Inferential Estimation of Polymer Melt Index Using Deep Belief Networks

Abstract

This paper presents using deep belief networks for the inferential estimation of polypropylene melt index in an industrial polymerization process. The polymer melt index is difficult to be measured online in practice. The relationship between easy-to-measure process variables and difficult-to-measure polymer melt index is found by using a deep belief network model. The development of a deep belief network model contains an unsupervised training process and a supervised training process. Deep belief networks use a novel semi-supervised learning method. The process operational data without corresponding quality measurements can be used in the unsupervised training process. The profuse information behind input data are captured by deep belief networks. It is shown that the deep belief network model gives very accurate estimation of melt index.