion

Different sets of experiments were performed to examine the relative performance of sensor

selection and adaptive routing model proposed here. K-fold stratified cross validation technique

has been used for performing experiments, with k=2, 5 and 10, based on the training data

available (using larger folds for larger training data). Further, to estimate the relative energy

efficiency achieved, we performed experiments with all sensors (without feature selection/sensor

selection) algorithm, and with sensors selected by feature selection algorithm. As mentioned

before, the feature selection algorithm allows selection of an optimal number of features or sensor

nodes needed to characterize or to classify the environment (which in turn leads to an energy

efficient scheme). Further, time taken to build the model is also an important parameter,

particularly for adaptive sensor routine scheme to be used for real time environment monitoring.

Chapter 5

100

Table 20 Temperature results from three experiments scenarios

Experiment # (Temperature)

Number of Sensors

Number of Samples

Features Selected

Time (No F selection)

Time (F selection)

RMSE No F selection

RMSE with F selection

Experiment 1 54 35 17,50 0.02 sec 0.01 sec 20.26 0.04

Experiment 2 53 2700 3,14,16,19,39 0.43 sec 0.02 sec 5.02 2.23

Experiment 3 53 5400 3,13,14,16,19,24,53 0.57 sec 0.03 sec 3.93 2.93

Table 21 Humidity results from three experiments scenarios.

Experiment # (Temperature)

Number of Sensors

Number of Samples

Features Selected

Time (No F selection)

Time (F selection)

RMSE No F selection

RMSE with F selection

Experiment 1 52 35 7,24,41,44,50 0.01 sec 0.01 sec 3.82 0.04

Experiment 2 52 2700 3,7,11,14,16,22,28,29,34,41 0.14 sec 0.01 sec 0.96 2.11

Experiment 3 52 5400 14,19,24,25,36 0.17 sec 0.03 sec 1.91 4.56

For the Temperature and humidity set of experiments, 54, 53, 52 sensors and a small set of

training samples (35 Humidity measurements) have been used. As can be seen from the sensor

locations shown in Humidity sensor selection map for 3 experiment scenario 1,2 and 3, sensor

50 is the sink node (emulating base station node), and sensors 1 to 49 participate in measuring

and transmitting the environment around them to the sink node, where the machine learning

prediction task is to estimate the measurement at sink node (sensor 50). The RMS error (root

mean squared error) at the sink node (node 50) provides a measure of prediction For all source

sensor nodes (1-49) in WSN participating in measuring the temperature in the environment and

sending it to sink node, the RMS error is 3.82%, and with sensor selection scheme used with only

5 sensors participating in routing scheme, the RMS error is 0.04%. As can be seen in Table 21,

with a moderate degradation in accuracy (3.82% to 0.04%), energy efficiency achieved is of the

order of 52 (52/5). The measure for energy efficiency, is the life time extension factor (LTEF)

metric, which can be defined as:

Life time Extension factor = Total number of features

Number of features used

With 2 sensors out of 54 sensor nodes in active mode, the LTEF achieved is around 27 times,

and 52 sensor nodes are in sleep mode. The trade- off is a slight reduction in accuracy. This could

be due to less training data used. We used only 35 temperature samples for prediction scheme.

Chapter 5

101

With more data samples used in the prediction scheme, performance could be better. To test this

hypothesis, we performed next set of experiments.

Figure 36 Temprature Experiment 1,2 and 3 results

Figure 37 Humidity experiment 1,2 and 3 results

For second set of experiments, we used 2700 training samples collected on different days. As can

be seen in Table 21, with larger training data size, we found that the participating sensors in the

Chapter 5

102

routing scheme are different, as the proposed feature selection algorithm chooses different set of

sensors (3, 14, 16, 19, 39). We used 53 sensors for this set of experiments, as two of the sensors

(sensor 5 did not have more than 35 measurements). With all 53 sensors in the routing scheme,

the RMS errors is 5.02%, and with 6 sensor nodes (3, 14, 16, 19, 39), the error is 5.02%. This is

a significant improvement in prediction accuracy (from 5.02% to 2.23%), with life time extension

of 10.6 (53/5). As is evident here, by using larger training data (2700 temperature measurements),

it was possible to achieve an improvement in prediction accuracy and energy efficiency as well.

To examine the influence of increasing training data size, we performed third set of experiments

with 5400 samples. The performance achieved for this set of experiments is shown in Table 20.

Here the adaptive routing scheme based on proposed feature selection technique selects 8 sensors

(3, 13, 14, 16, 19, 24, 34, 53). For this set of experiments, the RMS error varies from 5.02% for

all sensors participating in the scheme to 2.23% with LTEF of 6.6 (53/8). Though there is no

degradation in prediction accuracy, there is not much improvement in energy efficiency, with

doubling of training data size for the building the model this could be due to overtraining that has

happened, with the network losing its generalisation ability. So by increasing training data size,

it may not be just possible to achieve performance improvement, for pre-diction accuracy (RMS

error) and energy efficiency (LTEF), and a trade off may be needed. An optimal combination of

training data size, and number of sensors actively participating in routing scheme can result in

energy efficient WSN, without compromising the prediction accuracy.

Figure 38 Temperature, root mean square error

Chapter 5

103

Figure 39 Humidity, Root mean square error

Further, another important parameter is model building time, which represents learning time for

learning a new route, as for adaptive sensor routing scheme to be implemented in real time WSN

environment, routing scheme has to dynamically compute the sensors that are in active mode and

in sleep mode. Out of 3 experimental scenarios considered here, as can be seen from Table, the

model building time improves from 0.02 seconds to 0.01 seconds for experiment 1, from 0.43

seconds to 0.02 seconds for experiment 2, and from 0.57 seconds to 0.03 seconds for experiment

3. So, the proposed adaptive routing scheme for sensor selection provides an added benefit of

reduced model building times, suitable for real time deployment. Figure below shows the time

taken to build the model.

Chapter 5

104

Figure 40 Time taken to build the model, Temperature

Figure 41 Time taken to build the model, Humidity

5.6 Chapter Summary

In this Chapter, we proposed a joint sensor selection - adaptive routing model for sensor nodes

in WSN, based on machine learning with a feature selection algorithm based on Akaike criterion,

and can adapt them continuously as time evolves ( more data arrives). The experimental

evaluation for a real world publicly available WSN dataset, the Intel Berkeley Lab WSN test bed,

Chapter 5

105

validated our hypothesis, and allowed WSN operational and non-operational challenges to be

addressed including energy efficiency, prediction error, and QOS enhancements, such as

robustness to sensor failures and quick MAC layer adaptation ( with fast learning times) Next

Chapter concludes this work, with three major contributions for the integrated framework

proposed, and future directions of this research.

Chapter 6

107

Chapter 6 Conclusions and Future Directions

In this thesis a novel integrated framework for energy efficiency based on machine learning and data

mining techniques is proposed. The three stages of this framework, with joint sensor selection – event

detection model, adaptive models for energy efficiency, and joint sensor selection and adaptive routing

model, allow various functional and non-functional challenges in WSN to be addressed, including energy

efficiency, event detection accuracy, MAC layer routing adaptation, QoS enhancements, sensor failures

and model building or learning time.

In Chapter 3, a joint sensor selection and event detection model was proposed for WSN, based on machine

learning approach. As the method is data driven, and tries to learn the relationship between sensor data and

event detection capability, different types of publicly available datasets from UCI Machine Learning

repository were used to test the proposed joint model/scheme. Here, modelling of sensors in WSN is done

with features/attributes of a dataset, the output classes or variables as the application events, and sensor

measurements modelled with instances of dataset. A feature ranking algorithm was developed which ranks

the features or sensors in order of their significance in being able to predict the output or WSN state. Also,

an objective measure to determine the energy efficiency achieved was devised with a metric called life

time extension factor (LTEF), which needs to be improved by WSN learning from features and predict the

output class/variable accurately, to validate the hypothesis proposed in this work. An extensive

experimental evaluation with several publicly available datasets show, that proposed joint sensor selection

– event detection model allows learning from historical data, and meet the operational/functional WSN

challenges such as energy efficiency (LTEF), event detection (prediction accuracy) and QoS guarantees

(sensor failures).

As reiterated before, due to resource constraints on WSN and its sensor nodes, there is a need to come up

with light weight machine learning approaches, and the scheme such as the joint sensor selection – event

detection model proposed in this Chapter is one such simple and effective scheme, that can be implemented

in WSN nodes, amenable to both decentralised or centralised topologies.

In Chapter 4, an adaptive learning model was proposed to address the WSN challenges. This adaptive

classifier model performed extremely well, and along with adaptive feature selection scheme, it could

achieve energy efficiency and prediction accuracy targets, as well as address the resource management

issues. For experimental validation, two different types of large datasets, the forest cover type dataset

and Gas drift type data set was used to emulate a large physical environment instrumented with WSN, with

each attribute/feature from the data set modelling the node/sensor of a WSN set up to monitor or

Chapter 6

108

characterize a complex and large physical environment. With Gas sensor data set, it was showing

consistent results as was for Forest cover type data set. There was a significant performance improvement

with combined adaptive classifier - feature selection model with random forests, random tree, and random

committees, and with best first and greedy search type feature selection techniques. Further, using a

different learning scheme within the adaptive classifier mode - the ensemble learning scheme, it was

possible to pull up combined performance of weak classifiers, such as Naïve Bayes, and J48 decision trees.

This validates the hypothesis that the power machine learning/data mining approaches can indeed address

different WSN challenges, including, operational/functional challenges such as energy efficiency and

event detection accuracy, and non-operational/non-functional challenges such as failure recovery and

resource management.

In Chapter 5, a joint sensor selection - adaptive routing model was proposed for WSN, based on machine

learning with a feature selection algorithm based on Akaiki criterion, that selects few most significant

sensors to be active at a time, and adapts them continuously as time evolves. The experimental evaluation

for a real world publicly available WSN dataset, the Intel Berkeley Lab WSN test bed, validates our

hypothesis, and allows WSN operational and non-operational challenges to be addressed including energy

efficiency, prediction accuracy, and QOS enhancements, such as robustness to sensor failures and quick

MAC layer adaptation ( with fast learning times)

While the proposed integrated framework addressed some of the WSN key challenges were addressed

well. However, there are a myriad of challenges, operational, non-operational and application specific,

which can indeed be addressed by this framework, and can be extended with advanced machine learning

algorithms. Hence, there is still a need for future research in this interdisciplinary area, and some future

directions of this work include, use of emerging techniques that extract spatio-temporal correlations better

as compared to the techniques examined in this work, such as canonical correlation analysis, independent

component analysis, dictionary learning, and non-negative matrix factorization for sensor selection, as they

have proved to be highly efficient in other machine learning application contexts.

Another promising direction for further extending the proposed integrated framework, is to investigate,

some of unsupervised, self- learning and online approaches for addressing WSN challenges, and consider

alternate WSN topologies (instead of just centralised or decentralised WSN topology). Hierarchical

clustering is one such candidate, which uses unsupervised learning and can be deployed in hybrid WSN

topologies (combination of centralised and decentralised topology). Another potential extension of the

proposed framework is, instead of reducing the amount of data in the network, it could be possible to gain

more insight into WSN behaviour from the abundant data available, with some of recent big data analytics,

Chapter 6

109

and scalable machine learning approaches, and devise solutions to achieve energy efficiency, event

detection accuracy and QoS targets. Some of these aspects can be investigated in future.

Bibliography

111

Bibliography

1. Dargie, W. and C. Poellabauer, Fundamentals of wireless sensor networks: theory and

practice2010: John Wiley & Sons Inc.

2. Sohraby, K., D. Minoli, and T.F. Znati, Wireless sensor networks: technology, protocols, and

applications2007: Wiley-Blackwell.

3. David J. Stein, E., Wi

4. reless Sensor Network Simulator. 2006(1.1).

5. Chong, C.-Y. and S.P. Kumar, Sensor networks: evolution, opportunities, and challenges.

Proceedings of the IEEE, 2003. 91(8): p. 1247-1256.

6. Guo, B., D. Zhang, and M. Imai, Toward a cooperative programming framework for context-

aware applications. Personal and Ubiquitous Computing, 2011. 15(3): p. 221-233.

7. Baqer, M. and A. Khan. Energy-efficient pattern recognition approach for wireless sensor

networks. 2007. IEEE.

8. Nakamura, E.F. and A.A.F. Loureiro, Information fusion in wireless sensor networks, in

Proceedings of the 2008 ACM SIGMOD international conference on Management of data2008,

ACM: Vancouver, Canada. p. 1365-1372.

9. Bashyal, S. and G.K. Venayagamoorthy. Collaborative routing algorithm for wireless sensor

network longevity. 2007. IEEE.

10. Narasimhan, R. and D.C. Cox. A handoff algorithm for wireless systems using pattern

recognition. 1998. IEEE.

11. Song, M. and T. Allison, Frequency Hopping Pattern Recognition Algorithms for Wireless

Sensor Networks.

12. Wälchli, M. and T. Braun, Efficient signal processing and anomaly detection in wireless sensor

networks. Applications of Evolutionary Computing, 2009: p. 81-86.

13. Yu, W. and C. He. Resource reservation in wireless networks based on pattern recognition. 2001.

IEEE.

14. Dziengel, N., G. Wittenburg, and J. Schiller. Towards distributed event detection in wireless

sensor networks. 2008.

15. Schurgers, C. and M.B. Srivastava. Energy efficient routing in wireless sensor networks. 2001.

Ieee.

16. Ma, J., et al. Energy-efficient opportunistic topology control in wireless sensor networks. 2007.

ACM.

17. Ali, M. and Z.A. Uzmi. An energy-efficient node address naming scheme for wireless sensor

networks. 2004. IEEE.

Bibliography

112

18. Viera, M., et al. Scheduling nodes in wireless sensor networks: A Voronoi approach. 2003. IEEE.

19. Salhieh, A., et al. Power efficient topologies for wireless sensor networks. 2001. IEEE.

20. Ganesan, D., et al., Highly-resilient, energy-efficient multipath routing in wireless sensor

networks. ACM SIGMOBILE Mobile Computing and Communications Review, 2001. 5(4): p.

11-25.

21. Zytoune, Q., Y. Fakhri, and D. Aboutajdine, A balanced cost cluster-heads selection algorithm

for wireless sensor networks. International Journal of Computer Science, 2009. 4(1): p. 21-24.

22. Raicu, L., et al. e3D: an energy-efficient routing algorithm for wireless sensor networks. 2004.

Ieee.

23. Abdullah, N.Z.a.A.B., Different Techniques Towards Enhancing Wireless Sensor Network

(WSN) Routing Energy Efficiency and Quality of Service (QoS). World Applied Sciences

Journal 2011. 4.

24. Ping, S., Delay measurement time synchronization for wireless sensor networks. Intel Research

Berkeley Lab, 2003.

25. Asuncion, A. and D. Newman, UCI machine learning repository. University of California, Irvine,

School of Information and Computer Sciences, 2007. URL:< http://www. ics. uci.

edu/mlearn/MLRepository. html, 2010.

26. Alwadi, M.d. and G. Chetty, A novel feature selection scheme for energy efficient wireless

sensor networks, in Algorithms and Architectures for Parallel Processing2012, Springer. p. 264-

273.

27. MATLAB. 20/01/2012]; Available from: http://www.mathworks.com.au/.

28. Hall, M., et al., The WEKA data mining software: an update. ACM SIGKDD Explorations

Newsletter, 2009. 11(1): p. 10-18.

29. ALWADI, M. and G. CHETTY, Feature Selection and Energy Management for Wireless Sensor

Networks. IJCSNS, 2012. 12(6): p. 46.

30. Cortez, P. and A.J.R. Morais, A data mining approach to predict forest fires using meteorological

data. 2007.

31. Csirik, J., P. Bertholet, and H. Bunke. Pattern recognition in wireless sensor networks in presence

of sensor failures. 2011.

32. ALWADI, M. and G. CHETTY, Energy Efficiency Data Mining for Wireless Sensor Networks

Based on Random Forests.

33. Alwadi, M. and G. Chetty, Energy Efficient Data Mining Scheme for Big Data Biodiversity

Environment. 2014.

Bibliography

113

34. Karimi, K. and H.J. Hamilton, Logical Decision Rules: Teaching C4. 5 to Speak Prolog, in

Intelligent Data Engineering and Automated Learning—IDEAL 2000. Data Mining, Financial

Engineering, and Intelligent Agents2000, Springer. p. 85-90.

35. Hastie, T., et al., The elements of statistical learning: data mining, inference and prediction. The

Mathematical Intelligencer, 2005. 27(2): p. 83-85.

36. Alwadi, M. and G. Chetty, Energy Efficient Data Mining Scheme for High Dimensional Data.

Procedia Computer Science, 2015. 46: p. 483-490.

37. Richter, R., Distributed Pattern Recognition in Wireless Sensor Networks.

38. Bodik, P., et al., Intel lab data. Online dataset, 2004.

39. Hall, M.A., Correlation-based feature selection for machine learning, 1999, The University of

Waikato.

40. Akaike, H., Information theory and an extension of the maximum likelihood principle, in

Selected Papers of Hirotugu Akaike1998, Springer. p. 199-213.

41. Ashraf, M., et al., A New Approach for Constructing Missing Features Values. International

Journal of Intelligent Information Processing, 2012. 3(1).

42. T. O. Ayodele, “Introduction to machine learning,” in New Advances in Machine Learning.

InTech, 2010.

43. A. H. Duffy, “The “what” and “how” of learning in design,” IEEE Expert, vol. 12, no. 3, pp. 71–

76, 1997.

44. P. Langley and H. A. Simon, “Applications of machine learning and rule induction,”

Communications of the ACM, vol. 38, no. 11, pp. 54–64, 1995.

45. L. Paradis and Q. Han, “A survey of fault management in wireless sensor networks,” Journal of

Network and Systems Management, vol. 15, no. 2, pp.171–190, 2007.

46. B. Krishnamachari, D. Estrin, and S. Wicker, “The impact of data aggregation in wireless sensor

networks,” in 22nd International Conference on Distributed Computing Systems Workshops,

2002, pp. 575–578.

47. J. Al-Karaki and A. Kamal, “Routing techniques in wireless sensor networks: A survey,” IEEE

Wireless Communications, vol. 11, no. 6, pp. 6–28, 2004.

48. K. Romer and F. Mattern, “The design space of wireless sensor networks,” IEEE Wireless

Communications, vol. 11, no. 6, pp. 54–61, 2004.

49. J. Wan, M. Chen, F. Xia, L. Di, and K. Zhou, “From machine-to-machine communications

towards cyber-physical systems,” Computer Science and Information Systems, vol. 10, pp.

1105–1128, 2013.

50. Y. Bengio, “Learning deep architectures for AI,” Foundations and Trends in Machine Learning,

vol. 2, no. 1, pp. 1–127, 2009.

Bibliography

114

51. A. G. Hoffmann, “General limitations on machine learning,” pp. 345–347, 1990

52. M. Di and E. M. Joo, “A survey of machine learning in wireless sensor netoworks from

networking and application perspectives,” in 6th International Conference on Information,

Communications Signal Processing, 2007, pp. 1–5.

53. A. Forster, “Machine learning techniques applied to wireless ad-hoc networks: Guide and

survey,” in 3rd International Conference on Intelligent Sensors, Sensor Networks and

Information. IEEE, 2007, pp. 365–370.

54. A. Förster and M. Amy L, Machine learning across the WSN layers. InTech, 2011.

55. Y. Zhang, N. Meratnia, and P. Havinga, “Outlier detection techniques for wireless sensor

networks: A survey,” IEEE Communications Surveys & Tutorials, vol. 12, no. 2, pp. 159–170,

2010.

56. V. J. Hodge and J. Austin, “A survey of outlier detection methodologies,” Artificial Intelligence

Review, vol. 22, no. 2, pp. 85–126, 2004.

57. R. Kulkarni, A. Förster, and G. Venayagamoorthy, “Computational intelligence in wireless

sensor networks: A survey,” IEEE Communications Surveys & Tutorials, vol. 13, no. 1, pp. 68–

96, 2011.

58. S. Das, A. Abraham, and B. K. Panigrahi, Computational intelligence: Foundations, perspectives,

and recent trends. John Wiley & Sons, Inc., 2010, pp. 1–37.

59. Y. S. Abu-Mostafa, M. Magdon-Ismail, and H.-T. Lin, Learning from data. AMLBook, 2012.

60. O. Chapelle, B. Schlkopf, and A. Zien, Semi-supervised learning. MIT press Cambridge, 2006,

vol. 2.

61. S. Kulkarni, G. Lugosi, and S. Venkatesh, “Learning pattern classification-a survey,” IEEE

Transactions on Information Theory, vol. 44, no. 6, pp.2178–2206, 1998.

62. M. Morelande, B. Moran, and M. Brazil, “Bayesian node localisation in wireless sensor

networks,” in IEEE International Conference on Acoustics, Speech and Signal Processing, 2008,

pp. 2545–2548.

63. C.-H. Lu and L.-C. Fu, “Robust location-aware activity recognition using wireless sensor

network in an attentive home,” IEEE Transactions on Automation Science and Engineering, vol.

6, no. 4, pp. 598–609, 2009.

64. A. Shareef, Y. Zhu, and M. Musavi, “Localization using neural networks in wireless sensor

networks,” in Proceedings of the 1st International Conference on Mobile Wireless Middleware,

Operating Systems, and Applications, 2008, pp. 1–7.

65. J.Winter, Y. Xu, and W.-C. Lee, “Energy efficient processing of k nearest neighbor queries in

location-aware sensor networks,” in 2nd International Conference on Mobile and Ubiquitous

Systems: Networking and Services. IEEE, 2005, pp. 281–292.

Bibliography

115

66. P. P. Jayaraman, A. Zaslavsky, and J. Delsing, “Intelligent processing of k-nearest neighbors

queries using mobile data collectors in a location aware 3D wireless sensor network,” in Trends

in Applied Intelligent Systems. Springer, 2010, pp. 260–270.

67. L. Yu, N. Wang, and X. Meng, “Real-time forest fire detection with wireless sensor networks,”

in International Conference on Wireless Communications, Networking and Mobile Computing,

vol. 2, 2005, pp. 1214–1217.

68. M. Bahrepour, N. Meratnia, M. Poel, Z. Taghikhaki, and P. J. Havinga, “Distributed event

detection in wireless sensor networks for disaster management,” 2nd International Conference

on Intelligent Networking and Collaborative Systems. IEEE, 2010, pp. 507–512.

69. M. Kim and M.-G. Park, “Bayesian statistical modeling of system energy saving effectiveness

for MAC protocols of wireless sensor networks,” in Software Engineering, Artificial

Intelligence, Networking and Parallel/Distributed Computing, ser. Studies in Computational

Intelligence. Springer Berlin Heidelberg, 2009, vol. 209, pp. 233–245.

70. Y.-J. Shen and M.-S. Wang, “Broadcast scheduling in wireless sensor networks using fuzz

hopfield neural network,” Expert Systems with Applications, vol. 34, no. 2, pp. 900 – 907, 2008.

71. R. V. Kulkarni and G. K. Venayagamoorthy, “Neural network based secure media access control

protocol for wireless sensor networks,” in Proceedings of the 2009 International Joint Conference

on Neural Networks, ser. IJCNN’09. Piscataway, NJ, USA: IEEE Press, 2009, pp. 3437–3444.

72. D. Janakiram, V. Adi Mallikarjuna Reddy, and A. Phani Kumar, “Outlier detection in wireless

sensor networks using Bayesian belief networks,” in 1st International Conference on

Communication System Software and Middleware. IEEE, 2006, pp. 1–6.

73. W. Branch, C. Giannella, B. Szymanski, R. Wolff, and H. Kargupta, “In-network outlier

detection in wireless sensor networks,” Knowledge and information systems, vol. 34, no. 1, pp.

23–54, 2013.

74. S. Kaplantzis, A. Shilton, N. Mani, and Y. Sekercioglu, “Detecting selective forwarding attacks

in wireless sensor networks using support vector machines,” in 3rd International Conference on

Intelligent Sensors, Sensor Networks and Information. IEEE, 2007, pp. 335–340.

75. S. Rajasegarar, C. Leckie, M. Palaniswami, and J. Bezdek, “Quarter sphere based distributed

anomaly detection in wireless sensor networks,” in International Conference on

Communications, 2007, pp. 3864–3869.

76. A. Snow, P. Rastogi, and G. Weckman, “Assessing dependability of wireless networks using

neural networks,” in Military Communications Conference. IEEE, 2005, pp. 2809–2815 Vol. 5.

77. A. Moustapha and R. Selmic, “Wireless sensor network modeling using modified recurrent

neural networks: Application to fault detection,” IEEE Transactions on Instrumentation and

Measurement, vol. 57, no. 5, pp. 981–988, 2008.

Bibliography

116

78. Y. Wang, M. Martonosi, and L.-S. Peh, “Predicting link quality using supervised learning in

wireless sensor networks,” ACM SIGMOBILE Mobile Computing and Communications

Review, vol. 11, no. 3, pp. 71–83, 2007.

79. K. Beyer, J. Goldstein, R. Ramakrishnan, and U. Shaft, “When is “nearest neighbor”

meaningful?” in Database Theory. Springer, 1999, pp. 217–235.

80. T. O. Ayodele, “Types of machine learning algorithms,” in New Advances in Machine Learning.

InTech, 2010.

81. S. R. Safavian and D. Landgrebe, “A survey of decision tree classifier methodology,” IEEE

Transactions on Systems, Man and Cybernetics, vol. 21, no. 3, pp. 660–674, 1991.

82. R. Lippmann, “An introduction to computing with neural nets,” ASSP Magazine, IEEE, vol. 4,

no. 2, pp. 4–22, 1987.

83. W. Dargie and C. Poellabauer, Localization. John Wiley & Sons, Ltd, 2010, pp. 249–266.

84. T. Kohonen, Self-organizing maps, ser. Springer Series in Information Sciences. Springer Berlin

Heidelberg, 2001, vol. 30.

85. G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural

networks,” Science, vol. 313, no. 5786, pp. 504–507, 2006.

86. I. Steinwart and A. Christmann, Support vector machines. Springer, 2008.

87. Z. Yang, N. Meratnia, and P. Havinga, “An online outlier detection technique for wireless sensor

networks using unsupervised quarter-sphere support vector machine,” in International

Conference on Intelligent Sensors, Sensor Networks and Information Processing. IEEE, 2008,

pp. 151–156.

88. Y. Chen, Y. Qin, Y. Xiang, J. Zhong, and X. Jiao, “Intrusion detection system based on immune

algorithm and support vector machine in wireless sensor network,” in Information and

Automation, ser. Communications in Computer and Information Science. Springer Berlin

Heidelberg, 2011, vol. 86, pp. 372–376.

89. Y. Zhang, N. Meratnia, and P. J. Havinga, “Distributed online outlier detection in wireless sensor

networks using ellipsoidal support vector machine,” Ad Hoc Networks, vol. 11, no. 3, pp. 1062–

1074, 2013.

90. W. Kim, J. Park, and H. Kim, “Target localization using ensemble support vector regression in

wireless sensor networks,” in Wireless Communications and Networking Conference, 2010, pp.

1–5.

91. D. Tran and T. Nguyen, “Localization in wireless sensor networks based on support vector

machines,” IEEE Transactions on Parallel and Distributed Systems, vol. 19, no. 7, pp. 981–994,

2008.

Bibliography

117

92. B. Yang, J. Yang, J. Xu, and D. Yang, “Area localization algorithm for mobile nodes in wireless

sensor networks based on support vector machines,” in Mobile Ad-Hoc and Sensor Networks.

Springer, 2007, pp. 561–571.

93. G. E. Box and G. C. Tiao, Bayesian inference in statistical analysis. John Wiley & Sons, 2011,

vol. 40.

94. C. E. Rasmussen, “Gaussian processes for machine learning,” in in: Adaptive Computation and

Machine Learning. Citeseer, 2006.

95. S. Lee and T. Chung, “Data aggregation for wireless sensor networks using self-organizing map,”

in Artificial Intelligence and Simulation, ser. Lecture Notes in Computer Science. Springer

Berlin Heidelberg, 2005, vol. 3397, pp. 508–517.

96. R. Masiero, G. Quer, D. Munaretto, M. Rossi, J. Widmer, and M. Zorzi, “Data acquisition

through joint compressive sensing and principal component analysis,” in Global

Telecommunications Conference. IEEE, 2009, pp. 1–6.

97. R. Masiero, G. Quer, M. Rossi, and M. Zorzi, “A Bayesian analysis of compressive sensing data

recovery in wireless sensor networks,” in International Conference on Ultra Modern

Telecommunications Workshops, 2009, pp. 1–6.

98. A. Rooshenas, H. Rabiee, A. Movaghar, and M. Naderi, “Reducing the data transmission in

wireless sensor networks using the principal component analysis,” in 6th International

Conference on Intelligent Sensors, Sensor Networks and Information Processing. IEEE, 2010,

pp. 133–138.

99. S. Macua, P. Belanovic, and S. Zazo, “Consensus-based distributed principal component analysis

in wireless sensor networks,” in 11th International Workshop on Signal Processing Advances in

Wireless Communications, 2010, pp. 1–5.

100. Y.-C. Tseng, Y.-C. Wang, K.-Y. Cheng, and Y.-Y. Hsieh, “iMouse: An integrated mobile

surveillance and wireless sensor system,” Computer, vol. 40, no. 6, pp. 60–66, 2007.

101. D. Li, K. Wong, Y. H. Hu, and A. Sayeed, “Detection, classification, and tracking of targets,”

IEEE Signal Processing Magazine, vol. 19, no. 2, pp. 17–29, 2002.

102. T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman, and A. Y. Wu, “An

efficient k-means clustering algorithm: Analysis and implementation,” IEEE Transactions on

Pattern Analysis and Machine Intelligence, vol. 24, no. 7, pp. 881–892, 2002.

103. I. T. Jolliffe, Principal component analysis. Springer verlag, 2002.

104. D. Feldman, M. Schmidt, C. Sohler, D. Feldman, M. Schmidt, and C. Sohler, “Turning big data

into tiny data: Constant-size coresets for k-means, PCA and projective clustering,” in SODA,

2013, pp. 1434–1453.

105. C. Watkins and P. Dayan, “Q-learning,” Machine Learning, vol. 8, no. 3-4, pp. 279–292, 1992.

Bibliography

118

106. R. Sun, S. Tatsumi, and G. Zhao, “Q-MAP: A novel multicast routing method in wireless ad hoc

networks with multiagent reinforcement learning,” in Region 10 Conference on Computers,

Communications, Control and Power Engineering, vol. 1, 2002, pp. 667–670 vol.1.

107. S. Dong, P. Agrawal, and K. Sivalingam, “Reinforcement learning based geographic routing

protocol for UWB wireless sensor network,” in Global Telecommunications Conference. IEEE,

2007, pp. 652–656.

108. A. Förster and A. Murphy, “FROMS: Feedback routing for optimizing multiple sinks in wsn

with reinforcement learning,” in 3rd International Conference on Intelligent Sensors, Sensor

Networks and Information. IEEE, 2007, pp. 371–376.

109. R. Arroyo-Valles, R. Alaiz-Rodriguez, A. Guerrero-Curieses, and J. Cid-Sueiro, “Q-probabilistic

routing in wireless sensor networks,” in 3rd International Conference on Intelligent Sensors,

Sensor Networks and Information. IEEE, 2007, pp. 1–6.

110. C. Guestrin, P. Bodik, R. Thibaux, M. Paskin, and S. Madden, “Distributed regression: An

efficient framework for modeling sensor network data,” in 3rd International Symposium on

Information Processing in Sensor Networks, 2004, pp. 1–10.

111. J. Barbancho, C. León, F. Molina, and A. Barbancho, “A new QoS routing algorithm based on

self-organizing maps for wireless sensor networks,” Telecommunication Systems, vol. 36, pp.

73–83, 2007.

112. B. Scholkopf and A. J. Smola, Learning with kernels: Support vector machines, regularization,

optimization, and beyond. Cambridge, MA, USA: MIT Press, 2001.

113. J. Kivinen, A. Smola, and R. Williamson, “Online learning with kernels,” IEEE Transactions on

Signal Processing, vol. 52, no. 8, pp. 2165–2176, 2004.

114. G. Aiello and G. Rogerson, “Ultra-wideband wireless systems,” IEEE Microwave Magazine,

vol. 4, no. 2, pp. 36–47, 2003.

115. R. Rajagopalan and P. Varshney, “Data-aggregation techniques in sensor networks: A survey,”

IEEE Communications Surveys & Tutorials, vol. 8, no. 4, pp. 48–63, 2006.

116. G. Crosby, N. Pissinou, and J. Gadze, “A framework for trust-based cluster head election in

wireless sensor networks,” in 2nd IEEE Workshop on Dependability and Security in Sensor

Networks and Systems, 2006, pp. 10–22.

117. J.-M. Kim, S.-H. Park, Y.-J. Han, and T.-M. Chung, “CHEF: Cluster head election mechanism

using fuzzy logic in wireless sensor networks,” in 10th International Conference on Advanced

Communication Technology, vol. 1. IEEE, 2008, pp. 654–659.

118. S. Soro and W. Heinzelman, “Prolonging the lifetime of wireless sensor networks via unequal

clustering,” in 19th IEEE International Parallel and Distributed Processing Symposium, 2005,

pp. 4–8.

Bibliography

119

119. A. A. Abbasi and M. Younis, “A survey on clustering algorithms for wireless sensor networks,”

Computer communications, vol. 30, no. 14, pp. 2826–2841, 2007.

120. H. He, Z. Zhu, and E. Makinen, “A neural network model to minimize the connected dominating

set for self-configuration of wireless sensor networks,” IEEE Transactions on Neural Networks,

vol. 20, no. 6, pp. 973–982, 2009.

121. G. Ahmed, N. M. Khan, Z. Khalid, and R. Ramer, “Cluster head selection using decision trees

for wireless sensor networks,” in International Conference on Intelligent Sensors, Sensor

Networks and Information Processing. IEEE, 2008, pp. 173–178.

122. E. Ertin, “Gaussian process models for censored sensor readings,” in 14th Workshop on

Statistical Signal Processing. IEEE, 2007, pp. 665–669.

123. J. Kho, A. Rogers, and N. R. Jennings, “Decentralized control of adaptive sampling in wireless

sensor networks,” ACM Transactions on Sensor Networks (TOSN), vol. 5, no. 3, pp. 19:1–19:35,

2009.

124. S. Lin, V. Kalogeraki, D. Gunopulos, and S. Lonardi, “Online information compression in sensor

networks,” in IEEE International Conference on Communications, vol. 7. IEEE, 2006, pp. 3371–

3376.

125. C. Fenxiong, L. Mingming, W. Dianhong, and T. Bo, “Data compression through principal

component analysis over wireless sensor networks,” Journal of Computational Information

Systems, vol. 9, no. 5, pp. 1809–1816, 2013.

126. A. Förster and A. Murphy, “CLIQUE: Role-free clustering with q-learning for wireless sensor

networks,” in 29th IEEE International Conference on Distributed Computing Systems, 2009, pp.

441–449.

127. M. Mihaylov, K. Tuyls, and A. Nowe, “Decentralized learning in wireless sensor networks,” in

Adaptive and Learning Agents, ser. Lecture Notes in Computer Science. Springer Berlin

Heidelberg, 2010, vol. 5924, pp. 60–73.

128. W. B. Heinzelman, “Application-specific protocol architectures for wireless networks,” Ph.D.

dissertation, Massachusetts Institute of Technology, 2000.

129. M. Duarte and Y. Eldar, “Structured compressed sensing: From theory to applications,” IEEE

Transactions on Signal Processing, vol. 59, no. 9, pp. 4053–4085, 2011

130. A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood from incomplete data via

the EM algorithm,” Journal of the Royal Statistical Society. Series B (Methodological), pp. 1–

38, 1977.

131. M. H. DeGroot, “Reaching a consensus,” Journal of the American Statistical Association, vol.

69, no. 345, pp. 118–121, 1974.

Bibliography

120

132. B. Krishnamachari and S. Iyengar, “Distributed bayesian algorithms for fault-tolerant event

region detection in wireless sensor networks,” IEEE Transactions on Computers, vol. 53, no. 3,

pp. 241–250, 2004.

133. P. Zappi, C. Lombriser, T. Stiefmeier, E. Farella, D. Roggen, L. Benini, and G. Tröster, “Activity

recognition from on-body sensors: Accuracy-power trade-off by dynamic sensor selection,” in

Wireless Sensor Networks. Springer, 2008, pp. 17–33.

134. H. Malik, A. Malik, and C. Roy, “A methodology to optimize query in wireless sensor networks

using historical data,” Journal of Ambient Intelligence and Humanized Computing, vol. 2, pp.

227–238, 2011.

135. Q. Chen, K.-Y. Lam, and P. Fan, “Comments on "Distributed Bayesian algorithms for fault-

tolerant event region detection in wireless sensor networks",” IEEE Transactions on Computers,

vol. 54, no. 9, pp. 1182–1183, 2005.

136. K. Sha, W. Shi, and O. Watkins, “Using wireless sensor networks for fire rescue applications:

Requirements and challenges,” in IEEE International Conference on Electro/information

Technology, 2006, pp. 239–244.

137. H. Liu, H. Darabi, P. Banerjee, and J. Liu, “Survey of wireless indoor positioning techniques and

systems,” IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and

Reviews, vol. 37, no. 6, pp. 1067–1080, 2007.

138. Wang, R. Ghosh, and S. Das, “A survey on sensor localization,” Journal of Control Theory and

Applications, vol. 8, no. 1, pp. 2–11, 2010.

139. A. Nasipuri and K. Li, “A directionality based location discovery scheme for wireless sensor

networks,” in Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks

and Applications. ACM, 2002, pp. 105–111.

140. S. Yun, J. Lee, W. Chung, E. Kim, and S. Kim, “A soft computing approach to localization in

wireless sensor networks,” Expert Systems with Applications, vol. 36, no. 4, pp. 7552–7561,

2009.

141. S. Chagas, J. Martins, and L. de Oliveira, “An approach to localization scheme of wireless sensor

networks based on artificial neural networks and genetic algorithms,” in 10th International

Conference on New Circuits and Systems. IEEE, 2012, pp. 137–140.

142. Z. Merhi, M. Elgamel, and M. Bayoumi, “A lightweight collaborative fault tolerant target

localization system for wireless sensor networks,” IEEE Transactions on Mobile Computing,

vol. 8, no. 12, pp. 1690–1704, 2009.

143. E. Cayirci, H. Tezcan, Y. Dogan, and V. Coskun, “Wireless sensor networks for underwater

survelliance systems,” Ad Hoc Networks, vol. 4, no. 4, pp. 431–446, 2006.

Bibliography

121

144. A. Krause, A. Singh, and C. Guestrin, “Near-optimal sensor placements in gaussian processes:

Theory, efficient algorithms and empirical studies,” The Journal of Machine Learning Research,

vol. 9, pp. 235–284, 2008.

145. D. Gu and H. Hu, “Spatial Gaussian process regression with mobile sensor networks,” IEEE

Transactions on Neural Networks and Learning Systems, vol. 23, no. 8, pp. 1279–1290, 2012.

146. L. Paladina, M. Paone, G. Iellamo, and A. Puliafito, “Self organizing maps for distributed

localization in wireless sensor networks,” in 12th IEEE Symposium on Computers and

Communications, 2007, pp. 1113–1118.

147. G. Giorgetti, S. K. S. Gupta, and G. Manes, “Wireless localization using self-organizing maps,”

in Proceedings of the 6th International Conference on Information Processing in Sensor

Networks, ser. IPSN ’07. New York, NY, USA: ACM, 2007, pp. 293–302.

148. Hu and G. Lee, “Distributed localization of wireless sensor networks using self-organizing

maps,” in IEEE International Conference on Multisensor Fusion and Integration for Intelligent

Systems, 2008, pp. 284–289.

149. S. Li, X. Kong, and D. Lowe, “Dynamic path determination of mobile beacons employing

reinforcement learning for wireless sensor localization,” in 26th International Conference on

Advanced Information Networking and Applications Workshops, 2012, pp. 760–765.

150. C. Musso, N. Oudjane, and F. Le Gland, “Improving regularised particle filters,” in Sequential

Monte Carlo methods in practice. Springer, 2001, pp. 247–271.

151. Y.-X. Wang and Y.-J. Zhang, “Non-negative matrix factorization: A comprehensive review,”

IEEE Transactions on Knowledge and Data Engineering, vol. 25, no. 6, pp. 1336–1353, 2013.

152. H.-P. Tan, R. Diamant, W. K. Seah, and M. Waldmeyer, “A survey of techniques and challenges

in underwater localization,” Ocean Engineering, vol. 38, no. 14, pp. 1663–1676, 2011.

153. Y. Chu, P. Mitchell, and D. Grace, “ALOHA and q-learning based medium access control for

wireless sensor networks,” in International Symposium on Wireless Communication Systems,

2012, pp. 511–515.

154. A. Bachir, M. Dohler, T. Watteyne, and K. K. Leung, “MAC essentials for wireless sensor

networks,” IEEE Communications Surveys & Tutorials, vol. 12, no. 2, pp. 222–248, 2010.

155. Z. Liu and I. Elhanany, “RL-MAC: A reinforcement learning based MAC protocol for wireless

sensor networks,” International Journal of Sensor Networks, vol. 1, no. 3, pp. 117–124, 2006.

156. M. Sha, R. Dor, G. Hackmann, C. Lu, T.-S. Kim, and T. Park, “Self-adapting MAC layer for

wireless sensor networks,” Technical Report WUCSE- 2013-75, Washington University in St.

Louis, Tech. Rep., 2013.

157. W. Ye, J. Heidemann, and D. Estrin, “An energy-efficient MAC protocol for wireless sensor

networks,” in 21st Annual Joint Conference of the IEEE Computer and Communications

Bibliography

122

Societies, vol. 3, 2002, pp. 1567–1576 vol.3. T. van Dam and K. Langendoen, “An adaptive

energy-efficient MAC protocol for wireless sensor networks,” in Proceedings of the 1st

International Conference on Embedded Networked Sensor Systems, ser. SenSys ’03. New York,

NY, USA: ACM, 2003, pp. 171–180.

158. K. Klues, G. Hackmann, O. Chipara, and C. Lu, “A component-based architecture for power-

efficient media access control in wireless sensor networks,” in Proceedings of the 5th

International Conference on Embedded Networked Sensor Systems. ACM, 2007, pp. 59–72.

159. C. Doerr, M. Neufeld, J. Fifield, T. Weingart, D. C. Sicker, and D. Grunwald, “MultiMAC-an

adaptive MAC framework for dynamic radio networking,” in International Symposium on New

Frontiers in Dynamic Spectrum Access Networks. IEEE, 2005, pp. 548–555.

160. D. Moss and P. Levis, “BoX-MACs: Exploiting physical and link layer boundaries in low-power

networking,” Computer Systems Laboratory Stanford University, 2008.

161. Y. Sun, O. Gurewitz, and D. B. Johnson, “RI-MAC: A receiver-initiated asynchronous duty cycle

MAC protocol for dynamic traffic loads in wireless sensor networks,” in Proceedings of the 6th

ACM Conference on Embedded Network Sensor Systems. ACM, 2008, pp. 1–14.

162. Z. Alliance, “Zigbee-2007 specification,” Online:

http://www.zigbee.org/Specifications/ZigBee/Overview.aspx, 2007

163. T. Avram, S. Oh, and S. Hariri, “Analyzing attacks in wireless ad hoc network with self-

organizing maps,” in 5th Annual Conference on Communication Networks and Services

Research, 2007, pp. 166–175. L. N. De Castro and J. Timmis, Artificial immune systems: A new

computational intelligence approach. Springer, 2002.

164. G. J. Pottie and A. Pandya, Quality of service in wireless sensor networks. John Wiley & Sons,

Inc., 2008, pp. 401–435.

165. D. Chen and P. K. Varshney, “QoS support in wireless sensor networks: A survey,” in

International Conference on Wireless Networks, vol. 233, 2004.

166. M. A. Osborne, S. J. Roberts, A. Rogers, S. D. Ramchurn, and N. R. Jennings, “Towards real-

time information processing of sensor network data using computationally efficient multi-output

Gaussian processes,” in Proceedings of the 7th International Conference on Information

Processing in Sensor Networks. IEEE Computer Society, 2008, pp. 109–120.

167. N. Ouferhat and A. Mellouk, “A QoS scheduler packets for wireless sensor networks,” in

International Conference on Computer Systems and Applications, 2007, pp. 211–216.

168. M. Seah, C.-K. Tham, V. Srinivasan, and A. Xin, “Achieving coverage through distributed

reinforcement learning in wireless sensor networks,” in 3rd International Conference on

Intelligent Sensors, Sensor Networks and Information. IEEE, 2007, pp. 425–430.

Bibliography

123

169. R. Hsu, C.-T. Liu, K.-C. Wang, and W.-M. Lee, “QoS-aware power management for energy

harvesting wireless sensor network utilizing reinforcement learning,” in International

Conference on Computational Science and Engineering, vol. 2. IEEE, 2009, pp. 537–542.

170. X. Liang, M. Chen, Y. Xiao, I. Balasingham, and V. C. M. Leung, “A novel cooperative

communication protocol for QoS provisioning in wireless sensor networks,” in 5th International

Conference on Testbeds and Research Infrastructures for the Development of Networks

Communities and Workshops, 2009, pp. 1–6.

171. N. Baccour, A. Koubaa, L. Mottola, M. A. Zuniga, H. Youssef, C. A. Boano, and M. Alves,

“Radio link quality estimation in wireless sensor networks: A survey,” ACM Transactions on

Sensor Networks (TOSN), vol. 8, no. 4, p. 34, 2012.

172. A. Woo, T. Tong, and D. Culler, “Taming the underlying challenges of reliable multihop routing

in sensor networks,” in Proceedings of the 1st International Conference on Embedded Networked

Sensor Systems, ser. SenSys ’03. New York, NY, USA: ACM, 2003, pp. 14–27.

173. K. Shah and M. Kumar, “Distributed independent reinforcement learning (DIRL) approach to

resource management in wireless sensor networks,” in Internatonal Conference on Mobile Adhoc

and Sensor Systems, 2007, pp. 1–9.

174. A. Mainwaring, D. Culler, J. Polastre, R. Szewczyk, and J. Anderson, “Wireless sensor networks

for habitat monitoring,” in Proceedings of the 1st ACM International Workshop on Wireless

Sensor Networks and Applications. ACM, 2002, pp. 88–97.

175. Wiess and Indurkya, Predictive data mining, a [ractival guide, Morgan Kaufmann Publishers.

ISBN: 1-55860-403-0.