A new input variable selection method for soft sensor based on stacked auto-encoders

Abstract

Soft sensors have been widely used in industrial processes in the past two decades, using easy-to-measure process variables to predict hard-to-measure ones. Input variable selection is a significant step of soft sensor modeling to achieve a desired estimation performance while maintain a relative low complexity. In this paper, a hybrid variable selection method is proposed to select input variables for soft sensor model based on stacked auto-encoders. In each backward iteration, the input variables are ranked by the metric combining mutual information of input variables with prediction error of the soft sensor. Then, the least relevant variables are detected and deleted. To reduce the computational complexity, only the weights of first layer of the soft sensor model is updated after the removal of the irrelevant variables. Two practical industrial experiments are analyzed with implementation of the proposed input variable selection scheme and other four state-of-art variable selection methods are compared. With the proposed algorithm, an improvement of the system performance is demonstrated by the analytical and experimental results, while keep a low complexity.