pith. sign in

arxiv: 2605.08187 · v1 · submitted 2026-05-05 · 📡 eess.SP · cs.LG

Towards Interpretable Damage Detection based on Aerodynamic Pressure Measurements

Philip Franz , Max von Danwitz , Gregory Duth\'e , Alexander Popp , Eleni Chatzi This is my paper

Pith reviewed 2026-05-12 01:35 UTC · model grok-4.3

classification 📡 eess.SP cs.LG
keywords aerodynamic pressurestructural damage detectionwind turbine bladesconvolutional neural networkexplainable machine learningnon-intrusive sensingelastic structuresreal-time monitoring
0
0 comments X

The pith

Aerodynamic pressure measurements enable real-time detection and quantification of structural damage in elastic beam-like structures under turbulent flow.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper explores aerodynamic pressure data as a non-intrusive way to monitor damage in increasingly flexible wind turbine blades. Experiments mounted a NACA 633418 airfoil on a vibrating cantilever beam in an open wind tunnel and introduced progressive saw cuts as damage near the support. Pressure distributions were recorded across different inflow conditions and damage states, then fed to a convolutional neural network for detection and severity classification. Physics-based insights and explainable machine learning were added to interpret how damage alters both structural dynamics and the pressure field. A sympathetic reader would care because this offers a potentially economical, transparent alternative to traditional sensors for maintaining large, flexible structures in real-world aerodynamic environments.

Core claim

The results demonstrate that pressure measurements can effectively enable real-time detection and quantification of damage in elastic, beam-like structures subjected to mildly turbulent flow and varying operational conditions. Incorporating physics-based insights and explainable machine learning methods into the pipeline improves transparency, robustness, and physical consistency in data-driven monitoring of elastic, aerodynamically loaded structures.

What carries the argument

Convolutional neural network trained on aerodynamic pressure signals, enhanced by physics-based insights and explainable machine learning to link damage effects to changes in the pressure field and dynamic response.

Load-bearing premise

The small-scale laboratory setup with saw cuts on a cantilever beam accurately represents real structural damage and aerodynamic loading on full-scale wind turbine blades.

What would settle it

A test on a full-scale wind turbine blade or under stronger turbulence where the pressure-based model fails to detect or correctly quantify known damage would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.08187 by Alexander Popp, Eleni Chatzi, Gregory Duth\'e, Max von Danwitz, Philip Franz.

Figure 1
Figure 1. Figure 1: The experimental setup features an airfoil on a cantilever beam with an excitation motor and removable [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of barometers on the airfoil surface. The blue, circular markers indicate working sensors; the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Preprocessing pipeline and CNN architecture for pressure signal classification. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic blade section with a flap-wise degree of freedom [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Exemplary multivariate time series sample (sample 14) [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualizations of the TVB a) and MVB b) baselines for an exemplary sample and the signal of sensor 16. The [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: b) Attribution map for sample 14 (TS 1, damage class 0, split 1) using the APB. c) Sum of attributions per [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Violin plot presenting the distribution of the summed attribution vectors [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: b) Attribution map for sample 14 (split 1, 0 [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: b) Attribution map for sample 14 (split 1, 0 [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: STFT spectra of sensor 35 (trailing edge) for TS 8 with a 50% beam-width crack. Panels a) and b) show [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Example of possible integration paths in two dimensions between the basline [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Overview over attribution maps based on the APB for different inflow conditions for damage class 0 and 0 [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Overview over attribution maps based on the APB for different structural states for TS 1 and 0 [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Overview over attribution maps based on the APB for different inflow conditions for damage class 0 and 8 [PITH_FULL_IMAGE:figures/full_fig_p018_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Overview over attribution maps based on the APB for different inflow conditions for TS 5 and 8 [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Distribution of summed attribution vectors for the APB and 8 [PITH_FULL_IMAGE:figures/full_fig_p019_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Overview over attribution maps based on the TVB for different inflow conditions for damage class 0 and 0 [PITH_FULL_IMAGE:figures/full_fig_p020_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Overview over attribution maps based on the TVB for different inflow conditions for TS 1 and 0 [PITH_FULL_IMAGE:figures/full_fig_p021_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Distribution of summed attribution vectors for the TVB and 0 [PITH_FULL_IMAGE:figures/full_fig_p021_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Overview over attribution maps based on the TVB for different inflow conditions for damage class 0 and 0 [PITH_FULL_IMAGE:figures/full_fig_p022_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Overview over attribution maps based on the TVB for different inflow conditions for TS 5 and 8 [PITH_FULL_IMAGE:figures/full_fig_p023_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Distribution of summed attribution vectors for the TVB and 8 [PITH_FULL_IMAGE:figures/full_fig_p023_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Overview over attribution maps based on the MVB for different inflow conditions for damage class 0 and 0 [PITH_FULL_IMAGE:figures/full_fig_p024_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Overview over attribution maps based on the MVB for different inflow conditions for TS 1 and 0 [PITH_FULL_IMAGE:figures/full_fig_p025_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Distribution of summed attribution vectors for the MVB and 0 [PITH_FULL_IMAGE:figures/full_fig_p025_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Overview over attribution maps based on the MVB for different inflow conditions for damage class 0 and 8 [PITH_FULL_IMAGE:figures/full_fig_p026_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Overview over attribution maps based on the MVB for different inflow conditions for TS 5 and 8 [PITH_FULL_IMAGE:figures/full_fig_p027_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Distribution of summed attribution vectors for the MVB and 8 [PITH_FULL_IMAGE:figures/full_fig_p027_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Preprocessing and structure for the mean vector MLP. [PITH_FULL_IMAGE:figures/full_fig_p028_30.png] view at source ↗
read the original abstract

The increasing flexibility of modern large wind turbine blades necessitates cost-efficient and reliable structural monitoring solutions. For this purpose, we propose to use aerodynamic pressure measurements obtained via Aerosense, a novel, non-intrusive and economical sensing system. In former work [Franz et al., 2025], we investigated the potential of aerodynamic pressure measurements for structural damage detection on elastic and aerodynamically loaded structures. An experimental campaign was conducted on a NACA 633418 airfoil mounted on a vertically vibrating cantilever beam within an open wind tunnel. Structural damage was introduced progressively through controlled saw cuts near the beam support. Aerodynamic pressure distributions were recorded under varying inflow conditions and structural states. Based on this data set, we developed a convolutional neural network to detect structural damage and classify its severity using only aerodynamic pressure signals. The results demonstrate that pressure measurements can effectively enable real-time detection and quantification of damage in elastic, beam-like structures subjected to mildly turbulent flow and varying operational conditions. Recognizing the limitations of pure black-box classification, in this study, we further incorporate physics-based insights and explainable machine learning methods to interpret how structural damage influences both the dynamic response and the aerodynamic pressure field. This leads to an enhanced damage detection pipeline, aiming to improve transparency, robustness, and physical consistency in data-driven monitoring of elastic, aerodynamically loaded structures.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript extends prior experimental work on aerodynamic pressure-based structural health monitoring for flexible wind turbine blades. It describes wind-tunnel tests on a NACA 633418 airfoil mounted on a vertically vibrating cantilever beam, with progressive saw cuts introduced near the support to simulate damage. Pressure distributions are recorded across structural states and varying inflow conditions; a convolutional neural network is trained to detect damage presence and classify severity from pressure signals alone. The current contribution adds explainable AI methods to extract physics-based insights on how damage alters the dynamic response and pressure field, aiming for a more transparent and robust monitoring pipeline.

Significance. If the laboratory results hold, the work demonstrates a non-intrusive, economical sensing approach capable of real-time damage detection and quantification under mildly turbulent flow and changing operational conditions. The addition of XAI methods is a positive step toward interpretability and physical consistency. The experimental design is tightly scoped to the beam-like elastic structure in controlled conditions, avoiding unsupported extrapolation claims; however, the representativeness of saw-cut damage and the specific airfoil-cantilever setup for full-scale blades remains an open question for broader applicability.

major comments (2)
  1. [Abstract] Abstract: The central claim that pressure measurements 'can effectively enable real-time detection and quantification of damage' is asserted without any quantitative performance metrics (accuracy, F1-score, error bars, or cross-validation details) presented in the manuscript. The results are referenced to the prior unpublished work [Franz et al., 2025] without independent verification or new tabulated results here, which weakens the standalone evidential basis for the effectiveness assertion.
  2. [Abstract] The manuscript states that XAI methods are incorporated to interpret damage effects on dynamics and pressure fields, yet no specific XAI techniques (e.g., SHAP, LIME, or attention maps), no example interpretations, and no comparison of interpretability gains versus the prior black-box CNN are described. This makes it difficult to assess whether the enhanced pipeline delivers the claimed improvements in transparency and physical consistency.
minor comments (1)
  1. [Abstract] The abstract refers to 'mildly turbulent flow' and 'varying operational conditions' without defining the turbulence intensity levels or the specific inflow parameter ranges tested; adding these details would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review. We address the two major comments point by point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that pressure measurements 'can effectively enable real-time detection and quantification of damage' is asserted without any quantitative performance metrics (accuracy, F1-score, error bars, or cross-validation details) presented in the manuscript. The results are referenced to the prior unpublished work [Franz et al., 2025] without independent verification or new tabulated results here, which weakens the standalone evidential basis for the effectiveness assertion.

    Authors: We agree that the abstract would benefit from quantitative support to strengthen its standalone character. Although the full performance evaluation (including accuracy, F1-scores, and validation details) was the focus of the referenced prior work, we will revise the abstract and add a concise summary of the key CNN performance metrics in the introduction of the revised manuscript. This will provide the requested evidential basis without duplicating the prior analysis. revision: yes

  2. Referee: [Abstract] The manuscript states that XAI methods are incorporated to interpret damage effects on dynamics and pressure fields, yet no specific XAI techniques (e.g., SHAP, LIME, or attention maps), no example interpretations, and no comparison of interpretability gains versus the prior black-box CNN are described. This makes it difficult to assess whether the enhanced pipeline delivers the claimed improvements in transparency and physical consistency.

    Authors: We thank the referee for highlighting the need for greater specificity. The XAI component is described in the body of the manuscript, but we acknowledge that the abstract does not name the techniques or provide examples. We will revise the abstract to explicitly identify the XAI methods employed and note the interpretability improvements. We will also ensure that example interpretations and comparisons to the black-box baseline are clearly highlighted in the revised text. revision: yes

Circularity Check

0 steps flagged

Minor self-citation to prior work on data collection, but central claims remain independently grounded in new experimental results and XAI analysis

full rationale

The paper grounds its claims in a new experimental campaign using aerodynamic pressure data from a NACA 633418 airfoil on a vibrating cantilever beam with progressive saw cuts, recorded under varying inflow conditions. A CNN is trained on this dataset for damage detection, followed by incorporation of physics-based insights and explainable ML methods for interpretability. The sole self-citation [Franz et al., 2025] refers to the prior data-collection effort and does not bear the load of the current interpretability or quantification results. No derivation reduces by construction to fitted inputs, self-defined quantities, or unverified self-citations; the empirical pipeline is self-contained against the described laboratory setup.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach relies on standard assumptions about experimental damage simulation and ML applicability rather than introducing new free parameters, axioms beyond domain norms, or invented physical entities.

axioms (1)
  • domain assumption Controlled saw cuts near the beam support produce damage states representative of real structural degradation in aerodynamically loaded elastic structures.
    Invoked to generate the training data for damage detection and severity classification.

pith-pipeline@v0.9.0 · 5546 in / 1232 out tokens · 48622 ms · 2026-05-12T01:35:40.893125+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

  1. [1]

    and Duth\'e, G

    Franz, Philip and Abdallah, I. and Duth\'e, G. and Deparday, J. and Jafarabadi, A. and Jian, X. and von Danwitz, M. and Popp, A. and Barber, S. and Chatzi, E. , TITLE =. Wind Energy Science , VOLUME =. 2025 , NUMBER =

    work page 2025

  2. [2]

    , year =

    Bobek, Szymon and Nalepa, Grzegorz J. , year =. TSProto: Fusing deep feature extraction with interpretable glass-box surrogate model for explainable time-series classification , pages =. Information Fusion , doi =

    work page

  3. [3]

    Mathematics , doi =

    Fauvel, Kevin and Lin, Tao and Masson, V. Mathematics , doi =. 2021 , title =

    work page 2021

  4. [4]

    Shedding vortex characteristics analysis of NACA 0012 airfoil at low Reynolds numbers , pages =

    Chang, Jianlong and Zhang, Qingui and He, Liujing and Zhou, Yi , year =. Shedding vortex characteristics analysis of NACA 0012 airfoil at low Reynolds numbers , pages =. Energy Reports , doi =

    work page

  5. [5]

    2025 , editor =

    Jang, Hyeongwon and Kim, Changhun and Yang, Eunho , booktitle =. 2025 , editor =

    work page 2025

  6. [6]

    and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and

    Virtanen, Pauli and Gommers, Ralf and Oliphant, Travis E. and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and. Nature Methods , year =

    work page

  7. [7]

    Application of machine learning algorithms in prognostics and health monitoring of electronic systems: A review , pages =

    Bhat, Darshankumar and Muench, Stefan and Roellig, Mike , year =. Application of machine learning algorithms in prognostics and health monitoring of electronic systems: A review , pages =. e-Prime - Advances in Electrical Engineering, Electronics and Energy , doi =

    work page

  8. [8]

    Deep Learning-based Anomaly Detection in Cyber-physical Systems , pages =

    Luo, Yuan and Xiao, Ya and Cheng, Long and Peng, Guojun and Yao, Danfeng , year =. Deep Learning-based Anomaly Detection in Cyber-physical Systems , pages =. ACM Computing Surveys , doi =

    work page

  9. [9]

    Automation in Construction , doi =

    Cha, Young-Jin and Ali, Rahmat and Lewis, John and B. Automation in Construction , doi =. 2024 , title =

    work page 2024

  10. [10]

    Deep Learning Methods for Vibration-Based Structural Health Monitoring: A Review , pages =

    Wang, Hao and Wang, Baoli and Cui, Caixia , year =. Deep Learning Methods for Vibration-Based Structural Health Monitoring: A Review , pages =. Iranian Journal of Science

    work page

  11. [11]

    Workshop Report on Basic Research Needs for Scientific Machine Learning: Core Technologies for Artificial Intelligence , doi =

    Baker, Nathan and Alexander, Frank and Bremer, Timo and Hagberg, Aric and Kevrekidis, Yannis and Najm, Habib and Parashar, Manish and Patra, Abani and Sethian, James and Wild, Stefan and Willcox, Karen and Lee, Steven , year =. Workshop Report on Basic Research Needs for Scientific Machine Learning: Core Technologies for Artificial Intelligence , doi =

    work page

  12. [12]

    Optimal Strouhal numbers for oscillatory propulsion in density stratified fluids , volume=

    Wang, Jiadong and Kandel, Prabal and Deng, Jian , year=. Optimal Strouhal numbers for oscillatory propulsion in density stratified fluids , volume=. doi:10.1017/jfm.2025.283 , journal=

    work page doi:10.1017/jfm.2025.283 2025

  13. [13]

    Triantafyllou, M. S. and Triantafyllou, G. S. and Yue, D. K. P. , year =. Hydrodynamics of Fishlike Swimming , pages =. Annual Review of Fluid Mechanics , doi =

    work page

  14. [14]

    Combining physics-based and data-driven models: advancing the frontiers of research with scientific machine learning , pages =

    Quarteroni, Alfio and Gervasio, Paola and Regazzoni, Francesco , year =. Combining physics-based and data-driven models: advancing the frontiers of research with scientific machine learning , pages =. Mathematical Models and Methods in Applied Sciences , doi =

    work page

  15. [15]

    2026 , issue_date =

    He, Wenchong and Jiang, Zhe and Xiao, Tingsong and Xu, Zelin and Li, Yukun , title =. 2026 , issue_date =. doi:10.1145/3786319 , journal =

    work page doi:10.1145/3786319 2026

  16. [16]

    and Lu, Lu and Perdikaris, Paris and Wang, Sifan and Yang, Liu , year =

    Karniadakis, George Em and Kevrekidis, Ioannis G. and Lu, Lu and Perdikaris, Paris and Wang, Sifan and Yang, Liu , year =. Physics-informed machine learning , pages =. Nature Reviews Physics , doi =

    work page

  17. [17]

    , year =

    Bazilevs, Yuri and Takizawa, Kenji and Tezduyar, Tayfun E. , year =. Computational fluid-structure interaction: Methods and applications , price =

    work page

  18. [18]

    Random Forests , pages =

    Breiman, Leo , year =. Random Forests , pages =. Machine Learning , doi =

    work page

  19. [19]

    Friedman , journal =

    Jerome H. Friedman , journal =. Greedy Function Approximation: A Gradient Boosting Machine , urldate =

    work page

  20. [20]

    2019 , title =

    Data Mining and Knowledge Discovery , doi =. 2019 , title =

    work page 2019

  21. [21]

    Webb, Germain Forestier, and Mahsa Salehi

    Mohammadi Foumani, Navid and Miller, Lynn and Tan, Chang Wei and Webb, Geoffrey I. and Forestier, Germain and Salehi, Mahsa , title =. 2024 , issue_date =. doi:10.1145/3649448 , journal =

    work page doi:10.1145/3649448 2024

  22. [22]

    [Online]

    Zamanzadeh Darban, Zahra and Webb, Geoffrey I. and Pan, Shirui and Aggarwal, Charu and Salehi, Mahsa , title =. 2024 , issue_date =. doi:10.1145/3691338 , journal =

    work page doi:10.1145/3691338 2024

  23. [23]

    International Journal of Machine Learning and Cybernetics , doi =

    Kong, Xiangjie and Chen, Zhenghao and Liu, Weiyao and Ning, Kaili and Zhang, Lechao and. International Journal of Machine Learning and Cybernetics , doi =. 2025 , title =

    work page 2025

  24. [24]

    2017 International Joint Conference on Neural Networks (IJCNN) , year =

    Wang, Zhiguang and Yan, Weizhong and Oates, Tim , title =. 2017 International Joint Conference on Neural Networks (IJCNN) , year =

    work page 2017

  25. [25]

    2010 , editor =

    Understanding the difficulty of training deep feedforward neural networks, in: Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics , author =. 2010 , editor =

    work page 2010

  26. [26]

    Physics-Enhanced Machine Learning: a position paper for dynamical systems investigations , volume=

    Cicirello, Alice , title =. doi:10.1088/1742-6596/2909/1/012034 , year =

    work page doi:10.1088/1742-6596/2909/1/012034

  27. [27]

    Explainable AI for Time Series Classification: A Review, Taxonomy and Research Directions , year=

    Theissler, Andreas and Spinnato, Francesco and Schlegel, Udo and Guidotti, Riccardo , journal=. Explainable AI for Time Series Classification: A Review, Taxonomy and Research Directions , year=

    work page

  28. [28]

    Proceedings of the 34th International Conference on Machine Learning - Volume 70 , pages =

    Sundararajan, Mukund and Taly, Ankur and Yan, Qiqi , title =. Proceedings of the 34th International Conference on Machine Learning - Volume 70 , pages =. 2017 , publisher =

    work page 2017

  29. [29]

    and Marykovskiy, Y

    Barber, Sarah and Deparday, J. and Marykovskiy, Y. and Chatzi, E. and Abdallah, I. and Duth\'e, G. and Magno, M. and Polonelli, T. and Fischer, R. and M\"uller, H. , TITLE =. Wind Energy Science , VOLUME =. 2022 , NUMBER =

    work page 2022

  30. [30]

    The Fourteenth International Conference on Learning Representations , year=

    Geometric Autoencoder Priors for Bayesian Inversion: Learn First Observe Later , author=. The Fourteenth International Conference on Learning Representations , year=

    work page