International Conference on Innovations in Energy Engineering & Cleaner Production IEE CP22

An efficient model for induction motor fault

detection using a deep transfer learning network

Alasmer Ibrahim Fatih Anayi Michael Packianather

Wolfson Centre for Magnetics Wolfson Centre for Magnetics High-Value Manufacturing Group
Cardiff University Cardiff University Cardiff University
Cardiff CF24 3AA, UK Cardiff CF24 3AA, UK Cardiff CF24 3AA, UK
[email protected] [email protected] [email protected]

motor lifetime can be extended, and maintenance cost can be

minimized [1]. Many types of defects may occur in an induction
Abstract motor, which can be categorized under electrical and mechanical
defects which are caused by abrasion, unbalanced loads, electrical
The reliability and availability of induction motors is stress. Mechanical failure such as the bearing fault is the most
significant which can be achieved by monitoring the frequent fault in the induction motor which represents around 53%
performance of the motor regularly in the industry. of the motor faults [2]. Electrical fault such as the rotor fault appears
Knowledge-based approaches can be able efficiently to deal with with 10% of the motor failure [3]. In addition, stator fault stator
the sensor data for ensuring the reliability with high motor failure occurs typically at a rate of 38% of motor failure [4]. In recent
performance. Recently, deep learning networks based on years, many researchers have focused their attention on fault
machine learning structures have provided an accurate and diagnostics in different science fields. As a common type of fault
faster framework for fault diagnosis by ignoring feature diagnosis, data-driven fault diagnostic technique which has attracted
extraction process. However, training a deep convolutional the researcher’s attention may construct failure modes using the
neural network (CNN) is complex and time-consuming historical data of the application without the need of signal
procedure. For this reason, this paper proposes a novel deep symptoms [5]. This can make it particularly adapted for complex
learning procedure for fault diagnosis using thermal images systems with a large amount of data [6]. With the faster growth of
data of the induction motor applying residual neural network smart manufacturing, the amount of data produced by machines and
with 50 convolutional layers as feature extraction. The pre- devices is further increased and easier to be collected. The ability to
trained deep convolutional (ResNet-50) of the transfer learning learn about huge amount of historical data is the most important
is trained on ImageNet based weight. This work includes the aspect of the data-driven fault diagnostic technique [7]. A number
effect of data augmentation for enhancing the performance of of methods have applied in this field such as K-Nearest Neighbor
the proposed model and ensuring its robustness for fault (KNN), support vector machine (SVM), Random Forest algorithm
diagnosis. Firstly, the collected images are pre-processed resized (RF), Invasive Weed Optimization algorithm (IWO), and artificial
as input datatype of Resnet-50 network. Next, transfer leaning neural networks (ANN). In [3] KNN was suggested to build a
model based on convolutional neural network (ResNet-50) diagnostic system to detect the degradation of the bearing and the
structure is built to process the prepared images. Lastly, result was satisfactory based on the diagnostic performance. SVM
classifying the prepared images based on the related conditions is proposed in [8] to build a reliable fault diagnosis scheme for
of the induction motor. The experimental result shows that the incipient low speed rolling elements bearing failure. The
proposed model has achieved an accuracy of 99.98%. The Experimental results reported that the proposed model has achieved
presented model has further compared with recent deep a highest classification accuracy of 98.4%. In [9] RF was intended
learning applications, and it has proved its robustness in fault to achieve a novel hybrid approach for fault diagnosis of rolling
diagnosis. bearings. The obtained results showed that the proposed method
reached a classification accuracy rate of 88.23%. In [10] a new
Keywords: Thermal mages, ResNet-50 model, pre-trained diagnostic model is proposed applying the current and the vibration
model, fault diagnosis. signals. The model combines the invasive weed optimizer with three
different machine learning algorithms. The results have confirmed
I. INTRODUCTION that the performance of the proposed model is satisfactory. A novel
It is possible for several problems to arise unexpectedly in the model was proposed in [11] for diagnosing bearing faults. The
induction motors and causing a failure during even the normal experimental results showed that the proposed scheme is a
operation which can lead to production loss. However, by applying successful framework. As the aforementioned machine learning
condition monitoring to check the behavior of the motor. So, the methods are comprised of feature extraction and feature selection
processes [12]. It is however difficult to extract discriminative
features to build a robust classification model [13]. Therefore, Deep
IEECP’22, July 21-22, 2022, Oxford, United Kingdom learning (DL) has developed as a new area of research in the
Copyright: © 2022 by the author(s) | Licensee IEECP – SCI-INDEX machine-learning science in order to address the issues mentioned
Open access article distributed under the terms and conditions of CC BY license.
https://creativecommons.org/licenses/by/4.0/
above which it is capable of autonomously learning the
representation attributes from raw data [14]. The use of DL

1
International Conference on Innovations in Energy Engineering & Cleaner Production IEE CP22

approaches in the field of fault diagnosis has become increasingly in [30] and its results have indicated that more accurate and reliable
popular such as deep belief network (DBN) and convolutional than previous approaches. In [31] a CNN model-based approach for
neural network (CNN) [15]. The use of such DL network show great fault diagnosis was suggested for rotating machinery. The model
potential for fault diagnosis can achieve the reduction of the compared to traditional techniques that rely on manual feature
handcrafted features impact created by feature extraction extraction, the results have demonstrated that the suggested method
algorithms. However, fault identification requires a minimal number provides superior diagnostic performance. As reported by the
of labelled samples which limit the final prediction accuracies. In aforementioned methods, there are not any published studies using
addition, DL models can include up to 5 hidden layers [16]. So, it is to ResNet-50 model-based CNN tested on thermal images data of
difficult to train deep CNN models without a huge amount of the induction motor. Therefore, a novel intelligent model for fault
training datasets to train deep CNN models. Several studies have diagnosis is proposed in this work applying the pre-trained ResNet-
built deep CNN models by combining transfer learning techniques 50 model with an adjusted densely connected classifier. The
(TL) on ImageNet [17], then applied these CNN models as feature robustness of the proposed model was investigated using thermal
extractors on a small dataset in another domain [18]. Hence, deep images of different conditions of the induction motor.
TL can offer promising approaches to the problem of fault diagnosis.
TL model uses deep learning network (DL) to transfer high-level III.PROPOSED TECHNIQUE
properties from source data to target data [19]. One of the The presented work builds a novel application that uses a pre-
advantages of deep transfer learning is to use a layer-by-layer trained (ResNet-50) model based on CNN as a feature extractor. The
learning pattern to extract attribute from the input data, which allows whole dataset is divided into train, validation, and test subsamples
its deep architecture to produce high data representations with a which each subsample continuously plays the role of validating
potential enhancement of the diagnosis performance. dataset to achieve the reliable performance for fault diagnosis. This
As a result of this, this work proposes an efficient fault diagnosis model is trained by applying the weight of the ImageNet dataset.
model using transfer convolutional neural network based on residual The block diagram of proposed model is shown in figure 1. Images
neural network as a feature extractor considering 50 layers. ResNet- are pre-processed with class imbalance technique and data
50 can extract high-quality features from images on ImageNet. It is augmentation technique. This can affect the achieved results by
anticipated that the suggested (ResNet-50) would increase the final controlling the image zoom, horizontal flip, rotation, translation, and
prediction accuracy on fault diagnosis. The proposed model is image orientation. Then, the structure of resnet-50 network was
investigated using thermal images of the induction motor. initialized restoring the pre-trained weight in ResNet-50. Next, the
The rest of this paper is structured as follows: Section II images are taken as input to obtain the features and training the
defines the related work; Section III presents the proposed model; model for the classification purpose. SoftMax classifier layers were
materials and methods are illustrated in section IV; Section V added with 128 hidden neurons and determined with seven label
reports results; and the conclusion is drawn in section VI. classes. Adam optimizer, ReLU activation function, L2 regulation,
categorical cross-entropy (CE), and dropout layers.
II.RELATED WORK
This study includes data-driven fault diagnosis utilizing deep
CNN networks and feature transfer. Due to the rapid development
of smart manufacturing, data-driven fault detection has emerged as
a popular research topic in recent years[12]. In [20] a new fault
diagnosis based on the use the transfer learning of sparse
autoencoder method and the experimental result has achieved a high
accuracy of 99.82 %. In [21] The use of recurrent neural networks
and dynamic Bayesian modelling to detect faults in induction motors
was investigated. The model has carried out the real-time
experiments with three motors, estimating the probability
distributions for the motor's conditions and the model achieved an
efficient result. A new model based on a performance comparison
of sparse autoencoder with SoftMax regression was proposed in
[22], and the result was satisfactory. In [23] a new model was
proposed for intelligent fault diagnosis applying normalized sparse
autoencoder of air compressors. In [24] a fault diagnosis method was
suggested using stacked sparse autoencoder with ensemble
empirical mode decomposition and the result has proved the
robustness of the extracted features. In [25] deep learning model
based hierarchical diagnosis method to diagnose the rolling element
bearing fault, the obtained result has investigated the reliability of
the model. A new deep transfer learning model was suggested in
[26] to diagnosis the industry application faults, and the achieved
result has investigated several state-of-art transfer learning result
considering the operating condition and fault severities. Recently,
an adaptive deep CNN model was suggested in [27], and the result
was effective and robust. In [28] a hierarchical adaptive CNN was
investigated for weight updating by adding an adaptive
learning rate and a momentum component. In [29] an
intelligent fault diagnosis model using hierarchical CNN was
proposed for diagnosing rolling bearing faults. The result has proved
the effectiveness of the CNN model. A new fault diagnosis based on Figure1. Block diagram of the proposed network
the use of CNN with empirical mode decomposition was proposed

2
International Conference on Innovations in Energy Engineering & Cleaner Production IEE CP22

The categorical cross-entropy Table 1. Motor conditions

(CE) = − ∑𝐶𝑖 𝑡𝑖 log(𝑓(𝑠)𝑖 ) (1) Fault mode Motor load Images Class
(rpm) No. label
Where 𝑡𝑖 is the ground truth, and 𝑓(𝑠)𝑖 is the standard SoftMax.
𝑒 𝑠𝑖
𝑓(𝑠)𝑖 = ∑𝐶
𝑠𝑗 (2)
𝑖 𝑒
Where 𝑠𝑖 presents the given the class, 𝑠𝑗 is the scores derived
from the net for each class. Normal motor 1480/1450 350 0

IV.MATERIALS AND METHODS

IBF 1480/1450 350 1
A. Deep ResNet-50 network architecture
OBF 1480/1450 350 2
CNN models' ability to accurately diagnose faults is
constrained with the help of transfer learning that trained on
ImageNet [17]. This work uses ResNet-50 that trained on ImageNet 8BRBF 1480/1450 350 3
to classify different thermal images conditions of the induction
motor. Resent-50 model has achieved an efficient performance in
the field of images classification which extracts good quality IBF+1BRBF 1480/1450 350 4
features of images for building a strong fault diagnosis model.

The proposed ResNet-50 is consisted of 50-layers deep OBF+5BRBF 1480/1450 350 5

CNN[32] which includes one max- pooling layer, one average pool,
and 48 convolutional layers. The basic structure of ResNet-50 is
illustrated in figure 2 that used as feature extractor. The ResNet-50 BBF+8BRBF 1480/1450 350 6
is composed of five convolutional blocks of layers. The final
convolutional block of a 50-layer deep residual network produces
deep residual features that pre-trained on ImageNet. The
convolutional blocks of a ResNet-50 are different from those of the
traditional CNNs because of the introduction of a shortcut the D. Model evaluation
connection between the input and output of each block. Using
identity mappings as ResNet-50 shortcut connections can optimize
Some evaluation matrices have utilized to assess the
the training process and minimise complexity[33]. That can lead to performance of the proposed application such as the training
achieve a deeper model with fast training and less computational accuracy and loss curves and the parameters given in the
model if compared to i.e. VGG model [34]. following equations:

The proposed work extracts the features from thermal images

using last convolutional block of ResNet-50 (pre-trained model), TP+TN
Overall accuracy = (3)
and the output of the 5th Conv block is trained for fault classification. TP+FP+TN+FN

As a result, the output size of ReseNet-50 then is 2480. TP

Precision = (4)
TP+FP
B. Data Collection
TP
Sensitivity = (5)
The induction motor thermal images were captured in the TP+FN
Wolfson Magnetics Laboratory, School of Engineering Cardiff
Precision × Sensitivity
University, UK. The test rig is displayed in figure 3 which composes F1_score = 2× (6)
Precision+ Sensitivity
of the following components: induction motor, thermal camera
(FLIR C2), and dynamometer, which serves as the load. The thermal Where TP is the true positive prediction, FP is false positive
camera was located approximately 60 cm from the motor housing predictions, TN presents the true negative predictions, and the false
center. negative predictions is stated by FN.

The thermal images have been captured with seven motor V. RESULT
conditions considering the healthy and the faulty motor cases when
the motor running at two speeds as described in Table1. The result of the pretrained model is presented for fault
diagnosis in this section. Induction motor thermal images with
different cases were employed as an input of the pre-trained ResNet
C. Data pre-processing
model based on 50 layers network to extract the model features on
ImageNet dataset. These features were trained on FC layer to predict
Due to the input size of the prosed ResNet-50 are 224 × 224 and the correct class. The classification result is presented in Table 2.
the size of the collected images are 320×240, these images were The model has achieved classification accuracy of 99.98% with
resized to 224 × 224. Dataset was built which includes seven classes training error of 0.0015. In addition, the precision and F1-score have
based on different motor conditions and each class has 350 images. achieved great scores of 98.59, and 94.89; respectively. The overall
The images are proposed with class imbalance technique and data
trained accuracy and loss considering same epochs number are
augmentation technique to improve the model performance. This
layer affects the achieved results by controlling the image zoom, displayed in figure 4 and 5; respectively.
horizontal flip, rotation, translation, and image orientation.
3
International Conference on Innovations in Energy Engineering & Cleaner Production IEE CP22

presented in Table 3. It can be concluded that the suggested pre-

Figure 2. The architecture of the proposed ResNet-50 network

trained network ResNet-50 trained on ImageNet has successfully

achieved a satisfactory application for diagnosing induction motor
faults using motor thermal images.

Figure 3. Test rig of the experiments

The classification accuracy of this model is further compared

with recent published deep learning methos based (CNN).
Normalized SAE was proposed by Jia in [23], stacked sparse
autoencoder model was presented by Oi in [24], sparse filter (SF)
proposed by Lei in [35], and deep belief network (DBN) presented
by Gan [25]. Figure 4. Training accuracy curve applying ReseNet-50

Table 2. model classification result

Train ResNet-50 network with thermal
images
score Batch epochs
size
Accuracy (%) 99.98 64 100
Precision (%) 98.59 64 100
Sensitivity 95.99 64 100
(%)
F1-score (%) 94.89 64 100

Table 3. accuracy comparison with other methods

Model Accuracy
Proposed model 99.98 Figure 5. Training loss curve applying ReseNet-50
NSAE-LCN 99.92
SSAE 99.85
Sparse filter 99.66 IV. CONCLUSION
DBN 99.03
This study develops a new fault diagnosis model applying a
ResNet-50 CNN based transfer learning network. Thermal images
were applied and pre-processed using data augmentation techniques
for improving the final prediction accuracy, then provided as inputs
It has been observed that the proposed ResNet-50 network to deeper feature extraction network based pretrained model of
achieves the best results and the accuracy compassion results are ReseNet-50. The combination of the proposed pre-trained network

4
International Conference on Innovations in Energy Engineering & Cleaner Production IEE CP22

with densely connected classifier has given a highest classification 10.1109/UPEC50034.2021.9548171.

accuracy of 99.98%. in addition, the model has been compared with [11] H. M. Ertunc, H. Ocak, and C. Aliustaoglu, “ANN- and
other deep learning model, and the results show that the proposed ANFIS-based multi-staged decision algorithm for the
resent-50 is the best. detection and diagnosis of bearing faults,” Neural Comput.
Concisely, the overall accuracy of the classification method is Appl., vol. 22, no. SUPPL.1, pp. 435–446, 2013, doi:
satisfactory, suggesting that this model has potential applicability in 10.1007/s00521-012-0912-7.
the identification of induction motor faults utilizing the thermal [12] R. Liu, B. Yang, E. Zio, and X. Chen, “Artificial
imaging data. In future work, detection time is also considered for intelligence for fault diagnosis of rotating machinery: A
implementing the model. Moreover, the proposed model is trained review,” Mech. Syst. Signal Process., vol. 108, pp. 33–47,
using cross-validation technique for further accuracy improvement. 2018, doi: 10.1016/j.ymssp.2018.02.016.
Furthermore, this model is implemented for online application that [13] J. Wang, Y. Ma, L. Zhang, R. X. Gao, and D. Wu, “Deep
based on fault detection. learning for smart manufacturing: Methods and
applications,” J. Manuf. Syst., vol. 48, pp. 144–156, 2018,
doi: 10.1016/j.jmsy.2018.01.003.
VII. ACKNOWLEDGMENTS [14] Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,”
Nature, vol. 521, no. 7553, pp. 436–444, 2015, doi:
Alasmer Ibrahim acknowledges the sponsorship from the High Ministry 10.1038/nature14539.
of education in Libya. [15] L. Wen, X. Li, and L. Gao, “A transfer convolutional neural
network for fault diagnosis based on ResNet-50,” Neural
VIII. REFERENCES: Comput. Appl., vol. 32, no. 10, pp. 6111–6124, 2020, doi:
10.1007/s00521-019-04097-w.
[1] S. Ayas and M. S. Ayas, “A novel bearing fault diagnosis [16] R. Zhao, R. Yan, Z. Chen, K. Mao, P. Wang, and R. X.
method using deep residual learning network,” Multimed. Gao, “Deep learning and its applications to machine health
Tools Appl., no. 0123456789, 2022, doi: 10.1007/s11042- monitoring,” Mech. Syst. Signal Process., vol. 115, pp.
021-11617-1. 213–237, 2019, doi: 10.1016/j.ymssp.2018.05.050.
[2] Z. Gao, C. Cecati, and S. X. Ding, “A Survey of Fault [17] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How
Diagnosis and Fault-Tolerant Techniques—Part I: Fault transferable are features in deep neural networks?,” Adv.
Diagnosis With Model-Based and Signal-Based Neural Inf. Process. Syst., vol. 4, no. January, pp. 3320–
Approaches,” IEEE Trans. Ind. Electron., vol. 62, no. 6, 3328, 2014.
pp. 3757–3767, 2015, doi: 10.1109/TIE.2015.2417501. [18] T. D. Jeff Donahue∗, Yangqing Jia∗, Oriol Vinyals, Judy
[3] P. Baraldi, F. Cannarile, F. Di Maio, and E. Zio, Hoffman, Ning Zhang, Eric Tzeng, “A Deep Convolutional
“Hierarchical k-nearest neighbours classification and Activation Feature.,” vol. 32, 2014.
binary differential evolution for fault diagnostics of [19] L. Zuo, M. Jing, J. Li, L. Zhu, K. Lu, and Y. Yang,
automotive bearings operating under variable conditions,” “Challenging tough samples in unsupervised domain
Eng. Appl. Artif. Intell., vol. 56, pp. 1–13, 2016, doi: adaptation,” Pattern Recognit., vol. 110, 2021, doi:
10.1016/j.engappai.2016.08.011. 10.1016/j.patcog.2020.107540.
[4] K. Worden, W. J. Staszewski, and J. J. Hensman, “Natural [20] L. Wen, X. Li, X. Li, and L. Gao, “A New Deep Transfer
computing for mechanical systems research: A tutorial Learning Based on Sparse Auto-Encoder for Fault
overview,” Mech. Syst. Signal Process., vol. 25, no. 1, pp. Diagnosis,” Proc. 2019 IEEE 23rd Int. Conf. Comput.
4–111, 2011, doi: 10.1016/j.ymssp.2010.07.013. Support. Coop. Work Des. CSCWD 2019, vol. 49, no. 1, pp.
[5] L. Wen, X. Li, and L. Gao, “A New Two-Level 205–209, 2019, doi: 10.1109/CSCWD.2019.8791884.
Hierarchical Diagnosis Network Based on Convolutional [21] H. C. Cho, J. Knowles, M. S. Fadali, and K. S. Lee, “Fault
Neural Network,” IEEE Trans. Instrum. Meas., vol. 69, no. detection and isolation of induction motors using recurrent
2, pp. 330–338, 2020, doi: 10.1109/TIM.2019.2896370. neural networks and dynamic bayesian modeling,” IEEE
[6] S. Yin, S. X. Ding, X. Xie, and H. Luo, “A review on basic Trans. Control Syst. Technol., vol. 18, no. 2, pp. 430–437,
data-driven approaches for industrial process monitoring,” 2010, doi: 10.1109/TCST.2009.2020863.
IEEE Trans. Ind. Electron., vol. 61, no. 11, pp. 6418–6428, [22] N. K. Verma, V. K. Gupta, M. Sharma, and R. K. Sevakula,
2014, doi: 10.1109/TIE.2014.2301773. “Intelligent condition based monitoring of rotating
[7] X. Li, L. Gao, Q. Pan, L. Wan, and K. M. Chao, “An machines using sparse auto-encoders,” PHM 2013 - 2013
Effective Hybrid Genetic Algorithm and Variable IEEE Int. Conf. Progn. Heal. Manag. Conf. Proc., pp. 1–7,
Neighborhood Search for Integrated Process Planning and 2013, doi: 10.1109/ICPHM.2013.6621447.
Scheduling in a Packaging Machine Workshop,” IEEE [23] F. Jia, Y. Lei, L. Guo, J. Lin, and S. Xing, “A neural
Trans. Syst. Man, Cybern. Syst., vol. 49, no. 10, pp. 1933– network constructed by deep learning technique and its
1945, 2019, doi: 10.1109/TSMC.2018.2881686. application to intelligent fault diagnosis of machines,”
[8] M. Kang, J. Kim, and J. M. Kim, “Reliable fault diagnosis Neurocomputing, vol. 272, pp. 619–628, 2018, doi:
for incipient low-speed bearings using fault feature analysis 10.1016/j.neucom.2017.07.032.
based on a binary bat algorithm,” Inf. Sci. (Ny)., vol. 294, [24] Y. Qi, C. Shen, D. Wang, J. Shi, X. Jiang, and Z. Zhu,
pp. 423–438, 2015, doi: 10.1016/j.ins.2014.10.014. “Stacked Sparse Autoencoder-Based Deep Network for
[9] Z. Wang, Q. Zhang, J. Xiong, M. Xiao, G. Sun, and J. He, Fault Diagnosis of Rotating Machinery,” IEEE Access, vol.
“New Hybrid Invasive Weed Optimization and Machine 5, pp. 15066–15079, 2017, doi:
Learning Approach for Fault Detection,” IEEE Sens. J., 10.1109/ACCESS.2017.2728010.
vol. 17, no. 17, pp. 5581–5588, 2017, doi: [25] M. Gan, C. Wang, and C. Zhu, “Construction of
10.1109/JSEN.2017.2726011. hierarchical diagnosis network based on deep learning and
[10] A. Ibrahim, F. Anayi, M. Packianather, and O. Al-Omari, its application in the fault pattern recognition of rolling
“New Hybrid Invasive Weed Optimization and Machine element bearings,” Mech. Syst. Signal Process., vol. 72–73,
Learning Approach for Fault Detection,” 2021 56th Int. pp. 92–104, 2016, doi: 10.1016/j.ymssp.2015.11.014.
Univ. Power Eng. Conf. Powering Net Zero Emiss. UPEC [26] T. Han, C. Liu, W. Yang, and D. Jiang, “Deep transfer
2021 - Proc., no. Cm, 2021, doi: network with joint distribution adaptation: A new
5
International Conference on Innovations in Energy Engineering & Cleaner Production IEE CP22

intelligent fault diagnosis framework for industry

application,” ISA Trans., vol. 97, pp. 269–281, 2020, doi:
10.1016/j.isatra.2019.08.012.
[27] W. Fuan, J. Hongkai, S. Haidong, D. Wenjing, and W.
Shuaipeng, “An adaptive deep convolutional neural
network for rolling bearing fault diagnosis,” Meas. Sci.
Technol., vol. 28, no. 9, 2017, doi: 10.1088/1361-
6501/aa6e22.
[28] X. Guo, L. Chen, and C. Shen, “Hierarchical adaptive deep
convolution neural network and its application to bearing
fault diagnosis,” Meas. J. Int. Meas. Confed., vol. 93, pp.
490–502, 2016, doi: 10.1016/j.measurement.2016.07.054.
[29] C. Lu, Z. Wang, and B. Zhou, “Intelligent fault diagnosis
of rolling bearing using hierarchical convolutional network
based health state classification,” Adv. Eng. Informatics,
vol. 32, pp. 139–151, 2017, doi: 10.1016/j.aei.2017.02.005.
[30] R. Chen and S. Shen, “Fault Diagnosis for Rotating
Machinery Based on Convolutional Neural Network and
Empirical Mode Decomposition,” Zhendong Ceshi Yu
Zhenduan/Journal Vib. Meas. Diagnosis, vol. 25, no. 3, pp.
233–235, 2005.
[31] H. Li, J. Huang, X. Yang, J. Luo, L. Zhang, and Y. Pang,
“Fault Diagnosis for Rotating Machinery Using Multiple
Sensors and Convolutional Neural Networks,” Entropy,
vol. 22, no. 8, pp. 101–110, 2020, doi: 10.3390/E22080851.
[32] A. Mahmood et al., “Automatic hierarchical classification
of kelps using deep residual features,” Sensors
(Switzerland), vol. 20, no. 2, pp. 1–20, 2020, doi:
10.3390/s20020447.
[33] T. Gevers and A. Smeulders, “Identity Mappings in Deep
Residual Networks,” Lect. Notes Comput. Sci. (including
Subser. Lect. Notes Artif. Intell. Lect. Notes
Bioinformatics), vol. 9909 LNCS, p. V, 2016, doi:
10.1007/978-3-319-46493-0.
[34] V. Sangeetha and K. J. R. Prasad, “Deep Residual Learning
for Image Recognition Kaiming,” Indian J. Chem. - Sect. B
Org. Med. Chem., vol. 45, no. 8, pp. 1951–1954, 2006, doi:
10.1002/chin.200650130.
[35] Y. Lei, F. Jia, J. Lin, S. Xing, and S. X. Ding, “An
Intelligent Fault Diagnosis Method Using Unsupervised
Feature Learning Towards Mechanical Big Data,” IEEE
Trans. Ind. Electron., vol. 63, no. 5, pp. 3137–3147, 2016,
doi: 10.1109/TIE.2016.2519325.