Intelligent Fault Diagnosis of Broken Wires for Steel Wire Ropes Based on Generative Adversarial Nets

Round 1

Reviewer 1 Report

The author presents an interesting application of AI methods to detect rope damage.

The results should be discussed in more detail. Figure 13 shows the performance for different cases, and my question is why are cases 1 and 3 better than case 2? The case number is the number of wires, so why is it easier to detect two wires than one and three wires damaged? Of course, CNN, GAN + CNN are not XAI methods, but I would like to see some authors discussion.

Another question is: is the use of CNN sufficient, the performance difference between CNN and GAN + CNN is not very significant, would the difference be greater in real-world applications, what are your predictions in this subject?

Author Response

Response to Reviewer 1:

Thank you very much for your encouraging comments, suggestions, and careful attention paid to our manuscript. We have checked the manuscript and revised it according to the comments. The point-by-point responses to the comments and suggestions are listed as below:

 

Reviewer point #1: The results should be discussed in more detail. Figure 13 shows the performance for different cases, and my question is why are cases 1 and 3 better than case 2? The case number is the number of wires, so why is it easier to detect two wires than one and three wires damaged? Of course, CNN, GAN + CNN are not XAI methods, but I would like to see some authors discussion.

 

Author response #1: We agreed with your opinion that more discussion on the results is needed. We have made some supplementary explanations in Section 4.2 (line 221-223) in the revised paper.

 

Reviewer point #2: Another question is: is the use of CNN sufficient, the performance difference between CNN and GAN + CNN is not very significant, would the difference be greater in real-world applications, what are your predictions in this subject?

 

Author response #2: Thanks for your insightful criticism and advice about our paper. In this paper, we show that the GAN can improve the identification accuracy of wire rope damage through several types of fault experiments. CNN can realize adaptive feature extraction, but it is limited by small samples. So we used the GAN+CNN method to solve the problem of the sample limitation of CNN in real inspection. Of course, we agree with you that the comparison between CNN and GAN is not comprehensive enough. Meanwhile, a more comprehensive comparative study is of great significance for achieving practical high-precision detection. However, according to the experimental results, we believe that GAN will have better performance when facing complex actual detection. And now, more experimental studies are underway to improve our model and prove the generalization performance of the model.

 

Thank you again for your kind comments and consideration. They are valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We hope our revisions meet with your approval.

Author Response File: Author Response.docx

Reviewer 2 Report

* Originality / Novelty

1. The paper addresses an important research and innovation topic concering robustness of fault diagnistics with artificial intelligence. The work is novel as it device a method combining GAN and CNN with wavelet transfrom that shows improvements over the current state of practice. 

* Significance of Content

 

2. The contribution of new knowledge is good. However, the paper could benefit from a comparison with other machine learning techniques such as support vector machines or unsupervised methods to verify if the same results could be obtained by more lightweight methods that do not require large training sets. This would increase the significance of the scientific contribution.

* Quality of Presentation

3. The paper is overall well-structured and easy to follow. Several minor English grammar mistakes should be fixed (see annotations in the attached pdf). Other smaller issues related to the presentation are: 

3a. The first paragraph of Sec. 3 is a bit verbose. Check also the last sentence of the paragraph.

3b. I suggest skipping jargon as a "new intelligent method" (p. 4) and plainly using "new method". 

3c. Fig. 10 and 11: Indicate the time and the frequency axes in the figures.

* Scientific Soundness

4. My largest concern with the paper is the reproducibility of the results. The paper lacks sufficient details on the data processing pipeline and the experimental testbed. 

4a. Please, provide better clarity on the wavelet transformation step. What type of wavelet transform (mother wavelet) was used? How sensitive is the result of the choice of mother wavelet? 

4b. The data processing does not discuss the use of cross-validation. Did the study not use cross-validation and why not?

4c. I am missing details on the chosen magnetic flux leakage (MFL) signals detector. Was it a commercial device and in that case what type? 

* Interest to the readers

The topic is of interest in the automation and robotic industry. In addition, the paper addresses the important challenge of small training data sets for deep learning which has a more broad significance. 

* Overall Merit

Overall the paper has good merits. The paper has a modest contribution in its current form. It has the potential for a more scientific impact value by increasing the reproducibility of the results and from a comparison with alternative AI methods.

Comments for author File: Comments.pdf

Author Response

Response to Reviewer 2:

Thank you very much for your valuable comments, suggestions, and encouragements. We have carefully revised the manuscript according to the comments. The point-by-point responses to the comments and suggestions are listed as below:

 

Reviewer point #1: The paper addresses an important research and innovation topic concering robustness of fault diagnistics with artificial intelligence. The work is novel as it device a method combining GAN and CNN with wavelet transfrom that shows improvements over the current state of practice.

 

Author response #1: Thank you very much for your encouragement and affirmation of our research work.

 

Reviewer point #2: The contribution of new knowledge is good. However, the paper could benefit from a comparison with other machine learning techniques such as support vector machines or unsupervised methods to verify if the same results could be obtained by more lightweight methods that do not require large training sets. This would increase the significance of the scientific contribution.

 

Author response #2: Thanks for your insightful criticism and advice about our paper. We agree with you that the paper could benefit from a comparison with other machine learning techniques. For example, using support vector machine to classify damage images in the case of limited data and GAN extended data respectively. However, BP network is the most classical network in the field of wire rope fault diagnosis. BP network will be limited by artificial features when identifying faults. CNN can realize adaptive feature extraction, but it is limited by small samples. So we used the GAN+CNN method to solve the sample limitation of CNN, and compared it with the traditional BP network. Meanwhile, your suggestions are very instructive for our future research.

 

Reviewer point #3a: The first paragraph of Sec. 3 is a bit verbose. Check also the last sentence of the paragraph.

 

Author response #3a: We agreed with the reviewer, and some modifications were made in first paragraph of Section 3 (line 116-119).

 

Reviewer point #3b: I suggest skipping jargon as a "new intelligent method" (p. 4) and plainly using "new method".

 

Author response #3b: Thank you very much for your kind advice. We have modified "new intelligent method" to "new method" for clearer and more concise expression (line 121).

 

Reviewer point #3c: Fig. 10 and 11: Indicate the time and the frequency axes in the figures.

 

Author response #3c: We agreed with your opinion. We added time-frequency axes to Figure 10 and 11 in the revised paper.

 

Reviewer point #4a: Please, provide better clarity on the wavelet transformation step. What type of wavelet transform (mother wavelet) was used? How sensitive is the result of the choice of mother wavelet?

 

Author response #4a: Thank you for your kind comments and consideration. We have made relevant supplemented description about the wavelet transformation step in Section 4.2 (line 185-187).

 

Reviewer point #4b: The data processing does not discuss the use of cross-validation. Did the study not use cross-validation and why not?

 

Author response #4b: Thank you for your thoughtful review of our manuscript. In this paper, there is no discussion of cross-validation about data processing. The general approach of neural network training is adopted in this paper, and 70% of the data set is the training set, 30% for verification. Cross-validation can fully train the model to obtain better accuracy in the case of less data. However, in this paper, the GAN is used for data expansion to solve the problem of fault identification under the condition of small sample. Therefore, cross-validation is not discussed in this article.

 

Reviewer point #4c: I am missing details on the chosen magnetic flux leakage (MFL) signals detector. Was it a commercial device and in that case what type?

 

Author response #4c: We are sorry that we did not explain it clearly in the original manuscript. This detector is designed and manufactured by our laboratory, and we have supplemented relevant information in Section 4.1 in the revised paper (line 169-171).

 

Thank you again for your kind comments and consideration. They are valuable and helpful. We have revised the relevant contents and we tried our best to improve the expression of grammar, typos and professional representation of this manuscript and made some changes in the revised manuscript. We hope your questions have been clearly explained and our efforts meet with your approval.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The modifiacations are acceptable.