Abstract: Aiming at the issues of non-stationary,feature aliasing and low diagnostic accuracy of fault vibration signals from wind turbine gearbox,a fault diagnosis method for wind turbine gearbox based on graph attention networks（GAT）is proposed. Firstly,nodes and edges are defined by the frequency spectrum of the raw vibration signal to construct a visibility graph. Then,taking visibility graph data as input,the neighbor self-attention mechanism is embedded in GAT to adaptively extract node features and structure features of visibility graph signals. Finally,the classifier is used to classify and recognize the extracted node features. The experimental results of planetary gearbox dataset and wind turbine gearbox dataset show that the proposed method has higher accuracy,better robustness and noise immunity than machine learning,deep learning and other graph neural networks,which can effectively achieve end-to-end intelligent fault diagnosis.