Physics-Informed Anomaly Detection and Fault Localization for Wind Turbines Using Multi-Strategy DASVDD and Bidirectionally Adjusted Reconstruction Error

Abstract

Anomaly detection and fault localization are essential for the safe and economic operation of wind turbines and the mitigation of unplanned outages. A novel physics-informed anomaly detection and fault localization method is developed to address high false alarm rates and insufficient localization accuracy in small-sample scenarios. First, a physics-informed autoencoder, integrated with explicit algebraic physical constraints, is developed to verify the effective embedding of inherent physical relationships. Second, building upon this autoencoder, a multi-strategy deep autoencoder-based support vector data description (MS-DASVDD) for anomaly detection that incorporates physical constraint regularization, latent variable alignment, and transfer learning is developed, significantly reducing false alarms and missed detections with limited training data. Finally, a physics-informed method for bidirectional reconstruction error adjustment is proposed, which penalizes errors consistent with physical constraints while accentuating inconsistent ones, thus significantly boosting error-driven fault feature localization accuracy under limited fault samples. Case studies on two wind farms (40 MW and 30 MW) in Northern China show that the proposed physics-informed method markedly improves precision, recall, and F1 score, suppresses spiky anomaly scores, and enhances the contribution of fault features, confirming its effectiveness and robustness; sensitivity and comparative analyses further validate its superiority.