Bearing Fault Diagnosis Using ACWGAN-GP Enhanced by Principal Component Analysis
Round 1
Reviewer 1 Report
In this paper, a data enhancement model based on ACGAN and WGAN-GP optimized by principal component analysis is proposed for fault diagnosis of rolling bearings. The proposed method includes the conversion of temporal signals to grayscale images. The experimental verification in this paper is sufficient, but there are many technical disputes.
The major problems are given as follows,
1. The abstract is too long and it is recommended to revise it appropriately.
2. The format of equations is not standardized, the aesthetics are not enough, and there are problems of incomplete display, which need to be checked one by one. The general lack of explanation of the necessity of the variables in the formula reduces readability. In particular, the correctness of Eq. (4) and Eq. (5).
3. The specification of the variables needs to be strengthened, e.g. "for t = 1, ..., ncritic do" in Table 1, "K-L ipschitz" in line 206, "MVIDIA" in line 299.
4. It is interesting to use the features extracted by PCA as input to the generator instead of random noise as mentioned in the paper. It is necessary to give more detailed explanations in the following areas.
(1) It is mentioned in Section 4.2 that the input dimension of the generator is 100, which is not consistent with the 64*64 feature image shown in Figure 8.
(2) A fixed feature z of a certain fault class is bound to be misleading for fault generation in other classes. Is the input z of the proposed GAN changeable according to the class?
5. The cosine distance is used in Table 4 to measure the similarity. In general, the closer the cosine distance is to 1, the higher the similarity of the vectors in the direction. However, in Table 4, the CD of PCA-ACWGAN-GP is significantly and generally smaller than that of ACGAN and ACWGAN-GP. This contradicts the conclusion of section 4.3 and requires an explanation from the authors.
6. The authors' interpretation of TP, TN, FP and FN in Eq. (13) and (14) is ambiguous and should be redefined. Precision and Recall are both measures of the model's classification performance, related to the model's prediction class and independent of whether it is a generated sample or not. The interpretation of equations in Section 4.4 does not support the analysis of experimental results.
7. The sample numbers in Figure 10 are easily confused with Table 5. It is suggested that the authors add a description of the sample number in Table 8.
Author Response
Dear Reviewer:
We appreciate your positive and constructive comments on our manuscript. Your suggestions are very important to improve our manuscripts. Based on your detailed suggestions, we have made substantial changes to the manuscript. The point-by-point reply to each of your comments is as follows. If you have any questions about this article, please feel free to contact us. Thank you very much for your help.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper entitled "Fault Diagnosis of Rolling Bearing Using Improved ACGAN" presents an innovative approach to improving the accuracy of fault diagnosis in rolling bearings, which are essential components of rotating machinery. The study aims to address the issue of insufficient fault sample data in the field of rolling bearing fault diagnosis using deep learning techniques. Specifically, the authors propose an auxiliary classifier generative adversarial network (ACGAN) optimized by principal component analysis (PCA) method to generate an adversarial network model.
The proposed model uses Wasserstein distance and gradient penalty to enhance the stability of the network training process, thereby overcoming the problems of over-fitting and gradient disappearance during model training. Furthermore, the generator of the generative adversarial network is improved by transforming the one-dimensional time-domain signal into a two-dimensional grayscale image and applying the principal component analysis to reduce the feature dimension of the real sample. The label information of the fault sample is input into the generator as a condition, which transforms the generative adversarial network from unsupervised learning to supervised learning.
The results of the experiments show that the proposed model can produce high-quality samples that are similar to real samples, and effectively improve the classification accuracy of fault diagnosis in the case of insufficient fault samples. The study contributes to the field of fault diagnosis by presenting a novel approach that addresses the issue of insufficient fault sample data, which is a common problem in the field of rolling bearing fault diagnosis.
However, the introduction and literature review require more emphasis on the research work and its entire process, including past, present and future perspectives. The authors need to incorporate more recent and updated research papers related to the topic, particularly those from the past few years, to provide a comprehensive background of the study and its potential applications.
If the suggested changes are made to the introduction and literature review, the paper can be published.
1. How does the proposed PCA-ACWGAN-GP data enhancement model compare to other deep learning models in the field of rolling bearing fault diagnosis?
2. Can you provide a comprehensive background of the study and its potential applications?
3. How does the proposed model address the issue of insufficient fault sample data in the field of rolling bearing fault diagnosis?
4. What are the advantages of using Wasserstein distance and gradient penalty in the proposed model?
5. How does the generator of the generative adversarial network differ from other models used in rolling bearing fault diagnosis?
6. What is the role of the auxiliary classifier in the proposed ACGAN model?
7. How does the proposed model improve the stability of the network training process?
8. How does the proposed model improve the ability of the discriminator network to extract features in multi-classification scenarios?
9. How does the PCA method reduce the feature dimension of the real sample?
10. How does the label information of the fault sample input into the generator as a condition improve the performance of the generative adversarial network?
11. What is the significance of the proposed model in the field of rolling bearing fault diagnosis?
12. How does the proposed model contribute to the development of intelligent production equipment?
13. Can the proposed model be applied to other types of rotating machinery beyond rolling bearings?
14. How can the proposed model be optimized for better performance in scenarios with extremely small sample sizes?
15. How does the proposed model address the issue of over-fitting in deep learning models?
16. What are the limitations of the proposed model?
17. How can the proposed model be further improved in future research?
18. How can the proposed model be validated in real-world scenarios?
19. How can the proposed model be integrated into existing fault diagnosis systems for rolling bearings?
Author Response
Dear Reviewer:
We appreciate your positive and constructive comments on our manuscript. Your suggestions are very important to improve our manuscripts. Based on your detailed suggestions, we have made substantial changes to the manuscript. The point-by-point reply to each of your comments is as follows. If you have any questions about this article, please feel free to contact us. Thank you very much for your help.
Author Response File: Author Response.pdf
Reviewer 3 Report
This paper proposes an auxiliary classifier optimized by the principal component analysis method to generate an adversarial network model in order to evaluate the fault diagnosis of rolling bearing. The paper fits within the scope of this journal. The manuscript is appropriately structured. The introduction, methodology and results sections are well presented and discussed. Some points need to be revised, as follows:
1. The abstract is too long. Condense it. Also, revise the abstract providing specific results of the study.
2. Sub-section 4.1. The authors wrote that the data set contains 4000 samples (line 336). The test A contains 1500 samples for training data and test B contains of 1000 samples for test data (Table 3). What about the rest of 1500 samples? This is not clear. Please explain this part of the paper.
3. Sub-section 4.2. On what basis were the parameters determined? How did you obtain these values? It is not clarified. Did you perform a parameter grid search, or something else...?
4. Conclusions. In the conclusions, state the limitations of the study and provide further research directions.
Author Response
Dear Reviewer:
We appreciate your positive and constructive comments on our manuscript. Your suggestions are very important to improve our manuscripts. Based on your detailed suggestions, we have made substantial changes to the manuscript. The point-by-point reply to each of your comments is as follows. If you have any questions about this article, please feel free to contact us. Thank you very much for your help.
Author Response File: Author Response.pdf
Reviewer 4 Report
What’s the main scientific contribution? It should be addressed.
English grammar should be improved, for example:
”Therefore, the condition monitoring of the key core components of the rotating machinery, especially the rolling bearing, using the fault diagnosis technology to predict the possible faults, analyzing the causes of the existing faults and maintaining them in time before the loss is expanded, can ensure the safe and stable operation of the rotating machinery, can effectively reduce the maintenance cost during the working period of the mechanical equipment, reduce unnecessary financial expenses, avoid the occurrence of major accidents, and provide protection for the lives and property of the people”.
This sentence is too long to understand.
In figure 7, some types seem similar, is there clear classification criteria when labelling them manually?
Figure 9, it’s recommended to use a bar chart to make comparisons.
In table 6, what’s the design basis for the2D-CNN
Author Response
Dear Reviewer:
We appreciate your positive and constructive comments on our manuscript. Your suggestions are very important to improve our manuscripts. Based on your detailed suggestions, we have made substantial changes to the manuscript. The point-by-point reply to each of your comments is as follows. If you have any questions about this article, please feel free to contact us. Thank you very much for your help.
Author Response File: Author Response.pdf
Reviewer 5 Report
The manuscript proposes a Fault Diagnosis method for Rolling Bearing Using Improved ACGAN. The authors made a remarkable effort in the article and the content is written well. However, the manuscript can be improved by considering the following comments.
1. The abstract wasn’t written well. The abstract should present information in a direct concise way. Considering the background, research gap, methodology, and the findings obtained by performing the proposed model against the stat-of-art models.
2. I advise the authors to rewrite the abstract in a better way.
3. The manuscript’s organization for the next sections should be mentioned at the end of the introduction section.
4. Figure 2. Is not clear.
5. The structure of the generator and the discriminator looks similar with little modification, I think they should be presented in a better structure.
6. There is a self-attention layer in the model, please add an explanation about its function.
7. No details about the specifications of the used dataset, please consider this.
8. Can you explain what are the features extracted and used in the proposed model?
9. There are three state-of-arts models, it is better to add a summary of the models.
10. A comparison of Models’ efficiency (regarding execution time) should be considered.
11. The title considered an improvement of the ACGAN, this is not clear in the manuscript.
12. In the title, “for” should be used instead of “of”.
13. There are many English grammar and typos, therefore the whole manuscript should be checked.
Author Response
Dear Reviewer:
We appreciate your positive and constructive comments on our manuscript. Your suggestions are very important to improve our manuscripts. Based on your detailed suggestions, we have made substantial changes to the manuscript. The point-by-point reply to each of your comments is as follows. If you have any questions about this article, please feel free to contact us. Thank you very much for your help.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
I am thrilled to inform you that the revised version of your paper has successfully addressed the concerns identified in the initial version. The changes made have significantly improved the manuscript's clarity and coherence, leading me to believe that it is now fit for acceptance.
Reviewer 5 Report
The authors made remarkable efforts in revising the manuscript's content.
The manuscript can be accepted for publication in its current form.
