(1) The novelty of the predictive modeling component is not sufficiently clear. The manuscript mainly evaluates conventional machine learning models, with the best performance obtained from tree-based methods such as XGBoost and LightGBM. The authors should clarify why more advanced models for multivariate time-series analysis are not considered and better distinguish the methodological novelty from system-level integration.

(2) The role of the LLM and RAG components requires much clearer explanation. It is unclear what knowledge base is used for RAG, how retrieval is performed, and what retrieved information is provided to the LLM. Based on the current description, the LLM appears mainly to serve as a natural-language interface for presenting sensor and prediction results rather than providing a new diagnostic or reasoning capability.

(3) The manuscript lacks a clear unified framework or end-to-end workflow that connects the individual components. Although Figures 4 and 5 show the general system architecture and communication flow, the overall research methodology remains fragmented across SCADA preprocessing, model training, prototype sensing, pseudo-label generation, real-time inference, explainability, and the LLM interface. The manuscript would benefit from a unified methodological figure showing the complete workflow.

(4) The literature review is relatively limited and does not sufficiently establish the research gap. Rather than mainly listing individual studies and their performance, the authors should provide a more systematic overview of SCADA-based prediction, multimodal sensing, IoT-based monitoring, explainable AI, and recent LLM/RAG applications in wind-turbine maintenance. A comparison table could also help distinguish the proposed work from existing studies.

(5) The claim that the newly added gas and sound sensors significantly improve fault prediction is not sufficiently supported by the current experiments. Feature importance and SHAP results indicate that the model uses these features, but they do not demonstrate that adding these sensors improves predictive performance. The authors should conduct an ablation study comparing model performance with and without the newly added sensors to directly quantify their contribution.

(6) The current manuscript reads more like a technical project report than a research paper.