The manuscript "On the Potential of Aerodynamic Pressure Measurements for Structural Damage Detection" investigates whether aerodynamic pressure measurements via an airfoil can be used to detect structural damage in a controlled laboratory setting. Developing effective methods for monitoring structural damage in wind turbine blades is a crucial and open research problem. However, from the reviewer's point of view, the applicability of the study's findings to real-world wind turbines is highly questionable due to the assumptions made in the laboratory experiment and the analysis methodology.

Key concerns:

Material and Damage Representation: The structural damage was introduced in metal rather than in composite materials, which are more representative of real-world turbine blades. Furthermore, a saw cut does not replicate the characteristics of a crack as it would naturally occur.

Experimental Conditions: Both the wind excitation and imbalance excitation were kept constant throughout the experiments, a condition not reflective of the variable nature of real-world wind turbine environments.

Evaluation Methodology: A supervised classification method based on CNNs was employed. For real applications, datasets typically do not include labelled damage states, necessitating unsupervised methods that do not rely on such data.  

Given these points, the manuscript's findings have limited transferability to real-world settings. Furthermore, the impact and scope of the work may be misaligned with the target journal's focus. Specific points for improvement include:

- Line 281 mentions that the parameters of the CNNs are fewer compared to alternative methods. It would be helpful to specify the number of parameters used in the CNNs.

-Line 284 notes the use of the Adam algorithm. Given that AdamW is now the standard, why was the Adam algorithm chosen?

- Figure 8 suggests that labelled data is required. In real-world scenarios, where would labeled data come from? What can be done if no labelled data is available?

-Figure 10 is confusing due to the inconsistent decimal places, e.g. 0.99 and 0.011.

- While line 365 mentions fast classification times, how long does CNN training take? Showing a training loss curve would be beneficial.

- Section 5 would benefit from examining the model's performance when presented with data not included within the training data (e.g. other windspeed, bigger crack, etc.)

- Section 6.2 conducted studies under ambient excitation only, which depends on laboratory conditions and may not be comparable to the controlled excitation in wind and imbalance conditions in Section 5. This could also explain why certain frequencies were not identified for some damage states.

- The mode shapes in Figure 14 vary significantly between model orders. How were these complex mode shapes transformed into real space? The first mode shape seems improperly identified—what caused this?

- Line 482 suggests comparing identified mode shapes with those from the FE model using the MAC for clarity.

 - Line 490 raises doubts regarding the assumption that Mode 4 results from the cut, as it appears in the stabilization diagram in Figure 12 for the healthy state and could be probably identified at higher model orders. The saw cut does not have the same dynamic characteristics as a real crack, which would only become apparent at higher vibration amplitudes.

Due to these concerns, particularly the limitation in the methodology's applicability to real-world wind turbines since the authors just presented a supervised damage detection approach and the fact that the submitted manuscript does not really match the journal's scope, the reviewer recommends rejecting the manuscript in its current form and considering a submission to a more suitable journal.

The paper

“On the Potential of Aerodynamic Pressure Measurements for Structural Damage Detection”

By

Franz et al.

explores a proposed novel method for detecting structural damage in wind turbine blades (or similar elastic, aerodynamically loaded structures) using aerodynamic pressure measurements instead of traditional vibration-based sensing.

 the authors conduct a series of wind tunnel experiments using a heaving airfoil mounted on a cantilever beam, where structural damage is simulated by incrementally sawing the beam near its support. The aerodynamic pressure distribution across the airfoil is captured using a wireless, MEMS-based sensing system. A convolutional neural network (CNN) is then trained to classify damage severity based solely on these pressure time series, achieving high accuracy across various inflow conditions and damage states.

In summary, the proposed method is interesting and could be vey useful to researchers and practitioners alike. However, before being reconsider for full acceptance, the following remarks should all be addressed by the Authors.

1. The paper emphasises a controlled wind tunnel proof-of-concept, but it's important to better discuss how the proposed approach could scale to actual, real-life, real-size, field operational conditions.
2. Related to the first remark, further discussion of challenges or plans to generalize from airfoil-level to full blade implementation would add value to the scientific paper.
3. The paper acknowledges that sensor placement and distribution on full blades have not been explored; this aspect may be investigated a little more, or discussed in further detail with the current information.
4. The CNN is selected for its small parameter count, but its generalisation to unseen data (especially with small training sets) may need stronger justification or comparison with simpler ANN models.
5. The consistent confusion between damage classes 4 and 5 at higher AoA suggests model limitations or overlapping feature distributions. This should be discussed more explicitly and in more detail.
6. The concept of pressure-based damage indicators vs conventional vibration-based ones is very interesting from a conceptual perspective, however, a more quantitative comparison would be useful, even only in a small dedicated subset, to better compare the novel proposed and conventional methodologies. For instance, RMS acceleration (Section 6.1) and eigenfrequency evolution (Section 6.2), which are widely used damage-sensitive features, can be used similarly to train a CNN - or other kinds of ANNs.
7. Since the approach is fully data-driven, more insight into the physical interpretability of the CNN outputs and feature relevance (even only in a qualitative, descriptive manner) would improve trust in the method.
8. Including added mass as a damage class, to simulate damage occurrence without actually damaging the specimens, has been done several times by well-known Authors. However, this Reviewer has never been fully convinced about the reliability of this approach. Apart from similar shifts in natural frequencies (of course, can be manipulated by decreasing the n-th modal stiffness or increasing the corresponding n-th modal mass), the structural implications of adding masses differ significantly from reducing stiffnesses (which itself is only an approximation of actual crack initiation, growth, and coalescence). Hence, further justification to use this approach should be provided, and its limitations should be fully and explicitly discussed.
9. Apart from wind turbine blades, aerodynamic pressure plays a key, pivotal role in long-span suspended bridges – see e.g. https://doi.org/10.1016/j.proeng.2017.09.576, https://doi.org/10.1007/978-3-031-61425-5_46, https://doi.org/10.1061/JSENDH.STENG-12095. These works may be cited in the paper to provide a broader context.
10. Similarly to comment 8, referring to Fig 15, saw cutting a specimen to simulate damage is extremely common and acceptable, but the difference between this unrealistic damage scenario, which is straight cut and thick, and actual real-life cracks, which are irregular in shape and hair-like thin, should be explicitly discussed.
11. The reference https://doi.org/10.5281/ZENODO.8018677 is linked to the powerpoint of a presentation; it would be better to use a peer-reviewed journal or conference article instead.