SKDAN: A Signal Knowledge-enhanced Domain Adaptation Network for remaining useful life prediction and uncertainty quantification of rolling bearings

Domain adaptation-based methods are extensively applied to predict the Remaining Useful Life (RUL) of rolling bearings under complex operating conditions. However, the nonlinear degradation process of bearings gives rise to markedly non-stationary characteristics in vibration signals throughout the full life cycle. Although significant differences in fault features arise across different degradation stages, clearly identifying the critical degradation information remains a challenge. In this paper, a Signal Knowledge-enhanced Domain Adaptation Network (SKDAN) is proposed to learn domain-invariant features from non-stationary degradation processes, thereby improving cross-domain RUL prediction. Specifically, an adaptive short-time Fourier transform layer with a variable window is introduced to analyze the raw vibration signals in the time domain. This differentiable layer extracts time–frequency physical information with high energy concentration, which enhances the representation of degradation features. Subsequently, a novel discrepancy metric, termed Multi-Stage Maximum Mean Discrepancy (MSMMD), is proposed to replace the global average discrepancy with multiple local discrepancies. The MSMMD metric effectively increases the inter-class distance between cluster centers, which enables cross-domain feature alignment. Finally, an uncertainty measurement mechanism is constructed via a step-by-step training strategy, with the objective of quantifying the uncertainty in RUL results by calculating confidence intervals for prediction points. Comparative tests with other methods are conducted on two different bearing datasets, and the results demonstrate that SKDAN achieves superior performance and reliability in cross-domain RUL prediction. • An adaptive short-time Fourier transform layer enhances time–frequency feature representation. • A new discrepancy metric criterion is developed for domain adaptation. • The prediction uncertainty is quantified via a step-by-step training strategy. • A cross-domain prediction model is proposed for rolling bearings. • The effectiveness of the proposed method is verified by two different bearing datasets.