Robust variational Bayes analysis of linear change-point problem

Abstract

The deterioration of the condition of a physical system that produces output with linear relationship with the input can manifest in the data generated by such system via change-points. As a result, timely detection and analysis of a change-point in such systems form a significant element in providing pragmatic solutions towards the smooth operation of the system. In this regard, the thesis considered novel Variational Bayes methods for modeling, detection, and inference of change-point in linear systems. In particular, Variational Lower Bound Difference(VLBD), Variational Bayes Information Criteria (VBIC), and Variational Akaike Information Criteria (VAIC) ratio- based change-point detectors are developed for a single change-point detection in linear systems. The methods are assessed with linear change-point datasets in both simulation and real data of a refinery process, and their utility is soundly illustrated. Interestingly, the Variational lower bound difference- based detector shows robustness over its VBIC and VAIC counterparts in situations where there exist multiple change-points. This was evidenced by the real-data application.