See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/376047364

Deep Learning and Machine learning Techniques in Advanced Non-

Destructive Testing

Article in Harbin Gongcheng Daxue Xuebao/Journal of Harbin Engineering University · November 2023

CITATIONS READS

2 1,114

1 author:

Samira Gholizadeh
University of Cape Town
16 PUBLICATIONS 870 CITATIONS

SEE PROFILE

All content following this page was uploaded by Samira Gholizadeh on 06 December 2023.

The user has requested enhancement of the downloaded file.

Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

Deep Learning and Machine learning Techniques in

Advanced Non-Destructive Testing

S.Gholizadeh
Blast Impact and Survivability Research Unit (BISRU), Department of Mechanical Engineering, University of
Cape Town, Rondebosch, Cape Town, 7701, South Africa, Email: ghlsam002@myuct.ac.za

Abstract— There has been considerable evidence of the effectiveness of machine learning (ML) and artificial
intelligence (AI) algorithms as numerical tools; some of their applications include computational fluid dynamics,
control and automation, signal processing, and materials engineering. Recent advances in ML and AI can directly
benefit non-destructive testing, one of the most important industrial applications. With AI, data analysis can be
improved and better harnessed. NDT uses electromagnetic waves or material-based methods to acquire information
from a specimen. ML algorithms can interpret multiple signals or images to analyze, inspect, and examine the
integrity of the materials structure. An overview of the foundations and current applications of ML techniques in
advanced NDT is presented in this paper. In addition to explaining ML techniques, the most recent advances in ML
and AI, including deep learning, are also discussed.
Index Terms— Deep learning, Machine learning, Non-destructive testing, NDT

otherwise altering the material [1, 2]. In general NDT

uses electromagnetic waves or material-based methods
I. INTRODUCTION
to acquire information from a specimen. These methods
Non-destructive testing plays an essential role in can be categorized based on the factors that are
making sure that costs of all operations are highly evaluated. Some significant recent inspections with
minimized. NDTs ensure that there is protection of non-destructive testing methods have been summarized
mechanisms used in the entire operation such as and listed in Table 1.
railways, power generations, and petrochemical The era of big data has already begun. Existing IT
processes. In manufacturing industry, for instance, they and computing technologies and algorithms are
are used to detect flaws, such as cracks, gaps, leaks, challenged by big data. A traditional definition of big
pores, and fractures which are mostly caused by stress, data is an enormous amount of data (in excess of 1
contraction, processing, and hammering on the surface terabyte) produced at high velocity with high variety
or inside of materials, while maintaining durability and and veracity [22]. In data generation, velocity indicates
originality after inspection. In addition, non-destructive how fast data is generated, and the type of collected data
tests use physical principles to identify and evaluate is known as variety. The volume of the data can range
faults or destructive defects. The field of NDT or NDE from several petabytes (1015 bytes) to several exabytes
is cost effective methods of testing materials and are (1016 bytes). This data deluges are driving the need for
being used to identify and characterize damages on the automated data analysis, which ML offers.
surface and inside of materials without cutting apart or

977
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

Table 1: Category of NDT methods based on the between datasets. In particular, it is a method of
detecting factors [2, 3] automatically detecting patterns in data and then
Inspection Type NDT Method
predicting future data, or making other decisions under
Residual stresses in Ultrasonic Testing [4,
materials 5] uncertain conditions based on the uncovered patterns
Aircraft composites Shearography [6] and [23]. There are many ways in which uncertainty can
assessment Thermographic
present itself in ML, such as determining the best model
Health monitoring of Testing [7, 8],
aerospace composite Vibration Methods [7, for explaining some data; or based on past data, what is
structures 9] the best prediction for the future. It is common for NDT
Ultrasonic Testing, specialists to make mistakes during interpretation of
Renewable energy and Thermography,
industrial assets Acoustic emission signals and analysis of large amounts of data, resulting
[10] in defects being missed, and the sizing of the defect
Rebound Hammer being incorrect, or the defect not being identified. In
Evaluation of compressive (RH) and Ultrasonic
any NDT technique, consistency and reliability of the
strength in the structure Pulse Velocity (UPV)
[11, 12] results are strongly influenced by the personnel
Structural Health Ultrasonic Testing [13, conducting the inspections and evaluating the results.
Monitoring (SHM) 14]
With AI, data analysis can be improved and better
Optical (laser, LED) ,
Penetrant inspection, harnessed. By addressing the challenges posed by large
Ultrasonic testing , data analysis and processing, this innovation will
Acoustic emission reduce the amount of time that advanced NDT
Radiographic
Additive manufacturing methodologies must invest in a single analysis. The role
techniques
Electromagnetic of NDT becomes ever more important as industrial
techniques work and discovery increases and materials are used
Thermographic
beyond their physical limits. Thus, testing methods
techniques [15, 16]
Auto-detection of impact themselves are evolving [24].
damage in carbon fiber Thermographic
composites Testing [17, 18]
Characterizing damage in Radiography [19] ML ALGORITHMS CATEGORIES
CFRP structures
Impact damage in Each ML algorithm develops a model that is
Infrared
glass/epoxy with equivalent to a set of experiences. The following five
Thermography [20]
manufacturing defects categories can be used to categorize ML algorithms:
Multiple Cracks Detection Neutron Radiography
[21] [21]
a. Supervised Learning Algorithms: Labeled data is
ML is a subset of AI and computer science based on a used to train the algorithm. Supervised learning
set of statistical tools that focus on developing models algorithms are used when both predictors (inputs X)
with data learning and training that can predict new data and outcomes (outputs Y) are present in the training
in the future. In general, ML is a set of related data set.
algorithms and processes that create relationships

978
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

b. Unsupervised Learning Algorithms: A training process of determining an ideal behavior within a

algorithm that uses unlabeled data. Unsupervised specific context, thereby maximizing its efficiency.
learning algorithms are used when training data A sequential decision-making scenario can also be
contain only predictor variables (X) and without modeled using these algorithms.
outcome variables (Y). e. Evolutionary Learning Algorithms: Algorithms
c. Semi-Supervised Learning Algorithms: In terms of that mimic the learning process of humans and
learning method, semi-supervised learning is animals. Prescriptive analytics is their most
somewhere between supervised learning and common application. Evolutionary algorithms
unsupervised learning. A dataset may either have (EAs) are a subset of evolutionary computations,
no labels for all observations in unsupervised which are generic metaheuristic optimization
learning or may have labels for all observations in algorithms based on populations [29]. In EA,
supervised Learning. Since labeling requires skilled mutation, recombination, reproduction, and
human experts, the cost to label is quite high in selection are some of the mechanisms inspired by
many practical situations. Therefore, biological evolution. Genetic algorithms and
semi-supervised algorithms are the best candidates colony optimization are examples of these
for model building when labels are absent from techniques [30].
most observations but present in a small part.
In NDT, supervised learning algorithms are the f. Here is an overview of some of the types of
go-to methods for prediction, relying on labeled algorithms and techniques used in ML (Figure 1).
data to classify new datasets [25], while
unsupervised learning is more suitable for detecting
anomalies, dealing with unlabeled data for pattern
recognition [26]. NDT can utilize supervised
learning to detect severity and types of damage
through input and output [27], and unsupervised
learning to detect damage via clustering of
structural response data. Combining both of these
approaches leads to semi-supervised learning,
aiming to combine labeled and unlabeled data for
data classification [28].
d. Reinforcement Learning Algorithms: It is
common for the input variable X as well as the
output variable Y to be uncertain (for instance,
predictive keyboards or spell check). Through an
iterative learning process, reinforcement learning
algorithms (called agents) continuously learn from
their environment. The method automates the

979
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

Machine Learning

Figure 1: Common types of algorithms and techniques used in ML

maintenance [36].
I. DEEP LEARNING (DL)
ML and DL techniques have been highly successful
Deep learning is a subfield of ML that focuses on in extracting features, providing better performance
training and utilizing artificial neural networks with than many other techniques. Therefore, they are highly
multiple layers to learn and extract intricate patterns valued in many scientific fields. For example, Artificial
and representations from complex data. It is inspired by Neural Networks (ANNs) were employed to analyze
the structure and functioning of the human brain's fiber orientation in composite materials using neurons
neural networks. Deep learning models, also known as as connecting elements [37], a single hidden-layer
deep neural networks, are designed to automatically feedforward neural network (SLFN) was utilized for
learn hierarchical representations of data [31]. DL has defect classification [38], convolutional neural network
achieved remarkable success in various domains, (CNN) models were implemented for inner damage
including computer vision, natural language analysis and evaluation [39, 40], Support Vector
processing, speech recognition, and recommender Machine (SVM) automatic classification model was
systems. Its ability to automatically learn features and utilized to monitor microcracks using multi-sensor
representations from raw data, without the need for measurements [41], and k-means clustering method
manual feature engineering, has revolutionized many was used for automatic defect detection and classify
fields and led to breakthroughs in tasks such as image signal sources [42].
classification, object detection, machine translation, Other ML algorithms have been applied in various
and more. DL has various applications in NDT, offering fields, such as object recognition [43], speech
improved accuracy and efficiency in defect detection recognition [44], Probabilistic neural network (PNN)
[32, 33], classification [34, 35], and predictive [45], Least square support vector machines (LSSVM)

980
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

[46]. predictions or classifications based on what they have

learned. DL enables the construction of highly flexible
II. THE APPLICATION OF ML AND DL IN NDT
and powerful models that can learn complex patterns
ML and DL in NDT refers to the application of ML and make accurate predictions from large-scale,
and DL techniques and algorithms to analyze NDT high-dimensional data [47]. DL models are composed
data, extract patterns, and make predictions or of multiple layers of interconnected artificial neurons,
classifications related to the integrity, quality, and known as nodes or units. Each layer receives inputs
performance of materials, components, or structures. from the previous layer, performs computations, and
ML and DL algorithms can learn from historical data passes the outputs to the next layer. The layers closer to
and use that knowledge to improve defect detection, the input are responsible for learning low-level features,
classification, predictive maintenance, and while deeper layers learn high-level representations and
decision-making processes in NDT. In the context of abstractions. ML and DL have found several
NDT, ML algorithms are trained on labeled datasets applications and offers several advantages in the field
that consist of NDT data, such as images, waveforms, of NDT:
or sensor readings, along with corresponding
annotations or outcomes. The ML models learn to A. Automated Defect Detection and Defects

recognize patterns, relationships, and characteristics in Classification

the data through a process called training. Once trained, ML algorithms can be trained to automatically detect

the ML models can analyze new, unseen data and make and recognize defects in NDT data, such as images or

ML techniques can also classify and categorize effective generalization capability with a training
different types of defects or flaw patterns based on accuracy of 96.41% and testing accuracy 94.35%.
training data. This helps in accurately identifying and Risheh et al. [61] used K-means clustering algorithm
characterizing defects, improving the reliability of NDT for defect detection in thermal images of industrial
inspections. For example, ML can distinguish between materials, while Sophian et al. [62] utilized the same
different types of corrosion or classify different types of algorithms for defect detection using pulse eddy
surface cracks [54-58]. Such classification enables currents and Shrifan et al [63] used it in microwave
better understanding of defect types and facilitates NDT to detect defect.
appropriate remedial actions. For instance, K-means Similarly, in ultrasonic testing, ML can analyze
clustering was applied by Du et al. [59] to classify waveforms and identify specific flaw patterns [64, 65].
acoustic emission signals obtained from the analysis of Fei et al. [66] and Vaclav et al. [67] have demonstrated
stress-corrosion-cracking (SCC) on 304 stainless steel. the effectiveness of Support-Vector Machines (SVMs)
The authors conducted experiments with three clusters in classifying ultrasonic signals. Both shallow neural
(k = 3) that represented the types of damages networks and SVMs have achieved a high degree of
encountered in steel (crack propagation, pitting, and accuracy for A-scan. However, due to the necessity of
bubble break-up). Saenkhum et al. [60] classified feature engineering, these methods are not suitable for
corrosion using acoustic emission and Artificial Neural broader applications. Niu et al. [68] proposed the use of
Networks (ANNs). Testing phase of neural networks Surface Defect-Generation Adversarial Network
indicated very little misclassification rate and a very (SDGAN) for generating defect images to enhance

981
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

deep-learning-based surface defect recognition, improves the classification accuracy of deep learning
addressing the limitations of insufficient and costly classifiers. A typical defect detection study and test
labeled defect data. The SDGAN is applied to expand method of comparative analysis is shown in Table 2.
the commutator cylinder surface defect image data sets,
B. Predictive Maintenance (PdM), Condition and
both with labeled data (CCSD-L) and without labeled
Real-time Monitoring
data (CCSD-NL). For anomaly recognition, an error
ML models can analyze historical NDT data, sensor
rate of 1.77% and a relative improvement (IMP) of
readings, and other relevant parameters to predict the
49.43% for the CCSD-NL defect data set was reported.
likelihood of future defects or failures, or the remaining
In terms of defect classification, an error rate of 0.74%
useful life of components or structures. By considering
and an IMP of 57.47% for the CCSD-L defect data set
patterns and trends in the data, ML algorithms can
were achieved. Munir et al. [69] utilized deep neural
estimate when and how defects or failures might occur.
networks (autoencoder) to mechanize the ultrasonic
This enables proactive maintenance planning and
weldment defect classification system to identify and
minimizes the risk of unexpected equipment or
remove the noise from the counterbore, planer, and
component failures, scheduling repairs or replacements
volumetric weldment defect signals. Results revealed
before failures happen, and optimizing maintenance
that the autoencoder was successful in reducing noise
resources.
from ultrasonic weldment defect signals which in turn

Table 1: Defect detection, test methods and analysis

Defect detections DL defect detection
Application Precision Precision
technique technique
Optical Quality
Manufacturing Deep convolutional
Control (OQC) and 98.2% 99% [70]
and industries neural network
machine vision
Visual Inspection Artificial Neural
Steel materials 87% 98% [71]
System Networks (ANN)
Ultrasonic wave Deep Learning
Concrete 90% 96% [72]
propagation (AlexNet, ResNet)
Automatic optical Deep convolutional
95% 98% [73]
detection neural network
Granular crystal
Convolutional
Composite sensor - highly 90% -
89.3% neural network
materials nonlinear solitary 92% [74]
(CNN)
waves (HNSWs)
Convolutional
- - neural network 92% [75]
(CNN)
90% and
Deep convolutional
Rail - - 95% [76]
neural network
92% [77]
Deep convolutional
Road - - 87% [78]
neural network

982
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

In predictive maintenance, diagnostics and algorithms [87] have been extensively used in
prognostics play an important role. Diagnostics and data-driven PdM in industrial manufacturing. A
prognostics fall into three categories: i) the reliability detailed description of the implementation process for
data-driven approach that consist of determining the the data-driven PdM can be found in [88], which
expected lifetime of the average component under follows the design methodology. In general, there are
historically average conditions [79], ii) Stress-driven four stages in the process: operational assessment, data
models that estimate the lifespan of components in a acquisition, feature engineering, and modeling (Figure
particular environments [80], iii) In Condition-based, 3).
the component's lifetime is determined based on its
operating environment [81, 82]. Figure 2 demonstrates
a broad spectrum of diagnostic and prognostic
techniques, as well as different steps in data driven
diagnostics and prognostics.

Figure 2: Data driven PdM implementation process

[89]
Figure 1: Data driven diagnostics & prognostics:
broad categories & steps [81] Hybrid models: Hybrid approaches combine the two
approaches outlined above [83]. The development of
The numerous research works on PdM can be accurate PdM models in complex systems is now
classified into three approaches [83]: possible through data-driven and deep learning methods
Physical models: Through physical model methods, due to improvements in machine data collection. Figure
system degradation is mathematically described based 4 illustrates the frameworks of three different machine
on prior system knowledge [84, 85]. Despite being easy health monitoring systems (MHMS) including Physical
to understand on a physical level, they are difficult to Models, Conventional Data-driven Models and Deep
implement on a complex level. Learning Models.
Data-driven models: As explained above,
data-driven methods are used to predict the state of a
system by analyzing historical data and learning from it.
Due to the fact that they do not need to know how
complex systems work, they are suitable for complex
systems. ML algorithms [86] and deep learning

983
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

representations for mapping input patterns in measured

sensor data to output targets to assess damage at various
levels [96]. Gholizadeh et al. [96] evaluated damage
progression under fatigue loading on glass fiber
reinforced polyester composite materials using acoustic
emission (AE) with different approaches of ML;
ensemble learning methods namely; XGboost,
LightGBM, and CatBoost, and unsupervised learning
Figure 3: Frameworks of three different machine
k-means clustering. XGboost was found as a most
health monitoring systems (MHMS) [90]
reliable technique with the accuracy of 95%. Abhishek
et al. [97] investigated the SHM of Wind Turbine Blade
The prognostics stage of Structural Health
(WTB) using convolutional neural network (CNN) for
Monitoring (SHM) is used in predictive maintenance to
image classification. Accuracies of 94.94% of binary
predict the future evolution of the defect or degradation.
fault classification and 91% of multiple class fault
Originally, SHM systems utilized wired sensor
classification were achieved. Gomez et al. [95] focused
networks [91]. Wireless sensor networks (WSNs) are a
on the process of features extraction and patterns
viable option platform for SHM systems because of
recognition by SHM systems used in bridge
their superior trustworthiness and low installation and
construction. An evaluation of ML algorithms in SHM
maintenance costs [92, 93]. As part of the assessment of
systems for reinforced concrete bridges was conducted
structures with SHM systems, cracking, displacement,
by Fan et al. [98]. ML algorithms are explored in the
temperature, corrosion, force, fatigue, tilt, settlement,
areas of structural design, construction quality
vibration, strain, and water level and wind are taken into
management, bridge engineering, and inspection. To
account [94]. Figure 5 depicts the SHM systems areas,
identify, locate and predict structural health, feature
methods, techniques and algorithms.
extraction can be implemented using signal processing
techniques [99, 100]. Additionally, ML algorithms have
also been employed in the field of pattern recognition
[101-104]. The use of machine learning algorithms has
been demonstrated in the variety of applications
throughout the life cycle of structures, including design
optimization, maintenance, performance assessment,
SHM, damage detection and construction. [105-107].
As the name implies, condition-based maintenance
(CBM) plans maintenance based on condition
monitoring (CM) data such as temperature, vibration
Figure 4: Diagram of structural health monitoring and currents. As part of CBM, several disciplines are
systems areas [95] involved, such as failure analysis, on-line diagnostics,
interpreting diagnostic data and so forth. Widod and
ML in SHM is used to construct models or Yang [108] investigated machine condition monitoring

984
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

and fault diagnosis using support vector machine

(SVM). SVM has demonstrated excellent
generalization performance with high classification
accuracy for machine condition monitoring and
diagnosis. Bakar et al. [109] developed failure root
causes database using CBM actual site data of 33 kV
switchgears. Multiple methods were used to obtain the
CBM data, including ultrasound, thermoscanning, and
transient earth voltage (TEV). Based on the sequence of
failures in the database with the highest percentages, the
Figure 5: Illustrations of the induction motor fault
root cause was identified. On the basis of the highest
diagnosis using proposed unsupervised CNN [110]
percentage of occurrences, the ultrasonic method
successfully identified most root causes of failure. An
ML models can also be deployed for real-time
induction motor fault diagnosis model based on
monitoring of NDT data streams. By continuously
convolutional discriminative feature learning has been
analyzing incoming data, such as sensor readings or
proposed by Sun et al [110]. A back-propagation-based
streaming images, and compare it to established
neural network (BPNN) was used to pre-learn local
baselines or thresholds, ML algorithms can quickly
filters in a feed-forward convolutional pooling
detect deviations, anomalies, or potential defects,
architecture as shown in Figure 6. In the next step, SVM
enabling timely interventions or alerts to operators.
was used to classify fault conditions based on the
Real-time monitoring enhances safety, reduces
learned representation. Based on the BPNN-learned
downtime, and supports proactive maintenance
local filters, the following convolutional pooling
strategies.
architecture could extract invariant and discriminative
An important research topic is time series analysis,
features quickly from the raw vibration data. Since the
which involves two aspects: (1) identifying the nature
input data was a 1D vibration signal, their work was
of the phenomena represented by the time series of
also considered as a 1D CNN. Ben Ali et al. [111] used
observations[112], and (2) predicting future values of
an ANN-based model to monitor and diagnose rolling
the time series variable based on historical data [113].
bearing conditions. Empirical mode decomposition and
The preprocessing of time series data, the extraction of
energy entropy were used to extract features. As a
features, and the customization of algorithms are
result, the proposed method demonstrated that an ANN
critical to predictive maintenance. Prognostics have
can be effectively used by itself without human
been approached as a time series prediction problem by
intervention as a tool for assessing bearing degradation.
some researchers and forecasting methods have been
proposed such as: Artificial Neural Networks (ANNs)
[114], Auto Regressive Integrated Moving Average
(ARIMA) [115], Exponential smoothing model (ETS)
[116], Support vector Regression (SVR) [117].
Auto-Regressive Integrated Moving Average

985
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

(ARIMA), and Support Vector Regression (SVR) are predicted according to those vectors. Zhao et al. [124]
categorized as statistical models for anomaly detection Further, Zhao et al incorporated CNN and LSTM to
[81]. Sfar et al. [118] investigated SHM of a tube via formulate a more complicated deep learning model
ultrasonic guided waves by utilizing two models based called the Convolutional Bi-directional Long
on Empirical Mode Decomposition (EDM)- ARIMA to Short-Term Memory Network (CBLSTM). The robust
realize high-precision predictions, and predict local features of the sequential input were extracted
components; and Auto-Regressive Neural Networks with CNN, as shown in Figure 7. In the sequential
(ARNNET) technique to estimate functions based on a output of CNN, bi-directional LSTM was adopted to
huge volume of training data. EMD-ARIMA provides encode temporal information. The final step was to
better results than ARIMA for both modes. In the study predict the target value by adding fully-connected
by Hongzhan et al. [117], ARIMA was used to forecast stacked layers and using a linear regression. Compared
linear basic load, whereas SVMs were used to forecast with several baseline approaches, including
non-linear sensitive load, and a hybrid model of both. conventional LSTM models, the proposed model
The result showed that the hybrid ARIMA-SVMs excelled at tool wear tests. A LSTM-based
realized the mutual supplement with each other and encoder-decoder structure was developed by Malhotra
performed better than the two separate models of et al [125], in which a LSTM-based encoder
themselves with lowest error of 3.85%. transformed a multivariate input sequence to a
Deep learning has emerged as one of the most fixed-length vector, and then the LSTM decoder
popular techniques in the past years to detect anomalies produced the target sequence from the vectors. To
in time series data, delivering state-of-the-art results for predict RUL, their assumptions state that the model the
a range of unsupervised and supervised tasks due to its model can initially be trained unsupervised on raw
capability of learning high-level representations signals corresponding to normal behavior. After that,
automatically from data; its higher accuracy, greater the reconstruction error is used to calculate a health
flexibility, stronger generalization, and its lower index (HI) that can be used to estimate RUL.
dependency on domain knowledge [119]. long
short-term memory (LSTMs) and recurrent neural
networks (RNNs) are now the deep learning models of
choice for learning long-range patterns in sequential
and temporal data [120, 121].
Yuan et al. [122] investigated three RNN models for
aeroengine fault diagnosis and prognostics, including
vanilla RNN, LSTM and GRU. The LSTM and GRU Figure 6: Illustrations of the proposed Convolutional

advanced RNN models outperformed vanilla RNNs. Bi-directional Long Short-Term Memory Networks

Interstingly, LSTM performance was not improved by [124]

the ensemble model of the three RNN variants above. A

tool wear test utilizing LSTMs was presented by Zhao A real-time anomaly detection system was described

et al. [123]. In the applied LSTM model, raw sensory in Numenta [126], Aggarwal et al [127-129]. A

data was encoded into vectors and tool wear was potential breakthrough developed by Numenta is

986
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

Hierarchical Temporal Memory (HTM) [126], a [4] V. Dive and S. Lakade, Recent Research
Progress on Residual Stress Measurement
biologically inspired machine intelligence system that
Using Non-Destructive Testing. Materials
mimics neocortex processes and architecture. HTM Today: Proceedings, 2021. 47: p. 3282-3287.
[5] K. Kurashkin, V. Mishakin, and A. Rudenko,
stood out as a promising real-time predictive
Ultrasonic evaluation of residual stresses in
maintenance approach. Kukjin et al. [130] provided a welded joints of hydroelectric unit rotor frame.
Materials Today: Proceedings, 2019. 11: p.
broad explanation of how DL can be applied to anomaly
163-168.
detection. [6] R. Růžek, R. Lohonka, and J. Jironč, Ultrasonic
C-Scan and shearography NDI techniques
evaluation of impact defects identification.
NDT & E International, 2006. 39(2): p.
CONCLUSION 132-142.
[7] A. Katunin, K. Dragan, and M.
ML and DL represents power tools that can lead to Dziendzikowski, Damage identification in
significant improvements in NDT and related fields. aircraft composite structures: A case study
using various non-destructive testing
Overall, ML and DL in NDT empowers industries to techniques. Composite Structures, 2015.
leverage the power of data analysis and pattern 127(0): p. 1-9.
[8] N. P. Avdelidis, D. P. Almond, A. Dobbinson,
recognition to enhance inspection capabilities, increase B. C. Hawtin, C. Ibarra-Castanedo, and X.
accuracy and efficiency, enable and optimize predictive Maldague, Aircraft composites assessment by
means of transient thermal NDT. Progress in
maintenance strategies, support data-driven Aerospace Sciences, 2004. 40(3): p. 143-162.
decision-making in various industries and ultimately [9] T. H. Loutas, A. Panopoulou, D. Roulias, and
V. Kostopoulos, Intelligent health monitoring
contributing to safer and more reliable systems and of aerospace composite structures based on
structures. The rapid advancement of NDT dynamic strain measurements. Expert Systems
with Applications, 2012. 39(9): p. 8412-8422.
technologies and automation necessitates the urgent [10] M. Papaelias, F.P.G. Marquez, and A.
need for a precise assessment of test signals, data, Karyotakis, Non-Destructive Testing and
Condition Monitoring Techniques for
images, and patterns. In order to fulfill this requirement, Renewable Energy Industrial Assets. 2019:
an AI-powered knowledge-based system is desired to Elsevier Science.
[11] K. Ali-Benyahia, S. Kenai, M. Ghrici, Z.
process the NDT testing data and generate a Sbartaï, and S. Elachachi, Analysis of the
comprehensive output in the form of classified and accuracy of in-situ concrete characteristic
compressive strength assessment in real
systematic interpretation. structures using destructive and
non-destructive testing methods. Construction
and Building Materials, 2023. 366: p. 130161.
[12] R. Pucinotti, Reinforced concrete structure:
REFERENCES Non destructive in situ strength assessment of
concrete. Construction and Building Materials,
2015. 75: p. 331-341.
[1] C. D. Lockard, Anomaly detection in [13] W. J. Staszewski, S. Mahzan, and R. Traynor,
radiographic images of composite materials via Health monitoring of aerospace composite
crosshatch regression. 2015, Mills College: structures – Active and passive approach.
Ann Arbor. p. 76. Composites Science and Technology, 2009.
[2] S. Gholizadeh, A review of non-destructive 69(11–12): p. 1678-1685.
testing methods of composite materials. [14] V. Giurgiutiu and J. Bao,
Procedia Structural Integrity, 2016. 1: p. 50-57. Embedded-ultrasonics Structural Radar for In
[3] S.Gholizadeh, Impact behaviours and Situ Structural Health Monitoring of Thin-wall
Non-Destructive Testing (NDT) methods in Structures. Structural Health Monitoring, 2004.
Carbon Fiber Composites in Aerospace 3(2): p. 121-140.
Industry: A Review. Authorea, 2022.

987
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

[15] C. Mandache, Overview of non-destructive monitoring. CEUR Workshop Proceedings,

evaluation techniques for metal-based additive 2015. 1353: p. 133-138.
manufacturing. Materials Science and [28] K. Smarsly, K. Dragos, and J. Wiggenbrock,
Technology, 2019. 35(9): p. 1007-1015. Machine learning techniques for structural
[16] I. Ramírez, F. P. G. Márquez, and M. Papaelias, health monitoring in 8th European Workshop
Review on additive manufacturing and On Structural Health Monitoring (EWSHM
non-destructive testing. Journal of 2016). 2016: Spain, Bilbao.
Manufacturing Systems, 2023. 66: p. 260-286. [29] P. A. Vikhar. Evolutionary algorithms: A
[17] R. Usamentiaga, P. Venegas, J. Guerediaga, L. critical review and its future prospects. in 2016
Vega, and I. López, Automatic detection of International Conference on Global Trends in
impact damage in carbon fiber composites Signal Processing, Information Computing and
using active thermography. Infrared Physics & Communication (ICGTSPICC). 2016.
Technology, 2013. 58(0): p. 36-46. [30] H. Zheng, L. X. Kong, and S. Nahavandi,
[18] C. Maierhofer, P. Myrach, M. Reischel, H. Automatic inspection of metallic surface
Steinfurth, M. Röllig, and M. Kunert, defects using genetic algorithms. Journal of
Characterizing damage in CFRP structures Materials Processing Technology, 2002.
using flash thermography in reflection and 125-126: p. 427-433.
transmission configurations. Composites Part [31] I. Goodfellow, Y. Bengio, and A. Courville,
B: Engineering, 2014. 57(0): p. 35-46. Deep Learning (Adaptive Computation and
[19] A. Ataş and C. Soutis, Subcritical damage Machine Learning series). 2016: The MIT
mechanisms of bolted joints in CFRP Press; Illustrated edition
composite laminates. Composites Part B: [32] D. Medak, L. Posilović, M. Subašić, M.
Engineering, 2013. 54(0): p. 20-27. Budimir, and S. Lončarić, DefectDet: A deep
[20] C. Meola and G. M. Carlomagno, Infrared learning architecture for detection of defects
thermography to evaluate impact damage in with extreme aspect ratios in ultrasonic images.
glass/epoxy with manufacturing defects. Neurocomputing, 2022. 473: p. 107-115.
International Journal of Impact Engineering, [33] M. Mahdi Naddaf-Sh, S. Naddaf-Sh, H.
2014. 67(0): p. 1-11. Zargarzadeh, Sayyed M. Zahiri, M. Dalton, G.
[21] P. Zhang, F. H. Wittmann, T. J. Zhao, E. H. Elpers, and A.R. Kashani, 9 - Defect detection
Lehmann, L. Tian, and P. Vontobel, and classification in welding using deep
Observation and quantification of water learning and digital radiography, in Fault
penetration into Strain Hardening Diagnosis and Prognosis Techniques for
Cement-based Composites (SHCC) with Complex Engineering Systems, H. Karimi,
multiple cracks by means of neutron Editor. 2021, Academic Press. p. 327-352.
radiography. Nuclear Instruments and Methods [34] I.D. Made Oka Dharmawan, J. Lee, and S. Sim,
in Physics Research Section A: Accelerators, Defect Shape Classification Using Transfer
Spectrometers, Detectors and Associated Learning in Deep Convolutional Neural
Equipment. 620(2–3): p. 414-420. Network on Magneto-Optical Nondestructive
[22] U. Dinesh Kumar, Business Analytics: The Inspection. Advances in Nondestructive
Science of Data - Driven Decision Making Testing and Evaluation, 2022. 12(15): p. 7613.
2017: Wiley. [35] R. Neven and T. Goedemé, A multi-branch
[23] K.P. Murphy, Machine Learning: A U-Net for steel surface defect type and severity
Probabilistic Perspective. 2012: MIT Press. segmentation. Journal of Metals, 2021. 11(6):
[24] N. Meyendorf Nathan Ida, Handbook of p. 870.
Advanced Nondestructive Evaluation. 2019: [36] O. Serradilla, E. Zugasti, J. Rodriguez, and U.
Springer Cham. Zurutuza, Deep learning models for predictive
[25] I. H Sarker, ASM Kayes, Sh. Badsha, H. maintenance: a survey, comparison, challenges
Alqahtani, P. Watters, and A. Ng, and prospects. Journal of Applied Intelligence,
Cybersecurity data science: an overview from 2022. 52(10): p. 10934-10964.
machine learning perspective. Journal of Big [37] A. V. Lakshmi, V. Gopitilak, M. M. Parvez,
data . 2020. 7(1): p. 1-29. S.K. Subhani, and V.S. Ghali, Artificial neural
[26] J. Han, J. Pei, and M. Kamber, Data mining: networks based quantitative evaluation of
concepts and techniques. 2011: Elsevier. subsurface anomalies in quadratic frequency
[27] W. Nick, J. Shelton, K. Asamene, and A. modulated thermal wave imaging. Infrared
Esterline, A study of supervised machine Physics Technology, 2019. 97: p. 108-115.
learning techniques for structural health

988
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

[38] G. Huang, Y.Chen, and H.A. Babri, defects in horticultural products. Postharvest
Classification ability of single hidden layer Biology and Technology, 2017. 128: p. 33-43.
feedforward neural networks. IEEE [50] Fusaomi Nagata, Kenta Tokuno, Kazuki
transactions on neural networks, 2000. 11(3): p. Mitarai, Akimasa Otsuka, Takeshi Ikeda,
799-801. Hiroaki Ochi, Keigo Watanabe, and Maki K.
[39] J. Huang, Q. Hu, Zh. Song, G. Zhang, Ch.Zh. Habib, Defect detection method using deep
Qin, M. Wu, and Xia. Wang, Classification of convolutional neural network, support vector
cracking sources of different engineering machine and template matching techniques.
media via machine learning. Fatigue Fracture Artificial Life Robotics, 2019. 24: p. 512 - 519.
of Engineering Materials Structures, 2021. [51] I. Valavanis and D. Kosmopoulos, Multiclass
44(9). defect detection and classification in weld
[40] I. Tabian, H. Fu, and Z. Sharif Khodaei, A radiographic images using geometric and
Convolutional Neural Network for Impact texture features. Expert Systems with
Detection and Characterization of Complex Applications, 2010. 37(12): p. 7606-7614.
Composite Structures. Sensors, 2019. 19(22). [52] R.R. Silva and D. Mery, State-of-the-art of
[41] N. F. Ince, Chu-Shu Kao, M. Kaveh, A. weld seam inspection by radiography testing:
Tewfik, and J. F. Labuz, A Machine Learning Part I – Image Processing. Materials
Approach for Locating Acoustic Emission. Evaluation, , 2007. 65(2): p. pp. 643-647.
EURASIP Journal on Advances in Signal [53] Sh. Wang and R. M. Summers, Machine
Processing, 2010(1). learning and radiology. Medical Image
[42] P. Liu, Dong Xu, J. Li, Zh. Chen, Sh. Wang, J. Analysis, 2012. 16(5): p. 933-951.
Leng, R. Zhu, L. Jiao, W. Liu, and Zh. Li, [54] M. F. Sheikh, Kh. Kamal, F. Rafique, S. Sabir,
Damage mode identification of composite wind H. Zaheer, and K. Khan, Corrosion detection
turbine blade under accelerated fatigue loads and severity level prediction using acoustic
using acoustic emission and machine learning. emission and machine learning based approach.
SAGE, 2020. 19(4): p. 1092-1103. Ain Shams Engineering Journal, 2021. 12(4):
[43] M. Z. Alom, M. Alam, T. M. Taha, and K. M. p. 3891-3903.
Iftekharuddin. Object recognition using [55] M. Le, V. Luong, D. Kh. Nguyen, D.Kh. Le,
cellular simultaneous recurrent networks and and J. Lee, Auto-Detection of Hidden
convolutional neural network. in 2017 Corrosion in an Aircraft Structure by
International Joint Conference on Neural Electromagnetic Testing: A Machine-Learning
Networks (IJCNN). 2017. Approach. 2022. 12(10): p. 5175.
[44] V. Lakhani and R. Mahadev, Multi-Language [56] B. T. Bastian, J.N, S. Kumar Ranjith, and C. V.
Identification Using Convolutional Recurrent Jiji, Visual inspection and characterization of
Neural Network. arxiv, 2017. external corrosion in pipelines using deep
[45] M A. De Oliveira, N. V. S. Araujo, R. N. Da neural network. NDT & E International, 2019.
Silva, T. I. Da Silva, and J. Epaarachchi, Use of 107: p. 102134.
Savitzky–Golay Filter for Performances [57] A. Wronkowicz-Katunin, A. Katunin, M.
Improvement of SHM Systems Based on Nagode, and J. Klemenc, Classification of
Neural Networks and Distributed PZT Sensors. Cracks in Composite Structures Subjected to
Sensors, 2018. 18(1). Low-Velocity Impact Using
[46] N. Yue and Z. Sharif Khodaei, Assessment of Distribution-Based Segmentation and Wavelet
Impact Detection Techniques for Aeronautical Analysis of X-ray Tomograms. 2021. 21(24):
Application: ANN vs. LSSVM. Journal of p. 8342.
Multiscale Modelling, 2016. 07(04): p. [58] J.Zhang and J. Zhang, An Improved
1640005. Nondestructive Semantic Segmentation
[47] Y. LeCun, Y. Bengio, and G. Hinton, Deep Method for Concrete Dam Surface Crack
learning. Nature, 2015. 521(7553): p. 436-444. Images with High Resolution. Mathematical
[48] H. Kasban, O. Zahran, H. Arafa, M. El-Kordy, Problems in Engineering, 2020. 2020: p.
S. M. S. Elaraby, and F. E. Abd El-Samie, 5054740.
Welding defect detection from radiography [59] G. Du, J. Li, W. K. Wang, C. Jiang, and S. Z.
images with a cepstral approach. NDT & E Song, Detection and characterization of
International, 2011. 44(2): p. 226-231. stress-corrosion cracking on 304 stainless steel
[49] M. van Dael, P. Verboven, J. Dhaene, L. Van by electrochemical noise and acoustic emission
Hoorebeke, J. Sijbers, and B. Nicolai, techniques. Corrosion Science, 2011. 53(9): p.
Multisensor X-ray inspection of internal 2918-2926.

989
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

[60] N. Saenkhum, A. Prateepasen, and P. in industrial inspection. CIRP Annals, 2016.

Keawtrakulpong. Classification of corrosion 65(1): p. 417-420.
detected by acoustic emission. in ASME [71] L. A. O. Martins, F. L. C. Pádua, and P. E. M.
International Mechanical Engineering Almeida. Automatic detection of surface
Congress and Exposition. 2003. defects on rolled steel using Computer Vision
[61] A. Risheh, P. Tavakolian, A. Melinkov, and A. and Artificial Neural Networks. in IECON
Mandelis, Infrared computer vision in 2010 - 36th Annual Conference on IEEE
non-destructive imaging: Sharp delineation of Industrial Electronics Society. 2010.
subsurface defect boundaries in enhanced [72] A. Arbaoui, A. Ouahabi, S. Jacques, and M.
truncated correlation photothermal coherence Hamiane, Concrete Cracks Detection and
tomography images using K-means clustering. Monitoring Using Deep Learning-Based
NDT & E International, 2022. 125: p. 102568. Multiresolution Analysis. 2021. 10(15): p.
[62] A. Sophian, G. Tian, and M. Fan, Pulsed Eddy 1772.
Current Non-destructive Testing and [73] Y. Cha, W. Choi, and O. Büyüköztürk, Deep
Evaluation: A Review. Chinese Journal of Learning-Based Crack Damage Detection
Mechanical Engineering, 2017. 30(3): p. Using Convolutional Neural Networks.
500-514. Computed-Aided civil and infrastructure
[63] N. H. M. M. Shrifan, G. N. Jawad, N. A. M. Isa, engineering, 2017. 32(5): p. 361-378.
and M. F. Akbar, Microwave Nondestructive [74] S. Yoon, A. Kim, W. J. Cantwell, Ch. Yeun,
Testing for Defect Detection in Composites Ch. Cho, Y. Byon, and T. Kim, Defect
Based on K-Means Clustering Algorithm. detection in composites by deep learning using
IEEE Access, 2021. 9: p. 4820-4828. solitary waves. International Journal of
[64] O. Siljama, T. Koskinen, O. Jessen-Juhler, and Mechanical Sciences, 2023. 239: p. 107882.
I. Virkkunen, Automated Flaw Detection in [75] Q. Luo, W. Woo, and Y. Yang, Temporal and
Multi-channel Phased Array Ultrasonic Data spatial deep learning network for infrared
Using Machine Learning. Journal of thermal defect detection. NDT & E
Nondestructive Evaluation, 2021. 40(3): p. 67. International, 2019. 108: p. 102164.
[65] S. Sambath, P. Nagaraj, and N. Selvakumar, [76] H. Yuan, H. Chen, S. Liu, J. Lin, and X. Luo. A
Automatic Defect Classification in Ultrasonic Deep Convolutional Neural Network for
NDT Using Artificial Intelligence. Journal of Detection of Rail Surface Defect. in 2019 IEEE
Nondestructive Evaluation, 2011. 30: p. 20-28. Vehicle Power and Propulsion Conference
[66] C Fei, Zh Han, and J Dong, An ultrasonic (VPPC). 2019.
flaw-classification system with wavelet-packet [77] S. Faghih-Roohi, S. Hajizadeh, A. Núñez, R.
decomposition, a mutative scale chaotic genetic Babuska, and B. De Schutter. Deep
algorithm, and a support vector machine and its convolutional neural networks for detection of
application to petroleum-transporting rail surface defects. in 2016 International Joint
pipelines. Russian Journal of Nondestructive Conference on Neural Networks (IJCNN).
Testing, 2006. 42: p. 190-197. 2016.
[67] V. Matz, M. Kreidl, and R. Smid, Classification [78] K. Gopalakrishnan, S. K. Khaitan, A.
of ultrasonic signals. International Journal of Choudhary, and A. Agrawal, Deep
Materials Product Technology, 2006. 27(3-4): Convolutional Neural Networks with transfer
p. 145-155. learning for computer vision-based data-driven
[68] S. Niu, B. Li, X. Wang, and H. Lin, Defect pavement distress detection. Construction and
Image Sample Generation With GAN for Building Materials, 2017. 157: p. 322-330.
Improving Defect Recognition. IEEE [79] M. Baptista, Sh. Sankararaman, Ivo P. de
Transactions on Automation Science and Medeiros, C. Nascimento, H. Prendinger, and
Engineering, 2020. 17(3): p. 1611-1622. E. M. P. Henriques, Forecasting fault events for
[69] N. Munir, J. Park, H. Kim, S. Song, and S. predictive maintenance using data-driven
Kang, Performance enhancement of techniques and ARMA modeling. Computers
convolutional neural network for ultrasonic & Industrial Engineering, 2018. 115: p. 41-53.
flaw classification by adopting autoencoder. [80] L. Li, M. Liu, W. Shen, and G. Cheng, An
NDT & E International, 2020. 111: p. 102218. expert knowledge-based dynamic maintenance
[70] D. Weimer, B. Scholz-Reiter, and M. Shpitalni, task assignment model using discrete
Design of deep convolutional neural network stress–strength interference theory.
architectures for automated feature extraction Knowledge-Based Systems, 2017. 131: p.
135-148.

990
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

[81] P. Adhikari, H. Gururaja Rao, and Dipl.-Ing. [92] J. P. Lynch and K. J. Loh, A summary review
M. Buderath, Machine Learning based Data of wireless sensors and sensor networks for
Driven Diagnostics & Prognostics Framework structural health monitoring. Shock vibration
for Aircraft Predictive Maintenance in 10th digest, 2006. 38(2): p. 91-130.
International Symposium on NDT in [93] N. Xu, S. Rangwala, K. K. Chintalapudi, D.
Aerospace 2018: Germany. Ganesan, A. Broad, R. Govindan, and D.
[82] H. O. A. Ahmed, M. L. D. Wong, and A. K. Estrin. A wireless sensor network for structural
Nandi, Intelligent condition monitoring method monitoring. in Proceedings of the 2nd
for bearing faults from highly compressed international conference on Embedded
measurements using sparse over-complete networked sensor systems. 2004.
features. Mechanical Systems and Signal [94] M. Modares and N. Waksmanski, Overview of
Processing, 2018. 99: p. 459-477. structural health monitoring for steel bridges.
[83] L. Liao and F. Köttig, A hybrid framework Journal of Practice Periodical on Structural
combining data-driven and model-based Design Construction, 2013. 18(3): p. 187-191.
methods for system remaining useful life [95] A. Gomez-Cabrera, E. Ambrosio, and P. Jorge,
prediction. Applied Soft Computing, 2016. 44: Review of machine-learning techniques
p. 191-199. applied to structural health monitoring systems
[84] H. Oppenheimer Charles and A. Loparo for building and bridge structures. Journal of
Kenneth. Physically based diagnosis and Applied Sciences, 2022. 12(21): p. 10754.
prognosis of cracked rotor shafts. in Proc.SPIE. [96] S. Gholizadeh, Z. Leman, and B. T. H. T.
2002. Baharudin, State-of-the-art ensemble learning
[85] V. Venkatasubramanian, R. Rengaswamy, K. and unsupervised learning in fatigue crack
Yin, and S. N. Kavuri, A review of process recognition of glass fiber reinforced polyester
fault detection and diagnosis: Part I: composite (GFRP) using acoustic emission.
Quantitative model-based methods. Computers Ultrasonics, 2023. 132: p. 106998.
& Chemical Engineering, 2003. 27(3): p. [97] A. Reddy, V. Indragandhi, L. Ravi, and V.
293-311. Subramaniyaswamy, Detection of Cracks and
[86] Y. Jiang and S. Yin, Recursive Total Principle damage in wind turbine blades using artificial
Component Regression Based Fault Detection intelligence-based image analytics.
and Its Application to Vehicular Measurement, 2019. 147: p. 106823.
Cyber-Physical Systems. IEEE Transactions on [98] W. Fan, Y. Chen, J. Li, Y. Sun, J. Feng, H.
Industrial Informatics, 2018. 14(4): p. Hassanin, and P. Sareh. Machine learning
1415-1423. applied to the design and inspection of
[87] D. You, X. Gao, and S. Katayama, reinforced concrete bridges: Resilient methods
WPD-PCA-Based Laser Welding Process and emerging applications. in Structures. 2021.
Monitoring and Defects Diagnosis by Using Elsevier.
FNN and SVM. IEEE Transactions on [99] P. Jorge E. Ambrosio, X. Liu, N. Lieven, and
Industrial Electronics, 2015. 62(1): p. 628-636. JM. Ramírez-Cortés. Wavelet-fuzzy logic
[88] Ch. Farrar and K.Worden, Structural Health approach to structural health monitoring. in
Monitoring A Machine Learning Perspective. 2011 Annual Meeting of the North American
2013. Fuzzy Information Processing Society. 2011.
[89] W. Zhang, D. Yang, and H. Wang, Data-Driven IEEE.
Methods for Predictive Maintenance of [100] P. Jorge E. Ambrosio and N. Lieven. Soft
Industrial Equipment: A Survey. IEEE Systems computing feature extraction for health
Journal, 2019. 13(3): p. 2213-2227. monitoring of rotorcraft structures. in 2007
[90] R. Zhao, R. Yan, Zh. Chen, K. Mao, P. Wang, IEEE International Fuzzy Systems Conference.
and R. X. Gao, Deep learning and its 2007. IEEE.
applications to machine health monitoring. [101] H. Ibrahim and B. H Far. Simulation-based
Mechanical Systems and Signal Processing, benefit analysis of pattern recognition
2019. 115: p. 213-237. application in intelligent transportation
[91] A. Sofi, J. J. Regita, B. Rane, and H.Ho Lau, systems. in 2015 IEEE 28th Canadian
Structural health monitoring using wireless Conference on Electrical and Computer
smart sensor network – An overview. Engineering (CCECE). 2015. IEEE.
Mechanical Systems and Signal Processing, [102] A. Entezami, H. Shariatmadar, and S. Mariani,
2022. 163: p. 108113. Early damage assessment in large-scale
structures by innovative statistical pattern

991
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

recognition methods based on time series diagnosis based on vibration signals. Applied
modeling and novelty detection. Advances in Acoustics, 2015. 89: p. 16-27.
Engineering Software, 2020. 150: p. 102923. [112] E. Keogh, J. Lin, and A. Fu. Hot sax:
[103] A. Fiorini Morosini, S. Shaffiee Haghshenas, S. Efficiently finding the most unusual time series
Shaffiee Haghshenas, D. Yong Choi, and subsequence. in Fifth IEEE International
Z.Woo Geem, Sensitivity analysis for Conference on Data Mining (ICDM'05). 2005.
performance evaluation of a real water Ieee.
distribution system by a pressure driven [113] Zh. Ji, B. Wang, S. Deng, and Zh. You,
analysis approach and artificial intelligence Predicting dynamic deformation of retaining
method. Journal of Water, 2021. 13(8): p. 1116. structure by LSSVR-based time series method.
[104] G. Guido, S. Shaffiee Haghshenas, S. Shaffiee Journal of Neuro computing, 2014. 137: p.
Haghshenas, A. Vitale, V. Astarita, Y. Park, 165-172.
and Z. Woo Geem, Evaluation of contributing [114] Sh. Bhardwaj, E. Chandrasekhar, P. Padiyar,
factors affecting number of vehicles involved and V. Gadre, A comparative study of
in crashes using machine learning techniques in wavelet-based ANN and classical techniques
rural roads of Cosenza, Italy. Journal of Safety, for geophysical time-series forecasting.
2022. 8(2): p. 28. Computers & Geosciences, 2020. 138: p.
[105] G. Ferreira Gomes, Y. Mendéz, P. Silva Lopes 104461.
Alexandrino, S. Simões da Cunha Jr, and A. [115] N. Suhermi, Suhartono, D. Prastyo, and
Carlos Ancelotti Jr, The use of intelligent Baharuddin. A , Roll motion prediction using a
computational tools for damage detection and hybrid deep learning and ARIMA model.
identification with an emphasis on Procedia Computer Science, 2018. 144: p.
composites–A review. Composite Structures, 251-258.
2018. 196: p. 44-54. [116] S. Panigrahi and H. Behera, A hybrid
[106] R. Mikaeil, S. Sh. Haghshenas, S. Shaffiee ETS-ANN model for time series forecasting.
Haghshenas, and M. Ataei, Performance Eng. Appl. of AI, 2017. 66: p. 49-59.
prediction of circular saw machine using [117] H. Nie, G. Liu, X. Liu, and Y. Wang, Hybrid of
imperialist competitive algorithm and fuzzy ARIMA and SVMs for Short-Term Load
clustering technique. Journal of Neural Forecasting. Energy Procedia, 2012. 16: p.
Computing Applications, 2018. 29: p. 283-292. 1455-1460.
[107] E. Figueiredo, I. Moldovan, A. Santos, P. [118] M. Sfar, S. Yaacoubi, S. Dellagi, M.
Campos, and J.CWA Costa, Finite Mountassir, El, and S. Turki, Towards A
element–based machine-learning approach to Predictive Maintenance Based On Ultrasonic
detect damage in bridges under operational and Guided Waves Non-Destructive Testing
environmental variations. Journal of Bridge Technique: A Case Study, in 13ème
Engineering, 2019. 24(7): p. 04019061. Conference Internationale De Modelisation,
[108] A. Widodo and B. Yang, Support vector Optimisation Et Simulation (MOSIM2020),
machine in machine condition monitoring and 12-14 Nov 2020, AGADIR, Maroc. 2020:
fault diagnosis. Mechanical Systems and Signal AGADIR (virtual), Morocco.
Processing, 2007. 21(6): p. 2560-2574. [119] Zh. Ji, J. Gong, and J. Feng, A Novel Deep
[109] A. H. A. Bakar, H. A. Illias, M. K. Othman, and Learning Approach for Anomaly Detection of
H. Mokhlis, Identification of failure root causes Time Series Data. Scientific Programming,
using condition based monitoring data on a 2021. 2021: p. 6636270.
33kV switchgear. International Journal of [120] A. Sagheer and M. Kotb, Time series
Electrical Power & Energy Systems, 2013. 47: forecasting of petroleum production using deep
p. 305-312. LSTM recurrent networks. Journal of Neuro
[110] W. Sun, R. Zhao, R. Yan, S. Shao, and X. computing, 2019. 323: p. 203-213.
Chen, Convolutional Discriminative Feature [121] N. Jin, Y. Zeng, K. Yan, and Zh. Ji,
Learning for Induction Motor Fault Diagnosis. Multivariate air quality forecasting with nested
IEEE Transactions on Industrial Informatics, long short term memory neural network. IEEE
2017. 13(3): p. 1350-1359. Transactions on Industrial Informatics, 2021.
[111] J. Ben Ali, N. Fnaiech, L. Saidi, B. 17(12): p. 8514-8522.
Chebel-Morello, and F. Fnaiech, Application of [122] M. Yuan, Y. Wu, and L. Lin. Fault diagnosis
empirical mode decomposition and artificial and remaining useful life estimation of aero
neural network for automatic bearing fault engine using LSTM neural network. in 2016

992
Journal of Harbin Engineering University Vol 44 No. 11
ISSN: 1006-7043 November 2023

IEEE international conference on aircraft

utility systems (AUS). 2016. IEEE.
[123] R. Zhao, J. Wang, R. Yan, and K. Mao.
Machine health monitoring with LSTM
networks. in 2016 10th international
conference on sensing technology (ICST).
2016. IEEE.
[124] R. Zhao, R. Yan, J. Wang, and K. Mao,
Learning to Monitor Machine Health with
Convolutional Bi-Directional LSTM
Networks. 2017. 17(2): p. 273.
[125] P. Malhotra, V. Tv, A. Ramakrishnan, G.
Anand, L.Vig, P. Agarwal, and G. Shroff,
Multi-sensor prognostics using an
unsupervised health index based on LSTM
encoder-decoder. arXiv preprint arXiv:.06154,
2016.
[126] Numenta, The science of Anomaly Detection -
How HTM Enables Anomaly Detection in
Streaming Data White Paper, 2015.
[127] C. C. Aggarwal and K. Subbian, Event
Detection in Social Streams, in Proceedings of
the 2012 SIAM International Conference on
Data Mining (SDM). p. 624-635.
[128] C. C. Aggarwal, On Abnormality Detection in
Spuriously Populated Data Streams, in
Proceedings of the 2005 SIAM International
Conference on Data Mining (SDM). p. 80-91.
[129] C. C. Aggarwal, Y. Zhao, and P. S. Yu. Outlier
detection in graph streams. in IEEE 27th
International Conference on Data Engineering.
2011.
[130] K. Choi, J. Yi, C. Park, and S. Yoon, Deep
Learning for Anomaly Detection in
Time-Series Data: Review, Analysis, and
Guidelines. IEEE Access, 2021. PP: p. 1-1.

993
View publication stats