Semi-Supervised Soft Sensor Modeling for Industrial Dynamic Information Mining Based on Two-Path Weight Comparison

Deep learning methods can extract reliable feature representations from massive process data to build accurate soft sensor models. However, the data in actual industrial production is often nonlinear, dynamic, and even limited in label samples due to untimely sampling and analysis. Aiming at the above problems, this paper proposes a semi-supervised soft sensor modeling method (SSACF-LSTM) for industrial dynamic information mining based on two-path weight comparison. An unsupervised long short-term memory network (LSTM) encoder-decoder is utilized to extract the hidden dynamic nonlinear information in unlabeled samples. This information is then spliced and masked with the initial input, which is subsequently fed into a deep neural network encoder-decoder for pretraining to learn a self-supervised task that mines correlation information between sequences as well as between samples that is useful for the model. In addition, the LSTM with attentional convolutional fusion serves as an auxiliary path that mines correlation information between variables directly from the initial input to complement the information extracted by self-supervision. Finally, the results of both are linearly weighted to obtain the final prediction. The method is validated in a debutanizer process. The results show that SSACF-LSTM can utilize both labeled and unlabeled data to extract quality-related features for soft sensor modeling, which is better than the semi-supervised LSTM network with historical feature fusion (HFF-ssLSTM), semi-supervised stacked auto-encoder (SSAE), and variational auto-encoder regression (SVAER).