Towards Interpretable Damage Detection based on Aerodynamic Pressure Measurements

Technical Summary: Towards Interpretable Damage Detection Based on Aerodynamic Pressure Measurements

Problem Statement
Modern large wind turbine blades are increasingly flexible, necessitating cost-efficient and reliable structural health monitoring (SHM) solutions. While deep neural networks (DNNs), such as Convolutional Neural Networks (CNNs), have demonstrated success in pattern recognition for SHM, their "black-box" nature limits their utility in high-stakes decision-making. These models often lack interpretability, making it difficult to understand the physical mechanisms driving their predictions or to ensure they generalize reliably to real-world operating conditions that differ from training distributions. Previous work by the authors [Franz et al., 2025] established that aerodynamic pressure measurements (APM) obtained via the Aerosense system could effectively detect and classify structural damage on elastic, aerodynamically loaded structures. However, the underlying mechanisms by which damage influences the pressure field and the specific features the CNN relies upon remained opaque.

Methodology
This study employs Explainable Artificial Intelligence (XAI) to interpret the CNN-based damage detection pipeline proposed in [Franz et al., 2025]. The core methodology involves:

1. Experimental Data: The analysis utilizes data from an experimental campaign involving a NACA 633418 airfoil mounted on a vibrating cantilever beam in a wind tunnel. Structural damage was simulated via controlled saw cuts (reducing stiffness) and added mass. Aerodynamic pressure distributions were recorded at 37 functional sensor locations across varying angles of attack (0° and 8°), wind speeds, and heaving frequencies.
2. Integrated Gradients (IG): The authors apply Integrated Gradients, a model-agnostic, post-hoc attribution method, to the pre-trained CNN. IG calculates the contribution of each input feature (sensor channel and time step) to the model's classification output by integrating gradients along a path from a baseline to the input sample.
3. Baseline Analysis: To isolate specific signal properties, three distinct baselines were defined:
   • Ambient Pressure Baseline (APB): A constant zero signal (removing all aerodynamic loading and vibration).
   • Temporal Variations Baseline (TVB): Signals with channel offsets removed, preserving only temporal dynamics.
   • Channel-wise Mean Value Baseline (MVB): Signals replaced by their channel-wise mean values, preserving static inter-channel magnitude relationships but removing temporal dynamics.
4. Retraining and Evaluation: To validate insights gained from the attribution maps, the authors retrained models on the baseline data. A CNN was retrained on TVB data, and a Multi-Layer Perceptron (MLP) was trained on MVB data (mean vectors) to assess the information retention of static versus dynamic features.

Key Results

• Feature Attribution: Across all baselines, the CNN consistently prioritized sensors located near the leading edge of the airfoil (specifically channels 14–16).
• Static vs. Dynamic Features:
  • Temporal Dynamics (TVB): When the pre-trained CNN was tested on TVB data (static offsets removed), classification accuracy collapsed to approximately 16.7% (random guessing). This indicates the original model relies heavily on static features.
  • Static Magnitudes (MVB): Conversely, when tested on MVB data (temporal dynamics removed), the pre-trained CNN retained significant accuracy (~50% for 0° AoA and ~62% for 8° AoA).
  • Retrained Models: An MLP trained solely on mean values (MVB) achieved high accuracy (86.5% for 0° AoA and 83.7% for 8° AoA). In contrast, a CNN retrained on temporal variations (TVB) achieved significantly lower accuracy (47.4% and 34.8%, respectively).
• Damage Mechanism: The analysis suggests that the primary source of class-discriminative information is the static, inter-channel magnitude relationship, likely resulting from shifts in the equilibrium position of the structural system due to damage (e.g., downward deflection and twist). While temporal patterns (oscillation amplitudes) contribute to classification, their influence is secondary.
• Vortex Shedding: Spectral analysis confirmed that significant vortex shedding did not occur in the experimental conditions, suggesting that the temporal variations detected by the model are primarily driven by the structural oscillations of the beam rather than aerodynamic instabilities.

Significance and Claims
The paper claims that incorporating physics-based insights and XAI methods enhances the transparency and robustness of data-driven SHM. By demonstrating that the CNN relies primarily on static spatial relationships (inter-channel magnitudes) rather than complex temporal dynamics, the study:

• Validates the Physical Basis: It confirms that the model's predictions are grounded in physical shifts in the structure's equilibrium state caused by damage, rather than spurious correlations.
• Identifies Limitations: It highlights that the current CNN architecture may not be optimally capturing temporal dependencies, and that the "black-box" nature previously obscured the fact that static features were the dominant drivers of performance.
• Guides Future Design: The findings suggest that future monitoring systems could potentially rely on simpler, static feature extraction or that more sophisticated time-series specific XAI variants are needed to better isolate state-dependent characteristics.

The authors conclude that while the CNN approach is effective, its interpretability is crucial for understanding why it works. The reliance on static features implies that the model is detecting the permanent deformation and altered stiffness distribution caused by damage, offering a physically consistent explanation for the detection of structural integrity issues in aerodynamically loaded structures.