Neural networks for real-time estimation of signal parameters

The problem of estimating the amplitudes and frequencies of sinusoidal signals from noisy and distorted data has received considerable attention. Many sophisticated algorithms have been proposed including the Prony method, Pisarenko harmonic decomposition and the Yule-Walker method. Many of these algorithms lead to a large computation burden and are rather numerically time consuming. Fast determination of parameters of the fundamental waveform of voltages and currents is essential for control and protection devices. For this purpose various numerical algorithms have been developed, e.g. based on the Fourier and Kalman filtering. When using these algorithms, the speed of processing is quite limited. There has been a great interest in parallel algorithms and architectures based on the methods of artificial neural networks. The purpose of this paper is to present new algorithms and along with them new architectures of analogue neuron-like adaptive processors for on-line estimation of parameters of signal components, which are distorted by transient components, higher harmonics and noise. For steady-state conditions we have developed neural networks which enable us to estimate the amplitudes and the frequency of the fundamental component of signals. When estimating the basic waveform of currents during short circuits, the exponential DC component distorts the results. Assuming the known frequency, we have developed an adaptive feedback neural network which enables us to estimate the amplitudes of the basic component as well as the amplitudes and the time constant of the DC component