Abstract
The industrial adoption of machine learning for remaining useful life (RUL) prediction is constrained by data, not models. Run-to-failure records are rare and often noisy, incomplete, irregular, or inconsistent because full lifecycle testing is costly and time-consuming. Deep learning struggles under these constraints, demanding large datasets and offering limited uncertainty quantification. We address robust, uncertainty-aware RUL prediction in data-scarce settings, using aluminium electrolytic capacitors (AECs) as a representative case with complex, nonstationary degradation. Our approach targets scenarios with limited lifecycle observations—often fewer than 30 data points per unit—where traditional models exhibit reduced performance. We propose a Gaussian process regression (GPR) framework offering Bayesian uncertainty estimates and resilience under limited data. To handle nonstationary behavior, we introduce spectral mixture kernels (SMKs) to model frequency-domain structure. For calibrated uncertainty, we integrate quantile regression with split-conformal prediction, utilizing a Mondrian phase-stratified scheme. Under this scheme, local conformal calibration adapts intervals within early/mid/late regimes. For deployment on resource-constrained platforms, we implement sparse GPR (SGPR) with $k$-means inducing points. Evaluations on two accelerated ageing datasets show marked improvements: in Dataset 1 (105 $^{\circ }$C, 20 V), SMK-GPR reduces mean RMSE from 47.43 to 15.90 h (66.5%); in Dataset 2 (125 $^{\circ }$C, 22 V), from 11.24 to 3.56 h (68.3%) along with empirical coverage exceeding nominal targets. These results demonstrate robustness across degradation regimes and support accurate, interpretable, and deployable RUL predictions in real-world industrial settings.
Funding Information
  • Department of Atomic Energy, Government of India

This publication has 31 references indexed in Scilit: