The time-varying and multi-dimensional characteristics are major causes of the low performance of soft sensors in chemical processes. To solve the problem, an improved adaptive soft sensor modeling method is proposed. This method obtains predicted deviation by modular steps of moving window and evaluates deterioration of soft sensors via ttest adaptively. Besides, this paper combines the moving window-autoassociative neural network (AANN) method to update both the modeling auxiliary variable and the auxiliary variable data. Data simulation and result analysis obtained via a continuous stirred tank reactor (CSTR) and a debutanizer column process (DCP) show that the improved adaptive soft sensor modeling method proposed in this paper can evaluate the deterioration of soft sensors and update the soft sensor model adaptively, and improve the predicted performance of soft sensors for time-varying and multi-dimensional chemical processes.