Predictive maintenance of electromechanical systems based on enhanced generative adversarial neural network with convolutional neural network

IAES International Journal of Artificial Intelligence (IJ-AI)
Vol. 12, No. 4, December 2023, pp. 1704~1712
ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i4.pp1704-1712  1704
Journal homepage: http://ijai.iaescore.com
Predictive maintenance of electromechanical systems based on
enhanced generative adversarial neural network with
convolutional neural network
Azhar Muneer Abood, Ahmed Raoof Nasser, Huthaifa Al-Khazraji
Department of Control and Systems Engineering, University of Technology, Baghdad, Iraq
Article Info ABSTRACT
Article history:
Received Jun 29, 2022
Revised Dec 23, 2022
Accepted Jan 10, 2023
Predictive maintenance (PdM) is a cost-cutting method that involves avoiding
breakdowns and production losses. Deep learning (DL) algorithms can be
used for defect prediction and diagnostics due to the huge amount of data
generated by the integration of analog and digital systems in manufacturing
operations. To improve the predictive maintenance strategy, this study uses a
hybrid of the convolutional neural network (CNN) and conditional generative
adversarial neural network (CGAN) model. The proposed CNN-CGAN
algorithm improves forecast accuracy while substantially reducing model
complexity. A comparison with standalone CGAN utilizing a public dataset
is performed to evaluate the proposed model. The results show that the
proposed CNN-CGAN model outperforms the conditional GAN (CGAN) in
terms of prediction accuracy. The average F-Score is increased from 97.625%
for the CGAN to 100% for the CNN-CGAN.
Keywords:
Convolutional neural network
Deep learning
Generative adversarial neural
network
Predictive maintenance
This is an open access article under the CC BY-SA license.
Corresponding Author:
Azhar Muneer Abood
Department of Control and Systems Engineering, University of Technology
Baghdad, Iraq
Email: cse.20.09@grad.uotechnology.edu.iq
1. INTRODUCTION
The term predictive is defined as a state or behavior that will occur in the future. The task of
maintaining a machine is necessary to keep it in good working order. Therefore, predictive maintenance (PdM)
is defined as a way to predict the future failure of the machine’s tool before it fails [1]. PdM became a key role
in growing the productivity and the profitability of the industrial system. For this, it has obtained wide attention
in the last years in research.
Due to condition monitoring of all industrial equipment, get together with deep learning methods, the
maintenance task has been enhanced in modern production systems [2]. Besides, data acquisition, data
collected by smart sensors are available nowadays to make a file estimation and prediction of the current health
condition and machine tools [3]. Big data is not just the size of the collected data, but it also contains the
properties, the variety, and the velocity of data. The essential pattern of overall data becomes a major concern
for companies to investigate the utility of big data analytics. The main goal of large data analysis is to define
the attributes of data with the aim to derive patterns and connections in the data. In addition, big data analytics
aims to find the data functions that are descriptive such as classification, clustering, association, and logistic
regression analysis [4], [5].
Deep learning (DL) is more similar to the human brain. It is a subgroup of machine learning methods,
which is learning the many levels of representations of data with different levels of abstraction at each
stage [6]. DL is an efficient data feature extraction algorithm because it can overcome the problem of extracting

Int J Artif Intell ISSN: 2252-8938 
Predictive maintenance of electromechanical systems based on enhanced … (Azhar Muneer Abood)
1705
features that are involved in nonlinear big data by shallow learning. DL may be supervised or unsupervised.
The important features of data are extracted by using multiple levels of nonlinear processing units. The input
to the next layer is provided from the output of the current layer. The method of stochastic gradient descent is
used by the backpropagation algorithm to reach an ideal in the training set. Many types of deep structures are
used such as recurrent neural network (RNN), long short term memory (LSTM), convolutional neural network
(CNN), and deep belief network (DBN) [7], [8].
When compared to a regular or fully connected neural network (FCNN), the deep neural network has
a different structure. A straightforward neural network can be modified to a deep neural network by adding
two or more layers between the input and output layers [9], [10]. Conventional computing hardware that
depends on the central processing units (CPU) is not appropriate to handle the multilayer architecture because
this architecture deals with countless links between the layers. However, graphics processing unit (GPU)
techniques overcome this issue by executing and detecting multiple features and higher-order feature
relationships. Consequently, adding more layers between input and output in deep architectures was made
possible by the use of GPU processing and the presence of a sizable training set of data [11].
In this paper, a modified version of generative adversarial neural network (GAN) named conditional
GAN (CGAN) is introduced to predict the multiclass fault of an electromechanical system (motor) in the early
stage using a data set of asynchronous common motor fault (ACMF) in a normal (healthy) state and seven
abnormal (unhealthy) states. Next, CNN is used as extraction features of the training dataset of ACMF and
then passes it to CGAN. Finally, a comparison between the CGAN model and the hybrid CNN-CGAN model
is performed to evaluate the proposed model.
2. PROPOSED DEEP LEARNING ARCHITECTURE
This section presents the theoretical framework of the proposed deep learning model. The first
subsection explains the GAN model. The next subsection introduces the CNN model.
2.1. Generative adversarial neural network
GANs are a come-up approach for both semi-supervised and unsupervised learning. It was proposed
in 2014 by modeling high-dimensional distributions of data. There are various types of GAN networks like
CGAN, Cycle GAN, Wasserstein GAN, and Vanilla GAN. GANs can be described by the training of two
networks in competition with each other. The first network is known as a fake artist and the second as an art
expert. In the GAN literature, the fake called the generator (G) generates fakes, with the purpose of making
realistic data. The expert, called the discriminator (D) gets both fake and real data, to distinguish between them.
The G and D are trained simultaneously, and in competition with each other [12].
Definitely, the G has no direct connection to the real data, the only path it learns via interaction with
the D. On the other hand, D has reached both the fake and real samples taken from the stack of real data. The
mistaken signal to D is supplied by determining whether the data came from the real stack or from G. G received
the error from D, and then this error is used to train and make forgeries have better quality. The implementation
of the network includes the G and D by multilayer consisting of convolutional and/or fully connected layers.
The G and D network is not necessary to be directly invertible and must be differentiable. The G
network is an analysis of some description space, denoted a (latent space), to the space of the data [13].
Basically, in the GAN model, the D network may be similarly described as a function that maps from data to
an eventuality that the data is from the real data allocation, rather than the G allocation: D: (Dx) (0 or 1). For a
fixed G, the D may be trained to classify fault as either being from the training data (real, refer to 1) or from a
fixed generator (fake, refer to 0). The G may keep being learned so as to lower the accuracy of the D when the
D is optimal and it may be fixed. If the generator allocation is able to reach the real data allocation perfectly,
then the D will be maximally confused, predicting 50% for all inputs [14]. Figure 1 shows the GAN
architectures.
Depending on the binomial zero-sum game theory, networks of different types could exist for the G
and D architectures, such as a fully-connected layer, CNN, and autoencoder. Typically G and D are modeled
using nonlinear mapping equations. Through training, D tries to give a high likelihood based on real data, and
give data from G a low probability. Conversely, G creates false data while learning the distribution of real data
in order to fool D [15]. In this competitive learning, the enhancement of G and D efficiency can be introduced.
The minimax two-player game can be expressed by (1),
𝑚𝑖𝑛𝐺𝑚𝑎𝑥𝐷𝑉(𝐷, 𝐺) = 𝐸𝑥~𝑝𝑑𝑎𝑡𝑎 (𝑥)[𝑙𝑜𝑔𝐷(𝑥)] + 𝐸𝑧~𝑝𝑧 (𝑧) [𝑙𝑜𝑔 (1 − D(G(z)))] (1)

 ISSN: 2252-8938
Int J Artif Intell, Vol. 12, No. 4, December 2023: 1704-1712
1706
Signify the distribution of real data's probability specified in the data space X and the probability distribution
for the latent space Z pdata(x), pz(z), respectively. The binary cross-entropy function V (G, D) is often utilized
in binary classification issues [16].
Figure 1. GAN architecture
2.2. Convolutional neural network
A CNN is a type of deep neural network that have multi hidden layers that are connected. The main
advantage of CNN is to extraction feature of big data. It includes an input layer, convolutional layers, pooling
also known (as the subsampling layer), and output layers [17], [18]. The feature map is generated by using the
function in the convolution layer that is applied to input data, therefore the convolution layer is a mathematical
linear operation between matrixes that are using the kernel. The pooling layer decreases the dimension of the
extracted feature map. Many procedures are for pooling. The convolution and pooling layer represents the
feature extraction of row data. The main features are extracted by repeating the convolutional and the pooling
layer. After that, the extracted features are passed to the fully connected layer [19], [20]. Figure 2 presents the
CNN structures.
CNN executes a series of convolutions between the filters and the input signal in its typical technique.
A signal's weighted average as an input Xi is described as the convolution procedure in (2),
Si=Ki * Xi (2)
where Si stands for the i-th feature map, and a weighting factor, also known as a filter or kernel, is denoted
Ki [21].
Figure 2. CNN structure

1707
3. CASE STUDY
To improve the PdM strategy in an electromechanical system, this section presents an application of
the CGAN and CNN-CGAN DL algorithms to the motor with multi-fault diagnosis. A public dataset is utilized
to evaluate the proposed CNN-CGAN model. The comparison of the proposed model with the standalone
CGAN model based on Precision, Recall, and F-Score metrics is introduced in this section.
3.1. Dataset
Asynchronous motor common fault (AMCF) dataset is applied for standalone CGAN and CNN-
CGAN DL algorithms that is obtained from the Zenodo website. The dataset is generated by collecting the
vibration signal of eight motors with one healthy data where the example of the healthy (normal) signal is
shown in Figure 3(a) and seven unhealthy states where the example of the unhealthy (fault) signal is shown in
Figure 3(b). It consists of 8,000 overall instances (the normal data and other seven faults) each of them has
1,000 rows with each row 1,024 of vibration signal and 1 column has the label of the fault so it is becoming
(8,000×1,025). Table 1 shows the labels of each states [22].
(a) (b)
Figure 3. Vibration signal, (a) normal signal and (b) fault signal
Table 1. Types of the states of the AMCF
Label state State descriptions
0 Normal
1 Short circuit of 2 turns (SC2T)
4 Air-gap eccentricity (AE)
5 Rotor bar broken (RBB)
6 Bearing cage broken (BCB)
7 Bearing abrasion fault (BAF)
3.2. Predictive maintenance for CGAN
The CGAN is an extension of the GAN for supervised methods. It is different from general GAN by
extra information conditions (C) to manage the data production process in a supervised manner, like the class
label of data that is dependent on it for the classification of multiclass faults diagnosis. The conditional (class
label) is feeding into both the discriminator and generator as additional input data. The discriminator accepts
samples and the information vector C to differentiate fake samples given C, while the generator takes not just
a latent vector Z but also an extra information vector C. CGAN can regulate the number of Samples generated
in this way, which is impossible with normal GAN [16], [23]. Figure 4 illustrates the CGAN architecture. There
are different structures of CGAN such as multilayer perceptron, deep convolution neural, and autoencoder. The
deep convolution CGAN (DCCGAN) makes substantial contributions to GAN by adding CNN's convolution
layer and has more stability than it. The algorithm of CGAN that is used in this work is as:
− Input data.
− Split data into training data and testing data.
− Discriminator training=false.
− Generate noise (latent dim=m).
− Generated the fake data with (class label=c) in the generator network.

 ISSN: 2252-8938
1708
− Train the discriminator network with real data & (class label=c).
− Train both fake data & real data at the same time with a discriminator network.
Figure 4. CGAN architectures
3.3. Predictive maintenance for CNN-CGAN
The proposed structure of the hybrid CNN-CGAN is made by adding the CNN method to CGAN.
This work enhances the operation of CGAN so that it overcomes its limitations to classify the multi-label of
data input. CNN strategies of deep learning are good and strong for the extraction of feature maps of row data.
Therefore, the combination of two of these methods makes this model more effective and powerful for the
classification of multiclass fault prediction. Figure 5 illustrates the general architecture of the CNN-CGAN
hybrid model for PdM.
Figure 5. Architecture of the proposed CNN-CGAN model
In the D network, there are two dense layers. One for distinguishing real and fake output data contains
one output which is using binary cross-entropy loss function and sigmoid nonlinear activation function. The
other layer is for the classification of the eight fault classes of the motor machine, therefor the
sparse_categorical_crossentropy loss function and softmax nonlinear activation function are applied to classify
these faults. The optimizer in the D network is Adam.
The second network is G which is used to generate fake data. It begins in reverse order compared with
the D network. The dense layer contains a 128-sized filter which uses the rectified linear unit (ReLU) as an
activation function with an 8×8 window size for generating data as a result, the UpSampling2D layer doubles
the input dimensions to 16×16 window size, then it is as input to a 2-dimensional convolution layer with filter
size 128 and a kernel size of 3. Then a batch-normalization (BN) is applied. The UpSampling2D layer now
doubles the dimensions to 32×32 feature, then passes to the last two-layer having the 128 and 1 filter size with
a kernel size of 3. The binary cross-entropy loss function and (tanh) nonlinear activation function is used. The
optimizer in the D network is Adam. Figure 6 illustrates the details of the proposed model for hybrid
CNN-CGAN.

1709
Figure 6. Description of the proposed CNN-CGAN model
3.4. Training and evaluation of the CNN-CGAN PdM model
The dataset is split into two-part the training set of 70% and the testing set of 30%. The training set is
used to train the proposed CNN-CGAN model for PdM. The model is built on the TensorFlow platform [24]
and uses GPU acceleration. The model's hyper-parameters are chosen based on trial and error and used the best
values illustrated in Table 2.
Table 2. The D and G network hyper parameters of the CNN-CGAN model
Hyperparameter D G
Learning rate 0.0002 0.0002
Activation function Leaky ReLU/softmax/sigmoid/ ReLU/ tanh
Loss function Binary_crossentropy /sparse_categorical_crossentropy Binary_crossentropy
Dropout 0.25 0.25
Batch normalization 8 8
Epoch 15,000 15,000
Batch size 32 32
Optimizer Adam Adam
No. of convolution layer 4 3
No. of layer 8 7
No. of max-pooling layer 2 __
After the CNN-CGAN model has been trained, it is assessed using the testing dataset that was
previously produced. The model's performance is assessed using the model accuracy metrics Precision, Recall,
and F-score as described in (3)-(5). These facts are derived from the following of the confusion matrix in
Table 3,
− Precision is a metric that measures how exact a model is and can be determined as,

 ISSN: 2252-8938
1710
𝑃𝑟𝑒𝑠𝑠𝑖𝑜𝑛 = 𝑇𝑃/(𝑇𝑃 + 𝐹𝑃) (3)
− Recall is a metric that gauges the model's completeness and the calculation is,
𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃/(𝑇𝑃 + 𝐹𝑁)
(4)
− F-Score is calculated by weighing Recall and Precision. It's commonly applied to assess model accuracy
in situations where training dataset is unbalanced. The F-Score is given by [25],
𝐹 − 𝑠𝑐𝑜𝑟𝑒 = (2 ∗ 𝑃𝑟𝑒𝑠𝑠𝑖𝑜𝑛 ∗ 𝑅𝑒𝑐𝑎𝑙𝑙)/(𝑃𝑟𝑒𝑠𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙) (5)
Where, respectively, the letters true positive (TP), false positive (FP), false negative (FN), and true negative
(TN).
Table 3. Matrix of confusion
A true prediction A false prediction
Real true TP FN
Real false FP TN
Table 4 shows the performance metrics that were obtained using the test dataset to evaluate the CNN-
CGAN PdM model. Because the dataset used to train the model has numerous classes for prediction, the
evaluation metrics for each class were generated separately, and the model's overall effectiveness was
calculated by averaging all of the results. The regular CGAN deep learning approach is trained and assessed
with the CNN-CGAN model being applied to the same dataset to investigate the effectiveness of the proposed
deep learning method for PdM. Table 5 shows the evaluation results for the CGAN PdM model. Figure 7 shows
the confusion matrix of both the CNN-CGAN model as shown in Figure 7(a) and CGAN model as shown in
Figure 7(b).
Table 4. Evaluation of CNN-CGAN results
Failure type Precision Recall F-Score (%)
0 1.00 1.00 1.00
1 1.00 1.00 1.00
2 1.00 1.00 1.00
3 1.00 1.00 1.00
4 1.00 1.00 1.00
5 1.00 1.00 1.00
6 1.00 1.00 1.00
7 1.00 1.00 1.00
Average 1.00
Table 5. Evaluation of CGAN results
Failure type Precision Recall F-Score (%)
0 1.00 1.00 1.00
1 1.00 0.99 1.00
2 0.97 1.00 0.98
3 1.00 0.98 0.99
4 1.00 1.00 1.00
5 1.00 1.00 1.00
6 1.00 0.84 0.91
7 0.86 1.00 0.93
Average 0.97625
(a) (b)
Figure 7. The confusion matrix of, (a) CNN- CGAN model and (b) CGAN model

1711
Considering the assessment findings of the PdM CNN-CGAN model, it can be seen that the model
has a high f score accuracy of 100% in predicting the motor fault state in terms of stating which component is
broken, or will the machine continue to function without any problems. When the hybrid model CNN-CGAN
is compared to the CGAN model for PdM, it is seen that the hybrid model improves the typical prediction
accuracy of 2.375%. In terms of f score accuracy, the CGAN and hybrid model CNN-CGAN results are
compared to those of comparable works. Table 6 shows the outcomes of each model. The outcomes reveal that
the suggested hybrid model CNN-CGAN outperformed other relevant models in terms of fault prediction
accuracy.
Table 6. PdM comparison of CNN-CGAN, CGAN, and related works
Algorithm F-Score on average (%)
LSTM 92.45
CNN 94.24
CGAN 97.625
CNN- CGAN 100
4. CONCLUSION
The application of data-driven approaches in the field of PdM, such as DL algorithms has been made
possible by the use of sensor technologies to acquire information on the status of production equipment. To
construct a model in PdM of many related component production systems, a hybrid CNN-CGAN DL technique
is provided in this paper. The proposed algorithm mixes the reliability of a CNN network with the ability of
CGAN to classify the data as real and fake. The CNN-CGAN DL model is evaluated using an available real-
world industrial motor muli_fault data. The proposed model obtains an average of 100% Precision, 100%
Recall, and 100% F-Score on the testing dataset, according to the results. When compared to the normal CGAN
model, the proposed CNN-CGAN model improves prediction accuracy by 2.375%. The accuracy of the hybrid
model exceeds that of the real PdM works as well.
REFERENCES
[1] S. Sajid, A. Haleem, S. Bahl, M. Javaid, T. Goyal, and M. Mittal, “Data science applications for predictive maintenance and
materials science in context to Industry 4.0,” Materials Today: Proceedings, vol. 45, pp. 4898–4905, 2021, doi:
10.1016/j.matpr.2021.01.357.
[2] X. Bampoula, G. Siaterlis, N. Nikolakis, and K. Alexopoulos, “A deep learning model for predictive maintenance in Cyber-Physical
production Systems Using LSTM autoencoders,” Sensors, vol. 21, no. 3, p. 972, 2021, doi: 10.3390/s21030972.
[3] S. Zhai, B. Gehring, and G. Reinhart, “Enabling predictive maintenance integrated production scheduling by operation-specific
health prognostics with generative deep learning,” Journal of Manufacturing Systems, vol. 61, pp. 830–855, 2021, doi:
10.1016/j.jmsy.2021.02.006.
[4] C. K. M. Lee, Y. Cao, and K. H. Ng, “Big data analytics for predictive maintenance strategies.” pp. 50–74, 2017, doi: 10.4018/978-
1-5225-0956-1.ch004.
[5] H. S. Abdullah and S. B. Neama, “Proposed business intelligence system through big data,” Engineering and Technology Journal,
vol. 34, 2016, doi: 10.30684/etj.34.4B.10.
[6] A. M. Mahmood, A. Al-Yasiri, and O. Y. Alani, “Cognitive neural network delay predictor for high speed mobility in 5G C-RAN
cellular networks,” in 2018 IEEE 5G World Forum (5GWF), 2018, pp. 93–98, doi: 10.1109/5GWF.2018.8516715.
[7] F. Zhou, S. Yang, H. Fujita, D. Chen, and C. Wen, “Deep learning fault diagnosis method based on global optimization GAN for
unbalanced data,” Knowledge-Based Systems, vol. 187, p. 104837, 2020, doi: 10.1016/j.knosys.2019.07.008.
[8] A. Q. Albayati, A. S. Al-Araji, and S. H. Ameen, “A method of deep learning tackles sentiment analysis problem in Arabic texts,”
Iraqi Journal of Computers, Communications, Control and Systems engineering, vol. 20, no. 4, pp. 9–20, 2020, doi:
10.33103/uot.ijccce.20.4.2.
[9] A. Khamparia and K. M. Singh, “A systematic review on deep learning architectures and applications,” Expert Systems, vol. 36, no.
3, p. 12400, 2019, doi: 10.1111/exsy.12400.
[10] Z. A. Mohammed, M. N. Abdullah, and I. H. Al Hussaini, “Predicting incident duration based on machine learning methods,” Iraqi
Journal of Computers, Communications, Control and Systems Engineering, vol. 21, no. 1, pp. 1–15, 2021, [Online]. Available:
https://ijccce.uotechnology.edu.iq/article_168352.html.
[11] Y. Yamato, “Study and evaluation of automatic GPU offloading method from various language applications,” International Journal
of Parallel, Emergent and Distributed Systems, vol. 37, no. 1, pp. 22–39, 2022, doi: 10.1080/17445760.2021.1971666.
[12] A. Kusiak, “Convolutional and generative adversarial neural networks in manufacturing,” International Journal of Production
Research, vol. 58, no. 5, pp. 1594–1604, 2020, doi: 10.1080/00207543.2019.1662133.
[13] X. Wang, X. Liu, P. Song, Y. Li, and Y. Qie, “A novel deep learning model for mechanical rotating parts fault diagnosis based on
optimal transport and generative adversarial networks,” Actuators, vol. 10, no. 7, p. 146, 2021, doi: 10.3390/act10070146.
[14] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and A. A. Bharath, “Generative adversarial networks: An
Overview,” IEEE Signal Processing Magazine, vol. 35, no. 1, pp. 53–65, 2018, doi: 10.1109/MSP.2017.2765202.
[15] G. Ghosheh, J. Li, and T. Zhu, “A review of generative adversarial networks for rlectronic health records: Applications, evaluation
measures and data sources,” 2022.
[16] Y. Hong, U. Hwang, J. Yoo, and S. Yoon, “How generative adversarial networks and their variants work,” ACM Computing Surveys
(CSUR), vol. 52, no. 1, pp. 1–43, 2020, doi: 10.1145/3301282.

 ISSN: 2252-8938
1712
[17] B. K. O. C. Alwawi and L. H. Abood, “Convolution neural network and histogram equalization for COVID-19 diagnosis system,”
Indonesian Journal of Electrical Engineering and Computer Science, vol. 24, no. 1, pp. 420–427, 2021, doi:
10.11591/IJEECS.V24.I1.PP420-427.
[18] A. Nasser and H. Al-Khazraji, “A hybrid of convolutional neural network and long short-term memory network approach to
predictive maintenance,” International Journal of Electrical and Computer Engineering (IJECE), vol. 12, no. 1, 2022, doi:
10.11591/ijece.v12i1.pp721-730.
[19] J.-H. Lee, J.-H. Pack, and I.-S. Lee, “Fault diagnosis of induction motor using convolutional neural network,” Applied Sciences,
vol. 9, no. 15, p. 2950, 2019, doi: 10.3390/app9152950.
[20] S. A. H. Alazawi and M. N. Abdullah, “Online failure prediction model for open-source software system based on CNN.” pp. 599–
608, 2021, doi: 10.1007/978-981-16-2094-2_70.
[21] M. Jiménez-Guarneros, J. Grande-Barreto, and J. de Jesus Rangel-Magdaleno, “Multiclass incremental learning for fault diagnosis
in induction motors Using fine-tuning with a memory of exemplars and nearest centroid classifier,” Shock and Vibration, vol. 2021,
pp. 1–12, 2021, doi: 10.1155/2021/6627740.
[22] Liang, “The motor fault daignosis experiment dataset,” 2019, doi: 10.5281/ZENODO.3553755.
[23] J. Luo, J. Huang, and H. Li, “A case study of conditional deep convolutional generative adversarial networks in machine fault
diagnosis,” Journal of Intelligent Manufacturing, vol. 32, no. 2, pp. 407–425, 2021, doi: 10.1007/s10845-020-01579-w.
[24] M. Abadi et al., “TensorFlow: A system for large-scale machine learning,” In 12th USENIX symposium on operating systems design
and implementation, pp. 265–283, 2016, [Online]. Available: https://tensorflow.org.
[25] P. K. Mallick, S. Mishra, and G.-S. Chae, “Digital media news categorization using Bernoulli document model for web content
convergence,” Personal and Ubiquitous Computing, pp. 1–16, 2020, doi: 10.1007/s00779-020-01461-9.
BIOGRAPHIES OF AUTHORS
Azhar M. Abood was born in Bagdad, Iraq in 1983. She was given her B.Sc.
degree in Computer and Software Engineering Department, Al-Mustansiriya University,
Baghdad Iraq in 2005. Currently, she is a researcher at the University of Technology Control
and Systems Engineering to get an MSc degree. She works at the University of Karbala,
college of Engineering, Electrical and Electronic Eng. Department, laboratory unit. She can
be contacted at email: cse.20.09@grad.uotechnology.edu.iq.
Ahmed Nasser was born in Bagdad, Iraq in 1984. He got his BSc degree from
the University of Technology Control and Systems Eng. Faculty, Baghdad Iraq in 2006, MSc
degree in Computer Eng. from Istanbul University, Istanbul Turkey in 2012, PhD degree
from Hacettepe University, Ankara Turkey in Computer Engineering in 2018. Currently, he
is a Lecturer in the Control and Systems Engineering Department at the University of
Technology. His current interest is in “Data Mining”, “Natural Language Processing” and
“Machine and Deep Learning”. He can be contacted at email:
ahmed.r.nasser@uotechnology.edu.iq.
Huthaifa Al-Khazraji was born in Baghdad, Iraq in 1984. He received his B.S
degree in Control and Systems Engineering Department at the University of Technology,
Iraq in 2006. And His M.S Degree in Industrial Engineering and Management at Politecnico
di Torino University, Italy in 2010. From 2010 to 2014 he was Assistant Lecturer at the
Control and Systems Engineering Department at the University of Technology, Iraq. He was
accepted in Central Queensland University in Australia as a PhD student in March 2015. He
received his PhD degree in 2019. Currently, he is a Lecturer in the Control and Systems
Engineering Department at the University of Technology since 2019. His area of interest:
Control Engineering, Industrial Engineering, and Optimization. He can be contacted at email:
60141@uotechnology.edu.iq.

Predictive maintenance of electromechanical systems based on enhanced generative adversarial neural network with convolutional neural network

More Related Content

Similar to Predictive maintenance of electromechanical systems based on enhanced generative adversarial neural network with convolutional neural network (20)

More from IAESIJAI (20)

Recently uploaded (20)

Predictive maintenance of electromechanical systems based on enhanced generative adversarial neural network with convolutional neural network