基于自监督特征构建的轴承寿命预测

o improve the prediction accuracy of bearing remaining useful life (RUL) and address challenges of scarce labeled data and complex working conditions, this study proposes a dual-task collaborative self-supervised feature construction method. By designing a time-series prediction task and a contrastive learning task, bidirectional LSTM (Bi-LSTM) and CNN encoders are combined to extract temporal dependencies and local discriminative features from vibration signals. A multi-task joint optimization strategy is adopted to fuse these features. The time-series prediction task captures long-term dependencies in degradation trends using Bi-LSTM, while the contrastive learning task enhances feature discriminability through noise perturbation and cross-stage sampling to construct positive/negative sample pairs. Experiments on the IEEE PHM 2012 dataset demonstrate that the proposed method significantly reduces the MSE compared to traditional feature-based approaches, maintains low error rates under reduced data volumes, and exhibits superior performance in cross-device migration scenarios. Results indicate that the weighted fusion of self-supervised features and traditional statistical features significantly improves model adaptability to multi-condition and small-sample data, providing a highly robust solution for predictive maintenance in industrial applications.