Deep Residual Network Combined with Transfer Learning Based Fault Diagnosis for Rolling Bearing
Round 1
Reviewer 1 Report
Dear authors,
I appreciate your effort in repairing the work as required by the reviewers. The manuscript is now significantly improved.
Author Response
It was a pleasure to have your suggestion to improve the manuscript.
Reviewer 2 Report
Dear Authors,
It was a pleasure to review your paper. Find below some observations regarding it:
1. In the introductory part of your paper, you are considering only the possible fault detection methods using machine learning. There are several other effective methods based on the measurement of other physical quantities (such as stator current, stray fluxes, thermal image, etc.), which are not so much sensitive to outer disturbances. See the following papers:
- 10.1541/ieejjia.7.282
- 10.1109/ATEE.2013.6563406
- 10.1109/ASET48392.2020.9118361
It would help if you considered (at least mentioned) also these possibilities.
2. Related to the measurement results taken from the Case Western Reserve University; you should have their permission to use these data and especially to include in your paper the picture of the testbench. You should also provide the reference advised being used by the owners not only the address where these data were taken from.
3. You should provide a more detailed description of your testbench, including the types and precision of the used sensors and transducers, and of course the loading of the tested machine.
4. What do you mean by "A Typical view of the test rig".
5. More details are also needed on the way as the detection accuracies were obtained.
6. To make the conclusion section clearer, you have to include the point-by-point findings of this paper, not only write a simple overview of the paper and make a reiteration of what you had performed.
7. As you are using a lot of abbreviations a list of them should be included in the paper.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 3 Report
The article title is confusing and doesn't clearly state the main topic that deals with. So dos with the first sentence in the Abstract. Finally, the article and first sentence of Abstract misleadingly shifts main premise of the Article toward idea of some novel method that proposes and presents new approach in NN transfer learning based on "bearings-fault-principle" which is completely not truth and deviates from actual topic significantly.
Afterall, inspite of starting misunderstandings and confusing expressionism, the article in its very core deals with another application of DNN (with appropriate transferrial learning) but in aspect of fault diagnosis for roller bearings and shows its successful applicability even in systems where high stresses and loads prevents more direct methodological approach in fault prevention and diagnosis.
Further, introductory part where other methods and approaches in topics domain are presented should be broaden significantly by means of geographical viariety of relevant research papers. Especially in part of diagnosis of bearings fault where significant efforts are made, thus relevant papers published, by the western authors and corresponding researches. By citing and referencing papers and authors from a single country or region doesn't reflect significance nor seriousness of conducted research!
Be clear and concise in expressing your main idea and efforts you made. For example, you often refer that main feed data of NN id a 2D RGB image that misleadingly suggest core training data is a spatio-temporal information, by visual origins. Which is, of course, not truth. Real feed data is actually a in frequency domain solely, with temporal dynamics used to construct causal dataset that resembles a form of an image and interpret as that. To be short, as input data refer to CWT not to RGB term.
Why is chosen, by what criteria, the NN of structure given in Table 1, ResNet34 respectively? What are measurable reasons and justification of the choice among other NN architectures? Aside of that, there is nowhere quantitativelly supported in article content the claim that "training error is not much improved" in comparison to more complex/deep NN of ResNets so choosing ResNet34 is somewhat unsupported decision.
How do you split input data set, and why, to: training set and validation set, like in Table 2 and 4? State that clearly and support your decision quantitativelly! As you might know, right division ratio can significantly improve either training time or training deviance (error), also worsen it.
What NN framework and hardware setup you use for your experiment, under what condition/environment?
Also, explain influence of your playground setup onto practical aspects of real-time implementation? Besides of "NN training times" there is no clues what real-time properties of your solution are.
In confusion matrixes (Fig 6. and 9.) are shown training and recognition success rates in absolute values. It is customary and more acceptable to express the results/rates in percentage. Also, to emphasize and support validity of ResNet use, incorporate into same matrix the results of other NN you use for comparison, not only chosen ResNet!
At the end, authors should thoroughly proofread whole article content and lingually (expressionaly and grammatically) adjust to the main premise of it.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
Thank you for considering all my recommendations for improving your paper.
Author Response
It was a pleasure to have your suggestion to improve the manuscript.
Reviewer 3 Report
Despite prior flaws, now the article has been improved significantly with decent scientific soundness and corresponding sufficient quality of presentation.
It is my pleasure to read on achievement of another successful NN implementation in real life application.
Author Response
It was a pleasure to have your suggestion to improve the manuscript.
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
This paper combines deep residual network and Transfer Learning to solve fault diagnosis of rolling bearing problems, both of two methods have been proposed many years. The proposed methods lack originality and novelty from applied intelligence. Authors should improve the quality of this paper. This manuscript should be revised as follows:
- In the content of the article should be more clearly presented what is the original achievements.
- The author did not explain clearly why ResNet34 was used instead of other networks. What are the advantages of adopting ResNet34?
- There are some reference errors in the manuscript (line 181 to line 182) and (line 148 to line 149).
- Please give the computation time and network parameters for each method.
- The "Conclusions" section is too vague and should be extended. Also here would be to highlight the most important achievement.
- Finally, we deep encourage the authors to contribute to the research reproducibility and replicability by sharing the datasets and source codes employed.
Reviewer 2 Report
This paper presents a method for fault diagnosis of rolling bearings based on a deep residual neural network with the transfer learning technique.
The idea of the method is interesting. Authors transfer 1D signals into 2D images with continuous wavelet transform, and then use ResNet and similar NN to re-train the model on the new dataset. However, this idea is not new and has been used before in literature (for example, it has been discussed in "Extreme learning Machine-based classifier for fault diagnosis of rotating Machinery using a residual network and continuous wavelet transform", and others).
The quality of the manuscript should be significantly improved. The main contribution of the paper starts from page 7, and on page 9 there is already an experimental study. The authors spend a lot of time explaining the already familiar theory in Section 2, instead of focusing on their own methodology. Table 1 is not necessary. It is enough to mention which version of ResNet you are using.
Many references are missing in the paper (i.e. "Error! Reference source not found.")
Figures 7,8,9,10 are repetitive and do not add any significance to the quality of the paper.
Mandatory comparison with similar research works is missing.
Unfortunately, I do not see many contributions in this manuscript.
