Fault Diagnosis for Rolling Bearings Based on Novel Visibility Graph and GCN Scheme

Recently, the field of intelligent fault diagnosis has made great breakthroughs and achievements since feature extraction has a powerful ability to learn data. However, in non-Euclidean spaces, the types of bearing fault relationships are complex and the number of relationships is inconsistent, resulting in traditional deep learning methods that cannot accurately mine the potential relationships between fault information. To solve this problem, we propose a fault diagnosis method for rolling bearings based on a novel visibility graph (VG) and a new graph convolution neural (GCN) network. Specifically, a novel weighted visibility graph (WVG) method which can convert time series data into graph data is proposed. It can superiorly reflect the complex relationship between each factor in bearing fault diagnosis. In order to achieve fault diagnosis in the way of graph classification, we propose a new method SGIN+. It combines GraphSAGE and an improved graph isomorphic network (GIN), so that it can accurately learn the graph structure in large-scale classification tasks. The effectiveness of both WVG and SGIN+ is verified by a real bearing dataset.