TCN-BiLSTM bearing RUL prediction method based on fusion of TCN-CPC and statistical features

Abstract Bearings are prone to wear, fatigue, and other issues during long-term operation due to mechanical stress, insufficient lubrication, and other factors. Accurate Remaining Useful Life (RUL) prediction is crucial for equipment maintenance. Among existing methods, statistical features struggle to capture complex nonlinear relationships, while end-to-end models suffer from poor interpretability. To address this, this paper proposes a Temporal Convolutional Network-Bidirectional Long Short-Term Memory (TCN-BiLSTM) prediction method that integrates Temporal Convolutional Network-Contrastive Predictive Coding (TCN-CPC) with statistical features. The approach extracts long-term dependencies through TCN and enhances feature representation using CPC. SHapley Additive exPlanations (SHAP) analysis reveals that statistical features contribute significantly during the healthy stage, while features extracted by deep learning exhibit stronger discriminative power in the failure stage. Building on this, the TCN-BiLSTM model further improves the accuracy of bearing condition prediction for RUL estimation. Experimental results demonstrate that the proposed TCN-BiLSTM achieves superior performance with average improvements of 43.2% in RMSE, 58.8% in MAE, and 55.1% in R 2 compared to the strongest baseline BiLSTM-Attention. This validates the effectiveness of the feature fusion strategy.