Koopman-Informed Neural Network for Machinery's Nonlinear Dynamics Learning and Remaining Useful Life Prediction

In industrial field, prognostics and health management (PHM) techniques, particularly remaining useful life (RUL) prediction, are pivotal in guiding mechanical maintenance decisions, enhancing system safety, and reducing operational costs. Prevalent mechanical life prediction techniques encounter challenges arising from the lack of model transparency and limited interpretability, which can undermine confidence in the reliability and accuracy of forecast outcomes. To provide a fresh perspective for comprehending the behavior of complex systems and enhancing the robustness of prediction accuracy, a novel approach called Koopman-informed neural network (KINN) is proposed. This innovative approach adeptly captures complex data patterns and relationships via nonlinear transformations by amalgamating Koopman theory with neural networks. Unlike traditional approaches that rely on linear approximations or local analyses, the Koopman operator framework enables a global linear representation of a system's dynamics by lifting the state space to a higher dimensional eigenfunction space. This method employs both forward and backward Koopman dynamic operators to capture and reconstruct the system's temporal dynamics, with physical constraints helping to regularize system behaviors and ensure stability. Experimental assessments were conducted using both a synthetic N-CMAPSS dataset and a real-world bearing full-life dataset, revealing the robustness and generalization capabilities of the proposed methodology, leading to improvements in the accuracy of RUL prediction.