Publishing house Radiotekhnika

"Publishing house Radiotekhnika":
scientific and technical literature.
Books and journals of publishing houses: IPRZHR, RS-PRESS, SCIENCE-PRESS

Тел.: +7 (495) 625-9241


Neuroprocessor modeling of nonlinear surfaces of separating of object classes in the multidimensional continual space

DOI 10.18127/j19998554-201809-03


N.L. Dembitskiy – Ph.D. (Eng.), Associate Professor, Moscow Aviation Institute (National Research University)

The separation of classes is one of the most complex tasks of object detection and identification. In radar systems, high accuracy requirements are imposed on this task. The quality of the solution depends directly on the time of obtaining the results. The problem becomes especially difficult in conditions when the values of the object parameters differ insignificantly or are represented by intersecting sets. With increasing computational complexity, there are tangible delays in obtaining the result.
The advantage of neuroprocessor methods over algorithmic calculation systems is high speed, low power consumption and versatility. Neuroprocessors allow you to obtain real-time results in the analysis of multiparameter objects without resorting to the use of complex algorithms. The important advantage of neuroprocessors is its ability to adapt the computational process to the conditions of the problem being solved. Learning with examples of solutions allows you to configure the processor without programming.
The neuroprocessor model considered in the article is based on analogous schemes. It allows approximating nonlinear surfaces that can be used in problems of separating linearly inseparable classes of objects. The processor performs modeling of nonlinear surfaces in a multidimensional continual space. The article presents the scheme of its application in the task of separating classes of objects.
A special feature of the processor is the possibility of approximating the functions of many variables from experimental and theoretical data. The result of the calculations can be obtained in a time not exceeding a fraction of a microsecond. The speed is achieved by the elimination of analog-digital transformations, cyclic algorithms, parallelization of computations.
The technique is based on the partition of the domain Ω of a multidimensional continuum space into the set of adjoining subregions Ω*. Linear approximation is performed at the point specified by the parameters of the recognition object. The search for the result is carried out by parallel computations using the interpolating fuzzy processor developed by the author. High processor speed is achieved through parallelization and hardware implementation of calculations.
The important advantage of an analog neuroprocessor is elimination of programming phase. The setting is carried out in the learning mode.
The parameters of the neuroprocessor are adjusted according to experimental or theoretical data on the values of the function of separating objects at given points of the N-dimensional continual space. On the training input, the reference signals d1, d2, …, dS are alternately sent, and the corresponding input stimuli x1, x2, …, xN. The reference signals are selected such that in each subregions Ω* Ω there is at least one training sampling of signals. When repeating the learning epochs, the error of the function in the corners of the subregion Ω* is minimized.
The experimental verification of the proposed method showed a high convergence of the model parameters to nominal values.
The processor greatly simplifies the algorithmic and circuit design solutions. It can replace powerful digital processors in applications that require the use of small devices. The effect is achieved due to the gain in speed, mass and energy parameters in comparison with digital systems.

  1. Kul'chin Yu.N., Kim A.Yu., Notkin B.S., Lyukhtep A.B. Obpabotka signalov pasppedelennoj volokonno-opticheskoj seti dlya raspoznavaniya obrazov s primeneniem nejponnykh setej // Informatsionnye tekhnologii. 2015. T. 21. № 4. S. 312–318.
  2. Galushkin A.I. Nejpochipy i nejpomopfnye EVM: ppoblemy modelipovaniya // Informatsionnye tekhnologii. 2015. T. 21. № 12. S. 942–948.
  3. El-Khoribi R.A., Shoman M.A.I., Mohammed A.G.A. Automatic digital modulation recognition using artificial neural network in cognitive radio // International Journal of Emerging Trends & Technology in Computer Science (IJETTCS). 2014. V. 3. № 3. P. 132–136.
  4. Patent № 2269155 RF. Model' nejrona, realizuyushchaya logicheskuyu funktsiyu neravnoznachnosti / E.A. Samojlin. Opubl. 27.01.2006. Byul. № 3.
  5. Kruglov V.V., Borisov V.V. Iskusstvennye nejronnye seti. Teoriya i praktika. M.: Goryachaya liniya – Telekom. 2001.
  6. Anadigm Inc. Home Pages [Elektronnyj resurs] / URL: (data obrashcheniya: 19.02.2018).
  7. Polishchuk A. Programmiruemye analogovye IS Anadigm: ves' spektr analogovoj elektroniki na odnom kristalle // Sovremennaya elektronika. Dekabr' 2004. S. 8–11.
  8. Dembitskij N.L. Interpoliruyushchij nechetkij protsessor dlya avtomaticheskikh regulyatorov // Aviakosmicheskoe priborostroenie. 2017. № 7. S. 23–31.
  9. Osovskij S. Nejronnye seti dlya obrabotki informatsii. M.: Finansy i statistika. 2002.
  10. Bretthorst G.L. Bayesian spectrum analysis and parameter estimation. New York: Springer–Verlag. 1988.

© Издательство «РАДИОТЕХНИКА», 2004-2017            Тел.: (495) 625-9241                   Designed by [SWAP]Studio