New Method for Remaining Useful Life Prediction Based on Recurrence Multi‐Information Time‐Frequency Transformer Networks

ABSTRACT As the critical technology of prognostics and health management (PHM), the remaining useful life (RUL) prediction has received much attention. Deep learning algorithms based on data‐driven stand out among various methods. However, convolutional neural networks/recurrent neural networks (CNNs/RNNs) suffer from design gaps, making it challenging to achieve parallel processing of long‐time series data. In addition, most methods focus only on point prediction but lack uncertainty assessment of prediction results. Therefore, a recurrence multi‐information time‐frequency (RMTF) Transformer network is proposed in this paper, which is a kind of Transformer network based on a slice memory recurrence mechanism (SMRM), with multiple feature extraction encoders, and has the ability to extract time‐frequency features. RMTF Transformer network can realize effective extraction and fusion of long time series information, cross‐time period information, multi‐source information, and time‐frequency information. In addition, a Bayesian neural network (BNN) based on variational inference is used to predict the interval of RUL. The advancedness of our proposed method is verified and compared by the aero‐engines dataset and tool wear dataset. In particular, the RMTF Transformer demonstrates significant advantages over mainstream advanced models under complex operating conditions in two sub‐datasets of CMAPSS, FD002, and FD004. Specifically, for the FD002 dataset, the RMSE is reduced by 8.66%, and the SF is reduced by 23.37%. For the FD004 dataset, the RMSE is lowered by 6.36%, and the SF is decreased by 13.83%. The experimental results show that the proposed method can effectively predict RUL and assess the uncertainty of the prediction results.