A New Variable Selection Method for Soft Sensor Based on Deep Learning

Abstract

In recent years, soft sensing technique has been widely used in industry and other process control. Utilizing the easy-to-measure process variables directly related to target variables, a mathematical model called soft sensor is established to predict the value of the target variable. In the modeling process, the model performance is highly dependent on the selection of process variables. Selecting the process variables that are most related to target variable but lowest in quantity can effectively simplify the model structure, reduce the computational burden, and increase the convergence speed of the model. In this paper, a soft sensor is firstly built by stacking auto-encoders in the way of deep learning. Then, a new embedded and automatic variable selection method is proposed to rule out the process variables least related to the target variable and contributing least to the model performance. In addition, a new weight updating method for deep network is provided to reduce the computational burden. The result of case study reveals that compared with other common methods, the method proposed in this paper selects fewer process variables without damaging the prediction accuracy of soft sensing prediction.