An Unsupervised Recurrent Neural Network for Noise Identification

摘要

This article presents an unsupervised recurrent neural network that solves the noise identification problem. This work is based on an ARMA model and a two-layered linear neural network. The proposed scheme uses noise output samples as neural network inputs. Noise inverse dynamics and noise input samples are identified using a recurrent adaptive backpropagation algorithm. The neural network weights are calculated imposing on the input noise samples to be white noise. This structure can be used for off-line and on-line identification. Simulated results are presented for both cases.