Mechanism-based deep learning for tray efficiency soft-sensing in distillation process

Distillation is an important unit operation in the chemical industry. However, its process variables fluctuation can frequently cause abnormal conditions, resulting in the reduction of system reliability, and even causing safety accidents. Tray efficiency, as its key operation indicator, has been a long-term implicit variable that cannot be directly monitored so that the operators have insufficient information about the running status of the distillation system. Soft sensing for tray efficiency can greatly improve the safety, stability and reliability of the production system. In this paper, a mechanism-based deep learning method is proposed for the soft sensing of tray efficiency in distillation process. Firstly, based on the statistics of extreme alarm values and distillation process mechanism, the tray efficiency that is prone to anomalies is analyzed. The key trays that need to be monitored are identified. Secondly, the typical working conditions of the distillation system are focused by data clustering as the input of mechanism modeling. Then, the distillation system is simulated to obtain associated datasets of tray efficiency and process measurable variables. Finally, the LSTM-based deep learning model ex-tracts the mechanical characteristics of the distillation system to construct a surrogate model for the tray effi-ciency soft-sensing by using these datasets.