Markov Modeling via Spectral Analysis: Application to Detecting Combustion Instabilities

Devesh K. Jha, Nurali Virani, Asok Ray

Research output: Chapter in Book/Report/Conference proceedingChapter

Abstract

Effective representation of temporal patterns to infer generative models from measurement data is critical for dynamic data-driven application systems (DDDAS). Markov models are often used to capture temporal patterns in sequential data for statistical learning applications. This chapter presents a methodology for reduced-order Markov modeling of time-series data based on spectral properties of stochastic matrix and clustering of directed graphs. Instead of the common Hidden Markov model (HMM)-inspired techniques, a symbolic dynamics-based approach is used to infer an approximate generative Markov model for the data. The time-series data is first symbolized by partitioning the continuous domain to obtain a discrete-valued signal. The size of temporal memory of the discretized symbol sequence is then estimated using spectral properties of the stochastic matrix created from the symbol sequence for a first-order Markov model of the symbol sequence. Then, a graphical method is used to cluster the states of the corresponding high-order Markov model to infer a reduced-size Markov model with a non-deterministic algebraic structure. A Bayesian inference rule captures the parameters of the reduced-size Markov model from the original model. The proposed idea is illustrated by creating Markov models for pressure time-series data from a swirl stabilized combustor where some controlled protocols are used to induce instability. Results demonstrate complexity modeling of the underlying Markov model as the system operating condition changes from stable to unstable which is useful in combustion applications such as detection and control of thermo-acoustic instabilities.

Original languageEnglish (US)
Title of host publicationHandbook of Dynamic Data Driven Applications Systems
Subtitle of host publicationVolume 1: Second Edition
PublisherSpringer International Publishing
Pages129-145
Number of pages17
Volume1
ISBN (Electronic)9783030745684
ISBN (Print)9783030745677
DOIs
StatePublished - Jan 1 2022

All Science Journal Classification (ASJC) codes

  • General Computer Science

Cite this