Quality-driven deep feature representation learning and its industrial application to soft sensors

Establishing effective soft sensors relies on feature representation that is capable of capturing critical information. Stacked AutoEncoder (SAE) is able to capture the intricate structures of data characterized by high dimensionality and strong non-linearity by extracting abstract features layer by layer, making it widely used. However, the pretraining process of SAE is unsupervised, which means the features extracted cannot leverage label information to provide more actionable insights for downstream tasks. To extract more valuable feature representation, a new quality-driven dynamic weighted SAE (QD-SAE) is proposed in this paper. By incorporating supervised information dominated by the quality variable into the learned features during the pretraining of the SAE and weighting the abstract features layer by layer, the features that are beneficial to the prediction task are thus focused. In QD-SAE, the supervised information is computed by an improved attention score. In the initial state of the supervised fine-tuning process, the weighted features compose the hidden layers of the entire network. Finally, a benchmark function case and a real complex industrial process case are used to verify the effectiveness and advantages of QD-SAE. The experimental analyses demonstrate that the soft sensor constructed by the QD-SAE can predict the output variable with high accuracy and outperforms the conventional neural networks.