Wave Interpretation and Neural Net Monitoring of Noise in Voltage Signals on Power Lines

Abstract

Noise in voltage signals on power lines is determined by many factors. Therefore, at the standard sampling rate of signals in measuring instruments, it is considered, most often, to be Gaussian. At a high sampling rate, the noise is modulated, its distribution differs from the normal one. The analysis and control of its structure is of interest, for example, for damage diagnostics and determining the damage location. The purpose of the study is to show the possibility of neural network control of heterogeneous noise in voltage signals on power lines. Methods. Based on the wave analysis of signals in power lines, a noise model in industrial frequency voltage signals is described, which allows interpreting its modulation as a result of random spatial fluctuations in wave velocity. The control of noise heterogeneity over the harmonic signal period is carried out on the basis of a recurrent ANN in a sliding time window, the duration of which does not exceed 2 ms. Results. The noise model in power line voltage signals is proposed as a result of wave reflection from the wave velocity spatial inhomogeneities in the line. In the Born’s scattering approximation, noise is described by the simplest analytical formulas with random parameters. A neural network algorithm based on LSTM cells was tested on model signals recordings, which is used in a sliding time window and allows one to control the noise variance in units of percent of the industrial frequency signal amplitude. Estimates of the neural network algorithm accuracy are given. Conclusions. A comparison of the noise structure obtained using the proposed model with experimental signals recordings confirms the adequacy of the model at a qualitative level. The proposed neural network monitoring algorithm has high accuracy. The approach can be used to monitor the present state of power lines.