Towards Interpretable Damage Detection based on Aerodynamic Pressure Measurements
Philip Franz, Max von Danwitz, Gregory Duthé, Alexander Popp, Eleni Chatzi · 2026
Read Towards Interpretable Damage Detection based on Aerodynamic Pressure Measurements by Philip Franz, Max von Danwitz, Gregory Duthé, Alexander Popp, Eleni Chatzi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-08187.txt.
Introduction
Philip Franz, Max von Danwitz, Gregory Duthé, Alexander Popp, and coauthors detect damage in wind turbine blades from air pressure alone. Modern blades are long and flexible, and cheap, non-intrusive monitoring is in demand. Their sensor system, Aerosense, records aerodynamic pressure on the surface without touching the structure.
The underlying experiment mounted an airfoil on a vibrating beam in an open wind tunnel and cut the support in stages to mimic growing damage. A convolutional network trained only on the pressure records learned to detect the damage and grade its severity. The approach offers an interpretable route to blade health monitoring under changing inflow conditions.
Catalog record
| CREATOR | Philip Franz, Max von Danwitz, Gregory Duthé, Alexander Popp, Eleni Chatzi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.08187
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Towards Interpretable Damage Detection based on Aerodynamic Pressure Measurements Philip Franz, Max von Danwitz, Gregory Duthé, Alexander Popp, Eleni Chatzi 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 ( 9 ) , 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. Keywords integrated gradients \cdot aerodynamic pressure distribution \cdot CNN \cdot damage detection \cdot structural health monitoring \cdot explainable AI 1 Introduction Deep neural networks, such as convolutional or transformer architectures, have achieved remarkable success in pattern recognition tasks for multivariate time series, like classification, prediction or anomaly detection ( 11 ; 17 ; 14 ; 25 ) . This capability makes them particularly attractive for domains that rely on on extracting information from dense, multi-sensor measurements, such as structural health monitoring (SHM) ( 23 ; 5 ) . However, relying on opaque black-box models for high-stakes decision in SHM is risky, as deep neural networks typically lack interpretability. Furthermore, it remains unclear whether these models generalize reliably to real-world environments where operating conditions differ from the training distribution. Compounding this, inferring states from sparse and noisy measurements can be ill-posed, admitting a distribution of plausible solutions rather than a unique answer ( 21 ) . To address the lack of generalization and interpretability, the field of scientific machine learning (SciML) ( 1 ; 18 ) or physics-informed machine learning has emerged ( 13 ) . SciML integrates physical laws or domain-specific knowledge directly into the learning scheme, aiming to ensure adherence to the underlying physics, generalization beyond the training domain and to enhance the interpretability of the learned representations ( 13 ; 7 ) . However, SciML requires robust approaches and detailed physical insights and governing equations, which may not always be available for real-world systems. In such cases, explainable artificial intelligence (XAI) offers an alternative for increasing the transparency of black-box models. In our recent study ( 9 ) , we suggested a convolutional neural network (CNN)-based approach to detect and rank structural damage based on aerodynamic pressure measurements (APM) over the section of an airfoil, acquired with the Aerosense system ( 2 ) . Despite its accuracy, the CNN’s black-box nature limits its practical utility and prevents a thorough understanding of the underlying mechanism. Therefore, we employ an XAI method to enhance the intepretability of our appraoch and to identify damage-sensitive features within the APM. XAI methods can generally be categorized as ante-hoc or post-hoc, global or local, and model-specific or model-agnostic ( 20 ) . For this study, we select Integrated Gradients (IG) ( 19 ) , a model-agnostic, post-hoc attribution method ( 20 ) , which is suited for multivariate time series and provides attributions across all channels. While IG is a local method providing only sample-specific explanations, subsequence or instance-based approaches as XCM ( 8 ) , TSProto ( 3 ) , or ProtoTSNet ( 15 ) , can offer global, dataset-wide explanations via class-specific subsequences or prototypes. However, these methods require training new models, whereas our objective is to interpret the existing CNNs presented in ( 9 ) . The paper is organized as follows: Section 2 details the experimental setup, the previously trained CNNs and effects of damage on the APD. Section 3 presents the IG method and baselines, analyzing the resulting attribution maps. Section 4 describes the evaluation and retraining of the CNNs based on insights from the attribution analysis, while Section 5 provides a physics-grounded discussion of the results. We conclude with a summary of the main findings in Section 6 . 2 Experimental Setup & Effects of Damage 2.1 Experimental Setup The experimental setup (Figure 1 ) consists of a NACA 633418 airfoil mounted on a flexible aluminum cantilever beam within the open test section of the EPFL Unsteady Flow Diagnostics Laboratory wind tunnel. The wind tunnel test section measures 40 cm \times 40 cm and operates up to 35 m/s, while the beam itself features a rectangular cross-section (1 cm × 4 cm). Artificial damage is introduced by transversally sawing the beam in z-direction near the support to reduce local stiffness; the severity is quantified by the cut length across five levels (0%, 12.5%, 25%, 37.5%, and 50% of the beam width). An additional damage scenario is considered by attaching a mass to the undamaged beam between the airfoil and the motor (see Figure 1 ). Periodic excitation is applied at the beam’s tip via a motor-driven eccentric mass, primarily inducing flap-wise bending, although torsional motion also occurs. The Aerosense system ( 2 ) records the aerodynamic pressure distribution at 40 locations on the airfoil surface (only 37 were functioning correctly, see Figure 2 ) at the airfoil’s mid-span with a sampling rate of 100 Hz. Ultimately, this cantilever setup serves as a proxy for a rotating WTB oscillating flap-wise due to gravitational loading. Figure 1: The experimental setup features an airfoil on a cantilever beam with an excitation motor and removable mass, placed in front of the outlet of the wind tunnel. The Aerosense sensor node records aerodynamic pressure on the airfoil, and accelerometers measure beam vibration. Beam dimensions are indicated in the upper right. Figure taken from ( 9 ) . Table 1 summarizes the eight test series (TS) generated by combining the two AoAs, wind speeds V , and heaving frequencies f_{h} . Each TS comprises six structural states resp. damage classes: five crack lengths to simulate stiffness reduction and one added mass state. We conduct three 150 s runs per configuration: the first 15 s feature aerodynamic loading only, while the remaining 135 s involve simultaneous loading by aerodynamic and periodic forces. More details on the experimental setup, the data acquisition system and data normalization, can be found in Section 2 of ( 9 ) . Table 1: Boundary conditions of the eight test series. Test series 1 2 3 4 5 6 7 8 AoA [ ∘ ] 0 0 0 0 8 8 8 8 approximate f_{h} [Hz] 1.0 1.0 1.9 1.9 1.0 1.0 1.9 1.9 v [m/s] 12 24 12 24 12