Date of Award
Open Access Dissertation
College of Engineering and Computing
Fault diagnosis and prognosis (FDP) plays an important role in the modern complex industrial systems to maintain their reliability, safety, and availability. Diagnosis aims to monitor the fault state of the component or the system in real-time. Prognosis refers to the generation of long-term predictions that describe the evolution of a fault and the estimation of the remaining useful life (RUL) of a failing component or subsystem.
Traditional Riemann sampling-based FDP (RS-FDP) takes samples and executes algorithms in periodic time intervals and, in most cases, requires significant computational resources. This makes it difficult or even impossible to implement RS-FDP algorithms on hardware with very limited computational capabilities, such as embedded systems that are widely used in industries.
To overcome this bottleneck, this proposal develops a novel Lebesgue sampling-based FDP (LS-FDP), in which FDP algorithms are implemented “as-neede”. Different from RS-FDP, LS-FDP divides the state axis by a number of predefined states (also called Lebesgue states). The computation of LS-based diagnosis is triggered only when the value of measurements changes from one Lebesgue state to another, or “event-triggered”. This method significantly reduces the computation demands by eliminating unnecessary computation. This LS-FDP design is generic and able to accommodate different algorithms, such as Kalman filter and its variations, particle filter, relevant vector machine, etc.
This proposal first develops a particle filtering based LS-FDP for li-ion battery applications. To improve the accuracy and precision of the diagnosis and prognosis results, the parameters in the models are treated as time-varying ones and adjusted online by a recursive least square (RLS) method to accommodate the changing of dynamics, operation condition, and environment in the real cases. Uncertainty management is studied in LS-FDP to handle the uncertainties from inaccurate model structure and parameter, measurement noise, process noise, and unknown future loading.
The extended Kalman filter implemented in the framework of LS-FDP yields a more efficient LS-EKF algorithm. The proposed method takes full advantage of EKF and Lebesgue sampling to alleviate computation requirements and make it possible to be deployed on most of the distributed FDP systems.
All the proposed methods are verified by a study with the estimation of the state of health and RUL prediction of Lithium-ion batteries. The comparisons between traditional RS-FDP methods and LS-FDP show that LS-FDP has a much lower requirement on the computational resource. The proposed parameter adaptation and uncertainty management methods can produce more accurate and precise diagnostic and prognostic results. This research opens a new chapter for FDP method and make it easier to deploy FDP algorithms on the complicate systems build by embedded subsystem and micro-controllers with limited computational resources and communication band width.
Yan, W.(2017). A Lebesgue Sampling based Diagnosis and Prognosis Methodology with Application to Lithium-ion Batteries. (Doctoral dissertation). Retrieved from https://scholarcommons.sc.edu/etd/4325