Considering the performance of a predictive model is heavily depended on its training samples

a new on-line adaptive modeling approach based on incremental support vector regression (SVR) is presented.When a new sample arrives

it is firstly checked by the Karush-Kuhn-Tucker(KKT) condition of established model

only those which contain sufficient new information can be introduced into the training sample set.In this way

the model generalization ability will be maintained while the update frequency can be reduced.If the new sample cannot be described by the established model and

therefore

has a large prediction error

the model must be updated and the useless sample should be deleted from the model

to adapt the process characteristics.In this case

the useless sample

while not the oldest one

is selectively deleted from the model based on the similarity between samples.The proposed method is illustrated through the application to an industrial polypropylene unit to predict its melt index.The results show that

compared with other methods

the proposed method exhibits good generalization ability while the update frequency significantly lower

and therefore trace the process characteristics effectively.