A metabolism support vector machine (SVM) based modeling method was proposed to solve the problems that slowed down computing speed and caused poor stability during SVM incremental learning. Firstly

vector projection algorithm was used to pre-extract training samples

in order to reduce the number of samples and improve SVM modeling speed. Secondly

a new sample "metabolism"principle was pulled into the SVM incremental learning process

for addressing "explosion" of training samples number which was caused by adding new samples continuously. Finally

the vector projection metabolism SVM was utilized in ethylene distillation product quality soft-sensor modeling. The experiment results showed that vector projection metabolism SVM had better prediction result than SVM and LSSVM.