Deep Regression‐Based Encoder‐Decoder: A Novel Deep Learning Framework for Industrial Soft Sensor Application

ABSTRACT

A data‐driven soft sensor is indeed an empirical model which predicts the hard‐to‐measure quality variable from available easy‐to‐measure process variables. From the data‐driven perspective, the quality data inherently consists of both linear and nonlinear predictable information. Therefore, to exploit linear and nonlinear information useful for quality prediction, a novel module called regression‐based encoder‐decoder (RED) is first designed for deep learning. The RED utilizes a partial least square regression (PLSR) model to extract latent features highly correlated with the quality variable, and subsequently employs the encoder‐decoder framework to extract nonlinear hidden features that helpful for further reducing the prediction error. Moreover, in addition to the commonly used stacked structure of multiple RED modules, a regression layer with parameters pretrained by the least square regression (LSR) algorithm is proposed for fusing the multiple predictions generated through layer‐by‐layer propagation. The novel deep learning architecture proposed for soft sensor modeling is thus termed as deep RED (DRED), and its outstanding soft sensing performance is demonstrated through comparative experiments carried out on two real‐world industrial processes.