Quickest change detection in nonlinear hidden Markov models using a generalized CUSUM procedure with particle filters

Quickest change detection is a vital procedure of system monitoring that involves optimizing the tradeoff between detection delay and frequency of false alarms. From a filtering perspective, we study an effective method of tracking the generalized cumulative sum (CUSUM) statistics in practice to detect a change in the hidden Markov models (HMMs) with nonlinear dynamics. This method has been designed to enhance the accuracy of change detection in HMMs, which is a critical task in a variety of academic and industrial fields. Our research centers specifically on instances where the change excites a dynamical system from a pre-existing steady state to a new steady state or periodic oscillations in the post-change regime. The performance of the generalized CUSUMs using particle filters is examined with random simulations in the pure surge modes of Moore–Greitzer compressor models with external forcing. For various forms of the post-change limit cycles with different phases, we observe some similarities and differences between the generalized CUSUMs and a special CUSUM statistic that assumes the change occurs immediately.