Remaining Useful Life Prediction of Rolling Bearings Based on Time-Frequency Enhanced Transformer Model

Abstract Deep learning has achieved a lot of progress in predicting Remaining Useful Life (RUL). However, contemporary deep learning frameworks face inherent limitations, including constrained receptive fields, difficulties in capturing long-term dependencies, and singularities within the feature extraction domain. In response to these challenges, we propose a novel time-frequency enhanced Transformer model for predicting the remaining lifespan of rolling bearings. In this model, we utilize causal convolution operations to capture local features in the time-frequency domain and incorporate position encoding into the input signal. The time-frequency enhanced model exhibits end-to-end characteristics, feature fusion and selection can be performed without tedious feature extraction compared to existing methods. Furthermore, we introduce a custom Weighted Smooth Mean Squared Loss function to enhance the model’s prediction accuracy and interference resilience. In experiments, we comprehensively compare various advanced RUL prediction methods, with the results highlighting the superior performance and robust interference resistance of the time-frequency enhanced Transformer model in predicting the RUL of bearings.