Time-varying trajectory modeling via dynamic governing network for remaining useful life prediction. (1st January 2023)

• Record Type: Journal Article
• Authors: Zhou, Zheng; Li, Tianfu; Zhao, Zhibin; Sun, Chuang; Chen, Xuefeng; Yan, Ruqiang; Jia, Jide
• Is Part Of: Mechanical systems and signal processing. Volume 182(2023)
• Publication Date: 2023-01-01
• Subjects: Remaining useful life prediction -- Trajectory governing -- Ordinary differential equation -- Time-varying -- Deep reinforcement learning; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2022.109610
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: A novel time-varying trajectory modeling problem via physics analysis is developed for remaining useful life prediction. Neural ordinary differential equation guides the dynamic governing network of prediction trajectory. A nonnegative bounded function shrinks the disordered trajectory space to conform physical intuition. A dynamic learning scheme based on reinforcement learning endows the governing network with timing matching capability. Abstract: Remaining useful life (RUL) prediction is highly demanded in modern industry as it provides a scheduling basis for predictive maintenance. Recently, intelligent data-driven methods have been developed for RUL prediction due to their powerful performance. However, existing methods mostly ignore physical dynamics behind the RUL prediction problem, which predict a fluctuated RUL trajectory contrary to the physical intuition. To address this issue, we formulize RUL prediction as a time-varying trajectory modeling problem by analyzing the difference between stochastic degradation process and smooth RUL trajectory, and propose a dynamic governing network (DGN) to identify the RUL trajectory from life-span observation series. Specifically, a discretized ordinary differential equation (ODE) parameterized by neural networks is utilized to describe a governing equation of the RUL trajectory. To constraint the trajectory space, a nonnegative bounded function is inserted into each time step of the forward propagation of the ODE. To identify time-varying coefficients in the DGN, the ODE network is specified as a super-network with time-invariant parameters and a time-varying network architecture, which is dynamically determined by a deep reinforcement learning algorithm. Experimental results on two datasets demonstrate that the proposed DGN can capture underlying dynamics from observation series and can obtain state-of-the-art RUL prediction performance.