A Robust Infinite Gaussian Mixture Model and Its Application in Fault Detection of Nonlinear Multimode Processes

作     者：Pan, Yi Xie, Lei Su, Hongye Luo, Lin 

作者机构：Zhejiang Univ Inst Cyber Syst & Control State Key Lab Ind Control Technol Hangzhou 310027 Zhejiang Peoples R China Liaoning Shihua Univ Sch Informat & Control Engn 1 West Dandong Rd Fushun City Liaoning Peoples R China 

出 版 物：《JOURNAL OF CHEMICAL ENGINEERING OF JAPAN》 (日本化工杂志)

年 卷 期：2020年第53卷第12期

页      面：758-770页

核心收录：

学科分类：0817[工学-化学工程与技术] 08[工学] 

基　　金：National Key R&D Program of China [2018YFB1701102] National Nature Science Foundation of China 

主　　题：Multimode Process Modeling Infinite Gaussian Mixture Model Outliers Nonparametric Bayesian Methods Markov chain Monte Carlo Inference 

摘      要：Finite Gaussian mixture model (GMM) has recently proven to be a powerful unsupervised treatment for monitoring nonlinear processes with multiple operating conditions. The performance of GMM-based monitoring method largely depends on the number of mixture densities. However, the popular penalty method, such as Bayesian information criterion (BIC) and Akaike s information criterion (AIC), usually tend to yield noisy model size estimates. Moreover, the parameter estimates in GMM are susceptible to outliers. To overcome these deficiencies, this paper proposes a new process monitoring technique based on a robust infinite Gaussian mixture model (Ro-IGMM). Specifically, a separate weight at each point is assigned to the precisions as a measure of smoothness, representing the similarities to other data points. The Chinese restaurant process is then placed on a prior to turn into infinite groupings. The informations, such as a distribution over the number of clusters, the cluster assignments, and the parameters associated with each cluster, can be given by the posterior which is obtained by a collapse Markov chain Monte Carlo (MCMC) inference. Simulation results on the benchmark Tennessee Eastman process show that Ro-IGMM-based process monitoring method is more insensitive to outliers during process modeling, compared to traditional methods working with BIC model selection.