Data-based soft-sensing for melt index prediction

In practical chemical plants, melt index (MI) can only be measured by off-line analysis which cost 2-4 hours. So it is important to develop online analyzers using process data especially when the mechanism is complex. A novel data based soft-sensing method is proposed for MI estimating in propylene polymerization process. This approach is constructed under just-in-time modeling scheme, in which historical process datasets are searched but only the corresponding data samples relevant to the query data sample are used for local modeling, and partial least square (PLS) model is used as local model. The strategy is applied in propylene polymerization process with Spheripol technology, the results showed that feasible estimation for melt index during changeable process can be obtain by the proposed method and the root mean square error (RMSE) is much less than the PLS method.