Hohlov

pp. 62-79

pp. 137-153

pp.132-147

The article studies a neural network approach to obtain the statistical characteristics of the input vector signal and noise implementations at ill-conditioned matrices of correlation moments to solve the problems to select and reduce the vector dimensions of informative features at detection and recognition of signals and noise on the basis of regression methods. It considers a problem concerning the matrix condition of the correlation moment of informative signal features and a possibility to apply the neural networks in theoretical studies. Neural network algorithms are used to obtain the necessary statistical data for ill-conditioned matrices of correlation moments of informative features. It is shown that the weights obtained at training the neural network with no zero weights are the regression coefficients of the initial parameters of informative features input.

The article presents a method to form the non-periodic selection functions in the short-range coherent pulse Doppler radar information systems with random signal phase-shift keying (manipulation) in transmitting and heterodyne channels. It studies statistical characteristics of Doppler signals received when reflecting from the point object that is located inside the first and arbitrary remote strobes at synchronized random carrier phase-shift keying in transmitting and heterodyne channels. The article shows an expansion of the relative bandwidth of energy spectrum of a Doppler signal received after reflecting from the point object from an arbitrary remote strobe, except for the first one. To form a non-periodic selection function is possible via the recognition method for the Doppler signal implementations using the relative bandwidth of energy spectrum and via the application of a regression tract to process intervals between zeroes of implementations.

The authors propose a way to adapt the neural network algorithm based on an assessment of the average frequency of fluctuations in realization of the input signal and tuning of the clock frequency in estimating the number of samples in each interval between zeros of signal implementation.It allows you to achieve invariance of distribution of duration of intervals between zeros with respect to velocity of objects classified by the neural network.

 For solving classification problems of air and ground targets the authors justify informative signal features and robust neural network algorithms of formation of decision-making areas; these algorithms use prior information on initial statistical characteristics of acoustic signals.

 The article describes the possibility of obtaining waveform envelope readings of input implementations using the Hilbert conjugated process which can be used as informative parameters of the signals that are invariant to the energy spectrum medium frequency with a relative width of the band to 0.7-0.8. In this case automatic tracking of the sample rate is performed according to the changing medium frequency of the input implementation energy spectrum.

 The article considers algorithms of detection and recognition of random non-centered signals and disturbances in remote strobes of pulse locators. Neural network and regression algorithms were proposed for formation of decision-making areas. For single deterministic signals quasi-optimal filtering was considered.