Partial Least Squares and Neural Network Based Identification of Process Dynamics & Design of Neural Controllers

Abstract

Present study implemented the Neural network (NN) and Partial least squares (PLS) based identification of process dynamics for single-input single output (SISO) as well as multi-input multi-output (MIMO)systems. In the present study, the Neural network (NN) based controller
design has been implemented for a non-linear continuous bioreactor process. Multilayer feed forward networks (FFNN) were used as direct inverse neural network (DINN) controllers as well as IMC based NN controllers. The training as well as testing database was created by
perturbing the open loop process with pseudo random signals (PRS).DINN controllers performed effectively for set-point tracking. To address the disturbance rejection problems, which are very likely to be faced by the bioreactors, the IMC based neural control architecture was proposed with suitable choice of filter and disturbance transfer
function. To assess the controllability of the various configurations, like conventional turbidostat and nutristat& concentration turbidostat and nutristat, the offset or degree of disturbance rejection by the proposed
IMC based NN controllers were utilized. The ‘concentration turbidostat’using the feed substrate concentration as the manipulated variable was found to be the best control configuration among the continuous bioreactor configurations.A (2×2) distillation column was simulated to generate the time series data consisting various inputs and outputs of the process. Multivariate statistical technique PLS was used to relate the scores of input and output matrices. ARX as well as linear least squares techniques were used for inner relation development between the
input-output scores. The PLS model of the distillation column dynamics could simulate the process with reasonable accuracy.