Special Issue Article

Structural Health Monitoring

2019, Vol. 18(1) 123–142

Wavelet packet energy–based damage Ó The Author(s) 2018

Article reuse guidelines:
sagepub.com/journals-permissions
identification of wood utility poles DOI: 10.1177/1475921718798622
journals.sagepub.com/home/shm
using support vector machine
multi-classifier and evidence theory

Yang Yu1 , Ulrike Dackermann2, Jianchun Li1 and Ernst Niederleithinger3

Abstract
This article presents a novel assessment framework to identify the health condition of wood utility poles. The innovative
approach is based on the integration of data mining and machine learning methods and combines advanced signal pro-
cessing, multi-sensor data fusion and decision ensembles to classify different damage condition types of wood poles. In
the proposed framework, wavelet packet analysis is employed to transform captured multi-channel stress wave signals
into energy information, which is consequently compressed by principal component analysis to extract a feature vector.
Furthermore, support vector machine multi-classifier, optimized by genetic algorithm, is designed to identify the pole
condition type. Finally, evidence theory is applied to fuse different assessment results from different sensors for a final
decision. For validation of the proposed approach, the wood pole specimens with three common damage condition
types are tested using a novel multi-sensor narrow-band frequency-excitation non-destructive testing system in the
laboratory. The final experimental analysis results confirm that the proposed approach is capable of making full use of
multi-sensor information and providing an effective and accurate identification on types of conditions in wood poles.

Keywords
Wood pole, damage identification, wavelet packet energy, support vector machine, evidence theory

Introduction damage and evaluating the health states of untouchable

Round wood poles are extensively used all over the
world for electrical power and communication trans-
mission. In Australia, it was reported that around
5 million wood utility poles were installed to support
energy and communication networks.1 To routinely
inspect the structural integrity of these wood poles to
prevent failure, the government annually invests almost
50 million dollars in pole maintenance and asset man-
agement. In a conservative approach to avoid the age-
ing poles from failures, approximately 300,000 utility
poles are substituted annually in eastern Australia.1
Nevertheless, more than 80% of replaced poles are still
in a perfectly serviceable condition, resulting in enor-
mous waste of natural resources.
Restrictions of existing methods to assess the health
conditions of pile and pole structures contribute to
ongoing research and development in non-destructive
testing (NDT) methods based on stress waves (SW).
Such techniques have the capacity of detecting internal
sections of structures through wave propagation.2
Generally, SWs are produced by imparting an
1
Centre for Built Infrastructure Research, School of Civil and
Environmental Engineering, University of Technology Sydney, Sydney,
NSW, Australia
2
Centre for Infrastructure Engineering and Safety, School of Civil and
Environmental Engineering, University of New South Wales, Sydney,
NSW, Australia
3
Division 8.2, German Federal Institute for Material Research and Testing
(BAM), Berlin, Germany
Corresponding authors:
Yang Yu, Centre for Built Infrastructure Research, School of Civil and
Environmental Engineering, University of Technology Sydney, 15
Broadway, Ultimo, Sydney, NSW 2007, Australia.
Email: yang.yu@uts.edu.au
Jianchun Li, Centre for Built Infrastructure Research, School of Civil and
Environmental Engineering, University of Technology Sydney, 15
Broadway, Ultimo, Sydney, NSW 2007, Australia.
Email: jianchun.li@uts.edu.au
124
excitation force on the surface of a structure. The gen-
erated SWs propagate through the material and reflect
from the external surfaces, internal defects and bound-
aries of neighbouring materials. Wave reflections can
be identified either from the changes in time-series
acceleration responses or from peak values in the accel-
eration spectrums.3 Due to the slower wave propaga-
tion in damaged wood compared to intact wood, SWs
can be employed to evaluate the health condition of the
pile or pole through calculating the wave travelling
time at increasing positions along the structure.4–6 To
date, different SW-based assessment approaches have
been proposed such as the sonic echo method,7,8
impulse response method9 and bending wave (BW)
method.10–12 For the testing of first two methods, an
impact is imparted from the top of a pile/pole structure
(generating longitudinal waves) and reflection measure-
ments (echoes) from the bottom of the structure are
evaluated to assess the condition and determine the
length of the structure. In the BW method, a transver-
sal impact is imparted to the structure to generate flex-
ural waves, which are highly dispersive in nature.13,14
Therefore, to evaluate a structure by means of BW pro-
pagation, dispersive analysis is required to obtain the
wave information from chosen frequency bands.
In the practical integrity testing of wood poles,
because of the pole height as well as existing electrical
wires, it is neither practicable nor feasible to impart the
SW excitation from the top of the structure. Therefore,
the practical impact excitation should be induced from
the side of the structure at an accessible height with an
angle to generate both longitudinal and BW signals.
One challenge for this type of excitation is the compli-
cated wave propagation patterns: simultaneous genera-
tions of downward and upward travelling waves, which
are difficult wave behaviours to interpret and analyse.
In addition, both types of waves are excited over a
broad low-frequency bandwidth which can cause multi-
mode wave propagation in the structure. To overcome
this issue, advanced signal processing techniques can
play an important part in analysing SW signals to real-
ize the condition assessment and health monitoring of
in situ wood poles. However, current studies in this
area are rarely reported. Subhani et al.15 adopted the
continuous wavelet transform (CWT) and the short
kernel method (SKM) to calculate the phase velocity of
SWs and determine the embedded length of the pole
structures. Yan et al.16 proposed a wave energy index–
based method to identify different damage conditions
of wooden poles. Jozi et al.17 introduced a novel signal
processing method based on the combination of low-
pass finite impulse response (FIR), digital filtering, pre-
dictive deconvolution and frequency wavenumber anal-
ysis to separate bi-directional waves in the wood poles
with isotropic material properties. Nevertheless, the
Structural Health Monitoring 18(1)
above investigations are all based on numerical studies
and have not been verified experimentally due to vari-
ous constraints such as limited sensor numbers, uncer-
tainties caused by material properties and accessibility
of wood poles in the field. Recently, a study underta-
ken by the authors presented an experimental verifica-
tion of an SW-based testing method using a novel
narrow-band frequency-excitation system combined
with fast Fourier transform (FTT) signals and support
vector machine (SVM) for the damage assessment in
concrete and wood poles.18 While this study delivered
very encouraging results, the presence of false damage
classification outcomes and a lack in probability infor-
mation demands further improvement.
To avoid the false identification and improve classi-
fication accuracy, advanced signal processing and
machine learning techniques have been widely used for
damage detection and health monitoring of civil struc-
tures during past few years.19 Li and Hao20,21 employed
CWT to process dynamic relative displacement signals
and proposed a damage index based on wavelet packet
energy to identify the joint condition of the steel truss
bridge. Similarly, the wavelet transform technique was
applied in the damage evaluation of subsea pipelines
with satisfactory results.22 Bandara et al.23 developed a
hybrid method based on frequency response function
and principal component analysis (PCA) for structural
damage detection and health monitoring. In this study,
a numerical case on a two-storey framed structure was
employed to validate the effectiveness of the proposed
hybrid method, and the results showed it is capable of
detecting both single and multiple damages with low
error even though the data are contaminated with the
noise. Wang and Hao24 introduced the compressive
sensing (CS) technique into the identification of degra-
dation and free-spanning damages of the steel pipes.
All the techniques mentioned in above studies can be
regarded as potential tools to process the wave signals
for the purposed of damage classification and condi-
tion assessment of wood utility poles in this study.
Furthermore, artificial neural networks (ANN) and
adaptive neuro-fuzzy inference system (ANFIS) have
been adopted as the pattern recognition methods to
process the damage-sensitive feature for the condition
assessment of civil infrastructure. Tan et al.25 devel-
oped a three-layer network for damage detection of
steel beams, in which the modal strain energy was used
as inputs of the network. Hakim and Abdul Razak26
compared the capacities of ANN and ANFIS in the
aspect of structural damage identification. In this work,
the natural frequencies were used as the damage index
and both models were validated via the laboratory test-
ing on a steel girder bridge model. Even though the
ANN and ANFIS have been demonstrated to be useful
in structural damage detection, they still have the
Yu et al.
problems of network architecture design, connection
weight optimization, transfer function selection, fuzzy
rule option and training sample diversity, which will
directly affect the identification accuracy of the devel-
oped model.
In this work, an advanced assessment method is
investigated to identify health conditions of wood
poles. First, the previously proposed NDT system with
controlled frequency excitation is used to measure wave
signals and embedded damage signatures of the wood
pole specimens in the laboratory. This novel NDT sys-
tem effectively reduces the influence of multi-mode SW
signals. Then, wavelet packet transform (WPT) is
employed to extract energy information from the cap-
tured wave signals and the dominant energy compo-
nents are chosen to construct a feature vector used as
the input for the SVM classifier via PCA. A multi-label
classifier consisting of several genetic algorithm (GA)
optimized sub-SVM classifiers is proposed to enhance
the accuracy and applicability of the pole condition
assessment. Finally, D-S evidence theory is adopted to
combine results from different sub-SVM classifiers and
from different sensors, respectively. The experimental
results show that the proposed method including the
two-level decision fusion is able to offer more robust
and accurate identification results than conventional
approaches.
Experimental testing and condition
assessment framework of wood pole
Innovative NDT system
A novel SW excitation method is adopted in this arti-
cle, in which tactile transducers are adopted to control
the induction of sound waves in a pole structure. A tac-
tile transducer is an electro-mechanical device, very
similar to a traditional audio speaker, which is
mounted to a structure and driven by an amplifier.
While traditional audio speakers transfer sound waves
through the air, tactile transducers transfer vibrations/
sound waves through the structure, making it possible
to feel the sound. Therefore, tactile transducers are
able to induce controlled SWs into a structure such as
the utility pole. This audio technology is cost-effective
and can overcome the main issues associated with a
traditional hammer impact, that is, a controlled
narrow-band frequency range can be excited, the wave
excitation can be standardized and a symmetric SWs
can be produced by using an excitation ring with a
number of tactile transducers.18
In the testing system, four tactile transducers on the
sensor wedges are attached in a ring formation to the
pole structures providing the synchronized and sym-
metric narrow-band frequency excitations. The tactile
125
transducer used is a Vidsonix model VX-GH92 having
the frequency range between 0.65 and 16 kHz. A Hi-Fi
amplifier is employed to amplify and adjust the ampli-
tudes of the generated SWs, and a function generator is
used to produce the desired excitation frequency and
waveform. To measure the SW response of the struc-
ture, the accelerometers of PCB model 352C34 having
a frequency range between 0.5 Hz and 10 kHz are used.
The accelerometers are supplied with fixed currents
using a signal conditioner and are attached on the sur-
face of the pole structures in a line formation. A multi-
channel data acquisition system together with a com-
puter is employed to obtain and record the data. The
specific system components and setup are shown in
Figure 1.
Wood pole specimen testing
To obtain an advanced pole condition assessment
approach based on signal processing and machine
learning, it is important to have a reliable database of
the trained data to increase the accuracy of the devel-
oped model. However, the properties of wood vary
from species to species, which are also affected by tem-
perature, moisture, annular rings, boundary condition,
and so on. Accordingly, the confidence level to deter-
mine the condition of an unknown wood pole can be
quite limited while dealing with random data acquired
from less number of specimens. As a result, to develop
and evaluate the model via machine learning, it is nec-
essary to start the experimental test based on the speci-
mens of which materials properties are relatively
determined. In this study, the specimens are chosen
such a way that the uncertainties related to the material
and environmental properties of the specimen are mini-
mal. After investigating the accuracy of the developed
model on the simple specimens without uncertain fac-
tors, the proposed method can be easily extended to
other wood pole specimens by establishing a compre-
hensive database of timber material containing differ-
ent material properties and environmental conditions.
In this study, three wood pole specimens made of
pine, including healthy and damaged conditions, are
tested in the laboratory. Generally, damage to wood is
caused by climatic condition, fungi, termite attack and
mechanical damage from vehicles. In most cases, the
damage is invisible and happens below the ground.
Among different types of wood pole damage, decay
caused by fungi extends very fast because the fungi is
able to grow rapidly over the wood surface under ideal
condition. The fungi can penetrate or destroy the cell
walls of the wood, which would result in remarkable
strength loss.1 Besides, the termite voids is another
common damage type in wood poles. This type of dam-
age always occurs internally, which is difficult to detect
126
Figure 1. Testing system makeup: (a) sensor wedge, (b) tactile transducer, (c) transducer installed on sensor wedge, (d) function
generator, (e) Hi-Fi amplifier, (f) accelerometer, (g) data acquisition system, (h) signal conditioner, (i) computer and (j) schematic of
system setup.
Table 1. Details of wood pole specimens.
Pole type Damage type Damage dimension
Type 1 Undamaged
Type 2 Internal void damage Fvoid = 0.15 m, h = 0.7 m
(internal hollow section)
Type 3 External circumferential Fremain = 0.15 m, h = 0.7 m
cross-section loss damage
(circumferential cut)
until exposed. In this study, apart from the specimen
with intact condition, two specimens with artificial
damage are used to simulate these two common pole
damage types: inner termite voids and external circum-
ferential fungi decay, which are shown in Figure 2(c).
All the poles have 0.25 m diameters and 3 m lengths,
which are shown in Figure 2(b). The specific descrip-
tions of the pole specimens are given in Table 1.
The experimental setup including tactile transducers
and accelerometers is displayed in Figure 2(a). During
the testing, the wood pole specimen is freely supported
on a Styrofoam mat. Four transducers are installed to
sensor wedges through a screw connection and are
Structural Health Monitoring 18(1)
tightly attached to the specimen in a ring configuration
with uniform interval employing a ratchet strap. The
wave excitation is generated at the height of 1.6 m. The
twelve accelerometers (A1–A12) are attached to the
specimen in a line alignment using moulding clay to
measure the responses of the structure. The acceler-
ometer at the lowest location (A12) is located 0.4 m
above the ground and the spacing between two neigh-
bouring sensors is 0.1 m. During the testing, all the spe-
cimens are excited with the continuous Mexican hat
wave excitations with 1 kHz loading frequency. The
sampling rate and period are set as 1 MHz and 0.5 s,
respectively. All three pole specimens are tested five
times, generating five different sets of data.
As an instance, Figure 3 displays the sectional sam-
ples of SW signals of the 12 accelerometers tested from
the intact pole specimen. It is clearly seen that the sig-
nals have very similar wave patterns including fre-
quency and magnitude resulting from the continuous
Mexican hat wave excitation, in spite of the phase shift-
ing caused by the different sensor locations with
delayed upward and downward wave propagation.
However, it is unrealistic to directly distinguish differ-
ent damage types using raw SW signals due to various
Yu et al.
Figure 2. Experimental testing: (a) equipment setup, (b) damage configurations and dimensions of tested poles in longitudinal-
section and cross-section view and (c) real pictures of two common pole damage types.
factors including the high variability of the timber
material, measurement noise disturbances and the large
amount of sampling data. Accordingly, how to deal
with these noisy data to extract the useful information
for the purpose of pole condition identification
becomes an important and urgent issue, and the corre-
sponding assessment framework will be presented in
the next section.
Condition assessment framework
In this article, a novel scheme is proposed to deal with
captured complicated SW for pole condition assess-
ment. Figure 4 describes the overall framework of the
proposed method, which can be divided into two
phases: feature extraction and condition identification.
At first, the WPT is adopted to decompose the original
SW signals into different frequency components and
the energy information at each frequency component is
obtained and used to make up the WP energy vector.
Then, a novel feature index based on WP energy is pro-
posed to identify the damage pattern of wood poles. To
significantly reduce the feature dimension and avoid
redundant information, the PCA is employed to select
optimal components to reflect the previous features.
The selected principal components will be used as the
inputs to build up soft posterior probability SVM clas-
sifiers, which can generate the initial condition identifi-
cation result. To improve the identification accuracy of
127
the developed classifier, the GA is applied to optimize
the model parameters during the training procedure.
Finally, the D-S evidence theory is utilized to fuse the
initial results generated from different sensors and give
a final assessment decision for wood pole condition.
The main benefit of the proposed approach is that it is
capable of effectively dealing with wrong or conflicting
results from different sensors and improving the accu-
racy and robustness of identification result. The follow-
ing sections will give the specific procedures of feature
extraction and condition assessment for wood pole con-
dition identification with the validation and discussion.
Feature extraction based on wavelet
packet energy and PCA
WPT-based energy feature extraction
WPT is one of the most effective techniques to analyse
the SW signals acquired from the accelerometers for
damage detection.27,28 The waveform, which contains a
mean value of zero in the finite period, is called the
wavelet. Through WPT, a signal is able to be decom-
posed into two sections, approximation and detail, cor-
responding to the low-frequency and high-frequency
content, respectively.29,30 Then, both approximation
and detail can be further separated into the second-
level approximations and details, similar to the first-
level decomposition in WPT. The main benefit of WPT
128
Figure 3. One example of normalized time-series data of undamaged pole.
is that it can inspect the signals in narrow frequency
bands during a relatively short time window.
Based on the n-level WPT, the raw SW signal Y can
be decomposed into 2n signal frequency components,
denoted by fY1 , Y2 , . . . , Y2n g. Each component Yi
(i = 1, 2, ., 2n) can be expressed as follows
Yi = ½yi, 1 , yi, 2 , . . . , yi, m 
where m denotes the total number of sampling signal
points. Hence, the energy Ei of ith decomposed compo-
nent Yi is expressed as follows
Ei = kYi k22 = y2i, 1 + y2i, 2 +    + y2i, m
According to results from equation (2), the WP
energy vector of the SW signal can be defined as
follows
E = ½E1 , E2 , . . . , Ei 
To better demonstrate the difference between WP
energies of signals captured from intact specimen and
Structural Health Monitoring 18(1)
damaged specimen, this article proposes a novel feature
index called WP energy ratio, defined as follows
R = ½R1 , R2 , . . . , Ri  ð4Þ
where Ri denotes the energy ratio at the ith frequency
band, the value of which can be calculated by
ð1Þ
Ei
Ri = ð5Þ
Eiintact
where Eiintact denotes the corresponding energy ratio at
ith frequency band of the signal from the intact pole.
ð2Þ In general, the damage will shift the fundamental fre-
quencies of the structure and also change the ampli-
tudes of measured signals, compared with those from
undamaged structures, leading to different energy dis-
tributions of SW signals. Therefore, the value of the
ð3Þ proposed index should be close to one when there is no
damage in the structure, and it should vary with differ-
ent damage types and severities. Based on this theory,
the proposed vector of WP energy ratio is considered
Yu et al.
Figure 4. Proposed framework for condition assessment of wood pole.
as a promising index to classify the signal types from
pole structures with different damages. Figure 5 shows
the results of WP energy ratios at the first 10 frequency
bands from inner damage (Type 2) and outside damage
(Type 3) cases after 10-level WP decomposition of sig-
nals from accelerometer A1. It is noticeable that the
signal from the pole specimen with inner damage has
more energy than that from intact pole at second, third,
sixth, eighth and tenth frequency bands, while for the
specimen with outside damage, the main WP energies
distribute in the second, fourth, eighth and tenth bands.
The results in Figure 5 well illustrate the effectiveness
of the feature vector of the WP energy ratio on the clas-
sification of wood poles with different conditions.
Therefore, WPT can be considered as a good tool to
extract the damage features from SW signals for pole
condition identification.
Feature compression based on PCA
Features extracted using WP energy ratio do not only
include useful information on wave propagation or
129
damage type but also include redundant information
due to correlated features together with background
and measurement noises. If this feature vector is
directly employed to set up a predictive model (classi-
fier) for damage identification of wood poles, the rec-
ognition accuracy as well as generalization ability will
be compromised. For this issue, it is proposed to
replace these correlated and noisy features with fewer
comprehensive indices. In detail, PCA is proposed to
further process WP energy ratio features, in which the
original feature vector is substituted for several unre-
lated principal components. PCA is able to effectively
compress the original data while retaining most of its
information.31,32 Generally, the measurement noises
are not related to the global features of extracted fea-
ture data and can be expressed in less important
components.
To reduce the feature dimension of WP energy
ratios, a matrix including all extracted WP energy ratio
features from SW signals is formed and transformed
into their principal components using ‘princomp’ func-
tion in MATLAB v2012b. In this example, the row of
130 Structural Health Monitoring 18(1)

Figure 5. WP energy ratios at first 10 frequency bands after 10-level WPT of signals from sensor A1: (a) inner damage (Type 2)
case and (b) outside damage (Type 3) case.

Figure 6. Contributions of first 10 principal components of WP energy ratio features from pole specimen testing data: (a)
individual variance contribution and (b) cumulative variance contribution.

the established matrix consists of 180 samples (5 tests
3 3 pole specimens 3 12 accelerometers) of pole test-
ing results from 12 accelerometers while the column of
the matrix is made up of 1024 energy ratio
features after 10-level WPT. After PCA projection, all
the samples are replaced by 1024 principal components.
Figure 6 shows the individual and cumulative contribu-
tions of the first 10 principal components of WP energy
ratios of pole signals, respectively. It is clearly seen that
the first three principal components make up almost
95% of the whole feature information. Even if 5% of
energy feature will be lost, the feature dimension can be
greatly decreased from 1024 to 3, which will improve
the learning efficiency of the model training.
To validate the performance of selected principal
components to classify different condition types, differ-
ent groups of combinations of principal components
are compared according to different sensors. Figure
7(a) demonstrates the related results from the first
accelerometer (A1). It is clearly observed that the first
three principal components are able to perfectly distin-
guish the three common pole conditions. Especially for
Yu et al.
Figure 7. Identification results of testing samples based on first three principal components: (a) result from sensor A1 and (b)
result from sensor A2.
the second and third principal components, the five
testing results almost overlap together in the two-
dimension feature space, where intact samples are
located at the coordinate (2, 1) and the inner damage
and outside damage cases are situated at (22, 26) and
(0, 21), respectively. However, due to different sensor
locations and sensor self-defect, not all the outcomes
from the different accelerometers deliver satisfactory
classification results based on first three principal com-
ponents. Figure 7(b) shows the condition classification
results of the second accelerometer (A2). Obviously,
the principal components of inner damage cases are
easily distinguished from those of the other damage
types. Nevertheless, for the intact and outside damage
cases, their feature indices are located in a similar area
of feature space and may be difficult to be identified.
Therefore, based on the selected principal components,
it is necessary for the mutual cooperation of all the
accelerometers to provide an accurate condition assess-
ment result.
To illustrate the superiority of the proposed WP
energy ratio as feature indices for damage pattern clas-
sification, the conventional signal processing method
such as FFT, which was applied in previous study, is
also adopted in this part for performance comparison.
For the sake of fair assessment, the same testing data
are used for signal feature extraction using FFT.
Similarly, the PCA results of extracted features from
sensors A1 and A2 are provided, which are shown in
Figure 8. Different from the PCs extracted from WP
energy features that almost overlap together for each
condition scenario, the PCs calculated from the FFT
data scatter over a large area in the three-dimensional
131
(3D) feature space, which may result in wrong identifi-
cation of wood pole condition. Especially for the
results from sensor A2, although the intact cases are
well separated from the damaged cases, it is definitely
difficult to distinguish the inner damage cases from the
outside damage cases according to first three PCs.
Apparently, compared with the FFT-based features,
the features calculated based on WP energy ratio can
better reflect the SW pattern related to different types
of damage and will be considered as damage indices for
the pole condition identification in practice.
Condition identification based on SVM and
D-S evidence theory
Although compressed feature indices are effective in
identifying different condition types, a model or a clas-
sifier is still needed to automatically implement the
classification procedure in the practice. Here, SVM is
adopted to construct a multi-label classification model
and D-S evidence theory is also used to fuse the results
from different accelerometers.
SVM multi-classifier for pole condition assessment
The typical application of SVM is to set up a two-
category classifier, which is used to estimate the tested
samples belonging to positive or negative category. In
this work, the objective of the designed classifier is to
identify three different damage types in wood poles,
which can be considered as a multi-target classification
problem. The most direct solution for this issue is to
configure an objective function to calculate multiple
132
Figure 8. PCA results of testing samples using FFT method: (a) result from sensor A1 and (b) result from sensor A2.
hyper-planes at a time, which are employed to separate
the space into different areas corresponding to different
categories.33,34 However, this method is not feasible in
practice owing to enormous calculation requirements.
In practical applications, there are two common
ways to solve the multi-target classification problem:
one against rest (OAR) and one against one (OAO).
The main principle for both methods is to decompose a
multi-label classifier into several two-label sub-classi-
fiers. For an n-category classification problem, the
OAR method requires to set up n (Cn1 ) sub-classifiers.
In the ith sub-classifier, the samples with the ith cate-
gory are considered as positive, while the samples with
other categories are used as negative during the model
training. The final identification result is made based
on the positive outputs. The main advantage of the
OAR method is that it needs fewer sub-classifiers than
the OAO method, but it may lead to overlapped classi-
fication or unclassified problem. The OAO method, on
the contrary, constructs sub-classifiers according to
arbitrary two categories in the training samples, there-
fore contributing to a total of n(n 2 1)/2 (Cn2 ) sub-clas-
sifiers. The final classification result is decided via votes
from all the sub-classifiers. The main feature of OAO is
that the number of sub-classifiers will sharply ascend
with the increasing categories, which may result in low
efficiency of the model training.
In this work, both OAR and OAO methods are
investigated for the development of SVM sub-classifiers
to identify the pole condition. As a consequence, for
the OAR method, there are three classifiers to be built,
that is, intact-rest sub-classifier, inner damage-rest sub-
classifier and outside damage-rest sub-classifier. For the
OAO method, three sub-classifiers should be
Structural Health Monitoring 18(1)
established, that is, intact-inner damage sub-classifier,
intact-outside damage sub-classifier and inner damage-
outside damage sub-classifier.
Parameter selection of SVM classifier
The essence of nonlinear classification based on SVM is
to transform the high-dimensional feature space in the
classification problem into the original input space for
processing via inner product relationship between ker-
nel function and mapping function. According to
Amari and Wu,35 there are several kernel functions,
which can be used in tackling the practical problems. In
this article, the radial basis function (RBF) is selected
due to good capacity for nonlinear regression and clas-
sification, and its mathematical expression is given in
equation (6)
(  2 )
  xi  xj 
K xi , xj = exp ð6Þ
2s2
where s2 denotes the width of the RBF, and xi and xj
denote two samples of training.
After the kernel function is fixed, two important
parameters should be properly assigned before SVM
model training, that is, penalty coefficient C and kernel
function parameter s2. C indicates the recognition level
of misclassification during the model training, while s2
mainly affects the smoothness degree of the classifica-
tion hyperplane. The parameter selection in SVM is
directly related to the classification accuracy and gener-
alization capacity of the classifier model. In this work,
GA is employed to calculate the optimal parameter val-
ues of C and s2 because of its perfect performance in
Yu et al.
Figure 9. Parameter optimization in SVM by GA: (a) parameter value and (b) fitness value.
solving complicated optimization algorithms.36,37 The
main optimization procedure can be concluded as
follows:
1. Initialize the population. In this work, the chromo-
some is made up of penalty coefficient C and ker-
nel function parameter s2. The population number
of chromosomes is set as 30. Every chromosome
has two segments: one represents the penalty coef-
ficient C and the other represents the kernel func-
tion parameter s2. The initial chromosomes are set
randomly.
2. Calculate the fitness value. Here, the fitness func-
tion is designed based on the classification accuracy
using n-fold cross validation, defined as the ratio of
correct classification number to total number. In
this study, the constraints of the optimized para-
meters are set in the range of [0 100].
3. Selection operation. The roulette wheel method is
utilized to choose part of the chromosome to gen-
erate the new chromosome.
4. Crossover operation. Exchange the genes between
two chromosomes randomly via single point way.
The crossover probability is set as 0.8 in this
optimization.
5. Mutation operation. Mutation operation is used to
change the binary code. If the value is 0, it will
transform to 1. Inversely, if the value is 1, it will
transform to 0. Here, the mutation probability is
set as 0.01.
6. Optimization termination. If the algorithm iteration
number exceeds its maximum value (100), the algo-
rithm will be terminated. Otherwise, go to (2).
133
Figure 9 presents the optimization procedure of C
and s2 in intact-inner damage sub-classifier from the
first accelerometer (A1), where Figure 9(a) shows the
parameter change tendency during the optimization
and Figure 9(b) plots the convergence rate of GA to
optimize the SVM sub-classifier. The changing curves
in Figure 9(a) demonstrate that both C and s2 are able
to get close to their optimal values very quickly within
20 steps of iteration. In addition to using GA to opti-
mize the SVM parameters, optimal classification accu-
racy of the training model does increase to 100%
although the mean accuracy is just above 85%. Using
the optimal parameter values, the SVM sub-classifier
can be trained with best performance.
Posteriori probability output of SVM classifier
The outputs of standard SVM belong to hard decision,
which assigns the tested samples with specific category
information. However, for practical applications such
as the presented damage identification and classifica-
tion, it is of great practical value to pole asset managers
to provide a probability estimate to each category. For
such uncertain classification problems, the SVM classi-
fier with soft-decision capacity is necessary. Therefore,
Platt38 put forward a posterior probability model based
on the sigmoid function to realize the probability out-
put of SVM, which is defined as follows
1
p(x) = ð7Þ
1 + e(D1 f (x) + D2 )
where f(x) denotes the standard SVM classifier and its
expression is f (x) = v  x + b, in which v denotes the
134 Structural Health Monitoring 18(1)

Table 2. BPAs of sub-classifier corresponding to OAR mode.

Sub-classifier BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

SVM 1 m1 m1 (T1) 0 0 0 m1 (T2[T3) 0 m1 (Y) 0

SVM 2 m2 0 m2 (T2) 0 0 0 m1 (T1[T3) m2 (Y) 0
SVM 3 m3 0 0 m3 (T3) m1 (T1[T2) 0 0 m3 (Y) 0

BPA: basic probability function; OAR: one against rest; SVM: support vector machine.

connection weight vector and b denotes the bias.33D1 positive and negative categories can be expressed as
and D2 are configuration parameters and could be cal- p(x) and 1 2 p(x), respectively. According to the evi-
culated through dealing with the following optimization dence theory, Q denotes the uncertainty and its assign-
problem ment is of great significance to the final classification
result. Here, the concept of classification error expec-
X
l tance upper bound of testing samples is introduced,
minimize  ½ti ln (pi ) + (1  ti ) ln (1  pi ) ð8Þ which is expressed as average support vector number
i=1
divided by the total training sample number
where pi = (1=(1 + e½D1 f (xi ) + D2  )), ti = ((yi + 1)=2)and yi
denotes the label of ith sample. Ns
Eu (Perror ) = ð11Þ
N 1

Fusion decision based on evidence theory where Ns and N denote the support vector number and
training sample number, respectively. This upper bound
In this part, evidence theory is introduced to aggregate reflects the classification uncertainty of tested samples,
the identification results from different sub-classifiers which corresponds to the uncertainty Q in the frame of
and different sensors. Evidence theory was proposed by discernment u. To make sure that the summation of
Dempster39 and Shafer,40 which adopts the probability probability assignments is 1, the BPAs of the two-class
theory to combine the evidences from different sources. SVM classifier should be rewritten as follows
In evidence theory, a finite set u with possible
hypotheses is defined as the frame of discernment. In  
Ns
this article, for the case of wood pole condition identifi- m(C + ) = p(x) 1  ð12Þ
N 1
cation, the frame of discernment is defined as u = {T1,
T2, T3}, where T1 denotes the intact condition, T2  
 Ns
denotes the inner damage condition and T3 denotes the m(C ) = ½1  p(x) 1  ð13Þ
N 1
outside damage condition. Additionally, basic prob-
ability functions (BPAs) are defined as the primitives in Ns
m(Y) = ð14Þ
evidence theory, which are used to assign the values N 1
between 0 and 1 to the set 2u, satisfying the below
where C + and C2 denote the positive and negative
relationships
categories, respectively. Tables 2 and 3 give BPAs of
X sub-SVM classifiers in both OAR and OAO modes,
m(T ) = 1 ð9Þ
respectively.
T 2u
Evidence combination is also crucial to the imple-
m([) = 0 ð10Þ mentation of evidence theory and the combination rule
between two pieces of evidence B and C is a conjunc-
where m(T) denotes the BPA of the primitive T, corre- tive sum operation, defined as follows
sponding to a piece of evidence. [ denotes the empty
set. When the evidence theory is applied with SVM, the m(T ) = 0, T = [ ð15Þ
posteriori probability outputs of SVM can be consid-
P
ered as BPAs of the frame of discernment. However, m1 (Bi )  m2 (Cj )
despite that sigmoid function p(x) in equation (7) is able Bi \Cj = T
m(T ) = m1  m2 = , T 6¼ [
to satisfactorily solve the soft-decision problem in 1  KC
SVM, it does not follow that this function is close to ð16Þ
the posterior probability of SVM outputs.41,42 Actually, X
this model can just map the support information into KC = m1 (Bi )  m2 (Cj ) ð17Þ
the interval [0, 1], which means that the BPAs of Bi \Cj = [
Yu et al. 135

Table 3. BPAs of sub-classifier corresponding to OAO mode.

Sub-classifier BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

SVM# 1 m01 m01 (T1 ) m01 (T2 ) 0 0 0 0 m01 (Y) 0

SVM# 2 m02 m02 (T1 ) 0 m02 (T3 ) 0 0 0 m02 (Y) 0
SVM# 3 m03 0 m03 (T2 ) m03 (T3 ) 0 0 0 m03 (Y) 0

BPA: basic probability function; OAO: one against one; SVM: support vector machine.

Figure 10. Validation procedure of the proposed method.

where KC denotes the conflict degree between two Case study

pieces of evidence. Likewise, the combination result of
multiple pieces of evidence corresponding to three sub-
classifiers from different sensors is calculated according
to the following equation
m = m1  m2      mn
After combining multiple pieces of evidence, we
obtain the BPA of each primitive in the set 2u. Then, a
final decision is made based on the maximal confidence
degree method, which is defined as follows.
For the arbitrary primitives T1 and T2 in the set 2u, if
the below equations are satisfied
 
m(T2 ) = max m(Tk ), Tk  2u and Tk 6¼ T1
m(T1 )  m(T2 ).g1
m(Q)\g2
m(T1 ).m(Q)
where g1 and g 2 are the evaluation coefficients, and T1
is the final recognition result. In this study, the values
of g 1 and g2 are set as 0.3 following the suggestion of
Wan et al.43
To validate the effectiveness of the proposed method,
an additional test is conducted on each pole specimen.
The captured SW signals are processed as validation
samples and sent to the trained SVM-based sub-classi-
ð18Þ fiers, the outputs of which are combined to get the final
diagnosis results using two-level data fusion. The
detailed validation procedure is shown in Figure 10. It
is clearly seen that in the first-level fusion, the identifi-
cation results from different accelerometers are com-
bined using D-S method at each sub-classifier, and
then in the second-level fusion, the first-level fusion
results from different sub-classifiers are combined to
output a final decision.
ð19Þ Tables 4 to 6 show the BPAs from three sub-SVM
classifiers in OAR mode using the validation data of
ð20Þ
intact case, respectively. Table 4 lists the results of the
ð21Þ intact-rest sub-classifier to identify the intact wood
pole. It can be seen that most accelerometers assign
ð22Þ more confidence (BPA value) to intact primitive (T1)
rather than rest primitive (T2[T3), with the exceptions
of accelerometers A2 and A11. However, the decision
rule is not satisfied, because the uncertainty (Q) makes
up the largest BPA in each piece of evidence, which
136 Structural Health Monitoring 18(1)

Table 4. BPAs of intact-rest sub-classifier of the intact case validation data.

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m1,1 0.2493 0 0 0 0.0364 0 0.7143 0

m1,2 0.0951 0 0 0 0.1192 0 0.7857 0
m1.3 0.4258 0 0 0 0.0742 0 0.5000 0
m1,4 0.2172 0 0 0 0.0685 0 0.7143 0
m1,5 0.2053 0 0 0 0.0090 0 0.7857 0
m1,6 0.3227 0 0 0 0.0344 0 0.6429 0
m1,7 0.2624 0 0 0 0.0233 0 0.7143 0
m1,8 0.1792 0 0 0 0.0351 0 0.7857 0
m1,9 0.1305 0 0 0 0.0838 0 0.7857 0
m1,10 0.2084 0 0 0 0.0059 0 0.7857 0
m1,11 0.0743 0 0 0 0.1400 0 0.7857 0
m1,12 0.1887 0 0 0 0.0256 0 0.7857 0

BPA: basic probability function. Bold values demonstrate that the values from initial recognition are not consistent with the practical condition, which
will make it difficult for the system to make a correct decision.

Table 5. BPAs of inner damage-rest sub-classifier of the intact case validation data

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m2,1 0 0.0124 0 0 0 0.2019 0.7857 0

m2,2 0 0.0484 0 0 0 0.5230 0.4286 0
m2.3 0 0.0173 0 0 0 0.1970 0.7857 0
m2,4 0 0.0105 0 0 0 0.4181 0.5714 0
m2,5 0 0.0496 0 0 0 0.3790 0.5714 0
m2,6 0 0.0177 0 0 0 0.1966 0.7857 0
m2,7 0 0.0046 0 0 0 0.2811 0.7143 0
m2,8 0 0.0318 0 0 0 0.1824 0.7857 0
m2,9 0 0.0178 0 0 0 0.1965 0.7857 0
m2,10 0 0.0294 0 0 0 0.3277 0.6429 0
m2,11 0 0.0657 0 0 0 0.1486 0.7857 0
m2,12 0 0.0437 0 0 0 0.1706 0.7857 0

BPA: basic probability function. Bold values demonstrate that the values from initial recognition are not consistent with the practical condition, which
will make it difficult for the system to make a correct decision.

Table 6. BPAs of outside damage-rest sub-classifier of the intact case validation data.

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m3,1 0 0 0.0165 0.1977 0 0 0.7857 0

m3,2 0 0 0.0582 0.1561 0 0 0.7857 0
m3.3 0 0 0.0111 0.2031 0 0 0.7857 0
m3,4 0 0 0.0209 0.2648 0 0 0.7143 0
m3,5 0 0 0.0307 0.2551 0 0 0.7143 0
m3,6 0 0 0.0241 0.333 0 0 0.6429 0
m3,7 0 0 0.0112 0.2031 0 0 0.7857 0
m3,8 0 0 0.0338 0.3948 0 0 0.5714 0
m3,9 0 0 0.0434 0.4566 0 0 0.5000 0
m3,10 0 0 0.0373 0.3913 0 0 0.5714 0
m3,11 0 0 0.0400 0.5314 0 0 0.4286 0
m3,12 0 0 0.0152 0.3419 0 0 0.6429 0

BPA: basic probability function. Bold values demonstrate that the values from initial recognition are not consistent with the practical condition, which
will make it difficult for the system to make a correct decision.

makes it difficult for the sub-classifier to directly output to this phenomenon is the limited training samples, in
the identification result. The main reason contributing which most of them act as the support vectors. Tables 5
Yu et al. 137

Table 7. First-level evidence combination result of intact pole case (OAR).

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m1 0.9036 0 0 0 0.0558 0 0.0406 0

m2 0 0.0113 0 0 0 0.9719 0.0168 0
m3 0 0 0.0057 0.9863 0 0 0.0079 0

BPA: basic probability function; OAR: one against rest. Bold values denotes the the proposition with largest probability and uncertainty values, which
indicates that after evidence combination, the probabilities of correct proposition almost rises to 1 (100%) and uncertainty degree almost reduce to
0. In this case, the system is easy to make a correct decision.

Table 8. Second-level evidence combination result of intact pole case (OAR).

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m 0.9957 0.0022 0.0010 0.0007 0 0.0003 0 0

BPA: basic probability function; OAR: one against rest. Bold values denotes the the proposition with largest probability and uncertainty values, which
indicates that after evidence combination, the probabilities of correct proposition almost rises to 1 (100%) and uncertainty degree almost reduce to
0. In this case, the system is easy to make a correct decision.

Table 9. Evidence combination result of inner damage pole case based on OAR-based sub-classifiers.

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m 0.0018 0.9966 0.0006 0.0006 0.0002 0.0001 0 0

BPA: basic probability function; OAR: one against rest. Bold values denotes the the proposition with largest probability and uncertainty values, which
indicates that after evidence combination, the probabilities of correct proposition almost rises to 1 (100%) and uncertainty degree almost reduce to
0. In this case, the system is easy to make a correct decision.

and 6 provide the results of inner damage-rest
sub-classifier and outside damage-rest sub-classifier to
identify the intact specimen. Similar to intact-rest sub-
classifier, more BPA values are allocated to rest primi-
tives (T1[T3 and T1[T2) rather than inner damage
outside primitive (T2) and damage primitive (T3).
Likely, larger uncertainties lead to failure in making a
recognition result at each accelerometer.
As a consequence, a two-level evidence combination
operation should be implemented. As displayed in
Figure 8, in order to avoid a large amount of calcula-
tion, first we combine the BPA results with same type
of sub-classifiers corresponding to different acceler-
ometers first, and then we combine the BPA results
with different types of sub-classifiers. Table 7 shows
the first-level evidence combination results for intact
pole case. It is found that for the intact-rest sub-classi-
fier, the support probability of proposition T1 (intact)
has ascended to 0.9036 while the support probabilities
of proposition T2[T3 (rest) and uncertainty have
decreased to 0.0558 and 0.0406, respectively. In the
same way, for the inner damage-rest and outside dam-
age-rest sub-classifiers, the support probabilities of pro-
positions T1[T3 and T1[T2 have increased to 0.9719
and 0.9863, respectively. The uncertainties for both
sub-classifiers are declined to 0.0168 and 0.0079,
respectively. According to the results in Table 7, it is
noticeable that the identification result of the three sub-
classifiers from different sensors is prone to T1. Hence,
we continue to integrate the results of the three sub-
classifiers to obtain a final decision, which is given in
Table 8. It is clearly seen that the support probabilities
of right proposition T1 eventually rises to 0.9957, which
is close to 100%. The support probabilities of wrong
propositions T2 and T3 (inner damage and outside
damage) reduce to 0.0022 and 0.0010, respectively.
What is more, the uncertainty degree descends to 0.
Therefore, in accordance with the decision rule (equa-
tions (19) to (22)), the final decision result is T1, which
is in agreement with actual condition of the tested pole
(intact). This shows that after two-level evidence com-
bination, the support probability of right proposition is
greatly improved compared to the result from outputs
of the sub-classifiers.
In the same way, for the cases of inner damage and
outside damage, the BPA results after the two-level evi-
dence combination based on OAR-based sub-classifiers
are given in Tables 9 and 10. According to the decision
rule, the decision results for both cases are T2 (inner
138 Structural Health Monitoring 18(1)

Table 10. Evidence combination result of outside damage pole case based on OAR-based sub-classifiers.

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m 0.0011 0.0004 0.9980 0.0001 0.0001 0.0003 0 0

BPA: basic probability function; OAR: one against rest. Bold values denotes the the proposition with largest probability and uncertainty values, which
indicates that after evidence combination, the probabilities of correct proposition almost rises to 1 (100%) and uncertainty degree almost reduce to
0. In this case, the system is easy to make a correct decision.

Table 11. BPAs of intact-inner damage sub-classifier of the inner damage case validation data.

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m01, 1 0.0142 0.0969 0 0 0 0 0.8888 0

m01, 2 0.0455 0.2878 0 0 0 0 0.6666 0
m01, 3 0.0474 0.2860 0 0 0 0 0.6666 0
m01, 4 0.0459 0.2874 0 0 0 0 0.6666 0
m01, 5 0.0138 0.0973 0 0 0 0 0.8888 0
m01, 6 0.0135 0.0976 0 0 0 0 0.8888 0
m01, 7 0.0834 0.4721 0 0 0 0 0.4444 0
m01, 8 0.0372 0.0739 0 0 0 0 0.8888 0
m01, 9 0.0289 0.0822 0 0 0 0 0.8888 0
m01, 10 0.0061 0.1051 0 0 0 0 0.8888 0
m01, 11 0.0616 0.0495 0 0 0 0 0.8888 0
m01, 12 0.0078 0.1033 0 0 0 0 0.8888 0

BPA: basic probability function. Bold values demonstrate that the values from initial recognition are not consistent with the practical condition, which
will make it difficult for the system to make a correct decision.

Table 12. BPAs of intact-outside damage sub-classifier of the inner damage case validation data.

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m02, 1 0.0845 0 0.2488 0 0 0 0.6666 0

m02, 2 0.0569 0 0.0542 0 0 0 0.8888 0
m02, 3 0.0956 0 0.2378 0 0 0 0.6666 0
m02, 4 0.0135 0 0.0977 0 0 0 0.8888 0
m02, 5 0.1019 0 0.0092 0 0 0 0.8888 0
m02, 6 0.1602 0 0.1731 0 0 0 0.6666 0
m02, 7 0.0011 0 0.1101 0 0 0 0.8888 0
m02, 8 0.2321 0 0.1013 0 0 0 0.6666 0
m02, 9 0.2789 0 0.0544 0 0 0 0.6666 0
m02, 10 0.0875 0 0.0236 0 0 0 0.8888 0
m02, 11 0.2679 0 0.0654 0 0 0 0.6666 0
m02, 12 0.1760 0 0.1574 0 0 0 0.6666 0

BPA: basic probability function. Bold values demonstrate that the values from initial recognition are not consistent with the practical condition, which
will make it difficult for the system to make a correct decision.

damage) and T3 (outside damage), respectively, satisfy-
ing the actual conditions of the specimens.
Tables 11 to 13 provide the outputs (BPAs) from the
OAO-based sub-classifiers for the inner damage valida-
tion case, respectively. As it is evident, for the intact-
inner damage and inner damage-outside damage sub-
classifiers, more support probabilities are assigned to
T2 (inner damage) than T1 (intact) and T3 (outside
damage), which meets the actual pole condition.
Similar to the OAR-based sub-classifiers, the OAO-
based sub-classifiers also suffer from the problem of
difficulty in decision making due to more uncertainties.
Thus, the two-level evidence combination is also
applied here to integrate the identification results from
different sub-classifiers corresponding to different sen-
sors. Tables 14 and 15 demonstrate the first-level and
second-level evidence combination results, respectively.
After the two-level data fusion, the support probability
Yu et al. 139

Table 13. BPAs of inner damage-outside damage sub-classifier of the inner damage case validation data.

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m03, 1 0 0.0921 0.0191 0 0 0 0.8888 0

m03, 2 0 0.2882 0.0451 0 0 0 0.6666 0
m03, 3 0 0.0747 0.0364 0 0 0 0.8888 0
m03, 4 0 0.0952 0.0159 0 0 0 0.8888 0
m03, 5 0 0.2848 0.0485 0 0 0 0.6666 0
m03, 6 0 0.1001 0.0111 0 0 0 0.8888 0
m03, 7 0 0.0928 0.0184 0 0 0 0.8888 0
m03, 8 0 0.0961 0.0151 0 0 0 0.8888 0
m03, 9 0 0.4811 0.0744 0 0 0 0.4444 0
m03, 10 0 0.2914 0.0419 0 0 0 0.6666 0
m03, 11 0 0.2908 0.0425 0 0 0 0.6666 0
m03, 12 0 0.1031 0.0081 0 0 0 0.8888 0

BPA: basic probability function. Bold values demonstrate that the values from initial recognition are not consistent with the practical condition, which
will make it difficult for the system to make a correct decision.

Table 14. First-level evidence combination result of inner damage pole case (OAO).

Classifier BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

SVM# 1 m01 0.057 0.8696 0 0 0 0 0.0733 0

SVM# 2 m02 0.5311 0 0.3829 0 0 0 0.0861 0
SVM# 3 m03 0 0.9058 0.0399 0 0 0 0.0543 0

BPA: basic probability function; OAO: one against one. Bold values denotes the the proposition with largest probability and uncertainty values, which
indicates that after evidence combination, the probabilities of correct proposition almost rises to 1 (100%) and uncertainty degree almost reduce to
0. In this case, the system is easy to make a correct decision.

Table 15. Evidence combination result of different sub-classifiers of inner damage pole case (OAO).

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m0 0.0474 0.9144 0.0341 0 0 0 0.0041 0

BPA: basic probability function; OAO: one against one.

Table 16. Evidence combination result of intact pole case based on OAO-based sub-classifiers.

BPA T1 T2 T3 T1[T2 T2[T3 T1[T3 Y ;

m0 0.9220 0.0465 0.0277 0 0 0 0.0038 0

BPA: basic probability function; OAO: one against one. Bold values denotes the the proposition with largest probability and uncertainty values, which
indicates that after evidence combination, the probabilities of correct proposition almost rises to 1 (100%) and uncertainty degree almost reduce to
0. In this case, the system is easy to make a correct decision.

of right proposition (T2) ascends from less than 0.4721
to 0.9144 while the uncertainty degree declines to
0.0041. According to the decision rule, final decision
for this case is T2, which matches the real condition of
the validation case as well.
The final combination results for intact and outside
damage cases using the OAO-based sub-classifiers are
given in Tables 16 and 17, in which the probabilities of
right propositions are 0.9220 and 0.9546, respectively,
and thus the final decision can be made with high confi-
dence. Considering all the condition identification
results, it is clearly found that the evidence combination
is able to effectively improve the support probabilities
of right propositions and weaken the influence caused
140
Table 17. Evidence combination result of outside damage pole case based on OAO-based sub-classifiers.
BPA T1 T2 T3
m0 0.0214 0.0212 0.9546
BPA: basic probability function; OAO: one against one. Bold values denotes the the proposition with largest probability and uncertainty values, which
indicates that after evidence combination, the probabilities of correct proposition almost rises to 1 (100%) and uncertainty degree almost reduce to
0. In this case, the system is easy to make a correct decision.
by uncertainties. Compared to the results from the
OAO-based sub-classifiers, the evidence combination
from OAR-based sub-classifiers provides higher identi-
fication accuracy but requires more calculation costs.
On the contrary, with the adding classification cate-
gories, the number of trained sub-classifiers based on
OAO method (n(n 2 1)/2, n denotes the category num-
ber) will be sharply increased compared with that based
on OAR method (n, n denotes the category number).
Consequently, in the engineering application, the type
of SVM-based multiple classifier should be selected
according to the practical requirement.
Conclusion
This work presents a new wood pole condition identi-
fication approach based on advanced signal process-
ing, machine learning and decision-level data fusion
techniques. WPT together with PCA is used to extract
inherent damage features in SW signals obtained
from a novel narrow frequency band-excited NDT
system. Then, the SVM multi-classifier with posteriori
probability output is designed to provide the damage
classification decision at the sensor-level. D-S evi-
dence theory is subsequently employed to aggregate
the decisions from different sensors, contributing to a
final condition assessment result. The proposed
approach is experimentally validated on three wood
pole specimens with different types of damage. The
results demonstrate that the presented method is
capable of providing satisfactory damage condition
identification of wood utility poles along with practi-
cally useful probability information, meeting the
requirements of asset management and engineering
application. In this work, limited wood pole speci-
mens, with known material and environmental prop-
erties, are used for the method validation. In future
work, more specimens with different material and
environmental properties as well as boundary condi-
tions will be considered in the laboratory testing.
Moreover, to further validate and assess the perfor-
mance of the method, the proposed technique will be
tested in the field on in-situ wood poles with natural
defects.
Structural Health Monitoring 18(1)
T1[T2 T2[T3 T1[T3 Y ;
0 0 0 0.0027 0
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.
Funding
The author(s) disclosed receipt of the following financial sup-
port for the research, authorship, and/or publication of this
article: This work was funded by the Linkage Project (No.
LP110200162) from Australian Research Council and
Ausgrid as well as UTS International Research Development
Scheme and ATN-DAAD collaborative researcher exchange
grant.
ORCID iD
Yang Yu https://orcid.org/0000-0001-9592-8191
References
1. Nguyen M, Foliente G and Wang XM. State-of-the-prac-
tice and challenges in non-destructive evaluation of utility
poles in service. Key Eng Mater 2004; 270: 1521–1528.
2. Rizzo P and Scalea FLD. Feature extraction for defect
detection in strands by guided ultrasonic waves. Struct
Health Monit 2006; 5: 297–308.
3. Horáček P, Tippner J and Hassan KT. Nondestructive
evaluation of static bending properties of scots pine wood
using stress wave technique. Wood Res-Slovakia 2012; 57:
359–366.
4. Tallavo F, Pandey MD and Cascante G. Experimental
and numerical methods for detection of voids in wood
poles using ultrasonic testing. J Mater Civil Eng 2013; 25:
772–780.
5. Kasal B. Structural health assessment of in situ wood: an
interface between service life planning and wood engi-
neering. Wood Mater Sci Eng 2014; 9: 134–138.
6. Tallavo F, Cascante G and Pandey MD. Experimental
verification of an orthotropic finite element model for
numerical simulations of ultrasonic testing of wood poles.
Eur J Wood Wood Prod 2017; 75: 543–551.
7. Azari H, Nazarian S and Yuan D. Assessing sensitivity
of impact echo and ultrasonic surface waves methods for
nondestructive evaluation of concrete structures. Constr
Build Mater 2014; 71: 384–391.
8. Steinbach J and Vey E. Caisson evaluation by stress
wave propagation method. J Geotech Eng 1975; 101:
361–378.
Yu et al.
9. Sajid SH, Ali SM, Carino NJ, et al. Strength estimation
of concrete masonry units using stress-wave methods.
Constr Build Mater 2018; 163: 518–528.
10. Davis AG and Dunn CS. From theory to field experience
with the non-destructive vibration testing of piles. P I
Civil Eng 1974; 57: 571–593.
11. Sales A, Candian M and De Salles Cardin V. Nondes-
tructive evaluation of wood: the new Brazilian code for
the design of wood structures. Mater Struct 2010; 43:
213–221.
12. Dackermann U, Crews K, Kasal B, et al. In situ assess-
ment of structural wood using stress-wave measurements.
Mater Struct 2014; 47: 787–803.
13. Andreassen E, Manktelow K and Ruzzene M. Direc-
tional bending wave propagation in periodically perfo-
rated plates. J Sound Vib 2015; 335: 187–203.
14. Yin J and Hopkins C. Modelling bending wave transmis-
sion across coupled plate systems comprising periodic
ribbed plates in the low-, mid-, and high-frequency ranges
using forms of statistical energy analysis. Shock Vib 2015;
2015: 467875.
15. Subhani M, Li J and Samali B. A comparative study of
guided wave propagation in wood poles with isotropic
and transversely isotropic material models. J Civil Struct
Health Monit 2013; 3: 65–79.
16. Yan N, Li J, Dackermann U, et al. A numerical investi-
gation on the damage identification of wood utility
poles based on wavelet packet energy. In: Proceedings
of 23th Australasian conference on the mechanics of
structures and materials, Southern Cross University:
Lismore (City), Byron Bay, NSW, Australia, 6–9
December 2014, pp. 1185–1190.
17. Jozi B, Dackermann U, Braun R, et al. Application and
improvement of conventional stress-wave-based non-
destructive testing methods for the condition assessment
of in-service wood utility poles. In: Proceedings of 23th
Australasian conference on the mechanics of structures and
materials, Southern Cross University: Lismore (City),
Byron Bay, NSW, Australia, 6–9 December 2014, pp.
1197–1202.
18. Dackermann U, Yu Y, Niederleithinger E, et al. Condi-
tion assessment of foundation piles and utility poles based
on guided wave propagation using a network of tactile
transducers and support vector machines. Sensors 2017;
17: 2938.
19. Nguyen T, Chan THT, Thambiratnam DP, et al. Devel-
opment of a cost-effective and flexible vibration DAQ
system for long-term continuous structural health moni-
toring. Mech Syst Signal Pr 2015; 64–65: 313–324.
20. Li J and Hao H. Health monitoring of joint conditions in
steel truss bridges with relative displacement sensors.
Measurement 2016; 88: 360–371.
21. Li J and Hao H. Damage detection of shear connectors
under moving loads with relative displacement measure-
ments. Mech Syst Signal Pr 2015; 60: 124–150.
22. Peng XL, Hao H and Li ZX. Application of wavelet
packet transform in subsea pipeline bedding condition
assessment. Eng Struct 2012; 39: 50–65.
141
23. Bandara RP, Chan THT and Thambiratnam DP. Fre-
quency response function based damage identification
using principal component analysis and pattern recogni-
tion technique. Eng Struct 2014; 66: 116–128.
24. Wang Y and Hao H. Damage identification scheme based
on compressive sensing. J Comput Civil Eng 2013; 29:
04014037.
25. Tan ZX, Thambiratnam DP, Chan THT, et al. Detecting
damage in steel beams using modal strain energy based
damage index and Artificial Neural Network. Eng Fail
Anal 2017; 79: 253–262.
26. Hakim SJS and Abdul Razak H. Adaptive neuro fuzzy
inference system (ANFIS) and artificial neural networks
(ANNs) for structural damage identification. Struct Eng
Mech 2013; 45: 779–802.
27. Asgarian B, Aghaeidoost V and Shokrgozar HR. Dam-
age detection of jacket type offshore platforms using rate
of signal energy using wavelet packet transform. Mar
Struct 2016; 45: 1–21.
28. Tabrizi A, Garibaldi L, Fasana A, et al. Early damage
detection of roller bearings using wavelet packet decom-
position, ensemble empirical mode decomposition and
support vector machine. Meccanica 2015; 50: 865–874.
29. Facchini G, Bernardini L, Atek S, et al. Use of the wave-
let packet transform for pattern recognition in a struc-
tural health monitoring application. J Intel Mat Syst Str
2015; 26: 1513–1529.
30. Sun Z and Chang CC. Vibration based structural health
monitoring: wavelet packet transform based solution.
Struct Infrastruct E 2007; 3: 313–323.
31. Sierra-Pérez J, Torres-Arredondo M-A and Alvarez-
Montoya J. Damage detection methodology under vari-
able load conditions based on strain field pattern recogni-
tion using FBGs, nonlinear principal component analysis,
and clustering techniques. Smart Mater Struct 2018; 27:
015002.
32. Tibaduiza DA, Mujica LE, Rodellar J, et al. Structural
damage detection using principal component analysis and
damage indices. J Intel Mat Syst Str 2016; 27: 233–248.
33. Hasni H, Alavi AH, Jiao P, et al. Detection of fatigue
cracking in steel bridge girders: a support vector machine
approach. Arch Civ Mech Eng 2017; 17: 609–622.
34. Xu Q. Impact detection and location for a plate structure
using least squares support vector machines. Struct
Health Monit 2014; 13: 5–18.
35. Amari S and Wu S. Improving support vector machine
classifiers by modifying kernel functions. Neural Net-
works 1999; 12: 783–789.
36. Gui G, Pan H, Lin Z, et al. Data-driven support vector
machine with optimization techniques for structural
health monitoring and damage detection. KSCE J Civ
Eng 2017; 21: 523–534.
37. Plati C, Georgouli K, Cliatt B, et al. Incorporation of
GPR data into genetic algorithms for assessing recycled
pavements. Constr Build Mater 2017; 154: 1263–1271.
38. Platt JC. Probabilistic outputs for support vector
machines and comparisons to regularized likelihood
methods. In: Smola AJ, Bartlett PL, Scholköpf B, et al.
142
(eds) Advances in large margin classifiers. Cambridge,
MA: MIT Press, 1999, pp. 61–74.
39. Dempster AP. Upper and lower probabilities induced by
a multi-valued mapping. Ann Math Statist 1967; 38:
325–339.
40. Shafer G. A mathematical theory of evidence. Princeton,
NJ: Princeton University Press, 1976.
Structural Health Monitoring 18(1)
41. Tipping ME. Sparse Bayesian learning and the relevance
vector machine. J Mach Learn Res 2001; 21: 211–244.
42. Li M and Sethi IK. Confidence-based active learning.
IEEE T Pattern Anal 2006; 28: 1251–1261.
43. Wan J, Yu Y, Wu Y, et al. Hierarchical leak detection
and localization method in natural gas pipeline monitor-
ing sensor networks. Sensors 2012; 12: 189–214.