An Online Anomaly Monitoring Method Based on Multiscale Spatiotemporal Graph Learning for Wind Turbine
Qing Xu, Dazhong Ma, Haoran Zhao, Qiuye Sun · IEEE Transactions on Industrial Informatics · 2025
The variation in multivariate time series (MTS) under wind turbine (WT) operational conditions makes it challenging for traditional anomaly detection methods to model expected behavior under normal conditions, resulting in failure to identify anomalies. To address this, an online anomaly detection method based on multiscale spatiotemporal graph learning is proposed, enabling prompt anomaly detection. First, an adaptive multiscale graph correlation forecasting network is introduced, which autonomously learns temporal and feature dependencies at each scale. Next, a dynamic spatiotemporal graph variational autoencoder is presented to model the MTS’s spatiotemporal correlations and capture the normal operation patterns. In addition, we propose a nonparametric dynamic threshold updating mechanism using Welch’s t-test to adapt to changing operating conditions based on anomaly scores. The proposed method jointly optimizes the forecasting and pattern reconstruction networks to derive spatiotemporal graph representations and anomaly scores, effectively identifying anomalies that deviate from normal operating states. Experiments on real WT data demonstrate the method’s ability to detect anomalies earlier, with evaluation metrics showing at least a 2% improvement in anomaly detection accuracy compared to existing methods.