Inverter-Driven and Stator Winding Fault Detection in Permanent Magnet Synchronous Motors with Hybrid Deep Model
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsDetecting winding faults for PMSM can be of great importance. The choice of topic is in good alignment with the scope of the journal. The paper is OK in general. However, the layout and presentation, particularly, the demonstration of the methodology and result sections should be reinforced. The major issues and suggestions are:
- There are way too many abbreviations in the title, making reading it a headache, recommending revision. Please use general engineering terms whenever possible in the paper title.
- It is rather unorthodox to have a “2. Related Work” section. Generally, we should keep all the background information for the research in the introduction section.
- Subsections 3.12~3.15 are just plain reiteration of textbook/well-known knowledge, it is recommended to get rid of those. If the author insists to keep, you can make a new table, just like table 1 to store all the information.
But currently, it is unacceptable to have nearly 1/3 of the paper to be known knowledge, you need to focus on delivering new knowledge to the society.
- Regarding the details of the machine learning, this paper currently lacks a demonstration of the raw signals acquired for learning.
You need to plot the vibration/current signals for all different cases. Otherwise, it is untrustworthy in terms of research credibility.
- In the final comparison, recommend to include a figure/plot rather than just using table 7, it is more intuitive to see which method is better.
Author Response
I would like to thank the reviewer very much for the valuable comments about the paper. I take into account the comments, and the corresponding revisions/corrections are highlighted in the resubmitted files. The reviewer's suggestions are colored red in the resubmitted manuscript.
Comment 1) There are way too many abbreviations in the title, making reading it a headache, recommending revision. Please use general engineering terms whenever possible in the paper title.
Response 1) Thank the reviewer for pointing this out. I revised the title to remove abbreviations and use general engineering terms. The revised title is: ‘Inverter-Driven and Stator Winding Fault Detection in Permanent Magnet Synchronous Motors with Hybrid Deep Model”
Comment 2: It is rather unorthodox to have a “2. Related Work” section. Generally, we should keep all the background information for the research in the introduction section.
Response 2: Thank the reviewer for the valuable feedback. I have completely restructured this part of the submission. The "Related Work" section has been combined with the Introduction section. The Introduction includes the background information and the content of the literature review. I have also reviewed and refined all paragraph transitions throughout the Introduction and subsequent sections to ensure a coherent narrative flow.
Comment 3: Subsections 3.12~3.15 are just plain reiteration of textbook/well-known knowledge, it is recommended to get rid of those. If the author insists to keep, you can make a new table, just like table 1 to store all the information.
But currently, it is unacceptable to have nearly 1/3 of the paper to be known knowledge; you need to focus on delivering new knowledge to society.
Response 3: I appreciate the reviewer's insightful recommendation. To reduce redundancy and enhance the paper's focus on novel contributions, I have significantly shortened Subsections 3.1.2–3.1.5 and rearranged the related content into a new summary table (Table 1). The updated section now provides a brief overview of the background information necessary to understand the proposed approach.
Comment 4: Regarding the details of the machine learning, this paper currently lacks a demonstration of the raw signals acquired for learning.
You need to plot the vibration/current signals for all different cases. Otherwise, it is untrustworthy in terms of research credibility.
Response 4: I appreciate the reviewer's valuable comment. I fully agree that presenting the raw signals is essential to ensure credibility and reproducibility. In the revised version, I have included the full-length raw signals of all sensors for three representative cases (normal, short-circuit, and overheating faults) in the first dataset with Figures 3, 4, and 5, as well as three critical cases for the second dataset (healthy, inter-turn short circuit fault with 16.08% severity, and inter-coil short circuit fault with 23.48% severity.) with the Figures 7, 8, and 9 in the main body of the paper. The remaining fault cases are provided in the Appendix for both datasets.
Comment 5: In the final comparison, recommend to include a figure/plot rather than just using table 7, it is more intuitive to see which method is better.
Response 5: I thank the reviewer for the valuable feedback. I benchmark the results against recently published works using the same datasets, as shown in Tables 8 and 13. Furthermore, to provide a clearer and more intuitive illustration of the performance gap, "Figure 14. Comparison of the models for the same inverter-driven dataset," has been added and comprises the test accuracy and training time. This graphical representation immediately highlights the superiority and competitive edge of the proposed approach in achieving highly accurate and robust fault classification. Additionally, "Figure 19. Comparison of the models for the same stator winding faults dataset" has been added to illustrate the position of the proposed method in the literature, especially in terms of accuracy and fault class number.
Reviewer 2 Report
Comments and Suggestions for Authors- Notwithstanding the abstract's provision of a comprehensive introduction to the background information, the description of the proposed method makes only a cursory mention of the utilisation of a hybrid 1DCNN-BiGRU model for PMSM fault detection. However, the description of the method is concise and does not clearly articulate the innovation and advantages of this approach. It is therefore recommended that the abstract be supplemented with a concise introduction to the 1DCNN-BiGRU hybrid model, elucidating the manner in which it amalgamates the spatial feature extraction capabilities of CNN with the temporal modelling capabilities of BiGRU. This will serve to further emphasise the innovation and practical contributions of the method.
- The paper states that training the hybrid model is a time-consuming process (approximately 43 minutes for the inverter-driven dataset). While this can be attributed to the intricacy of the model, further exploration into the optimisation of training time is recommended in order to mitigate its impact on practical applications. Approaches such as model pruning or algorithmic improvements have the potential to enhance the efficiency of training.
- While the model demonstrates robust generalisation capabilities on two datasets, testing results for other fault types or real-world scenarios are limited. It is recommended that further testing be conducted in order to evaluate the model's performance across a range of fault types or under actual industrial conditions. This would allow verification of its stability and effectiveness in diverse environments.
- The document under review contains minor spelling and formatting issues, including the absence of units in certain tables and an inconsistent nomenclature for fault types. It is recommended that a thorough proofreading process be undertaken. Furthermore, the visual appeal of the figures is suboptimal and could be enhanced.
- Suggest citing and analyzing some of the latest references, such as 10.1109/TIE.2024.3485708.
The English could be improved to more clearly express the research.
Author Response
I would like to thank the reviewer very much for the valuable comments about the submission. I take into account the comments, and the corresponding revisions/corrections are highlighted in the resubmitted file. The reviewer's suggestions are colored blue in the resubmitted manuscript.
Comment 1: Notwithstanding the abstract's provision of a comprehensive introduction to the background information, the description of the proposed method makes only a cursory mention of the utilisation of a hybrid 1DCNN-BiGRU model for PMSM fault detection. However, the description of the method is concise and does not clearly articulate the innovation and advantages of this approach. It is therefore recommended that the abstract be supplemented with a concise introduction to the 1DCNN-BiGRU hybrid model, elucidating the manner in which it amalgamates the spatial feature extraction capabilities of CNN with the temporal modelling capabilities of BiGRU. This will serve to further emphasise the innovation and practical contributions of the method.
Response 1: I appreciate the reviewer's valuable feedback and well-founded suggestion to improve the abstract. The advice to clarify the novelty and technical necessity of the hybrid 1DCNN-BiGRU model is very appropriate. Following the recommendation, I have included a phrase in the abstract that provides a clearer explanation of how the 1DCNN and BiGRU components complement each other in the fault detection framework:
Inverter-driven short-circuit, open-circuit, and thermal faults, as well as stator faults, can cause electrical and thermal disturbances that affect PMSMs. Significant harmonic distortions, current and voltage peaks, and transient fluctuations are introduced by these faults. The proposed architecture utilizes handcrafted features, including statistical analysis, Fast Fourier Transform (FFT), and Discrete Wavelet Transform (DWT), extracted from the raw PMSM signals to efficiently capture these faults. 1DCNN effectively extracts local and high-frequency fault-related patterns that encode the effects of peaks and harmonic distortions, while BiGRU of this enriched representation models complex temporal dependencies, including global asymmetries across phase currents and long-term fault evolution trends seen in stator faults and thermal faults.
Comment 2: The paper states that training the hybrid model is a time-consuming process (approximately 43 minutes for the inverter-driven dataset). While this can be attributed to the intricacy of the model, further exploration into the optimization of training time is recommended in order to mitigate its impact on practical applications. Approaches such as model pruning or algorithmic improvements have the potential to enhance the efficiency of training.
Response 2: I thank the reviewer for the relevant comment regarding the training efficiency of the hybrid 1DCNN-BiGRU model, especially its impact on practical deployment. I agree that optimizing the training duration is crucial for an application-oriented system.
In response to this recommendation, I tried two distinct optimization strategies: The initial attemp was structured model pruning. This approach did not yield a satisfactory reduction in training time, resulting in a duration of 2247seconds ( minutes) with the accuracy of concluded that removing or resetting to zero of the individual, independent weights in the model is not a solution to the long duration. So, following the reviewer's suggestion to explore "algorithmic improvements,", I focused on training parameters, which have a more direct impact on epoch duration and convergence speed. The batch size was increased from its original value to he patience value for the Early Stopping mechanism was decreased to epochs. This combined adjustment resulted in a significant and practical reduction in training time while maintaining the model's performance. The revised training time is now 206.6 seconds with a classification accuracy of . This demonstrates that the bottleneck was primarily related to sub-optimal parameter selection, rather than the model architecture itself. So I have updated Table 4 (Training Performance) and the other related tables in the manuscript to reflect this superior training time.
Comment 3: While the model demonstrates robust generalisation capabilities on two datasets, testing results for other fault types or real-world scenarios are limited. It is recommended that further testing be conducted in order to evaluate the model's performance across a range of fault types or under actual industrial conditions. This would allow verification of its stability and effectiveness in diverse environments.
Response 3: I thank the reviewer for emphasizing the importance of testing under diverse conditions. I would like to clarify that the current study already employs a generalization assessment by utilizing two distinct datasets for evaluation. Specifically:
-
Dataset 1 focuses on Inverter-side faults (Power Electronics failures).
-
Dataset 2 focuses on PMSM internal faults (Stator Winding Inter-turn and Inter-coil short circuits for several power values).
This demonstrates the model's ability to generalize across different fault origins (electronics vs. machine core). Furthermore, the application of 5-fold cross-validation two distinct datasets in Sections 5.3.3 and 5.4.3 Cross Validation with 5-Fold ensures that the reported metrics accurately reflect the model's consistent stability across diverse subsets of the data used.
However, I agree that testing the model against actual industrial noise, varying load profiles, and novel fault types (e.g., bearing faults) is a crucial next step. I will explicitly mention this as the primary focus of the Future Work in the last paragraphs of the Conclusion section.
Comment 4: The document under review contains minor spelling and formatting issues, including the absence of units in certain tables and an inconsistent nomenclature for fault types. It is recommended that a thorough proofreading process be undertaken. Furthermore, the visual appeal of the figures is suboptimal and could be enhanced.
Response 4: I sincerely thank the reviewer for pointing out the spelling, formatting, and consistency issues. A comprehensive proofreading and technical editing process was performed throughout the entire manuscript. Specifically, I have corrected the observed spelling errors, ensured uniform nomenclature for all fault types, and checked all tables and text to ensure all units are consistently present where required.
All figures within the manuscript have been completely redrawn and updated. The new figures now feature enhanced resolution, improved color schemes, and clearer labeling to maximize their visual appeal and, more importantly, to ensure a highly intuitive presentation of the experimental results and analytical findings (e.g., SHAP and comparative analysis plots). This enhancement significantly aids the interpretation of the model's performance and contribution.
Comment 5: Suggest citing and analyzing some of the latest references, such as 10.1109/TIE.2024.3485708.
Response 5: I appreciate the reviewer's suggestion to include the latest references. I agree that incorporating recent publications significantly enhances the contextual relevance of the work. I have thoroughly analyzed the publication with DOI: 10.1109/TIE.2024.3485708 and have integrated its key findings into the Introduction, literature part as the 29th reference. Additionally, on the transfer learned and CNN-LSTM architecture, the References 35-39 have been added in the Introduction section as the state-of-the-art models.
The English has been revised again by using the Beta version of Grammarly.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe paper proposes a hybrid deep learning model (1DCNN–BiGRU) for detecting electrical faults in PMSMs. The 1DCNN extracts local patterns from raw sensor data, while the BiGRU captures bidirectional temporal dependencies. Handcrafted features (statistical, FFT, and DWT) are fused with deep features to enhance representation. The proposed hybrid model significantly outperforms baseline models (1DCNN, BiGRU, MLP, RF) across all metrics, achieving near-perfect accuracy. Overall, the work is interesting and well-motivated; however, several aspects require further clarification and analysis.
- The proposed architecture integrates several methods, but the specific motivation for each choice is not clearly explained, nor is their individual impact demonstrated. An ablation study is strongly required to show how each component contributes to the final performance. Which part of the model contributes most to the achieved results? Are all components necessary?
- To strengthen the credibility and practical impact of the study, the following additional analyses are recommended. These will help validate the robustness, generalization, and industrial applicability of the proposed approach.
a) Add k-fold cross-validation or use additional unseen test data to confirm model generalization.
b)Include sensitivity, specificity, or AUC metrics to better address class imbalance, and report training and inference times to assess industrial feasibility.
c)Expand the discussion by comparing the proposed approach with more recent Transformer-based and CNN–LSTM hybrid architectures.
d)Consider applying explainability techniques (e.g., Grad-CAM, SHAP) to identify which features most influence the model’s decisions.
e)Discuss the potential distortion introduced by oversampling temporal data (SMOTE) and analyze its effect on model performance.
f)Explicitly discuss practical constraints, including hardware requirements, data acquisition costs, and scalability of the proposed method.
Finally, ensure consistency and readability in all figures; in some, the font size is noticeably smaller (sometimes unreadable), while in others it appears much larger.
Author Response
I would like to thank the reviewer for the valuable comments on the paper. I take into account the comments, and the corresponding revisions/corrections are highlighted in the resubmitted files. The reviewer's suggestions are colored "green" in the resubmitted manuscript.
Comment 1: The proposed architecture integrates several methods, but the specific motivation for each choice is not clearly explained, nor is their individual impact demonstrated. An ablation study is strongly required to show how each component contributes to the final performance. Which part of the model contributes most to the achieved results? Are all components necessary?
Response 1: I appreciate the reviewer's perceptive comment and fully agree that clarifying the contribution of each component is essential for understanding the proposed system’s effectiveness.
I have conducted a comprehensive ablation study to explicitly evaluate the roles of both handcrafted features and architectural components in the revised manuscript. Specifically, for the first ablation study, I analyzed separately the impact of different feature groups (statistical, wavelet, and FFT) by training the same hybrid 1DCNN–BiGRU model under different input settings: (i) only raw data, (ii) only statistical features, (iii) only wavelet features, (iv) only FFT features, and (v) all features.
I have added subsections "5.3.1. Handcrafted Features Ablation Study" for the first inverter-driven dataset and "5.4.1. Handcrafted Features Ablation Study" for the second stator winding fault dataset. . These sections contain Table 4 and Table 9, which include all these cases, along with the resulting performance metrics (accuracy, precision, recall, F1-score, specificity, and AUC, with both macro and weighted values), as well as comprehensive paragraphs. This study demonstrated how each handcrafted feature category contributes to the model's discriminative power. All domains contribute positively to performance; however, the wavelet-based features showed the strongest standalone impact, reaching 98.70 % / 99.89 % accuracy and superior macro-level metrics. This indicates that time–frequency representations effectively capture discriminative transient patterns. Furthermore, multi-domain feature fusion outperforms any single domain (ACC = 99.44% / 99.98%), confirming that these domains are complementary and jointly enhance robustness.
Secondly, again for two datasets, I have performed architectural ablation studies to quantify the effect of each network module. I trained and compared three network variants: (i) only BiGRU, (ii) only 1DCNN, and (iii) the proposed hybrid BiGRU-1DCNN architecture. I have added "5.3.2. Architectural Ablation Study" for the first inverter-driven dataset and "5.4.2. Architectural Ablation Study" for the second stator winding faults dataset. The results are presented in Tables 5 and 10. The Only 1DCNN achieved results very close to those of the hybrid model (99.07% / 99.50% accuracy), confirming its strong capacity to extract representative temporal–spatial features. The addition of BiGRU slightly improves generalization and stability by modeling sequential dependencies. The hybrid 1D-CNN–BiGRU thus provides an optimal balance between accuracy, robustness, and computational cost, validating the suitability of the selected architecture.
For the last, a comparative baseline study has been done with the MLP and RF to place the proposed model in machine learning methods. The results are presented in Tables 6 and 11. This analysis highlights the complementary nature of convolutional and recurrent layers, showing that their combination yields the best generalization and accuracy.
The results clearly indicate that each component plays a significant role, and the integration of both the handcrafted feature sets and the hybrid architecture substantially enhances performance compared to individual parts.
Comment 2: To strengthen the credibility and practical impact of the study, the following additional analyses are recommended. These will help validate the robustness, generalization, and industrial applicability of the proposed approach.
Comment a: Add k-fold cross-validation or use additional unseen test data to confirm model generalization.
Comment b: I appreciate the reviewer's suggestion to strengthen the model's credibility. I have conducted 5-fold cross-validation on both datasets to validate the robustness and generalization capability of the proposed hybrid 1DCNN-BiGRU model. I have added sub-sections for both datasets, namely "5.3.3. Cross Validation with 5-Fold" for the inverter-driven dataset and "5.4.3. Cross Validation with 5-Fold" for the stator winding fault datasets. The 5-fold cross-validation results are presented in Table 7 and Table 12, showing minimal variance across folds, which validates the stability and practical applicability of the proposed approach for real-world scenarios.
This approach ensures that the findings are not dependent on a specific train-test partition, thereby enhancing the reliability and reproducibility of the results.
Comment b) Include sensitivity, specificity, or AUC metrics to better address class imbalance, and report training and inference times to assess industrial feasibility.
Response b) I have thoroughly addressed the class imbalance challenge by reporting sensitivity, specificity, and AUC metrics in both weighted and macro-averaged forms for all models and both datasets. These metrics are highlighted in green in the tables. These metrics provide a comprehensive evaluation of model performance across all classes. Furthermore, I have included inference time measurements in milliseconds to assess the industrial feasibility of the proposed approach. The results, , demonstrate that the model achieves excellent performance on minority classes while maintaining computational efficiency suitable for real-world deployment.
Comment c) Expand the discussion by comparing the proposed approach with more recent Transformer-based and CNN–LSTM hybrid architectures.
Response c) I appreciate the insightful comment regarding the necessity of comparing the proposed method with recent deep learning architectures, particularly those based on the Transformer and advanced LCNN-STM hybrids.
I have expanded the Introduction section to include a detailed comparative analysis with the transformer-based works of Zsuga et al. ( [35] Early Detection of ITSC Faults in PMSMs Using Transformer Model and Transient Time-Frequency Features), Li et al. ( [36] Permanent magnet synchronous motor inter-turn short circuit diagnosis based on physical-data dual model under oil-drilling environment), and Yu et al. ([37] Time-Frequency Domain Lightweight Dual-Branch MSCFormer for PMSM ITSC Fault Diagnosis) between the 146 - 161 rows. Additionally, a study ( [42] Bacha et al., Advanced Deep Learning Approaches for Fault Detection and Diagnosis in Inverter-Driven PMSM Systems) already included in the current submission, which proposes the integration of transformer-based architectures with physics-informed neural networks and tests the model on the same dataset as my model, has been highlighted in the Introduction section and Table 8.
The CNN-LSTM hybrid structure has been analysed in the Introduction again, after the transformer-based studies. Lale and Yüksek ( [38] Identification and classification of turn short-circuit and demagnetization failures in PMSM using LSTM and GRU methods), Feng et al. ( [39] Inverter Fault Diagnosis for a Three-Phase Permanent-Magnet Synchronous Motor Drive System Based on SDAE-GAN-LSTM), and Cheng et al. (1DCNN-Residual Bidirectional LSTM for Permanent Magnet Synchronous Motor Temperature Prediction Based on Operating Condition Clustering) studies have been examined.
The key inference from this discussion is as follows: While recent Transformer models demonstrate exceptional efficiency and performance, particularly for ITSC faults, my hybrid 1DCNN–BiGRU model, utilizing manually engineered statistical, FFT, and DWT features, provides a superior, generalized diagnostic solution across a broader spectrum of critical PMSM faults (inverter-driven, thermal, and stator faults). The proposed model architecture effectively leverages the complementary strengths of CNN and BiGRU, achieving competitive or superior accuracy on comprehensive, realistic datasets. While CNN is excellent at capturing fault-induced instantaneous current peaks and harmonic distortions (high-frequency patterns), BiGRU is particularly effective in modelling the slow evolution of thermal and stator faults by capturing the evolution of the fault over time or the cyclic asymmetry of the current signal. Even if transformer-based architecture enables fast inference and learning by processing temporal dependencies in parallel, rather than sequentially, Transformers naturally lack the inherent capability of CNNs to hierarchically extract local patterns; thus, they typically necessitate the use of patching or tokenization methods to encode local spatial information.
Furthermore, where many CNN–LSTM hybrids utilize the LSTM unit, my choice of the BiGRU offers a simpler, two-gate recurrent unit that provides performance comparable to BiLSTM while significantly reducing the number of parameters.
Comment d: Consider applying explainability techniques (e.g., Grad-CAM, SHAP) to identify which features most influence the model’s decisions.
Response d: I thank the reviewer for suggesting the use of explainability techniques. In response, I applied SHAP (SHapley Additive exPlanations) to both datasets to identify the most influential features contributing to model predictions. Dedicated subsection titled “5.3.4. Explainability (XAI) Analysis" for the first dataset, and "5.4.4. Explainability (XAI) Analysis" have been added to the revised manuscript. While the global SHAP-based feature importance graphic for the inverter-driven dataset is shown in Figure 13, the same graphic but for the second dataset i shown in Figure 18. The most influential features on the results are discussed in depth in the 5.3.4 and 5.4.4 subsections.
Comment e: Discuss the potential distortion introduced by oversampling temporal data (SMOTE) and analyze its effect on model performance.
Response e: I sincerely thank the reviewer for this insightful comment. I fully acknowledge that oversampling techniques such as SMOTE may potentially distort temporal continuity when applied directly to sequential datasets. To clarify this point, a discussion has been added to the Conclusion section in the revised manuscript.
In my study, each observation corresponds to a statistically independent segment extracted from real-world measurements rather than a continuously sampled time sequence. Therefore, the temporal correlation structure of the raw data is preserved. Furthermore, SMOTE was carefully applied only within the training partitions of each fold during the five-fold cross-validation process, ensuring that no synthetic samples were introduced into the validation or test subsets.
This design eliminates the risk of data leakage and prevents any artificial inflation of performance metrics. The cross-validation framework, combined with fold-wise isolation of the oversampling step, ensures that class balancing improves the model’s learning stability without compromising the temporal integrity or statistical validity of the evaluation. Additionally, the improvements observed in the metrics (in particular Recall, F1-Score, and overall Accuracy) show that synthetic examples do not blur the decision boundary, but rather resolve the imbalance problem, allowing a more reliable discrimination of the minority class (failure). The high performance proves that the synthetic data helps the model to accurately learn the overall feature distribution of the failure states.
While recognising the theoretical distortion concerns, in the study, I have practically eliminated this risk by strategically applying SMOTE to the frequency-domain feature space, significantly improving our model performance on the imbalanced dataset.
Comment f: Explicitly discuss practical constraints, including hardware requirements, data acquisition costs, and scalability of the proposed method.
Response f: I appreciate the reviewer’s comment regarding the discussion of practical constraints. In the revised manuscript, I have added a subsection (5.1) titled "Experimental Environment and Model Configuration" under Section 5. Experimental Verification section. I explicitly address hardware requirements considerations. The hardware on which I performed the study (Intel Core i7-14700HX CPU, 32 GB RAM, and an NVIDIA GeForce RTX 4060 GPU) represents the minimum requirements to train the model efficiently.
The proposed hybrid 1DCNN–BiGRU model is designed to be computationally lightweight. The real-time fault detection performance of the model depends on the inference speed. The inference times on two different datasets demonstrate the model's applicability. As detailed in Tables 4 and 9, the inference time per sample is 1.45 ms for the first dataset and 0.20 ms for the second one.
The millisecond-level speeds achieved in both scenarios are more than sufficient for PMSMs for most industrial real-time monitoring applications. However, the fact that the extraction time varies significantly from dataset to dataset (possibly depending on factors such as sample length or number of attributes) should be considered as a limitation. For low-latency and embedded (Edge Computing) applications, it may be necessary to apply model-lightening techniques to further improve the 1.45 ms/sample time. This discussion of limitations has been added to the Conclusion section.
My study shows the model generalization strength. My model is not restricted to a single motor type. The first dataset involves an inverter-driven PMSM with various faults (short-circuit, open-circuit, overheating), while the second includes data from motors of three different power ratings. This practically demonstrates the model's ability to generalize across different power levels and fundamental fault modes, not just a specific operational regime. Despite this strength, the entire range of PMSM variability encountered in industrial deployment cannot be fully covered (e.g., different manufacturers, bearing types, and extreme load/speed conditions). Therefore, a recalibration or minor fine-tuning cycle would still be required for entirely new and unseen motor types. zthe biggest practical constraint
The biggest practical constraint is the high cost of collecting rare failure data (forcing engines to fail or waiting for a long time). The main rationale for using SMOTE in the feature space is to overcome the lack of high-cost and scarce minority class data and to enable the model to learn even rare faults with high performance.
Regarding scalability, the method exhibits linear computational growth with dataset size and has been tested on two independent datasets to verify robustness. The lightweight design and modular feature extraction pipeline enable real-time operation and straightforward integration into IoT or edge-computing environments.
Finally, ensure consistency and readability in all figures; in some, the font size is noticeably smaller (sometimes unreadable), while in others it appears much larger.
Thank the reviewer for this warning. I have reviewed and rectified inconsistencies in figure readability. A global style configuration was implemented in the Python scripts used for figure generation, ensuring a uniform and appropriately sized font across figures. This addresses the previous issue of noticeably small or unreadable text in some graphical elements
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsI'm OK with the revised version. This edition has improved vastly from the last. The paper can be accepted for publication.
Reviewer 3 Report
Comments and Suggestions for AuthorsAll comments have been addressed.
