Fault Diagnosis Method of Special Vehicle Bearing Based on Multi-Scale Feature Fusion and Transfer Adversarial Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a rolling bearing diagnosis method based on multi-scale feature fusion and transfer adversarial learning. Firstly, a multi-scale convolutional fusion layer is designed to effectively extract fault features from the original vibration signals at multiple time scales. Through a feature encoding fusion module based on the multi-head attention mechanism, feature fusion extraction is performed, which can model long distance contextual information and significantly improve diagnostic accuracy and anti-noise capability. Secondly, based on the domain adaptation cross-domain feature adversarial learning strategy of transfer learning methods, the extraction of optimal domain-invariant features is achieved by reducing the gap in data distribution between the target domain and the source domain, addressing research on fault diagnosis across operating conditions, equipment, and virtual real migrations. Finally, experiments are conducted to verify and optimize the effectiveness of the feature extraction and fusion network. The public bearing dataset is used as the source domain data, and special vehicle bearing data is selected as the target domain data for comparative experiments on the effect of network transfer learning. Experimental results demonstrate that the proposed method exhibits exceptional performance in cross-domain and variable load environments. In multiple bearing cross-domain transfer learning tasks, the method achieves an average migration fault 26 diagnosis accuracy rate of up to 98.65%. However, there are some issues in this paper that need to be clarified and modified by the authors, as follows:

1. It is suggested to add an overall flow chart to better understand the work in this paper.

2. It is suggested to add a variety of working conditions to verify the effectiveness of the proposed method.

3. A multi-scale feature fusion network with attention mechanism is designed, which 100 implements one-dimensional multi-scale convolution and position encoding of the original bearing data through multi-scale one-dimensional convolutional network modules. Please explain in detail the theoretical innovativeness of the proposed method?

4. The analysis of experimental data results in this paper is too simple, so in-depth analysis is recommended.

5. I have not seen the results of comparison analysis between the proposed method and the current advanced method.

Comments on the Quality of English Language

The language of this manuscript needs further refinement.

Author Response

Comments 1: It is suggested to add an overall flow chart to better understand the work in this paper.
Response 1: Your suggestion is excellent. An overall flowchart can help readers better understand the methodology of the paper. I have made modifications based on Figure 1 in the paper, providing a complete framework diagram. Additionally, I have revised the explanatory statements below Figure 1, adding a section at the end to explain the process in practical applications for this framework diagram.

Comments 2: It is suggested to add a variety of working conditions to verify the effectiveness of the proposed method.
Response 2: Thank you for pointing this out. The CWRU bearing dataset used in the existing paper includes three operating conditions, and our foundational research is also based on this dataset. Your suggestion is very valuable. To more scientifically validate the effectiveness of the method, I will expand the data with different operating conditions and verify the method in future work.

Comments 3: A multi-scale feature fusion network with attention mechanism is designed, which 100 implements one-dimensional multi-scale convolution and position encoding of the original bearing data through multi-scale one-dimensional convolutional network modules. Please explain in detail the theoretical innovativeness of the proposed method?
Response 3: Thank you for pointing this out.This paper captures signal features simultaneously at multiple scales using convolutional kernels of different lengths, improving the comprehensiveness and accuracy of feature extraction. The position encoding enhancement model increases sensitivity to the positions in time series data, thereby improving fault diagnosis accuracy. The encoded inputs are then fed into an attention mechanism-based network, which adaptively adjusts the importance of different feature channels by computing attention weights, enhancing the model's noise resistance and discriminative feature extraction capability. These methods collectively form the core innovation of a multi-scale feature fusion network with an attention mechanism, providing an efficient and accurate solution for bearing fault diagnosis.

Comments 4: The analysis of experimental data results in this paper is too simple, so in-depth analysis is recommended.
Response 4: Your suggestion is excellent. We have added supplements to sections 4.3 and 4.4 in the paper.

Comments 5: I have not seen the results of comparison analysis between the proposed method and the current advanced method.
Response 5: The comparative analysis results between the proposed method in the paper and the current advanced methods are indeed insufficient. We have added supplementary information in section 4.3 of the paper.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presented a method for fault diagnosis in special vehicle bearings using multi-scale feature fusion and transfer adversarial learning. The authors proposed a multi-scale convolution fusion layer to extract fault features from vibration signals and a transfer adversarial learning strategy to improve diagnostic accuracy across different domains. Experiments demonstrated the method's high performance, with an average fault diagnosis accuracy of 98.65%.

Strengths

• High Accuracy: Demonstrates significant improvement in diagnostic accuracy compared to existing methods.
• Comprehensive Experiments: Validates the proposed method with extensive experiments, showing robust performance in cross-domain tasks.

Weaknesses

• Terminology Issues: The term "confrontation learning" should be revised to "adversarial learning."
• Introduction Quality: The introduction's logic is weak, and the first paragraph appears AI-generated. The relationship between special vehicles and bearings needs clarification.
• Lack of Research Gap: The manuscript doesn't clearly identify the research gap, considering transformers, adversarial learning, and transfer learning are well-established.
• Methodological Clarity: Descriptions of methods and model frameworks are confusing, and the logical flow between figures 1 to 4 needs improvement.
• Transformer Explanation: The manuscript lacks a clear introduction of the Q, K, V matrices in the transformer layers, raising concerns about potential errors in the description.
• Ablation Study Insufficiency: Due to the unclear research gap, the ablation study fails to convincingly demonstrate the paper’s innovative aspects.
• To better contextualize the research, it would be beneficial to mention recent advancements in the field and how this manuscript builds upon them. For instance: IEEE Transactions on Industrial Informatics, 2024, 20(6): 8628–8638; Mechanical Systems and Signal Processing, 2023, 202: 110680.

Recommendation

Revision Required: The manuscript shows significant promise with its innovative approach and high accuracy in fault diagnosis. However, it requires substantial revisions to address terminology issues, improve the introduction's logic, clearly define the research gap, and ensure clarity in the methodological descriptions and model frameworks.

Comments on the Quality of English Language

There are some grammar and typo errors in the manuscript. Please check the manuscript and revise them.

Author Response

Comments 1: Terminology Issues: The term "confrontation learning" should be revised to "adversarial learning."
Response 1: Thank you for your suggestion. I noticed my oversight and have made the necessary corrections in the title.

Comments 2: Introduction Quality: The introduction's logic is weak, and the first paragraph appears AI-generated. The relationship between special vehicles and bearings needs clarification.
Response 2: Thank you for pointing this out. We agree with this comment. I have revised the content of the first paragraph and explained the relationship between special vehicles and bearings.

Comments 3: Lack of Research Gap: The manuscript doesn't clearly identify the research gap, considering transformers, adversarial learning, and transfer learning are well-established.
Response 3: Thank you for pointing this out. We agree with this comment. Our research aims to achieve fault diagnosis for multiple categories of special vehicle rolling bearings by introducing multi-scale convolution and attention mechanism-based feature extraction and fusion, combined with adversarial transfer learning. The method can adapt to different special vehicles, requiring only a small amount of labeled fault data as the target domain data. Through adversarial learning, the diagnostic capability can be transferred, enabling an end-to-end application. The method's diagnostic accuracy was compared with various methods, achieving an average accuracy of 98.65% across six transfer tasks, which is higher than other comparative methods. Additionally, as per your suggestion, I have provided a clear explanation in the conclusion section.

Comments 4: Methodological Clarity: Descriptions of methods and model frameworks are confusing, and the logical flow between figures 1 to 4 needs improvement.
Response 4: Thank you for pointing this out. We agree with this comment. Your suggestion is excellent; an overall flowchart can help readers better understand the methodology of the paper. I have modified Figure 1 in the paper to provide a complete framework diagram. Additionally, I introduced the relationships between the various parts of the diagram in the statements below Figure 1 and added a final section explaining the practical application process based on this framework diagram.

Comments 5: Transformer Explanation: The manuscript lacks a clear introduction of the Q, K, V matrices in the transformer layers, raising concerns about potential errors in the description.
Response 5: Thank you for pointing this out. We agree with this comment. Thank you for your valuable feedback on our manuscript. We have noted that you mentioned the introduction of the Q (Query), K (Key), and V (Value) matrices in the transformer layer was not clear enough, which might cause readers to worry about potential errors in the description. In response, we have conducted a thorough review and made revisions. We have provided a detailed explanation of the Q, K, and V matrices below Equation (4) in the paper.

Comments 6: Ablation Study Insufficiency: Due to the unclear research gap, the ablation study fails to convincingly demonstrate the paper’s innovative aspects.
Response 6: Thank you for pointing this out. We agree with this comment. We noted that you mentioned the ablation study did not convincingly demonstrate the innovative aspects of the paper. In response, we have conducted a thorough review and made revisions. While there is no standalone ablation experiment section in the paper, we have included the ablation experiments in Section 4.4, providing detailed descriptions and analyses of the structural ablation. In Section 4.4, we conducted ablation experiments on two modules and performed feature visualization analysis. The comparative experiments highlight the advantages of our method, especially in the multi-scale feature convolution extraction and the feature extraction and fusion based on the multi-head attention mechanism.

Comments 7: To better contextualize the research, it would be beneficial to mention recent advancements in the field and how this manuscript builds upon them. For instance: IEEE Transactions on Industrial Informatics, 2024, 20(6): 8628–8638; Mechanical Systems and Signal Processing, 2023, 202: 110680.
Response 7: Thank you very much for your suggestions. The two papers you mentioned are excellent examples that explore fault diagnosis from different innovative perspectives, providing a valuable contribution to the research in this field. I have accepted your suggestions and have reorganized the background section, incorporating them at the end of the second paragraph in the background section of my paper.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

To address the issues of inadequate feature extraction for rolling bearings, inaccurate fault diagnosis, and overfitting in complex operating conditions, this paper proposes a rolling bearing diagnosis method based on multi-scale feature fusion and transfer adversarial learning. The proposed the DNMFF model for rolling bearing fault identification has a higher accuracy. The method proposed in the article has certain innovation and engineering application value. The main shortcomings of this article are as follows.

1. The title of the paper is special vehicle bearing, and the experiment in the paper is "a simulated laboratory". Please explain where the special vehicle bearing is reflected in your experiment? What are the differences between it and ordinary bearings?

2. The size of the convolution kernel paper selected 6（51, 101, 151, 201, 251, and 301 ）, why not choose the others, based on experience? Suggest selecting the optimal number of kernel convolutions for sampling optimization algorithm.

3. How to calculate the accurate value of model diagnosis in the paper? By chance or by averaging multiple measurements? Please explain the specific process

4. The DNMFF model mentioned in the paper has the highest recognition accuracy. Please explain why other methods are lower, and what are the advantages of your proposed DNMFF model method? Where can improvements be made in the future?

5. In order to demonstrate the superiority of the proposed method, the paper qualitatively illustrates it by the distance in the graph. It is suggested to add specific numerical quantification representation, and in addition, add horizontal axis annotation to Figure 7

Author Response

Comments 1: The title of the paper is special vehicle bearing, and the experiment in the paper is "a simulated laboratory". Please explain where the special vehicle bearing is reflected in your experiment? What are the differences between it and ordinary bearings?
Response 1: Thank you for pointing this out. We agree with this comment. In the experiment, we used two types of experimental data: one from the publicly available Case Western Reserve University bearing dataset, used for learning source domain data features; the other is the special vehicle bearing data generated by the simulation test bench shown in Figure 5, used for learning target domain features. In this simulation experiment environment, the bearings of a water tank fire truck (or any special vehicle rolling bearing) are placed on the test bench, and normal data and three types of fault data of the rolling bearings are collected by simulating different working conditions, forming the special vehicle bearing data. Finally, adversarial transfer learning is employed to learn the data from both the source target domains, enabling the final model to diagnose faults in the target domain data. 

Comments 2: The size of the convolution kernel paper selected 6（51, 101, 151, 201, 251, and 301 ）, why not choose the others, based on experience? Suggest selecting the optimal number of kernel convolutions for sampling optimization algorithm.
Response 2: Your suggestion is excellent, and we will attempt more sampling for optimization in the later stages. In the initial experiments, we conducted extensive single convolution scale experiments on the data and found that different convolution scales significantly affect bearing diagnosis. These six scales were chosen based on a comprehensive consideration of scale span and initial experimental results. We also attempted intelligent optimization of various hyperparameters within the network. However, for the convolution kernels, we only considered six types for parameter combination optimization. If there are too many parameters, the optimization process takes an exceptionally long time, so we did not conduct more in-depth research at the early stage.

Comments 3: How to calculate the accurate value of model diagnosis in the paper? By chance or by averaging multiple measurements? Please explain the specific process.
Response 3: Thank you for pointing this out. We agree with this comment. In the manuscript, accuracy is used to evaluate the overall performance of the model on its respective categories. The calculation method is as follows:
Acc=(TP+TN) / (TP+TN+FP+FN)
where the symbols are defined as follows:
TP (True Positive): The number of positive samples correctly predicted as positive by the model.
TN (True Negative): The number of negative samples correctly predicted as negative by the model.
FP (False Positive): The number of negative samples incorrectly predicted as positive by the model.
FN (False Negative): The number of positive samples incorrectly predicted as negative by the model.
The evaluation is performed on both training data and test data. The results presented in the manuscript are based on the evaluation of the test data. 
The specific process involves splitting the dataset into a training set and a test set with a ratio of 7:3. The test set is not involved in the training process and is used to test the model after the training is completed.

Comments 4: The DNMFF model mentioned in the paper has the highest recognition accuracy. Please explain why other methods are lower, and what are the advantages of your proposed DNMFF model method? Where can improvements be made in the future?
Response 4: Thank you for pointing this out. We agree with this comment. The advantage of the DNMFF model method is its ability to perform multi-scale feature extraction on the data, which allows for the best representation of data classification. Additionally, the model employs adversarial learning to learn invariant feature representations between source domain data and target domain data, enabling cross-domain and cross-condition fault diagnosis.
        Currently, the data only includes three types of operating conditions. If possible, we could create a more extensive bearing dataset with additional operating conditions to explore the diagnostic performance under multiple conditions.

Comments 5: In order to demonstrate the superiority of the proposed method, the paper qualitatively illustrates it by the distance in the graph. It is suggested to add specific numerical quantification representation, and in addition, add horizontal axis annotation to Figure 7

Response 5: Thank you very much for your meticulous review of my manuscript and for providing valuable comments and suggestions. I have carefully read your review comments and have responded to and revised each one accordingly. Below is a detailed response to the comments you raised:
The content shown in Figure 7 represents the data in the final fully connected layer of the network model, visualized using the method described in reference [39]. This method involves dimensionality reduction of multi-dimensional data features to a two-dimensional range of [-1, 1], displaying the results. The results are clustered into different colors by category, and the distance between clusters is an important reference indicator for evaluating the data feature extraction capability. I have provided a qualitative description of the distances in the analysis below Figure 7 in the paper. We have also added labels to the horizontal axis in Figure 7.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The quality of this manuscript has improved after revisions following the reviewer's commetns. However, there are still some limitations as follows, which should be carefully checked and revised:

(1) The first paragraph is still AI-generated. Please double-check the sentence: "The introduction should briefly place the study in a broad context... See the end of the document for further details on references." This sentence is irrelevant to the work in this manuscript. It should be removed.

(2) In response 7, you have added "Alhams, A. et al. propose a transparent operator network that incorporates a parameterized signal operator node. This node is implemented using a learnable Morlet wavelet operator in the frequency domain, complete with signal-wise gated matrices and skip connections." Although the description is correct, the author name in the main text and paper in the reference list are wrong. Please double-check the citation for transparent fault diagnosis models. 

(3) Please unify the tone and tense in the introduction.

(4) In revised Figure 1, please ensure the correction is made: the loss should be computed from the model output, not included within the model itself.

Author Response

Comments 1: The first paragraph is still AI-generated. Please double-check the sentence: "The introduction should briefly place the study in a broad context... See the end of the document for further details on references." This sentence is irrelevant to the work in this manuscript. It should be removed.
Response 1: Thank you for pointing this out. We agree with this comment. This issue was a serious oversight on my part, and I appreciate your understanding and the opportunity to make corrections. I have removed this entire section.

Comments 2: In response 7, you have added "Alhams, A. et al. propose a transparent operator network that incorporates a parameterized signal operator node. This node is implemented using a learnable Morlet wavelet operator in the frequency domain, complete with signal-wise gated matrices and skip connections." Although the description is correct, the author name in the main text and paper in the reference list are wrong. Please double-check the citation for transparent fault diagnosis models.
Response 2: Thank you for pointing this out. This issue was a serious oversight on my part, and I appreciate your understanding and the opportunity to make corrections. I have revised the name to "Li" and corrected the references accordingly.

Comments 3: Please unify the tone and tense in the introduction.
Response 3: Thank you for pointing this out. We agree with this comment. I have reviewed and revised the content of the introduction section.

Comments 4: In revised Figure 1, please ensure the correction is made: the loss should be computed from the model output, not included within the model itself.
Response 4: Thank you for pointing this out. We agree with this comment. We have made modifications to Figure 1. The main change is that we retained the overall structure of the model while removing the representation of the loss function. Corresponding changes were made to the description below Figure 1.

Author Response File: Author Response.pdf