Robust uncertainty quantification for online remaining useful life prediction with randomly missing and partially faulty sensor data

Uncertainty quantification (UQ) is crucial for accurate remaining useful life (RUL) prediction of equipment, playing a pivotal role in developing reliable maintenance strategies. However, in real-world online prediction scenarios, sensor monitoring data often exhibit incompleteness due to randomly missing and partially faulty sensor data, which can significantly compromise UQ effectiveness. Furthermore, existing methods like Bayesian neural networks and Gaussian process regression rely on specific distributional or model assumptions and struggle to produce statistically valid prediction intervals. To overcome these challenges, this paper introduces an innovative method — RobustUQ. It begins with a pre-training strategy to extract representations from sensor monitoring data, followed by a diffusion model to fit the distribution of missing values for effective imputation. Subsequently, conformal prediction is applied for UQ in RUL prediction, alongside deep metric learning to identify residual patterns and address the issue of non-exchangeability in spatiotemporally dependent sensor data. Theoretically, the prediction intervals constructed using this method are statistically valid. Experimental results demonstrate that RobustUQ accurately and robustly quantifies uncertainty in online RUL prediction. More importantly, RobustUQ surpasses existing methods by being distribution-independent and model-agnostic, enabling seamless integration with current RUL prediction models while delivering more precise and reliable prediction intervals.