Data rectification using recurrent (Elman) neural networks

Nonlinear programming techniques are used to train Elman-type simple recurrent neural networks to reconcile simulated measurements for a simple dynamic system (a draining tank). The networks are trained in a batch mode using the BFGS quasi-Newton nonlinear optimization algorithm. This makes it possible to avoid the trial and error associated with tuning the learning rate and momentum terms required in the various backpropagation algorithms. The random measurement errors used are uncorrelated and purely Gaussian in nature. Noisy data are used for both training the networks and testing network performance. Recurrent Elman networks are able to significantly reduce the noise level in the process measurements without explicit knowledge of the nonlinear dynamics of the system