Recurrent Use of a Perceptron for Signal Structural Analysis

Abstract

The paper is devoted to the use of an artificial neural network (ANN) of direct propagation (multilayer perceptron) for signal processing in electrical engineering and electric power industry. It is proposed to use such simple neural networks instead of ANN with a more complex structure (convolutional, recurrent), but within the framework of a sequential recurrent algorithm. This allows checking and controlling the quality of signal processing at each stage of calculations. The proposed algorithm is tested on the example of structural analysis of a signal with nonlinear distortions in a sliding time window. It is shown that the amplitude, frequency and phase of an industrial frequency signal with a high level of harmonics and an aperiodic component can be isolated with an accuracy of units of percent for a time not exceeding units of milliseconds. To increase the accuracy at each step of the calculations, traditional methods can be used, in addition to the ANN: averaging, median smoothing, etc.