Detection of Compound Faults in Ball Bearings Using Multiscale-SinGAN, Heat Transfer Search Optimization, and Extreme Learning Machine

Round 1

Reviewer 1 Report

1: The logic of the problem should be strengthened;

2: How robust is the algorithm?

3: How effective is the public dataset detection?

Author Response

Authors are thankful to anonymous reviewer 1 and special issue editors to give us chance to highlight the utility of our proposed methodology. The quality of our manuscript is enhanced considerably after incorporating the suggestion given by reviewer 1.

Author Response File: Author Response.docx

Reviewer 2 Report

In the paper the authors proposed a procedure for the classification of a damaged bearing by comparing different methods. Some parts are not clear, and the effect of some steps in the procedure is hidden.

1. In the abstract it is stated that HT is used, whereas in the text HHT is claimed. Which one has been used?
2. It is not clear and hidden how HHT has been used and how kurtograms have been obtained from signals after HHT. Kurtogram works on the raw signal and calculates the kurtosis of the signal in several frequency bands.
3. The details of the damaged bearing are hidden. Add pictures. Which kind of damage and have been introduced? Size (width, depth)? Lubrication? How is the load applied? How many bearings have been tested?
4. The effectiveness of creating artificial signal with SigGAN is not clear. From figure 8, very light differences can be highlighted at the same speed. It is obscure if it is necessary to use SigCAN, and if the results are good only due to the fact that artificial images are too similar to the original one. Test must be repeated with different damaged bearings, or signals can be acquired at different temperature states.
5. At page 13 the authors stated: “ It is observed that the computation cost and biasedness in classification are increased by the presence of redundant and irrelevant features, which significantly affects the misclassification accuracy”. Please give details of the computational time and the quantify the misclassification due to the redundant feature.
6. In tables 2,3 and 4, what does Class represent?
7. Figures 9 and 10 are difficult to be read. Please change the range of y-axis from 70% to 100%. Some figures can be removed and only one added as example. A simple table as table 5 is more readable. Please also compare result at the same speed
8. It seems that speed as a strong effect on the results. Why?
9. By looking at the results the slight difference between 99.6x and 100% on the three methods can justify the conclusions? What are the drawbacks (computational time, effect of SigGAN, etc.)?

Author Response

Authors are thankful to anonymous reviewer 2 and special issue editors to give us chance to highlight the utility of our proposed methodology. The quality of our manuscript is enhanced considerably after incorporating the suggestion given by reviewer 2.

Author Response File: Author Response.pdf

Reviewer 3 Report

In this paper, the authors proposed a hybrid framework for compound fault detection in ball bearings. Specifically, vibration signals are pre-processed using the Hilbert transform, and then a Kurtogram is generated. The multiscale-SinGAN model is adapted to generate additional faulty images from Kurtogram. Standard image quality parameters as a feature are extracted and to identify the relevant parameters. Metaheuristic optimization algorithms, Teaching learning-based optimization, and Heat Transfer Search are applied to the feature vector. Finally, selected features are fed into three machine-learning models for compound fault identifications. A case study is used to demonstrate the good performance of the proposed approach.

The detection results of case study are accurate, and I have some comments listed as follows:

1. This paper considers three types of defects: faults in the inner race, outer race, and rolling elements. What are the differences in vibration signals for them? Any graphical illustration?

2. There is a variety of existing models for bearing fault detection based on vibration signals. Can the authors add a few model comparisons to existing methods?

3. How is the overall computational time of the proposed approach? It seems 2000 images are generated via GAN for analysis during the process.

4. The generated images via GAN are quite similar to original images with very small variations. How does this affect the final detection accuracy?

5. Double check Equation (1). The integration should not be respective to t. Otherwise, on left side there will be no t.

6. The paper writing needs significant improvement. Just to name a few:

(1)   Page 4 line 114, “be determine” should be “be determined”.

(2)   Page 5 line 156, font sizes are not consistent.

(3)   Page 5 line 175, check “in two phase”. Should be “in two phases” or “in phase two”?

Author Response

Authors are thankful to anonymous reviewer 3 and special issue editors to give us chance to highlight the utility of our proposed methodology. The quality of our manuscript is enhanced considerably after incorporating the suggestion given by reviewer 3.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors improved the paper accordingly to reviewer comments.

Please add some details about the HHT for filtering the signal and a sentence in the text on the necessity of using HHT.

Author Response

Thank you for positive feedback. As per suggestion, necessary details has been added in revised version 2.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors have addressed my comments well. I recommend to accept.

Author Response

Thank you for reviewing and accepting manuscript for publication.

Author Response File: Author Response.pdf