Spatio-temporal feature fusion method for fault warning in wind turbine drivetrain systems using ALGCN-LSTM

The increasing faults and failures of wind turbine drivetrains has been attracting significant attention from manufacturers and operators. However, existing fault warning models for wind turbine drivetrains fail to effectively capture both spatial and temporal features of supervisory control and data acquisition system data within a single model, which reduces the confidence of fault warning and limits the model universality. In this paper, a spatio-temporal feature fusion method that integrates an adjacency matrix learnable graph convolutional network (ALGCN) with a long short-term memory (LSTM) network is proposed for incipient fault warning for wind turbine drivetrains. Initially, a maximal information coefficient based data-mechanism dual-driven graph structure construction method is proposed to embed physical information into the graph convolutional network (GCN). Subsequently, the GCN is improved by a learnable adjacency matrix, and both spatial and temporal features are extracted using the parallel arrangement of ALGCN and LSTM, significantly enhancing the feature extraction capability. Finally, prediction residuals from the corresponding measurements are employed to evaluate the condition of various drivetrain components within a single model. Experiments conducted on operational data from on-site wind turbines demonstrate that the proposed ALGCN-LSTM model effectively learns intrinsic healthy operating patterns and delivers robust performance under complex operating conditions, achieving accurate healthy-state prediction and reliable early warning across different fault types.