Soft Sensor For Polypropylene Plant Product Quality

Melt index is one of the most important quality variables in polypropylene plant. Studies on soft sensor of melt index have important theoretical and practical value on optimization and control of polymerization process. Aiming at many problems in polypropylene melt index soft sensor, this paper proposed many soft sensor modeling methods from the studies of mechanism models and data-based models and discussed the procedure and results of implementing soft sensor on industrial polypropylene plant.For spheripol II polypropylene unit,its reaction principle was analyzed and mechanism models of loop reactors were built. Then the effects of operating modes on plant variables were analyzed. Based on reactor mechanism models, two melt index prediction models on dual loop reactor were developed: logarithm model and power model. The recursive least square method with forgetting factor (FFRLS) and particle swarm optimization (PSO) algorithm were used to identify unknown model parameters respectively. And the validity of the models was verified through industrial data sets.Considering the difficulty of single mechanism soft sensor in multi-grade operation, two hybrid modeling methods were presented based on mechanism model and fuzzy technique: FP-EFCM and FPFIS. FP-EFCM combined mechanism models and enhanced fuzzy C means clustering and FP-FIS integrated adaptive neuron fuzzy inference system with mechanism models. The application results on industrial data showed these methods could predict the melt index effectively.To avoid the degradation of model precision because of the outliers in industrial data when support vector machine was applied in polypropylene soft sensor, clustering support vector machine (CSVM) was proposed. Clustering analysis on modeling data was performed and then the different kinds of data penalization parameters were weighted. Weighting value brought the information of outliers into soft sensor while reducing the disadvantage influence. The study results on plant data showed that CSVM could built more precise model than SVM.A new method based on adaptive kernel partial least squares (AKPLS) was studied to build melt index soft sensor for nonlinear and multi-grade polypropylene process. Nonlinear PLS method KPLS was applied to fit nonlinear function between melt index and secondary variables. Model parameters were updated online by optimizing soft sensor prediction error. The industrial data analysis indicated AKPLS could predict melt index more accurately than PLS and KPLS.In order to build nonlinear soft sensor with good generalization ability and flexibility in multi-grade modes, a new melt index soft sensor based on improved orthogonal least squares (IOLS) was proposed which applied sparse kernel model as model frame. Orthogonal signal correction (OSC) was applied to preprocess OLS model to reduce the noise information which was uncorrelated with output variables. In order to enhance model generalization and sparseness, parameter local regularization and leave-one-out mean square error were combined in OLS cost function. Furtherlly OLS model prarameter adaptive updating strategy was presented for multi-grade polypropylene operation, which updated the OLS model parameters online. The application results on real industrial process data showed that IOLS could predict the polypropylene melt index more accurately than partial least squares (PLS) and OLS.At last the implementation procedure and some key problems of industrial plant soft sensor were discussed. The industrial results of different soft sensor models were analyzed. Several difficulties were indicated during the industrial installation process of soft sensor.