Performance degradation prediction of rolling bearing based on temporal graph convolutional neural network

Abstract

Aiming at the prediction model of bearing performance degradation based on recurrent neural network (RNN) and its variants ignores the feature spatial correlation, and cannot effectively handle long time series data, this paper proposes a rolling bearing performance degradation prediction model based on temporal graph convolutional neural network (T-GCN). For non-stationary and nonlinear characteristics of vibration signals, this paper introduces a rolling bearing feature evaluation method based on multiscale dispersion entropy (MDE) to better characterize time series. To effectively solve the spatial correlation problem between samples and features, this paper uses the topological structure of a path graph to build a graph model and combines gated recurrent unit (GRU) and graph convolutional neural network (GCN) to build a T-GCN prediction model. Finally, this article established a rolling bearing fault prediction experimental platform and validated it using the University of Cincinnati public dataset. The experiment shows that compared with GRU, GCN, and LSTM models, the RMSE and the MAE evaluation indicators based on the T-GCN model have decreased by 6 % to 28 % and 11 % to 28 %, respectively, which suggests that the T-GCN model has a higher prediction accuracy and a better model fitting goodness.