An uncertainty-incorporated active data diffusion learning framework for few-shot equipment RUL prediction

Abstract

In predicting the remaining useful life (RUL) of critical equipment, the challenge of obtaining degradation data and the limitation of data volume lead to few-shot problems that significantly impact prediction accuracy. To address this issue, this paper introduces a reinforcement learning feedback loop mechanism for predicting the RUL of critical equipment. Initially, the framework uses a data diffusion model to generate a dataset that closely approximates the distribution of the labeled samples for data augmentation. Subsequently, Bayesian deep learning and Monte Carlo (MC) dropout inference provide uncertainty quantifications for RUL interval predictions. An active learning strategy, which is based on uncertainty and diversity, converts unlabeled samples into labeled samples, thereby selecting the most effective training dataset. In each iteration, the model adjusts its strategy for selecting and generating data based on the current state of learning, dynamically adapting to the needs of the learning process via Bayesian methods. The proposed prediction framework was validated through experiments using the C-MAPSS and NASA battery datasets. The results indicate that the application of data diffusion and active learning strategies significantly enhances prediction performance, increasing confidence by 42 %. Comparative experiments with other benchmark methods demonstrate that the proposed method reduces prediction uncertainty by at least 15 %.