A Semi-Supervised Approach for Partial Discharge Recognition Combining Graph Convolutional Network and Virtual Adversarial Training

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1.      The surface discharges mostly occurs at the rising point of sinusoidal waveform but in the current work some discharges are observed also at the maximum point of AC waveform. What is the reason?

2.      The GCN approach of classifying the different discharges resembles the same as PCA technique. What is the difference in approach with PCA and GCN?

3.      The change in percentage between CNN and other DL techniques is very high. What is the reason? Explain.

 

Author Response

Thank you for your attention to our manuscript entitled “A Semi-supervised Approach for Partial Discharge Recognition Combining Graph Convolutional Network and Virtual Adversarial Training” (ID: energies-3163938). We are very grateful for all your comments on the manuscript. Those comments are all valuable and very helpful for revising and improving our work. We have studied the comments carefully and have made corrections which we hope to meet with approval. Here below is our description of the revision according to the comments. The main corrections in the paper and the responses to the reviewer’s comments are as follows.

Comment 1:

The surface discharges mostly occurs at the rising point of sinusoidal waveform but in the current work some discharges are observed also at the maximum point of AC waveform. What is the reason?

Response 1:

Thank you for pointing this out. As experts said, surface discharges mainly occur on the rising edge of the sinusoidal waveform, but discharge phases are also associated with the uniformity of electric field in the discharge gaps. Literature [1] has elaborately investigated the diverse electric field non-uniformity and discovered that as the the uniformity of electric field in the surface discharge gaps gradually decrease, partial discharge will be observed at the maximum point of the AC waveform. In this manuscript, the surface discharge is simulated by a ball-plate gap with , and the obtained PRPD spectrum is consistent with the surface discharge of the ball-plate gap in reference [1].

*[1] ZHOU Yuanxiang, LI Yongyin, CHEN Jianning, BAI Zheng. Defect Recognition Method of Oil-paper Insulation Based on Information Fusion of PRPD Spectrum and Dissolved Gas Data. Insulating Materials, 2023,56(12): 43-53.

 

Comment 2:

The GCN approach of classifying the different discharges resembles the same as PCA technique. What is the difference in approach with PCA and GCN?

Response 2:

Thank you for your comments. PCA is primarily employed to mitigate redundancy in low-dimensional data while preserving essential information. It serves as a feature extraction technique that is often integrated with machine learning to facilitate classification tasks. Traditional machine learning methods extract features using PCA and train classifiers separately, which may yield suboptimal results. In contrast, GCN, a type of deep learning architecture, automatically extract features for classification. The classification results will be transferred back to guide feature extraction, improving partial discharge recognition..

 

Comment 3:

The change in percentage between CNN and other DL techniques is very high. What is the reason? Explain.

Response 3:

We apologize for the trouble that may have caused you. This study compares CNN with various DL methods, including supervised GCN and the new semi-supervised approach. The factors that account for the discrepancies in outcomes are found in: 1) with the full set of labeled samples, CNN and supervised GCN show comparable recognition rates; 2) the proposed approach uses unlabeled samples and the VAT technique to improve the supervised GCN, so it significantly improves the recognition capabilities; 3) when labeled samples are reduced by 50%, CNN's performance declines due to its dependence on large datasets, while GCN maintains better recognition by autonomously extracting numerical data and integrating topological relationships from the PRPD. 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

In this manuscript, the results of this research are conveyed thoughtfully and completely, and they are consistent with the experimental findings. However, the authors failed to explain and draw out the novelty of the work, this aspect needs to be improved. This work is worthwhile to be publish in this journal after major revision. The following issues should be addressed:

1. Introduction is not organized and the importance and novelty of the research should be highlighted and more clearly stated. The authors should give some examples of works in the bibliography, to clear the advantage of their work in comparison with those works. Some references that could be help:

 10.1109/MEPCON47431.2019.9008209

10.1109/CEIDP.2011.6232758

10.1109/CEIDP.2012.6378793

10.11591/ijece.v4i4.5693

10.15676/ijeei.2012.4.2.10

10.1109/ACCESS.2024.3420230

10.1109/MEPCON58725.2023.10462259

10.15676/ijeei.2014.6.4.5

2. The authors are responsible for the English, which should be polished throughout the manuscript to clear some minor typo/grammar errors.

3. Abbreviations should be added in its first appearance.

4. Could you elaborate on the specific challenges associated with labeling on-site PD data, and how these challenges affect the training of deep learning models?

5. The abstract is concise and provides a clear summary of the study's purpose, methods, and key findings.

6.  The abstract states that the approach was validated using test and on-site data. Could you clarify the types of data used for validation, and whether the results were consistent across different data sources?

7. ou mention that the proposed approach improves PD recognition rates by 6.14% to 14.72% compared to traditional methods. Could you explain the metrics used for this comparison

8. ou mention various deep learning techniques like DBN, SAE, and LSTM. How does the proposed GCN and VAT combination compare to these techniques in terms of performance and complexity?

9. The introduction discusses the use of unlabeled data in training deep learning models. Could you provide more details on how the proposed method specifically utilizes unlabeled data, and how this affects the model’s overall accuracy?

10. The introduction mentions that GCN has advantages over CNN in small-scale datasets. Could you provide specific examples or case studies where GCN outperformed CNN, and discuss any limitations observed in larger datasets?

Hence, I recommend it accepted for publication after Major revisions.

 

Author Response

Thank you for your attention to our manuscript entitled “A Semi-supervised Approach for Partial Discharge Recognition Combining Graph Convolutional Network and Virtual Adversarial Training” (ID: energies-3163938). We are very grateful for all your comments on the manuscript. Those comments are all valuable and very helpful for revising and improving our work. We have studied the comments carefully and have made corrections which we hope to meet with approval. Here below is our description of the revision according to the comments. The main corrections in the paper and the responses to the reviewer’s comments are as follows.

 

Reviewer:

In this manuscript, the results of this research are conveyed thoughtfully and completely, and they are consistent with the experimental findings. However, the authors failed to explain and draw out the novelty of the work, this aspect needs to be improved. This work is worthwhile to be publish in this journal after major revision. The following issues should be addressed:

Comment 1: 

Introduction is not organized and the importance and novelty of the research should be highlighted and more clearly stated. The authors should give some examples of works in the bibliography, to clear the advantage of their work in comparison with those works. Some references that could be help:

10.1109/MEPCON47431.2019.9008209

10.1109/CEIDP.2011.6232758

10.1109/CEIDP.2012.6378793

10.11591/ijece.v4i4.5693

10.15676/ijeei.2012.4.2.10

10.1109/ACCESS.2024.3420230

10.1109/MEPCON58725.2023.10462259

10.15676/ijeei.2014.6.4.5

Response 1: 

Thank you for pointing this out. We have revised the introduction of the manuscript to emphasize the importance and novelty of this research. For detailed modifications, please refer to the Section 1 (page 1 and 2) of the revised manuscript, in which these alterations are distinctly marked in yellow.

 

Comment 2:

The authors are responsible for the English, which should be polished throughout the manuscript to clear some minor typo/grammar errors.

Response 2: 

Thank you for your comments. We carefully checked the manuscript to clear some minor typo/grammar errors. At the same time, to ensure the quality of the article, we have invited a native speaker to check and revise the article. Relevant modifications are as follows:

1. Page 1, line 32

“manually extracting features” has been modified to “manually extracted features”.

2. Page 2, line 86

“the Laplacian matrix L (L=D-A) can serve as an alternative representation as well” has been modified to “the Laplacian matrix L (L=D-A) can also serve as an alternative representation”.

3. Page 13, line

“Firstly, obtain the foreground information of these PRPD images through image segmentation, and gradually remove the redundant elements such as the grid lines, the power-frequency sinusoidal waves and the coordinate axis. Finally, each image is converted into a PRPD matrix of size 128×128” has been modified to “Firstly, the foreground information of these PRPD images is obtained through image segmentation, gradually removing redundant elements such as grid lines, power-frequency sinusoidal waves, and coordinate axes. Finally, each image is converted into a PRPD matrix of size 128×128”

4. Others

This part will outline other modifications in the table, so we won't enumerate them individually.

Error type	Position
Articles & prepositions	1) Page 1, line 39; 2) Page 2, line50, line 84; 3) Page 3, line 98; 4) Page 5, line170; 5) Page 8, line 232
Spelling & expression	1) Page 2, line 77; 2) Page 2, line 86; 3) Page 5, line 157; 4) Page 6, line 206;

 

 

Comment 3:

All abbreviations must be defined and explained when they first appear in the text.

Response 3: 

Thank you for your comments. We have added abbreviations to some technical terms when they first appear. These revisions can be found on page 1, line 35, and page 2, line 50.

 

Comment 4:

Could you elaborate on the specific challenges associated with labeling on-site PD data, and how these challenges affect the training of deep learning models?

Response 4: 

Thank you for your comments. At present, on-site PD data labeling mainly relies on the rich experience and professional knowledge of the maintainers. Despite this, manually labeled data may have diagnostic inaccuracies due to the complexity of PD in transformers. If mislabeled data are mixed into the training set, it will mislead the training of DL method. In specific intricate scenarios, the most precise PD labels can only be acquired through the removal of the transformers' covers for internal inspection, which implies that the labeling of data will require substantial human and material resources. Consequently, such on-site data are often recorded in an unlabeled format, and it poses challenges for the conventional DL approach in effectively utilizing the valuable on-site PD data. This study aims to incorporate valuable unlabeled data into the training process to enhance the recognition capabilities of deep learning models, thereby mitigating the potential for misleading training outcomes that may arise from error labels.

 

Comment 5:

The abstract is concise and provides a clear summary of the study's purpose, methods, and key findings.

Response 5: 

Thank you for the recognition and encouragement.

 

Comment 6:

The abstract states that the approach was validated using test and on-site data. Could you clarify the types of data used for validation, and whether the results were consistent across different data sources?.

Response 6: 

We apologize for the trouble that may have caused you. The approach is validated using test and on-site data. The test data from test platform comprises of point discharge, surface discharge, gap discharge and floating discharge. However, the grid companies and equipment manufacturers typically do not disclose extensive on-site data publicly, so we have compiled a limited on-site data that includes three types: surface discharge, gap discharge and floating discharge.

Under the two data sources, the recognition rates are basically the same. The recognition rates for surface discharge and gap discharge are both about 90%. For floating discharge, the recognition rate of on-site data is marginally lower than that of test data, attributable to the intricate operating conditions of transformers. The majority of samples are accurately identified, which still proves the efficacy of the proposed approach. For details, please see Sections 5.4 and Section 5.5.

 

Comment 7:

ou mention that the proposed approach improves PD recognition rates by 6.14% to 14.72% compared to traditional methods. Could you explain the metrics used for this comparison.

Response 7: 

We apologize for the trouble that may have caused you. The metrics used for this comparison is the average recall rate of four PD types, where the recall rate represents the proportion of accurately identified samples relative to the total number of samples within the class. In the revised manuscript,we have added the description of the metrics, as shown in page 12, line 319.

 

Comment 8:

ou mention various deep learning techniques like DBN, SAE, and LSTM. How does the proposed GCN and VAT combination compare to these techniques in terms of performance and complexity?

Response 8: 

Thank you for your comments. DBN, SAE, and LSTM are primarily designed to process one-dimensional (1D) sequential data and cannot directly handle two-dimensional (2D) PRPD. During the study, we have previously conducted a comparative analysis of these DL methods by transforming the 2D PRPD into 1D inputs. The findings indicate that the above methods has minimal computational complexity because of its straightforward input data format and network design, while the proposed method demonstrates a slightly higher complexity. However, the proposed approach can get higher recognition rate by integrating the numerical and correlation information of PRPD and employing the VAT. For the diagnosis of discharge defects, accuracy is more important than computational complexity.

Additionally, it is important to note that several studies have evaluated the performance of DBN, SAE, LSTM, and CNN. These studies have indicated that CNN exhibits superior recognition rates when analyzing PRPD data. Hence, considering the limitations on the manuscript's length, the comparative analysis against DBN, SAE, and LSTM is not included.

 

Comment 9:

The introduction discusses the use of unlabeled data in training deep learning models. Could you provide more details on how the proposed method specifically utilizes unlabeled data, and how this affects the model’s overall accuracy?

Response 9: 

The use of unlabeled samples occurs in two phases. In the first phase, the GCN model identifies unlabeled samples and assigns pseudo-labels to those with high confidence. These samples are added to the training data, and the GCN model is updated in the next iteration. The second phase applies maximum perturbation to both labeled and unlabeled samples using the VAT technique and updates model to ensure minimal output variation before and after perturbation. Ablation experiments indicate that the proposed method improves recognition rates by 10.83% over the LS model and 5.41% over the PL model. For more details, please see Section 5.3 in manuscript.

The revised version adds the details of the special utilization of unlabeled data to Section 3 (Page 3, line 118) for facilitating readers' understanding.

 

Comment 10:

The introduction mentions that GCN has advantages over CNN in small-scale datasets. Could you provide specific examples or case studies where GCN outperformed CNN, and discuss any limitations observed in larger datasets?

Response 10: 

Thank you for your comments. A comparative analysis of CNN and GCN has performed with different sizes of labeled samples. With the complete labeled data, CNN achieved 88.74% accuracy, while GCN reached 89.57%, showing minimal difference. However, when labeled samples are halved, GCN's recognition rate exceeds that of CNN by about 3.5%, indicating a significant performance gap. For more details, please see Section 5.4 (Page 12) in this manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper provides a semi-supervised approach for partial discharge recognition, which is very important in electrical engineering. The result helps to save the time and labor costs for labeling on-site PD faults. My concerns can be found below;

(1) A comparison review between semi-supervised approach and other methods should be summarized.

(2) The resource of Equation (1), (2) should be given, for example, some important references are suggested to be provideds.

(3) what working conditions can this approach be used, for example transformer, substation etc.

Author Response

Thank you for your attention to our manuscript entitled “A Semi-supervised Approach for Partial Discharge Recognition Combining Graph Convolutional Network and Virtual Adversarial Training” (ID: energies-3163938). We are very grateful for all your comments on the manuscript. Those comments are all valuable and very helpful for revising and improving our work. We have studied the comments carefully and have made corrections which we hope to meet with approval. Here below is our description of the revision according to the comments. The main corrections in the paper and the responses to the reviewer’s comments are as follows.

 

Reviewer:

This paper provides a semi-supervised approach for partial discharge recognition, which is very important in electrical engineering. The result helps to save the time and labor costs for labeling on-site PD faults. My concerns can be found below:

Comment 1:

A comparison review between semi-supervised approach and other methods should be summarized.

Response 1: 

Thank you for pointing this out. In this revision, we have incorporated a comparative analysis of supervised learning, unsupervised learning, and semi-supervised learning within the introduction of the manuscript. Additionally, we have conducted a review of the many semi-supervised methods. For further details, please refer to the highlighted parts in the Page 2, line52.

 

Comment 2:

The resource of Equation (1), (2) should be given, for example, some important references are suggested to be provided.

Response 2:

Thank you for pointing this out. In this revision, we have incorporated references to Equation (1) and (2), which have been highlighted in yellow. The modification position is in Page 3, line 102 and Page 4, line 134 of the manuscript.

 

Comment 3:

What working conditions can this approach be used, for example transformer, substation etc.

Response 3:

The approach is primarily employed for the detection of partial discharge defects in oil-immersed transformers and reactors within substations. The internal insulation architecture of those equipment are subjected to the cumulative influences of elevated temperatures, high pressures, and mechanical stresses over extended durations. This complexity renders the insulation susceptible to the formation of weak points, which are often associated with PD phenomena. The proposed method provides a way to detect PD defects, which helps in promptly evaluating the condition of insulation and assists in pinpointing the defects.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

accepted in the present form