Quality prediction via semisupervised Bayesian regression with application to propylene polymerization

Statistical learning techniques are widely used for quality prediction in polymerization processes during the last decades. However, compared to operation variables, quality variables of polypropylene process are usually difficult to acquire resulting from the absence of measuring units. A semisupervised Bayesian regression method is therefore presented to improve the prediction accuracy by sufficient usage of unlabeled sampling data for melt index prediction in polypropylene processes. The developed model consists of Bayesian inference to predict quality variables and neighborhood kernel density estimation for finding relationships between unlabeled data and labeled samples, which takes advantages of a probability framework with iterative maximum likelihood techniques for parameter estimation and a sparse constraint for avoiding overfitting. The quality prediction regression method is compared with published models by applying to a real dataset of industrial propylene polymerization. The experiment results demonstrate the effectiveness of the proposed semisupervised Bayesian method. A semisupervised Bayesian inference framework is presented for predicting the quality variables in polypropylene processes. Relevance vector regression that is developed based on Bayesian inference is used to predict the quality variable via operation variables, while the neighborhood kernel density estimation is applied to produce the relationship between the unlabeled data and labeled samples. The proposed semisupervised learning method takes advantage of gathering information of unlabeled data to improve the prediction accuracy.