DEKF based Recurrent Neural Network for state estimation of nonlinear dynamical systems

In this paper decoupled extended kalman filter (DEKF) based Recurrent Neural Network (RNN) has been proposed for state estimation of nonlinear dynamical systems. The proposed state estimator uses cascading of recurrent neural network structures to learn the internal behavior of the dynamical system along with the measuring relations of the system from the input-output data through prediction error minimization. A dynamic learning algorithm for the recurrent neural network has been developed using DEKF. The performance of the proposed method is illustrated for an induction motor which is a typical nonlinear dynamical system and has been compared with that of the conventional state estimation method such as EKF.