HANOI UNIVERSITY OF SCIENCE AND TECHNOLOGY

SCHOOL OF ELECTRICAL AND ELECTRONICS

ENGINEERING

GRADUATION THESIS
DEEP LEARNING BASED APPROACHES FOR BEARING
FAULTS DIAGNOSIS

VU MANH HUNG
[email protected]
NGUYEN VAN QUANG
[email protected]

Department of Automation Engineering

Supervisor: Assoc. Prof., Ph.D. Pham Van Truong

Signature of Supervisor:

Department: Automation Engineering

School: School of Electrical and Electronics Engineering

Hanoi, 07/2023
 Acknowledgements

We would like to express my deepest gratitude to my thesis mentor Assoc.

Prof., Ph.D. Pham Van Truong as well as Assoc. Prof., Ph.D. Tran Thi Thao for
their unwavering support and guidance throughout our studies. Their expertise and
patience have been invaluable to me and have played a crucial role in the success of
this thesis. We consider ourselves fortunate to have been part of IPSAL, an experience
through which Assoc. Prof., Ph.D. Pham Van Truong and Assoc. Prof., Ph.D.Tran
Thi Thao not only ignited my enthusiasm but also acquainted us with the perpetual
challenges of deep learning. While our involvement with IPSAL commenced later
than anticipated, the unwavering backing of Assoc. Prof., Ph.D. Pham Van Truong
and Assoc. Prof., Ph.D. Tran Thi Thao played an indispensable role in our accomplishments
and progress.
Many thanks to Ha Noi University of Science and Technology (HUST) for
providing me with a transformative platform to nurture my aspirations over four
remarkable years. HUST has bestowed upon me the invaluable gift of passionate
educators, cherished companions, and captivating moments that are truly unparalleled.
Regardless of the path I tread in the future, I shall forever take pride in being a member
of the HUST community.
Lastly, we would be remiss in not mentioning our families, especially our parents.
Their belief in us has kept our spirits and motivation high during this process.

Hanoi, July 30th, 2023

Vu Manh Hung Nguyen Van Quang

 ABSTRACT

Bearing fault diagnosis plays a crucial role in the monitoring and maintenance
of rotating machinery, particularly in electrical machines. Accurately identifying
bearing failures has become increasingly important, prompting a surge of interest in
applying deep learning techniques to this domain. Convolutional neural networks
(CNNs) have shown promising results in bearing fault diagnosis, delivering high
accuracy rates. However, the main challenge is the need for a substantial amount
of training data to achieve optimal performance.
To address this limitation, we present a novel deep neural network approach that
combines the strengths of the MLP-mixer model and CNN. This approach has the
potential to enhance training performance even with a small dataset significantly. By
leveraging the advantages of both models, we aim to improve diagnostic accuracy
while reducing the dependency on a large training dataset.
Extensive experiments and thorough ablation studies have been conducted on
the widely used Case Western Reserve University (CWRU) database to validate the
effectiveness of our proposed approach. The evaluation results demonstrate that the
proposed model consistently achieves high accuracy even when the number of training
samples is reduced. This robustness in performance with limited data highlights the
practicality and reliability of our approach in real-world applications.
Comparisons with other state-of-the-art methods further validate the superiority
of our proposed model. It outperforms existing approaches, showcasing the advantages
of combining the MLP-mixer and CNN models for bearing fault diagnosis.
Overall, this thesis contributes novel and efficient deep-learning approaches for
bearing fault diagnosis that can achieve high accuracy even with limited training data.
The results from extensive experiments and comparisons with existing techniques
demonstrate the efficacy and potential of our proposed model. By reducing the
demand for large training datasets, this approach opens up new possibilities for implementing
fault diagnosis systems in real-world scenarios, ultimately improving the reliability
and maintenance of rotating machinery. Further research and exploration in this
direction hold promise for advancing the field of bearing fault diagnosis and deep
learning applications in industrial settings.
 TABLE OF CONTENTS

LIST OF ABBREVIATIONS i

LIST OF FIGURES iii

LIST OF TABLES iv

CHAPTER 1. INTRODUCTION 1
1.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Background and Problems of Research . . . . . . . . . . . . . . . . . 2
1.3 Contribution and outline of thesis . . . . . . . . . . . . . . . . . . . 3

CHAPTER 2. LITERATURE REVIEW 5

2.1 Multilayer Perceptron Network . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Multilayer perceptron . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Activation function . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Convolution Neural Network . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Convolutional layer . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2 Pooling layer . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 ConvMixer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 MLP-mixer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Fewshot Leanring . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6 Loss Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6.1 Contrastive Loss . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6.2 Cross-Entropy Loss . . . . . . . . . . . . . . . . . . . . . . . 19

CHAPTER 3. SIAMESE-BASED CONVMIXER ARCHITECTURE 20

3.1 General architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Comparative results . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

CHAPTER 4. HYBRID CNN-MLP ARCHITECTURE 26

4.1 Spectrogram input . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Hybrid CNN-MLP architecture . . . . . . . . . . . . . . . . . . . . . 26

7
4.3 Implemented details . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.4 Experiment results . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.4.1 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.4.2 Comparative results . . . . . . . . . . . . . . . . . . . . . . . 29

CHAPTER 5. Mixer-Former ConvaMnet architecture 34

5.1 Mixer-Former CovaMnet architecture . . . . . . . . . . . . . . . . . 34
5.2 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.3 Comparative results . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

CONCLUSION 42

LIST OF PUBLICATION 43

TÀI LIỆU THAM KHẢO 44

APPENDIX 48
 LIST OF ABBREVIATIONS

CNN Convolution Neural Network

MLP Multi-layer Perceptron
SGD Stochastic Gradient Descent
CWRU Case Western Reverse University
t-SNE t-distributed stochastic neighbor embedding
 LIST OF FIGURES

Figure 2.1. The general architecture of Multilayer perceptron networks . . 5

Figure 2.2. Sigmoid and its derivative function . . . . . . . . . . . . . . . 6
Figure 2.3. Tanh and its derivative function . . . . . . . . . . . . . . . . . 7
Figure 2.4. ReLU6 function . . . . . . . . . . . . . . . . . . . . . . . . . 8
Figure 2.5. Leaky ReLU function . . . . . . . . . . . . . . . . . . . . . . 9
Figure 2.6. The well-known Lenet5 architecture. It first applies to handwriting
recognition. The input is a handwritten digit, and the output is a
probability over 10 possible outcomes. . . . . . . . . . . . . . . . . . 10
Figure 2.7. Two-dimensional cross-correlation operation . . . . . . . . . . 10
Figure 2.8. Features map generated by applying various convolution kernels
to an image from MNIST datasets. . . . . . . . . . . . . . . . . . . . 11
Figure 2.9. Two-dimensional convolution with tried equal to 2 . . . . . . . 12
Figure 2.10. Two-dimensional convolution with zeros padding. . . . . . . . 13
Figure 2.11. The comparison between 2D dilated convolution with a spatial
size of 3×3 and a dilation factor of 2. . . . . . . . . . . . . . . . . . 14
Figure 2.12. There are two prevalent varieties of pooling operators available:
max pooling and average pooling . . . . . . . . . . . . . . . . . . . . 15
Figure 2.13. ConvMixer architecture . . . . . . . . . . . . . . . . . . . . . 15
Figure 2.14. MLP-mixer architecture: MLP-Mixer is made up of per-patch
linear embeddings, Mixer layers, and a classifier head. Mixer layers
include one token-mixing MLP and one channel-mixing MLP, each
with two fully-connected layers and a GELU nonlinearity. Other
components include skip-connections, dropout, and layer norm on the
channels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Figure 3.1. General architecture for predicting bearing failure . . . . . . . 20
Figure 3.2. The proposed Siamese-based Conv-Mixer model . . . . . . . . 21
Figure 3.3. The proposed Few-shot learning model . . . . . . . . . . . . . 21
Figure 3.4. Representative results when choosing 1 signal compared with
5 other random signals . . . . . . . . . . . . . . . . . . . . . . . . . 22
Figure 3.5. (a) Confusion matrix when training with 90 samples and (b)
Visualization of data while training with 90 samples . . . . . . . . . . 22
Figure 3.6. Representative results when choosing 1 signal compared with
6 other random signals . . . . . . . . . . . . . . . . . . . . . . . . . 23

ii