Uncertainty-Aware Bearing Remaining Useful Life Prediction via Tensor-Decoupled Latent Features and Bayesian Networks

Abstract The precise prediction of the remaining useful life for rolling bearings is fundamentally bottlenecked by profound uncertainties in degradation dynamics, encompassing highly nonlinear wear trajectories, profound cross-individual variability, and fluctuating operating conditions. Existing deterministic models often struggle to capture these complex spatiotemporal shifts. To overcome this limitation, we propose a novel uncertainty-aware probabilistic framework that elegantly synergizes variational autoencoders, tensor decomposition, and Bayesian neural networks. Unlike conventional feature extraction, our method explicitly isolates and quantifies degradation uncertainties. First, variational autoencoders are deployed to map high-dimensional raw monitoring data into a robust latent degradation space. Crucially, we introduce tensor decomposition to process these latent features, disentangling complex variations and yielding precise uncertainty estimates of the feature distributions. These uncertainty-enriched representations are fed into a Bayesian neural network, producing highly reliable, probabilistic remaining useful life predictions with rigorous confidence intervals. Experiments show that this strategy successfully navigates fluctuating operating conditions and degradation trends, delivering marked improvements in both prediction accuracy and system robustness.