An Uncertainty Quantification and Calibration Framework for RUL Prediction and Accuracy Improvement

Abstract

In prognostic and health management (PHM), predicting remaining useful life (RUL) and quantifying the uncertainties in predictions are necessary. This article proposes a Gaussian process (GP) autoregression-variational autoencoder (GPVAE) framework that can predict RUL based on degradation data, quantify predictive uncertainty, decompose this uncertainty into epistemic and aleatory types, and further quantify epistemic uncertainties on RUL-related features. Subsequently, uncertainty calibration is proposed to ensure that the quantified uncertainty matches the actual error of the model. The calibrated uncertainty is used for out-of-distribution (OOD) detection and active learning for the labeled and unlabeled data, which can improve the RUL prediction accuracy with limited computational resources and limited cost of degradation tests for obtaining RUL labels. The effectiveness of the proposed method is illustrated by the case study on lithium-ion batteries dataset.