Enabling generalizable RUL prediction for equipment: a dual-dimensional continual learning approach based on deep Gaussian process

Abstract

Abstract Accurately predicting Remaining Useful Life (RUL) is critical for the intelligent maintenance of high-end rotating machinery, which often operates under complex and variable conditions. However, conventional predictive models tend to degrade in performance when deployed in novel scenarios due to domain shifts. This study proposes a novel predictive framework that integrates Deep Gaussian Processes (DGP) with a dual-dimensional constrained continual learning (CL) strategy to address these challenges. Specifically, the framework leverages the probabilistic nature of DGP to quantify the inherent uncertainty during degradation. To preserve cross-domain degradation knowledge, we employ Elastic Weight Consolidation, which anchors critical weights, and Gradient Episodic Memory, which corrects update directions in the gradient space that conflict with established degradation patterns. Together, these techniques form a dual-dimensional constrained CL framework. This synergistic approach enables the DGP to incrementally assimilate knowledge from different operating conditions (cross-condition) and diverse machine types (cross-device) without suffering from catastrophic forgetting. Experimental results demonstrate that the proposed framework not only exhibits outstanding resistance to catastrophic forgetting and excellent multi-scenario adaptability, but also achieves superior predictive accuracy compared to state-of-the-art methods by leveraging the dual-dimensional continual learning strategy. Furthermore, the integrated DGP module provides robust uncertainty quantification, offering a reliable basis for intelligent maintenance decision-making in complex operating environments.