Rolling bearing fault diagnosis methods based on deep learning models usually apply regular data based on Euclidean space, and the relationship between data has not received corresponding attention. However, graph data based on non-Euclidean space has more comprehensive information expression ability and its intrinsic relationship is of great significance for fault identification. At present, the graph neural network diagnostic approach based on time domain cannot well capture the local features of the data adequately while constructing graph data, and the diagnosis accuracy is low. Therefore, a new method of bearing fault diagnosis based on graph neural network is proposed in this paper. Firstly, the time series is transformed by the visibility algorithm into non-Euclidean space graph structure data, which considerably enriches the internal relationship between data. Secondly, to identify and categorize rolling bearing failure characteristics, five different types of graph neural networks are fed with the generated graph data. At the same time, the effects of max, avg, and sum readout operations on the classification results of the graph neural network are analyzed. Finally, experiments are carried out on two bearing experimental data sets. The experimental findings demonstrate that the average accuracy of the five graph neural networks under the impact of max and avg readout operations is greater than 89 and 100, respectively. Graph Attention Network (GAT) has the highest classification accuracy among them in this research, with an average accuracy of more than 90. In general, graph neural networks can achieve good results in fault diagnosis of graph-level classification tasks, and the relationship information extracted by the visibility algorithm has certain advantages. The accuracy of this approach is somewhat enhanced when compared to previous time domain diagnostic methods of graph neural networks.
展开▼
