We thank the Reviewer for taking the time to provide detailed feedback on our manuscript. Our response is provided below:

Feedback:

Clarity of Presentation: Figures, especially 6–9, are vital for understanding the results but are difficult to interpret due to anonymized axes and scales. While confidentiality is necessary, providing normalized or generalized labels (e.g., "Normalized Time," "Fault Indicator Count") would make the visualizations more accessible without compromising sensitive data. Additionally, the Bayesian Ridge Regression model is only briefly introduced; a more detailed justification for its use compared to other regression techniques would improve the methodological section.

Response:

The data used to validate the method is private, and we are not allowed to share specific details regarding the machine configurations or the monitoring period. Therefore, anonymised axes are used to present the results. However, as highlighted in the feedback, the high-level health indicator in each result figure (specifically the (i)th subfigure) shows the fault indicator and counts the number of condition indicators that are exhibiting a fault trend at a given time.

Various regression techniques were evaluated before selecting Bayesian Ridge Regression. The details regarding the comparison of different regression models are now included in the manuscript, specifically in the "Normal Behavior Models" subsection (line 220).

Feedback:

Technical Gaps: While SCADA data is mentioned as an input, the integration process with vibration data is underexplored. Elaborating on preprocessing steps or synchronization challenges would provide a more comprehensive picture. Moreover, the paper does not report performance metrics such as precision, recall, or false alarm rates for the NBMs or hybrid system. Quantitative validation is crucial to assessing the method’s practical value.

Response:

As we are using real wind farm data, determining the exact time a fault is introduced into the system is challenging, making it difficult to provide a precise quantifiable performance measurement of the method. However, as suggested by the reviewer, we have conducted a performance analysis on data from 10 wind turbines (in the performance analysis subsection on line 310), based on the condition monitoring report and fault cases confirmed by technicians through manual inspection. This study provides a more quantitative assessment of fault prediction, including evaluation metrics such as precision, recall, and F1-score.

The integration of SCADA data and vibration data is described in the "Hybrid Condition Monitoring Fault Detection Method" section (line 130).

Feedback:

Scalability Concerns: The approach requires training individual NBMs for each condition indicator across different operating regimes, which is computationally intensive and may be impractical for large-scale deployment. Although the paper acknowledges this issue, it does not propose concrete steps to address it. Exploring techniques like transfer learning, ensemble models, or feature reduction could mitigate this limitation.

Response:

The current approach offers a detailed analysis of each condition indicator but requires significant computational resources since separate NBMs must be trained for each indicator in every operating regime. To address this challenge, future work will focus on developing an explainable deep learning model that can adapt to all condition indicators simultaneously. This will provide both a high-level assessment of the turbine’s overall health and a detailed analysis of individual fault trends while significantly reducing computational demands.

Feedback:

Limited Discussion of Related Work: The paper could better contextualize its contributions by comparing the proposed method to other state-of-the-art approaches, such as fully data-driven deep learning systems. Highlighting the relative strengths and weaknesses of the hybrid approach would provide a clearer perspective on its novelty and utility.

Response:

The related work section has been expanded by adding more information about SCADA-based condition monitoring in the introduction section (line 50).

Feedback:

Broader Applicability: The reliance on both vibration and SCADA data limits the applicability of the method to turbines equipped with these systems. The paper would benefit from discussing adaptations for turbines with only partial data availability or exploring how the method might generalize to other industrial systems.

Response:

This research focuses on vibration-based condition monitoring and presents a framework for assessing the health of wind turbine drivetrain components using vibration data. While SCADA-based methods exist in the literature, vibration-based approaches are generally more reliable. Additional details on SCADA-based condition monitoring have been incorporated into the introduction section (line 50). Moreover, future work will focus on extracting operational information directly from high-quality vibration data to reduce reliance on SCADA data. This update is included in the discussion section (line 370).

Feedback:

Results Presentation: While the paper discusses case studies qualitatively, it lacks numerical summaries or statistical analysis of the detection performance. Providing such data would strengthen the evidence for the method’s effectiveness. Additionally, the scalability of the method across a fleet of wind turbines needs to be demonstrated more convincingly.

Response:

To quantitatively evaluate the performance of the proposed method, we have added a confusion matrix in the performance analysis subsection (line 310), along with evaluation metrics such as precision, recall, and F1-score, to assess fault detection effectiveness.

We sincerely thank the Reviewer for taking the time to review our research and provide valuable feedback. Below is our response:

Feeback:

Would be interesting to see one of these studies beyond simulation.

Response:

The proposed method has been validated using real wind farm data, with results presented using anonymized axes due to confidentiality constraints. Additionally, to provide the quantitative evaluation, a performance analysis (detailed in the performance analysis subsection on line 310) is conducted on datasets from 10 wind turbines. This analysis includes a confusion matrix, which provides a quantitative assessment of the method.

We thank the Reviewer for providing constructive feedback on our research. Our response is provided below:

Feedback:

In the model validation session, the authors only showcase the prediction result in figure 7-9 without further statistics on the results. It is not straightforward to get a sense of the model performance. Hence, could the authors provide a more quantified measurement for the performance of the trained machine learning model? For example, what is the percentage of correctly predicted faults? What is the ratio of false positive and false negative?



Response:

As suggested by the reviewer, we have conducted a performance analysis on data from 10 wind turbines (in the performance analysis subsection on line 310), based on the condition monitoring report and fault cases confirmed by technicians through manual inspection. This study offers a more quantitative assessment of fault prediction, including false alarms and missed fault detections.