Date of Award


Degree Name

Doctor of Philosophy


Electrical Engineering


Tzu-Liang (Bill) Tseng

Second Advisor

Jianguo Wu


To fulfill the increasing demand on functionality and quality, modern engineering systems are usually built with overwhelming complexities. The more complex functions the system has been built, the higher reliability required of the system. This is mainly due to the fact that a single failure can result in catastrophic consequences. Therefore, methods that can predict and prevent such catastrophes have long been explored. Prognostics refer to the process of evaluating the current health of a system or a component and then predicting the remaining useful life (RUL) based on the information collected through condition monitoring. The fast development of information and sensing technologies offer great opportunities for real-time health condition monitoring, ensuring the safety, availability, and efficiency of various engineering systems. The condition monitoring signals, collected from sensors, also called degradation signals, are commonly used for system reliability assessment due to their direct relation with underlying physical degradation processes. The commonly applied statistical approach for RUL prediction is to fit degradation signals using parametric regression models to describe and predict how the currently available degradation signal evolves. However, these parametric models are often too rigid and not adequate or flexible enough to model the real degradation signals during the whole life cycle. In fact, degradation signals often show multiple phases in many applications, where the conventional parametric degradation models are often inadequate.

Motivated by the issues, a novel Bayesian multiple change-point modeling approach to characterize degradation signals for prognostics is proposed. Under the Bayesian framework, two stages are often required for prognostics: the offline modeling of historical degradation signals, and the online Bayesian individual model updating and RUL prediction of a new unit. To characterize the inherent unit-to-unit heterogeneity and make the model more flexible, In this Dissertation, all the model parameters are assumed to be random in the model, including the number of change-points, their locations, and all model parameters of each segment. This assumption brings several challenges on how to effectively apply the multiple change-point model and RUL prediction. To address these challenges, we propose a series of approaches in both offline modeling and online model updating and RUL prediction. The main contributions of this research include: (1) An innovative stratified particle filtering algorithm with partial Gibbs resample-move strategy is developed to improve modeling and prognostics. To improve the prediction accuracy, the priors are specified with a novel stochastic process and the multiple change-point model is formulated to a novel state-space model to facilitate online monitoring and prediction; (2) To reduce the model complexity, an exact Bayesian inference is developed. where the closed form of all posterior distributions can be sequentially obtained at online stage. To further control the computational cost, a fixed-support-size strategy in the online model updating and a partial Monte Carlo strategy in the RUL prediction are proposed; (3) To better capture the temporal uncertainties that are inherent in the degradation process, a multiple change-point Wiener process modeling is proposed. The advantages and effectiveness of the proposed methods have been demonstrated through extensive numerical studies and real-world case studies.




Received from ProQuest

File Size

135 pages

File Format


Rights Holder

Yuxin Wen

Included in

Engineering Commons