Dynamic data rectification using the extended Kalman filter and recurrent neural networks

The use of recurrent neural networks (RNN) for process modeling and data rectification is described from the viewpoint of system identification. RNNs are demonstrated to be a type of simple, nonparametric, nonlinear state-space model. The discrete extended Kalman filter (DEKF) is then introduced and combined with the RNN in order to estimate the states of the RNN model and hence the process measurements through the measurement equation. Simulation results are presented that indicate the combination of the RNN and DEKF provides superior results over the RNN alone