DATA-DRIVEN SYSTEM IDENTIFICATION USING NEURAL NETWORK WITH EXTENDED KALMAN FILTER FOR DISTILLATION COLUMN SYSTEM

• Main Author: Alifsyah Putra Nasution, M.
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78863
• Institution: Institut Teknologi Bandung

The development of the chemical industry in Indonesia has been continuously increasing up to the present time. One of the processes that is almost always present in the chemical industry is the distillation process. The plant used in this research is a batch-type distillation column system at the ITB Honeywell Control System Laboratory, which can be used to separate binary solution of ethanol and water. To apply the control scheme effectively, the plant’s dynamics model is required. However, the challenge lies in the fact that the distillation column plant is a highly nonlinear and multivariable system based on its physical laws, with a limited number of sensors available to collect all states information, where only 1 state, which is also an output, can be measured. As a result, this thesis research performs data-driven system identification through black-box modeling based on experimental input-output data using neural network to generate both nonlinear and linear models. This research employs the nonlinear NARX-FNN model and the linear ARMA-FNN model for online identification of the distillation column system, with a focus on studying the comparison of learning algorithms for FNN weight updating. The algorithms compared are Stochastic Mode Steepest Gradient Descent (SGD) and Extended Kalman Filter (EKF). Both algorithms are suitable for comparison because they both operate in an instance-by-instance mode and can be employed for online system identification. The best MSE result was achieved with the ARMA-FNN architecture using the EKF algorithm with a hyperparameter REKF = 1, resulting in an MSE of 5.5418e-05. Furthermore, with both the NARX- FNN and ARMA-FNN architectures, the EKF algorithm with a hyperparameter REKF = 0,05 was able to converge the fastest, around the 6th instance, although with a trade-off where the MSE results were slightly worse. However, the ability of the EKF algorithm to converge with only a small amount of data is highly advantageous when faced with real-world scenarios where there are only a few instances in the dataset.