Deep Learning of Process Data with Supervised Variational Auto-encoder for Soft Sensor

Abstract

Nonlinearity and uncertainty are two critical characteristics when process data modeling is applied to soft sensor. In this paper, a supervised variational auto-encoder (SVAE) is developed to capture both nonlinear and untcertain feature for regression modeling. SVAE, as a deep generative model, provides a probabilistic framework, based on which the deep nonlinear feature extraction is carried out and a probabilistic respresentation can be obtained. In this way, the probability distribution mapping between process data and key quality variables is learned so that quality prediction can be well achieved. The feasibility of the proposed method is illustrated by a numerical example and an industrial example, and the effectiveness of the proposed model is verified by comparing with the linear model.