Abstract
Design of static observers employing neural network has already appeared in the literature. In this paper neural networks are exploited to design nonlinear dynamic observers for estimating the states of a nonlinear system. A number of schemes using Multi-layered Feed-forward Neural Network (MFNN) are presented. In the first approach, the neural network is used to approximate the nonlinear Kalman gain of the observer. Two different training schemes are proposed in this structure. Full and reduced order observer schemes based on a more direct approach are then considered. These schemes utilize the neural nets to assume the nonlinear dynamic mapping from the system input and output in order to obtain the estimated states. The network training for all the schemes is based on a gradient algorithm that uses the recently proposed Block Partial Derivatives (BPD). Simulation results are presented to validate the usefulness of the proposed schemes.