A graph learning method based on image representation for fault diagnosis of wind turbine gearboxes

Data-driven fault diagnosis methods typically process images as grids or sequences in Euclidean space, which constrains the diagnostic model’s capacity to extract fault information and compromises its flexibility. This article proposes a graph learning-based fault diagnosis method. First, the vibration signal is converted into an image and divided into patches. Then, the MST-KNN graph is constructed through the minimum spanning tree algorithm and the K -nearest neighbor algorithm to achieve multi-dimensional representation of fault information. Second, the proposed adaptive graph learning module dynamically generates differentiated attention weights based on edge information, thereby fully exploiting the topological information of graph structures. Additionally, it fuses the original node features with geometric structural features to construct multi-level composite feature representations. Finally, the output layer globally aggregates the deep features and applies nonlinear mapping to project them into the fault label space for fault identification. The proposed model was experimentally validated on planetary gearbox and wind turbine gearbox (WTG) datasets. The accuracy of fault diagnosis for WTGs reached 96.84%, indicating that this method has certain advantages in fault diagnosis of WTGs.