Wind turbine fault detection and identification via self-attention-based dynamic graph representation learning and variable-level normalizing flow

Effective wind turbine (WT) condition monitoring is significant to improve wind power generation efficiency and reduce operation and maintenance costs. Supervisory control and data acquisition (SCADA) data are widely utilized for WT condition monitoring due to their low cost and accessibility. However, the intricate interdependencies among SCADA variables affect the accuracy of WT fault detection, and few methods provide identification for the anomaly cause. To solve these issues, this paper proposes an unsupervised fault detection and identification method based on self-attention-based dynamic graph representation learning and variable-level normalizing flow. Firstly, a dynamic graph representation learning model based on spatial and temporal self-attention mechanisms is proposed. It can effectively learn the dynamic and mutual relations among variables for early fault detection. Secondly, a variable-level normalizing flow is proposed for discriminative density estimation of variables, which can realize component fault localization. Finally, a node deviation index based on contrast graph is proposed to identify the root cause of anomalies. Experimental results using WT data from a wind farm in Northwest China prove that the proposed method has better accuracy and interpretability in WT fault detection and identification, which displays better effectiveness in practical wind energy applications.