Hybrid Feature Selection Framework for Bearing Fault Diagnosis Based on Wrapper-WPT
Round 1
Reviewer 1 Report (Previous Reviewer 2)
I have reviewed a previous submission of this manuscript. I think the authors did a good job at describing their methods and apply their method to other datasets. My following comments need to be addressed before accepting for publication:
1) I am not clear how the authors addressed the question of novelty from their previous submission. All well established techniques are weaved together in this paper - what is the significant novelty?
2) Can the authors provide a Table 7-like result for the CWRU dataset? There are lot of papers which provide their results for CWRU dataset and it will be interesting to see how the proposed method compares against some of other very different models (including deep learning).
3) Class labels 0, 1, 2, 3 are misleading. This would somehow mean that "Outer+Inner Ring Damage" is more related to "Inner Ring Damage" than "Outer Ring Damage" whereas in reality all of these should be independent. I see that authors use label encoding for model training - which is correct thing to do. I suggest using class labels from Fig 16 consistently across the manuscript to avoid any confusion.
4) I think the legend in Fig 1 and the pie chart is not matching ("rotor bar" and "others" seem to have switched). Moreover, I think Figure 1 doesn't add any new information - consider removing it. On a related note, a significant number of figures authors use are from other papers (cited). Consider reducing the number of figures for brevity.
5) (minor) Section 1.1 seems misplaced?
6) (minor) Figs 9, 10, 12, 13 etc. have a very Excel-like appearance. Please consider revising to make them appear more professional.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report (New Reviewer)
This paper proposed a hybrid feature selection framework for bearing fault diagnosis based on wrapper-wpt. The proposed method is evaluated on two different bearing benchmark vibration datasets with variable operating conditions. It is of interest. However, the authors should address the following points to further improve the quality.
1. The structure is not good. For example, the subsection 1.1 and section 2 should be reorganized.
2. The contributions of this paper should be highlighted at the end of the introduction. In addition, more information on the background and problem statement should be added.
3. Detailed comparison with prior works should be added. Please give reasoning of the results and a deeper explanation.
4. The validation is a big concern in this work. How authors can avoid bias validation? Authors can perform a significant test to show the efficiency of the method.
5. The linguistic quality needs further improvement.
6. Literature review on the data-based methods of fault diagnosis is limited. More recently-published papers in this field should be discussed. The authors may be benefited by reviewing more papers such as 10.1016/j.ymssp.2022.109569 and 10.1016/j.ymssp.2022.109834.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report (New Reviewer)
This work lie-s on the proposal of a framework that aims to improve the bearing fault diagnosis accuracy by using a hybrid feature selection method based on Wrapper-WPT. The proposal is interesting since most of the common faults that appear in rotating machinery are related with bearings. However, some issues must be addressed.
1.) In Section 1 Introduction, to many paragraphs are included, it is better if this section only is composed by three or four paragraphs.
2.) As it is stated, Deep Learning (Stacked Auto-Encoders) techniques have been recently proposed, in this regard, these techniques can be based on the use of Genetic Algorithms to improve the feature learning, as well as the use of different physical magnitudes. Can the authors include a brief discussion about the use of multiple physical magnitudes with Deep Stacked Autoencoders? Please consider the following references: https://doi.org/10.3390/s21175832 ; https://doi.org/10.3390/s21248453
3.) If Sub-section 1.1 is included as the theoretical background, please include this sub-section as a Section named “Theoretical background/basis”, and also include references to support the text.
4.) In general, for the whole manuscript, small paragraphs are included, please try to fusion or combine paragraphs to make more large paragraphs.
5.) Please improve the flow chart of Figure 11 to ensure that this representation matches with the steps that must be performed and explained between lines 431-461
6.) Also, it is mentioned that a feature normalization is carried out prior to the feature selection, can the authors include more details about this normalization? Include equations and its representative processing.
Most of the times advantages of the proposals are highlighted, but, what are the disadvantages of the proposed method? Please discuss them and include them in the conclusion section.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report (Previous Reviewer 2)
The authors have very carefully addressed my comments and significantly improved the quality of the manuscript. I suggest accepting the paper for publication.
My only minor suggestion is to color code the table cells identifying best accuracy. A sample of the table can be found in Table 6 of https://doi.org/10.1016/j.neucom.2021.12.035.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report (New Reviewer)
The revision has addressed all my issues. The quality has improved a lot after revision. I recommend it for publication.
Author Response
Thanks for your positive response.
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
The manuscript is good written and organized. The described methodology is easy to understand and follow.
However, it suffers from very a low novelty. The listed contributions on pages 3 and 4 do already exist in the literature and already known to the engineering community. All type of Wavelet Analysis (incl. Wavelet Packet Analysis) is very well-known. The Boruta algorithm and its application to the bearing fault diagnosis has been already investigated by other researchers. The Subspace k-NN method is very well-known as well. In addition, there are researches in the literature which have investigated, for example, the performance of the Boruta Feature Selection using a signal decomposition method (e.g. Wavelet Analysis).
Therefore, the manuscript can be considered as a case study, and may not be considered as a scientific publication which can contribute to the engineering community.
The publication of the manuscript is not recommended.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
In this paper the authors propose a novel method for bearing fault detection. This topic is current and has good readership. However, more work needs to be done to establish the generality of the proposed approach. I suggest a major revision before publishing and below are some comments that can guide the revisions:
1) The author's use the Boruta Feature Selection process. How does the use of other feature selection processes compare to Boruta? As of now it looks like this feature selection process works purely based on the final results. Rather, I suggest the authors to provide more insight at each step of novelty and compare what other features would have been selected if we were to use another feature selection process and why such a selection actually degrades the classification accuracy?
2) The proposed methodology is demonstrated on one case study. It is not clear if the dataset is obtained a part of this study or is publicly available? (citation seems missing). Either way I strongly recommend the authors to exercise their framework on other famous bearing diagnostic datasets. This not only helps generalize the approach, but also significantly boosts the credibility and readership.
3) Was any encoding used for the class labels? Simply using 1, 2, 3, 4 is not correct when posing a classification problem - because the classes need to be mutually independent whereas "4" is close to "3" and not "1".
4) The introduction can be significantly improved. There is an extensive bearing diagnostic literature and thorough discussion is needed to provide proper context for this paper.
5) In Section 3.4.1, the authors mention "it is evident...", "which is normally supposed to contain..", "these components..would have been omitted when using...". There are lot of claims without sufficient proof. Proper citations along with mathematical proofs are required for each of these sentences.
6) The figures need significant improvement. Clear y axis labels are needed (I see that the authors have the y-axis labels as titles in figs such as 9). "Figure 9" is repeated twice.
7) In Fig 9 "Boruta feature selection scheme" - shouldn't the "Shadow features" block be different from "Original Features" after "duplicating and shuffling"?
8) I recommend the authors to share their codes publicly of repeatability. If the proposed method cannot be disclosed publicly, the authors can consider posting the codes for the other methods used for comparison based on literature.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
In the proposed work the author’s claim of presenting a novel approach of bearing fault diagnosis method based on Wavelet Packet Transform (WPT) combined with Boruta Feature Selection and Subspace k-NN is proposed in this paper. In my opinion, the paper is worth being published in the “Machines” after the major review as given below:
Major comments:
1. The abstract of the paper should be revised carefully? I suggest making it short and concise it more? A maximum of 200 words would be enough.
2. The introduction is too long, it should be revised carefully and make it more concise and also cite the recent paper in this field. I would recommend including the following papers in the literature review.
3. Do the author use the de-noising method? Can the author explain it more?
4. Do the author normalize the data? Kingly overview of this?
5. The problem statement of this should be clearly revised in the 2nd last paragraph of the manuscript.
6. Can the author explain how the artificial fault is induced in the bearing? And how it can be correlated to the real fault in the bearing? Can it occur in a real environment?
7. What are the challenges in handling the vibration data for the fault detection of the bearing fault?
8. How can the author justify that the vibration data can be used in real application in an industrial environment, and how can alteration in the domain be justified?
9. How the vibration data can be used in a real environment of bearing fault detection? Can the author justify that “the working environment would be bulky in case of real application”? Can the author suggest some other techniques, like the embedded system of the working environment?
10. Why the author used these specific features? Is there any reason to use the list of different features? The author should clarify this.
11. Please explain the faults inflicted by accelerated lifetime tests. How this test is measured and labeled? How long is each type of bearing used, kindly specify the time for each bearing usage too.
12. I would recommend to revised the paper and section. For that I would suggest the style of the following paper, please follow this paper.
https://doi.org/10.1093/jcde/qwac015
13. The authors only showed the training accuracy and didn’t test the model. Why is the testing not considered? Can it be justified? The accuracies might be higher because of the overfitting issues. The authors are requested to highlight this issue.
14. The manuscript should be revised with the native English correction? Besides, plagiarism should be reduced, it is 20% please reduce under 15% to the maximum.
Minor comments:
1. The whole table or figure should be on the same page? Considered this in the whole manuscript? For instance, table 3 is covering pages 6 and 7, it should be on the same page? Same as table 4????
2. Please remove the comma (,) or full stop (.) sign after every equation. Revise it throughout the whole manuscript.
3. Please use the caption below the figure such as (a), (b), etc. which is written above the figure in figures 10, 11, 12, and 14.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors took care of most of my comments but did not address my suggestion of deploying their algorithm on benchmark bearing diagnostic datasets like the Case Western Reserve dataset. Without doing this, the generality of model application to other datasets is not well established and the 99% accuracy numbers could be interpreted as a result of the model tailored to a single dataset.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Accept as it is
Author Response
Thanks for the reviewer's positive response.
