Long Short-Term Memory Network-Based HVDC Systems Fault Diagnosis under Knowledge Graph

Round 1

Reviewer 1 Report (Previous Reviewer 3)

This paper version is in proper format and improved compared to the original version. I would like to thank the authors for providing adequate answers to the comments. I have some minor comments as follows:

If it is possible, I suggest authors improve the data description further. One should argue that this amount of data is unsuitable for neural networks.

Figure 4 is vague, and I strongly suggest authors change it instead of using the screenshot.

Please fix Table 11.

Figure 14 shows the RNN model is not well-designed, and its hyperparameters are not well-optimized.

I suggest authors improve the quality of Figure 13.

I still would like to see the results of SVM with non-linear kernel functions. Remember, due to the kernel tricks, the calculation amount would not be dramatically increased.

Considering Table 10, why do the models better detect some faults than others? Is it due to some physical meaning?

Have authors consider to compare their models to other methods such as random forest, XGboost, and convolution neural network?

Lastly, regarding the application of the proposed model, do authors suggest using this model instead of what model(s) is(are) used in the industry?  

Author Response

Thank you very much, please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report (Previous Reviewer 1)

Reviewer comment for electronics 2392223

Thank you for revising the paper as the previous comments. The paper is now improved. However, minor revision is required before it can be accepted. Below are some further comments:

1.       In figure 14, the unit and label of the y-axis do not make sense. As stated in the figure, the range of the Accuracy (%) is just within 0.1 – 0.45, which is very low. Please double check and correct accordingly.

2.       English style/writing, some sentences are confusing and not make sense, e.g., the sentence in line 482-484 is quite confusing. Please having proof reading by a native English speaker.  

Please having proof reading by a native English speaker.  

Author Response

Thank you very much, please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report (Previous Reviewer 3)

The authors addressed my comments adequately—congragilation on their manuscript.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Reviewer comment for electronics-2332527

The paper presented a fault diagnosis method for HVDC systems using LSTM network. In general, there is no significant contribution in this paper. Simply, a LSTM network is applied to predict the failure of HVDC system with the comparison to other methods. The authors are advised to hava major revision for the paper before further consideration. Some comments are given below:

1.       In the Abstract section, full term of the acronym: HVDC must be defined before its shorten can be used. The first sentence is quite lengthy, it should be rewritten to improve the presentation.

2.       Double check the acronyms: RMSprop,Adam,SGD, AdaGrad, … in lines 351 and 393, e.g., Win 10?. Make sure they are defined before being used.

3.       The authors should highlight the contribution of the paper as it is unclear in this version

4.       In section 1, the authors reviewed several AI and ML methods such as SVM, RNN, Naïve Bayes. However, a deeper review in this section is required with more relevant AI and ML methods showing the details of these approaches, their advantages, limitations and how they can be the candidates in comparing with the authors’ proposed method. Some relevant works recently discuss about these ML methods in details such as: doi.org/10.1016/j.est.2020.101271, 10.1109/TITS.2020.3028024, 10.1109/ICMT56556.2022.9997802. It’s necessary to review these papers to improve the quality of the paper.

5.       Figure 4 is so blurry; a better picture is required. Not sure what the authors trying to show and how it works in this Figure. A detailed explanation of this graph should be discussed.

6.       Section 4, the authors should differentiate between 239 samples of the fault data and 239 features used in the trained model. Are they presenting the same thing?

7.       Section 5, the authors are advised to discuss about the fault data’s features used for network training. What are they, how they are extracted and what is the training and testing accuracy? Have the authors performed the verification for such methods ?

8.       What are the units of Y-axis in Figure 12 and Figure 14?

9.       Section 6, the authors are suggested to have deeper discussion about the results, benefits of the proposed method. The indicators in this section need references (e.g., where is the reference for the 99.7116% value?)

10.   Section 7, future works should be presented to demonstrate the usefulness of the proposed model and outline the impacts of the study for future research.

11.   Language aspect: the manuscript should be proof-reading, grammar check required. Some sentences are difficult to understand. The term “HVDC systems” is repeated multiple times within the paper, it is necessary to use alternative terms.

Author Response

Thank you very much, Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposes a fault diagnosis method for HVDC systems based on long-short term memory network (LSTM) using data from a knowledge graph platform.

 The paper has nothing new, the LSTM network has been used for classification and regression in many papers. The more interesting is the application, but the paper need to be written again.

 First at all, the more novelty is the knowledge graph, but this is not explained in the paper. The authors talk and talk but do not explain what is the knowledge graph? And they do not put an example of this graph and how they can obtain it?

 You do not need to explain so exhaustively the LSTM network, this is a known technique.

 You must improve the figure 4 that is not possible to see it and explain it so much.

 So, I cannot understand what are the inputs to the LSTM network? Explain it more and put an example.  

 Also in table 3, do you have only 11 samples to test the first fault and so on?. Explain this better. This is not data to train the LSTM network.

 This is not a paper of optimization methods, you do not need to explain so in detail the optimization methods used to train the LSTM network, just to put that you have trained with different optimization methods and put the results.

 What happen with the batch size? You do not explain nothing about tit.

 Do you have optimized the parameters of the other techniques for the comparison? You must do it. Why do you use for example an SVM with linear kernel, and do not use a non-linear kernel?

 Figure 12 it not necessary

Author Response

Thank you very much ,Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

I strongly suggest authors re-write the abstract. Considering why there is a need for a new fault detection system and what is missing in the literature.

The literature review should be improved. 

Figure 4 has poor quality. It is impossible to read it. 

The authors should clearly explain each figure in the context with details.

A detailed description of the dataset and how it is slipt to train models must be included.

How did the authors choose the parameters of their model? How are its hyperparameters tuned? What about other comparison models?

Why is only a linear kernel for SVM chosen?

The authors must compare their model with state-of-the-art models presented in the literature and analytical methods as well.

Table 10 does not sound accurate.

Author Response

Thank you very much ,Please see the attachment.

Author Response File: Author Response.docx