Optimal online soft sensor for product quality monitoring in propylene polymerization process

In the real-time propylene polymerization manufacturing process, melt index (MI), as the key product quality variable, is hard to be measured on-line, which brings difficulties to the control and optimization of this process. However, a large amount of data of other relative process variables in this process can be routinely recorded online by the distributed control system (DCS). An optimal soft-sensor of least squares support vector machine (LS-SVM) is therefore proposed to implement the on-line estimation of MI with the above real-time DCS records, where LS-SVM is employed for developing a data-driven model of the above industry process. In view of that the input variable selection and parameter setting are crucial for the learning results and generalization ability of LS-SVM, the nonlinear isometric feature mapping technique and particle swarm optimization algorithm are then structurally integrated into the model to search the optimal values of those parameters. Considering the process time-varying nature, an online correction strategy is further switched on to update the modeling data and revise the model configuration parameters via adaptive behavior. Finally, the explored soft sensor model is illustrated with a real plant of propylene polymerization, and the results show the predictive accuracy and validity of the proposed systematic approach.