A new on-line identification approach for affine modeling of nonlinear processes using an adaptive neural network

This paper presents a new approach for on-line identification of an exact affine model for single-input, single- output (SISO) processes with nonlinear and time-varying behaviors. For this purpose, a modified growing and pruning algorithm for radial basis function (MGAP-RBF) neural network is used for affine modeling of the SISO nonlinear and time-varying processes. The extended Kalman filter (EKF) is utilized as an effective learning algorithm for parameter adaptation in the GAP-RBF neural network. The performances of the modified GAP-RBF and the original GAP-RBF are evaluated on a non-affine CSTR benchmark problem with nonlinear and time-varying behavior. Simulation results show good performance of the MGAP-RBF neural network for identification of an exact affine model for the CSTR.