Domain weighted distribution adaptation network: a novel remaining useful life prediction framework for machinery targeting time-varying operation conditions

Machinery typically operates with condition alternations throughout the degradation process. Commonly used deep learning-based methods for remaining useful life (RUL) prediction primarily focus on degradation under constant operation conditions, including studies that use domain adaptation and similar technologies to predict RULs across different but fixed conditions. However, the lack of training data under time-varying conditions limits the RUL estimation in condition alternation scenarios. To address this issue, this paper proposes a prediction framework targeting time-varying operation conditions, termed the domain weighted distribution adaptation network (DWDAN). It utilizes run-to-failure datasets from constant conditions to predict RULs under time-varying conditions, bridging the gap of prediction between these two scenarios. In the framework, discrepancies in feature distributions are attributed to degradation, operation conditions, and their nonlinear coupling. First, an RUL prediction model is developed using constant condition samples to capture degradation-related distributions. Then, domain weights for different conditions are optimized with normal stage samples from time-varying conditions. Finally, the feature distributions are adjusted via optimal transport (OT) to overcome the nonlinear coupling. The proposed method is validated on experimental run-to-failure datasets of gearboxes. The results demonstrate the superiority of the DWDAN in overcoming the impact of condition alternations and improving prediction performance.