A Novel Directed Graph Convolutional Neural Network for Rolling Bearings in Fault Diagnosis

As deep learning is broadly operationalized, fault detection and diagnosis are becoming increasingly intelligent. This has resulted in a significant increase in research efforts aimed at understanding and learning the different characteristics of data structured as graphs. However, in fault diagnosis, due to the vast and complex relationships between the data, we often overlook the important feature of the direction of relationships between data, resulting in confusion of some fault categories when performing classification. To address this problem, we first transform time series data into a directed graph structure using a weighted directed visualization graph (WDVG) method. To reduce the impact of distant nodes perceived as noise, the edges are assigned weights based on the difference between the sampling points. Secondly, in order to mine the directional features of relationships between nodes in large-scale graph structures, our proposed approach named DSGIN integrates graph isomorphism network (GIN) and GraphSAGE to extract directional features. Ultimately, we achieve the graph classification task. We validate the effectiveness of WDVG and DSGIN through actual datasets.