A Hybrid Contrastive Learning and Unscented Kalman Filtering Approach for Data-Efficient RUL Prediction

Remaining Useful Life prediction for rotating machinery is severely hindered by the scarcity of full-life data, which limits the efficacy of data-hungry end-to-end models. To address this data-efficiency challenge, this paper proposes a hybrid learning framework. First, a self-supervised health indicator is constructed from raw vibration signals using contrastive learning. The proposed model, trained with a service-time-based loss, learns a robust 1D degradation representation from minimal data. Second, an Unscented Kalman Filter is employed to model the nonlinear trajectory and stochastic fluctuations of the constructed health indicator. The UKF performs stepwise prediction to forecast when the health indicator will cross a predefined failure threshold. The framework is validated on public bearing datasets and a set of spiral bevel gear experiments, simulating a practical scenario by training on a single full-life sample. Results show the CL-based health indicator achieves superior monotonicity (0.716) and correlation (0.911) compared to baselines. The full framework demonstrates the highest prediction accuracy (0.392), validating its effectiveness and robustness for data-efficient prognostics.