To handle characteristics of nonlinearity

time-variation and multi-product of chemical processes

a self-adaptive soft sensing method was developed under ensemble learning framework. Initially

a self-adaptive localization technique was proposed to construct ensemble of high diversified local models by statistical hypothesis testing theory and k-nearest neighbor method. Subsequently

based on generalization capabilities of quantified local models with online query sample

primary process variables were estimated through selective ensemble learning. Furthermore

in order to measure estimation accuracy of primary process variables

a highly universal method of model performance assessment was presented by using local model's generalization error. Simulation study on a penicillin fermentation process demonstrated effectiveness of the proposed method.