Soft sensor modeling using deep learning with maximum relevance and minimum redundancy for quality prediction of industrial processes

Abstract

Deep learning techniques such as autoencoder (AE) and stacked autoencoders (SAE) have gained growing popularity in soft sensor applications. However, they often encounter several disadvantages, such as poor correlations between the extracted hidden features and the quality variable, inevitable information loss resulting from the layer-wise feature extraction, and information redundancy between the hidden features. Thus, a maximal relevance and minimal redundancy-based representation learning (MRMRRL) is proposed for quality prediction of industrial processes. MRMRRL obtains significant performance enhancement by combining the merits from three channels. First, the relevance between the input and output variable is taken into account for enabling quality-relevant feature extraction. Second, kernel principal component analysis (KPCA) is performed on the feature space of AE hidden layer for achieving redundancy reduction of hidden features. Third, inputs with high quality relevance are fed to the extension layer nodes for enabling information compensation. The experimental results show that, compared with the baseline SAE, the performance of MRMRRL is improved by about 37 % and 38 % for two application examples, respectively. Significant performance enhancement of MRMRRL can be also obtained compared to several state-of-the-art deep learning soft sensors. These results demonstrate the effectiveness and superiority of the proposed MRMRRL approach in extracting quality-related hidden features while ensuring automatic elimination of hidden feature redundancy and maintaining structure simplicity.