M²-STGN: A Missing Data-Tolerant Learning Model for Condition Monitoring of Wind Turbines

The field data of supervisory control and data acquisition (SCADA) systems are commonly missing, which increases the difficulty of feature extraction and reduces the effective utilization of data. This negatively affects the reliability and integrality of wind turbine condition monitoring (WTCM). To address the interference caused by missing data during feature extraction, this paper proposes a multiscale masked spatial-temporal graph networks (M2-STGN) model, mainly composed of multiscale spatial and masked temporal feature learning modules. Employing graph attention networks (GAT) with different heads, spatial features of diverse granularity are extracted from multiscale in parallel. Then, based on the absence of data and the degree of temporal correlation, a temporal masked attention module is specially designed to calculate differentiated attention scores. By collaborating with temporal convolutional networks (TCN), masked temporal features can be extracted from non-missing data. Therefore, the proposed model adapts to complex and changeable missing proportions and provides robust spatial-temporal features for WTCM tasks. Furthermore, to address missed alarms caused by the discontinuous WTCM process, a fault soft judgment strategy is proposed to enhance the sensitivity of fault detection. Two real-world SCADA datasets are utilized to validate the proposed model and strategy.