Reduction and Identification of Noise Signals Using Artificial Neural Networks with Various Activation Functions

Abstract

The scientific paper examined options for reducing the levels of noisy signals and their identification using artificia neural structures. Single linear neurons were applied to sinusoidal signals with added Gaussian White Noise and Periodic Random noise. The changes of the Sum Squared Errors are monitored by selecting their minimum values, which achieve the lowest noise levels. Artificial neural structures were created to identify square waveforms with superimposed Uniform Gaussian Noise and Periodic Random Noise. Various types of activation functions and neuronal units were tested in the hidden layer of neural models by examining the metrics - Accuracy and Mean Squared Error. The highest accuracy of 94.00% achieved was obtained by hyperbolic tangent sigmoid activation function.