Posters

Abstract

As wind energy assets age, operations and maintenance (O&M) costs for turbine blades have become a critical factor in the levelized cost of energy. In recent years, aeroacoustic monitoring has emerged as a promising, non-intrusive method for remote inspection. However, most existing approaches only detect the presence of damage, without quantifying its severity, which is a crucial requirement for planning inspections, prioritizing repairs, and minimising unplanned downtime. Operators need effective tools to distinguish between minor surface irregularities and critical structural damage that may require immediate action. This work addresses this industry need by introducing a physics-informed deep learning framework capable of quantifying trailing-edge crack widths from multi-microphone aeroacoustic measurements. Acoustic signals from a microphone array are converted into short-time Fourier transform spectra, which serve as multi-channel inputs to a convolutional neural network (CNN). To improve interpretability and performance, the model incorporates  an adaptive channel attention (ACA) mechanism enhanced by a physics-guided channel pooling (PGCP) strategy. Unlike conventional global average pooling [1], the PGCP strategy leverages domain knowledge to initialize channel weights. It evaluates five aeroacoustic indicators: global average, kurtosis, spectral slope, mean prominence and peak area, to assess the relevance of each microphone channel. This allows the network to automatically emphasize sensors capturing the most informative acoustic signatures of damage while suppressing background noise. By integrating PGCP into the squeeze-and-excitation process of the channel attention mechanism, the model ensures that the channel weights are physically meaningful and robust. The framework was validated using an experimental dataset from wind tunnel tests on downscaled airfoils [2], covering a variety of trailing-edge crack widths, flow velocities, angles of attack, and turbulence intensities. The study aimed to discriminate healthy airfoils from different trailing-edge crack severities, classifying the input data into five categories: no crack, 0.2mm, 0.5mm, 1.0mm and 2.0mm. Results demonstrate that the proposed adaptive channel attention network (ACA-Net) significantly outperforms standard CNN baselines, both with and without conventional channel attention mechanisms, achieving higher classification accuracy and improved generalization (indicated by a lower coefficient of variation) as shown in Table 1. Crucially for real-world deployment, the model exhibits strong robustness under high-turbulence intensity (7%), where traditional signal processing methods often struggle. Elevated turbulence intensity further enhances crack detectability by strengthening and stabilizing broadband trailing edge noise. By providing fast, reliable and non-intrusive estimation of the damage severity, this approach supports condition-based maintenance, enabling operators to optimize inspection schedules and extend blade service life. Future work will extend validation to lab-scale rotating wind turbine experiments to assess real world applicability. Table 1. Validation results for different model implementations. Model Average accuracy Coefficient of variation Backbone CNN 0.795 10.8% SENet [1] 0.816 7.1% ACA-Net 0.896 3.6% References [1] Hu, J., Shen, L., & Sun, G. (2018). Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 7132-7141). [2] Zhang, Y., Avallone, F., & Watson, S. (2022). Wind turbine blade trailing edge crack detection based on airfoil aerodynamic noise: An experimental study. Applied Acoustics, 191, 108668.