ALMA MATER STUDIORUM

UNIVERSITÀ DI BOLOGNA

Dottorato di Ricerca in Monitoraggio e Gestione delle Strutture e

dell’Ambiente (SEHM2 )
Dipartimento di Ingegneria dell’Eenergia Elettrica e
dell’Informazione “Gulgielmo Marconi” - DEI

MACHINE LEARNING FOR

STRUCTURAL MONITORING AND
ANOMALY DETECTION

Settore Concorsuale: 09/F2 - Telecomunicazioni

Settore Scientifico Disciplinare: ING-INF/03 - Telecomunicazioni

Presentato da: Supervisore:

ELIA FAVARELLI Prof. Ing.
ANDREA GIORGETTI
Coordinatore Dottorato: Co-supervisore:
Prof. Ing. Prof. Ing.
ALESSANDRO MARZANI ALESSANDRO MARZANI

CICLO XXXIII
ESAME FINALE ANNO 2021
 ALMA MATER STUDIORUM
UNIVERSITY OF BOLOGNA

Ph.D. Programme in Structural and Environmental Health

Monitoring and Management (SEHM2 )
Department of Electrical, Electronic, and Information Engineering
“Gulgielmo Marconi” - DEI

MACHINE LEARNING FOR

STRUCTURAL MONITORING AND
ANOMALY DETECTION

09/F2 - Telecommunications
ING-INF/03 - Telecommunications

Ph.D. Candidate: Advisor:

ELIA FAVARELLI Prof. Eng.
ANDREA GIORGETTI
Ph.D. Coordinator: Co-advisor:
Prof. Eng. Prof. Eng.
ALESSANDRO MARZANI ALESSANDRO MARZANI

CYCLE XXXIII
FINAL EXAM YEAR 2021
 Keywords

Machine Learning
Deep Learning
Neural Network
Anomaly Detection
Structural Health Monitoring
Abstract

Autonomous structural health monitoring (SHM) of a large number of struc-

tures became a topic of paramount importance for maintenance purposes
and safety reasons in the last few decades. Civil infrastructures are the back-
bone of modern society, and the assessment of their conditions is of renowned
importance. This aspect is even more exacerbated because of the existing
systems, such as bridges, that are fast approaching their service life. Since the
replacement of those structures is functionally and economically demanding,
maintenance and retrofitting operations must be planned wisely.
Moreover, edge computing, a key part of the upcoming 6G technologies,
will offer cloud applications whilst providing more resources and reduced
latency. This paradigm is grounded on mobile application-specific compu-
tations between the cloud, the data-producing devices, and the network in-
frastructure components at the edges of wireless and wired networks. The
increasing amount and variety of data generated by users and sensors inter-
connected to the future 6G network requires new strategies to manage several
types of data with highly different characteristics and also requires solutions
to power the wireless network with renewable energies [1–7]. In the last
few years, the rise of Internet of things (IoT) showed us how pervasive and
widespread the applications of electronic devices and intelligent sensors could
be, and now heading towards 6G, the idea of Internet of everything (IoE)
has become anything but visionary [8–13].
As far as bridge monitoring is concerned, some statistics highlight the
relevance of the problem. For example, in Italy, there are almost 2000 bridges
that require special monitoring; in France, 4000 bridges need to be restored,
and 840 are considered to be in critical conditions; in Germany, 800 bridges

vii
viii Abstract

are reputed at risk. According to conservative estimates, at least 1 % of

the bridges are considered deficient in the entire world. Besides, historic
structures and highly populated buildings need special attention in terms of
monitoring.
In this scenario, the adoption of artificial intelligence (AI) and in par-
ticular machine learning (ML) strategies represents a flexible and poten-
tially powerful solution that must be investigated. To manage the big and
widespread amount of data generated by the extensive usage of multiple
types of sensors, several ML techniques can be investigated, with the aim to
perform data fusion and reduce the amount of data. Furthermore, the usage
of anomaly detection techniques to identify potentially critical situations in
infrastructures and buildings represents a topic of particular interest that
still needs a significant investigation effort.
In this research activity, we provide the fundamental guidelines to per-
form automatic damage detection, which combines SHM strategies and ML
algorithms capable of performing anomaly detection on a wide set of struc-
tures. In particular, several algorithms and strategies capable of extracting
relevant features from large amounts of data generated by different types of
sensors are investigated. Finally, to effectively manage such an amount of
data in communication constraints, we obtained some design rules for the
acquisition system for bridge monitoring for the first time.
List of Figures

1.1 An example of wireless networks for structural health moni-

toring (SHM) in an outdoor scenario. . . . . . . . . . . . . . . 2

2.1 Block diagram of a general machine learning (ML) algorithm. 8

2.2 Gradient descent applied to a convex function, red points rep-
resent the iterative evaluation of the error function E(θ) due
to the update of the weight vector θ. . . . . . . . . . . . . . . 11
2.3 Example of polynomial fitting, with underfitting and overfit-
ting phenomena due to different values of φ. . . . . . . . . . . 13
2.4 General neural network (NN) structure; grey neurons repre-
sent the input layer, white neurons the hidden layers, and
yellow neurons the output one. . . . . . . . . . . . . . . . . . . 15
2.5 NN structure for regression, different neuron colours stays for
different activation functions. . . . . . . . . . . . . . . . . . . 17
2.6 NN structure for classification, different neuron colours stands
for different activation functions. . . . . . . . . . . . . . . . . 18
2.7 auto-associative neural network (ANN) structure for dimen-
sionality reduction, different neuron colours stands for differ-
ent activation functions. . . . . . . . . . . . . . . . . . . . . . 19
2.8 Example of principal component analysis (PCA); blue points
represent the starting points in the original feature space, the
projected ones are depicted in red. . . . . . . . . . . . . . . . 20

ix
x LIST OF FIGURES

2.9 Example of kernel principal component analysis (KPCA); on

the left the points in the original feature space are depicted,
on the right the same points are projected in a new feature
space through radial basis function (RBF). . . . . . . . . . . . 22

2.10 Example of Gaussian mixture model (GMM); blue points rep-

resent the training data, the surface represents the Gaussian
mixture of order 2 that fit the data. . . . . . . . . . . . . . . . 23

2.11 Block diagram of the one class classifier neural network (OCCNN)
algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.12 Average accuracy calculated varying α1 and NP . . . . . . . . . 27

2.13 Accuracy as a function of α2 for different distributions. . . . . 28

2.14 At the top some examples of normalised dataset points dis-

tribution. Blue circles denote features corresponding to nor-
mal conditions, red circles denote features corresponding to
anomaly conditions. Estimated boundaries by the OCCNN
algorithm are showed at the bottom. . . . . . . . . . . . . . . 29

3.1 Block diagram of the operational modal analysis (OMA) strat-

egy [14]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3.2 Example of stabilization diagram for the first hour monitoring:

a) through stochastic subspace identification (SSI), b) after
mode selection and clustering. . . . . . . . . . . . . . . . . . . 41
LIST OF FIGURES xi

3.3 a) Residual modes after mode selection over time with one
acquisition per hour (first 200 acquisitions). b) Natural fre-
quencies selected after the initial phase of the tracking algo-
rithm. c) First two natural frequencies estimation after the
density-based tracking algorithm. Blue points represent the
residual modes, red and yellow tracks represent, respectively,
the first and second fundamental frequency estimated after the
tracking algorithm; vertical dashed lines highlight the period
when the average measured temperatures are below 0 ◦ C [15],
in particular, it has been demonstrated in [16] that when the
temperature goes below 0 ◦ C the natural frequencies of the
bridge increase. Blue and green backgrounds highlight the
acquisitions made during the normal condition of the bridge,
used respectively as training and test sets, while red back-
ground stands for damaged condition acquisitions used in the
test phase. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4.1 Data acquisition setup along the Z-24 bridge: the selected
accelerometers, their positions, and the measured acceleration
direction [17]. . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 The block diagram for signal acquisition, processing, feature
extraction, and detection. . . . . . . . . . . . . . . . . . . . . 50
4.3 Error function evolution over the epochs during training. . . . 54
4.4 Comparison of the classification algorithms in terms of F1
score, recall, precision, and accuracy. . . . . . . . . . . . . . . 56
4.5 F1 score varying the number of points, Nx , used for training,
with Ny = 854 and Nu = 854. . . . . . . . . . . . . . . . . . . 57
4.6 F1 score varying the number of points during damage, Nu ,
used to test the algorithms, with Nx = 2399 and Ny = 854. . . 58

5.1 F1 score varying the considered fundamental frequencies, ver-

tical dashed red lines indicate the best configuration with 4,
3, and 2 frequencies. . . . . . . . . . . . . . . . . . . . . . . . 61
xii LIST OF FIGURES

5.2 Modal frequencies variance for different points; vertical dashed

red lines indicate the minimum number of points for the cor-
rect frequency sorting. . . . . . . . . . . . . . . . . . . . . . . 62
5.3 Modal frequencies skewness varying the number of points; ver-
tical dashed red lines indicate the minimum number of points
for the correct frequency sorting. . . . . . . . . . . . . . . . . 63
5.4 Modal frequencies entropy varying the number of points. . . . 64
5.5 F1 score varying the dimensionality of the feature space, dashed
lines correspond to frequency extraction technique, and con-
tinuous lines are referred to as the frequency extraction ap-
proach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.6 Comparison of the classification algorithms in terms of F1
score, recall, precision, and accuracy with the best feature
configuration. . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

6.1 Examples of feature transformation due to the effect of low

number of sensors, low number of bits, and low number of
samples with respect to the standard measurement condition
reported on the left. . . . . . . . . . . . . . . . . . . . . . . . 68
6.2 Error produced removing the selected sensor. . . . . . . . . . . 69
6.3 Error varying the number of sensors available. . . . . . . . . . 70
6.4 Error varying the number of samples. . . . . . . . . . . . . . . 71
6.5 Error varying the number of bits. . . . . . . . . . . . . . . . . 72

7.1 Illustration of the processing data chain to extract features

from the geophone signals and perform classification. . . . . . 77
7.2 Accuracy varying PCA components of the vertical and hori-
zontal channels, with Tob = 20 s. . . . . . . . . . . . . . . . . . 82
7.3 Accuracy varying the observation window duration, Tob . For
PCA-based classifiers Dh = Dv = 5. . . . . . . . . . . . . . . . 84

8.1 The four scenarios considered with the anomaly represented

by a human being. . . . . . . . . . . . . . . . . . . . . . . . . 89
LIST OF FIGURES xiii

8.2 Illustration of the pre-processing data chain to extract features

from beacon packets. . . . . . . . . . . . . . . . . . . . . . . . 92
8.3 Examples of PCA outputs, in the four scenarios, when P = 2
principal components are selected. Blue circles refer to empty
room points while red circles refer to the presence of a target
in the room. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8.4 Accuracy in scenarios (a) and (b), varying M . . . . . . . . . . 99
8.5 Accuracy in scenarios (c) and (d), varying M . . . . . . . . . . 100
8.6 Accuracy of PCA, KPCA, and received signal strength (RSS)-
based method varying M , in the four scenarios. . . . . . . . . 101
xiv LIST OF FIGURES
List of Acronyms

ADC analog to digital converter

AI artificial intelligence

ANN auto-associative neural network

AP access point

ARMA auto-regressive moving average

BFD basic frequency domain

CSI channel-state information

DNN deep neural network

EVD eigenvalues decomposition

FA false alarm

FFT fast Fourier transform

FIR finite impulse response

GMM Gaussian mixture model

IoT Internet of things

IoE Internet of everything

ISM industrial, scientific and medical

xv
xvi List of Acronyms

LDA linear discriminant analysis

LOS line-of-sight

k-NN k-nearest neighbours

KPCA kernel principal component analysis

MAC modal assurance criterion

ML machine learning

MP mean phase

MPD mean phase deviation

MSE mean square error

NLL negative log likelihood

NLOS non-line-of-sight

NN neural network

OCC one class classifier

OCCNN one class classifier neural network

ODS operational deflection shape

OMA operational modal analysis

PCA principal component analysis

PSD power spectral density

RBF radial basis function

RF radio-frequency

ReLU rectified linear unit

RMSE root mean square error

 xvii

RSS received signal strength

SDR software defined radio

SHM structural health monitoring

SoOp signals of opportunity

SSE sum of squares error

SSI stochastic subspace identification

SSID service set identifier

SVD singular value decomposition

SVM support vector machine

ToA time of arrival

WOSA weighted overlapped segment averaging

xviii List of Acronyms
Contents

Abstract vii

List of Figures xiii

List of Acronyms xv

1 Introduction 1
1.1 Scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Main Contributions . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Thesis Structure and Notation . . . . . . . . . . . . . . . . . . 4

2 Machine Learning Algorithms 7

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.2 Minimization Strategies . . . . . . . . . . . . . . . . . 10
2.1.3 Example: Polynomial curve fitting . . . . . . . . . . . 12
2.2 Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.1 Regression . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.2 Classification . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.3 Dimensionality Reduction . . . . . . . . . . . . . . . . 18
2.3 Anomaly Detection . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.1 Principal component analysis . . . . . . . . . . . . . . 20
2.3.2 Kernel principal component analysis . . . . . . . . . . 21
2.3.3 Gaussian mixture model . . . . . . . . . . . . . . . . . 22
2.4 OCCNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

xix
xx CONTENTS

2.4.1 Density estimator . . . . . . . . . . . . . . . . . . . . . 24

2.4.2 Adversarial points generator . . . . . . . . . . . . . . . 25
2.4.3 Network structure . . . . . . . . . . . . . . . . . . . . . 26
2.4.4 Dimensioning . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.5 OCCNN2 . . . . . . . . . . . . . . . . . . . . . . . . . 29

3 Structural Health Monitoring 31

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Operational Modal Analysis . . . . . . . . . . . . . . . . . . . 32
3.2.1 Peak-Picking . . . . . . . . . . . . . . . . . . . . . . . 35
3.2.2 Auto-regressive moving average . . . . . . . . . . . . . 37
3.2.3 Stochastic subspace identification . . . . . . . . . . . . 38
3.3 Mode selection . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4 Features Extraction . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4.1 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.4.2 Density-based mode tracking . . . . . . . . . . . . . . . 44

4 Z-24 Bridge 47
4.1 System Configuration and Data Collection . . . . . . . . . . . 47
4.1.1 Data collection . . . . . . . . . . . . . . . . . . . . . . 48
4.1.2 Data pre-processing . . . . . . . . . . . . . . . . . . . . 50
4.2 Algorithmic Complexity and Processing Time . . . . . . . . . 51
4.3 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3.1 Algorithm Comparison . . . . . . . . . . . . . . . . . . 56
4.3.2 Impact of the training set and responsiveness . . . . . 57

5 Dimensionality Reduction 59
5.0.1 Feature extraction . . . . . . . . . . . . . . . . . . . . 59
5.0.2 Feature selection . . . . . . . . . . . . . . . . . . . . . 60
5.1 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.1.1 Frequencies selection . . . . . . . . . . . . . . . . . . . 62
5.1.2 Algorithm Comparison . . . . . . . . . . . . . . . . . . 64
CONTENTS xxi

6 Data Management 67
6.1 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.1.1 Sensors Relevance . . . . . . . . . . . . . . . . . . . . . 69
6.1.2 Number of Sensors . . . . . . . . . . . . . . . . . . . . 70
6.1.3 Number of Samples . . . . . . . . . . . . . . . . . . . . 71
6.1.4 Number of Bits . . . . . . . . . . . . . . . . . . . . . . 72
6.2 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

7 Human Activities Classification Using Biaxial Seismic Sen-

sors 75
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
7.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
7.2.1 PSD estimation . . . . . . . . . . . . . . . . . . . . . . 77
7.2.2 PCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
7.3 Classification Techniques . . . . . . . . . . . . . . . . . . . . . 78
7.3.1 Support vector machine . . . . . . . . . . . . . . . . . 79
7.3.2 k-nearest neighbours (k-NN) . . . . . . . . . . . . . . . 80
7.3.3 Template matching . . . . . . . . . . . . . . . . . . . . 80
7.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.4.1 Accuracy vs. number of PCA components . . . . . . . 82
7.4.2 Accuracy vs. observation duration . . . . . . . . . . . . 83
7.5 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

8 Anomaly Detection Using WiFi Signals of Opportunity 87

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
8.2 System Overview and Problem Setup . . . . . . . . . . . . . . 90
8.2.1 Data pre-processing . . . . . . . . . . . . . . . . . . . . 90
8.2.2 Features extraction . . . . . . . . . . . . . . . . . . . . 92
8.3 Survey of ML Techniques . . . . . . . . . . . . . . . . . . . . . 93
8.3.1 Principal Component Analysis . . . . . . . . . . . . . . 93
8.3.2 Kernel Principal Component Analysis . . . . . . . . . . 94
8.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . 96
8.4.1 AP-sensor link in line-of-sight . . . . . . . . . . . . . . 97
xxii CONTENTS

8.4.2 AP-sensor link in non-line-of-sight . . . . . . . . . . . . 98

8.4.3 Beacon average . . . . . . . . . . . . . . . . . . . . . . 100
8.5 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

9 Conclusion 103

Acknowledgements 125
Chapter 1

Introduction

1.1 Scenario
Nowadays, the widespread adoption of automation and sensors in a broad
set of applications (i.e., industry, telecommunication, structural monitoring,
autonomous driving, crowdsensing) generated the opportunity to access a
variety of data that can be used for new types of services and to increase the
reliability of the existing ones (see Fig. 1.1). Moreover, the increasing number
of devices per km2 that with the 6G aims to become greater than 10 million
leads to new networks paradigms that must be properly designed [18–23],
and represents a tremendous source of data that must be organized and
compressed.
n this scenario, where several new research areas have grown, both in
data analysis and big data management, this thesis focuses on the adoption
of innovative ML strategies applied to the important field of SHM.
The study of methods able to determine the state of health of a structure,
and the localization and quantification of the damage, generated a vast liter-
ature, reviewed in several works [24–28]. Over the years, with the aim to per-
form damage detection on structures, several techniques have been developed
to extract the most significant damage-sensitive features. Such techniques
can be divided into model-free and model-based. In model-free methods, the
only information is gathered by measurements (e.g., acceleration, temper-

1
2 Introduction

Figure 1.1: An example of wireless networks for SHM in an outdoor scenario.

ature, position), while in model-based approaches, information comes from

measurements and prior knowledge of a model of the structure [14]. With the
aim to develop a general strategy that can be applied to a broad set of struc-
tures and situations, this thesis is focused on the investigation of model-free
methods.
Since the whole process is quite complicated and requires fine-tuning of
several parameters, specifically for the structure at hand, recently, some
works put forward the idea to adopt ML techniques to detect changes in
the damage sensitive features [29–32]. However, despite the considerable
amount of literature on SHM, much effort is still necessary to analyze data
automatically and ensure the massive monitoring of a large number of infras-
tructures. Unfortunately, performing automatic SHM is not trivial because
of the variety of structures and their peculiar characteristics, which, again,
tend to require specific settings.
ML represents a comprehensive set of powerful and flexible tools that can
1.2 Main Contributions 3

solve general tasks if adequately used. After the extraction of the most sig-
nificant damage-sensitive features performed by SHM techniques, a wide set
of ML algorithm able to perform anomaly detection on a general dataset will
be presented and extensively discussed. Moreover, a collection of innovative
tools with increased detection capabilities, compared to classical approaches,
will be proposed and meticulously examined.

1.2 Main Contributions

In this thesis, we attempt to provide a method to automatically detect
anomalies in structures starting from vibrational data. The proposed frame-
work starts from a feature extraction phase, performed through OMA, a
widely known SHM strategy, to extract the fundamental frequencies of the
structure, followed by their clustering and tracking in the time domain. The
processed modal frequencies are then used as features for a one class classi-
fier (OCC) to perform anomaly detection. In particular, the main contribu-
tions of this thesis can be summarized as follows:
• We present a complete processing chain capable of monitoring the struc-
tural health of a structure autonomously using model-free system iden-
tification based on accelerometric data.

• After performing modal frequencies clustering in the stabilization di-

agram, we propose a new time-domain tracking algorithm based on
modal frequencies density.

• We investigate strategies to select the best features among all the ex-
tracted ones, comparing feature extraction and feature selection tech-
niques in order to provide the best solution for bridge monitoring.

• We propose a new anomaly detector, namely OCCNN, which finds the

normal class (the boundary of the features space in normal operating
conditions) through a two-step approach.

• We propose a variant of OCCNN, named OCCNN2 , which combines

the two-step approach of OCCNN with an ANN.
4 Introduction

• We applied the proposed framework on data collected from the Z-24

bridge to assess its performance, and we made comparisons with several
anomaly detection techniques such as PCA, KPCA, GMM, and ANN,
in terms of F1 score, accuracy, recall, and precision.

• We test the robustness and responsiveness of the proposed solution to

different sensor configurations.

• We investigate the effect of sensor failures in order to provide the mini-

mum number of sensors necessary to ensure predefined anomaly detec-
tion performance.

• We investigate the minimum observation time and the minimum num-

ber of bits that must be used to achieve the target performance, to
reduce the amount of data stored for anomaly detection.

1.3 Thesis Structure and Notation

The rest of the thesis is organized as follows:

Chapter 2 presents an overview of the ML literature, considering the prob-

lems of regression, classification, dimensionality reduction, and anomaly
detection, and several algorithms able to accomplish these tasks. More-
over, two novel anomaly detectors OCCNN and OCCNN2 are proposed
and extensively analyzed.

Chapter 3 investigates some strategies of SHM, whose aim is to extract

damage sensitive features, with particular focus on the model-free meth-
ods. Among the several possibilities, SSI has been selected as a good
candidate to solve this task because of its characteristics and wide ap-
plicability.

Chapter 4 proposes a connection between ML and SHM, the complete

chain for acquisition, feature extraction, and damage detection is pre-
sented and also tested on a real structure, the Z-24 bridge, a benchmark
widely known in the literature.
1.3 Thesis Structure and Notation 5

Chapter 5 analyzes strategies to select the most reliable frequencies among

all the extracted ones, comparing feature extraction and feature selec-
tion techniques.

Chapter 6 explores the reliability of the monitoring system, and the min-
imum operating conditions necessary to ensure a target performance
of the anomaly detection chain, varying the number of sensors avail-
able, the acquisition time, and the number of bits per sample used to
quantize the measurements.

Chapter 7 proposes a method for passive human activity classification ex-

ploiting ground vibrations observed by a biaxial geophone. The solu-
tion is grounded on the idea that some activities can be better analyzed
by the horizontal channel while others by the vertical one.

Chapter 8 investigates the problem of detecting the presence of a target in

a room to perform intrusion detection, evaluating the power spectral
density (PSD) variation of signals of opportunity via anomaly detection
techniques.

Throughout the thesis, capital boldface letters denote matrices and ten-
sors, lowercase bold letters denote vectors, (·)T stands for transposition, (·)+
indicates the Moore-Penrose pseudoinverse operator, || · ||2 is the `2 -norm of
a vector, || · ||F is the Frobenius norm of a matrix, <{·} and ={·} are the real
and imaginary parts of a complex number, respectively, V{·} is the variance
operator, and 1{a, b} is the indicator function equal to 1 when a = b, and
zero otherwise.
6 Introduction
Chapter 2

Machine Learning Algorithms

2.1 Introduction
Recognising patterns, classifying elements, clusterize data, and perform re-
gression from sampled data, are fundamental task to solve problems in en-
gineering. In the last few decades many efforts have been spent to develop
algorithms able to accomplish these tasks with few hyper-parameters, i.e.,
few parameters to be tuned.
Generally, a ML algorithm workflow is composed by a training phase and
a test phase. During the training phase the algorithm is fed with a training
dataset with the aim to set some parameters θ typical of the considered
algorithm in order to minimize a cost function. After that, a validation
dataset is used to set some hyper-parameters φ of the algorithm and finally
a test dataset is used to evaluate the performance. More in detail, given a
matrix of data D of dimension Nd × D where Nd represents the number of
observations and D stands for the dimensionality of the observation (number
of features), we can define the following subsets:

• Training set X of dimension Nx × D, where Nx is the number of ob-

servations used to train the algorithm and set the parameters θ, with
the aim to minimize an error function E(θ, φ) with a direct solution
(where possible) or with some minimization strategies (i.e., stochastic
descent algorithms and Newton-Raphson method).

7
8 Machine Learning Algorithms

Figure 2.1: Block diagram of a general ML algorithm.

• Validation set V of dimension Nv × D, where Nv is the number of

observations used to set the hyper-parameters φ of the algorithm, in
this case the hyper-parameters are set manually with trial and error
strategies or with some metrics that can be defined depending on the
problem.

• Test set Y of dimension Ny × D, where Ny is the number of obser-

vations used to test the algorithm performance; in this phase neither
the parameters nor the hyper-parameters can be modified, in order to
preserve generalization with respect to the training points.

It can be noted that the dimensionality D is constant among the sets of data
and the total number of observations can be written as Nd = Nx + Nv + Ny .
It is good practice to partition the data with the following proportions:
Nx = 60% Nd , Nv = 20% Nd , Ny = 20% Nd . When the number of obser-
vations Nd is low, some strategies like k-fold validation, cross validation,
and one-vs-rest can be implemented to use a greater number of observations
for the training set [29]. Another important step to apply before using the
data is the normalization. Let us define the offset x̂ as the column vec-
tor containing the row-wise mean of the matrix X, and the rescaling factor
xm = maxn,d |x̄n,d − x̂n |. Before proceeding with the application of ML algo-
rithms, the matrices X, Y and V are centered and normalized subtracting
the offset x̂ row-wise and dividing each entry by the rescaling factor xm , in
this way the training data will fall in the interval [+1, −1]. A target matrix
T(·) is defined for the whole dataset of dimensions Nd × Dout where Dout
stands for the dimensionality of the output; this matrix represents the out-
put of the observations. For instance, in a classification problem, for the
training, the target matrix is actually a vector t(X) of dimension Nx × 1 that
contains the classes to which each training point belong.
2.1 Introduction 9

The general procedure of a ML algorithm can be summarized as shown in

Fig. 2.1, and the goal is to set the parameters θ and the hyper-parameters φ
to make a prediction g(X, θ, φ) of a target matrix T(X) in order to minimize
an error function E(θ, φ).

2.1.1 Metrics
Several metrics can be defined to estimate the performance of a ML algo-
rithm, usually related to the particular application. Generally, the most
common metrics for regression purpose are the following:

• Sum of squares error (SSE)

N
X
SSE(θ, φ) = (g(xn , θ, φ) − T(xn ))2
n=1

• Mean square error (MSE)

N
1 X
M SE(θ, φ) = (g(xn , θ, φ) − T(xn ))2
N n=1

• Root mean square error (RMSE)

Ã
N
1 X
RM SE(θ, φ) = √ (g(xn , θ, φ) − T(xn ))2
N n=1

• Negative log likelihood (NLL)

N
X
N LL(θ, φ) = − g(xn , θ, φ) ln T(xn ).
n=1

In several applications a regularization term is introduced in order to

make the error function convex. This term helps the minimization algorithms
to converge to an optimal solution, usually written in the form λ||θ||22 where
10 Machine Learning Algorithms

λ represents the regularization hyper-parameter that belongs to the hyper-

parameters set φ. For classification purposes we can define a different set of
metrics, more useful to understand the performance of our algorithms:

• Accuracy
TP + TN
Acc =
TP + TN + FP + FN

• Precision
TP
Prec =
TP + FP

• Recall
TP
Rec =
TP + FN

• F1 score
Rec · Prec
F1 = 2 ·
Rec + Prec

where TP , TN , FP , and FN , represent respectively true positive, true negative,

false positive, and false negative predictions. Such indicators are obtained
comparing the actual labels t(X), with those predicted by the ML algo-
rithm g(X, θ, φ). In the case of unbalanced classes, the F1 score represents
a more reliable metric to evaluate the performance with respect to the accu-
racy. These metrics can be defined only for two-class classification problems;
for multi-class classification problems, the accuracy is still valid, defined as
number of correct predictions over the total number of points, and confusion
matrix can be adopted to estimate the correct prediction over the different
classes.

2.1.2 Minimization Strategies

Now that we have defined the most common methods to evaluate the algo-
rithm performance, we are ready to define some generic strategies to set the
weights θ in order to minimize the selected error function. The most com-
mon strategy is the gradient descent. The basic idea with gradient descent
is to build a linear model of the selected error function E(θ) (to simplify the
2.1 Introduction 11

Figure 2.2: Gradient descent applied to a convex function, red points repre-
sent the iterative evaluation of the error function E(θ) due to the update of
the weight vector θ.

notation we neglected φ, because it is not influent in this process), deter-

mine the downward direction on this function, travel a short distance along
this direction, hop back on to the error function, and repeat until conver-
gence. Formally, beginning at an initial configuration of weights θ (0) the
linear model of E(θ) at this point is given precisely by the first order Taylor
series approximation centered at θ (0) :

E(θ)
‹ = E(θ (0) ) + ∇E(θ (0) )T (θ − θ (0) ). (2.1)

Now we take the first step by travelling in the direction in which the
tangent hyper-plane most sharply angles downward (referred to as the steep-
est descent direction). It can be shown that this steepest descent direction
is given precisely as −∇E(θ (0) ). Thus, we descend in the direction of the
negative gradient (hence the name of the algorithm gradient descent) taking
our first step to a point θ (1) where

θ (1) = θ (0) − ρ∇E(θ (0) ). (2.2)

12 Machine Learning Algorithms

The same process can be repeated iteratively; in general the number of iter-
ation represents the number of epochs Ne (if all the dataset is used for each
iteration of the algorithm) and the hyper-parameter ρ represents the learning
rate. For the kth iteration we can write the update rule as follows:

θ (k) = θ (k−1) − ρ∇E(θ (k−1) ). (2.3)

The number of epochs Ne can be substituted with several convergence criteria

based on the error value or the error decreasing rate [30]. Several variations
of this technique have been introduced over the years to reduce the oscillation
problems and increase the velocity of convergence [30]. Furthermore, when
the second derivative of E(θ) can be computed, the Newton-Raphson method
can be implemented to strongly increase the convergence rate [29]. An exam-
ple of gradient descent algorithm applied in a two-dimensional weight vector
space is shown in Fig. 2.2.

2.1.3 Example: Polynomial curve fitting

To clarify the notation, a toy regression problem is presented. For instance,

observe a one-dimensional (D = 1) phenomena described by the following
equation: f (z) = y(z) + n = cos (4πz) + 2 sin (2πz) + n where n represents
a generic additive noise. Let’s now observe a set of Nd acquisitions of the
phenomena that form the observation matrix D. In this example the dimen-
sionality of the problem is equal to one, hence the observation matrix is a
column vector d of dimension Nd × 1. Partitioning this vector in training
set, validation set, and test set as described previously, we obtain respec-
tively the training vector x of dimension Nx × 1, the validation vector v of
dimension Nv × 1, and the test vector y of dimension Ny × 1. In this exam-
ple the target vector corresponds to the collection of samples of the function
output t(d) = f (d), with dimension Nd × 1. An example of training set
for the given problem, with Nx = 60, is reported in Fig 2.3. We can now
define a polynomial model, in order to implement the curve fitting. A generic
2.1 Introduction 13

Model Underfitting

2 2
1 1
0 0
f(z)

f(z)
−1 −1
−2 −2
−3 −3
−4 Training y(z) −4 g(z θ 10)
−1 −0,5 0 0,5 1 −1 −0,5 0 0,5 1
z z
Correct Fitting Overfitting

2 2
1 1
0 0
f(z)

f(z)

−1 −1
−2 −2
−3 −3
−4 g(z θ 15) −4 g(z θ 40)
−1 −0,5 0 0,5 1 −1 −0,5 0 0,5 1
z z

Figure 2.3: Example of polynomial fitting, with underfitting and overfitting

phenomena due to different values of φ.

one-dimensional polynomial model can be defined as follows:

φ
X
2 φ
g(z, θ, φ) = θ0 + θ1 z + θ2 z + · · · + θφ z = θj z j
j=0

where φ represents the model order of the polynomial model, the only hyper-
parameter for this model, and z is a generic point. The values of the coef-
ficients θ can be determined by fitting the polynomial to the training data.
This can be done by minimizing an error function that measures the misfit
between the function g(x, θ, φ), for any given value of θ, and the training set
data point targets t(x). One simple choice of error function, which is widely
used in literature, is given by the sum of the squares of the errors between the
predictions g(xn , θ, φ) for each data point xn and the corresponding target
14 Machine Learning Algorithms

values t(xn ), so that we minimize:

NX
X
E(θ, φ) = (g(xn , θ, φ) − t(xn ))2 .
n=1

It is possible to solve the curve fitting problem by choosing the value of θ

for which E(θ, φ) is as small as possible. Because the error function is a
quadratic function of the coefficients θ, its derivatives with respect to the
coefficients will be linear in the elements of θ, and so the minimization of the
error function has a unique solution, denoted by θ ∗ , which can be found as a
direct solution. The resulting polynomial is given by the function g(x, θ ∗ , φ∗ ).
The optimal φ∗ must be set in order to prevent overfitting and underfitting,
the first one is a situation in witch the error is low over the training set
but high over the test set, whereas in the second case the error is high both
in the training and the test set. An example of overfitting, underfitting,
and optimal configuration is illustrated in Fig 2.3 to clarify these possible
situations. In the example shown, the noise presented is a Gaussian noise
with zero mean and variance equal to 0.1, i.e., n ∼ N (0, 0.1).

2.2 Neural Networks

NNs represent a powerful and flexible tool belonging to the ML algorithms.
This kind of algorithm can be adopted to solve regression, classification, and
dimensionality reduction problems with slightly differences in its structure,
and without particular limitations [29, 31]. The general structure of a NN is
reported in Fig. 2.4. As we can see in the picture the structure provides the
following elements:

• Neurons, represented as circles, each one has an activation function

used to propagate the input of the neuron to the next layer.

• Weights, represented as arrows, used to connect the neurons giving to

them different relevance.

The layers of a neural network can be divided into the following groups:
2.2 Neural Networks 15

Figure 2.4: General NN structure; grey neurons represent the input layer,
white neurons the hidden layers, and yellow neurons the output one.

• Input layer, equal to the number of features of the problem (D) that
are used to feed the network with the dataset.

• Hidden layers, that are used to create non linear function able to follow
the data function.

• Output layer, that is used to communicate the results.

In the next chapter we will see how to set the activation functions, and the
number of neurons and layers for several problems. Generally, in this thesis
the number of hidden layers and the relative number of neurons in each
hidden layer will be represented with the vector h of dimension Nh × 1 where
Nh stands for the number of hidden layers and the elements of h represent
the number of neurons in the relative hidden layer.
The NN provide, as all the ML algorithms, a training phase and an online
phase. During the training phase the network is fed with the training dataset
X with the training target T(X) to set the network weights θ in order to
minimize the selected error function E(θ, φ). To set the hyper-parameters
φ (i.e. λ, ρ, h, and Ne ) several trainings can be performed to maximize the
performance on the validation set V and finally the real performance can be
evaluated on the test set Y. As we can infer from the previous observations,
the training phase can be computationally onerous, and requires a long time
16 Machine Learning Algorithms

to be performed. In order to reduce the computational cost, the gradient

descent algorithm is usually replaced by the backword propagation algorithm
that accomplishes the same gradient descent task but with lower complexity.
The description of this strategy is beyond the scope of this thesis, but detailed
description can be found in [29]. The online phase (after the training phase)
for the NN is faster and this represents another remarkable aspect of this
kind of algorithms that are prone to be used in real time applications after
a time consuming training phase.

2.2.1 Regression
As shown previously, a regression problem is a possible application for ML
algorithms and particularly for NNs. In this kind of problem, the goal is
to reconstruct a curve starting from a set of measurements. This type of
application can be useful for:

• filtering, when the goal is to reduce the noise present on a set of mea-
surements.

• interpolation, when the goal is to reconstruct a curve in order to esti-

mate missing points.

• prediction, when the objective is to predict some values out of the

boundaries of the given points.

The generic structure of a NN able to perform regression is reported

in Fig. 2.5. For this application the number of input neurons (depicted in
blue) is equal to the number of input features D, the neurons present in the
hidden layers (in the example h = [3, 4, 3]) exhibit a non linear activation
function (the most common is the rectified linear unit (ReLU) function, but
many other possible solutions are presented in [29, 30]). In these layers the
activation functions are non-linear in order to catch non-linear dependencies
among the data, and finally the output neurons (depicted in green) equal to
the dimensionality of the output and adopt a linear activation function, able
to follow all the possible excursion of the target t(X). The white neurons
2.2 Neural Networks 17

Figure 2.5: NN structure for regression, different neuron colours stays for
different activation functions.

represent the bias terms, they are not connected to the previous layer and
are able to follow the bias values among the data.

2.2.2 Classification
As said before, a quite similar structure can be used to perform classification
on a given dataset. A classification problem is an application of NN where
the target function t(X) is discrete, and represents the class to which the
points belong. In this case the number of input neurons is still equal to the
number of input features D, the neurons in the hidden layers remain the same
as the previous application, but the output layer has a number of neurons
equal to the number of classes of the problem C. The activation function is
a sigmoid with the following expression:

ez
S(z) = (2.4)
ez + 1

where z represents the input of the considered neuron and S(z) its output.
In this case a sigmoid activation function with excursion between 0 and 1
is preferred to a linear one. That is because the output of each network
can be interpreted as a value related to the probability that the input point
belongs to the class represented by the considered neuron, hence an excursion
18 Machine Learning Algorithms

Figure 2.6: NN structure for classification, different neuron colours stands

for different activation functions.

between 0 and 1 is preferred. To assign an input to the relative class, the

neuron with the highest activation function is selected. An example of NN
structure for classification purpose is shown in Fig. 2.6.

2.2.3 Dimensionality Reduction

The last application considered is the dimensionality reduction performed by
the ANN [31]. This particular NN structure is composed by the following
sections:

• Mapping layers, which consist of one or more hidden layers, with the
number of neurons in each layer decreasing progressively till the last
one, called bottleneck. In the bottleneck the number of neurons is
usually lower than the number of input features;

• Demapping layers, composed of one or more hidden layers where the

number of neurons increases progressively.

The input and output layers have the same dimension of the feature space,
and the labels during the training phase must be set equal to the input
data point. With this structure, the data is mapped in a lower-dimensional
feature space (with dimension equal to the number of neurons present in
2.3 Anomaly Detection 19

Figure 2.7: ANN structure for dimensionality reduction, different neuron

colours stands for different activation functions.

the bottleneck layer), and then reconstructed through the demapping layers
minimizing the error with respect to the input data. After the training phase,
if we feed the network with a new data point, the activation function in the
bottleneck layer outputs a mapped version of the starting data point in a
lower dimensional feature space. Afterwards, we can reconstruct the data
using the demapping layers with a low reconstruction error. An example of
ANN is shown in Fig. 2.7.
Now, focusing on the specific problems we are investigating the following
two questions arises:

Q1: How can we use tools for dimensionality reduction to perform anomaly
detection?

Q2: What are the most effective algorithms able to perform anomaly detec-
tion?

2.3 Anomaly Detection

In this section we review PCA, KPCA, and GMM, which are often im-
plemented for anomaly detection. Usually, the algorithms that implement
dimensionality reduction can also be used as anomaly detectors. One class
20 Machine Learning Algorithms

0
f2

-1

-2

-3
-3 -2 -1 0 1 2 3
f1

Figure 2.8: Example of PCA; blue points represent the starting points in the
original feature space, the projected ones are depicted in red.

classification or anomaly detection is a particular case of the classification

problem, where the training data is present only for one class (the standard
one), the algorithm is trained only with this class, and the goal is to detect
whether a new point belongs to the same class or is an anomaly [33–40]. To
use a dimensionality reduction technique as anomaly detector, it is enough
to evaluate the reconstruction error between the input data and the recon-
structed ones. Then, through a threshold, that can be set in order to ensure
a certain false alarm rate on the training set, evaluate if a new point exper-
iments a reconstruction error greater than the threshold: in this case it is
considered as anomalous, otherwise it is considered standard. These strate-
gies can be also used with ANN which represents an important and useful
tool for anomaly detection.

2.3.1 Principal component analysis

This technique remaps the training data from the feature space RD in a
subspace RP (where P < D is the number of components selected) that
2.3 Anomaly Detection 21

minimizes the error (defined as Euclidean distance) between the data in the
feature space and their projection in the selected subspace [41]. More in
detail, to find the best subspace over which to project the training data, we
need to evaluate the D × D sample covariance matrix

XT X
Σx = . (2.5)
Nx − 1

By eigenvalue decomposition Σx can be factorized as Σx = Vx Λx Vx T , where

Vx is an orthonormal matrix whose columns are the eigenvectors, while Λx
is a diagonal matrix that contains the D eigenvalues. The eigenvalues mag-
nitude represents the importance of the direction pointed by the relative
eigenvector. The projection into the subspace is obtained by multiplying the
data with VP , i.e., XP = XVP , yP = YVP , and uP = UVP where VP
represent the P greatest eigenvector selected among the D extracted.
To evaluate the error between the projected points and the starting ones
it is necessary to reconstruct the data in the original feature space (i.e. X
‹=
‹ = YP VT , and U
XP VPT , Y P
‹ = UP VT ). After the reconstruction, it is possible
P
to evaluate the error as the Euclidean distance between the original data and
the reconstructed data. An example of PCA applied on a set of data is
reported in Fig. 2.8. An important limitation of this technique is that it is
able to find just linear boundaries in the original feature space.

2.3.2 Kernel principal component analysis

In many cases, the linear boundaries found by PCA represent a severe limi-
tation [42]. KPCA firstly maps the data with a non-linear function, named
kernel, then applies the standard PCA to find a linear boundary in the new
feature space. Such boundary becomes non-linear in the original feature
space. A crucial point in KPCA is the selection of the kernel that leads to
linearly separable data in the new feature space. In [43], when data distribu-
tion is unknown, the RBF kernel is proposed as good candidate to accomplish
this task. Given a generic point z that corresponds to a 1 × D vector, we can
22 Machine Learning Algorithms

4 3

3 2

2 1

1 0

g2
f2

0 -1

-1 -2

-2 -3

-3 -4

-4 -5
-3 -2 -1 0 1 2 3 -4 -2 0 2 4
f1 g1

Figure 2.9: Example of KPCA; on the left the points in the original feature
space are depicted, on the right the same points are projected in a new feature
space through RBF.

apply the RBF as

2
Kn(z) = e−γ||z−xn || with n = 1, 2, . . . , Nx (2.6)

where γ is a kernel parameter (which controls the width of the Gaussian

(z)
function) that must be set properly, xn is the nth row of X, and Kn is the
nth component of the point z in the kernel space. Overall, the vector z is
(z) (z) (z)
mapped in the vector k(z) = [K1 , K2 , . . . , KNx ]. Remapping all the data
in the kernel space, we obtain the subsequent matrices Kx of size Nx × Nx
for training, Ky of size Ny × Nx for validation, and Kv of size Nv × Nx for
test, respectively. Applying now the PCA to the new data sets, it is possible
to find non-linear boundaries in the original feature space. An example of
KPCA with RBF kernel applied on a set of data is shown in Fig. 2.9.

2.3.3 Gaussian mixture model

Another important algorithm that can be used to solve OCC problems is
the GMM [44]. This approach assumes that data can be represented by
a mixture of M multivariate Gaussian distributions. The outputs of the
2.4 OCCNN 23

Figure 2.10: Example of GMM; blue points represent the training data, the
surface represents the Gaussian mixture of order 2 that fit the data.

algorithm are the covariance matrices, Σm , and the mean values, µm , of the
Gaussian functions, with m = 1, 2, . . . , M. The GMM algorithm finds the
set of parameters Σm and µm of a Gaussian mixture that better fit the data
distribution. This strategy can also be used to perform clustering on a set of
data [29, 30]. An example of GMM applied on a set of data is shown in 2.10.
Now, considering the high versatility of NNs, and the need for effective
OCCs, the following question arises:

Q3: Is it possible to use classical NN structures (instead of the ANNs) to

perform anomaly detection?

2.4 OCCNN
In this section, a novel technique based on NN classifiers is presented. For
the sake of clarity, this section is divided in several parts, each one describing
a block of the diagram depicted in Fig. 2.11.
24 Machine Learning Algorithms

Figure 2.11: Block diagram of the OCCNN algorithm.

2.4.1 Density estimator

The first fundamental step in the classification chain is the density estimation
of the training points. As proved in [45] by Pollard, a minimum-variance un-
biased estimator of the density exists when the data is distributed according
to a Poisson point process. Such estimator has the form [45]
ÄPN ä
p
n=1 kn −1
λ
bx = PNp 2 (2.7)
π n=1 rn

where λbx is the estimated density for a dataset X, rn is the Euclidean distance
between the the n-th point and its kn -th neighbour, and Np is the number
of points considered for the estimation. It is possible to show that the mean
value and variance of the estimator are, respectively

¶ ©
E λ
b x = λx (2.8)

bx = ÄP λx ä
¶ ©
V λ Np
. (2.9)
n=1 kn − 2

From equation (2.9) it is clear that by increasing the number of points Np

and the value of kn (hence considering farther points) the estimator variance
2.4 OCCNN 25

decreases and accordingly the accuracy increases. However, when the spatial
distribution of data points deviates from Poisson, some countermeasures are
taken especially to account for the finite boundaries of the normal class:

• the set of points from which evaluating the distances rn is a subset of

the training data X;

• low values of kn are selected;

• A high number of points Np is used to increase the estimator accuracy.

More precisely, a subset of Np = 200 points is selected from the training

set (when X has less points, all the training set is used for the density es-
timation), and for all these points the second nearest point is considered so
kn = 2 for n = 1, 2, . . . , Np .

2.4.2 Adversarial points generator

This block solves the task to generate uniformly distributed random points
with density λi = λ bx αi in a specific portion of the feature space defined
by θ i−1 where i is the iteration index. This is fundamental to create an
adversarial class Z from which the NN can learn the boundaries between the
training data X and the adversarial ones. The adversarial points must be
generated in the portion of space where the training points are absent, hence
where the output layer activation function is greater than 0.5 (we suppose
to associate label 1 to the training set points and 0 to the adversarial ones).
The function gets as input the estimated training point density multiplied by
a factor αi , that we will discuss in Section 2.4.4, and the NN weights matrix
θ i−1 that represents the network state after the previous training iteration
and defines the actual estimated boundaries. At the first iteration (i = 1)
we suppose θ 0 equal to a null matrix; this means that the adversarial points
will be generated in all the feature space including in the area filled by the
training points.
26 Machine Learning Algorithms

2.4.3 Network structure

A feed-forward NN with two hidden layers and M neurons in each layer
is trained to find the boundaries between the normal class data and the
adversarial ones [30]. All the layers are fully connected, and the activation
functions are ReLU for the hidden layers and softmax for the output layer.
The network is trained with all the training set points for Ne epochs with
learning rate ρ. At the end of the training phase, the network provides the
weight matrix θ i that defines the boundaries estimated at the ith iteration
of the OCCNN algorithm that will be used in the next iteration by the
adversarial points generator block to generate points outside the boundaries
defined. Finally, at the last iteration, the boundaries estimated by the NN
are the ones used for anomaly detection.

2.4.4 Dimensioning
In this section, a wide set of tests is presented with the aim to find the best
configuration of hyper-parameters. The main goals are:

• testing the proposed algorithm with different numbers of training points,

Nx , varying α1 ;

• finding the value of α2 that maximize the classification accuracy;

• comparing the OCCNN solution with PCA, KPCA, GMM, and ANN.

The OCCNN hyper-parameters are M = 50 neurons, Ne = 5000 epochs, and

the learning rate is ρ = 0.05. The error function for a generic training point
xn is defined as Ã
XD
exn = en,d )2 .
(xn,d − x (2.10)
d=1

The same metric can also be used to evaluate the reconstruction error of all
the test set points. A target false alarm rate on the training set is fixed equal
to 0.01, consequently a threshold is selected for each algorithm to guarantee
such constraint.
2.4 OCCNN 27

Figure 2.12: Average accuracy calculated varying α1 and NP .

Dataset size and test on α1

The first important step is to find the value of α1 that maximizes the accuracy
of the first iteration evaluated on a validation set and checked on the test
set. Validation and test sets have the same distribution, but in order to avoid
generalization error it is good practice to define two different datasets [31].
The value of Nx = Ny = Nv = Np is tested between 200 and 1000, for lower
values the density estimator variance becomes too high and the dataset too
small to train the NN. Again, the value of α1 is varied in the range between
0.01 and 1, higher values would generate too many adversarial points in
the area of the training set and the NN would classify all the feature space
as adversarial. In Fig. 2.12 a heatmap that shows the average accuracy
(evaluated over the six distributions shown in Fig. 2.14) of the OCCNN
algorithm varying Np and α1 is reported. Note that accuracy increases with
the number of points in the training set and the best value for α1 is 0.3,
irrespectively of the number of training points. From now on the parameter
α1 is set to 0.3 and the number of training, test and validation points for the
next simulations is fixed at 500.

Test on α2

Another important test is the definition of the parameter α2 that sets the
adversarial point density after the first iteration. This test is executed with
28 Machine Learning Algorithms

Figure 2.13: Accuracy as a function of α2 for different distributions.

a fixed number of points in the training set, and the accuracy, when α2
varies between 0.01 and 2, is shown in Fig. 2.13. It is important to highlight
that in this case α2 can assume values greater than 1 because in this phase of
the OCCNN algorithm the boundaries are already defined, so the adversarial
points will only be generated outside the boundaries estimated by the NN and
defined by the weight matrix θ 1 . We can see this step as a fine-tuning of the
boundaries roughly estimated in the first iteration. All six curves are shown
along with the average accuracy (thick blue curve). It can be noted that
for values of α2 lower than 0.6, the accuracy becomes unpredictable because
the low number of adversarial points generated is not enough to define the
boundaries correctly. On the other hand, increasing α2 beyond 1.5 slightly
decreases the accuracy because the boundary region tends to be shaped by
the adversarial class. The value of α2 = 0.8 is selected to maximize the
accuracy and improve the performances with respect to the first step. With
these parameters, an image of the boundaries estimated by the OCCNN at
the end of the two iterations is shown at the bottom of Fig. 2.14. Comparing
these with the points distributions at the top of Fig. 2.14 it is possible to see
2.4 OCCNN 29

1 1 1

0 0 0

−1 0 1−1 0 1−1 0 1
1 1 1

0 0 0

−1 0 1−1 0 1−1 0 1

Figure 2.14: At the top some examples of normalised dataset points dis-
tribution. Blue circles denote features corresponding to normal conditions,
red circles denote features corresponding to anomaly conditions. Estimated
boundaries by the OCCNN algorithm are showed at the bottom.

that the estimated boundaries are really close to the real ones. Now that all
the parameters are set we are ready to compare the various algorithms.

2.4.5 OCCNN2
The OCCNN performance is rather sensitive to the density estimation; there-
fore, depending on the data set, its ability to detect anomalies may be dom-
inated by this first step [36]. Moreover, the Pollard’s estimator (2.7) may
exhibit accuracy degradation when the data set points distribution devi-
ates from Poisson. Based on these considerations, a possible solution is the
30 Machine Learning Algorithms

OCCNN2 , an anomaly detector that shares the same strategy of OCCNN

only the first NN of the classification chain (the red box in Fig. 2.11) is
replaced with another OCC to have a rough estimate of the normal class
boundaries. As shown in Section 4.3, a good choice for this step is to use an
ANN.

Now that a set of tools for detection of anomalies are available, two ques-
tions concerning the goal of this thesis arise:

Q4: How can we perform anomaly detection based on machine learning tools
to detect a damage in a structure?

Q5: Is there an effective technique to extract features able to highlight a

damage in a structure?
Chapter 3

Structural Health Monitoring

3.1 Introduction
In this section we present some tools that belong to the so-called OMA able
to perform structural monitoring and extract damage sensitive features from
some structure characteristics. Over the years several strategies have been
developed to extract relevant parameters that can represent the healthy state
of a structure (i.e., natural frequencies, fundamental modes, dumping ratios,
etc.). These strategies can be divided in two groups:

• Model based methods, that require a simulated model of the structure,

hence needing an accurate knowledge of the structure elements and
materials.

• Model-free methods, that can be assumed as a blind technique not

needing a detailed description of the structure.

Among the model based techniques we can find the finite element method,
this strategy provides a simulated model of complex structure as a sum of
basic elements that are easier to model with respect to the whole system.
This strategy provides an accurate estimation of the modal parameter of
the structure as long as its simulation is accurate. The main problem with
this strategy is the computational cost and the impossibility to generalize
it to different structures. In fact, the need of a model require a targeted

31
32 Structural Health Monitoring

study of the particular infrastructure and cannot be generalized [14]. For

this reason we decided to focus our efforts on the model-free methods; to this
class belong the peak-picking method, the auto-regressive moving average
(ARMA) models, and the SSI approach.

3.2 Operational Modal Analysis

The expression OMA means the class of modal identification methods based
on response measurements only. This discipline has been systematized in the
last two decades but early applications can be traced back to the beginning
of modal testing in the 1960s.
The first applications of OMA were mainly based on the analysis of PSD
and the identification of operational deflection shapes (ODSs). An ODS
represents the deflection of a structure at a particular frequency under a
generic input and it is usually the result of the contribution of different
modes.
However, under certain assumptions, ODSs are a close estimate of the
actual mode shapes. The mode shapes represent the displacement of a struc-
ture subject to a stress with respect to its repose position.
In the 1990s, a number of methods working in time domain were devel-
oped and applied in combination with correlation functions, leading to the
so-called output-only modal analysis. In the same period the use of ARMA
models for modal parameter estimation, first suggested in the late 1970s,
started spreading. An increasing number of applications appeared in the
literature but output-only modal identification was not fully developed and
widely accepted as a reliable source of information, yet.
However, at the end of the 1990s new effective output-only modal identi-
fication techniques, such as the SSI, became available, overcoming the limi-
tations of the previous techniques in dealing with closely spaced modes and
noise [46].
Nowadays, OMA is a widely accepted tool for modal identification, with
several successful applications in civil engineering (bridges, buildings, pedes-
trian bridges, historical structures, offshore platforms, wind turbines, dams,
3.2 Operational Modal Analysis 33

stadia), mechanical and industrial engineering (ships, trucks, car bodies, en-
gines, rotating machineries), aerospace engineering (in-flight modal identifi-
cation of aircrafts and shuttles, studies about flutter).
OMA is based on the following assumptions:

• Linearity (the response of the system to a given combination of inputs

is equal to the same combination of the corresponding outputs).

• Stationarity (the dynamic characteristics of the structure do not change

over time, so that the coefficients of the differential equations governing
the dynamic response of the structure are time independent).

• Observability (the sensor layout has been properly designed to observe

the modes of interest, avoiding, for instance, nodal points).

Moreover, unlike the traditional modal testing where the input is con-
trolled, OMA is based on the dynamic response of the structure under test
to non-controllable and immeasurable loadings such as environmental and op-
erational loads (traffic, wind, microtremors, and so on). As a consequence,
some assumptions about the input are needed. If the structure is excited
by white noise, that is to say, the input spectrum is constant, all modes are
equally excited and the output spectrum contains full information about the
structure [14]. However, this is rarely the case, since the excitation has a
spectral distribution of its own. Modes are, therefore, weighted by the spec-
tral distribution of the input and both the properties of the input and the
modal parameters of the structure are observed in the response.
Additionally, noise and spurious harmonics due to rotating equipment
are observed in the response. Thus, generally, the structure is assumed to
be excited by unknown forces that are the output of the so-called excitation
system loaded by white noise (see also Fig. 3.1). Under this assumption, the
measured response can be interpreted as the output of the combined system,
made by the excitation system and the structure under test in series, to a
stationary, zero mean, Gaussian white noise.
Since the excitation system and the structure under test are in cascade,
the frequency response function of the combined system is the product of
34 Structural Health Monitoring

Figure 3.1: Block diagram of the OMA strategy [14].

their respective frequency response functions [14]:

Hc (ω) = Hf (ω)Hs (ω)

where Hc (ω), Hf (ω), and Hs (ω) are the frequency responses of the combined
system, the excitation system, and the structure under test, respectively.
In fact, for each subsystem, output and input are related by the following
equations [14]:
F (ω) = N (ω)Hf (ω)

Y (ω) = F (ω)Hs (ω) (3.1)

where N (ω), F (ω), and Y (ω) denote the Fourier transforms of the white
noise input to the excitation system, the excitation system output, and the
structure output, respectively. In this context, the measured response in-
cludes information about the excitation system and the structure under test,
but the modal parameters of the structure are preserved and identifiable, and
the characteristics of the excitation system have no influence on the accuracy
of modal parameter estimates.
The discrimination between structural modes and properties of the ex-
citation system is possible since the structural system has a narrowband
response and time invariant properties, while the excitation system has a
broadband response and it may have either time variance or time invariance
properties.
The estimation of the modal model of the structure gives the opportunity
to estimate also the unknown forces, according to (3.1). The assumption of
broadband excitation ensures that all the structural modes in the frequency
range of interest are excited. Assuming that the combined system is excited
3.2 Operational Modal Analysis 35

by a random input, the second order statistics of the response carries all
the physical information about the system and plays a fundamental role in
output-only modal identification. The focus on second order statistics is
justified by the central limit theorem. In fact, the structural response is
approximately Gaussian in most cases, regardless of the distributions of the
(independent) input loads, which are often not Gaussian.
The spatial distribution of the input also affects the performance of OMA
methods, especially in the presence of closely spaced modes. A random dis-
tribution of inputs in time and space provides better modal identification
results. The presence of measurement noise and spurious harmonics in re-
sponse to measurements requires appropriate data processing to mitigate
their effects and discriminate them from actual structural modes. These
strategies take the name of mode selection and will be discussed in the fol-
lowing section.

3.2.1 Peak-Picking

The most undemanding method for output-only modal parameter identi-

fication is the basic frequency domain (BFD) method, also known as the
peak-picking method. The name of the method comes from the fact that the
modes are identified by picking the peaks in the PSD plots. Given a pair
of sample records x(t) and y(t) of finite duration T from stationary random
processes, their Fourier transforms (which exist as a consequence of the finite
duration of the signals) are:
Z T
X(f ) = x(t)e−j2πf t dt
0

Z T
Y (f ) = y(t)e−j2πf t dt
0

and the auto-spectral and cross-spectral density functions are defined as fol-
lows: Z +∞
Sxx (f ) = Rxx (τ )e−j2πf τ dτ
−∞
36 Structural Health Monitoring

Z +∞
Syy (f ) = Ryy (τ )e−j2πf τ dτ
−∞
Z +∞
Sxy (f ) = Syx (f ) = Rxy (τ )e−j2πf τ dτ
−∞

where Rxy (τ ) represents the cross-correlation function between the signals

x(t) and y(t). The BFD technique is based on the assumption that, around
a resonance, only one mode is dominant. A reference sensor for the com-
putation of the cross-spectral densities with all other measurement channels
has to be selected to ensure that most of the modes can be observed. As a
consequence, sensors close to nodes of the mode shapes cannot be adopted
as reference sensors. The ideal choice for the reference sensor makes possi-
ble the identification of all the modes through the computation of a single
column of the PSD matrix (the column made by the cross-spectral densities
between the selected reference sensor and all other sensors).
However, depending on the geometry of the structure and the adopted
sensor layout, a single reference sensor could be insufficient to identify all the
modes and at least a couple of reference sensors with different orientation
have to be adopted. For instance, in the case of a building-like structure
characterized by two bending modes in two orthogonal directions x and y
and sensors parallel to these directions, the selection of a sensor measuring
along x as reference permits the identification of the bending modes in the
x direction and eventually torsional modes, but it is inadequate to identify
the bending modes in the y direction. This can be observed only through
the selection of an additional reference sensor measuring along y [14].
However, the success of the identification process heavily depends on the
geometry of the structure and the skill of the analyst. The results of modal
identification suffer a certain degree of subjectivity also in the case of noisy
measurements, when the peaks in the spectra are not clear. The BFD method
provides local estimates of the modal properties. Moreover, the accuracy of
the estimated eigenfrequencies depends on the frequency resolution of the
spectra. A fine frequency resolution is fundamental to obtain good natu-
ral frequency estimates, but this can be increased with well known signal
processing techniques (i.e., zero padding) [47].
3.2 Operational Modal Analysis 37

In principle, the BFD should be applied to evaluate natural frequencies

and mode shapes only. However, the BFD technique is effective when damp-
ing is low and modes are well separated. If these conditions are violated
it may lead to erroneous results. Thus, this technique can be used only in
certain circumstances, suffer of subjectivity, and does not provide an ex-
haustive and accurate estimation of the modal parameters. However it is
computationally inexpensive.

3.2.2 Auto-regressive moving average

This technique starts with the equation of motion of a randomly excited
linear time-invariant system, that can be written as follows:

d2 y(t) dy(t)
M 2
+D + Ky(t) = w(t) (3.2)
dt dt

where y(t) stands for the measured system response, w(t) is a continuous-
time, zero mean Gaussian white noise, M represents the mass matrix, D the
dumping matrix, and K the stiffness matrix [14]. It is possible to show that
this system can be also described by a discrete-time ARMA vector model
(the ARMA model is referred as such an ARMA vector model to point out
its multivariate character) by approximating the differential operator with
finite differences over a finite time step ∆t.
Historically, ARMA vector models have been used for the estimation of
the modal parameters of civil structures. Due to a number of shortcomings
(in particular for systems with many outputs and many modes, where the
large set of parameters to be estimated leads to large computational burden
and convergence problems), stochastic state-space models have progressively
replaced them in the context of modal identification.
In order to explain how modal parameters can be extracted from an
ARMA model, assume that a continuous-time system is observed at discrete
time instants k with a sampling interval ∆t. Since the input on the structure
is not available (it is immeasurable), the equivalent discrete-time system can
be obtained only by requiring that the covariance function of its response
38 Structural Health Monitoring

to a Gaussian white noise input be coincident at all discrete time lags with
that of the continuous-time system. This implies that the first and second
order moments of the response of the discretized model are equal to the
first and second order moments of the response of the continuous-time sys-
tem at all the considered discrete time instants. Under the assumption that
the response of the system is Gaussian distributed, the covariance equiva-
lent model is the most accurate approximated model, since it is exact at all
discrete time lags. When the dynamic response of the system is driven by
the Gaussian white noise w(t) but there are also some disturbances (process
and measurement noise), the latter must also to be taken into account by
the equivalent discrete-time model. In the presence of such disturbances, an
ARM A(nα, nγ) model has the following form:

yk + Ω1 yk−1 + · · · + ΩNΩ yk−NΩ = ek + Γ1 ek−1 + · · · + ΓNΓ ek−NΓ (3.3)

where yk is the vector of the output at the time instant tk , and ek is the
innovation modeled as a zero mean Gaussian white noise. The left-hand side
of (3.3) is the auto-regressive part, while the right-hand side is the moving
average part. The matrices Ωi contain the auto-regressive parameters, while
the matrices Γj contain the moving average parameters; NΩ and NΓ represent
the auto-regressive and moving average order of the model. It is possible to
show that a covariance equivalent ARMA vector model can be converted into
a forward innovation state space model, and vice versa. If the order of the
state-space model is too large, the model will contain redundant information;
on the contrary, if the state-space dimension is too small, a certain amount
of information about the modelled system will be lost. This transformation
provides a good estimation of the modal parameters but ARMA models
usually present stability issues and are computationally expansive.

3.2.3 Stochastic subspace identification

Among all the possibilities, this strategy has been selected as the best can-
didate to extract modal parameters from the accelerometric measurements
present for the Z-24 bridge. In the following will be explain how this tech-
3.2 Operational Modal Analysis 39

nique has been applied to the Z-24 environment. The SSI is a time-domain,
parametric, covariance-driven procedure for blind modal analysis (i.e., it can
extract the modal frequencies from the accelerometric measurements without
any a priori knowledge about the structure [14]).
SSI is characterized by the system order n ∈ N and the time-lag i ≥ 1. To
apply the algorithm it is enough to satisfy the following constraint, l · i ≥ n
[14], where l stands for the number of sensors. In the Z-24 monitoring, the
system order n is considered unknown, so it is kept as a parameter varied in
the range n ∈ [2, 160] (with step 2), while the time-lag is i = 60 [14].
Let us define the block Toeplitz matrix for a given time-lag i and shift s
 
(a) (a) (a) (a)
Ri Ri−1 . . . Rs+1 Rs
 (a)
 Ri+1 R(a)
i
(a) (a) 
. . . Rs+2 Rs+1 
(a)
Ts|i = . (3.4)
 
 .. .
.. .. .. .. 
 . . .  
(a) (a) (a) (a)
R2i−1 R2i−2 . . . Ri+1 Ri

of dimensions li × li where

(a) 1
Ri = D(a,1:N −i,:) DT
(a,i:N,:) (3.5)
N −i

is a correlation matrix of dimension l × l, and the data matrix D(a,b:c,:) is

obtained by the tensor D selecting a particular acquisition a from all the
sensors in the time interval [b, c], where tensor D of dimensions Na × l ×
Ns , contains all the data collected after a pre-processing phase that will
be described in the following, with Na number of acquisitions, l number
of sensors, and Ns number of samples. For the sake of this presentation
we will drop the acquisition index a, considering that all the subsequent
operations are performed for each acquisition. Now, by the singular value
decomposition (SVD) we can factorize the block Toeplitz matrix (3.4), with
s = 1, as
T1|i = U(n) Σ(n) V(n)T (3.6)

where U(n) and V(n)T are matrices of dimensions respectively li×n and n×li
and Σ is a diagonal matrix, of dimension n × n, that contain the singular
40 Structural Health Monitoring

values arranged in descending order. For the sake of this presentation, we

will drop the model order index n, considering that all the subsequent obser-
vations can be applied for each model order. The SVD allows selecting the
correct number of singular values that must be considered, and therefore to
split the matrix T1|i in two parts

T1|i = Oi Γi (3.7)

Oi = UΣ1/2 S (3.8)

Γi = S−1 Σ1/2 VT (3.9)

where  
C
 
 CA  î ó
Oi =  .  and Γi = Ai−1 G . . . AG G (3.10)
 
 .. 
 
i−1
CA
represent, respectively, the observability matrix and the reversed controlla-
bility matrix. In (3.8) and (3.9) the matrix S plays the role of a similarity
transformation applied to the state-space model, therefore it can be set equal
to the identity matrix I. The matrices A, C, and G, in equation (3.10) repre-
sent the state matrix, the output influence matrix, and the next state-output
covariance matrix, respectively. Note that C and G can be easily extracted
from the matrices Oi and Γi , while A can be calculated by (3.4) as

A = O+ +
i T2|i+1 Γi . (3.11)

Applying the eigenvalues decomposition to A we get

A = ΨΩΨT (3.12)

where Ψ is an orthonormal matrix whose columns are the eigenvectors, while

Ω = diag(λ
e1 , . . . , λ
en ) is a diagonal matrix that contains the n eigenvalues of
the state matrix. Therefore, reintroducing the previously omitted indices, it
3.3 Mode selection 41

160 160

140 140

120 120

100 100

80 80
n

n
60 60

40 40

20 20

0 0
0 5 10 15 20 25 0 5 10 15 20 25
f [Hz] f [Hz]
(a) (b)

Figure 3.2: Example of stabilization diagram for the first hour monitoring:
a) through SSI, b) after mode selection and clustering.

is possible to estimate the continuous-time damage sensitive parameters of

the pth mode as follows:
(a,n) (a,n)
• eigenvalues λp = fs ln(λ
ep );
(a,n) (a,n)
• natural frequencies µp = |λp |/(2π);
(a,n) (a,n) (a,n)
• dumping ratios δp = −<{λp }/|λp |;

• mode shapes φ(a,n)

p = C(a,n) ψ (a,n)
p ;

where φ(a,n)
p is a l × 1 vector, and ψ (a,n)
p is the pth column vector of Ψ(a,n)
defined in (3.12). The natural frequencies extracted through this approach
in the Z-24 environment for the first acquisition (a = 1) varying the order n
is depicted in the stabilization diagram shown in Fig. 3.2a.

3.3 Mode selection

Through SSI, a large number of spurious modes are generated. To sift them
out, there are several approaches to distinguish between the real modes and
the spurious ones [48]. In this work we used four criteria, which must be satis-
fied to select a physic mode and reject the spurious ones. The four metrics ex-
tracted are: modal assurance criterion (MAC), mean phase deviation (MPD),
42 Structural Health Monitoring

dumping ratio check, and complex conjugate poles check. Among several
criteria, we choose to use these four metrics as a good compromise between
computational cost and accuracy of the spurious modes detection. The met-
rics selected are widely used and discussed in literature [46, 48–50].

MAC

It is a dimensionless squared correlation coefficient between mode shapes,

defined as [46]

|φ(a,n)T
p φ(a,j)
q |2
MAC(φ(a,n)
p , φ(a,j)
q ) = (3.13)
||φ(a,n)
p ||22 ||φ(a,j)
q ||22

with values between 0 and 1. A MAC larger than 0.9 indicates a consis-
tent correspondence between the modes and so, very likely physical modes,
whereas small values indicate poor resemblance of the two shapes and so, a
spurious mode. A validation criteria based on MAC is the following [51]

(a,n) (a,j)
|λp − λq |
dm (a, n, j, p, q) = (a,n) (a,j)
+ 1 − MAC(φ(a,n)
p , φ(a,j)
q ) (3.14)
max(|λp |, |λq |)

where the first term is a measure of the distance between the pth and qth
eigenvalues. With (3.14) a mode is considered physical when dm (a, n, j, p, q) <
0.1.

MPD

It measures the deviation of a mode shape components from the mean phase
(MP) of all its components. A widely used technique to evaluate the MP is
through the SVD [51]

PΛQT = [<{φ(a,n)
p } ={φ(a,n)
p }] (3.15)
3.4 Features Extraction 43

where P is l × 2, Λ is 2 × 2, and Q is 2 × 2. From this decomposition the

MPD can be evaluated as follows
(a,n) (a,n)
Pl (a,n) <{φr,p }q22 −={φr,p }q12
r=1 |φr,p | arccos
√ (a,n)
2 +q 2 |φ
q12 22 r,p |
MPD(φ(a,n)
p )= Pl (a,n)
. (3.16)
r=1 φr,p

When the ratio MPD(φ(a,n)

p )/(π/2) > 0.75 the deviation of the mode shape
components is stronger than MP and probably the mode is spurious, hence
it is rejected.

Damping ratio and complex conjugate poles

(a,n)
For each mode the dumping ratio δp in real structures must be positive
and lower than 0.2 (otherwise the structure is unstable) hence only modes
(a,n)  (a,n)
with 0 < δp < 0.2 are considered. Moreover, if < λp > 0 the mode
represents an unstable structure hence it is considered spurious.
In Fig. 3.2b, showing the stabilization diagram after mode selection, the
(a,n)
remaining modes are represented with µ̄p .

Q6: Is it possible to perform frequencies tracking automatically after the

fundamental frequencies extraction?

3.4 Features Extraction

After mode selection, we propose clustering and a novel tracking algorithm
to extract the time-series of the first two fundamental frequencies. To clarify
the process, this section will be presented in the Z-24 bridge environment,
but can be easily extended to other structures.

3.4.1 Clustering
The stabilization diagram obtained after mode selection needs to be further
processed to select the natural frequencies. For this purpose, K-means can
be used to cluster the data, associating each cluster to a prospective natural
44 Structural Health Monitoring

frequency corresponding to its centroid [29, 52]. For each acquisition the
algorithm starts with an initial number of centroids (e.g., K = 4) initialized
at random frequencies within the range of interest, and then their position
will be updated until convergence.
Since the number of natural frequencies is unknown, the approach is re-
peated for different K values, ranging between 2 and 6; a large K tends to
produce many spurious modes while small values may discard real funda-
mental frequencies.
Finally, the centroids configuration that leads to the solution with the
lowest error (evaluated as the sum of the Euclidean distances between the
centroids and the associated points) is selected. The output of K-means is the
number of natural frequencies and their estimation (red lines in Fig. 3.2(b)).
Repeating this procedure for each acquisition the blue points depicted in
Fig. 3.3a are obtained.
At the end of clustering, the estimated frequencies are no more dependent
(a)
on the model order n, consequently the notation µ̄p is adopted (see Fig. 4.2).

3.4.2 Density-based mode tracking

The tracking phase is the final step in the natural frequencies extraction
chain. There are several algorithms able to find the frequency traces starting
from the estimation made through the clustering phase [49]. In this thesis,
a novel technique, that needs neither the frequency starting points nor their
number, is proposed, contrary to several approaches that infer the number
and the starting position of the fundamental frequencies through simulation
of the physical structure. The proposed tracking algorithm consists of two
steps: a starting phase and an online phase.

Initial phase

Without any assumption about the structure, the idea is to analyze the data,
(a)
µ̄p , and find some clusters of points that could be the starting ones. To per-
(a)
form this task, the first Nt = 200 acquisitions (i.e. µ̄p with a = 1, 2, . . . , Nt ,
are considered (see Fig. 3.3a)). From this initial data, the number of points
3.4 Features Extraction 45

f1
15 15 f2
f3
f4
(0)
f4
Frequency [Hz]

(0)
10 10 f3

(0)
5 5 f2
(0)
f1

0 100 200 0 250 500 0 1000 2000 3000 4000

Time [h] N Points Time [h]
(a) (b) () (c)

Figure 3.3: a) Residual modes after mode selection over time with one acqui-
sition per hour (first 200 acquisitions). b) Natural frequencies selected after
the initial phase of the tracking algorithm. c) First two natural frequencies
estimation after the density-based tracking algorithm. Blue points represent
the residual modes, red and yellow tracks represent, respectively, the first
and second fundamental frequency estimated after the tracking algorithm;
vertical dashed lines highlight the period when the average measured tem-
peratures are below 0 ◦ C [15], in particular, it has been demonstrated in [16]
that when the temperature goes below 0 ◦ C the natural frequencies of the
bridge increase. Blue and green backgrounds highlight the acquisitions made
during the normal condition of the bridge, used respectively as training and
test sets, while red background stands for damaged condition acquisitions
used in the test phase.

that fall into frequency bins of bandwidth Bf = 0.4 Hz are counted. The his-
togram obtained is depicted in Fig. 3.3b. Selecting the largest values of the
histogram, the number of starting points and the corresponding frequencies,
(0)
fs , are estimated. Specifically, the first estimated frequency is evaluated as
the average values of the frequencies that fall into the respective bins. For
(0)
example, according to Fig. 3.3b the values of the starting points, fs , in this
case s = 1, . . . , 4, are estimated and correspond to 4.0, 5.2, 10.1 and 12.8 Hz.
46 Structural Health Monitoring

Online phase

In this phase, for each starting point a Gaussian kernel of the form

(a) 2
Ä ä
1
− θ−fs
2σ 2
θ; fs(a) , σf


G =e f (3.17)

is used to track the frequencies evolution over time. The parameter σf con-
trols the kernel width and has been chosen equal to σf = 0.01 Hz; larger val-
ues of σf make the system more reactive to fast frequency changes during the
tracking but more sensitive to outliers due to the noisy measurements. The
(0)
tracking algorithm is initialized with fs and iteratively updated through
the following rule
fs(a) = (1 − )fs(a−1) + fes(a) (3.18)

where the parameter  ∈ [0, 1] controls the impact of the new observation.
Large values of  reduce smoothness but allow capturing sudden changes
of modal frequencies. For the specific data set  = 0.7 is selected. The
(a)
innovation term fes in (3.18) is evaluated through the Gaussian kernel as
P (a) (a) (a−1)

pµ̄p G µ̄p ; fs , σf
fes(a) = P (a) (a−1)
. (3.19)
p G µ̄p ; fs , σf

(a)
The four tracks fs = {fs }Na
a=1 are shown in Fig. 3.3 and stored in the
following matrix
 T  (1) (2) (N ) T

f1 f1 f1 ... f1 a
 (1) (2) (N ) 
f2 a 
 
 = f2 f2 ...
f2  
F=  . (3.20)
f  f (1) (2)
f3 ...
(N )
f3 a 
 3  3 
(1) (2) (Na )
f4 f4 f4 ... f4

Now that a tool to extract damage sensitive features is introduced and

designed the following question arises:
Q7: How can we test these strategies? Is there in literature a reference
structure with accelerometric measurements both in standard and in
damaged conditions?
Chapter 4

Z-24 Bridge

4.1 System Configuration and Data Collec-

tion
The Z-24 bridge was located in the Switzerland canton of Bern. The bridge
was part of the road connection between Koppigen and Utzenstorf, over-
passing the A1 highway between Bern and Zurich. It was a classical post-
tensioned concrete two-cell box girder bridge with a main span of 30 m and
two side spans of 14 m. The bridge was built as a freestanding frame with
the approaches backfilled later. Both abutments consisted of triple concrete
columns connected with concrete hinges to the girder. Both intermediate
supports were concrete piers clamped into the girder. An extension of the
bridge girder at the approaches provided a sliding slab. All supports were
rotated with respect to the longitudinal axis that yielded a skew bridge. The
bridge was demolished at the end of 1998 [53]. Before complete demolition,
the bridge was subjected to a long-term continuous monitoring test and sev-
eral progressive damage tests (see also Tab. 4.1):

• A long-term continuous monitoring test took place during the year be-
fore demolition. The aim was to quantify the environmental variability
of the bridge dynamics.

• Progressive damage tests took place over a month, shortly before com-

47
48 Z-24 Bridge

Table 4.1: Chronological overview of applied damage scenarios

Date (1998) Scenario
4 August Undamaged condition
9 August Installation of pier settlement system
10 August Lowering of pier, 20 mm
12 August Lowering of pier, 40 mm
17 August Lowering of pier, 80 mm
18 August Lowering of pier, 95 mm
19 August Lifting of pier, tilt of foundation
20 August New reference condition
25 August Spalling of concrete at soffit, 12 m2
26 August Spalling of concrete at soffit, 24 m2
27 August Landslide of 1 m at abutment
31 August Failure of concrete hinge
2 September Failure of 2 anchor heads
3 September Failure of 4 anchor heads
7 September Rupture of 2 out of 16 tendons
8 September Rupture of 4 out of 16 tendons
9 September Rupture of 6 out of 16 tendons

plete demolition. The aim was to prove experimentally that realistic

damage has a measurable influence on the bridge dynamics. Progressive
damage tests were alternated with short-term monitoring tests while
the continuous monitoring system was still running.

4.1.1 Data collection

The database of accelerometric measurements is freely available for research
purposes on the website [17]. Such a dataset was built within the framework
of the European Brite EuRam research project BE-3157, “System Identifica-
tion to Monitor Civil Engineering Structures” (SIMCES). One of the main
objectives of the SIMCES project was to deliver a proof of feasibility for
vibration-based structural health monitoring of civil engineering structures
by full-scale, long-term tests and progressive failure tests of a representative
structure, the Z-24 bridge.
The accelerometers used are force balance type FBA-11 by Kinemet-
rics [54]; they are triaxial devices with high sensitivity and are characterized
by high reliability and low current consumption. The main specifications
4.1 System Configuration and Data Collection 49

Figure 4.1: Data acquisition setup along the Z-24 bridge: the selected ac-
celerometers, their positions, and the measured acceleration direction [17].

are: full scale range ±9.81 m/s2 , output ±2.5 V, natural frequency critical
damping 0.7, and supply voltage ±12 V DC. The A/D converter used is the
IOtech ADC 488/8SA [55], an 8 channel sample and hold with 8 differential
inputs, 16 bit, and 100 kHz sampling rate.

The accelerometers position and their measurements axis are shown in

Fig. 4.1. In this work, we considered l = 8 accelerometers, identified as 03,
05, 06, 07, 10, 12, 14, and 16, which are present in both long-term continuous
monitoring phase and in the progressive damage one.1 Longitudinal accelera-
tion is collected by sensors 03 and 06, transversal acceleration is measured by
sensors 14 and 16, and all the remaining sensors gather vertical accelerations.
Every hour Ns = 65536 samples are acquired from each sensor with sam-
pling frequency fsamp = 100 Hz which corresponds to an acquisition time
Ta = 655.36 s. Since the measurements are not always available, there are
Na = 4107 acquisitions collected in a period of 44 weeks. All the data col-
lected by the selected sensors are organized in a tensor D raw of dimensions
Na × l × Ns .
During the long-term monitoring, environmental parameters like temper-
ature, wind speed, and traffic were also considered before and after each ac-
celerometer measurement. In this work, only the temperature is considered,
consisting of 20 measurements acquired by different thermometers placed
1
Some accelerometers that experienced failures during the long-term monitoring have
been avoided. Moreover, we selected a subset of accelerometers present in both phases to
ensure data consistency.
50 Z-24 Bridge

Figure 4.2: The block diagram for signal acquisition, processing, feature
extraction, and detection.

along with the structure.

4.1.2 Data pre-processing

The block diagram depicted in Fig. 4.2 represents the sequence of tasks per-
formed for the fully automatic anomaly detection approach presented in this
work.
To reduce the computational cost and memory needs of the subsequent
elaborations, some pre-processing steps have been applied to the data D raw .
First, a decimation by a factor 2 is applied to each acquisition; thus the
sampling frequency is scaled to fsamp = 50 Hz. Such sampling frequency
is deemed sufficient because the Z-24 fundamental frequencies fall in the
[2.5, 20] Hz frequency range [53]. After decimation, the data are processed
with a finite impulse response (FIR) band-pass filter of order 30 with band
[2.5, 20] Hz, to remove disturbances. The pre-processed data are then stored
in a tensor D of size Na × l × N , where N = 32000 ' Ns /2. Regarding the
environmetal parameters, the temperatures collected are averaged over time
and among sensors to have one estimate for the whole bridge each hour. The
main matrices, vectors and scalars introduced are summarised in Table 4.2.
4.2 Algorithmic Complexity and Processing Time 51

Table 4.2: Data symbol, dimension, unit of measure, and description

Variable Dimension Unit Description
2
D raw Na × l × Ns m/s Raw accelerometer data
2
D (a,:,:) Na × l × N m/s Preprocessed accelerometer data
(a,n)
λp 1×1 Hz Eigenvalues
(a,n)
µp 1×1 Hz Natural frequencies
(a,n)
δp 1×1 Dumping ratios
φ(a,n)
p l×1 Mode shapes
(a,n)
µ̄p 1×1 Hz Natural freq. after mode selection
(a)
µ̄p 1×1 Hz Natural frequencies after clustering
(a)
f1 1×1 Hz First natural freq. after tracking
(a)
f2 1×1 Hz Second natural freq. after tracking
ζb(a) 1×1 Decision labels

4.2 Algorithmic Complexity and Processing

Time
In this section, we discuss the computational complexity of the algorithms
adopted and the corresponding processing time when they are applied on the
Z-24 dataset. The most computationally expensive blocks are SSI and the
anomaly detectors during the training phase: PCA, KPCA, GMM, ANN,
OCCNN, OCCNN2 . Their complexity is evaluated as follows:

• SSI requires the SVD of matrix T1|i (3.6), of size li×li, and eigenvalues
decomposition of matrix A (3.12), of size n × n. SVD complexity is
O(k(li)2 + k 0 (li)3 ) (where k and k 0 are constants which are 4 and 22,
respectively, for the R-SVD algorithm) [56]; the complexity of eigenval-
ues decomposition is O(n3 ). All these procedures are applied for each
system order n considered in the range n ∈ [2, 160] (with step 2, for
a total of ns = 80), and repeated for all the Na acquisitions. Thus,
the overall complexity is O(Na n ((k(li)2 + k 0 (li)3 ) + n3 )), and can be
P

rewritten as O(Na ns (k(li)2 + k 0 (li)3 + 2n3s )).

52 Z-24 Bridge

• PCA needs the evaluation of the covariance matrix of the training

points with complexity O(Nx s2 ), and its eigenvalues decomposition,
O(s3 ), where s is the number of features and Nx the number of train-
ing points. The total complexity of PCA results to be O(Nx s2 + s3 ).

• KPCA requires the evaluation of the kernel with complexity O(Nx2 ),

its covariance matrix evaluation, O(Nx3 ), and its eigenvalues decompo-
sition, O(Nx3 ). The overall complexity of KPCA is O(Nx2 + 2Nx3 ).

• GMM has complexity O(Nx M2 ) where M is the order of the model,

according to [57].

• ANN is a particular kind of NN with 5 layers of L1 , L2 , L3 , L4 and

L5 neurons each, respectively. The computational cost for training the
ANN using the backpropagation algorithm is O(Nx Ne (L1 L2 + L2 L3 +
L3 L4 + L4 L5 )), where Ne is the number of epochs.

• OCCNN requires a two-step training with a different number of training

points in each phase, Nx1 and Nx2 respectively, with Nx1 > Nx2 as
explained in the previous section. Considering a NN of 4 layers with
L1 , L2 , L3 , and L4 neurons, respectively, the complexity is O((Nx1 +
Nx2 )Ne (L1 L2 + L2 L3 + L3 L4 )).

• OCCNN2 exploits the combination of ANN and OCCNN. Therefore,

O(Nx Ne (L1 L2 + L2 L3 + L3 L4 + L4 L5 ) + Nx2 Ne (L6 L7 + L7 L8 + L8 L9 ))
is the overall complexity, where L1 , L2 , L3 , L4 and L5 are the neurons
in the ANN layers, and L6 , L7 , L8 , and L9 are the ones in the OCCNN
layers.

In Table 4.3 a summary of the complexity of all the algorithms is reported,

along with the processing times experienced on a computer with 2, 4 GHz
Intel Core i5 processor and 2133 MHz LPDDR3 RAM.
4.3 Performance 53

Table 4.3: Computational complexity and processing time of the algorithms

Algorithm Complexity Time [s]

SSI O(Na ns (k(li)2 + k 0 (li)3 + 2n3s )) 12 334.29

2 3
PCA O(Nx s + s ) 3.55
KPCA O(Nx2 + 2Nx3 ) 8.71
2
GMM O(Nx M ) 2.50
ANN O(Nx Ne (L1 L2 + L2 L3 + L3 L4 + L4 L5 )) 26.26
OCCNN O((Nx1 + Nx2 )Ne (L1 L2 + L2 L3 + L3 L4 )) 135.65
2
OCCNN O(Nx Ne (L1 L2 +L2 L3 +L3 L4 +L4 L5 )+Nx2 Ne (L6 L7 + 54.86
L7 L8 + L8 L9 ))

4.3 Performance
In this section, the proposed algorithms are applied to the Z-24 bridge
data set to detect anomaly based on the fundamental frequencies estima-
tion [14,58,59]. The performance is evaluated through the following metrics,
described previously, considering only the test set through accuracy, preci-
sion, recall and F1 score. In these numerical results, the dataset is divided
into a training set and a test set. For this comparison, we decided not to
use a validation phase because it is not needed in OCCNN and OCCNN2
as there are no hyper-parameters to set. For the other algorithms, we set
their parameters to ensure the maximum accuracy on the test set. With this
methodology, we slightly overestimate the accuracy of PCA, KPCA, and
GMM. Despite this setup favor PCA, KPCA and GMM, in the following
paragraph it is shown that the proposed solution, OCCNN2 , overcame their
performance.
The feature space has dimension D = 2 because only the first two funda-
mental frequencies are considered (this decision will be widely discussed in
the next chapter), and unless otherwise specified the three data sets used for
training, test in normal condition, and damaged condition, have cardinality
Nx = 2399, Ny = 854, and Nu = 854, respectively. For PCA, the number of
components selected is P = 1. For KPCA, after several tests the values of P
and γ that ensure the minimum reconstruction error are P = 3 and γ = 8.
54 Z-24 Bridge

1
Step 1 Step 2 OCCNN
0.8
0.6
EX

0.4
0.2
0
1 5, 000 10, 000
1
Step 1 Step 2 OCCNN2
0.8
0.6
EX

0.4
0.2
0
1 5, 000 10, 000
Epochs

Figure 4.3: Error function evolution over the epochs during training.

For GMM the order of the model that maximize performance is M = 10.
Regarding the ANN, we adopted a fully connected network with 5 layers of,
respectively, 100, 50, 1, 50 and 100 neurons, with ReLU activation functions.
Note that, the same ANN is also used for the first step of the OCCNN2 al-
gorithm. The parameters of the OCCNN and OCCNN2 are set accordingly
to [36], resulting in α1 = 0.3, α2 = 0.8, and are largely independent on the
spatial distribution of the feature points; this allows to presume that such
OCCs can work on different structures and bridges. The NN has 2 hidden
layers with L = 50 neurons each. All the NNs are trained for a number of
epochs Ne = 5000 with a learning rate ρ = 0.05. The error function adopted
for a training set X is

Nx X
X C
EX = − tn,c ln e
tn,c (4.1)
n=1 c=1

where Nx is the number of points in the training set, C is the number of

classes (C = 2), tn,c = 1 if the nth acquisition belongs to the cth class and
zero otherwise, and etn,c is the activation function value for point n of the cth
4.3 Performance 55

Table 4.4: Algorithm, parameter, value, and description

Algorithm Parameter Value Description

SSI n [2, 160] System order

Clustering K [2, 6] Number of clusters
Tracking Nt 200 Window size
Tracking Bf 0.4 Hz Frequency bin bandwidth
Tracking φf 0.01 Hz Kernel width
Tracking  0.7 Innovation
PCA P 1 Number of selected components
KPCA γ 8 Gaussian kernel width
KPCA P 3 Number of selected components
GMM M 10 Gaussian mixture model order
ANN layers 5 Number of hidden layers
ANN neurons {100, 50, 1, 50, 100} Number of neurons
NN layers 2 Number of hidden layers
NN neurons {50, 50} Number of neurons
Training Ne 5000 Number of training epochs
Training ρ 0.05 Learning rate
OCCNN α1 0.3 Density factor first iteration
OCCNN α2 0.8 Density factor second iteration
2
OCCNN α2 0.8 Density factor second iteration

output neuron [30]. As can be seen in Fig. 4.3 for each step of the algorithms
the networks are trained for Ne = 5000 epochs, quite enough for the error
function to reach the minimum. A target false alarm rate on the training set
is fixed equal to 0.01, from which a threshold is selected for each algorithm
to guarantee such constraint. In Table 4.4 the principal hyper-parameters
are summarized.
56 Z-24 Bridge

OCCNN OCCNN2 PCA

KPCA GMM ANN

0.89
0.95
F1 0.95
0.45
0.96
0.94
0.8
0.93
Rec 0.92
0.29
0.94
0.95
0.99
0.98
Prec 0.99
0.95
0.98
0.94
0.89
0.95
Acc 0.95
0.62
0.96
0.94

Figure 4.4: Comparison of the classification algorithms in terms of F1 score,

recall, precision, and accuracy.

4.3.1 Algorithm Comparison

The performance comparison of the algorithms is reported in Fig. 4.4. As we

can see, considering the F1 score OCCNN outperforms PCA and ANN but
is overtaken by KPCA and GMM. This happens because the non-uniform
data distribution of the training set influences Pollard’s estimator which over-
estimates the density. As a result, there are too many adversarial points
generated, which lead to an underestimation of the normal class boundary.
The OCCNN2 overcomes this limitation and shows the best F1 score and
accuracy. This represents a remarkable result considering the low number of
hyper-parameters to be tuned.
4.3 Performance 57

0,8

0,6
F1 score

0,4

0,2
OCCNN OCCNN2 PCA
KPCA GMM ANN
0
500 1000 1500 2000 2500
Training Points
Figure 4.5: F1 score varying the number of points, Nx , used for training, with
Ny = 854 and Nu = 854.

4.3.2 Impact of the training set and responsiveness

In Fig. 4.5 and Fig. 4.6, the F1 score dependence on the number of points
used to train the algorithms, Nx , and the number of anomaly points, Nu ,
used to test the algorithms, are depicted.
The number of training points, Nx , in Fig. 4.5 varies from 250 to 2399 with
steps of 250 points (except for the last step, equal to 149). This plot shows
how many points are necessary for the correct estimation of the normal class
boundaries. In these results, all the numbers of points used for test are kept
constant (i.e. Ny = Nu = 854) to maintain the same boundaries between
the classes. As it can be seen in the figure, all the algorithms require a high
number of data points corresponding to long-lasting monitoring (almost one
year [14]) to provide the highest F1 score. The algorithms that show a sharper
increase in their performance, hence requiring less training points, are KPCA
and GMM.
In Fig. 4.6, the number of training points is kept constant, Nx = 2399,
58 Z-24 Bridge

0.8

0.6
F1 score

0.4

0.2

0 OCCNN OCCNN2 PCA

KPCA GMM ANN
200 400 600 800
Damaged Points
Figure 4.6: F1 score varying the number of points during damage, Nu , used
to test the algorithms, with Nx = 2399 and Ny = 854.

and the number of anomaly test points, Nu , varies from 86 to 854 with 10
points steps. The plot shows that OCCNN2 exhibits the best responsiveness
(i.e. it requires the lowest number of damaged points to achieve higher values
of the F1 score).

Now that the performance has been analyzed in detail, the following
question arises:

Q8: Is it possible to automatically select the most reliable fundamental fre-

quencies, among all the ones extracted, in order to reduce the problem
dimensionality without affecting the performance?
Chapter 5

Dimensionality Reduction

In this section, we introduce two possible strategies to reduce the dimension-

ality of the damage-sensitive features extracted, with the intention of reduc-
ing the amount of data stored and transmitted through the sensor network
and increase the anomaly detection performance or at least do not deteriorate
it. Moreover, in case of wireless monitoring network, several strategies can
be adopted to manage the sensors and increase the network life time [60–63],
but firstly, a good practice can be to define the minimum number of sensors
necessary to accomplish the anomaly detection task.

5.0.1 Feature extraction

This technique consists of mapping a set of data in a low-dimensional fea-
ture space, trying to reduce an error function that represents the distance
between the original data and the remapped data obtained from the map-
ping subspace. In this work, we decided to use the ANN as auto-encoders to
accomplish this task [29, 31]. In this case, the normalized feature matrix X
is fed to an ANN with the classic bottleneck structure that provides a map-
ping set of layers and a consequent set of demapping layers. The input and
output layers have the same dimension of the feature space, and the labels
during the training phase must be set equals to the input data point. With
this structure, the data are mapped in lower-dimensional feature space (with
dimension equal to the number of neurons present in the bottleneck layer)

59
60 Dimensionality Reduction

and then reconstructed through the demapping layers minimizing the error
with respect to the input data.

5.0.2 Feature selection

This approach proposes to select the most reliable features among all the
available ones with some metrics to eliminate noisy components that can
deteriorate the damage detection capability of the OCC algorithms. In this
thesis, we decided to evaluate some statistical features (i.e., variance, kurto-
sis, and entropy) to identify the most reliable fundamental frequencies among
the four extracted, which are widely discussed in the next section.

5.1 Performance
The performance is evaluated through the following metrics, described pre-
viously, considering only the test set:

TP + TN
Accuracy =
TP + TN + FP + FN

TP
Precision =
TP + FP
TP
Recall =
TP + FN
Recall · Precision
F1 score = 2 ·
Recall + Precision
where TP , TN , FP , and FN , represent, respectively, true positive, true nega-
tive, false positive, and false negative predictions.
The feature space has dimension D = 4, the three data sets used for
training, test in normal condition, and damaged condition, have cardinality
Nx = 2399, Ny = 854, and Nu = 854. For PCA, the number of components
selected is P = 1. For KPCA, after several tests the values of P and γ that
ensure the minimum reconstruction error are P = 3 and γ = 8. Regarding
the ANN we adopted a fully connected network with 7 layers of, respectively,
5.1 Performance 61

0,8

0,6
F1 score

0,4

0,2

PCA KPCA ANN

0
f4321 f432 f431 f421 f321 f43 f42 f41 f32 f31 f21
Considered Frequencies
Figure 5.1: F1 score varying the considered fundamental frequencies, vertical
dashed red lines indicate the best configuration with 4, 3, and 2 frequencies.

50, 20, 10, k, 10, 20 and 50 neurons with k number of features extracted,
with ReLU activation functions for the feature extraction task, and a fully
connected network with 5 layers of, respectively, 100, 50, 1, 50 and 100 for
the anomaly detection. All the NNs are trained for a number of epochs
Ne = 5000 with a learning rate ρ = 0.05. The error function adopted for a
training set X is
XNx X C
EX = − tn,c ln e
tn,c , (5.1)
n=1 c=1

where Nx is the number of points in the training set, C is the number of

classes (C = 2), tn,c = 1 if the nth acquisition belongs to the cth class and
zero otherwise, and etn,c is the activation function value for point n of the cth
output neuron [30].
62 Dimensionality Reduction

0,02
f1 f2 f3 f4
Variance

0,01

0,00
2 20 40 60 80 100 120 140 160 180 200
Nc

Figure 5.2: Modal frequencies variance for different points; vertical dashed
red lines indicate the minimum number of points for the correct frequency
sorting.

5.1.1 Frequencies selection

First of all, a brute force approach is implemented to evaluate the fundamen-

tal frequencies that provide the best performance of the anomaly detection
algorithms in terms of F1 score with all the possible combinations of features.
As reported in Fig. 5.1, the best performance is achieved with algorithms with
the same feature configuration, highlighted by the red vertical dashed lines,
that allows sorting the features with increasing importance as f4 , f3 , f2 , and
f1 . This is why a good metric to describe the fundamental frequencies re-
liability must sort the frequencies in the same order. With this aim, three
statistical metrics are reported as a good candidate to accomplish this task.
The first one is the sample variance of the frequencies extracted defined
as
(i)
− f¯s )2
PNc
i=1 (fs
Variance =
Nc − 1
5.1 Performance 63

2
f1 f2 f3 f4

1
Skewness

−1

−2
2 20 40 60 80 100 120 140 160 180 200
Nc

Figure 5.3: Modal frequencies skewness varying the number of points; verti-
cal dashed red lines indicate the minimum number of points for the correct
frequency sorting.

(i)
where Nc is the number of acquisition considered, fs is the ith acquisition
of the sth fundamental frequency, and f¯s stands for the mean value of the
sth frequency evaluated in the interval {1, . . . , Nc }. As shown in Fig. 5.2,
this feature works well after Nc = 30 observations; as depicted, the variance
of reliable features is lower with respect to the noisy ones; so, this method
can be successfully used to sort the frequencies in the correct order.

The second metric proposed is the skewness that measures the asymmetry
of the probability distribution, i.e.

(i)
− f¯s )3
1
PNc
Nc i=1 (fs
Skewness = » 3 .
(i)
− f¯s )2
1
PNc
Nc i=1 (fs

As shown in Fig. 5.3, this metric can also be used to sort the frequencies
correctly; the method becomes reliable after around Nc = 100 observations.
64 Dimensionality Reduction

0
f1 f2 f3 f4

−250
Entropy

−500

−750

−1000
2 20 40 60 80 100 120 140 160 180 200
Nc

Figure 5.4: Modal frequencies entropy varying the number of points.

Finally, in Fig. 5.4 the entropy is evaluated as a further metric

Nc
X
Entropy = − P (fs(i) ) log10 P (fs(i) )
i=1

(i)
where the probability density function P (fs ) is evaluated numerically im-
plementing data binning. In this case, the trend is descendent because the
information introduced by new measurements decreases by increasing the
number of observations. This metric can also be used to sort the frequencies,
but in this case, f3 and f4 are inverted for some values of Nc and f2 is very
close to the previous two and can be missorted.

5.1.2 Algorithm Comparison

This paragraph provides a performance comparison of the anomaly detection
algorithms and dimensionality reduction techniques. In Fig. 5.5 the perfor-
mance of PCA, KPCA, and ANN is evaluated in term of F1 score, varying
the dimensionality of the features considered. Dashed curves are referred to
5.1 Performance 65

0,8

0,6
F1 score

0,4

0,2

PCA KPCA ANN

0
2 3 4
Features
Figure 5.5: F1 score varying the dimensionality of the feature space, dashed
lines correspond to frequency extraction technique, and continuous lines are
referred to as the frequency extraction approach.

features extracted with auto-encoder implemented with ANN. As we can

observe from the plot, the dimensionality reduction slightly decreases the
algorithms’ performance; hence, for this application, feature extraction does
not improve the system’s anomaly detection capability. On the contrary, con-
tinuous lines are referred to feature selected through the metrics described
in the previous paragraph. In this case, selecting the correct fundamental
frequencies tends to increase the detection performance, so it is strongly rec-
ommended for this application because both reduce the dimensionality of
the problem and increase the anomaly detection capability. In Fig. 5.6 the
performance of the algorithms in terms of F1 score, recall, precision, and
accuracy is reported in the best configuration, hence considering only the
first two fundamental frequencies selected by the described metrics (f1 and
f2 ). As we can see, the F1 score of both KPCA and ANN are around 95%,
and the accuracy for ANN is greater than 96% that represents a remarkable
66 Dimensionality Reduction

0.95
F1 0.95
0.59

0.95
Rec 0.94
0.43

0.95
Prec 0.97
0.98

0.96
Acc 0.95
0.68

PCA KPCA ANN

Figure 5.6: Comparison of the classification algorithms in terms of F1 score,
recall, precision, and accuracy with the best feature configuration.

result in this anomaly detection application.

The performance assessment of the automatic detection of anomalies re-

vealed promising results. However, a widespread use of such techniques is a
large-scale monitoring scenario poses the following research questions:

Q9: Is it possible to further reduce the amount of data that must be stored
by the network?

Q10: Is there a possibility to reduce the network infrastructure costs, reducing

the amount of sensors necessary to detect a damage or using low-cost
sensors?
Chapter 6

Data Management

This chapter analyzes the minimum amount of data that must be stored to
perform anomaly detection on the vibrational waveforms, and some strategies
that can be implemented to reduce such volume of data, both representing
important topics often studied in literature [6, 64].
Considering a network of l = 8 synchronized sensors interconnected to a
coordinator that store the accelerometric measurements, where each sensor
acquire Ns = 65536 samples each acquisition with Nb = 16 resolution bits
for Na = 4107 acquisition, it is trivial to observe that the total amount of
data stored by the coordinator is equal to Mt = Ns Nb Na l ' 32 Gbit = 4 GB.
This considerable amount of data has been stored in a year of non continuous
measurements. In fact the effective acquisition time can be estimated as Tt =
Ta Na ' 44860 m ' 448 h, that shows how huge could be the volume of data in
a continuous measurement system; in a year it is around 47 GB. To reduce the
amount of data the first step is the decimation approach. In this application
the fundamental frequency of the bridge falls in the interval [0, 20] Hz, so
to respect the Shannon’s theorem with a guard band of 5 Hz a sampling
frequency fsamp = 50 Hz it is enough to capture the bridge oscillations, since
the measurements are acquired by accelerometers with fsamp = 100 Hz, a
decimation by factor 2 can be adopted and the total amount of data is halved:
Md = Mt /2 ' 2 GB. Moreover, starting from the decimated acquisitions,
three other parameters can be modified to find some possible configuration

67
68 Data Management

5 Sensors 8 Bit 3000 Samples

Reference: 8 Sensors, 16 bit, 32000 samples
1 1 1
1

0 0 0
[Hz]

−1 −1 −1
0 −1 0 1 −1 0 1 −1 0 1
200 Samples
(a)

2 Sensors 3 Bit
f2

1 1 1
Training
Standard Test
Anomaly Test 0 0 0
−1
−1 0 1
(a)
f1 [Hz] −1 −1 −1
−1 0 1 −1 0 1 −1 0 1

Figure 6.1: Examples of feature transformation due to the effect of low num-
ber of sensors, low number of bits, and low number of samples with respect
to the standard measurement condition reported on the left.

that does not deteriorate the performance of the OCC algorithms, but can
reduce the volume of data:

• The number of sensors l can be reduced in order to contain the network

costs and the amount of data produced by the sensor network.

• The number of samples Ns or equivalently the acquisition time Ta can be

reduced to limit the network’s operative time, the energy consumption,
and the data produced.

• The number of bits Nb can be reduced to drop the storage occupation

of the single acquisition and contain the accelerometer cost.

All these possibilities will be analyzed and widely discussed in the next
section. In Fig. 6.1 some working points of the system are reported and
compared to the reference working condition after decimation (l = 8, Nd =
32768, Nb = 16).

6.1 Performance
The performance is evaluated through the accuracy, considering only the test
set. The feature space has dimension D = 2, and the three data sets used for
6.1 Performance 69

Figure 6.2: Error produced removing the selected sensor.

training, test in normal condition, and damaged condition, have cardinality

Nx = 2399, Ny = 854, and Nu = 854. For PCA, the number of components
selected is P = 1. For KPCA, after several tests, the values of P and γ that
ensure the minimum reconstruction error are P = 3 and γ = 8. For GMM
the order of the model that maximizes performance is M = 10. Regarding
the OCCNN2 the first step boundary estimation is made by a fully connected
ANN with 7 layers of, respectively, 50, 20, 10, 1, 10, 20 and 50 neurons, with
ReLU activation functions, and a fully connected NN with 2 hidden layers
with L = 50 neurons, each one for the second step. All the NNs are trained
for a number of epochs Ne = 5000 with a learning rate ρ = 0.05. The error
adopted to evaluate the points displacement in the feature space from the
original position due to the different configurations is the RMSE, defined as
Ã
Na
Ns X
1 X (n) (n)
Ef = √ (fs − f¯s )2 (6.1)
Na Ns s=1 n=1

(n)
where Ns is the number of features (Ns = 2), fs is the sth feature of the
(n)
nth acquisition in the initial configuration, and f¯s is the relative data point
in the modified configuration.

6.1.1 Sensors Relevance

Before evaluating the error introduced by reducing the number of sensors,
it is fundamental to evaluate the importance that each one has in the fun-
70 Data Management

0.1 1

0.08 0.8

Accuracy
0.06 0.6
Ef [Hz]

Error
OCCNN2
0.04 PCA 0.4
KPCA
GMM
0.02 0.2

0 0
1 2 3 4 5 6 7 8
l

Figure 6.3: Error varying the number of sensors available.

damental frequencies estimation. It is widely known in literature that the

sensor position strongly affects the mode estimation [14, 65, 66]. To verify
the sensors’ relevance, we removed them one-by-one and evaluated the error
resulting in the feature space points with respect to the standard condi-
tion. The error is calculated as the RMSE defined previously. As it can
be seen in Fig. 6.2, sensor S10 generates the most significant error in the
fundamental frequencies extraction when removed. With this technique it
is possible to sort the sensors from the most relevant to the least one as
follows: S10 , S03 , S16 , S14 , S05 , S12 , S06 , S07 . To evaluate the performance
with respect to the number of sensors used to extract fundamental frequen-
cies, the sensors will be removed in the same order, to always consider the
worst condition with the given number of sensors.

6.1.2 Number of Sensors

Now that the sensor relevance is defined, we are able to verify the perfor-
mance by varying the number of sensors used on the structure to derive the
fundamental frequencies with low error. As we can see in Fig. 6.3, the ac-
6.1 Performance 71

0.1 1

0.08 0.8

Accuracy
0.06 0.6
Ef [Hz]

Error
OCCNN2
0.04 PCA 0.4
KPCA
GMM
0.02 0.2

200 500 1000 1500 2000 2500 3000

Ns

Figure 6.4: Error varying the number of samples.

curacy of the algorithms remains almost the same as long as the number of
sensors available is greater than 2, as the error presents a significant increase
in correspondence of the gap between 2 and 3 sensors. Thus we can deduce
that the minimum number of sensors that can be used to monitor the Z-24
bridge is equal to 3. In this configuration it is easy to notice that the amount
of data stored is reduced to Msen = Ns Nb Na 3 ' 0.8 GB.

6.1.3 Number of Samples

To evaluate the effect of the acquisition time on the algorithms’ anomaly
detection performance, we progressively reduced the number of samples used
to extract the structure’s fundamental frequencies. As we can see in Fig. 6.4,
the performance of the algorithms remains almost constant as long as the
number of samples Ns is greater than 600, which corresponds to an acquisition
time of 12 s with a sampling frequency fsamp = 50 Hz. By reducing the
acquisition time drastically without relevant loss of performance, we achieve
a strong reduction of data occupation, that in this configuration is equal to
Msam = 600Nb Na l ' 0.04 GB without performance degradation.
72 Data Management

0.1 1

0.08 0.8

Accuracy
0.06 0.6
Ef [Hz]

Error
OCCNN2
0.04 PCA 0.4
KPCA
GMM
0.02 0.2

0 0
3 4 5 6 7 8 9 10 11 12 13 14 15 16
Nb

Figure 6.5: Error varying the number of bits.

6.1.4 Number of Bits

The number of bits can also be dropped to reduce the volume of data stored
and the single accelerometer cost. To test their impact on the performance,
we progressively reduced the number of bits used to encode the waveforms
extracted from the accelerometers as reported in Fig. 6.5. As we can see,
the error remains contained as far as the number of bits used to encode the
samples is greater than 6; likewise, the accuracy of the algorithms remains
high as long as the error is contained. Several low-cost accelerometers are
available on the market with a resolution Nb = 8, and these results show
that this type of sensor could accomplish the anomaly detection task. In this
case, the data occupation is Mbit = 8Ns Na l ' 1 GB.

6.2 Observations
Three different approaches are proposed to reduce the volume of data stored
and limit the costs of sensors and network infrastructure necessary to mon-
itor the structure. In this sense, when the goal is reducing the amount of
6.2 Observations 73

data stored, it is good practice cut down the observation time using a number
of accurate sensors; when the target is to minimize the sensor cost, a good
practice is to adopt low-cost sensors (with low resolution or number of bits
per sample), combined with long observation time with a network of several
sensors; when the objective is to contain the network infrastructure cost, a
low number of accurate sensors and long observation time can be considered.
To evaluate the error introduced from these strategies and the performance
of the algorithms, the RMSE and the accuracy are used as metrics. The
results show that these strategies can be adopted without significant loss of
performance; in fact, all the algorithms except the PCA, ensure accuracy
greater than 94% in all the proposed configurations, with the maximum per-
formance reached by OCCNN2 whose accuracy never goes down below 95%.

The effectiveness of this approach led us to poses the following question:

Q11: Is it possible to exploit accelerometric measurements acquired by senors

installed on a structure also to infer human activities?
74 Data Management
Chapter 7

Human Activities Classification

Using Biaxial Seismic Sensors

7.1 Introduction
The problem of identifying and classifying the presence of a target in a par-
ticular environment with low-cost sensors remains a key issue for outdoor
security applications [35, 36, 67, 68]. The variability of the ground and en-
vironment characteristics (i.e., weather conditions, humidity, temperature,
wind speed) makes the target detection more complicated than a controlled
indoor environment.
In literature, many works propose to use networks of geophones (which
present weak dependencies from the environment) to capture the ground
vibration to detect the presence of persons in indoor scenarios [69] or to clas-
sify several vehicles with different weights in a well defined outdoor area [70].
In [71], a method for detecting intruders and predicting their activities out-
door using a seismic sensor is presented. Similarly, in [72], the objective is
to detect and classify different targets (e.g., humans, vehicles, and animals
led by a human) using seismic and passive infrared sensors. Other solutions
exploit cooperative sensors (i.e., microphones and geophones) and data fu-
sion techniques to improve vehicle classification accuracy and estimate their
velocity [73,74]. In [75], a system architecture for the classification of moving

75
76 Human Activities Classification Using Biaxial Seismic Sensors

objects using both scalar and multimedia sensors is proposed.

In this work, we aim to distinguish between four different human activ-
ities, ride a bike, drive a car, walk, run, and investigate the possibility of
doing it using only a biaxial geophone (i.e., with two channels, horizontal
and vertical). In particular, the intuition behind this work is that forces in-
volved in the activities based on sliding contact with the ground solicit mainly
horizontal ground vibrations, while activities characterized by footsteps are
responsible for vertical vibrations. The consequences of this consideration on
the solution proposed are explained more in-depth in Section 7.3. The pre-
sented solution is based on the dimensionality reduction of frequency-domain
features of the data, based on PCA, to ensure good classification performance
at low computational cost [76]. Then, two different classification methods,
using two distinct classifiers for each one, are presented.
The main contributions of the letter are the following. i) We propose a
human activities classification system that uses only a biaxial geophone; ii)
We propose the PSD as a sensitive feature for the activity identification, and
PCA to remap the data in an advantageous feature space; iii) We propose two
different classification methods: classification made with a single classifier or
using a cascade of two classifiers; iv) For both methods, we compare two dis-
tinct classifiers, support vector machine (SVM) and k-NN, and we measure
the impact of vertical and horizontal components on classification perfor-
mance. Moreover, we present a comparison with conventional algorithms
that exploit cross-correlation, named template matching, and an alternative
classification method based on linear discriminant analysis (LDA).

7.2 System Model

As depicted in Fig. 7.1, a passive geophone with two channels (vertical and
horizontal) captures ground vibrations in an outdoor environment. The ac-
quired data are then processed to classify between four human activities: ride
a bike, drive a car, walk, and run.
At first, the analog signals produced by the horizontal and vertical chan-
nels of the geophone, xh (t) and xv (t), pass through an analog to digital
7.2 System Model 77

FEATURES EXTRACTION
CONDITIONING
x h (t) xh Zh Sh S̃h ŷ
STACKING
ADC & FFT PCA CLASSIFIER
NORM.
x v (t) xv Zv Sv S̃v

Figure 7.1: Illustration of the processing data chain to extract features from
the geophone signals and perform classification.

converter (ADC) with sampling frequency fs and resolution Nbit . The out-
put of the conversion consists in two time series xh = {xh (k/fs )}K−1 k=0 and
K−1
xv = {xv (k/fs )}k=0 of K samples. Subsequently, the time series are firstly
split in Nw row vectors, obtained through a partially overlapped sliding win-
dow of length W samples, with W ≤ K, and a sliding step ∆w , and then
rearranged into the matrices Zh and Zv , of size Nw × W , by stacking them.
From now on, since the processing stages apply to Zh and Zv separately, we
indicate both with Z for the sake of conciseness. In this phase, the samples
of each observation window are normalized column-wise such that the results
are zero mean row vectors with maxj |zn,j | = 1, n = 1, . . . , Nw . Finally, the
data are labeled by activity for the classification during the training phase.

7.2.1 PSD estimation

For the identification of relevant features for human activity classification, we
analyze the signal in the frequency domain [34]. In particular, we estimate
the PSD of the acquired samples using the weighted overlapped segment
averaging (WOSA) for each row of Z, with Hanning window and 50% overlap.
The number of points, D, of the fast Fourier transform (FFT) regulates the
trade-off between the frequency resolution and the PSD estimation accuracy
[77]. The estimated PSDs {sj }N w
j=1 are then organized in the matrix S of size
Nw ×D, generically used at this stage to indicate both Sh and Sv . Afterward,
PCA is applied to S to reduce the dimensionality of the data and extract
the features. From now on, we split Nw into two subsets No and Nt , for the
training and the test phases, respectively.
78 Human Activities Classification Using Biaxial Seismic Sensors

7.2.2 PCA
Principal component analysis (PCA) distills the essential information from
the dataset, which is then represented as a set of new orthogonal variables
called principal components obtained from a linear combination of the orig-
inal data [78]. For the calculation of the principal components, we consider
only the No points of the training subset. After centering the matrix S by
subtracting its column-wise sample mean, we evaluate the sample covari-
ance matrix Σ = N1o ST S. Then, Σ is factorized by eigenvalues decompo-
sition (EVD) Σ = QΛQT , where Λ is the diagonal matrix of eigenvalues,
ordered from the largest to the smallest, and Q is the matrix of eigenvec-
tors [79].
In order to perform dimensionality reduction, we only keep the first Dh
or Dv eigenvalues of Λ and the corresponding eigenvectors Q̃ (i.e., selected
columns of Q). The projections S̃ of the observations in the components
subspace through the new projection matrix Q̃ are S̃ = SQ̃, where S̃ is a
Nw × Dh or Nw × Dv matrix, and Dh , Dv ≤ D are the number of principal
components considered for S̃h and S̃v . Note that, while the principal com-
ponents are calculated solely over the training points No , all the Nw points
are projected in the components subspace. These two matrices represent the
selected features used to train and test the classifiers described in Section 7.3.

7.3 Classification Techniques

After dimensionality reduction, the features S̃h , S̃v are used by a classifier
to determine the type of human activity. In particular, two classification
methods are proposed.

• Single classifier. For each observation window, the data of the two
channels are jointly processed. The Nw × (Dh + Dv ) matrix S̃hv is built
concatenating the matrices S̃h and S̃v as S̃hv = [S̃h S̃v ]. The monoaxial
(i.e., horizontal or vertical) geophone configuration is obtained setting
Dv = 0 or Dh = 0, respectively.
7.3 Classification Techniques 79

• Cascade classifier. The classification is carried out in two steps.

Firstly, a three-class classifier uses the data acquired from the horizon-
tal channel S̃h to classify between bike, car, and footsteps (i.e., run and
walk are treated as if they are the same class). Then, if the first classi-
fier does not choose for the bike or car classes, a second one uses the S̃v
features to discriminate between run and walk. The rationale behind
this approach is that based on the dominant forces during the inter-
action between the target and the ground, all the activities based on
sliding contact (i.e., car and bike) tend to stimulate horizontal ground
vibrations. In contrast, activities characterized by footsteps tend to
excite vertical vibrations.

Hereafter, the two classifiers used in this work are briefly described. Be-
sides, ordinary cross-correlation based classification method used both in
time and frequency domain is reviewed. For the sake of clarity, y is the vec-
tor of actual classification labels of length Nw , ŷ is the vector of classification
labels estimated by the algorithms, s̃n is the n-th row of S̃h (either S̃v or S̃hv ),
and S is the feature space of dimension D (either Dh or Dv ). Thus, each
point, both for training and test, is represented by the pair (s̃n , yn ).

7.3.1 Support vector machine

The SVM constructs a set of hyperplanes in high-dimensional space that
can be used for tasks like classification or regression [29]. Hence, it is a
parametric learning algorithm whose error function includes a regularization
term as follows:
No
T
X Ä ä
g(w̃) = ln 1 + e−yn (s̃n w̃) + λ||w̃||22
n=1

where w̃ are the weights of the parametrical model, and λ is the regularization
parameter [29, 30]. The hyperplane that performs a good separation is the
one that maximizes its distance from the nearest training points of each class,
and it is identified by the set of weights w̃ that minimizes the error function.
Given a test point (s̃m , ym ), the estimated label ŷm is given by ŷm = s̃m w̃T .
80 Human Activities Classification Using Biaxial Seismic Sensors

7.3.2 k-NN

In k-NN a set of No pairs {(s̃n , yn )}No

n=1 is given as training set, where s̃n takes
values in the feature space S upon which is defined the metric d(s̃n , s̃m ); the
Euclidean distance in this work. Given a test point (s̃m , ym ), the estimate of
ym is given by the nearest neighbor training point with respect to the test
point as

ŷm = yk : s̃k = arg min d(s̃n , s̃m ) . (7.1)
s̃n

With (7.1) we can assign s̃m to the same class of the nearest s̃n . If the number
of training points is large enough, it makes sense to use the majority rule of
the nearest k neighbors, instead of the single nearest neighbors [80]. In the
case of binary classification, k is chosen to be odd to avoid that a point is
assigned to two different classes.

7.3.3 Template matching

Classification methods that use cross-correlation as a similarity metric, often

called template matching, are extensively used in signal processing [81, 82].
Similarity can be searched both in time and frequency domain. In the
frequency-domain case, at first the estimated PSDs, {sj }N w
j=1 , of both hor-
izontal and vertical channel, are normalized to have zero mean and unitary
standard deviation. Then, given a set of No pairs ({sn , yn )}N o
n=1 as train-
ing set, and a test point (sm , ym ), we define the vector r such that its n-th
element is
rn = max{corr(sn , sm )} n = 1, . . . , No (7.2)

where corr(sn , sm ) is the cross-correlation vector, of length 2D − 1, between

the sequences sn and sm [81, 82]. Carrying out the same operations for both
the horizontal and the vertical channel, we obtain the vectors rh and rv ,
respectively, which are then concatenated to obtain rhv = [rh rv ]. The esti-
mated label ŷm , is thus

ŷm = y(k−1)modNo +1 : k = arg max(rhv ) . (7.3)
n
7.4 Numerical Results 81

Similar operations are performed in the time-domain case.

7.4 Numerical Results

In this section, we present several tests to compare the performance of
the classifiers in different settings. For the measurements, a biaxial two-
channel geophone, with natural frequency of 4.5 Hz, frequency bandwidth
of 0.2 − 240 Hz, and sensitivity of 78 V/m/s has been used. The geophone
has been placed in a flowerbed next to a car park. All the activities have
been performed within a distance of 6 m from the geophone. The seismic
sensor was interfaced with the PC using an ADC with sampling frequency
fs = 400 Hz and resolution Nbit = 14 bit. Furthermore, for each channel
a LM 386 amplifier has been adopted to increase the signal-to-quantization
noise ratio.
Each activity has been repeated several times for a whole period equal
to Ta = 20 min, using different vehicles and involving people with various
height and gait. In particular, six people took part in the run and walk
measurements, while two bikes and a car, driven by two persons, were used.
Then, the acquisition period has been split into several observation windows
of duration Tob = W/fs ranging between 5 s to 30 s, depending on the test,
and using a sliding step of ∆w = 5 s; the PSDs was computed with a D = 64
points FFT.
To properly study the classifiers’ performance, the acquisitions have been
randomly split in training and test sets as:

• Training and validation: 60% of all PSDs of each activity are ran-
domly chosen to calculate the projection matrix of PCA, Q̃. Then,
the projected points are used to train the classifiers. To set the hyper-
parameters of SVM and k-NN, 10% of these points are randomly chosen
and then used to perform cross-validation. The resulting optimal values
are k = 3 and λ = 0.1.

• Test: the other 40% of PSDs are used to test the performance of the
algorithms. In this case, the projection in the principal components
82 Human Activities Classification Using Biaxial Seismic Sensors

0.98

0.94

0.9

0.86

0.82 C-SVM
η

S-SVM
0.78 S-SVM (V)
S-SVM (H)
0.74
C-k-NN
0.7 S-k-NN
S-k-NN (V)
0.66 S-k-NN (H)

2 3 4 5 6 7 8 9 10
Dv , Dh

Figure 7.2: Accuracy varying PCA components of the vertical and horizontal
channels, with Tob = 20 s.

subspace through PCA is made using the projection matrix computed

during the training phase.

The same ratio of training and test points has also been used for the template
matching based approaches and the LDA. However, no cross-validation is
required in these cases. As a figure of merit, to evaluate the performance of
the classifiers, we consider the accuracy, defined as

number of points correctly classified

η= . (7.4)
number of total points

7.4.1 Accuracy vs. number of PCA components

For both the horizontal and the vertical channel, the number of selected
components Dh and Dv have been varied and the effect on the accuracy of
the algorithms has been studied. In this test the observation window has been
set to W = 8000 samples, corresponding to an observation time Tob = 20 s.
7.4 Numerical Results 83

As shown in Fig. 7.2, the single SVM (S-SVM) classifier provides always
the best performance, for each value of Dh and Dv ; in particular, with only
Dh = Dv = 3 components, it reaches an accuracy greater than 96%, while
to achieve the same accuracy with the cascade SVM (C-SVM) configuration,
it is necessary to use a higher number of PCA components (Dh , Dv ≥ 5).
On the contrary, the single k-NN (S-k-NN) converges to an accuracy roughly
equal to 93%, much lower than that of the cascade k-NN (C-k-NN), which
stands at almost 97%. These results prove that, in the case of k-NN and
for this experiment setup, it is better to adopt the cascade solution and
use the samples coming from the channels separately. With regard to the
monoaxial scenario, we can see that for the SVM, the performance in case
we use only the samples of the vertical channel is comparable with those of
C-SVM and S-SVM solutions, when Dv = Dh = 10. On the contrary, if we
use only the horizontal channel, the accuracy is always worse. Differently,
the performance of k-NN is always better when using a biaxial geophone.

7.4.2 Accuracy vs. observation duration

In this test, the observation window duration varies between 5 s and 30 s. The
selected principal components are Dh = Dv = 5. As shown in Fig. 7.3, here
too, the performance of the single k-NN is worse than the cascade solution.
On the contrary, the accuracy S-SVM is always better than the cascade
solution, with a higher gap when Tob = 5 s. Moreover, we can notice that
for Tob > 20 s the accuracy of the cascade k-NN and of the single SVM are
comparable, while that of the cascade SVM decreases. The results for the
monoaxial geophone confirm what has been experimented in Section 7.4.1.
The performance of template matching based approaches, both in time
and frequency -domain, are included in Fig. 7.3 for comparison. As can
be noted, the traditional techniques which do not exploit data structure
are outperformed by the proposed solutions. In addition, the accuracy of
LDA in both cascade (C-LDA) and single (S-LDA) configurations, are also
included [83]. While for very long (> 30 s) observation windows, the S-
LDA classifier reaches the same performance as the proposed solution based
84 Human Activities Classification Using Biaxial Seismic Sensors

0.98

0.94

0.9

0.86

0.82
η

0.78 C-SVM S-SVM

S-SVM (V) S-SVM (H)
0.74
C-k-NN S-k-NN
0.7 S-k-NN (V) S-k-NN (H)
f. corr. t. corr.
0.66 C-LDA S-LDA

5 10 15 20 25 30
Tob [s]

Figure 7.3: Accuracy varying the observation window duration, Tob . For
PCA-based classifiers Dh = Dv = 5.

on PCA, followed by SVM, for short observations some of the proposed

solutions (C-SVM, S-SVM, and S-SVM (V)) outperform LDA. Note that,
C-LDA classification always exhibit poor performance almost equivalent to
template matching.

7.5 Observations
A passive human activity classification method exploiting the ground vibra-
tions observed by a two-channel geophone is proposed. Data collected by
the two channels are processed by a PCA-based dimensionality reduction
to extract significant features in the frequency domain. The classification
step is performed by either a single classifier or a cascade of two classifiers.
The analysis considered the most important parameters (observation window
and the number of PCA components) and setups (joint processing or cascade
processing) to provide a complete set of results which may assist the system
7.5 Observations 85

designer. Based on performance assessment on real waveforms, the extensive

numerical results show that the number of selected PCA components and
the observation window duration strongly impact the performance of the al-
gorithms. For example, for large observation windows, the best solution is
represented by the single SVM classifier that jointly processes both vertical
and horizontal data. However, as complexity may have a role in designing
autonomous low-power devices, if the k-NN classifier is preferred, the cascade
solution performs the best.

The flexibility of these tools led us to poses the following question:

Q12: Is it possible to apply similar techniques to solve other problems?

86 Human Activities Classification Using Biaxial Seismic Sensors
Chapter 8

Anomaly Detection Using WiFi

Signals of Opportunity

8.1 Introduction
With the advent of the technological revolution named Internet of things
(IoT), increasingly pervasive and context-adaptive communication systems
are conquering the radio-frequency (RF) spectrum [10]. Since spectrum
population may represent an issue in some frequency bands, e.g., the over-
crowded industrial, scientific and medical (ISM) ones, there is an increasing
interest in exploiting existing over-the-air signals, devised for some specific
purpose, to perform other tasks thus avoiding dedicated radio emissions.
WiFi routers, broadcast stations, and mobile cellular networks are only a
few examples of such signals of opportunity (SoOp) [84–87].
Security in homes, industrial environments, and facilities is becoming a
critical aspect of modern society, and for such reasons, ambient intelligence
is gaining attention recently [37]. Video-based surveillance systems using, for
example, cameras are the dominant technology in such scenarios. However,
the personal privacy issue is still a reason for deterring users. The ambient
intelligence paradigm is not only beneficial for security purposes but more
generally as an enabler for context-aware applications like smart homes, to
name one example.

87
88 Anomaly Detection Using WiFi Signals of Opportunity

The capability to extract information from the effects of the propagation

on RF signals opens up a way to acquire knowledge about an environment
by the observation of SoOp. In this context, there are two main charac-
teristics of the observed signal used for detection, one is the RSS, and the
other is channel-state information (CSI). The RSS is very easy to get with
simple hardware, so it gained considerable attention in the last decade. A
human motion localization method that exploits standard deviation of RSS
is presented in [88], [89], and the detection and tracking of multiple persons
in an indoor environment is proposed in [90]. However, techniques that ex-
ploit received power are susceptible to multipath propagation and need a
multitude of devices to be effective, even when confined in indoor environ-
ments [91], [92].
Channel estimation allows greater precision when used in motion detec-
tion compared to RSS measurements. In [93], fine-grained subcarrier infor-
mation (i.e., channel frequency response) is exploited to design a device-free
passive human detection. In [94], a scheme for adaptive indoor passive de-
tection is proposed, where the CSI amplitude measured in an indoor environ-
ment is shown to vary in the presence of human motion. In [95] the authors
propose a device-free RF environmental vision system based both on RSS
and CSI, while in [96] a crowd counting system that uses SoOp is presented.
Target/change detection can also be performed with radar techniques,
either using dedicated sources with large bandwidth [97], or using sources
of opportunities [98] with smaller bandwidth but in large environments that
ensure target/anomaly spatial resolution. However, our goal here is to avoid
dedicated signals with very large bandwidth and perform anomaly detection
with SoOp even when the target/anomaly is not spatially resolvable.
This work proposes a ML approach for anomaly detection in an indoor
environment using WiFi SoOp. In particular, the main contributions are the
following.

• We use inexpensive RF sensors to collect SoOp [35]. Environmental

changes are detected through RF channel modifications without de-
modulating the signal.
8.1 Introduction 89

Figure 8.1: The four scenarios considered with the anomaly represented by
a human being.

• In particular, we record and analyze samples that belong to beacon

packets transmitted by an access point (AP).

• We compare the performance of two ML classifiers such as PCA and

KPCA, as a function of the number of beacon packets collected [36].

• The tests have been performed in both line-of-sight (LOS) and non-
line-of-sight (NLOS) conditions.

• Finally, we show that the proposed approach exhibits superior perfor-

mance than a well-known RSS-based solution in terms of accuracy, even
using just a single sensor.
90 Anomaly Detection Using WiFi Signals of Opportunity

Table 8.1: Scenarios considered for anomaly detection.

Condition Scenario Target/Anomaly
Static in the middle of
(a)
LOS the direct path
(b) Moving in the room
Static in the middle of
(c)
NLOS the direct path
(d) Moving in the room

8.2 System Overview and Problem Setup

Let us consider a scenario of the ones depicted in Fig. 8.1, where a RF
sensor performs down-conversion followed by an analog-to-digital conversion
to capture samples of over-the-air WiFi signals. In particular, without loss
of generality, let us focus on the case where there is an AP operating in the
2.4 GHz band. While, nowadays, there are several high-performance devices
for spectrum monitoring, it is interesting to investigate the possibility of
using very low-cost devices [35]. Using inexpensive devices the measurement
quality is usually affected by front-end impairments, low sampling frequency,
low resolution, and low sensitivity. However, as is shown in Section 8.4, the
samples have been acquired with enough accuracy to allow the ML algorithms
to achieve high accuracy.
As case studies, let us focus on the four different scenarios depicted in
Fig. 8.1, where the visibility of the AP and the position of the target are
varying. In Table 8.1, a synthetic description of the scenarios is presented.
For the sake of conciseness, we refer to such scenarios as (a), (b), (c), and
(d), respectively.

8.2.1 Data pre-processing

The samples of the complex envelope of the received signal at the RF sensor,
ri , i ∈ N, are pre-processed to detect the time of arrival (ToA) of the beacon
8.2 System Overview and Problem Setup 91

packets emitted by the AP. In the RF sensor adopted, the sampling rate,
fs , can be varied according to the needs. On the one hand, a high sampling
frequency can guarantee a larger signal bandwidth at the cost of an increasing
computational rate required to process the samples. On the other hand, a
low sampling rate may alleviate the computational burden but result in poor
detection performance.

To detect changes in the propagation environment, we need to compare

the received packet samples (or their frequency representation) in the pres-
ence and absence of the anomaly. To facilitate this task, in principle, we
should have the same transmitted SoOp in both situations so that any change
in the received signal characteristics can be ascribed to a change in the chan-
nel. In the IEEE 802.11 standard, the AP sends beacons at regular intervals.
Beacons are special packets repeated over time, which contain the same infor-
mation, e.g., the network service set identifier (SSID). Our idea is to exploit
beacon packets as sources of opportunity for our anomaly detector. In litera-
ture, the most common approaches for intrusion detection are based on RSS
estimation. A non-parametric kernel-function based anomaly detection has
been used in several applications, such as [99], [92]. These methods do not
require to have the same transmitted SoOp because RSS estimation reflects
the energy superposition of multiple paths of the signal. In section 8.4, the
performance of the RSS-based method (named RASID) proposed in [99] is
compared to our novel solution based on PCA.

As shown in Fig. 8.2, the first step in the proposed approach is to extract
beacon packets from the observed samples. In order to keep the computa-
tional burden low, it is possible to use non-coherent detection of beacons
starting from the envelope of the received samples, |ri |, i ∈ N. After bea-
con detection from the ultra-dense over-the-air packet flow, NP samples are
extracted. In particular, denoting with NB the number of beacons detected,
we have
ToAj +NP
bj = {ri }i=ToA j
with j = 1, . . . , NB (8.1)

where bj is a vector containing the first NP samples of the j-th beacon while
ToAj is the j−th beacon time of arrival. We want to emphasize that precise
92 Anomaly Detection Using WiFi Signals of Opportunity

ri bj sj
PREAMBLE PSD
features
EXTRACTION ESTIMATION

Beacons as SoOp

|ri | |r̃i |
|·| BEACON
DETECTION

j =1
j =2
j =3
|ri|

|r‹i|
j =4

sj,n
i i n

Figure 8.2: Illustration of the pre-processing data chain to extract features

from beacon packets.

ToA estimation is not necessary as the detection algorithm proposed is rather

insensitive to time shifts and therefore to ToA uncertainty. Note that when
NP is relatively small, only a fraction of the beacon is captured, e.g., its
preamble.

8.2.2 Features extraction

Starting from the basic idea that any change in the environment reflects
on the radio channel response, we estimate the PSD of the received beacon
signals and train two well-known ML algorithms for change detection.
The PSD is estimated from the received samples {bj }N B
j=1 using WOSA
with Hanning window and 50% overlap. The segment length D < NP used
in WOSA is a parameter that can be tuned to trade-off between frequency
resolution and PSD estimation accuracy (D is the FFT length). The esti-
mated PSDs {sj }N j=1 are then organized in a matrix S of size NB × D. To
B

further mitigate the impact of outliers in training the algorithms, PSDs have
been averaged over M points and organized in vectors xk , as

kM
1 X
xk = sj with k = 1, . . . , N (8.2)
M
j=(k−1)M +1

where N is the number of points after the averaging operation, i.e., N =

NB /M . For a given M it is possible to estimate the time needed for target
8.3 Survey of ML Techniques 93

detection as
M
TD = (8.3)
BR EC
where BR is the beacon rate (number of beacons per second), and EC rep-
resents the number of beacons extracted over the number of beacons trans-
mitted.1 Finally, the features are organized in the matrix

X = [xT T T T
1 , x2 , . . . , xN ] (8.4)

of size N × D, where, from now on, N is the number of input points and D
is their dimension. Note that if M = 1 then X = S.

8.3 Survey of ML Techniques

In this section, we briefly review PCA and KPCA which are usually adopted
for classification problems, especially for pattern recognition and anomaly
detection. Both of them need a zero-mean training set Xm , that must be
obtained subtracting the mean value m = 1N ⊗ [m1 , m2 , ..., mD ], calculated
as follows
N
1 X
md = xn,d with d = 1, 2, ..., D (8.5)
N n=1

and therefore evaluated as Xm = X − m.

8.3.1 Principal Component Analysis

PCA is a widely known algorithm in exploratory data analysis. Given the
centered training set Xm , the algorithm remaps the training data from the
feature space RD in a subspace RP (where P < D is the number of principal
components selected) that minimize the information loss between the pro-
jected data and the original ones. The best subspace over which to project
the data depends on the training set distribution and the number of com-
ponents selected P . The algorithm starts evaluating the sample covariance
1
Beacon extraction may fail because of collisions so the actual fraction of beacons
selected can be less than one.
94 Anomaly Detection Using WiFi Signals of Opportunity

matrix C of the training set as

XT
m Xm
C= . (8.6)
N −1

The matrix C can be factorised by eigenvalue decomposition as

C = VΛVT (8.7)

where V is an orthonormal matrix whose columns are the eigenvectors, and

Λ is a diagonal matrix that contains the D eigenvalues. The eigenvalues mag-
nitude represent the importance of a particular component, hence selecting
the P eigenvectors corresponding to the P highest eigenvalues, we obtain the
best P -dimensional linear subspace over which to project the data. To do
this, we multiply the portion of eigenvector matrix selected

VP = [v1 , v2 , . . . , vP ] (8.8)

by the data that we want to project

XP = XVP . (8.9)

An illustration of the data projected in a two-dimensional subspace is re-

ported in Fig. 8.3. For each scenario, the blue points represent the features
extracted in the empty room case, while the red points refer to features ob-
served in the presence of a target. It is easy to see that as the two groups of
points are overlapped, anomaly detection can be challenging.

8.3.2 Kernel Principal Component Analysis

This approach takes inspiration by the standard PCA and overcomes the lim-
itation of the linear mapping that corresponds to finding linear boundaries
in the original feature space. In many applications, this constraint repre-
sents a severe limitation and can sharply decrease the classification accuracy.
KPCA firstly maps the data with a non-linear function, after which applies
8.3 Survey of ML Techniques 95

Figure 8.3: Examples of PCA outputs, in the four scenarios, when P = 2

principal components are selected. Blue circles refer to empty room points
while red circles refer to the presence of a target in the room.

the standard PCA to find a linear boundary in the new feature space. Such
boundary becomes non-linear, going back to the original feature space. A
crucial point in KPCA is the selection of a non-linear function that leads
to linearly separable data in the new feature space. In literature, when the
data distribution is unknown, the RBF kernel is often proposed as the right
candidate to accomplish this task [43]. Suppose we have a generic point z
that corresponds to a vector of length D, we can apply the RBF as follows

2
Kzj = e−γ||z−xj || with j = 1, 2, . . . , N (8.10)

where γ is a kernel parameter (inversely proportional to the width of the

Gaussian function) that must be appropriately set, and Kzj is the j-th com-
96 Anomaly Detection Using WiFi Signals of Opportunity

ponent of the point z in the kernel space. Overall the starting vector z is
mapped in a vector Kz of length N . Applying now the PCA described in
section 8.3.1 to the new data set obtained remapping all the training points,
it is possible to find non-linear boundaries in the starting feature space for
a better classification. It is good practice to center the points mapped with
the RBF because the mapping in the new feature space could be non-zero
mean.

8.4 Numerical Results

In this section, we present several tests intending to select the best parameters
setting in different working conditions and compare the performance of the
algorithms.
The RF sensor is represented by the software defined radio (SDR) device
HackRF One operating in receiving mode in the 2.4 GHz ISM band with a
bandwidth of 20 MHz which corresponds to fs = 20 MS/s. The sensor output
is composed by the in-phase and the in-quadrature baseband signals, each one
represented with 8 bit/sample. In the setting considered BR = 10 packets/s
and EC = 0.83. The scenario is represented by a room of size 4.8×4.8×2.8 m3
and the target considered is a person of 1.78 m height and 65 kg weight.
The next subsections present the accuracy varying the number of averaged
points M , used to make the detection in two different working conditions
and compare the performance also with the RSS-based technique [99], in
two cases, M = 2 and M = 8, respectively.2 The number of features is set
to D = 32, which corresponds to the number of frequency bins. In all the
experiments, the data are partitioned in the following sets:

• Training: composed of 50% of the data extracted in the empty room

case;

• Validation: composed of 25% of the empty room points and with the
same number of adversarial points (the adversarial class depends on
2
The accuracy is defined as the number of correct classifications over the number of
points classified.
8.4 Numerical Results 97

the working condition and the target type);

• Test: composed of the remaining empty room points and the same
number of adversarial points.

In this configuration, training, validation, and test sets have the same number
of points, N = 400, when averaging is not performed (M = 1). For each
simulation point, 100 Monte Carlo iterations are performed, and the data
partitioning is repeated, selecting random points in each iteration. Given a
generic point z and its reconstructed version e z = zP VPT from its projection
zP , the error function ez is defined as

ez = ||z − e
z||. (8.11)

Evaluating the error function over the validation set, we selected a threshold
to ensure a false alarm (FA) probability lower than 0.01 on the training set.

8.4.1 AP-sensor link in line-of-sight

Firstly, the scenarios showed in Fig. 8.1a and Fig. 8.1b, where the AP and the
RF sensor are in LOS, are analyzed. In these scenarios, the PCA, KPCA,
and RSS-based algorithms are trained varying the number of beacons M
considered for the classification of one point. For each value of M the num-
ber of components P of the PCA algorithm is varied from 1 to D − 1, and
the accuracy is tested on the validation set. Hereafter, the value of P that
maximizes the validation accuracy with a particular value of M is used to
evaluate the performance on the test set. The same procedure is taken for
the KPCA to select the best configuration of the parameters P and γ (the
kernel parameter is varied from 5 to 40 with steps of 5). The parameters used
are listed in Table 8.2. A Gaussian kernel for RSS-based method has been
chosen, and the normal distribution approximation for the bandwidth has
been adopted [100]. The RSS has been calculated over the acquired beacon
samples. As shown in Fig. 8.4 both PCA and KPCA provide comparable
98 Anomaly Detection Using WiFi Signals of Opportunity

Table 8.2: PCA and KPCA parameters for the considered scenarios.

M
1 2 3 4 5 6 7 8 9 10 11 12
PCA P 4 2 1 1 1 1 1 1 1 1 1 1

(a) P 6 2 1 1 2 2 2 2 2 2 2 2
KPCA
γ 30 10 5 5 5 5 5 5 5 5 5 5
PCA P 4 2 1 1 1 1 1 1 1 1 1 1

(b) P 10 2 1 1 2 2 2 2 2 2 2 2
KPCA
γ 25 20 15 15 10 10 5 5 5 5 5 5
PCA P 2 1 1 1 1 1 1 1 1 1 1 1

(c) P 4 4 3 3 2 2 2 2 2 2 2 1
KPCA
γ 10 10 10 20 10 10 15 5 15 5 25 5
PCA P 2 1 1 1 1 1 1 1 1 1 1 1

(d) P 4 3 3 3 2 2 2 2 2 2 1 2
KPCA
γ 10 15 10 10 5 10 10 5 10 5 20 10

performance for values of M greater than 3, that corresponds to data acqui-

sition of 0.4 s and overcome the accuracy of the RSS-based technique both
in scenarios (a) and (b). For lower values of M , the KPCA presents higher
accuracy with respect to PCA.

8.4.2 AP-sensor link in non-line-of-sight

In the NLOS case depicted in Fig. 8.1c and Fig. 8.1d, an obstacle between
the AP and the RF sensor is present. Even in this scenario PCA, KPCA,
and RSS-based method are trained and optimized with the same procedure
described in the previous subsection. As we can see from Fig. 8.5, the per-
formances are different from the LOS case. In this configuration, the PCA
8.4 Numerical Results 99

Figure 8.4: Accuracy in scenarios (a) and (b), varying M .

accuracy with static target slightly decrease, instead KPCA works well also
for low M . For M = 2 in the case of a static object, we experience a remark-
able performance degradation for PCA while KPCA preserves high accuracy.
This effect is not present in the case of a moving target. This happens be-
cause the absence of a direct path between the transmitter and the RF sensor
makes the channel transfer function strongly influenced by the target that
intersects the reflected paths. Therefore, the detection of a moving target,
for low values of M , is favored by a NLOS scenario: presumably, the inter-
section of a reflected path is more likely to happen than the intersection of
the direct path in the LOS case. Even in these scenarios, the RSS-based
technique accuracy is overwhelmed by the proposed approaches, particularly
for low values of M .
100 Anomaly Detection Using WiFi Signals of Opportunity

Figure 8.5: Accuracy in scenarios (c) and (d), varying M .

8.4.3 Beacon average

For the sake of selecting the best solution in each scenario, in this subsection,
we compare the performance of the algorithms, with M = 2 when the main
goal is a fast detection of the target and M = 8 when the objective is the
maximization of the classifier accuracy. As we can see in Fig. 8.6, the KPCA
algorithm provides a better solution when a timely detection is needed (blue
curve), in this case, the accuracy is always greater than 95%. The worst-case
for the PCA algorithm is the NLOS scenario with a static target where the
accuracy goes down to 75%. In the case of M = 8, the accuracy generally
increases in all the scenarios for all the algorithms, because the average over a
higher number of points reduces the outliers and hence increases the distance
between points in different classes, particularly PCA sharply increase its
performance becoming comparable with the KPCA, instead the RSS-based
solution experiences a saturation effect with no increment in the performance.
With large values of M , the PCA performance is comparable with that of
8.5 Observations 101

Accuracy
PCA(a)
KPCA(d) PCA(b)
100

90
KPCA(c) PCA(c)
80

70

60

KPCA(b) 50 PCA(d)

KPCA(a) RSS(a)

RSS(d) RSS(b)
RSS(c)

M =2 M =8

Figure 8.6: Accuracy of PCA, KPCA, and RSS-based method varying M , in

the four scenarios.

KPCA in all the scenarios, suggesting the use of PCA that is less complex
and computationally faster. Finally, it is also important to highlight that the
PCA presents lower accuracy in the case of NLOS with respect to the LOS
one when M is low. This happens because the interception of the direct path
strongly changes the channel transfer function when the target is present.

8.5 Observations
In this work, we proposed and studied an RF-based automatic indoor anomaly
detector exploiting SoOp from WiFi devices and machine learning techniques.
The numerical results based on real waveforms, collected by a RF sensor,
102 Anomaly Detection Using WiFi Signals of Opportunity

demonstrated that the detection of changes in the environment is possible

and that it does not require expensive devices. Moreover, the accuracy of
PCA reaches 90% at relatively low observation intervals (i.e., 0.4 s), while for
KPCA the observation of one beacon is enough to ensure an accuracy greater
than 95%. Finally, numerical results show that the proposed approaches ex-
hibit superior performance than the RSS-based solution in terms of accuracy,
even using only one RF sensor.
Chapter 9

Conclusion

With this thesis, we proposed an automated approach to perform anomaly

detection for SHM. Several ML algorithms have been tested and compared
on a widely known dataset, the Z-24 accelerometric measurements.
More in detail, in chapter 2, we provided an overview of several machine
learning algorithms and their general applications (regression, classification,
dimensionality reduction) with a particular focus on anomaly detection. In
this branch of ML, we contributed with two new anomaly detectors named
OCCNN and OCCNN2 that in low dimensional feature spaces provide better
performance compared to the classic solutions available in the literature.
Both approaches exploit the flexibility of the NN classifiers in the anomaly
detection setting. Still, the second one overcame the limitation of OCCNN
related to the density estimation of the points in the feature space.
In chapter 3, we presented some important SHM algorithms and strategies
able to extract modal parameters from the input data. We focused our efforts
on the output-only model-free approaches to extend our strategy to different
kinds of structures without any assumption about materials or models. The
goal was to develop a blind system that does not need prior knowledge about
the structure, except the accelerometer data.
In chapter 4, we presented how to apply the ML algorithms described
in chapter 2 in the context of SHM. To demonstrate the proposed solu-
tions’ effectiveness, we considered a widely known benchmark, the dataset

103
104 Conclusion

of accelerometric measurements taken from the Z-24 bridge. In particular,

we applied the SSI algorithm to the Z-24 bridge measurements to extract
the structure’s modal parameters. Then, we proposed a new density-based
mode tracking able to extract the fundamental frequencies from the modal
parameter via a Gaussian kernel. Finally, we applied the ML algorithms on
the extracted frequencies to perform anomaly detection, and we proved that
OCCNN2 provides the best performance among all the algorithms tested, in
terms of accuracy and F1 score.
In chapter 5, we proposed some strategies to select the most reliable
features among the ones extracted. More specifically, we compared the fea-
ture extraction technique performed by an ANN with the feature selection
paradigm. We proved that the second option is the most appropriate in
the SHM environment, where the fundamental frequencies are already sig-
nificant features. Moreover, we demonstrated that noisy modal frequencies
could deteriorate the anomaly detector performance; hence using only the
most reliable features is recommended in this scenario.
In chapter 6, we investigated the minimum operational condition that can
be adopted to ensure some target performance, with the aim to reduce the
amount of data stored by the network during monitoring. We showed that
it is possible to reduce the number of bits used to encode the accelerometric
measurements, the number of sensors used to monitor the structure, and the
acquisition time for each measurement. In particular, the third solution is the
one that reduces the amount of data that must be stored more significantly.
Still, when the aim is to reduce the network complexity or the sensors cost,
the other two solutions can be a good compromise between complexity and
data volume.
A related work that exploits some of the signal processing tools developed
during the research is presented in chapter 7. Such work consists of classifying
the activities performed by a human being from accelerometric data collected
through a biaxial geophone located in its vicinity. Finally, in chapter 8, we
propose anomaly detection tools in the radio-frequency domain to exploit
over-the-air SoOp to assess the presence of a non-collaborative intruder in
an indoor scenario. While deviating a little bit from the primary research
 105

area, these two last chapters present original results that originated from it.
This thesis presents the endeavor to explore the use of ML algorithms in
the area of SHM. While most of the results are based on the Z-24 database,
the extension of the proposed solutions to a wide class of structures appears
feasible and supported by the analysis of their robustness. Therefore, an in-
teresting direction to pursue is applying and validating the designed method-
ologies to other structures. Moreover, since deep learning is a growing area
capturing the attention of many researchers worldwide, the application of
deep learning in this context seems a thriving research area that could lead
to attractive solutions.

Questions Recall
Q1: How can we use tools for dimensionality reduction to perform anomaly
detection?

A1: Dimensionality reduction techniques can always be used to accomplish

anomaly detection tasks. The strategy maps the data in a low dimen-
sional feature space and remaps them in the original one. This proce-
dure introduces a reconstruction error; depending on its magnitude, a
point can be defined as standard or anomalous.

Q2: What are the most effective algorithms able to perform anomaly detec-
tion?

A2: It is not possible to define the best algorithm to perform anomaly

detection. The performance of the algorithms depends on the data
structure; for instance, some algorithms work well with a large number
of points in a low dimensional feature space. On the contrary, other
strategies achieve good performances with a low number of points and
high dimensionality. In general, there are algorithms more flexible than
others, which can provide good results in a wide set of problems.

Q3: Is it possible to use classical NN structures (instead of the ANNs) to

perform anomaly detection?
106 Conclusion

A3: Yes, it is. In this thesis, we propose the OCCNN and OCCNN2 algo-
rithms that exploit the NN flexibility in the anomaly detection task.
These algorithms generate adversarial dummy points to train the anomaly
detector as a two-class NN classifier.

Q4: How can we perform anomaly detection based on machine learning tools
to detect damage in a structure?

A4: Anomaly detection can be applied in several scenarios, one of this

is SHM. To achieve good performance, it is necessary to effectively
extract damage-sensitive features from the sensors’ measurements in-
stalled on the structure.

Q5: Is there an effective technique to extract features able to highlight dam-

age in a structure?

A5: Yes, there is. In this thesis, we focused our efforts on processing ac-
celerometric measurements gathered by sensors installed on bridges.
In order to maintain generality, we decide not to use strategies that
require a model of the monitored structure; hence we only exploited in-
formation extracted by the data (model-free or data-driven strategies).
Among several approaches offered by the literature, we decide to use
SSI that satisfies all the requirements described above.

Q6: Is it possible to perform frequencies tracking automatically after the

fundamental frequencies extraction?

A6: Yes, it is. In this thesis, we propose several strategies to accomplish

this task. We proposed a novel density-based time-domain tracking
algorithm to exploit the time correlation between consecutive measure-
ments, filtering new measurements with different kernels (linear and
Gaussian). These strategies provide satisfactory and reliable tracking
results.

Q7: How can we test these strategies? Is there in literature a reference

structure with accelerometric measurements both in standard and in
damaged conditions?
 107

A7: The crucial point is to find an actual structure with an extensive

database of measurements both in standard and in damaged condi-
tion to test the algorithm performance. Fortunately, the Z-24 bridge,
a known benchmark in literature, is a real structure whose accelero-
metric and environmental measurements are available online both in
standard and damaged conditions. We decided to test our algorithms
on this structure.

Q8: Is it possible to automatically select the most reliable fundamental fre-

quencies, among all the ones extracted, to reduce the problem dimen-
sionality without affecting the performance?

A8: Yes, it is. We proposed several metrics to accomplish this task, and we
showed that variance and skewness of the extracted frequencies could
be effectively used to determine the most reliable frequencies among all
the extracted ones.

Q9: Is it possible to further reduce the amount of data that the network must
store?

A9: Yes, it is. In the thesis, we proposed three different strategies to reduce
the amount of data stored without effect the anomaly detectors perfor-
mance: i) reduce the number of sensors used to monitor the structure;
ii) reduce the number of resolution bits of the sensors; iii) reduce the
acquisition time for each measurement.

Q10: Is there a possibility to reduce the network infrastructure costs, reducing

the number of sensors necessary to detect damage, or using low-cost
sensors?

A10: Yes, it is. Experimental results showed that when the objective is to
reduce the network infrastructure costs, the best strategy is to reduce
the number of sensors using a low number of accurate accelerometers to
achieve good performance. Instead, when the goal is to reduce sensors
cost, a large network of low-cost sensors should be the best solution to
maintain high anomaly detection capability.
108 Conclusion

Q11: Is it possible to exploit accelerometric measurements acquired by sensors

installed on a structure also to infer human activities?

A11: Yes, it is. As shown in chapter 7, measurements gathered by seismic

sensors installed on the ground can be used to infer human activities
with high accuracy.

Q12: Is it possible to apply similar techniques to solve other problems?

A12: Yes, it is. As said before, ML and anomaly detection techniques are
flexible strategies that can be used in different fields. For instance, in
chapter 8, we propose an application that exploits WiFi signals of op-
portunity to detect the presence of a target in an environment through
anomaly detection algorithms. Experimental results show that this ap-
proach works pretty well and provides very good results in terms of
accuracy.
 109

Publications
1 E. Favarelli, A. Giorgetti, “Machine Learning for Automatic Processing
of Modal Analysis in Damage Detection of Bridges,” IEEE Transac-
tions on Instrumentation and Measurement (TIM), 2020.

2 E. Favarelli, E. Testi, A. Giorgetti, “Data Management in Structural

Health Monitoring,” 8th Civil Structural Health Monitoring Workshop
(CSHM-8), Online, Mar. 29 - 31, 2021.

3 E. Favarelli, E. Testi, A. Giorgetti, “Dimensionality Reduction in Modal

Analysis for Structural Health Monitoring,” International Conference
on Smart Urban Infrastructures and Artificial Intelligence Technologies
(ICSUIAIT), Rome, Italy, Feb. 18-19, 2021.

4 E. Testi, L. Pucci, E. Favarelli, A. Giorgetti, “Blind Traffic Classifi-

cation in Wireless Networks,” European Signal Processing Conference
(EUSIPCO), Amsterdam, Jan. 18-21, 2021.

5 E. Testi, E. Favarelli, A. Giorgetti, “Blind Source Separation for Wire-

less Networks: a Tool for Topology Sensing,” EAI International Confer-
ence on Cognitive Radio Oriented Wireless Networks (CROWNCOM),
Online, Nov. 25-26, 2020.

6 E. Testi, E. Favarelli, A. Giorgetti, “Reinforcement Learning for Con-

nected Autonomous Vehicle Localization via UAVs,” IEEE Interna-
tional Workshop on Metrology for Agriculture and Forestry, Online,
Nov. 4-6, 2020.

7 L. Pucci, E. Testi, E. Favarelli, A. Giorgetti, “Human Activities Classi-

fication Using Biaxial Seismic Sensors,” IEEE Sensors Letters, vol. 4,
no. 10, pp. 1-4, Oct. 2020.

8 E. Favarelli, E. Testi, A. Giorgetti, “One Class Classifier Neural Net-

work for Anomaly Detection in Low Dimensional Feature Spaces,” In-
ternational Conference on Signal Processing and Communication Sys-
110 Conclusion

tems (ICSPCS), Surfers Paradise, Gold Coast, Australia, Dec. 16-18,

2019.

9 E. Favarelli, E. Testi, L. Pucci, M. Chiani, A. Giorgetti, “Anomaly De-

tection Using WiFi Signals of Opportunity,” International Conference
on Signal Processing and Communication Systems (ICSPCS), Surfers
Paradise, Gold Coast, Australia, Dec. 16-18, 2019.

10 E. Testi, E. Favarelli, L. Pucci, A. Giorgetti, “Machine Learning for

Wireless Network Topology Inference,” International Conference on
Signal Processing and Communication Systems (ICSPCS), Surfers Par-
adise, Gold Coast, Australia, Dec. 16-18, 2019.

11 E. Testi; E. Favarelli; A. Giorgetti, “Machine Learning for User Traf-

fic Classification in Wireless Systems,” in European Signal Processing
Conference (EUSIPCO), pp. 1-5, Rome, Italy, Sep. 3-7, 2018.
Bibliography

[1] D. Brunelli, C. Moser, L. Thiele, and L. Benini, “Design of a solar-

harvesting circuit for batteryless embedded systems,” IEEE Transac-
tions on Circuits and Systems I: Regular Papers, vol. 56, no. 11, pp.
2519–2528, 2009.

[2] D. Brunelli, L. Benini, C. Moser, and L. Thiele, “An efficient solar en-
ergy harvester for wireless sensor nodes,” in 2008 Design, Automation
and Test in Europe, 2008, pp. 104–109.

[3] A. S. Weddell, M. Magno, G. V. Merrett, D. Brunelli, B. M. Al-

Hashimi, and L. Benini, “A survey of multi-source energy harvesting
systems,” in 2013 Design, Automation Test in Europe Conference Ex-
hibition (DATE), 2013, pp. 905–908.

[4] D. Carli, D. Brunelli, D. Bertozzi, and L. Benini, “A high-efficiency

wind-flow energy harvester using micro turbine,” in SPEEDAM 2010,
2010, pp. 778–783.

[5] D. Pianini, A. Elzanaty, A. Giorgetti, and M. Chiani, “Emerging dis-

tributed programming paradigm for cyber-physical systems over lo-
rawans,” in IEEE Globecom Workshops (GC Wkshps), 2018, pp. 1–6.

[6] A. Elzanaty, A. Giorgetti, and M. Chiani, “Lossy compression of noisy

sparse sources based on syndrome encoding,” IEEE Transactions on
Communications, vol. 67, no. 10, pp. 7073–7087, 2019.

[7] S. Sindaco, S. Nanni, C. Aguzzi, L. Roffia, and T. Salmon Cinotti, “En-

abling context aware tuning of low power sensors for smart agriculture,”

111
112 BIBLIOGRAPHY

in 2020 IEEE International Workshop on Metrology for Agriculture and

Forestry (MetroAgriFor), 2020, pp. 114–118.

[8] F. G. Brundu, E. Patti, A. Osello, M. D. Giudice, N. Rapetti,

A. Krylovskiy, M. Jahn, V. Verda, E. Guelpa, L. Rietto, and A. Ac-
quaviva, “IoT software infrastructure for energy management and sim-
ulation in smart cities,” IEEE Transactions on Industrial Informatics,
vol. 13, no. 2, pp. 832–840, 2017.

[9] E. Lattanzi, E. Regini, A. Acquaviva, and A. Bogliolo, “Energetic sus-

tainability of routing algorithms for energy-harvesting wireless sensor
networks,” Computer Communications, vol. 30, no. 14, pp. 2976 – 2986,
2007.

[10] M. Chiani and A. Elzanaty, “On the LoRa modulation for IoT: Wave-
form properties and spectral analysis,” IEEE Internet of Things Jour-
nal, vol. 6, no. 5, pp. 8463–8470, 2019.

[11] B. Al Homssi, A. Al-Hourani, S. Chandrasekharan, K. M. Gomez, and

S. Kandeepan, “On the bound of energy consumption in cellular IoT
networks,” IEEE Transactions on Green Communications and Net-
working, vol. 4, no. 2, pp. 355–364, 2020.

[12] B. Al Homssi, A. Al-Hourani, K. G. Chavez, S. Chandrasekharan, and

S. Kandeepan, “Energy-efficient IoT for 5G: A framework for adaptive
power and rate control,” in 12th International Conference on Signal
Processing and Communication Systems (ICSPCS), 2018, pp. 1–6.

[13] L. Sciullo, C. Aguzzi, M. Di Felice, and T. S. Cinotti, “WoT store: En-

abling things and applications discovery for the W3C web of things,” in
16th IEEE Annual Consumer Communications Networking Conference
(CCNC), 2019, pp. 1–8.

[14] G. Fabbrocino and C. Rainieri, Operational modal analysis of civil en-

gineering structures. New York: Springer-verlag, May 2014.
BIBLIOGRAPHY 113

[15] R. M. Azzara, G. D. Roeck, M. Girardi, C. Padovani, D. Pellegrini, and

E. Reynders, “The influence of environmental parameters on the dy-
namic behaviour of the San Frediano bell tower in Lucca,” Engineering
Structures, vol. 156, pp. 175–187, 2018.

[16] E. Reynders, G. Wursten, and G. D. Roeck, “Output-only structural

health monitoring in changing environmental conditions by means of
nonlinear system identification,” Structural Health Monitoring, vol. 13,
no. 1, pp. 82–93, 2014.

[17] Z-24 bridge data. [Online]. Available:

https://bwk.kuleuven.be/bwm/z24

[18] A. Al-Hourani, S. Chandrasekharan, and S. Kandeepan, “Path loss

study for millimeter wave device-to-device communications in urban
environment,” in IEEE international conference on communications
workshops (ICC), 2014, pp. 102–107.

[19] K. Sithamparanathan and A. Giorgetti, Cognitive radio techniques:

spectrum sensing, interference mitigation, and localization. Artech
house, 2012.

[20] W. Ejaz, M. Naeem, M. Basharat, A. Anpalagan, and S. Kandeepan,

“Efficient wireless power transfer in software-defined wireless sensor
networks,” IEEE Sensors Journal, vol. 16, no. 20, pp. 7409–7420, 2016.

[21] S. Chandrasekharan, K. Gomez, A. Al-Hourani, S. Kandeepan,

T. Rasheed, L. Goratti, L. Reynaud, D. Grace, I. Bucaille, T. Wirth,
and S. Allsopp, “Designing and implementing future aerial communi-
cation networks,” IEEE Communications Magazine, vol. 54, no. 5, pp.
26–34, 2016.

[22] B. Al Homssi, A. Al-Hourani, R. J. Evans, K. G. Chavez, S. Kandeepan,

W. S. T. Rowe, and M. Loney, “Free spectrum for IoT: How much
can it take?” in IEEE International Conference on Communications
Workshops (ICC Workshops), 2018, pp. 1–6.
114 BIBLIOGRAPHY

[23] V. Yadav, L. Kumar, and P. Kumar, “Evolution and development of

wireless communication system,” in International Conference on Com-
puting, Power and Communication Technologies (GUCON), 2019, pp.
53–57.

[24] G. D. Roeck, “The state-of-the-art of damage detection by vibration

monitoring: the SIMCES experience,” Journal of Structural Control,
vol. 10, no. 2, pp. 127–134, May 2003.

[25] K. Worden, C. Farrar, J. Haywood, and M. Todd, “A review of nonlin-

ear dynamics applications to structural health monitoring,” Structural
Control and Health Monitoring, vol. 15, no. 4, pp. 540–567, Jul. 2008.

[26] F. Zonzini, C. Aguzzi, L. Gigli, L. Sciullo, N. Testoni, L. De Marchi,

M. Di Felice, T. S. Cinotti, C. Mennuti, and A. Marzani, “Structural
health monitoring and prognostic of industrial plants and civil struc-
tures: A sensor to cloud architecture,” IEEE Instrumentation Mea-
surement Magazine, vol. 23, no. 9, pp. 21–27, 2020.

[27] F. Zonzini, A. Girolami, L. De Marchi, A. Marzani, and D. Brunelli,

“Cluster-based vibration analysis of structures with GSP,” IEEE
Transactions on Industrial Electronics, vol. 68, no. 4, pp. 3465–3474,
2021.

[28] F. Zonzini, M. M. Malatesta, D. Bogomolov, N. Testoni, A. Marzani,

and L. De Marchi, “Vibration-based SHM with upscalable and low-cost
sensor networks,” IEEE Transactions on Instrumentation and Mea-
surement, vol. 69, no. 10, pp. 7990–7998, 2020.

[29] C. M. Bishop, Pattern Recognition and Machine Learning. Springer

Verlag, Aug. 2006.

[30] J. Watt, R. Borhani, and A. K. Katsaggelos, Machine Learning Refined.

Cambridge University Press, 2016.

[31] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press,

2016.
BIBLIOGRAPHY 115

[32] D. Y. Harvey and M. D. Todd, “Structural health monitoring fea-

ture design by genetic programming,” Smart Materials and Structures,
vol. 23, no. 9, p. 095002, jul 2014.

[33] A. Santos, M. Silva, C. Sales, J. Costa, and E. Figueiredo, “Applica-

bility of linear and nonlinear principal component analysis for damage
detection,” in IEEE International Instrumentation and Measurement
Technology Conference (I2MTC) Proceedings, Pisa, Italy, May 2015,
pp. 869–874.

[34] E. Favarelli, E. Testi, L. Pucci, and A. Giorgetti, “Anomaly detection

using WiFi signals of opportunity,” in 13th International Conference
on Signal Processing and Communication Systems (ICSPCS), Surfers
Paradise, Gold Coast, Australia, Dec. 2019, pp. 1–7.

[35] E. Testi, E. Favarelli, and A. Giorgetti, “Machine learning for user traf-
fic classification in wireless systems,” in 26th European Signal Process-
ing Conference (EUSIPCO), Rome, Italy, Sep. 2018, pp. 2040–2044.

[36] E. Favarelli, E. Testi, and A. Giorgetti, “One class classifier neural

network for anomaly detection in low dimensional feature spaces,” in
13th International Conference on Signal Processing and Communica-
tion Systems (ICSPCS), Surfers Paradise, Gold Coast, Australia, Dec.
2019, pp. 1–7.

[37] E. Testi, E. Favarelli, L. Pucci, and A. Giorgetti, “Machine learning

for wireless network topology inference,” in 13th International Con-
ference on Signal Processing and Communication Systems (ICSPCS),
Gold Coast, Australia, Dec. 2019.

[38] P. Perera and V. M. Patel, “Learning deep features for one-class classi-
fication,” IEEE Transactions on Image Processing, vol. 28, no. 11, pp.
5450–5463, Nov. 2019.

[39] R. Chalapathy, A. K. Menon, and S. Chawla, “Anomaly detection using

one-class neural networks,” Computing Research Repository (CoRR),
Aug. 2018.
116 BIBLIOGRAPHY

[40] P. Schlachter and B. Yang, “Active learning for one-class classifica-

tion using two one-class classifiers,” in 26th European Signal Processing
Conference (EUSIPCO), Rome, Italy, Sep. 2018, pp. 1197–1201.

[41] H. Abdi and L. J. Williams, “Principal component analysis,” Wiley

Interdisciplinary Reviews: Computational Statistics, vol. 2, no. 4, pp.
433–459, 2010.

[42] B. Schölkopf, A. Smola, and K.-R. Müller, “Kernel principal component

analysis,” in Proceedings International Conference on Artificial Neural
Networks, vol. 1327, no. 6. Lausanne, Switzerland: Springer, Oct.
1997, pp. 583–588.

[43] B. Schölkopf, A. Smola, E. Smola, and K.-R. Müller, “Nonlinear com-

ponent analysis as a kernel eigenvalue problem,” Neural Computation,
vol. 10, pp. 1299–1319, Jul. 1998.

[44] A. Santos, E. Figueiredo, M. Silva, R. Santos, C. Sales, and J. C.

W. A. Costa, “Genetic-based EM algorithm to improve the robustness
of gaussian mixture models for damage detection in bridges,” Structural
Control and Health Monitoring, vol. 24, no. 3, p. e1886, May 2017.

[45] J. H. Pollard, “On distance estimators of density in randomly dis-

tributed forests,” Biometrics, vol. 27, no. 4, pp. 991–1002, Dec. 1971.

[46] D. Brigante, C. Rainieri, and G. Fabbrocino, “The role of the modal

assurance criterion in the interpretation and validation of models for
seismic analysis of architectural complexes,” in Proceedings Interna-
tional Conference on Structural Dynamics (Eurodin), vol. 199, Rome,
Italy, Sep. 2017, pp. 3404–3409.

[47] L. Calandrino and M. Chiani, Lezioni di Comunicazioni Elettriche.

Pitagora, 2013.

[48] C. Wu, H. Liu, X. Qin, and J. Wang, “Stabilization diagrams to dis-

tinguish physical modes and spurious modes for structural parameter
BIBLIOGRAPHY 117

identification,” Journal of Vibroeng., vol. 19, no. 4, pp. 2777–2794, Jun.

2017.

[49] A. Cabboi, F. Magalhães, C. Gentile, and Á. Cunha, “Automated

modal identification and tracking: Application to an iron arch bridge,”
Structural Control and Health Monitoring, vol. 24, no. 1, p. e1854, Feb.
2017.

[50] M. Pastor, M. Binda, and T. Harčarik, “Modal assurance criterion,”

Procedia Engineering, vol. 48, pp. 543–548, 2012.

[51] E. Reynders, J. Houbrechts, and G. D. Roeck”, “Fully automated (op-

erational) modal analysis,” Mechanical Systems and Signal Processing,
vol. 29, pp. 228–250, May 2012.

[52] E. P. Carden and J. M. W. Brownjohn, “Fuzzy clustering of stability

diagrams for vibration-based structural health monitoring,” Computer-
Aided Civil and Infrastructure Eng., vol. 23, no. 5, pp. 360–372, May
2008.

[53] E. Reynders and G. D. Roeck, “Continuous vibration monitoring and

progressive damage testing on the z 24 bridge,” Encyclopedia of struc-
tural health monitoring, vol. 10, pp. 127–134, May 2009.

[54] (2001) FBA11 datasheet. [Online]. Available:

ftp://aplica.at/Manuals/Sensors/FBA11.pdf

[55] (1998) ADC488 datasheet. [Online]. Available:

https://www.artisantg.com/info/IOtech ADC488 Manual.pdf

[56] G. H. Golub and C. F. V. Loan, Matrix computations. JHU press,

2012, vol. 3.

[57] J. J. Verbeek, N. Vlassis, and B. Kröse, “Efficient greedy learning of

gaussian mixture models,” Neural computation, vol. 15, no. 2, pp. 469–
485, 2003.
118 BIBLIOGRAPHY

[58] M. Silva, A. Santos, E. Figueiredo, R. Santos, C. Sales, and J. Costa,

“A novel unsupervised approach based on a genetic algorithm for struc-
tural damage detection in bridges,” Engineering Applications of Arti-
ficial Intelligence, vol. 52, pp. 168–180, Jun. 2016.

[59] M. Silva, A. Santos, R. Santos, E. Figueiredo, C. Sales, and J. C. Costa,

“Agglomerative concentric hypersphere clustering applied to structural
damage detection,” Mechanical Systems and Signal Processing, vol. 92,
pp. 196–212, Feb. 2017.

[60] G. Park, T. Rosing, M. D. Todd, C. R. Farrar, and W. Hodgkiss,

“Energy harvesting for structural health monitoring sensor networks,”
Journal of Infrastructure Systems, vol. 14, no. 1, pp. 64–79, 2008.

[61] S. Taylor, K. Farinholt, E. Flynn, E. Figueiredo, D. L. Mascarenas,

E. A. Moro, G. Park, M. Todd, and C. Farrar, “A mobile-agent-based
wireless sensing network for structural monitoring applications,” Mea-
surement Science and Technology, vol. 20, no. 4, pp. 1–14, Jan. 2009.

[62] D. Mascarenas, E. Flynn, C. Farrar, G. Park, and M. Todd, “A mo-

bile host approach for wireless powering and interrogation of structural
health monitoring sensor networks,” IEEE Sensors Journal, vol. 9,
no. 12, pp. 1719–1726, 2009.

[63] A. Girolami, F. Zonzini, L. De Marchi, D. Brunelli, and L. Benini,

“Modal analysis of structures with low-cost embedded systems,” in
IEEE International Symposium on Circuits and Systems (ISCAS),
2018, pp. 1–4.

[64] A. Elzanaty, A. Giorgetti, and M. Chiani, “Weak RIC analysis of finite

gaussian matrices for joint sparse recovery,” IEEE Signal Processing
Letters, vol. 24, no. 10, pp. 1473–1477, 2017.

[65] E. B. Flynn and M. D. Todd, “A bayesian approach to optimal sensor

placement for structural health monitoring with application to active
sensing,” Mechanical Systems and Signal Processing, vol. 24, no. 4, pp.
891 – 903, 2010.
BIBLIOGRAPHY 119

[66] E. B. Flynn and M. Todd, “Optimal placement of piezoelectric actua-

tors and sensors for detecting damage in plate structures,” Journal of
Intelligent Material Systems and Structures, vol. 21, no. 3, pp. 265–274,
2010.

[67] Q. Zhou, G. Tong, D. Xie, B. Li, and X. Yuan, “A seismic-based fea-

ture extraction algorithm for robust ground target classification,” IEEE
Signal Processing Letters, vol. 19, no. 10, pp. 639–642, Oct. 2012.

[68] J. Huang, Q. Zhou, X. Zhang, E. Song, B. Li, and X. Yuan, “Seis-

mic target classification using a wavelet packet manifold in unattended
ground sensors systems,” Sensors, vol. 13, no. 7, pp. 8534–8550, Jul.
2013.

[69] S. Pan, K. Lyons, M. Mirshekari, H. Y. Noh, and P. Zhang, “Multiple

pedestrian tracking through ambient structural vibration sensing,” in
Proceedings of the 14th ACM Conference on Embedded Network Sensor
Systems, New York, NY, USA, Nov. 2016, pp. 366–367.

[70] J. Jackowski and R. Wantoch-Rekowski, “Classification of wheeled mil-

itary vehicles using neural networks,” in 18th International Conference
on Systems Engineering (ICSEng), Las Vegas, NV, USA, Aug. 2005,
pp. 212–217.

[71] B. Mukhopadhyay, S. Anchal, and S. Kar, “Detection of an intruder

and prediction of his state of motion by using seismic sensor,” IEEE
Sensors Journal, vol. 18, no. 2, pp. 703–712, Jan. 2018.

[72] X. Jin, S. Sarkar, A. Ray, S. Gupta, and T. Damarla, “Target detec-

tion and classification using seismic and PIR sensors,” IEEE Sensors
Journal, vol. 12, no. 6, pp. 1709–1718, Jun. 2012.

[73] M. Zubair and K. Hartmann, “Target classification based on sensor

fusion in multi-channel seismic network,” in IEEE International Sym-
posium on Signal Processing and Information Technology (ISSPIT),
Bilbao, Spain, Dec. 2011, pp. 438–443.
120 BIBLIOGRAPHY

[74] A. Sutin, H. Salloum, A. Sedunov, and N. Sedunov, “Acoustic detec-

tion, tracking and classification of low flying aircraft,” in IEEE In-
ternational Conference on Technologies for Homeland Security (HST),
Waltham, MA, USA, Nov. 2013, pp. 141–146.

[75] M. Civelek and A. Yazici, “Automated moving object classification in

wireless multimedia sensor networks,” IEEE Sensors Journal, vol. 17,
no. 4, pp. 1116–1131, Feb. 2017.

[76] W. D. Fisher, T. K. Camp, and V. V. Krzhizhanovskaya, “Anomaly

detection in earth dam and levee passive seismic data using support
vector machines and automatic feature selection,” Journal of Compu-
tational Science, vol. 20, pp. 143–153, May 2017.

[77] P. Welch, “The use of fast Fourier transform for the estimation of power
spectra: A method based on time averaging over short, modified peri-
odograms,” IEEE Transactions on audio and electroacoustics, vol. 15,
no. 2, pp. 70–73, Jun. 1967.

[78] I. Jolliffe, Principal Component Analysis. Springer Verlag, 1986.

[79] H. Abdi and L. J. Williams, “Principal component analysis,” WIREs

Comput. Stat., vol. 2, no. 4, pp. 433–459, Jul. 2010.

[80] T. Cover and P. Hart, “Nearest neighbor pattern classification,” IEEE

transactions on information theory, vol. 13, no. 1, pp. 21–27, Jan. 1967.

[81] R. C. Gonzalez, R. E. Woods, and S. L. Eddins, Digital image process-

ing using MATLAB. Pearson Education India, 2004.

[82] K. Briechle and U. D. Hanebeck, “Template matching using fast nor-

malized cross correlation,” in Optical Pattern Recognition XII, vol.
4387, 2001, pp. 95–102.

[83] Y. Guo, T. Hastie, and R. Tibshirani, “Regularized linear discriminant

analysis and its application in microarrays,” Biostatistics, vol. 8, no. 1,
pp. 86–100, 2007.
BIBLIOGRAPHY 121

[84] S. Bartoletti, A. Conti, and M. Z. Win, “Device-free counting via wide-

band signals,” IEEE Journal on Selected Areas in Communications,
vol. 35, no. 5, pp. 1163–1174, May 2017.

[85] S. Bartoletti, A. Conti, and M. Z. Win, “Passive radar via LTE signals
of opportunity,” IEEE International Conference on Communication
Workshops (ICC), pp. 181–185, Jun. 2014.

[86] S. Bartoletti, A. Conti, and M. Z. Win, “Device-free counting via

OFDM signals of opportunity,” IEEE International Conference on
Communication Workshops (ICC), pp. 1–5, May 2018.

[87] M. Leng, W. P. Tay, C. M. S. See, S. Gulam Razul, and M. Z. Win,

“Modified CRLB for cooperative geolocation of two devices using sig-
nals of opportunity,” IEEE Transactions on Wireless Communications,
vol. 13, no. 7, pp. 3636–3649, Jul. 2014.

[88] A. E. Kosba, A. Abdelkader, and M. Youssef, “Analysis of a device-

free passive tracking system in typical wireless environments,” in 3rd
International Conference on New Technologies, Mobility and Security,
Cairo, Egypt, Dec. 2009, pp. 1–5.

[89] Y. Jin, Z. Tian, M. Zhou, Z. Li, and Z. Zhang, “A whole-home

level intrusion detection system using WiFi-enabled IoT,” in 14th In-
ternational Wireless Communications Mobile Computing Conference
(IWCMC), Limassol, Cyprus, Jun. 2018, pp. 494–499.

[90] I. Sabek and M. Youssef, “Multi-entity device-free WLAN localiza-

tion,” in IEEE Global Communications Conference (GLOBECOM),
Anaheim, California, Dec. 2012, pp. 2018–2023.

[91] J. Yang, Y. Ge, H. Xiong, Y. Chen, and H. Liu, “Performing joint

learning for passive intrusion detection in pervasive wireless environ-
ments,” in Proceedings IEEE International Conference on Computer
Communications (INFOCOM), San Diego, California, Mar. 2010, pp.
1–9.
122 BIBLIOGRAPHY

[92] Y. Jin, Z. Tian, M. Zhou, Z. Li, and Z. Zhang, “An adaptive and
robust device-free intrusion detection using ubiquitous WiFi signals,”
in IEEE International Conference on Digital Signal Processing (DSP),
Shanghai, China, Nov. 2018, pp. 1–5.

[93] Z. Zhou, Z. Yang, C. Wu, L. Shangguan, and Y. Liu, “Omnidirectional

coverage for device-free passive human detection,” IEEE Transactions
on Parallel and Distributed Systems, vol. 25, no. 7, pp. 1819–1829, Jul.
2014.

[94] Z. Tian, Y. Li, M. Zhou, and Z. Li, “WiFi-based adaptive indoor pas-
sive intrusion detection,” in IEEE International Conference on Digital
Signal Processing (DSP), Shanghai, China, Nov. 2018, pp. 1–5.

[95] S. Savazzi, S. Sigg, M. Nicoli, V. Rampa, S. Kianoush, and U. Spag-

nolini, “Device-free radio vision for assisted living: Leveraging wireless
channel quality information for human sensing,” IEEE Signal Process-
ing Magazine, vol. 33, no. 2, pp. 45–58, Mar. 2016.

[96] E. Cianca, M. De Sanctis, and S. Di Domenico, “Radios as sensors,”

IEEE Internet of Things Journal, vol. 4, no. 2, pp. 363–373, Apr. 2017.

[97] M. Chiani, A. Giorgetti, and E. Paolini, “Sensor radar for object track-
ing,” Proceedings of the IEEE, vol. 106, no. 6, pp. 1022–1041, Jun.
2018.

[98] F. Colone, D. Pastina, P. Falcone, and P. Lombardo, “WiFi-based pas-

sive ISAR for high-resolution cross-range profiling of moving targets,”
IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 6,
pp. 3486–3501, Jun. 2014.

[99] A. E. Kosba, A. Saeed, and M. Youssef, “Rasid: A robust wlan device-

free passive motion detection system,” in IEEE International Con-
ference on Pervasive Computing and Communications, Lugano,Swiss,
Mar. 2012, pp. 180–189.
BIBLIOGRAPHY 123

[100] B. W. Silverman, Density Estimation for Statistics and Data Analysis.

Chapman and Hall/CRC, Apr. 1986.
124 BIBLIOGRAPHY
Acknowledgements

There are several people I should thanks for where I am today, several people
who gave me support when I wanted to give up, several people who showed
me the light when the only thing I could see was the darkness.

First of all Enrico Testi, a friend, a colleague, a teammate. He was al-

ways by my side cheering in the best moments and standing next to me in
the worst ones. Enrico I own you all my success and I will be with you to
share all of our failures. I could not hope for a better friend to walk with me
down this path, I can just hope one day I will become for you what you are
for me.

My family, my mother Barbara, Massimo, and my little sister Nicole (the

greatest person I know). My mother, my first fan always there, since she
gave me the first Lego piece till my Ph.D. in engineering, she never failed
to show me her support. Massimo, that reminds me everyday that does not
matter what destiny takes from your life, you can always find a way to create
your own path. My sister, Nicole, she taught me how to dream and how to
fill my mind with a million dreams to create the world I vision and that I
want to create. They are my safe place where to find refuge from a storm, I
own you my life, my strength and my hope. Thank you.

My father, Daniele, my green light at the end of the pier, a man who
had that rare smile that probably you meet once in your life, an indelible
example of the person that I want to be.

125
126 Acknowledgements

All my friends and relatives that support me every day, sometimes with
simple gesture, sometimes with great supports, I hope to spend as much time
as possible with them and to give them back at least half of the good they
gave me.

I want to thanks my co-advisor Prof. Alessandro Marzani, always present

when I needed suggestions or help, he represents a great support for me.

Finally I would like to thanks my advisor and mentor Prof. Andrea Gior-
getti, a great person and an illuminating professor, he always supported me
during these years and I hope to continue to work with him as long as pos-
sible. I have the merit of having followed him, he has the merit of having
believed in me.

Thank you all, for your trust, your support and your love. Is not al-
ways easy to believe in yourself during life, there are moments that the only
thought you have in your mind is “I am not gonna make it I am not good
enough” well today I finally proved to myself, that I am good enough. Thank
you.
“A true master is an eternal student”