A dynamic CNN for nonlinear dynamic feature learning in soft sensor modeling of industrial process data

Abstract

Hierarchical local nonlinear dynamic feature learning is of great importance for soft sensor modeling in process industry. Convolutional neural network (CNN) is an excellent local feature extractor that is suitable for process data representation. In this paper, a dynamic CNN (DCNN) strategy is designed to learn hierarchical local nonlinear dynamic features for soft sensor modeling. In DCNN, each 1D process sample is dynamically augmented into 2D data sample with lagged unlabeled process variables, which contains both spatial cross-correlations and temporal auto-correlations. Then, the convolutional and pooling layers are alternately utilized to extract the local nonlinear spatial–temporal feature from the 2D sample data matrix. Moreover, the principle is analyzed for DCNN on how it can learn the local nonlinear spatial–temporal feature from the network. The effectiveness of proposed DCNN is verified on an industrial hydrocracking process.