Data Driven Modelling for Dual Composition Estimation in Binary Distillation Process

In distillation process, critical process variable like composition is not viable for direct online measurement and control. In such cases data driven soft sensors offer a cheap and effective replacement for traditional sensors to estimate these primary process variables by inferring to measurable secondary variables. In the present work a data driven soft sensor is developed to estimate product compositions of a laboratory distillation column UOP3CC inferring to selective tray temperature and mixture vapour liquid equilibrium data. Initially an Auto Regressive Moving Average Exogenous (ARMAX) process identification method is used to formulate a Two Input Two Output (TITO) linear state space process models using process input-output data. These models are then used to estimate top and bottom compositions using Kalman Filter under additive Gaussian process and sensor noises. The results are reported for a non-ideal mixture separation with a continuous feed input. The performance of the estimator is tested for step change in reflux and reboiler power inputs under feed flow disturbances. Results show that the Kalman filter is able to optimally estimate the composition under the presence of feed flow disturbances.