Processing and Analysis of Signals with Superposed Noises by Artificial Neural Networks

Sine, square and triangle signals with added impact of Uniform White Noise (UWN) and Inverse F Noise (IFN) are studied in LabVIEW. Spectral characteristics of the signals based on the Fast Fourier Transform (FFT) algorithm are obtained. A potential task for identification of the type of disturbing impacts with application of artificial intelligence, trained by Scaled Conjugate Gradient technical approach in partial data samples for signals in the time and frequency domains is considered. Target probabilities of membership for UWN signals and IFN signals with and without FFT processing are defined. Procedures for synthesis of Feed-forward neural architectures with “Sigmoid” and “Softmax” activation functions in the hidden and output layers at different amounts of hidden neural units within the limits of 5 to 25 were performed. The data on basic criteria for cross-entropy synthesis and classification accuracy are analyzed. Neural networks for identification of spectral signals and signals with superposed noise without FFT with 20 and 11 hidden neurons and with the best accuracy ratios of 73.40% and 98.00% were selected. It is established that the use of FFT in the signal processing is not achieve the desire effect for improvement of classification parameters. An analogous approach is applied to an attempt to identification of sine and square signals without preprocessing, were the presence of UWN are replaced by Gaussian White Noise (GWN), and where the highest achieved identifiable accuracy is 94.20%.