Graph Neural Network for Fault Diagnosis with Multi-Scale Time-Spatial Information Fusion Mechanism

Due to the long-term operation of planetary gearboxes in strong noise environments and changing working conditions, the collected vibration signals exhibit weak fault characteristics and variable signal patterns, making them difficult to identify. Intelligent fault diagnosis of planetary gearboxes under these conditions remains a challenging task. In order to achieve high diagnostic accuracy and strong model generalization performance, a fault diagnosis method using a graph neural network with a multi-scale time-spatial information fusion mechanism is proposed. The method first uses convolution kernels of different scales to extract features from the original vibration signal, reducing the masking effect of strong noise signals on valuable information and enhancing its feature expression ability. A channel attention mechanism is then constructed to adaptively assign different weights among different channels to features of different scales, enhancing features in segments of information containing crucial fault characteristics. Finally, the multi-scale features of the convolution module output are used to construct graph data with spatial structure information for graph convolution learning. This approach allows for the full utilization and deep fusion of multi-dimensional time domain information and spatial correlation information, effectively improving the accuracy of diagnosis and the generalization performance of the model. The proposed method was verified using a fault dataset of wind power equipment with planetary gearbox structure. The average diagnosis accuracy of the proposed method was found to reach 98.85% and 91.29% under cross-load and cross-speed conditions, respectively. These results are superior to other intelligent diagnosis methods, including deep convolutional neural networks with wide first-layer kernels (WDCNN), long short-term memory network (LSTM), residual network (ResNet), and multi-scale convolution neural network (MSCNN). Therefore, the strong generalization performance and superiority of the proposed method were confirmed.