A Multi-Source Consistency Domain Adaptation Neural Network MCDANN for Fault Diagnosis

Round 1

Reviewer 1 Report

This paper proposes a multi-source consistency domain adaptation neural network for fault diagnosis in bearings. In general, this paper is well presented. The idea is interesting with experimental validations.

The following issues can be further considered:

-        The domain mismatch problem can be more discussed. Also, can you add more cites or a review paper citation?

-        The contributions listed in the paper are not novel enough. The authors must highlight the proposals they make as well as the problems they intend to solve.

-        Figures 1, 2, 3 are not very descriptive. The explanation of each module can be improved and illustrated in a better way.

-        The signals to train the model are processed using wavelet transform and converted into an image. The authors use 128 datapoints to generate these images. Some images can be presented as a comparison between the different fault modes.

-        In this regard, using 128 data points does not seem like an appropriate choice. The speed of the motor and the sampling frequency must be considered. A wiser choice is to consider at least one full turn datapoints. This to have a correct characterization of the state of the motor.

-        The model hyperparameters are not properly described.

-        The authors experimentally demonstrate that the use of multiple sources improves diagnostic performance compared to a single source. In addition, according to the experimental results, MCDANN is superior in all test scenarios compared to other domain adaptation methods. However, in terms of accuracy, the results are very close to the DANN method. In this sense, what other advantage do the authors suppose that the proposed method gives compared to the rest of the state-of-the-art methodologies?

-        The authors mention the need to identify anomalies to minimize the occurrence of dangerous situations. Some related recent works on this topic can be reviewed, such as "Deep-Compact-Clustering based Anomaly Detection applied to Electromechanical Industrial Systems", "Sparse Autoencoder-based Multi-head Deep Neural Networks for Machinery Fault Diagnostics with Detection of Novelties", etc.

Author Response

Point 1: The domain mismatch problem can be more discussed. Also, can you add more cites or a review paper citation?

 

Response 1: Thanks for your notice us this problem. We find that figure 11, the confusion matrix in our paper is a good meterial to discuss and explain the domain mismatch problem, and we add some analysis based on the result which is shown in figure 11. Besides, we rewrite the Introduction section and cited more domain adaptation literature related to fault diagnosis.

 

Point 2: The contributions listed in the paper are not novel enough. The authors must highlight the proposals they make as well as the problems they intend to solve.

 

Response 2: LMMD is an greate metric-based method in recent years, it’s very simple and effective. But when we conduct our experiment using our way, this approach achieves poor results in some domains because sometimes the fine-grained alignment doesn’t work well, and we found that muti-source method and prediction consistency can eliminate this phenomenon, that’s way we proposed our method.

 

Point 3: Figures 1, 2, 3 are not very descriptive. The explanation of each module can be improved and illustrated in a better way.

 

Response 3: We have modified the figures or added more description to make it more clear.

 

Point 4: The signals to train the model are processed using wavelet transform and converted into an image. The authors use 128 datapoints to generate these images. Some images can be presented as a comparison between the different fault modes.

 

Response 4: We added some wavelet transform image(Figure 7) in our paper.

 

Point 5: In this regard, using 128 data points does not seem like an appropriate choice. The speed of the motor and the sampling frequency must be considered. A wiser choice is to consider at least one full turn datapoints. This to have a correct characterization of the state of the motor.

 

Response 5: It was an oversight of ours and thanks for your correction. We just mentioned the overlap size of 128, but we forgot to mention that one sample actually include 1024 signal points and now we fixed this mistake. In CWRU dataset, one full turn include about 400 signal points, so the sample we chose include at least two full turn datapoints.

 

Point 6: The model hyperparameters are not properly described.

 

Response 6: We added a figure(figure 6) to describe the network parameters of our model.

 

Point 7: The authors experimentally demonstrate that the use of multiple sources improves diagnostic performance compared to a single source. In addition, according to the experimental results, MCDANN is superior in all test scenarios compared to other domain adaptation methods. However, in terms of accuracy, the results are very close to the DANN method. In this sense, what other advantage do the authors suppose that the proposed method gives compared to the rest of the state-of-the-art methodologies?

 

Response 7: We use majority voting and LMMD to achievement fine-grained alignment, and conbine fine-grained alignment and multi-source to deal with domain mismatch problem. Other method which don’t consider about this issue may face severe performance degradation in some domain once the domain is mismatched during training.

 

Point 8: The authors mention the need to identify anomalies to minimize the occurrence of dangerous situations. Some related recent works on this topic can be reviewed, such as "Deep-Compact-Clustering based Anomaly Detection applied to Electromechanical Industrial Systems", "Sparse Autoencoder-based Multi-head Deep Neural Networks for Machinery Fault Diagnostics with Detection of Novelties", etc.

 

Response 8: We reviewed these paper mentioned and add some intersting view to our paper.

 

 

Reviewer 2 Report

This paper proposes a multi-source consistency domain adaptation neural network MCDANN to achieve fine-grained domain matching and improve the transfer accuracy of the model, which uses sub-domain division alignment and multi-source prediction consistency. this study is interesting. Before Accepted, however, some questions must be addressed as follows:

1. in Abstract, too many descriptions on background are given, please simplify. besides, main results and data should be given to support the conclusions. 

2. in this study, the diagnostic accuracy is 96%. it is very high. becasue the ideal dataset CWRU is applied. the reviewer suggest that the authors should use the real data from industrial system to validate your method. 

3. In the field of fault diagnosis, numerous advaned methods on singnal processing and disgnosis have emergied, such as Applied Science, 2022,12(9): 4668; Structural Control and Health Monitoring, 2021, 28(12): e2839; Shock and Vibration, 2021, 8828317; entropy, 2020, 22(1):57 and Journal of Vibration and Control, 2019, 25(8): 1436-1446.  please comment them and compare them with your methods. 

4.  in this study, sub-domain division alignment and multi-source prediction consistency are only used. no new thing is presented. the authors should give or emphasize the contributions and innovations. 

Author Response

Point 1: in Abstract, too many descriptions on background are given, please simplify. besides, main results and data should be given to support the conclusions.

 

Response 1: Thanks for your advice, we reduced background content and added more data.

 

Point 2: in this study, the diagnostic accuracy is 96%. it is very high. becasue the ideal dataset CWRU is applied. the reviewer suggest that the authors should use the real data from industrial system to validate your method.

 

Response 2: We added another bearing dataset to prove the effectiveness of our model now.

 

Point 3: In the field of fault diagnosis, numerous advaned methods on singnal processing and disgnosis have emergied, such as Applied Science, 2022,12(9): 4668; Structural Control and Health Monitoring, 2021, 28(12): e2839; Shock and Vibration, 2021, 8828317; entropy, 2020, 22(1):57 and Journal of Vibration and Control, 2019, 25(8): 1436-1446.  please comment them and compare them with your methods.

 

Response 3: We reviewed these papers and found that they are good signal-based method, our reserch is data-based method, so we comment them in our introduction as example of signal-based method, which belongs to an another develop direction.

 

Point 4:  in this study, sub-domain division alignment and multi-source prediction consistency are only used. no new thing is presented. the authors should give or emphasize the contributions and innovations

 

Response 4: LMMD is an greate metric-based method in recent years, it’s very simple and effective. But when we conduct our experiment using our way, this approach achieves poor results in some domains because sometimes the fine-grained alignment doesn’t work well, and we found that muti-source method and prediction consistency can eliminate this phenomenon, that’s way we proposed our method, and we updated our contribution to emphasize on this point.

Round 2

Reviewer 1 Report

The authors have addressed properly most of my previous concerns.

 

The following suggestions can be further considered:

The scheme of Figure 6 is not clear enough. Authors are suggested to improve the figure and a more appropriate explanation.

Report if the new DIRG test bench is a public database or an authors test bench. Complement the information with references.

Author Response

Point 1: The scheme of Figure 6 is not clear enough. Authors are suggested to improve the figure and a more appropriate explanation.

Response 1: The original Figure 6 is ambiguous and its description is short as well as ambiguous. We updated this figure and added some sentences to the paragraph above Figure 6 to correctly describe what we want to express here, that is, Figure 6 shows the network structure of the feature extractor, label classifier and domain classifier in MCDANN. Besides, we also added some text descriptions on them.

Point 2: Report if the new DIRG test bench is a public database or an authors test bench. Complement the information with references. 

Response 2: DIRG dataset is a public database acquired on the rolling bearing test rig of the Dynamic and Identification Research Group (DIRG) from Department of Mechanical and Aerospace Engineering at Politecnico di Torino. We added some descriptions and included a reference describing this dataset in our paper.