Semi‐Supervised Learning of Parsimonious Spatiotemporal Representations for Industrial Quality Prediction

Abstract Stacked autoencoder (SAEs) are widely applied in industrial processes. However, issues such as insufficient correlation between hidden layer features and quality variables, feature redundancy caused by increased stacking layers, and inadequate utilization of unlabeled samples compromise prediction accuracy. This article proposes a method called semi‐supervised learning of parsimonious spatiotemporal representations for industrial quality prediction (SQTDSTF). A semi‐supervised spatiotemporal feature extraction network (SST) processes labeled and unlabeled samples; a quality‐related module selects key features to enhance feature‐quality correlation; kernel principal component analysis (KPCA) is applied to each hidden layer to suppress redundancy. The effectiveness of the SQTDSTF model is validated through simulations of debutaning towers and thermal power generation.