InvMOE: MOEs Based Invariant Representation Learning for Fault Detection in Converter Stations
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript introduces a novel and highly applicable model, InvMOE, combining Multi-task Learning and Invariant Representation Learning to address specific challenges in fault detection at HVDC converter stations. However, the authors need to provide more detailed explanations of the proposed model and include more in-depth analyses to enhance the study's credibility. Below are suggestions for revisions before the manuscript is considered for publication:
1. The research uses Swin Transformer for feature extraction but does not explain in detail why Swin Transformer was chosen over other popular architectures like Vision Transformer (ViT) or CNN. The authors should include diagrams or illustrations showing how Swin Transformer extracts features (compared to other models) to clarify its role in analyzing images from converter stations.
2. In the description of Multi-task Learning, the "adaptive routing" mechanism of MOE is presented too briefly and does not clearly explain how the "gating network" operates. Figure 1 provides an overview of the framework but needs further elaboration as it is the backbone of the study:
- "Gating Network" that routes inputs to the "Experts" should be described in detail -> Explain by Figures
- "Experts" should be specifically explained in relation to the tasks and associated datasets, such as fault types, and the relationships between the experts should also be made clear. -> Explain by Figures
3. The dataset (Table 1) shows data imbalance (Task 5 only has 100 samples), but the authors do not address this issue. The study does not propose any solution for handling data imbalance during model training. Additionally, the authors should provide information on the scale and representativeness of the dataset. Is the experimental data sufficient to ensure generalization and applicability across different converter stations worldwide?
4. In experimental results, the conclusions are rather general, focusing on numerical results without detailed analyses, such as why InvMOE outperforms baseline models, especially for challenging tasks like "valve cooling water leakage detection." The authors should provide more in-depth analyses, clarifying the performance of the models in each specific task to better highlight the proposed model.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsComments:
1. Section 1 is regular in content considering that the authors added in the same section the contributions that should be contained in a different document or section, it should be improved.
This section should contain the main bases of the article, from concepts to related works, as well as a brief description of the sections that make up the article.
2. In section II the authors do not show or validate the relation of their work with the related works where they specify the advantages or disadvantages of their proposal.
3. The authors put figure 1 in the article but it is not referenced or mentioned in any paragraph to know specifically what it shows or what they want to express.
4. Section III is good but not very friendly, especially when the mathematical part is shown, where the symbolic part and the numerical part would be expected to show the results of the models or equations obtained that define the proposed system. Improve it.
5. Section IV is good but lacks some more depth in the results obtained, and has the ability to compare with related works.
The authors do not show all the results obtained, which is an essential part of knowing if the proposal reported in the article is as expected or is good and can be validated.
6. Section V is good but could be improved to be more interesting to the reader.
7. Before reaching section VI, a discussion section should be added.
8. Section VI for the moment could improve the discussion section.
The references are all located in the article.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsIt should be noted that the paper addresses the fascinating topic of energy conversion between AC and DC. This is important because of the increasing use of high-voltage renewable energy sources, such as solar cells, wind turbines, or fuel cells. One can also see a rather modern approach to fault diagnosis of energy conversion stations. Such stations will be used increasingly so that failures will be an increasing problem for maintenance engineers.
Good points of the work:
1. The paper has an excellent theoretical introduction and a well-prepared literature chapter.
2. The algorithm of the proposed solution is very well presented.
3. A summary of the experimental results is sufficient.
The work also has several discussion points, which I list below:
1. What is the reason for such a short list of literature? I assume that the topic presented is widely discussed in the literature due to the use of agent-based systems in the prepared algorithm. Is it not possible to expand the literature a little?
2. Is it not possible to post larger graphics showing the damage to energy conversion stations? They would be more readable.
3. Why do the authors not consider the use of current overload as another fault in the prepared algorithm? Non-contact temperature measurement with bolometer sensors could be used for this purpose.
4. Why do the authors not consider the measurement of coolant temperature in converter stations? This could have been another factor in the prepared algorithm.
5. The paper does not indicate whether the prepared algorithm will work in real mode at the energy conversion station to help prevent failures of such stations. If so, please indicate what elements will be used to implement the prepared algorithm. Perhaps a block diagram of such a device should be prepared.
I believe that the guidance provided will help to improve the work submitted for assessment.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe author's response thoroughly addressed all issues raised by Reviewer. The revisions and additions were detailed and well-justified, including clarifications on the selection of the Swin Transformer, a comprehensive explanation of the "adaptive routing" mechanism in Multi-task Learning with Mixture of Experts (MOE), and proposed solutions for handling data imbalance. Additionally, the deeper analysis of experimental results enhanced the scientific rigor and clarity of the manuscript. Based on these comprehensive revisions, Reviewer recommends the publication of this manuscript.
Author Response
Thank you.
Reviewer 2 Report
Comments and Suggestions for AuthorsComments:
1. The abstract complies with what is indicated within the guidelines of the defined template.
2. Section I:
a. The authors added text giving a better foundation to the section.
b. The authors also added a very brief description of the sections that make up the article, but it is too brief, I suggest improving it.
Section II:
a. The authors added a paragraph at the end of the section, giving it a better foundation for the reader to have more information and justification of the section itself.
4. Section III:
a. The authors begin with a brief but punctual description of the subsections that compose this section.
b. In subsection 3.1 the authors added a somewhat lengthy paragraph on strengthening, giving a better foundation to the writing.
c. In subsection 3.2.1 the authors added a better explanation to give the reader better information and better explained for its understanding.
5. Section IV: the authors kept the content as in the first revision version, but if it is necessary to add more results obtained, this comment was already given in the first revision.
6. Section V: the authors kept the content as the first version under review.
7. Section VI: the authors kept the content as the first version to be revised.
8. The authors have not added a formal discussion section to provide validation and comparative grounds for their proposal with respect to related works; this is important to include.
9. The references are those indicated and used in the article.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 2 Report
Comments and Suggestions for AuthorsComments:
1. The article is not in the journal format.
2. Section I was updated to give the section a better foundation.
3. The authors updated section II by adding a summary of the section giving better understanding for the readers.
4. Section III was updated in the first paragraphs, as well as updated in other sections of the section giving better focus and interest to the reader.
5. Section IV of experimental results was updated in section 4.2.2, updated tables 2 and 3 of section 4.2.3, updated section 4.3; generating a better foundation to the section.
6. The authors updated section 5, although it is noteworthy that at the beginning of the section there is an introductory paragraph; they updated section 5.1.
7. The discussion section in text, expected a strong comparative discussion of their proposal vs. related works, improve this section.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf