Semi-Supervised Transfer Learning Method for Bearing Fault Diagnosis with Imbalanced Data
Round 1
Reviewer 1 Report
- Figure 1, please make the font size consistent with the font size used in the main content. And change 'Figure 1.' to Fig. 1 or change the content to be Figure 1 for consistency.
- Figure 2, and other figures, please make the font size similar to the font size in the main content.
- Figure 2, please make the x, y axis label readable, and label the units.
- Suggest adding some examples of the CWRU data and a table summary, and discuss some differences between those two datasets.
- How is accuracy defined?
- Figure 6, please make the label the actual fault for easy-to-read.
Author Response
Please kindly refer to the attachment for point-by-point response to the reviewer’s comments.
Author Response File: Author Response.pdf
Reviewer 2 Report
In this paper, the author proposed an efficient bearing fault diagnosis method based on Domain Adversarial Training of Neural Networks (DANN) and uncertainty-aware pseudo-label selection (UPS). And public datasets have been utilized to prove the brilliant performance of the proposed method. The proposed model is innovative in the related areas but the paper needs significant improvement before acceptance for publication. My detailed comments are concluded as follows:
- What’s the main contributions of this paper?
- How to determine the margins between Phase â… , Phase â…¡, and Phase â…¢? Previous experience?
- It’s recommended to condense the discussion on Section 2.2. Or it may be better to remove related discussions to the Section Introduction. And in Section 2.2.1, “Different from the negative learning (NL) proposed by [34], the NL in UPS is designed to include additional unlabeled samples into the training phase and generalizing pseudo-labeling for multi-label classification settings.” A brief introduction on the negative learning (NL) proposed by [34] is recommended to supply. Then, what’s the main points in Section 2.2? In Section 2.2.3, the bearing diagnostic model is proposed, and what’s the exact logical relationship between uncertainty-aware pseudo-label selection and diagnostic model?
- Since α in Equation 7 is adaptive and learned by gradient descent. It seems useless to show the loss function figure, which is shown in Figure 5(b).
- In Section 3.1, citation is required while introducing XJTU-SY dataset, just as CWRU dataset.
- How to split source and target domain dataset? What’s the data size in total for CWRU?
- Please introduce the structure of baseline model.
- In-depth discussion on experiment results should be conducted in relation to the objective problem statement introduced in Section 1. Confusion matrix of other compared models should also be added.
- In Section 4, it seems improper to mention “DANN, DAAN, and UPS” as “traditional independent approaches”.
- There are some minor grammar errors, for example, Line 10 in Section Abstract “Therefore, the data 4 imbalance and the deficiency of labels is a practical challenge in the fault diagnosis of machinery 5 bearings.”, and “STFT applies the window and shifts it to have a 103 fixed temporal resolution for time-domain signals, which construct the spectrogram for the 104 subsequent data input.” The authors should revise the manuscript carefully.
Author Response
Please kindly refer to the attachment for point-by-point response to the reviewer’s comments.
Author Response File: Author Response.pdf
Reviewer 3 Report
The article describes one of the learning methods of the artificial neural networks used for bearing fault diagnosis. The authors offer their original solution to the problem of insufficient amount of data in training set and imbalance between «normal» and «abnormal» samples in it. However, the style of the manuscript, the compactness of some sections and misspellings make the article rather a redrafted text of the conference report than an article. Therefore, it requires certain improvements. General comments include the following:
1. Line 18. The reference to [1-3] at the beginning of the introduction is not clear. The authors note the general role of the rotating machinery and their application in modern industry but refer to the papers devoted to private issues of machine condition diagnosis. Of course, the cited articles contain some references to the role of rotating machinery, but such statements are of general importance and do not require special confirmation.
2. Line 25. The need to use "intelligent" diagnostic methods is unreasonable. Why is it necessary to use exactly "intelligent" diagnostic methods to reduce labor and time costs? Aren’t other "non-intellectual" methods of automatic monitoring and diagnostics aimed at solving the same problems? Perhaps it should simply be pointed out that intelligent diagnostic techniques are one of the most advanced and “trendy” approaches in recent years aimed at solving these problems.
3. Lines 111-112. What are the structures of CWRU Dataset and XJTU-SY Dataset? Why do the authors transfer CWRU Dataset into the form of XJTU-SY Dataset? The reader can only guess about it.
4. Line 143. What VIDEO do the authors of the article have in mind?
5. Figure 4. Given the formula (7), there is no need to write the entire expression at the output of the “Cross Entropy” block.
6. Table 1 location before its explanation in Lines 205-212 raises some questions and misunderstandings at the first reading of the article since the paragraph before the table contains the description of the experiment with 15 bearings (line 200) and the table presents data for 8 bearings only. Of course, Line 208 explains that this is a sample from the original set but it would be more reasonable to place table 1 after its first mention before table 2.
7. Figure 5. Bad location of the legend. It overlaps the informative part of the graphs.
8. Why do the authors compare the effectiveness of their approach with those indicated in Table 3? An explanation is required.
9. Editing of English language and style is required.
Author Response
Please kindly refer to the attachment for point-by-point response to the reviewer’s comments.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors took my comments into account, so I suggest accepting the paper in the current version.
Reviewer 3 Report
The manuscript is sufficiently improved and can be published in Machines.