Because batch processes have strong non-linearity

multi-stage

slow time-evolution

and batch-to-batch variation

conventional single prediction model cannot effectively capture characteristics of multi-stage and inter-stage transition. A novel multi-model soft sensor method was proposed on the basis of Gath-Geva clustering and kernel extreme learning machine (KELM). First

principal component analysis (PCA) was used to extract features of input variables. Then

Gath-Geva algorithm was used to classify different operating stages of the batch process and local KELM model was built for each operating stage. For a query sample

every local KELM predictions were calculated and final predictions were obtained by integrating fuzzy membership of each local KELM as weight and its corresponding prediction value. The numeric simulation results on data of penicillin fermentation show that this multi-model approach has more accurate prediction than single model.