Combined use of multiple entropy and weighted square envelope for remaining useful life prediction of railway gearbox with Bayesian deep learning

Abstract

The accurate prediction of the remaining useful life (RUL) of rail train gearboxes is significant for the safety of rail transportation. However, traditional RUL prediction methods have two main limitations: the health indicator construction is insufficient in noise suppression and the point-based prediction paradigm struggles to quantify model uncertainty. To settle these issues, a weighted squared envelope multi-entropy Bayesian deep learning (WSEME-BDL) method is proposed. Above all, the weighted squared envelope entropy is used to suppress noise and enhance degradation features in vibration signals, breaking through the sensitivity bottleneck of traditional feature extraction methods to early weak faults. Second, an adaptive degradation stage segmentation is designed to accurately identify critical point in degradation stage. Finally, a Bayesian deep neural network model is constructed to predict RUL of railway train gearboxes by integrating variational distribution and sampling strategies. This model outputs the uncertainty confidence intervals while predicting the RUL. The data obtained from a self-built parallel gearbox test platform over its full-life cycle is utilized to validate the accuracy of this method. Experiments show that compared to deterministic models such as SVM, LSTM and GRU, and BDL prediction under single entropy feature extraction, WSEME-BDL significantly improves the accuracy and reliability of prediction results.