Deep relevant representation learning for soft sensing

Abstract

Soft sensing provides a reliable estimation of difficult-to-measure variables and is important for process control, optimization, and monitoring. The extraction of beneficial information from the abundance of available data in modern industrial processes and the development of data-driven soft sensors are becoming areas of increasing interest. In addition, the use of deep neural networks (DNNs) has become a popular data processing and feature extraction technique owing to its superiority in generating high-level abstract representations from massive amounts of data. A deep relevant representation learning (DRRL) approach based on a stacked autoencoder is proposed for the development of an efficient soft sensor. Representations from conventional DNN methods are not extracted for an output prediction, and thus a mutual information analysis is conducted between the representations and the output variable in each layer. Analysis results indicate that irrelevant representations are eliminated during the training of the subsequent layer. Hence, relevant information is highlighted in a layer-by-layer manner. Deep relevant representations are then extracted, and a soft sensor model is established. The results of a numerical example and an industrial oil refining process show that the prediction performance of the proposed DRRL-based soft sensing approach is better than that of other state-of-the-art methods.