Deep Learning Framework for Data-driven Soft Sensor Modeling

Deep learning has been widely used in industrial processes, which automatically learns hidden knowledge from process data and detects quality variables that are difficult to measure. Therefore, accurate extraction and utilization of effective features and elimination of useless features are still one of the most important research issues in soft sensor modeling. In this paper, an attention-based gated supervised encoder-decoder BiLSTM (AGSED-BiLSTM) is proposed. AGSED-BiLSTM first encodes input variables to obtain feature expressions of different abstract levels with high correlation with quality parameters, and links the outputs of each layer together through gating neurons, so as to make full use of the information of different hidden layers to obtain the final output, and improves learning efficiency through bidirectional architecture. Finally, the proposed model is applied to penicillin fermentation process, which proves the effectiveness and superiority of the proposed soft sensor model based on AGSED-BiLSTM network, and is superior to the most advanced and traditional soft sensor models based on deep learning.