A Novel Attentional Feature Fusion with Inception Based on Capsule Network and Application to the Fault Diagnosis of Bearing with Small Data Samples
Round 1
Reviewer 1 Report
This paper presents a new method for fault diagnosis of rolling bearing with small samples. The signals are decomposed into different frequencies by EEMD, and then the fusion features are extracted by using initialization module and CABM module. Finally, EEMD features and fusion features are processed as the input of capsule network, and the rolling bearing fault diagnosis under the condition of small samples is completed. The method flow of the article is clear and the content structure is perfect. The author only needs to deal with the following problems:
1. In part 2.1 of the article, the first four order IMF of EEMD is used as the characteristic input. What is the basis for such selection? It is suggested that the author explain the influence of the number of EEMD modes on the overall methodï¼›
2. Pay attention to the relevant format in the article. There is a problem that the formula code is not aligned in the article. The author is recommended to modify it;
3. Whether the content of Fig. 2 in the article is different from the usual network structure representation method, the author is suggested to adjust the content of Fig. 2;
4. It is suggested that the author adjust the text in some pictures in the article to ensure that the pictures are clear;
5. It is suggested that the author mark the horizontal and vertical coordinates in the relevant pictures of the experimental part of the article to clarify the experimental content.
6. As for the use of the EEMD method, it seems that there is no discussion on the signal IMF extraction method in the article. It can be appropriately discussed and explained in the subsequent research or in the relevant parts of the article.
Author Response
thanks for your comments, please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors studied fault diagnosis problem based on deep learning methods in the condition of small samples. This paper is overall interesting and has some merits. The following issues need taken into consideration in a revision.
1. All the abbreviations need to be explained with the corresponding full spellings, including those in the abstract and in the main text.
2. What is the definition of small sample in the context of this work? What is the scale of datasets considered as small samples and did you learn this from industrial practice?
3. The introduction needs major improvement. It shall start from the practical and scientific problems, rather than discussing directly about the solutions.
4. A formal problem formulation section is needed before entering the proposed solution (part 2). How is the problem defined and what are the necessary assumptions, e.g., in terms of data distributions?
5. It is unclear what is the novelty that is originally proposed by the authors in sections 2.1-2.4. They seem to be a combination of existing methods. What are the improvements and how to prove its performance?
6. It is recommended to analysis with SOTA methods, such as in A review on soft sensors for monitoring, control and optimization of industrial processes.
7. English writing can be further polished. There are many very long sentenses that can be shorten or divided into several sentences.
Author Response
please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
This version is better, it is recommended to accept
Author Response
There is not notes to this reviewer.
Reviewer 2 Report
The authors made rather brief response to the review comments. However, it is a bit unclear how the previous comments #2, # 4, #6 were addressed in the revised paper (note, in the main text, rather than in the response).
To further improve, it is recommended to #2 add a formal definition of small sample in the context of this work in the main text; #4 add a PROBLEM STATEMENT/FORMULATION section in between part 1 and part 2; #6 analyze and discuss more related works. Furthermore, English writing needs major revision. E.g., "And" should not be used to start a sentence.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 2 Report
An attentional feature fusion with inception based on Capsnet is proposed in this paper. Overall, the article is interesting, but there are still some problems. I accept its publication after the authors considering my suggestion. Some suggestions and comments are as follows:
1. The research motivation of the article needs to be further organized.
2. Many pictures in the article are not clear enough after being enlarged. It is recommended to modify the format.
3. The novelty of the article is not clear, it seems to be just a combination of existing methods, it is recommended that the authors further reflect the novelty of the method and emphasize in the revised version.
4. The second part of the article is too small to be a separate Section, either merged or expanded.
5. A convolutional block attention module (CBAM) based attentional feature fusion with inception module based on capsule network (Capsnet) is proposed in the paper. However, the authors are recommended to analyse more related work, such as convolutional neural networks and deep learning methods of attention mechanisms in industry application. Consequently, it is recommended to analyze or compare more recent new work, such as Prediction of material removal rate in chemical mechanical polishing via, residual convolutional neural network, Prediction of remaining useful life based on bidirectional gated recurrent unit with temporal self-attention mechanism.
6. The writing in many places in the article is not standardized enough, it is recommended to make corrections.
7. The comparison experiment of the article is not perfect, it is recommended to compare with more existing methods.
Author Response
please see the attachment.
Author Response File: Author Response.pdf
