An adaptive filter using neural networks approach

A new adaptive filter for system state estimation, based on a recurrent neural networks approach, has been proposed in the paper. A general procedure for defining the desired output signal dynamics in the training algorithm, based on the methodology of projecting an identity observer for deterministic dynamic systems, has been developed. An alternative approach for designing the desired output vector, based on a specific three-state track model with position only measurement and physical nature of the state vector components, has been also considered. Results of simulation demonstrating the robustness of the proposed filter, in the sense of its low sensitivity to the imprecise knowledge of noise statistics and the presence of unmodelled dynamics, are included.