Multi-Model Fusion Modeling Method And Its Application In Soft Sensor

Considering that key quality index of chemical process is difficult to measure online, soft sensor technology is proposed to solve this problem. The researches on soft sensor technology nowadays mainly focus on the modeling method. Because it is difficult to establish accurately soft sensor model of chemical process by single model, a multi-model fusion modeling method is proposed. Stacked neural networks (SNNs) is developed by reasonable fusion multiple single neural network. Considering that there is strong nonlinear correlation between each variables of chemical process, kernel principal component analysis (KPCA) is applied to compress the input data of model and select the nonlinear principle component. The modeling method based on KPCA-SNNs is proposed and applied to soft sensor modeling of polypropylene melt index.The specific research content is as follows:(1) Considering that there are many factors which affect chemical process and nonlinear correlation between each variable, if auxiliary variable is served as model input directly, prediction accuracy is influenced and model calculated quantity is increased greatly. Because traditional principal component analysis is not suitable for nonlinear chemical process, KPCA is applied to compress the nonlinear input data of soft sensor model by using strong nonlinear feature extraction ability of KPCA.(2) Because prediction accuracy and robustness of single model is bad, the modeling method based on SNNs is proposed by reasonable fusion multiple BP neural networks. Proper determination of the stacking weights is essential for good model performance, ridge regression and minimizing the maximum absolute error about determination method of appropriate weights is studied. KPCA is used to compress the input data of SNNs, and then the results will be used as model inputs, the soft sensor modeling method based on KPCA-SNNs is established.(3) The modeling method based on KPCA-SNNs is applied to soft sensor modeling of polypropylene melt index. Firstly, auxiliary variable is selected according to propylene polymerization mechanism. Secondly, sample data is collected and pretreated, including random error processing, abnormal data processing, normalization and KPCA, etc. According to pretreated training data, polypropylene melt index soft sensor model based on KPCA-SNNs is developed. Simulation results indicate that the melt index soft sensor model has a better robustness and prediction accuracy than traditional SNNs, and calculating speed is improved. Finally, the process characteristics analysis is carried out by using the polypropylene melt index soft sensor model. The influence of each process variables about polypropylene melt index is studied.Multi-model fusion modeling method based on KPCA-SNNs is proposed and applied to develop polypropylene melt index soft sensor model. The research results indicated that it is not only provides a good performance modeling method for soft sensor modeling, but also solves the problem of chemical process important quality index can not measure online. The research will be benefit to promoting the application of soft sensor technology in chemical process and is of great practical and theoretical value.