Motor Fault Diagnosis Based on BiLSTM and GNN

In recent years, Graph neural network(GNN) have gained significant popularity in fault diagnosis applications due to their outstanding capability to aggregate features and model non-linear data. However, the performance of GNN networks is usually affected by interference information and the limitation of network depth, which makes it difficult to cope with the complex diagnostic environment in the real world. Therefore, this paper proposes a motor fault diagnosis method based on Bidirectional Long Short-Term Memory(BiLSTM) and GNN. Firstly, by leveraging the special gating mechanism and bi-directional modeling capability of BiLSTM, harmful information in the motor signal is effectively filtered and the backward and forward correlated temporal features are extracted; Secondly, the aggregation and propagation capabilities of GNN networks for features are utilized to extract spatial features at different levels and scales, Finally, the classification of fault characteristics is achieved through a fully connected layer. Interference experiments were conducted on seven different GNN models on two publicly available motor datasets. The results indicate that our proposed method significantly improves the recognition rate of motor faults. In the node-level diagnostic task, the average accuracy is 89.71%, representing a 43.56% improvement over the baseline GNN models. In the graph-level fault diagnosis task, the average accuracy is 91.71%, demonstrating a 48.59% improvement over the baseline GNN models.