Few-Shot Probabilistic RUL Prediction With Uncertainty Quantification of Slurry Pumps

Abstract

Slurry pumps are crucial in the mining industry, as their performance and reliability directly affect the efficiency and safety of mining production systems. However, existing remaining useful life (RUL) prediction models face challenges due to the scarcity of degradation samples caused by the difficulty of obtaining degradation data in industrial settings, and their inability to provide prediction result confidence intervals (CIs). This article proposes a meta transformer with uncertainty quantification (MTUQ) based on an approximate Bayesian framework. The model enhances the capability to quickly adapt to new tasks in few-shot scenarios through a dual-loop meta-learning strategy, addressing the issue of sample sparsity. Additionally, random subnetwork sampling (RSNS) is proposed to achieve approximate Bayesian posterior distribution and combines Kernel density estimation (KDE) to quantify the model’s prediction uncertainty. Experimental results on the few-shot RUL prediction of slurry pumps in actual production scenarios demonstrate that MTUQ outperforms baseline methods in handling sparse samples and quantifying uncertainty, improving its prediction accuracy and reliability.