Window Test-Time Adaptation for Lifetime Prediction Considering Non-Stationary Noise

Abstract

Accurate lifetime prediction is critical for prognostics and health management (PHM) of safety-critical industrial equipment such as rolling bearings. However, degradation monitoring signals collected from run-to-failure processes often suffer from continuous distribution drift under complex and nonstationary operating conditions, which significantly undermines the robustness and deployment reliability of data-driven lifetime prognostics models. To this end, we propose Window Test-Time Adaptation (WTTA), a time-series forecasting framework for robust lifetime prediction that improves generalization under nonstationary and condition-varying environments. WTTA quantifies operational condition shifts through a sliding-window mean absolute error metric, and introduces a dynamic adaptation strategy that automatically adjusts the update step size to avoid overand under-adaptation during online inference. Furthermore, we develop a weighted prediction fusion mechanism to integrate original and adapted model outputs, effectively mitigating prediction deviations caused by noise and short-term fluctuations along degradation trajectories. Extensive experiments on public run-to-failure benchmarks, including rolling bearing degradation datasets and the NASA CMAPSS turbofan engine dataset, demonstrate that WTTA consistently improves both prediction accuracy and stability across baseline models. This indicates that WTTA provides a reliable and practical solution for industrial lifetime prognostics and predictive maintenance applications.