Augmented Multidimensional Convolutional Neural Network for Industrial Soft Sensing

Abstract

In the era of industrial big data, data-driven soft-sensor models have become an important method to guide production and optimize control. However, due to the limitation of data acquisition methods, the industrial data obtained in practice have unavoidable defects that affect the performance of soft sensors. Aiming at the unbalanced sampling, inaccurate matching, and partial missing of industrial process data, this article proposes an augmented multidimensional convolutional neural network (CNN) for industrial soft sensing. For complete process data information, we stitch the fine-grained data to obtain coarse-grained data and then use a CNN to extract deep features. On the basis of analyzing the physical meaning of data, the structure of multidimensional convolution is designed to focus on different details of process information. In this framework, the problem of partial missing data is emphasized. Then, mix-grained data augmentation strategies are invented to solve this problem and improve the performance of multidimensional CNN (MDCNN) soft sensors. The proposed augmented MDCNN is applied to the f-CaO prediction of the cement clinker production process, results of which show superiority compared to existing methods.