GA-LSTM-Based Degradation Prediction for IGBTs in Power Electronic Systems

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Good afternoon! Dear authors,
The relevance of the research topic in the manuscript is beyond doubt. The widespread use of power converter technology necessitates monitoring the condition of IGBTs. IGBTs are one of the key components.
The manuscript is at the intersection of scientific fields.
The review is well written and the problematic issues are well-defined. However, I find the approach of researchers taking "some" data and basing their entire research on it unacceptable. The authors merely outlined the problem in the manuscript. The remainder of the manuscript is devoted to the application of the LSTM-based approach.
What is the article about IGBTs and LSTMs?

What is the process that needs to be studied to predict? But the degradation process of IGBTs is understudied, isn't it?
The entire article is devoted to improving the LSTM algorithm. Isn't there something special about its application to IGBTs?
You could take any process and get the same article.
The author needs to change their approach! After all, the object of the study is IGBTs, right? And LSTM is a research tool. The end result is that the entire manuscript is devoted to LSTMs. The results obtained by improving the LSTMs are of no scientific interest. And the IGBT study itself is completely absent from the manuscript.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript addresses an important and timely topic, namely the prediction of IGBT degradation using a GA-LSTM model. The subject is relevant to predictive maintenance in power electronics. The manuscript is generally well structured, and the experimental validation based on NASA PCoE data is appropriate. However, several issues must be addressed before the manuscript can be considered for publication:

1.  GA-based optimization of LSTM has been studied extensively. The authors should clearly emphasize what is new in this work for IGBT degradation prediction, especially compared with PSO-LSTM, Bayesian optimization, or other recent neural architecture search methods.

2. The link between physical degradation mechanisms (solder fatigue, bond wire delamination) and the selected degradation feature is not fully justified. Please strengthen the discussion and, if possible, include other parameters to improve robustness.

3. The model is trained solely on NASA PCoE accelerated aging data. Discuss limitations for generalization to real industrial conditions and provide more details on the impact of data augmentation.

4.  The reported improvements in RMSE, MAE, and MAPE are relatively modest. Statistical tests across multiple runs should be provided to demonstrate significance.

5. Ensure that CNN-LSTM, ARIMA, SVR, and EMD-LSTM baselines are tuned fairly and report their hyperparameters. Otherwise, comparisons may be misleading.

6.  Figures 4–5 and 7–11 require higher resolution and clearer captions (with axis units, legends, and consistent terminology). Table 7 mixes very different models; clarify runtime and environment assumptions.

7. The manuscript is lengthy and contains repetitions (e.g., in Sections 3–5). Some sentences are grammatically incorrect or incomplete and should be revised for clarity.

8. The conclusion should highlight specific engineering implications (e.g., predictive maintenance schedules, early failure prevention) rather than reiterating the methodology.

9.  Please include more recent studies (2022–2024) on physics-informed machine learning and IGBT prognostics

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have thoroughly revised the manuscript and substantially improved clarity, figure quality, and methodological transparency. The novelty positioning is now well defined, focusing on the IGBT-specific problem formulation and event-centered forecasting strategy rather than algorithmic innovation. The physical interpretation and engineering relevance have also been strengthened. Overall, the paper is much clearer and closer to publication in Energies.

However, several refinements would further enhance the technical depth and completeness.

Major Comments

1. Quantitative Significance of Improvement
   The improvement (ΔRMSE = 0.073; ΔMAPE = 0.726 %) is meaningful but modest. Please report either 95 % confidence intervals or the variance across several random seeds to demonstrate statistical reliability. Even a concise appendix or supplementary figure would be sufficient.

2. Deployment Practicality
   The revised conclusion adds useful predictive-maintenance guidance. It would strengthen the engineering value if the authors briefly mention approximate inference latency (e.g., milliseconds / sample) or computational feasibility on embedded hardware for real-time monitoring.

3. Physical Mechanism → Observable Mapping
   The linkage between structural degradation and electrical indicators has improved notably. Further emphasize this connection using recent thermo-mechanical studies:

   • Xu et al., Micromachines 14 (7) (2023) 1344 (https://doi.org/10.3390/mi14071344) quantitatively analyzed how void damage in the solder layer raises junction temperature, establishing the chain structural defect → thermal rise → electro-thermal stress → waveform/indicator change. This work provides strong physical evidence supporting your use of VCEV_{CE}VCE​ as a degradation-sensitive observable.

   • Tan et al., Micromachines 13 (4) (2022) 554 (https://doi.org/10.3390/mi13040554) used flow–solid coupling to characterize the interaction between the cooling channel and device thermal field. Their results support your discussion on deployment feasibility and the influence of cooling conditions on indicator stability.

   Adding a brief citation or sentence referring to these works in Section 2.3 and the Conclusion will reinforce the physical grounding of your model.

4. Physics-Informed Integration (Future Work)
   The new discussion in Section 5.5 is valuable. Consider including a small schematic showing how electro-thermal consistency terms could be embedded in the loss function for future physics-informed extensions—this would make the idea more accessible to general readers.

5. Figures 10–11 and Table 8
   Figures are clearer but could use stronger color contrast for grayscale printing. Ensure Table 8 explicitly states that all models share identical preprocessing and data splits.

6. Language Polishing
   The English has improved significantly. A final proofreading is advised to fix residual typos (e.g., “cfoure,” “forecaster”) and unify terminology (use “turn-off overvoltage” consistently).

Minor Comments

• In the Abstract (lines 26–35), specify that improvements are achieved “under identical preprocessing and validation-based protocol.”

• Remove template placeholder text in Reference [29].

• When discussing electro-thermal consistency in Section 2.3, directly cite Xu 2023 to substantiate the mechanism pathway.

• Ensure all figure axes include physical units (V, °C, cycles).

• Provide the exact NASA PCoE dataset version or release date for traceability.

Author Response

Please see the attachment.

Author Response File: Author Response.docx