Barbosh et al.

J Infrastruct Preserv Resil (2022) 3:6 Journal of Infrastructure

https://doi.org/10.1186/s43065-022-00051-8
Preservation and Resilience

RESEARCH Open Access

Acoustic emission‑based damage

localization using wavelet‑assisted deep
learning
Mohamed Barbosh, Kyle Dunphy and Ayan Sadhu*

Abstract
Acoustic Emission (AE) has emerged as a popular damage detection and localization tool due to its high performance
in identifying minor damage or crack. Due to the high sampling rate, AE sensors result in massive data during long-
term monitoring of large-scale civil structures. Analyzing such big data and associated AE parameters (e.g., rise time,
amplitude, counts, etc.) becomes time-consuming using traditional feature extraction methods. This paper proposes a
2D convolutional neural network (2D CNN)-based Artificial Intelligence (AI) algorithm combined with time–frequency
decomposition techniques to extract the damage information from the measured AE data without using standalone
AE parameters. In this paper, Empirical Mode Decomposition (EMD) is employed to extract the intrinsic mode func-
tions (IMFs) from noisy raw AE measurements, where the IMFs serve as the key AE components of the data. Continu-
ous Wavelet Transform (CWT) is then used to obtain the spectrograms of the AE components, serving as the “artificial
images” to an AI network. These spectrograms are fed into 2D CNN algorithm to detect and identify the potential loca-
tion of the damage. The proposed approach is validated using a suite of numerical and experimental studies.
Keywords: NDT, Damage detection and localization, AE, EMD, WT, CNN, Acoustic sensor

Introduction propagation [27]. Depending on the nature of structures

Civil structures are prone to damage due to structural and loading conditions, a suite of AE parameters such as
ageing, faulty construction, natural disasters, acciden- duration, signal strength, amplitude, rise time, counts,
tal loads or lack of adequate, timely maintenance. Their and energy indices reflect their as-is state and detect any
existing condition can be monitored using sensor-driven anomalies. However, these parameters are often sensi-
structural health monitoring (SHM) strategies to avoid tive to operational conditions, the severity of damage and
any catastrophic failure [18]. SHM involves data-driven measurement noise present in the data. In this paper, an
technology to monitor and detect the damage and pro- improved damage detection and localization technique is
vide the infrastructure owners opportunities for timely developed using a deep learning method augmented with
maintenance. Over the last several years, Acoustic Emis- the time–frequency decomposition of AE data.
sion (AE) has shown great promise to detect and localize Over the last several years, AE technique has been
different internal damage patterns such as crack, fatigue, applied to detect and localize damage in various struc-
corrosion in civil structures [32]. In general, AE is defined tural elements such as beam, slab, wall, and full-scale
as a transient elastic wave generated as an outcome of structures [1, 2, 5, 9, 41]. For example Abouhussien and
a material deformation due to damage initiation and Hassan [3], proposed an AE monitoring system to evalu-
ate the performance of bonds of reinforcement rebars
subjected to corrosion in reinforced concrete structures.
*Correspondence: [email protected] The traditional AE parameters were considered to detect,
Department of Civil and Environmental Engineering, Western University,
ON, London, Canada
quantify, and classify the bond damage. The delamination

© The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 2 of 24

defects in concrete containment structures were identi- suitable selection of features to identify the severity and
fied using the AE-based technique [12]. AE parameters location of the damage using AE data.
obtained from an instrumented concrete wall were con- With the advancement in deep learning (DL)-based
sidered to detect and classify the damage using a novel Artificial Intelligence techniques, the SHM researchers
clustering and visualization approach. Quy and Kim [26] have explored various Convolutional Neural Networks
proposed a method based on AE data to detect and local- (CNNs) for crack localization [11, 35]. In these methods,
ize the defect in pipeline systems. The time–frequency the actual images of the crack can be used as the train-
decomposition method was used to detect the AE event ing data for the identification and classification of cracks
as well as localize the AE source by determining the time without requiring any specific feature selection. Recently,
difference of arrival. An experimental test was conducted fast Fourier transform (FFT) and various other time–fre-
on a pipeline system subjected to internal pressure while quency (TF) methods such as short-term Fourier trans-
collecting AE data using AE sensors. It was concluded form (STFT), wavelet transform (WT) [4, 28], empirical
that the proposed method could be applied for both WT (EWT) [13, 43] transform (HHT [6, 15] empirical
offline and online monitoring of ageing pipelines. mode decomposition (EMD) [7, 31], and synchro-squeez-
In a recent study, Manthei and Plenkes [22] provided ing transform (SST) [19, 24, 34], have been explored to
a comprehensive review of AE-based structural con- obtain TF images that were fed into DL techniques in
dition assessment in underground structures. In this both mechanical and structural systems [38]. explored
research, the AE technique was applied to detect the dif- three different TF methods (i.e., STFT, WT and HHT),
ferent levels of deformation in mines. On the other hand, and the resulting TF images of bearing diagnostic data
Calabrese and Proverbio [10] provided a review of the were used as the input to the CNN. It was shown that
AE techniques to detect the damage due to corrosion in the resizing of TF-images greatly minimized the training
various engineering systems. In addition, the study dis- time but significantly reduced the classification accuracy
cussed the importance of the AE technique as a damage [40]. used eight different TF methods, and the TF images
detection and localization tool for material or systems were used as the input to AlexNet-based DL architec-
subjected to corrosion-induced damage. On the other ture. Two bearing datasets with different loads and
hand, Verstrynge et al. [39] discussed the application speeds have been used to validate the proposed method.
of AE-based methods to detect, assess, and identify the Whereas STFT provides low-resolution TF images due to
severity of damage in masonry structures. The study pro- the fixed window size, CWT can provide high-resolution
vided a detailed discussion regarding the current chal- images in both time and frequency domain due to its self-
lenges, findings, and future works that can be conducted adaption of window size [29, 38, 40].
to improve the performance of AE-based methods as a Pandhare et al. [25] used spectrograms as the input
damage detection tool for masonry structures. to the CNN. The performance of CNN was compared
Ma and Du [23] integrated AE parameters into an with conventional machine learning methods [14].
advanced deep neural network to detect cracks in pre- used EWT to convert current signals of a three-phase
stressed concrete elements. A set of experimental and induction motor to 2D grayscale images, which were
full-scale studies were undertaken to extract the crack used in CNN to classify several types of faults, includ-
signature of the prestressed concrete samples using AE ing bearing axis deviation, stator and rotor friction,
features. In a recent study, Wang et al. [42] developed a rotor aluminum end ring break, bearing noise and poor
new technique based on AE measurements to evaluate insulation. Sun et al. [36] adopted Multi-SST to ana-
the damage in concrete elements. The wavelet packet lyze time-varying signals for obtaining TF matrices
transform method was utilized to decompose the meas- that could accurately reflect the raw signal information.
ured AE signal into a set of frequency bands, followed by Sparse Feature Coding and Dictionary Learning were
calculating the energy of the coefficient in each frequency then implemented to extract the low dimensional and
band. The variation in energy across various coefficients most discriminative features. Liu et al. [20] proposed
was used to estimate the damage level. On the other a fault diagnosis method based on CNN and transfer
hand, Barbosh et al. [8] applied a new feature extraction learning, where CWT was used to extract the features
technique by combining EMD with Shannon Entropy in of the sampled signal data, and the resulting scalograms
AE data to detect and localize minor damage. The pro- were used as input to the LeNet-5 CNN architecture.
posed technique was used to monitor and evaluate the It was shown that the results of CNN converged very
existing condition of the walls in a dam under various fast with reasonable accuracy. Zhang et al. [44] used 2D
operational conditions. However, most of these studies spectrograms which were used as input to the LeNet-5
are either based on pattern recognition or signal process- CNN structure. It was demonstrated that the proposed
ing techniques that require a computationally intensive method had a high fault-diagnosis accuracy rate and

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 3 of 24

strong classification ability for various health condi- experimental studies is conducted next to show the per-
tions. In another study, Kumar et al. [17] implemented formance of the proposed method.
grayscale scalograms of an accelerometer installed on
a bearing housing as the input to the CNN. Shao et al.
[33] proposed a DL-based multi-signal fault diagno- Proposed AE‑based damage localization method
sis method in which the acquired data was converted Continuous wavelet transform
into scalograms that were used as input to the CNN. WT provides an improved TF representation of the sig-
However, most of the research associated with TF nals in a multi-resolution framework. Continuous wave-
image-based deep learning has been explored only in let transform (CWT) is considered one of the powerful
mechanical systems; there has been a very limited focus signal processing techniques that are applied in various
on SHM. applications such as image compression, noise removal,
Contrasting machine health monitoring, there has and pattern recognition [29] CWT of a signal x(t) is
been minimal research conducted in SHM where the TF defined as:
images were fed into CNN for damage or anomaly detec-  ∞
tion. Tang et al. [37] proposed a data anomaly detection r t−m
cm (x) = x(t)ψ ∗ ( )dt (1)
method based on a 2D CNN. FFT was implemented to −∞ r
generate frequency domain images of the segmented
The inverse CWT (ICWT) can be determined as
time series, which were analyzed in CNN for data anom-
follows:
aly classification. The results showed that the proposed
approach could detect the multipattern anomalies (such ∞  
1 1 r   ∗ t−m dr

as outlier, minor, missing data, trend and drift) of SHM x(t) = cm y ψ dm 2 (2)
data with high accuracy. Acceleration data of a real long‐ Cϕ |r| r r
−∞
span cable‐stayed bridge was used for validation, and
the results were compared with a deep neural network- where Cϕ can be expressed as:
based framework. Li et al. [21] used SST to represent the
|ψ(f )|2
 ∞
energies of AE signals in the TF domain. A multi-branch Cϕ = df < ∞ (3)
|f |
CNN model with Adam parameter optimization algo- −∞

rithm was developed as the feature extractor and clas- where r and m represent scale and translation of the
sifier to automatically distinguish multiple types of AE mother wavelet, respectively. The basis function is called
waves resulting from crack, impact and operational con- mother wavelet ψ(t), where superscript (∗) denotes its
ditions. The proposed model utilized the TF energy dis- complex conjugate. With the appropriate choice of r
tribution features of AE waves in a rail track and achieved and m, CWT utilizes the shifted and scaled versions of
higher accuracy than the traditional CNN models. How- ψ and subsequently estimates its inner product with x(t).
ever, most of these applications were employed directly In this paper, CWT is applied to generate spectrograms
using the vibration data, which are only suitable for of key AE components (i.e., IMFs) that are used to feed
global damage detection and have a limited application the CNN model to detect and identify the potential loca-
with respect to localized damage detection such as crack tion of the damage. A mixture of sine signals with three
localization. different frequencies, including f1= 1.4 Hz, f2= 3.5 Hz,
This study aims to propose a new AE data-based 2D and f3 = 7.0 Hz is considered to show the performance
CNN model to automate the detection and identifica- of CWT as shown in Eq. 5. CWT is applied to the mixed
tion of the potential location of the damage. In this paper, signal using two popular mother wavelet basis func-
an EMD-based TF method is developed to eliminate the tions (i.e., Morse and Morlet) to investigate and compare
presence of noise and extract the key AE components the resolution of extracted CWT spectrograms of both
of the measured data obtained from limited AE sensors. basis functions. Figure 1 (a) shows the Fourier spectrum
Continuous Wavelet Transform (CWT) is then used of the mixed-signal X(t), whereas Fig. 1 (b-c) represents
to extract the spectrograms of the key AE components. the spectrogram of X obtained from CWT using Morse
The resulting spectrograms form the training and testing and Morlet function as a mother wavelet, respectively.
database to validate the proposed 2D CNN, which can It can be seen that the signal contains three frequencies,
serve as a damage detection and classification tool with- which appear in the FFT and spectrogram. Moreover, it
out considering the standalone AE parameters and fea- can be observed in Fig. 1 (b) that the CWT spectrogram
ture extraction. The paper is outlined as follows. A brief of Morse wavelet provides a better resolution compared
background of CWT and 2D CNN is presented first, fol- with the spectrogram of Morlet wavelet, as shown in
lowed by the proposed method. A suite of numerical and Fig. 1 (c). Therefore, this study uses the Morse wavelet

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 4 of 24

Fig. 1 (a) Fourier spectrum of the mixed-signal and (b) and (c) its spectrogram obtained from the CWT using Morse and Morlet basis functions,
respectively

as the mother wavelet to extract the CWT spectrograms classification layers consisting of fully connected layers,
herein. softmax layers and classification output layers.
      The convolution layer is the first layer of feature extrac-
x1 = sin 2π f1 t , x2 = sin 2π f2 t , x3 = sin 2π f3 t tion from an input image using the convolution operation
(4) between an image and a filter (or kernel). A 2D convo-
X = x1 + x2 + x3 (5) lution layer applies to slide convolution filters to the
input. The layer convolves the input by moving the filters
along the input vertically and horizontally and comput-
Convolutional neural network ing the dot product of the weights and the input, and
A Convolutional Neural Network (CNN), as depicted in then adding a bias term. The convolution operation is an
Fig. 2, is a Deep Learning algorithm that assigns learnable element-wise product of images and filters, followed by
weights and biases to various features present in 1D and the summation of these values. The output of the convo-
2D data in order to differentiate them from one another. lution operation represents a low-dimensional invariant
The architecture of CNN is similar to that of the con- feature space of high-level feature data. The batch nor-
nectivity pattern of neurons in the human brain, and it malization layer normalizes each input channel across
was inspired by the organization of the visual cortex. The a mini-batch. In order to speed up the training of CNN
three major parameters, including local receptive fields, and reduce the network initialization sensitivity, batch
weight sharing and sub-sampling in the spatial domain, normalization layers are used between convolution layers
make CNN highly suitable for feature extraction from 2D and nonlinearities such as Rectified Linear Unit (ReLU)
data such as images. A typical CNN architecture com- layers. The layer first normalizes the activations of each
prises feature extraction layers, including convolution channel by subtracting the mini-batch mean and divid-
layers, batch-normalization layers, activation layers, and ing by the standard deviation of the mini-batch. Then, the

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 5 of 24

Fig. 2 An example of a typical CNN architecture

layer shifts the input by a learnable offset ‘β’ and scales it sigmoid or hyperbolic tangent function; however, ReLUs
by a learnable scale factor ‘γ’. are often used to introduce nonlinearity. A ReLU layer
Activation layers are implemented to verify the activa- performs a threshold operation to each element of the
tion of various elements of the feature space in the pres- input, where any value less than zero in this layer is set to
ence of specific features at a given spatial location. This zero. Mathematically, it can be expressed as follows:
is accomplished by introducing nonlinearity to the layers 
through the use of an activation function. Various acti- 0 for x < 0
f (x) =
x for x ≥ 0 (6)
vation functions have been used in literature, such as

Fig. 3 The proposed modified VGG-16 architecture used in the study

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 6 of 24

Fig. 4 Flowchart of the proposed method

The objective of a fully connected layer is to take the a large probability, but it will always remain between 0
results of the convolution process and use them to clas- and 1. A classification layer computes the cross-entropy
sify the images into labels. The output of a convolution loss for multiclass classification problems with mutu-
is flattened into a 1D feature vector of values, each rep- ally exclusive classes. This layer determines the number
resenting a probability of a certain feature belonging to of classes from the output size of the preceding layer.
a label and is fed into a fully connected layer such as a This layer takes input from the softmax function and
neural network. The fully connected layer multiplies the allocates each input to one of the mutually exclusive
input by a weight matrix and then adds a bias vector. In classes using the cross-entropy loss function.
this layer, softmax function is provided as an activation
function to determine the probability of feature clas-
sification for the data analyzed. The softmax function Proposed methodology
is a normalized exponential function that transforms The proposed methodology explores the capability of
a vector of K real values into a vector of K real values TF signal processing methods (e.g., EMD and CWT)
that sum to 1. If the input vector of K real numbers is to extract key AE components along with the feature-
applied to the softmax layer, it normalizes into a prob- free classification capability of a deep learning algo-
ability distribution comprising of K probabilities. The rithm in a unified manner. First, EMD is undertaken
input values can be positive, negative, zero, or greater to eliminate the presence of noise in the measured AE
than one, but the softmax function transforms them data and decompose it into a set of IMFs that repre-
into values between 0 and 1 so that they can be inter- sent the undamaged and damaged AE signals. The time
preted as probabilities. If one of the inputs is small or response of the AE signal is considered non-station-
negative, the softmax function turns it into a small ary, which can be decomposed using EMD due to its
probability, whereas if the input is large, it turns it into capability of analyzing nonlinear and non-stationary

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 7 of 24

Fig. 5 (a) CWT spectrogram of y(t), and (b) IMF-1, (c) IMF-2, and (d) IMF-3 of y(t) obtained from EMD

Fig. 6 Training stage: (a) performance and (b) loss evaluation of the proposed CNN architecture using CWT TF images

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 8 of 24

Fig. 7 Confusion matrices for (a) validation and (b) testing data of the sine signal

Fig. 8 (a) AE monitoring system and the experimental setup: (b) actual and (c) schematic

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 9 of 24

Fig. 9 Time-history of the measured AE data collected from sensors ­A1 and ­A2 of damage at D

Fig. 10 Spectrograms of IMFs using (a) undamaged, (b) damaged AE data obtained from A
­ 1, and (c) undamaged, (d) damaged AE data from ­A2 for
damage at D

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 10 of 24

Fig. 11 Training stage: (a) performance and (b) loss of the proposed method using CWT images of undamaged and damaged cases

Fig. 12 Comparison of clusters for training and testing datasets for wooden beam

signals [15]. More details such as the systematic steps n


and criteria of EMD and its performance can be found x(t) = Hi (t) + vn (t) (7)
elsewhere [7]. Once the IMFs belong to the crack are i=1
identified, their spectrograms can be obtained from the CWT is then applied to each IMF obtained from EMD
CWT method. Subsequently, the extracted spectro- as expressed in Eq. 8, where Hi(i = 1, 2, 3, ……,n) is the
grams are fed into the 2D CNN model to automate the IMF of the original signal and vn is a residue.
process of detection and localization of the damage.  ∞
Assume x(t) is a measured AE signal, which is decom- r t−m
cm (Hi ) = Hi (t)ψ ∗ ( )dt (8)
posed using EMD. For a given AE sensor, a number of −∞ r
IMFs are extracted from EMD using Eq. 7.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 11 of 24

Fig. 13 Confusion matrices (a-b) for validation and testing using damaged and undamaged data from A
­ 1 and (c-d) for validation and testing using
damaged and undamaged data from ­A2

The resulting CWT images of each IMF are then fed compared to the original network. Additionally, the num-
into 2D CNN to localize the damage. In this study, a ber of filters for each network block increases by a fac-
modified VGG-16 [30], as depicted in Fig. 3, is used to tor of 2, starting with 4 filters in the first network block
classify the CWT images obtained from AE data. The and ending with 32 filters in the final network block.
architecture is comprised of five network blocks that Zero padding is introduced in the convolutional layers
consist of a varying number of convolutional layers that to ensure that the input and output layer sizes remain
are finalized with a max-pooling layer. The first two net- constant regardless of the convolutional operation. Each
work blocks consist of two convolutional layers followed convolutional layer is followed by a Batch Normaliza-
by a max-pooling layer, while the three network blocks tion (BN) layer and ReLU to normalize and activate the
consist of three convolutional layers followed by a max- extracted features. The max-pooling layer has a kernel
pooling layer. The number of filters of each convolutional size of 2 × 2 with a stride value of 2 to reduce the inputted
layer from the traditional VGG-16 network is reduced size to half through the max-pooling operation. Moreo-
by a factor of 16 as the number of classifications and ver, the length of intermediary fully connected layers is
complexity of the features are dramatically reduced as reduced from 4096 to 512, with the fully connected layer

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 12 of 24

Fig. 14 Time history of AE data of a typical segment for damage at D

having a length equivalent to the number of classes used Numerical Illustration

in the study. Each fully connected layer is proceeded by a The mixed-signal shown in Eq. 5 is added with random
ReLU layer, and a dropout layer with a probability value noise to develop a suite of training databases required for
set to 0.2 was implemented to reduce overfitting during a 2D CNN model, as shown in Eq. 9:
the training process. The proposed method is illustrated
using the flowchart, as shown in Fig. 4.
y(t) = (x1 + x2 + x3 ) + ε (9)

Fig. 15 Training stage: (a) performance and (b) loss of the proposed method using the experimental data

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 13 of 24

Fig. 16 Confusion matrices for (a) validation and (b) testing data of the wooden beam

where ε is a noise component that is a time-series

Table 1 Performance indicators calculated from the sequence of a normal random variable. Then, EMD is
classification of CWT images of close ­(A1) and far damage (­ A2)
applied to y(t) to decompose the signal and extract the
Performance Indicator Validation Testing mono-component IMF. Finally, a spectrogram of each
IMF is obtained, which belongs to each frequency class.
Accuracy (%) 93.7 93.2
Figure 5a shows the spectrograms of the mixed-signal y
Recall 0.93 0.93
(t), and Fig. 5b-d shows the spectrograms of IMF-1, IMF-
Precision 0.95 0.94
2, and IMF-3 obtained from EMD, respectively. It can be
F1 Score 0.94 0.93
seen that EMD is able to decompose the mixed-signal
and extract the mono-component signal.
In this study, 1000 images of each of the three fre-
quency classes (i.e., a total of 3000 images) are used
in the training, validation and testing process. 70% of
the images are used for training, 20% are used for val-
idation, while the rest, 10%, are used for testing. This
resulted in 2100 images, 600 images and 300 images for
training, validation and testing, respectively. The train-
ing is conducted over 30 epochs using a Stochastic Gra-
dient Descent with Momentum (SGDM) solver with an
initial learning rate of 0.0005, minibatch size of 128 and
L2-regularization of 0.0005. The parameters of the CNN
were determined by conducting a grid search of the
parameters and comparing the training accuracy across
all iterations to determine which parameters resulted
in the most robust and accurate model. All numerical
simulations and CNN training and classifications are
conducted using a Lenovo ThinkStation with NVIDIA
Quadro GPU and 32 GB RAM. Figure 6 shows the vari-
ation of the accuracy and loss throughout the train-
ing and validation process. The network is trained for
Fig. 17 Experimental setup of the plate 30 epochs resulting in 63 minibatch iterations. It can

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 14 of 24

Fig. 18 Time history of the measured AE data collected using sensors ­A1 and ­A2 of damage at D

Fig. 19 Spectrograms of IMFs using (a) undamaged, (b) damaged AE data obtained from sensor A
­ 1, and (c) undamaged, and (d) damaged AE data
from sensor ­A2 for damage at D

be observed that due to the high performance of the able to achieve 100% classification accuracy for all fre-
network, the accuracy reaches 100% relatively quickly quency categories.
while minimizing the loss function. The trained net-
work is tested using 100 images of each frequency class. The selected AE data acquisition system
The confusion matrix is shown in Fig. 7 (a-b), repre- Figure 8(a) illustrates a typical setup of the proposed
senting the classification accuracy for the validation AE monitoring system used in the experimental study.
and testing datasets, respectively. The CNN model is The AE system is used to record AE signals from a

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 15 of 24

Fig. 20 Comparison of clusters for training and testing datasets of the wooden plate

Fig. 21 Training stage: (a) performance and (b) loss of the proposed method using CWT images of undamaged and damaged conditions of the
wooden plate

wooden beam and plate subjected to minor damage acquisition (DAQ) system to collect the AE data. It is
using AE sensors. Two AE sensors with a characteristic also attached with a direct current supply to power the
frequency band extending from 20–450 kHz are placed AE sensor. The DAQ is used to transfer the measured
on the specimen, as shown in Fig. 8(b). A pre-amplifi- AE signals to the computer. It has a sampling rate of
cation of the AE signals is added using preamplifiers 200 Msamples/s, which is capable of dealing with high-
with 34—40 dB gain and plug-in bandpass filters from frequency AE data collected by the sensors. In the pro-
2.5–2400 kHz to magnify the AE signal. A decoupling posed experimentation, the sampling frequency of the
box is utilized to connect the preamplifier with the data sensors is set to 20 kHz.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 16 of 24

Fig. 22 Confusion matrices (a-b) for validation and testing using damaged and undamaged data from A
­ 1 and (c-d) from ­A2

length, 6.5 cm width, and 2 cm thickness. Two AE sen-

Table 2 Performance indicators calculated from the sors ­(A1 and ­A2) are used to collect AE data from the
classification of CWT images of undamaged and damaged beam. The damage location is considered close to A ­ 1 to
conditions check the capability of the proposed method for local-
Performance indicator Validation Testing izing the damage (e.g. location D). Figure 8(b-c) shows
the location of sensors and damage on the beam. The
Accuracy (%) 100 97.5
damage is created by simulating a hole using a drill-
Recall 1.00 0.98
ing machine, while the data was collected using the AE
Precision 1.00 0.98
sensors.
F1 Score 1.00 0.98

Damage detection
In order to validate the performance of the proposed
Damage identification in a one‑dimensional method as a damage detection tool, the proposed
structural member method is applied to the AE data of the undamaged and
In order to evaluate the performance of the proposed damaged system. Figure 9 shows the time-history of
method, a wooden beam is considered in the experi- AE data collected from wooden beam using ­A1 and ­A2.
mental study. The beam has a dimension of 62 cm It can be seen that the damage progression begins near

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 17 of 24

Fig. 23 Training stage: (a) performance and (b) loss of the proposed method using the experimental data

Fig. 24 Confusion matrices for (a) validation and (b) testing data of the wooden plate for damage at D
­ 1

t = 5 s. Therefore, the AE data collected between 0–5 s

is considered as an undamaged state and data collected
after 5 s is considered as a damaged state. The AE sen-
Table 3 Performance indicators calculated from the
sors have produced big data due to the higher sampling
classification of CWT images of close and far damage
frequencies that can be used to create enough datasets to
Indicator Damage localization feed the CNN model. Therefore, AE data of undamaged
Validation Testing and damaged conditions is divided into a finite number
of segments. EMD is applied to each segment separately
Accuracy (%) 90.0 87.5 to extract the IMFs. Then, the CWT is used to generate
Precision 0.90 0.80 the spectrogram of each IMF obtained from EMD. The
Recall 0.90 0.95 spectrograms of IMFs have a size of 936 × 1920, which
F1 Score 0.90 0.86 are used as the training and testing data of the 2D CNN.

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 18 of 24

Fig. 25 (a) Experimental setup of the concrete beam and (b) AE monitoring system

Figure 10 shows a typical spectrogram of IMFs using


P
undamaged and damaged AE data collected from (a-b) PC = (13)
N
­A1 and (c-d) A­ 2, respectively.
In this study, 195 randomly selected images of undam- where xc is the centroid value along the width of the
aged and damaged conditions collected using ­A1 are used image in pixels, yc is the centroid value along with the
in the training process, as shown in Fig. 11. An additional height of the image in pixels, xm is the centroidal value
56 images and 28 images of each case are used in the vali- of the width at pixel m. yn is the centroidal value of the
dation and testing process. height at pixel ‘n’, P is the pixel intensity for the pixel
To ensure that the training, validation and testing located at the 2D location represented by the integers ‘m’
datasets were diverse, each dataset was represented by and ‘n’, and N is the number of pixels in the image. For
a 2D cluster and compared with each other to deter- example, at a pixel location of (m,n) = (1,1), the values of
mine the diversity of the datasets. To convert each 3D yn and xm would be considered 0.5 pixels, respectively, as
image matrix to a 2D point for clustering, the RGB we assumed each pixel is a unit pixel in width and height.
images were first converted to grayscale images by con- Once all the 3D points have been determined, the cen-
verting the pixel intensity using, troid of each cluster is determined by the average of all
values across each of the dependent variables represented
P = 0.2989PR + 0.5870PG + 0.1140PB (10) in Eqs. 10–12 as represented by Fig. 12.
From the figure above, it can be observed that the cen-
where P is the intensity of any pixel of the grayscale
image represented as an integer from 0 to 255 and PR, troids of both the training and testing datasets along the
PG and PB is the intensity of any pixel of the RGB image x and y-axis of the image are consistent with each other,
which suggests that the images belong to the same clus-
for the red, green and blue channel, respectively. Once
the images were converted, the grayscale image was ter. Though these images belong to the same cluster, the
converted to a 3D point by determining the centroid of spread of the images around the centroid point of the
the image. The centroid of the image can be considered datasets suggests that the datasets are sufficiently diverse
a point in 2D image space at which all pixel intensities and that though the testing dataset has similar features to
are balanced. To determine the width, height, and pixel the training dataset, they do not have significantly similar
intensity which represents the centroid, the following features that would result in overfitting.
equation can be used: The resulting confusion matrix is shown in Fig. 13(a-
b), displaying the classification accuracy for the valida-
tion and testing datasets extracted from AE data using

xm P
xC =  (11) ­A1. The same process is used to detect the damage using
P
AE data collected from A ­ 2. Figure 13(c-d) shows the con-
 fusion matrices that display the classification accuracy
yn P
yC =  (12) for the validation and testing datasets obtained from AE
P data collected using ­A2. It can be observed that the CNN

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 19 of 24

model is able to achieve ~ 100% classification accuracy structural elements using a limited number of AE sen-
for damaged and undamaged categories. Therefore, the sors (e.g., two sensors). The damage location is identi-
proposed method can be considered a reliable tool for fied by using the term damage-close or far with respect
damage detection in structural elements. to the sensor location. In order to identify the potential
location of damage at D, spectrograms obtained from
Damage localization ­A1 are treated as damage-close, whereas spectrograms
Figure 14 shows the time-history of measured AE data from ­A2 are treated as damage-far.
of a typical segment for damage at D. It can be observed In this study, 473 randomly selected images of each
that the amplitude of undamaged and damaged signals sensor class are used in the training process. An addi-
significantly overlap. Subsequently, the raw amplitude tional 135 images and 67 images of each sensor class are
of the time-history of AE data will provide faulty infor- used in the validation and testing process. The training
mation for the damage location, which motivates the process is shown in Fig. 15. The confusion matrix shown
need for the proposed method to automate the process in Fig. 16 (a-b) shows the classification accuracy for the
of identifying the potential location of the damage. The validation and testing datasets. The accuracy, recall, pre-
spectrogram of IMFs of each AE segment is extracted cision and F1 scores of the validation dataset and test-
from CWT. These spectrograms are then used to feed ing dataset are summarized in Table 1. The performance
the CNN model to identify the location of damage in indicators calculated from the confusion matrices sug-
the structural element. The proposed method is applied gest an excellent generalization of the CWT images by
to identify the approximate location of damage in the proposed algorithm.

Fig. 26 Time history of the measured AE waveforms and first five IMFs extracted from sensors A1 and A2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 20 of 24

Fig. 27 Spectrograms of IMFs using AE waveform obtained from sensor ­A1 and sensor A
­2

Damage identification in a two‑dimensional The spectrograms are used as the training and testing
structural member data of the 2D CNN. Figure 19 shows a typical CWT
In this section, an experimental study using a 2D struc- spectrogram of IMFs using undamaged and damaged AE
tural element such as a wooden plate is undertaken to data collected from (a-b) ­A1 and (c-d) ­A2.
validate the performance of the proposed method. The In order to verify the capability of the 2D CNN model
dimensions of the plate are 1 m length, 1 m width, and for classifying the undamaged and damaged state of the
2 cm thickness, respectively, as shown in Fig. 17. Two wooden plate, 140 images of undamaged and damaged
sensors ­(A1 and ­A2) are used to measure the AE signal conditions collected using ­A1 are used in the training
while the plate is subjected to active damage. The plate process. An additional 40 images and 20 images of each
is subjected to damage at location D. Figure 17 shows the case are used in the validation and testing process. As
proposed location of sensors and damage on the plate. A per the process described in Sect. 5.1, the centroid of
drilling machine is used to simulate the damage by creat- each image was determined the clusters for the training
ing a hole, while the signal is measured using the selected and testing datasets were compared as shown in Fig. 20.
AE sensors. Similarly, though the centroids of both the training and
testing images across the entire dataset are similar, the
spread of the points around the centroid is relatively high,
Damage detection suggesting there is good robustness within the dataset.
In this study, the proposed method is applied to undam- Furthermore, though both the beam and plate are made
aged and damaged AE signals to classify the undamaged of similar material and are damaged in a similar manner,
and damaged cases of the wooden plate. Figure 18 rep- each damage scenario generates unique clusters, sug-
resents the time-history of AE data collected from the gesting geometric properties of the element and relative
wooden plate using ­A1 and A ­ 2. It can be observed that damage location have significant influence with regards
the damage progression begins near t = 1 s and ends at to the AE signals generated.
t = 19 s. Therefore, the AE data collected between 0–1 s Figure 21 shows the performance of the proposed
is considered as undamaged and between 1–19 s is con- method for identifying the undamaged and damaged
sidered as damaged data. Then, the CWT is used to gen- states of the plate. The resulting confusion matrix is
erate the spectrogram of each IMF obtained from EMD. shown in Fig. 22 (a-b), displaying the classification

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 21 of 24

Fig. 28 Comparison of clusters for the training and testing datasets of the concrete beam

accuracy for the validation and testing datasets extracted excellent generalization of the CWT images by the pro-
from AE data using ­A1. The same process is executed to posed algorithm.
detect the damage using AE data collected from ­A2. Fig-
ure 22 (c-d) shows the confusion matrices that display Crack identification in a concrete beam
the classification accuracy for the validation and test- In this section, the performance of the proposed method
ing datasets obtained from AE data collected using ­A2. is validated experimentally using a concrete beam. The
Table 2 summarizes the accuracy, recall, precision and F1 beam size is 355 mm in length and 78 mm in width and
scores of the validation dataset and testing dataset of the thickness. Ordinary Portland cement, fine aggregate, and
undamaged and damaged condition of the plate. It can be coarse aggregate are used in the concrete mixture. MTS
observed that the CNN model is able to achieve ~ 100% testing machine is used to carry out a three-point flex-
classification accuracy for damaged and undamaged ural loading test. Two AE sensors (­A1 and A ­ 2) are used
cases. Therefore, the proposed method shows excellent to collect AE data from the beam. During the experi-
capabilities to perform damage detection in a structural ment, the crack opening is visually monitored, where
element such as a plate. the crack initiates at the center bottom of the beam and
then propagates to the top. Figure 25 (a) shows the loca-
Damage localization tion of sensors and crack on the concrete beam. AE sig-
In this section, the proposed method is extended to nal is collected by a Micro-SHM acquisition system that
localize the damage using limited AE sensors. In order has four input channels. The Micro-SHM acquisition
to identify the potential location of damage, the spec- system is used to transfer the measured AE waveforms
trograms of IMFs of AE signal obtained from ­A1 and ­A2 from sensors to the computer using A ­ Ewin software, as
(defined as damage-close and damage-far, respectively) shown in Fig. 25 (b). The system has a sampling rate of
are used to feed into the 2D CNN. In this study, 140 10 Msamples/s, which can deal with high-frequency AE
randomly selected images of each sensor class are used signals collected by the sensors due to the micro-cracks
in the training process. An additional 40 images and in concrete material.
20 images of each sensor class are used in the valida-
tion and testing process. The training process is shown
Crack localization
in Fig. 23. The confusion matrix shown in Fig. 24(a-b)
In this study, the AE waveforms collected from the beam
displays the classification accuracy for the validation
using two AE sensors are firstly decomposed to n num-
and testing datasets. The accuracy, recall, precision and
ber mono-components signals (IMFs) using the EMD
F1 scores of the validation dataset and testing dataset
method. Figure 26 represents the time-history response
are summarized in Table 3. The performance indica-
of a typical AE waveform and its first five IMFs obtained
tors calculated from the confusion matrices suggest an
from EMD using sensors A ­ 1 (first column) and A ­2

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 22 of 24

Fig. 29 Training stage: (a) performance and (b) loss of the proposed method using the experimental data

Fig. 30 Confusion matrices for (a) validation and (b) testing data of the concrete beam under three-point flexural loading

Table 4 Performance indicators calculated from the (second column). It can be seen that the amplitude of the
classification of CWT images of close and far damage AE waveform obtained from A1 and A2 are the same.
Indicator Damage Localization Thus, the amplitude of the time-history of raw AE data
might provide inaccurate information for the crack loca-
Validation Testing
tion, which motivates to apply the proposed method to
Accuracy (%) 95.0 95.0 automate the process of identifying the potential loca-
Precision 0.95 0.97 tion of the crack in the concrete beam. In order to predict
Recall 0.95 0.94 the approximate location of the crack, the CWT spec-
F1 Score 0.95 0.95 trograms of IMFs extracted from A ­ 1 and ­A2 (defined as

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 23 of 24

damage-far and damage-close, respectively) are used to Authors’ contributions

MB conceived the proposed study, conducted a detailed literature review for
feed into the 2D CNN. Figure 27 shows the CWT spec- the paper and completed the experimental work. MB and KD participated in
trograms of the first five IMFs of typical AE waveform the research design, analyzed, and interpreted the data of the experimental
obtained from A ­ 1 (considered as crack-far) and A
­ 2 (con- and full-scale study. MB drafted the manuscript, KD contributed to the write-
up, and AS provided a detailed review of the manuscript. AS supervised the
sidered as crack-close). entire research. All authors approved the manuscript for publication in the
In this study, 210 randomly selected images of each journal.
sensor class are used in the training process. An addi-
Funding
tional 60 images and 30 images of each sensor class are The authors would like to thank the University of Tripoli for funding the
used in the validation and testing process. The centroi- research through the Libyan Ministry of Education. Also, the authors thank
dal clustering, as demonstrated in Fig. 28, demonstrates Mitacs for providing financial support for this research through the Mitacs
Research Training Award program.
that similar to wooden beam and plate, the concrete
beam has a robust dataset that has unique clustering Availability of data and materials
characteristics that are indicative of the uniqueness of The datasets used in this study are available upon request.
the damage caused by the three-point bending test.
The training process is shown in Fig. 29. The confu- Declarations
sion matrix shown in Fig. 30 (a-b) displays the classifica- Ethics approval and consent to participate
tion accuracy for the validation and testing datasets. The Not applicable.
accuracy, recall, precision and F1 scores of the validation
Consent for publication
dataset and testing dataset are summarized in Table 4. All authors have consented to the publication of the manuscript.
The performance indicators calculated from the confu-
sion matrices suggest an excellent generalization of the Competing interests
The authors declare that they have no competing interests.
CWT images by the proposed algorithm.
Received: 28 January 2022 Accepted: 3 March 2022

Conclusions
In this paper, a 2D CNN framework is proposed to auto-
mate the process of detecting and localizing the damage
References
using AE data collected from structures. EMD is used to 1. Aggelis DG, Shiotani T, Terazawa M (2010) Assessment of construction
suppress the presence of noise and extract the key AE com- joint effect in full-scale concrete beams by acoustic emission activity. J
Eng Mech 136(7):906–912
ponents, followed by CWT to generate the spectrograms
2. Anay R, Cortez TM, Jáuregui DV, El Batanouny MK, Ziehl P (2016) On-site
of the extracted IMFs. Then, the resulting noise-free spec- acoustic-emission monitoring for assessment of a Prestressed con-
trograms are utilized to feed the CNN model for improved crete double-tee-Beam bridge without plans. J Perform Constr Facil
30(4):04015062
classification accuracy. CNN model is employed to auto-
3. Abouhussien AA, Hassan AA (2016) Acoustic emission-based analysis of
mate the process of identifying the potential location of bond behavior of corroded reinforcement in existing concrete structures.
damage using the AE data. A set of numerical and experi- Struct Control Health Monit 24(3):e1893
4. Abdulazeez AM, Zeebaree DQ, Zebari DA, Zebari GM, Adeen IMN (2020)
mental studies are conducted to validate the performance
The applications of discrete wavelet transform in image processing: a
of the proposed approach as a damage detection tool using review. J Soft Comput Data Mining 1(2):31–43
a limited number of AE sensors. 5. Benavent-Climent A, Gallego A, Vico JM (2011) An acoustic emission
energy index for damage evaluation of reinforced concrete slabs under
The results show the capability of the proposed approach
seismic loads. Struct Health Monit 11(1):69–81
to identify the approximate location of the damage in 6. Bahar O, Ramezani S (2012) Enhanced Hilbert Huang transform and
various structural elements. For example, the extracted its applications to modal identification. Struct Des Tall Special Build
23(4):239–253
results obtained from the wooden beam showed the high
7. Barbosh M, Singh P, Sadhu A (2020) Empirical mode decomposition and
performance of the proposed approach in identifying the its variants: A review with applications in structural health monitoring.
potential location of damage with 93% accuracy. Also, the Smart Mater Struct 29(9):093001
8. Barbosh M, Sadhu A, Sankar G (2021) Time–frequency decomposition-
proposed technique can identify the approximate loca-
assisted improved localization of proximity of damage using acoustic
tion of damage in a 2D wooden plate with an accuracy of sensors. Smart Mater Struct 30(2):025021
88%. Moreover, the results obtained from the experimen- 9. Carpinteri A, Lacidogna G, Niccolini G (2010) Damage analysis of rein-
forced concrete buildings by the acoustic emission technique. Struct
tal study using concrete beam represented the capability
Control Health Monit 18(6):660–673
of this approach in identifying the crack location where 10. Calabrese L, Proverbio E (2020) A review on the applications of acoustic
the accuracy reached 95%. Future studies are reserved for emission technique in the study of stress corrosion cracking. Corrosion
Mater Degrad 2(1):1–33
validating the proposed method in a wide range of full-
11. Dunphy K, Sadhu A (2022) Autonomous crack detection approach for
scale structures and identifying the location of multiple masonry structures using artificial intelligence. Recent Developments
damages. in Structural Health Monitoring and Assessment–Opportunities and

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Barbosh et al. J Infrastruct Preserv Resil (2022) 3:6 Page 24 of 24

Challenges: Bridges, Buildings and Other Infrastructures, World Scientific. 34. Sony S, Sadhu A (2020) Synchrosqueezing transform-based identification
pp 253–283 of time-varying structural systems using multi-sensor data. J Sound Vibrat
12. Ebrahimkhanlou A, Choi J, Hrynyk TD, Salamone S, Bayrak O (2020) 486:115576
Acoustic emission monitoring of containment structures during post- 35. Sony S, Dunphy K, Sadhu A, Capretz M (2021) A systematic review of
tensioning. Eng Struct 209:109930 convolutional neural network-based structural condition assessment
13. Gilles J (2013) Empirical Wavelet Transform. IEEE Trans Signal Process techniques. Engineering Structures, Elsevier 226:111347
61(16):3999–4010 36. Sun G, Gao Y, Lin K, Hu Y (2019) Fine-grained fault diagnosis method of
14. Hsueh Y-M, Ittangihal V, Wu W-B, Chang H-C, Kuo C-C (2019) Fault rolling bearing combining multisynchrosqueezing transform and sparse
Diagnosis System for Induction Motors by CNN Using Empirical Wavelet feature coding based on dictionary learning. Shock Vibrat 2019:1531079
Transform. Symmetry 11(10):1212 37. Tang Z, Chen Z, Bao Y, Li H (2018) Convolutional neural network-based
15. Huang NE, Shen Z, Long SR, Wu MC, Shih HH, Zheng Q, Yen NC, Tung data anomaly detection method using multiple information for structural
CC, Liu HH (1998) The empirical mode decomposition and the Hilbert health monitoring. Struct Control Health Monit 26(1)
spectrum for nonlinear and non-stationary time series analysis. Proc R 38. Verstraete D, Ferrada A, Droguett EL, Meruane V, Modarres M (2017) Deep
Soc Math Phys Eng Sci 454(1971):903–995 learning enabled fault diagnosis using time-frequency image analysis of
16. Huang NE, Shen Z, Long SR, Wu MC, Shih HH, Zheng Q, Liu HH (2002) The rolling element bearings. Shock Vib 2017:1–17
empirical mode decomposition and Hilbert spectrum for nonlinear and 39. Verstrynge E, Lacidogna G, Accornero F, Tomor A (2021) A review on
non-stationary time series analysis. Proc R Soc A 454:903–995 acoustic emission monitoring for damage detection in masonry struc-
17. Kumar A, Zhou Y, Gandhi CP, Kumar R, Xiang J (2020) Bearing defect size tures. Construct Build Mater 268:121089
assessment using wavelet transform based Deep Convolutional Neural 40. Wang J, Mo Z, Zhang H, Miao Q (2019) A Deep Learning Method for
Network (DCNN). Alex Eng J 59:999–1012 Bearing Fault Diagnosis Based on Time-Frequency Image. IEEE Access
18. Kralovec C, Schagerl M (2020) Review of structural health monitoring 7:42373–42383
methods regarding a multi-sensor approach for damage assessment of 41. Worley R, Dewoolkar MM, Xia T, Farrell R, Orfeo D, Burns D, Huston DR
metal and composite structures. Sensors 20(3):826 (2019) Acoustic emission sensing for crack monitoring in prefabri-
19. Li Z, Park HS, Adeli H (2016) New method for modal identification of cated and Prestressed reinforced concrete bridge girders. J Bridg Eng
super high-rise building structures using discretized synchrosqueezed 24(4):04019018
wavelet and Hilbert transforms. Struct Des Tall Spec Build 26(3):e1312 42. Wang Z, Ding K, Ren H, Ning J (2021). “Quantitative acoustic emission
20. Liu Q, Huang C (2019) A Fault Diagnosis Method Based on Transfer Con- investigation on the crack evolution in concrete prisms by frequency
volutional Neural Networks. IEEE Access 7:171423–171430 analysis based on wavelet packet transform.”, Structural Health Monitor-
21. Li D, Wang Y, Yan WJ, Ren WX (2020) Acoustic emission wave classification ing, 147592172110188.
for rail crack monitoring based on synchrosqueezed wavelet transform 43. Yuan M, Sadhu A, Liu K (2017) Condition assessment of structure with
and multi-branch convolutional neural network. Struct Health Monit tuned mass damper using empirical wavelet transform. J Vib Control
20(4):1563–1582 24(20):4850–4867
22. Manthei G, Plenkers K (2018) Review on in situ acoustic emission moni- 44. Zhang Y, Xing K, Bai R, Sun D, Meng Z (2020) An enhanced convolutional
toring in the context of structural health monitoring in mines. Appl Sci neural network for bearing fault diagnosis based on time-frequency
8(9):1595 image. Measurement 157:107667
23. Ma G, Du Q (2020) Structural health evaluation of the prestressed con-
crete using advanced acoustic emission (AE) parameters. Construction
and Building Materials 250:118860 Publisher’s Note
24. Perez-Ramirez CA, Amezquita-Sanchez JP, Adeli H, Valtierra-Rodriguez M, Springer Nature remains neutral with regard to jurisdictional claims in pub-
Camarena-Martinez D, Romero-Troncoso RJ (2016) New methodology for lished maps and institutional affiliations.
modal parameters identification of smart civil structures using ambient
vibrations and synchrosqueezed wavelet transform. Eng Appl Artif Intell
48:1–12
25. Pandhare V, Singh J, Lee J (2019) “Convolutional Neural Network Based
Rolling-Element Bearing Fault Diagnosis for Naturally Occurring and
Progressing Defects Using Time-Frequency Domain Features,” Prognostics
and System Health Management Conference (PHM-Paris). France, Paris, pp
320–326
26. Quy TB, Kim J (2021) Crack detection and localization in a fluid pipeline
based on acoustic emission signals. Mechanical Systems and Signal
Processing 150:107254
27. Sarfarazi MP (1992) Acoustic emissions and damage constitutive charac-
teristics of paper. Institute of paper science and technology
28. Sadhu, A. (2013), “Decentralized ambient modal identification of struc-
tures.”, PhD Thesis, Department of Civil and Environmental Engineering,
University of Waterloo.
29. Sadhu A, Sony S, Friesen P (2019) Evaluation of progressive damage in
structures using tensor decomposition-based wavelet analysis. J Vib
Control 25(19–20):2595–2610
30. Simonyan K, Zisserman A (2014) Very deep convolutional networks for
large-scale image recognition. ICLR Conference 1409:1556
31. Singh P, Keyvanlou M, Sadhu A (2021) An improved time-varying empiri-
cal mode decomposition for structural condition assessment using
limited sensors. Eng Struct 232:111882
32. Saeedifar M, Zarouchas D (2020) Damage characterization of lami-
nated composites using acoustic emission: A review. Composit B Eng
195:108039
33. Shao S, Yan R, Lu Y, Wang P, Gao RX (2020) DCNN-based multi-signal
induction motor fault diagnosis. IEEE Trans Instrum Meas 69(6):2658–2669

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

[email protected]