Graph-based predictable deep transfer network for soft sensing of dynamic industrial processes

Abstract

Due to their advantages in high-level abstract feature extraction, stacked auto-encoders and their supervisory variants have been widely used to develop soft sensors for industrial processes. However, since they fail to provide analytical solutions for the autoregression embedded in networks, the difficulty in modeling the temporal correlation in latent features of supervisory stacked auto-encoders greatly increases. Moreover, the normal assumption for the design of soft sensors of an independent, identical distribution does not hold when there are changes in the training and testing data. Thus, a graph-based, predictable, deep transfer network (GPDTN) for soft sensing of dynamic industrial processes is proposed. To effectively learn the dynamic information of past data, a new loss function is proposed to reconstruct the data and simultaneously learn the predictable latent space using joint mutual information and graph embedding. Then, using deep domain adaptation, a new regular term of dynamic alignment is added to narrow the differences of the predictive information between source and target domains, enabling the graph-based predictable structure to be adaptable to concept drift in dynamic processes. Finally, the performance and effectiveness of the GPDTN-based soft sensors are demonstrated through experimental results on the industrial debutanizer and sulfur recovery unit.