Bearing remaining useful life prediction using spatial-temporal multiscale graph convolutional neural network

Abstract

Listen

Translate article

Remaining useful life (RUL) prediction of bearings is important to guarantee their reliability and formulate a maintenance strategy. Recently, deep graph neural networks (GNNs) have been applied to predict the RUL of bearings. However, they usuallylack dynamic features, use manual stage identification, and experience the over-smoothing problem, which will have a negative effect on the prediction accuracy. This paper proposes a new framework for bearing RUL prediction based on spatial-temporal multi-scale graph convolutional neural network (STMSGCN), which can improve the accuracy of prediction by solving the above-mentioned problems. Specifically, different to the most-used static feature of bearings, a dynamic feature that can capture the time-varying change of vibration energy is proposed. A sliding window alarm method is proposed to detect the fault occurrence time (FOT), which can offer an accurate healthy stage rather than human-defined methods. Then, the STMSGCN is proposed to predict the RUL of bearings, which solves the over-smoothing problem of the deep GNN model. The PRONOSTIA platform is adopted to verify the proposed method. The results verify that the sliding window alarm method can detect the FOT faster for slowly degrading bearings, and the proposed STMSGCN structure gives higher prediction accuracy compared to the existing methods.