The paper entitled “Fault Detection in Wind Turbines Using Health Index Monitoring with Variational Autoencoders” deals with an interesting and timely topic, which is definitely adequate for the scientific objectives of the journal.

The quality of the presentation is in general high. Indeed, the rationale for the proposed methods is clearly explained. The test case is well discussed. The method is rigorously applied and the obtained results are convincing and sound. I have particularly appreciated the discussion of the icing vs. pitch imbalance case.

Despite the overall high quality of the paper, there are minor issues which in my opinion could be addressed in order to further improve the quality of the paper.

• In Table 1, aren’t Crest Factor and Peak Factor the same quantity or am I wrong?
• In Table 1, you employ the x symbol for the signal in the definition of Crest Factor and Clearance Factor, but in the caption of the Table you declare that the signal is indicated with a.
• The features in the frequency domain require more explanation. I guess that their computation passes through a Fourier Transform. How is it computed? On the sub-chunks indicated in Figure 8? Please clarify.
• How have you selected the division of the data sets indicated in Figure 6?
• In the case of pitch drive failure (Section 4.3.1), I appreciate the results, given that this kind of fault is very difficult to detect in advance. Nevertheless, I consider too strong the statement “The framework demonstrates exceptional early warning capability for this progressive fault. As shown in Figure 9, an alarm is triggered at 14:32, providing a 2.5-hour lead time before the actual turbine shutdown at 17:12. This substantial lead time enables proactive maintenance planning and minimizes unexpected downtime.”. An advance time of 2.5 hours is impractical for proactive maintenance planning. In my opinion, it would be more important to emphasize the robust detection of the aerodynamic imbalance because this kind of error, being substantially static and with a time-to-failure potentially infinite in line of principle, can be treated timely by minimizing the performance losses.

Dear authors,

Thank you for writing this paper, which I've found interesting and well written. You will find several comments in the attached pdf, and I report hereafter the main ones.

• The features that you have selected are completely general and have nothing specific of wind turbines. You should acknowledge that there are other studies that focused on correlating damages with wind turbine-specific features, such as: 1P, 3P, modes frequency, damping and mode shapes.
• All the models presented in this paper are trained on 1 wind turbine during 1 measurement champaign. Therefore, they will not perform as well on other copies of the same turbine, or even after a few years.
• The threshold for detecting failures is not only a function of the turbine and features, but also of the machine learning model (see Fig. 10). Therefore, it lacks generality and might require expert tuning.
• Stemming from the previous point, I'm getting the impression that you have conveniently selected the training time and threshold to detect the failures that you knew were there.
• Please specify how you have computed the frequency domain features.
• I agree with the other reviewers that little can be done by detecting problems 2.5 hours in advance. This is not necessarily an issue of your framework, because maybe the failure was simply not there before.

Thanks to the authors for a very clear and well-written paper. I enjoyed reading it. I only have a few minor comments related to the method:

1. It would be useful to clearly explain in the overview the motivation for using a probabilistic model for this application and not a deterministic approach, which is often simpler. It is briefly mentioned in Sec. 3.3, but could appear earlier in the text. You could highlight some other benefits, such as better generalization, being less prone to overfitting, dealing with noise in the data, etc.

In relation to that, could you comment in the result section on what it is about VAE that makes it a robust tool and performs better than the other methods you have compared it to? 

2. Some more details on the feature extraction process would be appreciated, especially for the frequency domain terms.

3. It would be good to differentiate between the vector and scalar notations by using bold for vectors, for example. Introduce the feature, output, and the latent space with mathematically rigorous notations (For example, instead of x, it is more complete to introduce it as \boldsymbol x \in \mathbb R^N). Similarly, \boldsymbol z \in \mathbb R^M, where M<N.

4. Line 146: Reference variational inference.