Domain adaptive fault diagnosis algorithm based on multi-graph convolution for rotating machinery

In real industrial production, bearings usually operate under variable operating conditions. However, existing deep learning-based bearing fault diagnosis methods overlook the complex structural relationships between fault signal data, leading to limitations in generalization ability. To address the above problems, a novel domain adaptation multi-graph convolutional network (DAM-GCN) bearing fault diagnosis method is proposed in this paper. First, we extract the basic fault signal features with the help of a convolutional neural network (CNN). Subsequently, Top-k graph, k-NN graph and Radius graph are used to generate graph structures, which utilize the local, similarity and density information in the data respectively, enabling the network to deeply capture the fault structure characteristics from multiple perspectives. Second, to ensure that these features can be effectively compared on the same scale, a contrastive learning strategy is employed to minimize the feature similarity within the same feature tensor to improve the distinguishability and expressiveness of the features. Finally, we jointly consider the classification loss and domain alignment loss. By minimizing the distribution and graph structure differences between the target and source domains, the fault diagnosis ability of the model under different working conditions is enhanced. Numerous experimental findings show that the proposed domain-adaptive multi-graph neural network-based approach outperforms existing SOTA methods.