2024 Prognostics and System Health Management Conference (PHM)

Machine Learning techniques for Vibration-based

Condition Monitoring–A review
Ahmed Shehata Omar D.Mohammed
Estonian Maritime Academy Mechanical Engineering Department
Tallinn University of Technology Prince Mohammad bin Fahd University
Tallinn, Estonia Khobar, Saudi Arabia
2024 Prognostics and System Health Management Conference (PHM) | 979-8-3503-6058-5/24/$31.00 ©2024 IEEE | DOI: 10.1109/PHM61473.2024.00049

[email protected], [email protected] [email protected]

Abstract—Applying Health Management (HM) and endurance and effectiveness in engineering applications, was
specifically Condition Monitoring (CM) techniques is necessary to discussed.
maintain the healthy condition of the system. Machine Learning
(ML), which is a subset of Artificial Intelligence (AI) where a In the same context, Engineers are armed with a powerful
computerized program is learned to perform a task by finding weapon in their toolkits with the advent of advanced and
patterns in data, can be useful for this purpose. Recently ML has sophisticated ML algorithms that possess the ability to identify
been applied for CM, and due to the huge development in cloud complex patterns from large datasets. In the last decade, these
computing and high-performance computers, the ability of the techniques have gained popularity due to the huge advancement
machine to learn patterns in a huge amount of data becomes in cloud computing and big data Tools and techniques. Using
possible. The advantage is that computers can now process big Google Trends [5], Figure 1 shows that the increase in the search
data in real time and suggest some possible actions to overcome for the term machine learning was associated with the sudden
challenges. One of the challenges is the application of ML for increase in the search for the term big data. Moreover, big data
Vibration-based CM. Unwanted vibrations in mechanical systems allows for the storage and analysis of a large amount of data
may harm the system's structural integrity and diminish the points which facilitates the process of finding patterns and trends
equipment's efficiency. In the current article, the application and in the data understudy.
significant influence of ML algorithms on the CM of mechanical
systems are reviewed and discussed. The review gives a good
understanding of the application of the recent ML techniques in
the HM strategies of mechanical vibrations in different
engineering applications.

Keywords—Health Management, Condition Monitoring,

Machine Learning, AI, Mechanical Vibrations, Supervised
Learning

I. INTRODUCTION
In the realm of engineering that is always evolving
constantly, one must seek a delicate connection between
efficiency and precision and this approach to striving for greater Figure 1: Big data is the driving force for the development in the machine
learning field
effectiveness of mechanical systems has contributed to never-
ending research for novel approaches. Health Management (HM)
These techniques offer a paradigm change in vibration
with the existence of mechanical vibrations constitutes one of
management, substituting conventional rule-based systems with
the prime challenges that engineers must confront and deal with
intelligent, adaptive ones. ML makes it easier to analyse
regularly. Unwanted oscillations in mechanical systems may
complicated vibration patterns in real-time signals with the help
jeopardize the structural integrity of the systems they infiltrate
of advanced tools and algorithms. This makes it possible for
and diminish equipment efficiency. In this aspect, the realm of
systems to self-correct and suppress unwanted oscillations [2,6].
Machine Learning (ML) and Artificial Intelligence (AI) has
Multiple methods were studied in [7], that ML could be used to
constituted a beacon of hope, and this has become apparent with
regulate vibrations, spanning predictive maintenance, active
the incorporation of machine learning in this complicated
dampening systems, and adaptive control schemes.
interaction of forces and dynamics. The combined approach of
Understanding the different types and applications of ML in
artificial intelligence in conjunction with vibration control
vibration-based health monitoring can be difficult in some cases.
collectively offers a revolutionary approach that can allow us to
prepare for and actively regulate vibrations as opposed to merely In the current article, the types, and applications of ML in
reacting to them. The objective of introducing a few of the ML vibration-based condition monitoring (CM) are reviewed and
methods and techniques and studying the significant influence discussed. This paper will shed light on the recent advancement
that ML algorithms offer on our understanding was covered in and research of using data-driven methods to mitigate the effect
[1,2,3,4]. Vibration management, which results in a new era of of unwanted oscillation in the structural integrity of a

2166-5656/24/$31.00 ©2024 IEEE 229

DOI 10.1109/PHM61473.2024.00049
Authorized licensed use limited to: Birla Institute of Technology & Science. Downloaded on January 22,2025 at 05:34:26 UTC from IEEE Xplore. Restrictions apply.
 mechanical system. ‫ܡ‬૚ (2)
‫=ܡ‬൥‫ڭ‬൩
II. TYPES OF MACHINE LEARNING FRAMEWORKS. ‫ܖܡ‬
ML algorithms serve as a training algorithm for AI system
that is able to make inferences by analysing and making Thus, the algorithm is tasked to learn a mapping or a function
deductions and inferences from available information and that can accurately predict the output for new, unseen inputs. To
historical past data without requiring explicit programming. do so, the learning process becomes an optimization process
Mitchell presented a traditional definition of the term ML in [8]. where the objective is to find the model parameters that
According to the presented definition, machine learning minimizing that objective function. For instance, in support
constitutes a class of computer systems that improve their vector machine, the objective function can be formulated as in
efficiency in tasks (T) by gaining experience (E) using a equation 3 where ࣂ is the model parameter [13].
performance measure (P). The three basic components that
comprise a machine learning algorithm are discussed as follows ௡
(3)
[1,9]: 1 ԡߠԡଶ
‫ = ݁ݒ݅ݐ݆ܾܿ݁݋‬෍ ݈‫ ݏݏ݋‬+
1) Representation: The representation defines the hypothesis ݊ 2
௜ୀଵ
space (H), that will encompass the potential models (m ‫ א‬H) that
can be used to model and also represent patterns in the dataset In SHM, supervised learning can be leveraged to assist the
(D). The different layers and neurons in an Artificial Neural prevalent issue of vibrations that is inherent in almost all
Networks (ANN) and hyperplane representation in a Decision engineering applications and mechanical systems and help
Tree or support vector machine (SVM) are examples of this. curtail them as discussed in [14] where the author used neural
network for fault diagnose in rotating machines. The machine
2) Evaluation: The evaluation component typically learning model employed in this case will then learn to
comprises the cost function that can assess the performance of generalize and anticipate vibrations under various settings by
the model. The direct approach is to compute how much the training a model on labelled datasets that correlate input features
prediction deviates from the real label in the test data. A deeper with corresponding vibration levels or frequencies. Having such
analysis could be to compute the correlation matrix where recall, a model trained on the operating data of the machine will help
precision, sensitivity, and specificity evaluate the model the device successfully reduce unwanted vibrations in real time
performance. However, these techniques require a deeper by using its predictive capability to engage damping
understanding of the model. For instance, if the model is right mechanisms or proactively alter parameters [1,14,15,16].
skewed that means more negative points, which will lead to a
higher number of false negatives and higher recall in our data. x Unsupervised learning: Unsupervised learning is a
machine learning approach that is applied to those
3) Learning algorithm: The learning algorithm is the datasets that do not have defined outputs or explicit
component that will be used for the model identification (m ‫ א‬H) labels, hence the name “unsupervised”. In this case, the
that best fits the training dataset, based on the evaluation primary objective of the method is to examine the
criterion. Examples of this algorithm such as stochastic gradient underlying structure or patterns in the data without any
decent (SGD) in neural network, SVM or Decision trees. specified labels as in the case of supervised learning.
III. APPLICATION IN VIBRATION CONTROL This kind of unsupervised learning modelling approach
typically involves tasks that include dimensionality
The three main types of machine learning approaches are reduction and clustering. Clustering aims to group or
supervised learning, unsupervised learning, and reinforcement cluster similar data points together to identify groups
learning. These approaches were introduced in [10] and can be within the data, and dimensionality reduction seeks to
summarised as following: simplify the dataset by identifying its most important
x Supervised learning: supervised learning is a very features. In situations like customer segmentation,
popular approach and has applications in almost all anomaly detection, and data compression,
domains such as image, speech, classification, and unsupervised learning is frequently used since the
regression tasks [11,12].Furthermore, it is a machine algorithm may identify hidden correlations or
learning approach in which the algorithm is provided structures in the input data. In our context of vibration
with a labelled dataset (࢞, ࢟) where x is a matrix n*m analysis and controlling them via means of machine
where n is the number of data points and m is the learning, Clustering can prove to be useful for
number of input variables. As introduced in equations classifying various vibration modes and identifying
(1) and (2). any potential anomalies that can indicate future
problems. This data will equip engineers with the
proper tools to better optimize the overall performance
‫ܠ‬૚૚ ‫ ܠ ڮ‬૚‫ܕ‬ (1) and dependability of the engineering systems and the
‫=ܠ‬൥ ‫ڭ‬ ‫ڰ‬ ‫ ڭ‬൩ devices by creating targeted control methods for certain
‫ ܠ‬૚‫ܖ‬ ‫ܕܖ ܠ ڮ‬ harmful vibration profiles [1,4,6].
x Reinforcement Learning: This is a machine learning
approach in which an agent interacts with its

230

Authorized licensed use limited to: Birla Institute of Technology & Science. Downloaded on January 22,2025 at 05:34:26 UTC from IEEE Xplore. Restrictions apply.
 surroundings and decides in iterations to accomplish a make adaptive decisions in dynamic settings to complete certain
target goal using awarded rewards and also in the tasks or objectives. A guideline in how to use deep learning
course of incorrect actions, penalties are imposed on it. combined with reinforcement learning in robotics training was
Reinforcement learning is a method in which the model introduced in [19]. Deep learning is used to evaluate the actions
learns from interaction with the environment. The that might be taken by the agent and provide generalised
model is based on Markov decision process [16]. In this formulation for the features. In the context of Vibration control,
model, there is a learning agent, states (S), policy (࣊), it can be studied by using the machine learning paradigm
set of possible actions (A), transitions (T), a reward (R) Reinforcement learning as it lends a dynamic approach to such
defined based on the desired state and a value function mechanical systems that generate large oscillations to adapt and
(࢜) . In this model, the learning agent is given a reward learn optimal control policies through interaction with their
for every possible action and the agent is not given a environment. The problem can then be framed as the agent
clear instruction to move from a state to another, but it would need to learn the optimal policy in this regard which
explores all the possible actions and assign an optimal would be to adapt and adjust the operating parameters and the
Q value which will be notated here as ࡽ‫࢙( כ‬, ࢇ) for damping strategies that will result from the observed
every action based on Bellman equation described in consequences of its actions, optimizing its response to varying
equation (4) [17].The objective of the approach is then vibration patterns. In situations that are unpredictable and
to find the best course of action that maximizes the complicated and where the probability of being subjected to
cumulative benefit over time. In mathematical external disturbances is high, this method is particularly
donations, the objective is to find ࢓ࢇ࢞ ࡽ‫࢙( כ‬, ࢇ) [18] valuable as the system learns to continuously refine its control
ࢇ
that will maximize the reward. strategy for effective vibration mitigation [1,3,4,6].
Structural health monitoring (SHM) is one widely used area
(4) where machine learning aids a lot to curtail machine vibrations.
‫ݏ( כ‬, ᇱ)
ܳ ܽ) = ෍ ܶ(‫ݏ‬, ܽ, ‫ݏ‬ (ܴ(‫ݏ‬, ܽ, ‫ ) ݏ‬+ ߛܸ (‫)) ݏ‬
ᇱ ‫כ‬ ᇱ
Recently the number of publications that includes the term
௦ ´ᇲ (machine learning and structural health monitoring) is
increasing rapidly. Figure 2 shows that since 2019 the number
of publications that include both terms almost doubled. And one
ܸ ‫ ݏ( כ‬ᇱ ) = ݉ܽ‫ݏ( כ ܳ ݔ‬, ܽ) (5) of our objectives here is introduce the framework that can be
௔
used to build a machine learning model in structural health
monitoring.
Reinforcement learning is widely used in domains such as
autonomous systems, robotics, and games, where agents must

Figure 2: The number of the publication that contained the search term (machine learning and structural health monitoring ) From 2000 to 2023 [20].

In this context of SHM, the conventionally used workflow principal component analysis (PCA). Feature selection is then
always follows a systematic method that utilizes Machine accomplished by eliminating some features or adding more
Learning. The first stage is the acquisition of sensor data from weights to others using a correlation matrix [21] or enhanced
the structure that is monitored. The next stage is to extract recursive feature elimination [22]. Finally, the selected features
crucial information from the data using multiple transformations are utilized as inputs to a Machine Learning regressor or
that include frequency, or wavelet domains. Following the classifier such as a Decision tree or Random Forest (RF) [23]
preliminary analysis, the next important step is the extraction of and support vector machine SVM [24]. These classifiers
features from the data. In this aspect, it is important to generate produce output related to the monitored structure, such as failure
an understanding of the distribution of the features using status, failure class, or Remaining Useful Life (RUL). Thus, the
statistical measures such as variance, mean, wavelet energy, problem was framed as a multi-class classification [1,2,3].
skewness, maximum values, harmonics, dispersion ratio, and

231

Authorized licensed use limited to: Birla Institute of Technology & Science. Downloaded on January 22,2025 at 05:34:26 UTC from IEEE Xplore. Restrictions apply.
 As opposed to the conventional workflow, we also have a tedious process as in certain cases of high dimensional data, the
deep learning-based workflow for SHM as shown in Figure 3. task is not trivial. The model learns to discern patterns and
The prime difference in this aspect is the automation of feature relationships within the data, allowing it to make predictions
extraction and feature steps in the deep-learning workflow. without explicit feature engineering. This end-to-end approach
These models are designed to produce predictions that are end- can enhance the efficiency of the SHM process, especially when
to-end. Thus, for such deep learning models, we directly feed dealing with large and complex datasets, as it leverages the
the raw sensor data, represented in the time or frequency domain power of neural networks to autonomously extract informative
into the deep learning model. They have the inherent ability to features for accurate predictions [1,4].
extract relevant features from the input that is fed to it while the
Both of these workflows can be summarized in the image
training process is running [1,6]. This saves the burden of
below.
additionally handcrafting the features manually which is a

Figure 3: Workflow of SHM [1].

Machine learning algorithms can also be used for the novel systems that are evolving with limited data and the
identification of anomalies resulting in higher vibrations and can possibility of emergence of anomaly over time is present. By
well in advance predict an anomaly thereby preventing them integrating ML-based anomaly detection into vibration control
from happening [1,2]. For this use case, the machine learning systems, engineers can enhance the reliability of machinery,
algorithms are fed data from the normal operational mode and reduce downtime, and optimize maintenance strategies through
thus the model learns the operating conditions, behaviours and proactive identification of potential issues [5,9,11,12].
patterns that pertain to normal operation conditions. Leveraging
these supervised learning techniques, the algorithm becomes IV. THE LACK OF DATA CHALLENGE AND GAN
adept at recognizing the expected vibration signatures associated In most engineering applications and structural health
with optimal system performance. After the model is monitoring, there is not enough data available to train the model
successfully trained, it can be deployed to real-time data streams, and as a result, a data augmentation model is needed [25]. One
wherein continuous monitoring of the produced vibrations of the data augmentation techniques is generative models.
during system operation is conducted. Whenever the system Generative models are used to generate data that is hard to be
detects any unwarranted deviation from the learned normal distinguished from the authentic data. the framework to develop
patterns, these can then be flagged as an anomaly, signifying that model was introduced by Ian Goodfellow et.al [26]. The
potential issues such as structural damage, wear, or malfunction. generative adversarial network consists of two models. The first
Using this approach makes it possible to effectively combat is known as the generator. The generator consists of one multi
mechanical vibrations and help control them. This also leads to layers perceptron neural network and its weights is ߠீ The
early detection of abnormal vibrations, enabling timely generator could be described mathematically as G (z; ߠீ ) where
intervention and preventive maintenance to mitigate potential the input to the generator is an input noise variable ‫݌‬௭ (‫)ݖ‬. The
failures [4,8]. second neural network is known as the discriminator D (x; ߠௗ ).
In cases where labels are absent, vibration control for The output of the discriminator is a scalar value that estimates
anomaly detection can be done by algorithms like Clustering the probability of x, the probability of the generator output, to be
algorithms, such as K-means, or density-based approaches, like authentic. The framework to build such a model is introduced in
Isolation Forests, which can be utilized to identify patterns in the Fig. 4. In conclusion, G is trained to generate authentic data and
vibration data. Deviations from the established cluster of the D is trained to distinguish between the real authentic and the
algorithm can be flagged as potential anomalies and can be generated data. The loss function in this process is described as
dissected further for analysis. This unsupervised approach helps follows:
in many vibration control scenarios for real-time engineering
systems, obtaining labelled data for anomaly is scarce or for

232

Authorized licensed use limited to: Birla Institute of Technology & Science. Downloaded on January 22,2025 at 05:34:26 UTC from IEEE Xplore. Restrictions apply.
 ݈‫ = ݏݏ݋‬log (1 െ D(G(z)) (6) ranking to the generated sample and then ranks that sample
based on their realness. the loss function in that model is the
Therefore, in this case a model is trained if the condition in Wasserstein distance [28]. The framework to use generative
equation 2 is met: models in structural health monitoring has some small variations
in comparison to the framework described earlier with
discriminative models. In [29] a generative model to generate
lim log (1 െ D(G(z)))  data in structure health monitoring was introduced. The
௫՜଴
vibration data was collected in [30] and the evaluation of the
The framework to train a generative adversarial network as generated data showed good promises.
described in [27] can be summarized with the the below figure . Therefore, in the differences between generative models and
discriminative models can be summarized in very few points:
Input noise variable
p(z) x Similar to any neural network training process, the
objective is to optimize the weights of each neuron to
optimize a loss function. The difference here is that the
input to the model is a noise vector as in Figure 2 and
there are two neural networks with two different
The generator structures.
x Furthermore, unlike discriminative models, the
Multi-layer Output variable from the
performance of a GAN is hard to evaluate. One Famous
perceptron generator ‫݌‬௚ approach to evaluate it is performance is the usage of
‫) ீߠ ;ݖ(ܩ‬. Fréchet inception distance, known as the FID score,
which using the means and the covariance to evaluate the
performance of the GAN as described in the following
equation:
The
discriminator ଶ
‫ݔ(ܦܫܨ‬, ݃) = ห|ߤ௫ െ ߤ௚ |หଶ + ܶ௥ (‫ܥ‬௫ + ‫ܥ‬௚ 
Multi-layer െ 2(‫ܥ‬௫ ‫ܥ‬௚ )଴.ହ
perceptron
‫ߠ ;ݖ(ܦ‬ௗ ) In equation 2 the donation x means real sample and the
donation g means generated samples. Moreover, ߤ is the means
while C is the covariance of the data. Finally,ܶ௥ is the trace of
No the matrices or the sum of all diagonal elements in the matrix.
The lower the FID score, the more identical the generated
sample with the real one [29,31]. Another Evaluation matrix is
the structural similarity index measure (SSIM) which indicate
lim log (1 െ D(G(z))) how identical are wo samples from scale to 0 to one with 1 is
௫՜଴
identical and 0 is totally different. one study conducted in
structural health monitoring for electrical machine used the
SSIM score with a creativity factor to distinguish between
Yes identical generated samples and samples that is similar but
different in some characteristics and for that they used a
The model creativity factor of 0.8 [29,32].
trained. VI. CONCLUSION
Figure 4: Generative adversarial neural network flow chart the input data is In this paper, a systematic review of the usage of the machine
feeded into a neural network known as the generator . the generator will feed learning models in structural health monitoring was introduced.
the data to the discriminator which will output a scalar that shown the In conclusion, the development of the big data tools was the
probability of the output data to be generated from the input. the purpose of the driving force for the development of the machine learning
training algorithm is to minimize a loss = lim log (1 െ D(G(z))) .
௫՜଴ models. Furthermore, Machine learning main categories are
supervised, unsupervised and reinforcement learning.
V. THE IMPLEMENTATION OF GAN IN SHM Supervised learning is the most common approach in SHM. The
GAN is hard to train. In some cases, the discriminator over reason for that is in structural health monitoring the objective is
power the generator which leads the gradient to vanish as was to classify faulty data from unfaulty data. Unsupervised learning
discussed in [26]. To overcome such a challenge, the WGAN could be used in clustering vibration modes. The Framework to
was introduced. In WGAN, the discriminator in WGAN has use machine learning in SHM was described and analysed. The
replaced the discriminator with a critic. The critic gives a distinguish between discrimination models and generative
models was described. Moreover, the lack of data challenge was

233

Authorized licensed use limited to: Birla Institute of Technology & Science. Downloaded on January 22,2025 at 05:34:26 UTC from IEEE Xplore. Restrictions apply.
 introduced and discussed and the framework to build generative [17] Bellman, R. (1957). "A Markovian Decision Process". Journal of
models were also described and analysed to tackle this challenge. Mathematics and Mechanics. 6 (5): 679–684. JSTOR 24900506
[18] R. A. Howard, Dynamic programming and Markov processes.
REFERENCES Technology Press of Massachusetts Institute of Technology, 1960.
[19] Ibarz J, Tan J, Finn C, Kalakrishnan M, Pastor P, Levine S. How to train
[1] Cunha, B. Z., Droz, C., Zine, A. M., Foulard, S., & Ichchou, M. (2023).
your robot with deep reinforcement learning: lessons we have learned.
A review of machine learning methods applied to structural dynamics and
The International Journal of Robotics Research. 2021;40(4-5):698-721.
vibroacoustic. Mechanical Systems and Signal Processing, 200, 110535.
doi:10.1177/0278364920987859
[2] Sharma, A., Abid, K., Srivastava, S., Velasquez, A. F. B., & Teodoriu, C.
[20] T. D. Team. Dimensions publication trends. [Online]. Available:
(2023). A review of torsional vibration mitigation techniques using active
https://app.dimensions.ai/discover/publication
control and machine learning strategies. Petroleum.
[21] Chen, X. and Jeong, J.C. (2007) ‘Enhanced recursive feature elimination’,
[3] Radmilovic, Z., Zehetner, J., & Watzenig, D. (2014). Vibration comfort
Sixth International Conference on Machine Learning and Applications
control for HEV based on machine learning (No. 2014-01-2091). SAE
(ICMLA 2007) [Preprint]. doi:10.1109/icmla.2007.35.
Technical Paper.
[22] Tin Kam Ho (no date) ‘Random decision forests’, Proceedings of 3rd
[4] Nick, W., Asamene, K., Bullock, G., Esterline, A., & Sundaresan, M.
International Conference on Document Analysis and Recognition
(2015). A study of machine learning techniques for detecting and
[Preprint]. doi:10.1109/icdar.1995.598994.
classifying structural damage. International Journal of Machine Learning
and Computing, 5(4), 313. [23] Sonbul, O.S. and Rashid, M. (2023) ‘Algorithms and techniques for the
structural health monitoring of bridges: Systematic Literature Review’,
[5] Google Trends : https://trends.google.com/trends/
Sensors, 23(9), p. 4230. doi:10.3390/s23094230.
[6] O. Avci, O. Abdeljaber, S. Kiranyaz, M. Hussein, M. Gabbouj, D. J.
[24] Chang, C.-M., Lin, T.-K. and Chang, C.-W. (2018) ‘Applications of
Inman, A review of vibration-based damage detection in civil structures:
neural network models for structural health monitoring based on derived
From traditional methods to machine learning and deep learning
modal properties’, Measurement, 129, pp. 457–470.
applications, Mechanical systems and signal processing 147 (2021)
doi:10.1016/j.measurement.2018.07.051.
[7] K. Vos, Z. Peng, C. Jenkins, M. R. Shahriar, P. Borghesani, W. Wang,
[25] Bansal, Ms.A., Sharma, Dr.R. and Kathuria, Dr.M. (2022) ‘A systematic
Vibration-based anomaly detection using lstm/svm approaches,
review on data scarcity problem in deep learning: Solution and
Mechanical Systems and Signal Processing 169 (2022) 108752
applications’, ACM Computing Surveys, 54(10s), pp. 1–29.
[8] Mitchell, T. (1997), Machine Learning , McGraw-Hill Education . doi:10.1145/3502287.
[9] X. W. Ye, T. Jin, C. B. Yun, A review on deep learning-based structural [26] Goodfellow, I. et al. (2020) ‘Generative Adversarial Networks’,
health monitoring of civil infrastructures, Smart Structures and Systems Communications of the ACM, 63(11), pp. 139–144. doi:10.1145/3422622.
24 (2019) 567–585
[27] Regenwetter, L., Nobari, A.H. and Ahmed, F. (2022) ‘Deep generative
[10] I. H. Sarker, ‘Machine Learning: Algorithms, Real-World Applications models in Engineering Design: A Review’, Journal of Mechanical Design,
and Research Directions’, SN Comput. Sci., vol. 2, no. 3, p. 160, Mar. 144(7). doi:10.1115/1.4053859.
2021, doi: 10.1007/s42979-021-00592-x
[28] M. Arjovsky, S. Chintala, and L. Bottou, ‘Wasserstein Generative
[11] He, K. et al. (2016) ‘Deep residual learning for image recognition’, 2016 Adversarial Networks’, in Proceedings of the 34th International
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Conference on Machine Learning, D. Precup and Y. W. Teh, Eds., in
[Preprint]. doi:10.1109/cvpr.2016.90. Proceedings of Machine Learning Research, vol. 70. PMLR, Aug. 2017,
[12] Sepulveda, N. E., & Sinha, J. (2020). Parameter optimisation in the pp. 214–223.
vibration-based machine learning model for accurate and reliable faults [29] Luleci, F., Catbas, F.N. and Avci, O. (2022) ‘Generative adversarial
diagnosis in rotating machines. Machines, 8(4), 66. networks for Data Generation in Structural Health Monitoring’, Frontiers
[13] J. Cervantes, F. Garcia-Lamont, L. Rodríguez-Mazahua, and A. Lopez, in Built Environment, 8. doi:10.3389/fbuil.2022.816644.
“A comprehensive survey on support vector machine classification: [30] Abdeljaber, O., and Avci, O. (2016). Nonparametric Structural Damage
Applications, challenges and trends,” Neurocomputing, vol. 408, pp. 189– Detection Algorithm for Ambient Vibration Response: Utilizing Artificial
215, Sep. 2020, doi: 10.1016/j.neucom.2019.10.118. Neural Networks and Self-Organizing Maps. J. Architectural Eng. 22 (2),
[14] Y. Bel-Hadj, W. Weijtjens, Anomaly detection in vibration signals for 04016004. doi:10.1061/(ASCE)AE.1943-5568.0000205
structural health monitoring of an offshore wind tur36 bine, in: European [31] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter,
Workshop on Structural Health Monitoring, Springer, 2023, pp. 348–358 ‘GANs Trained by a Two Time-Scale Update Rule Converge to a Local
[15] (VNL6Ü<ÕOGÕUÕP9LEUDWLRQFRntrol of vehicle active suspension system Nash Equilibrium’, in Advances in Neural Information Processing
using a new robust neural network control system, Simulation Modelling Systems, I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S.
Practice and Theory 17 Vishwanathan, and R. Garnett, Eds., Curran Associates, Inc., 2017.
[16] Lu, W. et al. (2016) ‘Data Correlation Analysis for optimal sensor [32] Sabir, R. et al. (2021) ‘Signal generation using 1d deep convolutional
placement using a bond energy algorithm’, Measurement, 91, pp. 509– generative adversarial networks for fault diagnosis of Electrical
518. doi:10.1016/j.measurement.2016.05.089. Machines’, 2020 25th International Conference on Pattern Recognition
(ICPR) [Preprint]. doi:10.1109/icpr48806.2021.9413119.

234

Authorized licensed use limited to: Birla Institute of Technology & Science. Downloaded on January 22,2025 at 05:34:26 UTC from IEEE Xplore. Restrictions apply.