In most complex chemical processes

measurements are often collected with noises and some outliers. These contaminated data would have negative effect on the accuracy of data-based process modelling and fault detection. A new robust semi-supervised PLVR model (RSSPLVR) was proposed by consideration of the real measuring environment in chemical processes and extended to a nonlinear model K-RSSPLVR with a kernel methodology. In both RSSPLVR and K-RSSPLVR

a weighted coefficient based on sample similarity among all observations was used as prior checking parameter of probability model to effectively eliminate influence of outliers on modelling. Model parameter training was accomplished by analysis of the weighted dataset with EM algorithm and a fault detection scheme was developed. Finally

TE process simulation demonstrated effectiveness of the proposed modelling methods.