TC-BiKAN: A domain-adaptive framework with Kolmogorov-Arnold networks for rolling bearing RUL prediction

Remaining useful life (RUL) prediction for rolling bearings is critical for maintaining the reliability and safety of rotating machinery. Under complex operating conditions, cross-domain distribution drift and nonlinear degradation dynamics pose considerable challenges for existing methods in achieving both reliable degradation stage division and accurate RUL regression. To address these issues, this paper proposes an integrated online prediction framework based on transfer clustering and Kolmogorov-Arnold Networks (KAN), termed TC-BiKAN. For first prediction time identification, a domain-adaptive fuzzy clustering algorithm with boundary identification (DAFC-BI) is developed to attenuate domain shift and quantify boundary ambiguity through membership differences, thereby enabling robust detection of degradation onset across varying operating conditions. To reduce supervision bias arising from stage misalignment in cross-domain scenarios, a trend-adaptive multistage labeling (TAML) strategy is introduced to reconstruct stage-wise RUL labels by estimating physical degradation rates. To enhance nonlinear regression during accelerated late-stage degradation, a BiKAN predictor is proposed that replaces the conventional fully connected regression head of bidirectional long short-term memory networks with KAN-based learnable spline functions. Experiments on the PHM2012 and XJTU-SY datasets demonstrate that TC-BiKAN achieves competitive prediction accuracy while providing, on average, more conservative RUL estimates than representative baselines, thereby supporting practical deployment for health management under non-stationary operating conditions.