Dynamic system identification for looper tension control Using BPNN & Kalman Filtering

Abstract

System identification is an important area in control system. This paper discuss some of the reasons that causes the slow convergence of Back Propagation Neural Network (BPNN). This paper also comment on the effect of the number of neurons in the hidden layer when apply BPNN with Kalman Filtering to dynamic system identification for a looper tension control. BPNN is based on LMS, and uses steepest descent method to find the optimum weights to the adjacent layer. It always consumes much of time while training and not easy to get a global optimal value while applying to on-line training. The Levenberg–Marquardt algorithm (LMA) provides a numerical solution to the problem of minimizing a function, generally nonlinear, over a space of parameters of the function. Kalman Filtering is a better linear and discrete method for parameters estimation. By this way to solve a problem, it can involve the initial conditions, and also can apply to stationary and non-stationary system. So, applying BPNN and Kalman Filtering together to the dynamic system identification will give a better result both on convergent efficiency and stability.