A meta-learning parameter-space conditional flow matching method for remaining useful life prediction of rolling bearings with limited sensor data

Abstract Reliable remaining useful life (RUL) prediction of bearings is crucial for ensuring the safe and efficient operation of rotating machinery in industrial environments. However, in industrial scenarios, RUL prediction is often constrained by limited sensor measurement data and incomplete life-cycle observations. Distribution shifts under varying operating conditions further reduce prediction reliability. Gradient-based meta-learning methods enable rapid task adaptation. Their reliance on predefined optimization trajectories may lead to instability and limited generalization. To overcome these limitations, this paper proposes Meta-ParaFlow for RUL prediction under limited sensor measurement data. Meta-ParaFlow is a meta-learning framework based on parameter-space conditional flow matching. Meta-ParaFlow learns a task-conditioned velocity field in parameter space. Source-trained reference parameters are used to construct the flow-matching supervision. At inference, task-specific model parameters are generated from Gaussian noise by numerically integrating the learned velocity field. This process requires no task-specific gradient updates. A conditioned parameter flow network (CPFN) is designed to estimate the velocity field. A multi-candidate sampling strategy generates alternative parameter sets under the same task condition and selects the final predictor according to their prediction errors. Experiments are conducted on the IEEE PHM 2012 and ABLT-1A bearing vibration datasets. Meta-ParaFlow achieves the lowest average RMSE of 0.0458 on the IEEE PHM 2012 dataset and 0.0467 on the ABLT-1A dataset. These results demonstrate the effectiveness of Meta-ParaFlow in improving RUL prediction reliability under limited sensor measurement data.