Classification of Multiple Partial Discharge Sources Using Time-Frequency Analysis and Deep Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

in attached file

Comments for author File: Comments.pdf

Comments on the Quality of English Language

good

Author Response

Comment 1
 In the introduction must give a literature of time-frequency methods that used in partial discharge.
 Response : We appreciate this suggestion. The introduction has been enhanced to incorporate a detailed review of time-frequency methods used in partial discharge signal analysis. Methods such as the Short-Time Fourier Transform (STFT), Wavelet Transform (WT), Wigner-Ville Distribution (WVD), and Empirical Mode Decomposition (EMD) are specifically examined. The revision includes a comprehensive literature review, highlighting key works in the field, discussing the strengths and  limitations of these techniques, and situating the current work within the existing academic discourse.
 Comment 2 In Table 6 and 7, the author must add the figures.
 Response : We have updated Tables 6 and 7 to include the corresponding figures
 to visually represent the tabular data, enhancing clarity by providing an intuitive
 understanding of classification results and performance metrics. This addition ad
dresses previous omissions by offering both numerical and graphical information,
 thereby improving the interpretability of the results in the revised manuscript.
 Comment 3:  The author can give the references of the time-frequency methods the part of Time
 Frequency Analysis for Feature Extraction.
 Response : Additional references related to the application of time-frequency analysis for feature extraction have been added to the manuscript. These citations now explicitly support the discussion of TF methods, ensuring that the literature review
 is comprehensive and well-grounded.
 Comment 4: The comparison between time-frequency methods must applied in signal and calculated after the MSE in the first part for giving the powerful method. The author can use the equations existing in the paper above.
 
Comment 5:
 The images of matrix confusion are not appeared good... The author can’t declare how much used the images...
 Response : The confusion matrix images have been updated with higher resolution and improved formatting. We also enlarged the figures for better readability, and we now clearly describe the number of images used in training, validation, and testing datasets.
 Comment 7: The author can give the diagram of the classification explain step by step the methoduses (CNN and time-frequency technique) and give the percent of image uses of each part in the classification.
 Response : Adetailed block diagram has been incorporated into the manuscript, presented as Figure 2. This diagram encapsulates the entire classification process, starting from data acquisition and TF feature extraction, through to CNN analysis. Additionally, Figure 1 has been added, showing the percentage distribution of image contributions at various stages. Comment 8: How does the integration of FEM simulations, TFA, and DL address the challenges
 of distinguishing overlapping PD signals from multiple sources in real-world HV systems, and to what extent can this approach be generalized to different types of insulation materials and defect configurations?
Response This text discusses how combining Finite Element Method (FEM) simulations, Time-Frequency Analysis (TFA), and Deep Learning (DL) addresses two primary challenges: Effective Distinction of Overlapping PD Signals:
 1. FEMSimulations: These generate detailed datasets that simulate the behavior of partial discharge (PD) signals in high-voltage (HV) systems, capturing interactions and overlaps from various PD sources.
2. Time-Frequency Analysis (TFA): TFA extracts key features by analyzing
 signals in time and frequency domains, crucial for separating overlapping signal characteristics.
 3. Deep Learning (DL): Guided by TFA features and FEM data, DL models, such as Convolutional Neural Networks, are trained to identify and distinguish overlapping PD signal patterns accurately.  Adaptability to Varied Insulation Materials and Defect Configurations:
 1. Calibration and Adaptability: By adjusting DL model parameters, this method caters to a wide range of insulation materials and defect setups.
 2. Experimental Validation: Experiments under different conditions validate the approach’s robustness across various materials and real-world defect scenarios in HV systems.
 

Reviewer 2 Report

Comments and Suggestions for Authors

In this paper the classification of multiple partial discharge sources using time-frequency analysis and deep learning based on FEM PD models. The topic is interesting; however, the main limitation of the proposed analysis is the complete lack of experimental validation. The novelty is poor, the topic has already been widely analyzed in the literature. Just the use of FEA technique for this application is not a novel enough contribution. Following are some comments:

1. It would be useful if the authors provide specific examples where your method offers a significant advantage in real-world HV system diagnostics.

2. Why was polycarbonate chosen as the insulation material, and how does this choice affect the generalizability of the results?

3. Could you provide more detail on how the stochastic models were implemented in the FEM simulations, including the selection of parameters?

4. It would be useful if the authors included a table comparing the performance of your method against traditional PD classification techniques using relevant metrics.

5. More details about the software used and the training dataset to improve reproducibility.

6. Table 10 is impossible to read. The same problems with several figures that are too small.

7. Is there any problem with Table 7? It looks like some figures are missing.

Author Response

Comment 1: It would be useful if the authors provide specific examples where your method offers a significant advantage in real-world HV system diagnostics.
Response: Examples of applications, including the identification of concurrent in
ternal and surface discharges in GIS systems, as well as in power transformers and
 cables, have been included to demonstrate the method’s efficacy in managing signal
 overlap and noise in practical diagnostics.
 Comment 2: Why was polycarbonate chosen as the insulation material, and how does this choice
 affect the generalisability of the results?
 Response: Polycarbonate was selected because its electrical and thermal properties are well-characterized, making it suitable for benchmarking in simulations.
 We have clarified that, while polycarbonate serves as a representative material, the
 methodology is general and can be adapted to other materials by modifying the
 material parameters accordingly.
 Comment 3: Could you provide more detail on how the stochastic models were implemented in
 the FEM simulations, including the selection of parameters?
 Response : The revised manuscript now includes a detailed explanation of the
 stochastic models used in our FEM simulations. This includes specific information on
 parameter settings and the rationale behind their selection, supported by references
 from previous studies.
 Comment 4: It would be useful if the authors included a table comparing the performance of your
 method against traditional PD classification techniques using relevant metrics.
 Response : We have incorporated a comprehensive comparative table that meticulously contrasts the results obtained from our simulations against those derived from
 our experimental procedures.
Comment 5: More details about the software used and the training dataset to improve repro
ducibility.
Response: Software tools used (COMSOL Multiphysics, MATLAB) and dataset characteristics have been detailed in the Reproducibility subsection to enhance clar
ity for replication.
Comment 6: Table 10 is impossible to read. The same problems with several figures that are too small. Is there any problem with Table 7? It looks like some figures are missing.

Response : We apologise for the formatting issues. Table 10 and other affected
 figures have been reformatted and resized to ensure all visual elements are presented
 clearly and are easily readable.

Reviewer 3 Report

Comments and Suggestions for Authors

The relevance of the topic is due to the growing need for reliable diagnostics of high-voltage systems, where partial discharges (PD) are an early indicator of insulation material degradation. Traditional diagnostic methods work effectively in the presence of a single PD source, but in the presence of several simultaneously occurring discharges, there is a significant overlap of signals, which complicates their analysis. Modern industrial systems require automated solutions that can accurately distinguish and classify PDs, which directly affects the safety and reliability of equipment operation.

The paper presents an integrated approach combining finite element modeling (FEM), time-frequency analysis (TFA) methods, and modern deep learning algorithms (CNN). The main contributions are:

• Developing a comprehensive methodology that allows generating synthetic PD data for model training.

• Applying various TFAs (including WT, FRWT, and WST) to extract informative features from signals with overlapping PD sources.

• Demonstration of high classification accuracy (up to 96.67% in accuracy) in complex scenarios, indicating potential practical application in high-voltage system diagnostics.

Comments requiring improvement:

1. There is no comparison of the simulation results with real data. It is recommended to supplement the study with experimental measurements to confirm the practical applicability of the developed approach.
2. It is necessary to describe in more detail the criteria for selecting the model parameters, such as dimensions, material, and boundary value conditions, and justify their compliance with real operating conditions.
3. The description of the convolutional neural network structure requires more detail. It is recommended to provide a detailed description of the parameters, number of layers, types of activation functions used, and regularization strategy to improve the reproducibility of the results.
4. The work does not sufficiently address the issue of the influence of external noise and interference on the quality of PD classification. An analysis of the model's robustness to noisy data should be conducted, which is critical for practical application.
5. It is recommended to expand the comparative analysis with other modern methods of PD diagnostics in order to more clearly demonstrate the advantages of the proposed approach.
6. The authors should consider the possibility of integrating the methods of localization of discharge sources, as this will significantly increase the diagnostic value of the study for practical applications.
7. It is necessary to justify in more detail the choice of specific methods of time-frequency analysis and their parameters, as well as to conduct an analysis of the sensitivity of the model to these parameters.
   These comments are aimed at strengthening the theoretical validity, practical applicability and reproducibility of the research results.

Author Response

Comment 1: There is no comparison of the simulation results with real data.

 Response :An extensive laboratory experiment has been performed to verify the validity of our model and the accompanying simulation results. To achieve experimental validation and enhance our dataset, a thorough validation protocol is carried out, which includes an in-depth comparison of simulation outputs with empirical observations.
 Comment 2: It is necessary to describe in more detail the criteria for selecting the model parameters, such as dimensions, material, and boundary value conditions, and justify their compliance with real operating conditions.
 Response : We now justify the choices for dimensions, material properties, and boundary conditions by aligning them with realistic HV system operating conditions. References to industry standards and previous experimental results have been added  to support our choices.
 Comment 3: The description of the convolutional neural network structure requires more detail.
 Response: We have significantly enhanced our description of the Convolutional Neural Network (CNN) architecture. The revised version now includes specifics on the number of layers, kernel sizes, activation functions (e.g., ReLU), dropout rates, and any applied regularization strategies, ensuring that every aspect of the architecture is clear and reproducible.
 Comment 4: The work does not sufficiently address the issue of the influence of external noise and interference on the quality of PD classification.
Response : We have performed additional experiments by introducing various levels of external noise into the simulation data. The results of these experiments are discussed in the revised manuscript, demonstrating that our methodology retains robust performance even under noisy conditions.
 Comment 5: It is recommended to expand the comparative analysis with other modern methods
 of PD diagnostics.
 Response : In our forthcoming research endeavors, we aim to incorporate attention mechanisms along with LSTM architectures into our convolutional neural network (CNN) model.
 Comment 6: The authors should consider the possibility of integrating the methods of localization of discharge sources.
 Response : We thank the reviewer for this suggestion. While the current study focuses on the classification of PD sources, we have now included a discussion on potential strategies for integrating source localization techniques. This outline serves as a promising direction for future research that could further enhance the diagnostic framework.
 Comment 7: It is necessary to justify in more detail the choice of specific methods of time-frequency analysis and their parameters.
 Response : We have expanded our explanation regarding the selection of time frequency analysis methods and their respective parameters. In addition, a sensitivity analysis has been incorporated to assess how variations in these parameters affect the overall classification performance. This addition provides deeper insight into the
 robustness and flexibility of our approach

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The comparison between time-frequency methods must applied in signal and calculated after the MSE in the first part for giving the powerful method. The author can use the equations existing in paper above:
El Khadiri, K., Elouaham, S., Nassiri, B., El Melhaoui, O., Said, S., El Kamoun, N., Zougagh, H. “A Comparison of the Denoising Performance Using Capon Time-Frequency and Empirical Wavelet Transform Applied on Biomedical Signal” (2023) International Journal on Engineering Applications, 11 (5), pp.358- 365. doi: 10.15866/irea.v11i5.23395 

Author Response

Comment: The comparison between time-frequency methods must applied in signal and calculated after the MSE in the first part for giving the powerful method. The author can use the equations existing in paper above:
5. El Khadiri, K., Elouaham, S., Nassiri, B., El Melhaoui, O., Said, S., El Kamoun, N., Zougagh, H. “A  Comparison of the Denoising Performance Using Capon Time-Frequency and Empirical Wavelet Transform Applied on Biomedical Signal” (2023) International Journal on Engineering Applications, 11 (5), pp.358365.  doi: 10.15866/irea.v11i5.23395 
Response: We appreciate the insightful comment regarding the comparison of time-frequency methods using MSE metrics, such as those outlined in El Khadiri et al. (2023). However, we would like to clarify that our primary goal in applying time-frequency analysis is to transform the raw PD signals into image representations that effectively capture discriminative features for CNN-based classification—not to perform denoising. In our work, the time-frequency representations are designed to preserve key attributes such as amplitude variations and repetition rates, which are critical for distinguishing between one PD source and multiple PD sources. There is a strong argument against denoising in the context of partial discharge signal analysis. Denoising methods, while beneficial in many signal processing tasks, may inadvertently remove or attenuate subtle signal components that are important for PD classification. In PD signal analysis, even what might normally be considered noise can contain valuable information useful for characterizing transient PD phenomena. As such, our evaluation framework is based on how well the time-frequency representations enhance the CNN's ability to correctly classify PD patterns, rather than on the MSE metric used in denoising performance. We have thus compared different time-frequency methods based on the quality and discriminability of the resulting images and ultimately their impact on classification performance (e.g., accuracy, precision, and recall). This approach ensures that we retain all the diagnostically relevant information necessary for accurate PD source identification.

Reviewer 2 Report

Comments and Suggestions for Authors

All the requested improvements have been performed.

Author Response

We sincerely thank the reviewer for their time and positive feedback. We are pleased to know that all the requested improvements have been satisfactorily addressed. We appreciate your support and consideration of our revised manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors tried to take into account the comments and recommendations of the reviewers. After revision, the quality of the article has improved significantly. I recommend the academic editor to consider accepting the article for publication.

Author Response

We sincerely thank the reviewer for the thoughtful comments and positive feedback. We are glad that the revisions have improved the quality of the article.