ĐIỀU KHIỂN PID MỘT NƠ-RON HỒI QUY LƯU LƯỢNG CHẤT LỎNG

Abstract

This study aims to experiment a recurrent single neural PID controller on RT020 liquid flow control system of the Gunt-Hamburg. A digital PID controller is organized as a recurrent single neuron. The weights of neuron are corresponding to three parameters K_p, K_d and K_i of the PID, can be trained online during control action. This training algorithm allows to self-adjust the parameters of the PID according to the change of condictions, based on the controlled plant's sensitivity, so called Jacobian information. An radial basis function (RBF) neural network is also trained online for non-parametric model identification, in order to estimate the Jacobian information. Experimental results on the RT020 device, and comparison with the PID controller provided by the manufacturer show that the recurrent single neural PID controller can be self-tuning during control; the transient response is improved with reducing setting time, low overshoot and eliminated steady-state error.