Unsupervised Wind Turbine Blade Damage Detection With Memory-Aided Denoising Reconstruction
Xiaodong Jia, Xiao Chen
Abstract
AI-based automated wind turbine blade damage detection has significant economic value. This article proposes a novel memory-aided denoising autoencoder for unsupervised blade damage detection, which detects structural damages with a denoising autoencoder and detects logical damages with a designed memory system. Specifically, by training a denoising autoencoder on blade data with synthesized artificial damage as input and corresponding damage-free data as output, it learns to differentiate between damage and normal domains, thereby detecting structural damage. Furthermore, a memory system encompassing memory reading, writing, and management is designed for the denoising autoencoder, enabling the approach to detect logical damages effectively by leveraging information read from memory. Experimental results on a blade damage dataset, MvTec anomaly detection dataset, and MvTec logical constraints anomaly detection not only show our approach outperforms state-of-the-art methods by 4.7% on logical damages without compromising its performance on structural damages but also verify its generalization and robustness.