A physics-of-failure-based steady-state transfer regression modeling for electronic components under incomplete data conditions

Degradation modeling of target-specification electronic components is often limited by test cost and duration, leading to scarce labeled target samples. In practical condition monitoring, degradation indicators are further affected by missing and noisy sensor measurements caused by sensing, acquisition, and transmission, which amplifies the identification uncertainty of physics-of-failure (PoF) coefficients and degrades the credibility of lifetime prediction. To enable robust inference under small-sample and sensor-incomplete conditions, we propose a steady-state feature transfer regression framework that leverages manufacturing features as transferable carriers for sensor-driven degradation estimation. First, a conditional mapping from manufacturing parameters and operating conditions to PoF coefficients is learned in a joint space of multiple candidate source domains and the target domain. A sample-based predictive entropy criterion is then introduced to quantify source-domain applicability from stochastic PoF-coefficient samples under target-domain inputs, enabling reproducible source-domain selection. Next, a physics-consistency-embedded steady-state transfer regression model is constructed with structural-shift regularization and perturbation-suppression terms to enhance robustness to inter-domain discrepancies and measurement perturbations. For severe target data scarcity, a collaborative training strategy couples trend-based pseudo labels with posterior uncertainty estimates to constrain boundary-region learning and improve extrapolation reliability. Experiments on an electromagnetic-relay degradation dataset show that the proposed method achieves RMSE = 0.032 and R² = 0.92 on the target test set. After variance-scaling calibration, the nominal 95% prediction interval achieves PICP(95%) = 0.89 with mean prediction-interval width (95%) = 0.085, indicating improved empirical coverage while maintaining a relatively narrow interval width. These results demonstrate the effectiveness of the proposed framework for PoF-informed degradation estimation and uncertainty-aware decision support under small-sample target-domain conditions.