RECONSTRUCTION OF BINARY ..SIGNALS USING ARTIFICIAL NEURAL NETWORKS

Abstract

This dissertation work considers the problem of reconstructing the digital signals which have been passed through a dispersive channel and corrupted with additive noise, a problem which is of considerable importance in digital communication including digital television transmission. Here the channel equalization problem is viewed as a geometric spatial decision problem and it is shown that the optimal decision boundary is highly nonlinear. Traditional techniques for solving equalization problems are based on linear finite filters which can only generate linear decision boundaries. Since the channel equalization is an inherently nonlinear problem conventional linear equalizers need to be replaced by some nonlinear architectures capable of realizing highly nonlinear boundaries to achieve fully or near optimal performance. Here, Artificial Neural Networks are chosen as nonlinear structures for channel equalization and it is demonstrated how the problems encountered by linear equalizers under adverse conditions on signal-to-noise ratio and channel phase can be overcome using these nonlinear classifiers. Neural networks of the type Feed Forward Error Backpropagation are utilized as adaptive channel equalizers. The neural network simulations are implemented in `C' and run on a UNIX system. The training of the neural networks occurred on a set of samples chosen from {-1,1 } with equal probability and corrupted with zero mean white Gaussian noise. The simulation results are presented which suggest that the neural network equalizers offer a performance which exceeds that of the linear structures, particularly in the high-noise environment.