When soft sensor models were constructed for complicated chemical processes by traditional stochastic gradient Boosting (SGB)

improper selection of shrinkage parameters would reduce convergence rate of the algorithm

engender overfitting

and sometimes make it difficult to obtain a satisfactory generalization performance. In order to solve this problem

a modified SGB ensemble learning soft sensor was proposed

in which Gaussian process regression (GPR) was adopted as base learner and shrinkage parameters were automatically adjusted according to feedback of a weak learner in each iteration such that both estimation accuracy and learning efficiency were improved. Simulation results in an industrial process of bisphenol A production showed that the modified integration algorithm had higher learning efficiency and generalization performance than traditional SGB models.