NEURAL NET USING TO DETERMINE DEPTH AND FREQUENCY OF SIGNALS’ MODULATION FOR ELECTRICAL EQUIPMENT ULTRASONIC VIBROCONTROL

Abstract

The apparatus of artificial neural networks (ANN) is proposed to be used for signal processing in active ultrasonic (US) vibration control of electrical equipment. A feature of the applied neural network algorithm is that the required information about vibration parameters is embedded in the ultrasound signal’s phase change at its constant amplitude. Under these conditions, traditional spectral analysis of signals requires a high sampling rate and a significant recording duration. When using the direct propagation’s ANN with three hidden layers, it was shown that it is sufficient to use a sampling frequency of 5-6 points for the period of an ultrasonic wave and a recording duration of 4-5 periods to estimate the nonstationary frequency and amplitude of the vibration signal. Estimates of the error in determining the amplitude, frequency and phase of vibrations are obtained. The root-mean-square errors of the neural network algorithm do not exceed units of percent. The possibilities of using a trained neural network for signal processing in a «sliding window» are demonstrated. The accuracy characteristics of the proposed neural network algorithm of signal processing and the possibility of its optimization for electrical equipment’s vibration control are discussed.