Multi-Source Domain Generalization for Machine Remaining Useful Life Prediction via Risk Minimization-Based Test-Time Adaptation

Due to unnecessary access to the target data during training, domain generalization (DG) has received great attention in remaining useful life (RUL) prediction for rotating machines. However, existing methods often fail to estimate the ubiquitous adaptation gap, which intractably minimizes the generalization risk. In this study, a novel multi-source DG method is proposed for cross-domain RUL prediction, which considers adaptation gap and performs test-time adaptation to minimize the risk of generalization. Initially, multi-head domain-specific regressors are pretrained to learn the hypothesis from multi-source domains separately. Afterward, the test-time model selection and ensemble is utilized to collaboratively minimize adaptation gap, wherein two strategies of domain similarity and predictive indicator are presented to dynamically integrate the optimal regressor adapted to target domain. Meanwhile, the multioutputs integrated pseudo-labels are used to retrain and optimize the model. Experimental studies indicate that the proposed approach is promising with a maximum 13.05% improvement on prediction performance.