Prediction of Remaining Service Life for Rolling Bearings Using a CapsTCN‐Transformer Hybrid Model

ABSTRACT This paper proposes a rolling bearing remaining useful life (RUL) prediction method based on a hybrid Capsule‐enhanced Temporal Convolutional Transformer with Degradation Stage Embedding (CapsTCN‐Transformer) model. Addressing the limitations of traditional RUL prediction methods that rely on manual feature extraction, this study innovatively constructs an end‐to‐end prediction model architecture. First, a Convolutional Neural—Bidirectional Long Short‐Term Memory (CNN‐BiLSTM) network is employed for automatic detection of degradation onset points. By integrating the local feature extraction capability of 2D convolutions with the temporal modeling advantages of bidirectional LSTMs, this approach accurately identifies critical bearing degradation thresholds. Building upon this foundation, the proposed CapsTCN‐Transformer model achieves automatic extraction of multi‐level features from bearing vibration signals and precise modeling of temporal dependencies through the synergistic interaction of multi‐scale capsule layers, causal dilated convolutions, and Transformer attention mechanisms. Experimental results demonstrate that on the XJTU‐SY bearing dataset, the model achieves significantly superior average MSE (0. 0076), MAE (0. 0521), and R 2 (0. 9200) compared to traditional Transformer and TCN‐LSTM models. Particularly in Task2₄, it exhibits outstanding performance with MSE = 0. 0013 and R 2 = 0. 9797. This study provides a high‐precision RUL prediction solution for intelligent maintenance of industrial equipment, demonstrating significant engineering application value.