IMPLEMENTATION OF SOFT SENSORS USING SOFT COMPUTING TECHNIQUES

Abstract

In the monitoring and control of distillation columns, on-line product composition measurement offer challenges. The product composition in the distillation column is difficult to measure on-line due to the limitations such as time delay, high cost and reliability. In this dissertation, the "soft sensors for estimating product compositions, are constructed. The construction of soft sensors makes use of secondary variables, such as temperature and pressure, to infer the value of non-measurable primary variables such as distillate composition. As a case study, a binary distillation column with steady state data and a multicomponent distillation column with dynamic time series data are considered. Both the plants are simulated using HYSYS plant simulator. To provide the necessary process insights, analyses of dynamic behavior has been carried out. Appropriate secondary measurements with significant relationships with the product composition are then identified for the construction of the inferential estimator within MATLAB environment. Four techniques have been considered to construct soft sensors for the cases mentioned above. Initially, the properties of linear, nonlinear regression techniques and neural network are exploited to construct soft sensors to estimate the product composition. Later, a Genetic Algorithm (GA) based Neural Network has been constructed as a soft sensor for the estimation of composition. This is particularly useful for cases involving changing operating conditions as well as for highly nonlinear processes. The Genetic Algorithm based Neural Network, employed as inferential estimator is successful in providing product estimates to a reasonable accuracy. Finally, a Generalized Regression Neural Network (GRNN) has been used to construct a soft sensor. The performance criterions, Mean Square Error (MSE) and Mean Absolute Percentage Error (MAPE), are considered to compare all four techniques mentioned above. The results so obtained are then tabulated, plotted, interpreted and compared for these techniques. It is observed that the hybrid GA-Neural Network estimator provides better results than Regression and Neural Network techniques, while the GRNN based soft sensor offers best results.