Review Reports - Bearing Fault Diagnosis Based on Multi-Scale CNN and Bidirectional GRU

Round 1

Reviewer 1 Report

In Table 5, while mentioning the method, I can see a particular reference number has been cited. The name of the method should also be included, along with citations.

How to deal with the data diversity of the present moment and moment in the future?

Figures 5 and 6 are the same. Is there no misclassification? Is accuracy 100%? I can’t trust the results. Training data accuracy can be 100%. However, that also represents overfitting which is undesirable.

Include a confusion matrix for training data and test data. Also, while representing training data, specify the classification by the actual number of instances and not its percentage. The same applies to testing data.

What is the possibility of misclassification of a normal condition as a faulty condition depending on the degree of fault in a confusion matrix?

What is the possibility of misclassifying a faulty condition as normal depending on the degree of fault (type II error)?

If the model is deployed in real-time and such a situation arises, how will you identify that the component is in the failure zone and showcased as normal by your system?

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

· Though the authors have used Multi-Scale CNN and Bidirectional GRU, unless there is an adaptive control strategy, implementing the proposed method in a dynamic environment is challenging.

· Why was the transfer learning approach not considered?

· How to tackle unknown random vibrations moments in the future?

· Surprise to see accuracy = 100% in confusion matrices

· Confusion matrices are confusing and hypothetical

· Suppose a machine shop is situated near the railway track. So comment on the feasibility of your framework in such a situation.

· Why did you go for only classification rather than a combination of regression and classification?

· Reinforcement learning approach would be helpful; you may suggest it in the future scope.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper studies the bearing fault diagnosis approach based on the parallel convolutional neural network and bidirectional Gated Recurrent Unit. The introduction states the purpose of the paper, and the relation between the paper and the previous works is clearly explained. It is illustrated with the CWRU dataset that the presented method can identify the bearing fault with high accuracy.

The following are my specific comments:

(1) In the introduction, it is better to provide some remarks about the merits and limitations of the model-based methods, the machine learning-based methods and the combination of the above two methods. Some practical examples for the industrial applications of the machine learning-based methods could be described briefly.

(2) The full names of abbreviations, such as KNN, should be given when they are first presented.

(3) It is stated in the introduction that one of the contributions of the paper is that the convergence speed of the method will be very fast. However, the convergence speed is not clearly demonstrated in Experiments and Results.

(4) I cannot find the definition of k in equation (6). In addition, some parameters or functions are not defined in equation (19).

(5) In Section 2, it is better to provide some explanations about the key idea of the adaptive gradient algorithm.

(6) The authors are suggested to evaluate the computation load of the presented bearing fault diagnosis scheme in comparison with other existing methods in Section 4.

(7) There may be a typo error in the sentence "this experiment is reported in Table ??".

(8) Which is the main approach for the design of the bearing fault diagnosis method to obtain a superior performance in comparison with the existing methods? Try to provide some remarks.