This manuscript presents a well-structured and methodologically rigorous approach to scalable failure prediction in offshore wind turbines using SCADA data, autoencoder-based normal behavior modeling, and fleet median filtering. The authors have developed and validated a cloud-based, modular pipeline and propose a post-processing technique to reduce false positives in anomaly detection. While the work is timely and technically sound, several aspects could benefit from further clarification.

1. The filtering method is described as novel, however similar fleet based anomaly filtering strategies have been discussed in prior work (Hendrickx et al. 2020, Li et al. 2020). A clearer articulation of what distinguishes this work is needed.

2. The fleet median filtering method assumes most turbines operate under the same conditions at any given time. This assumption may break down, when turbines are shut down for maintenance. Furthermore, in region I downstream turbines produce less power due to wake losses, hence their generator and gearbox temperatures are lower than those of upstream turbines. The authors should discuss how such conditions might affect the effectiveness of the filtering method.

3. The scalability of the pipeline is asserted and architecturally supported, but not empirically demonstrated in the manuscript. If this is claimed as a major contribution, the authors should have included for example:

- Report runtime performance under different fleet sizes

- Demonstrate linear or sublinear scaling

- Show cost, memory or latency metrics as functions of load

This paper presents an autoencoder-based anomaly detection approach for failure prediction in offshore wind turbines that analyzes temperature signals from SCADA data. The proposed approach also uses a fleet-level median filtering technique to reduce non-relevant anomalies. While the work addresses important challenges in wind turbine condition monitoring, several aspects require clarification and additional validation.

1. The term "physics-informed" used to describe the filtering method could benefit from further clarification. The description of the filtering method in section 3.3 (distance to fleet median, windowing, multidimensional distances) appears to be primarily statistical and temporal, rather than directly incorporating physical models or principles. It would enhance clarity if the authors could explicitly detail how "physics-informed" aspects are integrated into the filtering logic.
2. The paper describes its cloud-based pipeline, highlighting its modularity and scalability for managing anomaly detection across wind farms. However, the contribution of this solution remains unclear as the results section focuses solely on the autoencoder and filtering methods. There are no empirical data or quantitative metrics presented to validate the pipeline's actual performance, scalability, or efficiency.
3. There is not enough detail about the specific failure types examined in this work. The authors mention gearbox and generator failures, but more information is needed about the failure sub-types and their locations for enhanced clarity.
4. While the paper acknowledges that data can differ greatly across the fleet and emphasizes the importance of having a large enough fleet for reliable median calculation, I think more discussion is needed about specific sources of variability that could affect the fleet median approach. The paper assumes that a large fleet size will normalize variations, but factors like seasonal variation, turbine location within the wind farm (wake effects, wind exposure differences), and individual operational patterns might create systematic rather than random variations. It would be helpful to have more analysis of how these location-based and operational differences are distinguished from actual anomalies, especially since some turbines might consistently operate differently due to their position rather than equipment issues.
5. Figures 5-7 show negative reconstruction errors and Figures 12-14 show negative anomaly scores. Since reconstruction errors are typically positive differences between predicted and actual values, it's unclear how to interpret negative values in this anomaly detection context.
6. I would recommend comparing the autoencoder model with other normal behavior modeling approaches, especially since a cloud-based solution has been provided in this work and deployment of autoencoder models could be expensively high. Alternative models like isolation forest, one-class SVM, or statistical approaches might offer better cost-effectiveness and computational efficiency for cloud deployment while achieving similar anomaly detection performance.