Nonlinear noise filtering with neural networks: comparison with Weiner optimal filtering

This paper reports the application of a multi layer perceptron for the filtering of noisy time-series signals. The signals employed for the investigation are frequency modulated sine waves corrupted by different kinds of Gaussian and non-Gaussian noise. In general, overlapping spectra of signal and noise require nontrivial solutions for removing the noise with minimal degradation of the signal. For this task a multi layer perceptron is trained with corresponding pairs of noise corrupted and noise free signal patterns. The trained neural network is tested by independent test signals from the same signal/noise generator used for training. As a reference an optimal Wiener filter is designed using the same information with which the neural network was trained. The performance of both approaches is quantitatively measured by the signal to noise ratio at the filter output for different signal to noise ratios at the filter input