Neural Network Signal Processing: The Tasks Without “Deep Learning”

Abstract

The purpose of the study is to evaluate the possibility of using artificial neural networks for electrical signals processing.
 
 Methods. The application of a multilayer perceptron for these purposes as the simplest neural feed forward network is considered. Its difference as the basis of neural network algorithms is that when analyzing dynamic processes, signal processing should be carried out in a “sliding time window”.
 
 Results. It is shown that neural network processing makes it possible to approximate the shape of the signal with high accuracy and determine its parameters in real time. Using the example of periodic signals and transients in electrical circuits, accuracy assessments are made and the features of neural network processing are analyzed. The necessary sizes of the training sample of signals and the level of errors that occur when testing a neural network are discussed. Estimates of the required signal sampling frequency, the “sliding window” duration and the variation range of signal parameters when creating a training sample are given.
 
 Conclusions. It is shown that the proposed approach does not require “deep learning” of neural networks with complex architecture. It enables to create a signals training sample based on simple analytical formulas and to control the neural network algorithm quality at intermediate stages of calculations.