Prototype-routed multi-source unsupervised domain adaptation framework for RUL prediction via online fine-tuning

The domain shift issue is often encountered in prognostics and health management (PHM) domain due to variable operating conditions, load fluctuations, environmental changes, etc. A common limitation lying in the existing transfer learning methods for remaining useful life (RUL) prediction is their reliance on a rigid offline training paradigm, which leads to distinct challenges for two commonly used transfer learning methods: the domain generalization (DG) methods suffer from unreliable performance, while unsupervised domain adaptation (UDA) methods are constrained by high- and costly-data requirements. To overcome these limitations, we propose a prototype-routed multi-source unsupervised domain adaptation framework via online fine-tuning (PR-OFT) for RUL prediction in this paper. Specifically, a prototype-routed degradation stage identification (PDSI) model is designed firstly, which integrates an efficient Mamba-based feature extractor with a novel prototype-guided supervised contrastive learning strategy to precisely identify the health state stage of target sample in real time. Based on the online identified stage, a candidate knowledge base is then dynamically constructed and enriched via a pseudo-domain augmentation strategy. Furthermore, the framework precisely routes and matches the optimal knowledge for the target sample, executing a one-time and distribution-aligned online fine-tuning to instantly generate a personalized predictor. The proposed PR-OFT framework follows a novel paradigm of generalizing at training and adapting at testing, i.e., it requires no exposure to target domain data during training and dynamically constructs personalized predictors for individual unlabeled target samples arriving in a data stream during inference. The proposed PR-OFT method is experimentally validated on two publicly available bearing datasets as well as a hard disk drive dataset, where the proposed method consistently outperformed several state-of-the-art methods, demonstrating its outstanding predictive performance and strong generalization capability.