Expectation Maximization method for multivariate change point detection in presence of unknown and changing covariance

• Published in: Computers & chemical engineering Vol. 69; pp. 128 - 146
• Main Authors: Keshavarz, Marziyeh, Huang, Biao
• Format: Journal Article
• Language: English
• Published: Kidlington Elsevier Ltd 03.10.2014; Elsevier
• Subjects: Algorithms; Applied sciences; Bayesian inference; Change point detection; Chemical engineering; Computer science; control theory; systems; Computer simulation; Control system analysis; Control theory. Systems; Covariance; Data processing; Exact sciences and technology; Expectation Maximization; Mathematics; Maximization; Monitoring; Optimization; Parametric inference; Probability and statistics; Sciences and techniques of general use; Statistics; Bayesian inference; Change point detection; Expectation Maximization; Bayes estimation; Data analysis; Event detection; Time series; Change point; Experimental study; Modeling; Optimization; Data acquisition system; Covariance; regime switching model; EM algorithm; System monitoring
• ISSN: 0098-1354; 1873-4375
• DOI: 10.1016/j.compchemeng.2014.06.016

•Developed simultaneous detection of multiple change points.•Handled change point detection in presence of unknown and varying covariance.•Solved change detection problem under Expectation Maximization framework. Data analysis plays an important role in system modeling, monitoring and optimization. Among those data analysis techniques, change point detection has been widely applied in various areas including chemical process, climate monitoring, examination of gene expressions and quality control in the manufacturing industry, etc. In this paper, an Expectation Maximization (EM) algorithm is proposed to detect the time instants at which data properties are subject to change. The problem is solved in the presence of unknown and changing mean and covariance in process data. Performance of the proposed algorithm is evaluated through simulated and experimental study. The results demonstrate satisfactory detection of single and multiple changes using EM approach.