Condition monitoring of wind turbine based on a novel spatio-temporal feature aggregation network integrated with adaptive threshold interval

Condition monitoring (CM) technology based on supervisory control and data acquisition (SCADA) data is crucial for ensuring reliable operation and reducing maintenance costs of wind turbine (WT). However, SCADA data is highly dimensional and sophisticated, making it challenging to extract features and determine fault thresholds to improve CM accuracy. In this paper, we propose a new approach for effectively estimating the operating condition and diagnose the faults of WTs, which is founded on the integration of spatio-temporal feature aggregation network and adaptive threshold interval. Briefly, the Kalman filter is used to preprocess the SCADA data and the variables are selected by Pearson correlation coefficient to improve data quality. Then, a novel stacked sample convolution and interaction network (SCINet) embedded in a coordinate attention mechanism (CA) is designed to extract and merge the multi-dimensional and multi-resolution spatial and temporal features. Moreover, an adaptive threshold interval is proposed for early warning of fault occurrence based on Monte Carlo (MC) dropout and kernel density estimation (KDE). The effectiveness of the proposed method is demonstrated through two real cases from a wind farm, which show superior performance compared to other comparative methods.