E-BMAML: An enhanced Bayesian meta-learning framework with adaptive asymmetric calibration for uncertainty-aware few-shot industrial diagnostics

• Proposes E-BMAML framework for uncertainty-aware industrial RUL prediction. • Designs an Asymmetric Dual-Scale Calibration loss function for more reliable uncertainty quantification. • Builds a front-end module to process non-ideal industrial data like non-uniform sampling. • Reduces calibration error by 92.4% on public benchmarks and prediction error by 66.5% on real-world wind turbine data. • Reduces risk rate by 65.1% via conservative bias in wind turbine RUL tasks Reliable uncertainty quantification in remaining useful life (RUL) estimation is essential for predictive maintenance (PdM) in industrial settings. However, existing methods often struggle to produce well-calibrated confidence estimates under heterogeneous operating conditions. We propose E-BMAML, a Bayesian meta-learning framework that incorporates an Asymmetric Dual-Scale Calibration (ADSC) loss to jointly optimize predictive accuracy and uncertainty calibration, along with an adaptive preprocessing module to accommodate non-uniform sampling and nonlinear degradation patterns commonly found in industrial data. Experiments on the PHM and XJTU benchmarks show a 92.4% reduction in mean absolute calibration error relative to the compared methods. Validation on real-world wind turbine data further shows a 66.5% reduction in RMSE and an 82.4% reduction in MACE over traditional linear labeling approaches, with E-BMAML attaining the lowest Negative Log-Likelihood among all baselines. In addition, the prediction risk rate is reduced from 98.3% to 33.2% compared to GPR, reflecting a shift in the error distribution toward a conservatively biased region. These results suggest that the proposed framework offers a practically viable approach to risk-aware RUL prediction in complex industrial environments.