Cross regime bearing remaining useful life estimation via domain adversarial meta survival learning

Accurate estimation of the remaining useful life (RUL) of rolling element bearings is the cornerstone of condition-based maintenance for rotating machinery operating under heterogeneous speed–load regimes. Two obstacles persist in the current literature: (i) the severe distributional mismatch across operating conditions degrades source-trained models when target-regime run-to-failure data are scarce, and (ii) existing approaches overwhelmingly yield point-valued predictions devoid of reliability-grounded confidence bounds. This paper presents DAMSNet, a Domain-Adversarial Meta-Survival Network that simultaneously overcomes both obstacles through a tightly integrated architecture. Situated at the intersection of meta-learning-based domain generalization and few-shot adaptation, a condition-steered multi-scale fusion (CSMF) encoder maps speed and load metadata into channel-wise attention scores that dynamically blend three parallel dilated convolution branches, extracting degradation features whose scale emphasis is explicitly governed by the prevailing operating state. To eliminate residual regime-specific distributional artifacts, a gradient reversal discriminator is attached to the encoder output, enforcing feature invariance through adversarial optimization without erasing degradation-discriminative structure. A recurrent temporal module then summarizes the aligned feature trajectory, and a Weibull survival head converts the resulting hidden representation into scale and shape parameters via maximum likelihood, producing a full lifetime distribution—and hence both a point RUL estimate and a calibrated prediction interval—at every inference step. A model-agnostic meta-learning wrapper organizes training into leave-one-regime-out episodes with learnable per-parameter adaptation rates, so that as few as one to five target-condition trajectories enable effective deployment-time transfer. Strict bearing-level data partitioning ensures that no test trajectory is exposed during adaptation, preventing trajectory-level information leakage. Experiments on the XJTU-SY and Pronostia accelerated degradation benchmarks yield a within-regime root mean square error of 0.128 and a 3-shot cross-regime error of 0.152, representing improvements of 23.4% and 16.9% over the strongest Weibull-recurrent baseline. The accompanying 80% prediction intervals reach 84.1% empirical coverage with a mean width of only 16.3 cycles. Multi-level calibration analysis, reliability diagrams, and continuous ranked probability scores further validate the uncertainty estimates. Interpretability analyses—including Grad-CAM feature attribution and Weibull parameter trajectory inspection—confirm that the network learns physically meaningful failure dynamics. Ablation analysis, including combined-component ablations that reveal synergistic interactions, t-SNE feature visualization, and statistical significance testing over ten independent trials collectively validate the necessity of every architectural component and the robustness of the reported gains.