S.V. Chelebaev – Ph.D. (Eng.), Associate Professor, Department «Automated control systems»,  V.F. Utkin Ryazan State Radio Engineering University

E-mail: sergeychelr@yandex.ru

Yu.A. Chelebaeva – Post-graduate Student, Department «Information-measuring and biomedical engineering», V.F. Utkin Ryazan State Radio Engineering University

E-mail: chel-juliya@yandex.ru

Now a need for implementation of functional dependence of two and more variables for conversion facilities of biosignals is had.

One of the main problems is increase in reproduction accuracy of nonlinear functional dependence of two variables. One of ways of increase in accuracy of function biosignals converters is the mathematical apparatus of artificial neural networks.

Neural network structures of converters of time-and-frequency biosignals parameters in the digital code on the basis of two-layer and three-layer perceptron at the mathematical description level are offered in work. Several options of neurons distribution between hidden layers of 3-layer network of the biosignals converter are considered.

The technique of converters adjustment of time-and-frequency biosignals parameters in the code of two variables consisting of 6 stages is offered. Adjustment results of neural network converters of two variables are given for the selected function. Influence of activation function parameter on a conversion error extent is analyzed. Influence of the training selection volume on a conversion error extent is investigated. The dependence of conversion error on the neurons number of hidden layer for a neuronet with one hidden layer is obtained. The dependence of conversion error on the neurons number of hidden layers for a neuronet with two hidden layers is obtained. Questions of neurons optimum allocation of converter network between hidden layers are investigated.

Implementation questions of the offered models on the programmable logic integrated circuits (FPGAs) are considered. It is offered to use piecewise linear approximation as a way of implementation of sigmoidal activation function of neurons on FPGAs. The dependence of sections quantity of approximation on reproduction accuracy of sigmoidal activation function of neurons is found. Coefficients of piecewise linear approximation for sigmoidal activation function are given as an example. Hardware costs of implementation of activation function taking into account the obtained coefficients are given.

The obtained results allow to draw a conclusion that offered converters structures of time-and-frequency parameters of biosignals in the digital code of two variables have a low error of nonlinear dependence reproduction. Therefore the offered structures of converters can be implemented on Field Programmable Gate Arrays with insignificant hardware expenses.

1. Golovko V.A. Nejronnye seti: obuchenie, organizaciya i primenenie. M.: IPRZHR. 2001. 256 s.
2. Bereznyj E.A., Rubin A.M., Utekhina G.A. Prakticheskaya kardioritmografiya. M.: Neo. 2005. 140 s.
3. Daponte P., Grimaldi D., Michaeli L. Gray code ADC based on analog neural curcuit // Radioengineering. Apr. 1995. V. 4. № 1. P. 7–12.
4. Loktyuhin V.N., CHelebaev S.V. Nejrosetevye preobrazovateli impul'sno-analogovoj informacii: organizaciya, sintez, realizaciya / Pod obshch. red.  A.I. Galushkina. M.: Goryachaya liniya–Telekom. 2008. 144 s.
5. Loktyuhin V.N., CHelebaev S.V., Antonenko A.V. Nejrosetevye analogo-cifrovye preob-razovateli. M.: Goryachaya liniya–Telekom. 2010. 128 s.
6. Galushkin A.I. Nejronnye seti: osnovy teorii. M.: Goryachaya liniya–Telekom. 2010. 496 s.