Lightning Damage Detection Method Using Autoencoder: A Case Study on Wind Turbines with Different Blade Damage Patterns

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Author.

The study addresses a relevant issue in the wind energy sector, and your proposal to use autoencoders for remote damage detection represents an interesting contribution with practical potential.

However, I have identified several areas that could be strengthened to improve clarity, methodological rigor, and overall impact of the study. I recommend expanding the review of related work, providing more detail on the model architecture and evaluation process, and better contextualizing the actual scope of the study, given that it was applied to only two wind turbines.

I have attached a PDF file with specific comments. I hope these comments prove helpful in further developing your manuscript. I commend your efforts and encourage you to continue exploring this line of research.

Kind regards.

Comments for author File: Comments.pdf

Author Response

Response to Reviewer 1 Comments

 

I would like to express my gratitude for the constructive comments that have been provided.

The following responses to the aforementioned comments are provided below:

 

Point 1: [Title]Although the title states that the method is applied to "multiple wind turbines," the study only considers two turbines. This may be misleading or an overstatement of the actual scope of the work.

Additionally, while "lightning damage" is mentioned, it would be more informative to specify that the study focuses on blade damage, which provides greater clarity regarding the type of anomaly being addressed.

 

Response 1: As previously stated, this is indeed correct. The title aims to emphasise that these accident cases possess distinct characteristics. Therefore, the title has been revised to elucidate the points more clearly:

Lightning Damage Detection Method Using Autoencoder: A Case Study on Wind Turbines with Different Blade Damage Patterns

 

 

Point 2: [Abstract]The abstract does not mention any performance metrics for the model, which is a critical component of a scientific summary.

Although the use of an autoencoder and SCADA data is referenced, the specific variables used and the evaluation procedure are not described, limiting the reader’s understanding of the applied methodology.

 

Response 2: The abstract has been augmented with an explanation of the features employed, the methodology, and the assessment method.

 

 

Point 3: [Introduction]The impact of the study could be broadened by maintaining a more global perspective, rather than focusing solely on the Japanese context. Over-localization may limit the applicability of the proposed solution if it is not adequately framed in an international setting.

The justification of the study should not rely exclusively on regulatory mandates, such as those issued by the Ministry of Economy, Trade and Industry. Instead, the rationale should be grounded in technological needs and clear research gaps.

 

Response 3: We have also expanded the references to include overseas cases. In addition, we have rewritten it so that it is simply an example of Japan.

The regulatory mandates were presented as merely an example of lightning protection measures in Japa. We have rewritten the text to make this intention clearer.

 

 

Point 4: [Introduction]While a solution is suggested, the paper does not explicitly formulate a working hypothesis or pose a clear research question—both of which are essential to guide and structure the investigation.

 

Response 4: The paper was revised to clarify the research question, which emphasised the novelty of the study.

Point 5: [Introduction]The paper does not specify the type of autoencoder used, nor does it describe the data acquisition and processing architecture, which weakens the technical transparency of the proposed approach.

 

Response 5: We redefined it the Basic Autoencoder to indicate that this is the simplest type of autoencoder. The data used this paper was simply provided by the wind turbine operator in the form of a dataset. Consequently, there is no established framework for its acquisition. we have added an explanation to the paper to make this clear.

 

 

Point 6: [Introduction] The introduction lacks a comprehensive review of the state of the art, and it does not clearly justify how the proposed method improves upon existing alternatives.

 

Response 6: An explanation of the results of previous research was added, and the novelty of this paper was emphasised.

 

 

Point 7: [The Accident Overview and SCADA Data] The descriptions of the accidents are primarily narrative in nature but lack technical depth. It would be valuable to specify, for instance, which sensors exhibited significant changes or how the damage was reflected in the SCADA data, to establish a clearer connection between the physical event and the digital evidence.

 

Response 7: A more specific description of the circumstances of the accident has been added. Additionally, we have included an explanation of the data that will be affected by the circumstances of the accident.

 

 

Point 8: [The Accident Overview and SCADA Data] The study is based on only two wind turbines, which constitute a very limited sample size. This significantly restricts the external validity of the proposed model and limits its potential to be generalized to other contexts or turbine types.

 

Response 8: You are correct. In order to enhance the reliability of the model, it is necessary to apply it to a greater number of wind turbine accident cases. However, the SCADA data at the moment a lightning accident occurs is extremely rare and currently only the two cases discussed in this paper can be analysed. As you correctly identified, this is an important question in this research, and we have added an explanation to the conclusion about the need to apply this to other wind turbines.

 

 

Point 9: [Anomaly Detection Model] Only two variables—wind speed and blade rotational speed—were used. For a system as complex as a wind turbine, this selection is limited and may restrict the model’s ability to detect subtle or multifactorial anomalies

 

Response 9: While a simple model is preferable, but we can see your point. We are currently researching methods to utilise other variables.

 

 

Point 10: [Anomaly Detection Model] The network construction methodology is not clearly explained. It is mentioned that the input and output layers have 12 nodes, which is confusing given that only two variables were considered. This raises important questions: Were additional temporal features extracted using the sliding window? How exactly were the input vectors constructed?

 

Response 10: Please find attached a diagram that illustrates sliding windows and clarifies the reason for the input being 12-dimensional.

 

 

Point 11: [Anomaly Detection Model] The complete autoencoder architecture is not described, including the number of hidden layers, activation functions, optimizer used, or any regularization techniques. This omission significantly

limits the study's reproducibility.

 

Response 11: A table has been created to summarise the hyperparameters of the autoencoder.

 

 

Point 12: [Anomaly Detection Model] Potential issues related to overfitting are not discussed, which is especially important given the small dataset size.

 

Response 12: In this study, 10% of the training data is used as validation data to check for overfitting to the training data. We have included a learning curve plot in the results to demonstrate that there has been no overfitting.

 

 

Point 13: [Anomaly Detection Model] While the threshold for anomaly detection is said to be derived from the training data, the method for determining this threshold is not explained (e.g., percentile-based, standard deviation, crossvalidation). This is a critical component in anomaly detection and deserves more thorough treatment.

 

Response 13: We have added to the explanation to make it clear that the threshold is determined using percentile-based methods.

 

 

Point 14: [Anomaly Detection Model] The paper does not mention whether the dataset was partitioned for validation or if any hyperparameter tuning was performed, which is essential in machine learning workflows.

 

Response 14: Added the following sentence: Note that 10% of the normal data is utilized for the purpose of validation data.

 

 

Point 15: [Anomaly Detection Model] The implementation environment is not specified: there is no mention of the programming language or platform used (e.g., Python, MATLAB, TensorFlow), nor whether specialized hardware (e.g., GPU) or standard CPU computation was employed.

 

Response 15: Added the following sentence: These processes were defined using the Python programming language and the Pytorch framework, with the training of the models being carried out using graphics pro-cessing units (GPUs).

 

 

Point 16: [Anomaly Detection Model] There is also no mention of the tools used for data analysis, visualization, or evaluation, making it unclear how key elements such as the ROC curve or confusion matrix were generated.

 

Response 16: Added the following sentence: The visualization of these assessment methods was performed by the utilization of the Python programming language and the Matplotlib plotting library.

 

 

Point 17: [Assessment Result] In the case of wind turbine B, the intermediate data between lightning strikes were excluded due to uncertainty regarding which strike caused the damage. While this decision is methodologically understandable, it results in the loss of potentially valuable information that could over insights into the pre-damage behavior. Moreover, no alternative techniques are roposed to address this uncertainty, such as weak labeling, unsupervised analysis, or probabilistic approaches that could leverage ambiguous data more effectively.

 

Response 17: Although the intermediate data was not used for assessment, anomaly detection was performed. Consequently, it became evident that the anomaly score began to increase around 15,000 min, thereby providing the insight that it was highly probable that the accident occurred at this time.

We would like to explore alternative techniques to address this uncertainty, such as those you have suggested, in the future. we would like to express my gratitude for the valuable input you have provided.

 

 

Point 18: [Conclusions] The conclusions section mainly reiterates what has already been stated in the results, without providing a deeper analysis or synthesis. For instance, it does not highlight the key insights gained from the model's behavior, nor does it explore any unexpected findings that could open new research directions. Moreover, the section fails to mention the study's structural limitations, such as the fact that it relied on only two wind turbines, which significantly constrains the generalizability of the model. The limited dataset size is also not addressed—an important factor that should have been acknowledged to strengthen the scientific rigor and transparency of the work.

 

Response 18: We have expanded the discussion to include the applicability of this technology to other wind turbines, and we invite you to review it.

 

 

Thank you for your cooperation.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents a topic of interes: ligthninf damage detection on wind turbines rotor blades.

The abstract contains statements of a much too general nature, for exemple: ”... utilization of wind power generation has recently garnered significant attention as a cost-effective renewable energy ...” or ”This development is currently underway across Japan.” Please by more precisly, try to avoid such well-known statements.

The authors refer, several times, in a generic way, to SCADA data. Please be precisly and indicate which parameters are subject to SCADA monitoring and highlight there connection within lightning damage! Also, some remarks present, again, an very general character, i.e. ”SCADA system is a monitoring control system that remotely monitors basic information about wind turbines, ...”.  

The paper needs to describe and provide much better input to the readers, so that the improvements done should be clearly traceable.

Author Response

I would like to express my gratitude for the constructive comments that have been provided.

The following responses to the aforementioned comments are provided below:

 

Comment: The abstract contains statements of a much too general nature, for exemple:”... utilization of wind power generation has recently garnered significant attention as a cost-effective renewable energy ...” or” This development is currently underway across Japan.” Please by more precisly, try to avoid such well-known statements.

The authors refer, several times, in a generic way, to SCADA data. Please be precisly and indicate which parameters are subject to SCADA monitoring and highlight there connection within lightning damage! Also, some remarks present, again, an very general character, i.e. ”SCADA system is a monitoring control system that remotely monitors basic information about wind turbines, ...”.

 

Response: The general explanation has been removed and a bibliography added instead.

However, we have taken the precaution of including some general explanations in the introduction to ensure that readers without any prior knowledge can receive the information accurately.

Thank you for your cooperation.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Lightning strike accidents are common disasters in plain-area wind farms. This study proposes an approach: a SCADA data analysis method using an autoencoder to assess wind turbine integrity without visual inspection. Case studies suggest that wind turbine availability could be improved by employing this remote automatic verification method after lightning detection. The paper demonstrates sufficient academic value and engineering significance, aligning with the scope of Wind journal. It is recommended for acceptance after minor revisions. The authors could improve manuscript quality in the following aspects:

1. The abstract dedicates excessive space to research background introduction. The emphasis should be on presenting the methodology, research content, conclusions, and work value.
2. In the introduction section, the summary of existing achievements in lightning damage detection is insufficient. The authors should provide a more comprehensive review and evaluation of existing work.
3. The introduction section should more explicitly summarize the limitations of existing research, the objectives of this study, and its potential contributions or value.
4. Section 2.2 presents two cases of wind turbine lightning damage. Are there more reported cases of lightning disasters? What are the specific failure modes and their respective probabilities of occurrence?
5. Table 1 presents the training and evaluation datasets for deep learning. Please provide detailed descriptions of data contents and clarify the deep learning parameters, input data, and output data in Section 3.
6. In Section 4, the authors identify operational anomalies through rotor speed for two wind turbines. What degree of damage would cause rotational speed abnormalities?
7. In the conclusion section, can this model be further applied to identify specific damage types, such as different levels of blade damage: blade fracture versus surface skin damage?

Author Response

Response to Reviewer 3 Comments

 

I would like to express my gratitude for the constructive comments that have been provided.

The following responses to the aforementioned comments are provided below:

 

Point 1: The abstract dedicates excessive space to research background introduction. The emphasis should be on presenting the methodology, research content, conclusions, and work value.

 

Response 1: The background information regarding wind turbines has been removed. In its place, an explanation of feature, anomaly detection methods and assessment methods has been added.

 

 

Point 2: In the introduction section, the summary of existing achievements in lightning damage detection is insufficient. The authors should provide a more comprehensive review and evaluation of existing work.

 

Response 2: An explanation of the results of previous research was added, and the novelty of this paper was emphasised.

 

 

Point 3: The introduction section should more explicitly summarize the limitations of existing research, the objectives of this study, and its potential contributions or value.

 

Response 3: The paper was revised to clarify the research question, which emphasised the novelty of the study.

 

Point 4: Section 2.2 presents two cases of wind turbine lightning damage. Are there more reported cases of lightning disasters? What are the specific failure modes and their respective probabilities of occurrence?

 

Response 4: For further examples see references [4]. Additionally, Added the following sentence: According to reference[6], wind turbines in Japan are subject to shutdowns for an average of more than 50 days during the winter months due to lightning strikes.

 

Point 5: Table 1 presents the training and evaluation datasets for deep learning. Please provide detailed descriptions of data contents and clarify the deep learning parameters, input data, and output data in Section 3.

 

Response 5: A table has been created to summarise the hyperparameters of the autoencoder.

 

 

Point 6: In Section 4, the authors identify operational anomalies through rotor speed for two wind turbines. What degree of damage would cause rotational speed abnormalities?

 

Response 6: In order for this method to detect blade damage, it is believed that the damage must cause a change in the lift coefficient of the blade. The reasons for this are discussed in the conclusion, so please take a look.

 

Point 7: In the conclusion section, can this model be further applied to identify specific damage types, such as different levels of blade damage: blade fracture versus surface skin damage?

 

Response 7: In order to classify damage types, it is necessary to have a sufficient number of examples for each type. This is considered unrealistic. For the same reason, the level of damage cannot be determined. However, by adjusting the threshold, operators can specify the level of damage they wish to detect.

 

Thank you for your cooperation.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Author,

Thank you for your detailed responses to the comments provided. You have implemented relevant corrections, particularly about clarifying the title, improving the abstract, and expanding the technical discussion of the model’s implementation.

However, while you have acknowledged limitations such as sample size and the scope of the selected variables, some responses could be further strengthened by elaborating on how these issues could be addressed in future research, beyond simply expressing interest or intent.

Additionally, a typographical error was found in the sentence: "Note that 0% of the normal data is utilized for the purpose of validation data." Please correct this accordingly.

Overall, the revisions represent a substantial improvement, which certainly contributes to making the article significantly more robust.

Kind regards.

Author Response

Response to Reviewer 1 Comments

 

I would like to express my gratitude for the constructive comments that have been provided.

The following responses to the aforementioned comments are provided below:

 

Point 1: while you have acknowledged limitations such as sample size and the scope of the selected variables, some responses could be further strengthened by elaborating on how these issues could be addressed in future research, beyond simply expressing interest or intent.

 

Response 1: Please be advised that a new paragraph has been added to the conclusion to provide specific information on future research.

 

Point 2: Additionally, a typographical error was found in the sentence: "Note that 0% of the normal data is utilized for the purpose of validation data. "Please correct this accordingly.

 

Response 2: Thank you for highlighting that. I have rectified the issue.

 

 

Thank you for your cooperation.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The response obtained from the authors adresses just some of the recomandations made to improve the papers quality. 

In the section ”Recommendations for Authors” the boxes were checked indicating that the aspects related to research design, adequately description of methods, clearly presentation of results and conclusions supported by the results need also improvement! Those aspects were not been addressed or the changes/improvements made do not result.

Author Response

Response to Reviewer 2 Comments

 

I would like to express my gratitude for the constructive comments that have been provided.

The following responses to the aforementioned comments are provided below:

 

Comment: In the section ”Recommendations for Authors” the boxes were checked indicating that the aspects related to research design, adequately description of methods, clearly presentation of results and conclusions supported by the results need also improvement! Those aspects were not been addressed or the changes/improvements made do not result.

 

Response: Thank you for highlighting this matter. I will explain the changes I made, focusing on the items in the "Recommendations for Authors" section.

 

Point 1: Does the introduction provide sufficient background and include all relevant references?

Response: In the background section, we have added reference on overseas cases of lightning damage to wind turbines. We also added reference on current lightning damage countermeasures to emphasise the novelty of this research.

 

Point 2: Is the research design appropriate?

Response: The paper was revised to clarify the research question, which emphasised the novelty of the study. Specifically, we explained the background of the research question and described the verification method.

 

Point 3: Are the methods adequately described?

Response: We have included specific information about the methodology (types of AE, list of various parameters, learning methods, etc.). We have also included comprehensive instructions on pretreatment to enhance the reproducibility of the research.

 

Point 4: Are the results clearly presented?

Response: Please be informed that a graph showing the model loss has been added to demonstrate that overfitting is not occurring. Additionally, the scope of applicability of the model derived from the results is outlined in the conclusion. Please refer to this document for further details.

 

Point 5: Are the conclusions supported by the results?

Response: As previously stated, the conclusion provides insights drawn from the results. Please also note that a paragraph regarding future prospects has been added.

 

 

Thank you for your cooperation.

Author Response File: Author Response.docx

Round 3

Reviewer 2 Report

Comments and Suggestions for Authors

The paper can be accepted for publishing.