A Neural Net Based Scheme for Narrowband Signal Extraction

Abstract

This paper proposes an alternative method for extraction of narrow band signal contaminated with varying white noise power. A three layred feedforward network with one hidden layer is suggested in this paper. The internal parameters are updated by employing the widely used backpropagation algorithm and the statistical Cauchy's algorithm. The network uses the tangent hyperbolic function to provide the desired nonlinearity to the scheme. During learning the net is exposed to signals with high SNR ranging from +40 dB to +20 dB. However, after learning quite interesting results are obtained for very low SNR of the input signal even up to 0 dB. The training samples for fundamental, 3rd harmonic and 5th harmonic are chosen to be different with a view to exploit the pattern classification feature of the net. Therefore, after termination of training the proposed scheme accomplishes the filtering action successfully. The phenomenon of network paralysis as encountered in BP algorithm is circumvented by opting for the statistical Cauchy's algorithm with Boltzmann's probability feature. The extraction capability of the proposed scheme is highlighted by providing inadequate noisy input samples to the net. Besides, a performance comparison is presented for both the algorithms. Results presented emphasizes the substitution of the scheme for adaptive filter for a particular zone of operation.