Peculiarities of creation and training of the neural network direction finder

We consider the problem of passive radio direction finder radio source (RS) in the short wave (SW) range by means of the system of separated receiving aerials, the number of which is N. Generally, location of the points of receiving can be random, at every point of receiving the nondirectional aerial is used. To solve the problem, we use artificial neural network (ANN) in the form of three-layered perceptron. The input signals of ANN are N values of useful signal, measured at every point of receiving, relative to the phase center of the aerial system. The output signal of ANN is a value of bearing RS.

Methods of creation, training and testing of ANN are described in details which were made in the MATLAB environment. The results of the static simulating of useful signal at different values of the signal to noise ratio (SNR) were used during training and testing the direction finder. Satisfactory training of ANN showed that it is possible to measure bearing in the sector of angles different from the full circle. Otherwise function approximated with the help of ANN has a break, which can lead to the impossibility of training ANN. Two or more ANN, trained for work in the appropriate sectors of bearing, are required to use for work of ANN direction finder within the limit of 360°.

For research of ANN-direction finder, the Aerial system is used in the form of ring aerial grid, containing N=8 aerials: one aerial is in the center and seven others are evenly placed on the circle. Optimization of such parameters as the number of neurons in the hidden layer, the number of the training patterns and values of STP training selection patterns were held during training ANN-direction finder. By means of statistical modeling analysis of work of the ANN-direction finder is carried out mean root square errors (MRSE) of the bearing accuracy depending on value of SNR is calculated.

Comparison between the ANN-direction finder and optimal direction finder, which was based on getting of the bearing value by the maximum likelihood criterion was performed. This comparison showed that the mean value of bearing MRSE, averaged over all angle measurement, for the ANN-direction finder loses to the optimal direction finder no more than on 3-5%.