Bayesian Nonparametric Modeling of Markov Chains for Detection of Thermoacoustic Instabilities

This paper proposes a Bayesian nonparametric method for detecting thermoacoustic instabilities in gas turbine engines in real-time, where the underlying algorithms are formulated in the symbolic domain and the resulting patterns are constructed from symbolized pressure measurements as probabilistic finite state automata (PFSA) that is built upon a finite-memory Markov model, called D-Markov machine. The Bayesian nonparametric structure is adopted for: (i) automated selection of parameters in the D-Markov machine, and (ii) online sequential testing, to provide a data-driven and coherent statistical analysis of combustion instability phenomena without relying on numerically intensive models of combustion dynamics. The proposed method has been experimentally validated on the time series generated from a laboratory-scale combustion apparatus. The results of instability prediction, derived from the time series, have been compared with those of other existing techniques.