This paper proposes a hybrid machine learning framework combining feature engineering with classification algorithms to detect operational failures in wind turbines using vibration and environmental data. However, the paper suffers from structural deficiencies and lacks clear motivation. While machine learning approaches for wind turbine fault detection are valuable, this work does not adequately differentiate itself from existing literature or demonstrate sufficient novelty for publication. The experimental design contains several methodological flaws that compromise the reliability of the reported results. I have the following comments on the detailed assessment of these issues:

1. The introduction does not adequately establish a clear research gap or compelling motivation for this work. The authors need to articulate more clearly what distinct advantages their proposed approach offers compared to existing methodologies. 
2. The manuscript contains numerous grammatical errors and inconsistencies that impede readability, for example: 
   • Line 20: Incorrect citation formatting "...those turbines Veers et al. (2023)"
   • Algorithm 1, Step 1: "Receive structural ... from the time-domain responses"
   • Line 128: ... to capture the most information about the damage ...
   • Undefined abbreviations (RHS, LHS in Figure 4 caption, DT algorithm) 
3. The statement on line 35 that "unlike traditional methods that rely on hand-crafted features, machine learning enables ..." requires clarification and justification. The shallow machine learning models implemented in this work are fundamentally dependent on hand-crafted features rather than learned representations. 
4. Table 2 lists numerous correlated features (RMS, variance, standard deviation, energy) that likely exhibit multicollinearity. Including all these features appears arbitrary and may degrade model performance due to redundant information. Additionally, clarification on "spectral features" is needed in this table. The spectral section still computes time-domain features. 
5. Section 2.2 appears to describe routine data preprocessing rather than a methodological contribution. Furthermore, the overall feature engineering pipeline lacks clear explanation of how the final feature vector is constructed. 
6.  The relative change damage index (Equation 1) lacks theoretical foundation. Why normalize by max(Δf) specifically? How does this normalization enhance fault sensitivity? The threshold of 0.6 for feature selection appears arbitrary without statistical justification.
7. The canonical correlation analysis for sensor ranking (Equations 2-3) requires clearer explanation. The mathematical formulation doesn't clearly translate to practical sensor selection criteria.
8. Table 4 reports perfect results for SVM/kNN/RF/DT, yet the text states only SVM achieves perfect performance. This contradiction requires clarification. How do you justify these perfect accuracy claims, and have you conducted statistical validation? Such results typically suggest potential overfitting.
9. Figures 3b-c demonstrate clear environmental dependencies, but the model's ability to distinguish between environment-induced changes and actual failures remains unvalidated.

The paper

“Failure classification of wind turbine operational conditions using hybrid machine learning”

By

Machado et al.

presents a hybrid machine learning framework for classifying wind turbine operational conditions, integrating supervised and unsupervised learning techniques.

More specifically, the authors propose a novel relative change damage index for feature normalisation and apply canonical correlation analysis (CCA) for feature and sensor selection. The framework is implemented within the PyMLDA open-source platform and validated on experimental data from a small-scale Aventa AV-7 wind turbine. Results show excellent classification performance (up to 100% accuracy) in both binary and multiclass scenarios, with SVM outperforming other classifiers.

In summary, the proposed approach is relevant and well-motivated, and the results indicate strong potential for wind turbine monitoring. However, before being reconsidered for full acceptance, the following remarks should all be addressed by the Authors:

1. The title is a bit too generic and should be better detailed and circumscribed to the specific Machine Learning approach finally selected (Support Vector Machine).
2. The study relies on a single small-scale 6.7 kW turbine dataset, partly with simulated faults (e.g., aerodynamic imbalance via roughness tape). The generalisability of the findings to utility-scale turbines in field conditions should be better acknowledged and discussed.
3. Related to the first remark, the robustness of the model under more complex environmental and operational variability (e.g., turbulence, large-scale icing, mixed faults) is not clear. More discussion on scaling the framework to real-world turbines is needed.
4. The claimed methodological novelties (relative change damage index and CCA-based feature/sensor selection) are interesting but not sufficiently differentiated from existing approaches in the literature. A stronger justification and comparative analysis would clarify their true contribution.
5. The reported near-perfect accuracies (often 100%) raise concerns of overfitting or data leakage. The authors should explain in more detail how independence between training and testing sets was ensured, and ideally provide confidence intervals or statistical robustness checks.
6. K-means clustering is used to initialise labelling, but the validation of clusters relies solely on the elbow method. Additional cluster quality indices (e.g., silhouette score, Davies–Bouldin) should be reported to strengthen confidence in the unsupervised stage.
7. The threshold of 0.6 for feature selection appears arbitrary; further justification or sensitivity analysis would be valuable.
8. The pipeline of Figure 1 is quite generic and, as it is now, does not really differentiate itself from any other ML-based Condition Monitoring approach.
9. The finding that SVM performs other classifiers fits well with recent findings in the SHM literature. However, Relevance Vector Machine (RVM) can be a feasible alternative, see e.g. https://doi.org/10.1016/j.oceaneng.2024.117692
10. It would be useful to expand the context of Condition and Structural Health Monitoring for Wind Turbines, mentioning research review works, e.g. https://doi.org/10.3390/s22041627
11. While many figures are detailed, some are dense or partially redundant. Simplifying or condensing visual material, and clarifying captions to emphasise the main insights, would improve readability.
12. The conclusions somewhat overstate the generalisability of the proposed framework. A more balanced discussion of limitations (especially regarding dataset size, simulated faults, and applicability to large-scale turbines) would be appropriate.