Research On Probabilistic Principal Component Analysis In Multimode Process Monitoring

Because of the features and needs of modern industry, the process always has many steady modes, not only one mode. If traditional multivariate statistics methods are used in process monitoring directly, it may cause high missing rate. As a widely-used method of process monitoring, multimodel method is easy to understand, and it can build models to fit every modes to detect faults of the multimode process.This dissertation is based on PCA method,we apply MPPCA model in TE process and CSTR process. The details of this dissertation are:(1) Introduce PCA method and PCA-based monitoring method in details, and study the knowledge of PPCA model, then analyze the estimate principle of EM algorithm. And to fit multimode process, use a method to expend PPCA model to MPPCA model. At last, give statistics of monitoring.(2) Based on EM algorithm, variational-Bayes is introduced to expend EM algorithm to VBEM algorithm. This algorithm can take advantage of prior knowledge. Parameters are considered as random variables, we use super parameters to estimate posterior probability of parameters which is based on mean field theory. At last, we apply VBEM algorithm to a set of data to prove it useful to build multiple models.(3) We apply EM algorithm and VBEM algorithm to TE process and CSTR process,then use Bayesian inference to integrate monitoring statistics. The results shows that:VBEM algorithm is more effective then EM algorithm in multimode process monitoring.