Novel Semi-Supervised Deep Probabilistic Slow Feature Extraction for Online Chemical Process Soft Sensing Application

Nonlinear and slowness characteristics commonly appear with increasing complexity and dynamics in chemical processes, which have brought challenges for soft sensor modeling. To handle these characteristics, a feature extractor Siamese variational autoencoder (SVAE) model is designed for obtaining probabilistic slow features (PSFs) by resampling from the probability distribution, which can be a better solution to the uncertainty existing in the chemical process data. Furthermore, a semi-supervised SVAE (ss-SVAE) regression method is proposed to handle the issue of multiple sampling rate existing between quality and auxiliary variables. In addition, a feature augmentation method is developed, utilizing the previous quality variable value to introduce quality information into regressor. While in online prediction process, PSFs are augmented with the prediction value at previous moment to enhance the predictive performance. To avoid the cumulative errors, the prediction value will be replaced with the actual value after quality variable collected. Finally, the proposed method is applied to a debutanizer column case and a CO2 absorption column case to verify the feasibility and effectiveness.