The paper “Machine Learning Framework for Scour Detection Across Multiple Offshore Wind Farms” proposes a domain generalisation framework – specifically a domain-adversarial neural network using acceleration spectral content as input – for scour detection across multiple offshore wind farms. To this end, a coupled OpenFAST-SSI model was used to simulate responses of turbines in three hypothetical wind farms sharing the same turbine type and with varying soil and structural characteristics, subject to different environmental conditions. The main novelty of the contribution lies with the extending scour detection into fleets. Apart from minor typos, the paper is generally well-written.

However, I believe some further results are required to strengthen the paper and its assertions.

Firstly, appropriate baselines are missing. Without these, some of the conclusions of the contribution can be called into question. Namely:

• The Domain-Adversarial NN (DANN) is only compared to Causal Disentanglement Domain Generalisation (CDDG), but there is no baseline present without Domain Generalisation. Therefore, we cannot extrapolate what is gained by this approach compared to simply training a classifier. This is especially relevant given the results for Tasks 2-4, where the scour detection appears to be highly training turbine-dependent and generalisation appears to be lacking. In this context, it is important to understand if there is any generalisation gain from adversarial training or if the results depend solely on the training data.
• Also very relevant, is the absence of any non-scour case: without it, we cannot say if the model can accurately distinguish between normal operation and scour, or if it’s just triggering for different turbines because their dynamics are somewhat different from the training turbines’.

Therefore, the conclusion that model training is best performed for the shallowest farms/turbines with lowest stiffness foundation could be called into question (if, i.e., the model is just triggering in different dynamics and not on actual scour). Given this assertion, it could also be considered as an additional baseline into the ‘population-based’ approach to train in a single turbine (e.g. T1.1., the shallowest and with lowest soil stiffness). A further suggestion related to this conclusion would be to fully flesh it out in terms of the what it means for the structures’ eigenfrequency (lower soil density, lower stiffness, lower eigenfrequency, but conversely, shallower water depth, higher eigenfrequency).  

The setup of the experiment could also be strengthen in the introduction: in a landscape where there are few labelled scour data (and therefore an incentive for self-supervised learning) and OWT fleets more often than not consist of different wind turbine types, how would you set up the current experiment. These aspects could also be reflected in the conclusions/future work.

Regarding the modelling in OpenFAST in sections 2.3. and 3., more details could be provided:

• How long is the simulated time period?
• How much time was cut off as an initialisation period?
• Which turbulence model or intensity was used for the wind field?
• Which wave spectrum was used?
• How was SSI modelled? soil-springs? linear or non-linear? PISA or API p-y?
• How was scour included in the model?

Some further clarifications could also be provided:

• Line 254: what is meant by PSD features;
• Line 256: The t-SNE visualization is of the model’s latent space?
• Line 315: what’s the mix between the classification and domain losses?

Finally, some suggestions:

• Present scour depths as a fraction of the tower diameter (D). Naturally for some cases D varies, but the fraction makes results more interpretable.
• CDDG should first appear explained under the methodology and not in section 4.4.
• Add a column to Table 13 with a brief description of each task.
• Tables 3-12 could be in annex to improve readability and text flow.

This paper aims to apply machine learning to detect scour across multiple wind turbines using reinforcement learning and a single accelerometer measurement. I am not able to comment in great detail on the different algorithms, but to me it seems that the work at present lacks sufficient explanation related to physical parameters, and there is insufficient detail regarding the generation of data. In order for such an approach to be believable, the numerical simulations providing the data should be sufficiently realistic: there needs to be enough stochastic variation in the wind fields/small errors in yaw and pitch control/rotor imbalance/noise in the measurements/misalignment of the accelerometer/wakes etc. Otherwise, given the approach, it is not easy to know if other signatures from the generated wind fields or other inputs are being detected rather than the desired output. 

It is not entirely clear to me how the effect of scour has been included in the OpenFAST/SSI model: is the soil interaction simply removed over a certain depth, or are other changes also made? 

The manuscript does not document the natural periods of the scoured and healthy structures - this would be useful to understand what magnitude of changes are being detected. The "high energy" content region ranges from 0 - 8 Hz. It would be useful to see examples of the acceleration spectra for healthy and scoured structures. The upper end of this range seems fairly high, and I am missing discussion of what the important physical processes at these frequencies are. I don't know if the time series in Figure 3 are purely illustrative or if they are representative - there appears to be significant resonant response, based on the very narrow-banded characteristics, which also leads to the question of damping. What types of damping are included and what is exciting the resonance? 

Wind field generation and wave field generation are not discussed.  I am assuming that the authors are using "significant wave height" and "peak period" rather than "wave height" and "wave period", but this should be corrected as well. How many different wind fields are generated? How is the turbulence intensity determined? How are wake effects included? Is there any relationship in the wind field generation for turbines in the same farm? 

The conditions are primarily above rated - which leads to similar 3P for many cases. How does this affect the algorithm's applicability? 

Some of the tables could be shown with color coding similar to Fig 5. 

In Fig 5, the quantity of comparison is not very well defined. It is also somewhat remarkable that - for many cases - increasing scour depth does not lead to better identification after approximately 4 m scour depth. 

The fact that the performance is generally better for the tower base sensor rather than tower top suggests that scour has an important effect on the second mode and that this can be useful for detection. This would be useful to discuss and show.