Presentations

Abstract

Blade damage inspection without stopping the normal operation of wind turbines has significant economic value for wind farm operation. This study proposes a physics-informed AI approach that can accurately, robustly, and in real-time detect blade surface and structural damages from optical and thermal videos taken by drones in the field without stopping the normal operation of turbines. This approach first fuses optical and thermal videos taken by drones from normal operating wind turbines and achieves near real-time blade damage segmentation by utilizing multimodal and temporal complementarity. Then, it employs a Swin-Transformer-based denoising autoencoder model to differentiate between normal and anomaly areas in blade images, thereby detecting damage. Furthermore, a physics-informed curriculum learning strategy is designed to train this model, enabling it to learn in an easy-to-hard way. To train and evaluate the approach, we collected a large-scale blade video dataset that contains 100 optical thermal video pairs and over 55,000 images. Experimental results demonstrate that our approach achieves, for the first time, near real-time surface and structural blade damage detection in the field without stopping the normal operations of wind turbines. Our results can be found at https://aquada-go.github.io/.