【报告摘要】Multistate models are effective for monitoring, prediction, and optimal decision-making based on sequentially observed data across different states or stages. Traditionally, multistate models such as Hidden Markov Model and Markov Decision Process assumes that observations at each state follow a stationary distribution. However, with the growing technological and sensing capabilities for real-time condition monitoring (CM) of industrial equipment, the relationship between the underlying system states and the observations is often time-variant in many applications. This time-dependent relationship in multistate models has not been sufficiently addressed in the quality and reliability engineering literature. This talk presents several problems we have studied over recent years to address this shortcoming by modeling observed CM signals and underlying system states, where the linkage is time-dependent and piecewise linear with jumps. Specifically, I will highlight the following research: (1) A robust and computationally efficient method using multiple change-point models for online steady-state detection, which is critical for system/process performance assessment, optimization, fault detection, and process automation and control; (2) a Bayesian self-starting monitoring scheme for a process subject to various types of random drift and jumps; (3) A prognostic framework that jointly models CM signals and failure event data based on a Hidden Markov model, where the CM signals are linked to the hidden states through a time-dependent functional relationship; and (4) An optimal maintenance framework based on a Partially Observable Markov Decision Process (POMDP) with time-dependent observations at a given state. We will also briefly discuss ongoing and potential future works and extensions in this research direction.