Remaining useful life prediction for rotating machinery based on dynamic graph and spatial–temporal network

Abstract

Deep learning-based rotating machinery remaining useful life (RUL) prediction approaches rarely consider spatial dependencies and global temporal correlation of monitoring signals simultaneously. Superiorly, graph convolutional networks (GCNs) learn relationship information among graph nodes, considering the spatial dependencies of signals. It is beneficial for constructing high-quality graphs to improve the prediction performance in single-sensor monitoring scenarios. In this paper, an RUL prediction approach for rotating machinery based on a dynamic graph and a spatial–temporal network (STNet) is proposed. Short-time Fourier transform is introduced to obtain node features, and dynamic edge connections are established through node importance weights. Furthermore, an STNet is constructed to learn graph features, in which the GCN is used to mine spatial dependencies of input graphs, and a bi-directional long short-term memory network is applied to capture global temporal correlations. Finally, an autoencoder-based graph readout layer is designed to pass learned graph features. Case studies are conducted to demonstrate its effectiveness.