A novel multi-task fault detection model embedded with spatio-temporal feature fusion for wind turbine pitch and drive train systems

Abstract

Fault detection based on supervisory control and data acquisition (SCADA) data is crucial for ensuring reliable operation of wind turbine (WT). However, the current research of fault detection mainly focuses on the entire machine or one certain subsystem of WT, which is difficult to monitor the faults of different subsystems at the same time. We propose a novel multi-task model embedded with spatio-temporal feature fusion to detect the faults of WT pitch and drive train systems simultaneously. Briefly, SCADA data is preprocessed to improve the data quality firstly, including data interpolation and variables selection. In the proposed multi-task model, the multi-scale convolution encoding network (MSCEN) is constructed as shared layer to extract multi-dimensional and multi-resolution spatial features. And the multi-branches structure is designed to extract temporal features based on dynamic temporal attention module (DTAM) and bidirectional Gated Recurrent Unit (Bi-GRU). Moreover, dynamic weight average (DWA) is used to optimize the training process of multi-task models. Four actual cases from one wind farm are used to illustrate the effectiveness of the proposed method, and they perform better than other comparative methods.