RU2780606C1 - Method for detecting and classifying naval targets based on neural network technologies and artificial intelligence elements - Google Patents

Abstract

FIELD: hydroacoustics.

SUBSTANCE: invention relates to hydroacoustics and can be used to implement neural network recognition operations of target classes (surface or underwater object) detected by signs of amplitude-phase modulation of low-frequency signals of pumping the marine environment by radiation and fields of objects. Essence: in the method for detecting and classifying marine targets based on neural network technologies and artificial intelligence elements, when the device is turned on in the marine environment, a zone of nonlinear interaction and parametric transformation of pumping waves with object signals is formed, for which the emitter and the receiving antenna are placed on opposite borders of the controlled area of the marine environment, then the pumping waves modulated by the object signals are received and amplified in the parametric transformation band, their frequency-time scale is transferred to the high-frequency region, a narrow-band spectral analysis is carried out, parametric components of the total or difference frequency are isolated, according to which, taking into account the time and parametric transformation of waves, the characteristics of the object’s signals are restored, which are fed into the neural network recognition and classification path, consisting of a target class recognition unit based on amplitude-frequency characteristics, implementing computational operations of an artificial neural network and covered by feedback with the training unit, in whose memory the data of mathematically processed images of spectrograms of marine targets are recorded, moreover, the first input of the target class recognition unit according to amplitude-frequency characteristics receives data from the output of the spectrum analyzer of the signal reception, processing and registration path, and its second input receives data from the training unit of the neural network recognition and classification path. In this case, the data is fed into the additionally introduced synthesis path of neural network and neural fuzzy recognition models with grouping of features at the input, the unit for analyzing information about the features and topology of the training sample, and then fed to the input of the neural network synthesis unit, covered by feedback with the neural fuzzy network synthesis unit, the output of which is connected to the input of a logical device, the function of which is performed by a six-layer Wang-Mendel neural fuzzy network, where neural and neural fuzzy models are synthesized in a non-iterative mode with linearization, factor grouping and convolution of signs of a marine target and a fuzzy logical conclusion is formed for the training unit of the neural network recognition and classification path, after which an artificial neural network is configured and a conclusion is formed about the degree of belonging of the studied spectral region to the classification object (surface or underwater object).

EFFECT: an increase in the efficiency of classification of marine targets by 5-7% relative to the prototype and other classification methods using modulation of low-frequency signals of pumping the marine environment by radiation and fields of objects.

1 cl, 11 dwg

Description

The invention relates to hydroacoustics and can be used to implement operations of neural network recognition of target classes (surface or underwater object) detected on the basis of amplitude-phase modulation of low-frequency pump signals of the marine environment by radiation and object fields.

The principle of operation of parametric antennas with both high-frequency (tens-hundreds of kHz) and low-frequency (tens-hundreds of Hz) pumping of the marine environment is based on the use of the natural nonlinear properties of the marine environment.

There are known methods and parametric antennas that implement them, using high-frequency pumping of the marine environment (see Nonlinear hydroacoustics in monitoring systems for hydrophysical and geophysical fields of marine areas: monograph / M.V. Mironenko, A.M. Vasilenko, V.A. Pyatakovich, [and - Vladivostok: VUNC VMF "VMA" (branch, Vladivostok), 2013. - 324 p. - Text: direct .; Technologies of nonlinear transmissive hydroacoustics and neuro-fuzzy operations in the problems of recognition of marine objects: monograph / V.A. Pyatakovich, A. M. Vasilenko, M. V. Mironenko - Vladivostok: Far East Federal University, 2016. - 190 pp. - ISBN 978-5-7444-3790-9 - Text: direct.; Recognition and classification of sources of formation of fields of various physical nature in the marine environment: monograph / V. A. Pyatakovich, A. M. Vasilenko, O. V. Khotinsky - Vladivostok: G. I. Nevelskoy Marine State University, 2017. - 255 pp. - ISBN 978-5-8343-1066-2. - Text: direct.; Parametric optimization designation of a neural network system for classifying marine targets according to the criterion of reliability. / V.A. Pyatakovich, V.F. Rychkov. - Text: direct // Marine intelligent technologies. - St. Petersburg: 2018. No. 4-5 (42). - S. 153-161.; A method for classifying underwater technical objects by an expert intelligent system with a receiving parametric antenna / V.A. Pyatakovich, A.M. Vasilenko, V.F. Rychkov. - Text: direct // Marine intelligent technologies. - St. Petersburg: 2018. No. 2(40). Volume 2. - S. 121-126 .; Patent No. 2593625. C2. Russian Federation, IPC G10K 11/00 (2006.01). A method of transmitting information waves from the marine environment to the atmosphere and vice versa. Application No. 2015115231/28 : Appl. 04/22/2015 : publ. (registered) 10.08.2016 Bull. No. 22 / Mironenko M.V., Vasilenko A.M., Pyatakovich V.A.; - 17 s. : ill. - Text : direct.; Calculation of the spatial structure of a multibeam parametric antenna using software modules of the integrated marine monitoring system. / V.A. Pyatakovich, A.M. Vasilenko. - Text: direct // Results of modern scientific research and development: Sat. articles based on the materials of the III International Scientific and Practical Conference - Penza: ICNS "Science and Education", 2017. - P. 50-62 .; Processing of hydroacoustic signals in paths with a parametric receiving antenna for solving problems of classification of sources of formation of fields by a neural network expert system. / V.A. Pyatakovich, O.V. Khotinsky. - Text: direct // Prospects for the development of modern science: Sat. articles based on the materials of the International Scientific and Practical Conference - Volgograd: TsPM, 2017. - P. 32-42.).

Known methods of receiving an elastic wave in the marine environment, increasing the value of the parameter of the nonlinearity of the environment in the zone of signal reception. Why, in addition to the elastic wave, an additional electromagnetic signal is introduced into the zone of parametric signal reception, subjected to frequency-time modulation, or a modulated hydrodynamic fluid flow. The modulation of the hydrodynamic flow of the liquid is carried out by changing the density due to controlled gas saturation, or changing the temperature, or changing its chemical composition. (Patent No. 2158029. C2. Russian Federation, IPC G10K 11/00(2006.01), G10K 15/02(2006.01). Method for receiving an elastic wave in the marine environment (variants). Application No. 98122520/28 : Appl. : 15. 12.1998 .: publ. 20.10.2000. / Korochentsev V.I., Mironenko M.V., Zvonarev [et al.] - 11 pp.: ill. - Text: direct.; Patent No. 2167454. C2. Russian Federation , IPC G10K 11/00 (2006.01), G10K 15/02 (2006.01) Method for transmitting an elastic wave in sea water (variants) Application No. 98122980/28,: Appl.: 15. 12.1998. : Published on 20.05.2001. / Korochentsev V.I., Mironenko M.V., Zvonarev M.I. [et al.] - 11 pp.: illustration - Text: direct.; Patent No. 2602763 C2. Russian Federation, IPC G01H 3/00 (2006.01) G10K 11/00 (2006.01) Method for parametric reception of waves of different physical nature in the marine environment Application No. 2014152012/28 : Application 22.12.2014 : Published 20.11.2016 Bulletin No. 32 / Mironenko M.V. , Malashenko A.E., Karachun L.E. [et al.] -32 pp.: illustration - Text: direct.; Patent No. 2694848 C1. Russian Federation, IPC G10K 11/18 (2006.01) G01S 13/86 (2006.01) G01S 3/00 (2006.01) G01H 3/00 (2006.01). A method for forming a scalable system for detecting and classifying marine targets with elements of artificial intelligence. Application No. 2018143169: Appl. 12/05/2018: publ. (reg.) 07/17/2019. Bull. No. 20 / Pyatakovich V.A., Vasilenko A.M., Mironenko M.V.; - 23 s. : ill. - Text : direct.).

The disadvantages of the methods are low sensitivity and, as a result, limited range (a few kilometers) of parametric reception of information waves of various physical nature in the infrasonic and fractional (units-fractions of hertz) frequency ranges.

The closest in technical essence to the claimed invention is the "Method for detecting and neural network recognition of signs of fields of various physical nature generated by marine targets" (Patent No. 2682088. C1. Russian Federation, IPC G01S 15/02 (2006.01). Application No. 2018120829: Appl. 06/05/2018 : published 03/14/2019 Bulletin No. 8. / Vasilenko A.M. - 16 pp. : ill. - Text : direct), which consists in the fact that first a zone of nonlinear interaction and parametric transformation is formed in the marine environment pump waves with object signals, for which the emitter and receiving antenna are placed on opposite boundaries of the controlled area of the marine environment, then pump waves modulated by object signals are received and amplified in the parametric conversion band, their time-frequency scale is transferred to the high-frequency region, and a narrow-band spectral analysis, allocate the parametric components of the sum or difference frequency, according to which, taking into account ohm of temporal and parametric wave conversion restore the characteristics of the signals of the object, which are fed into the neural network recognition and classification path, consisting of a target class recognition unit by amplitude-frequency characteristics, which implements the computational operations of an artificial neural network and is covered by feedback with a learning unit, in whose memory data of mathematically processed images of spectrograms of sea targets (MC), and the first input of the target class recognition unit by amplitude-frequency characteristics receives data from the output of the spectrum analyzer, covered by feedback with the recorder, the path for receiving, processing and recording signals, and its second input receives data from the training block of the path of neural network recognition and classification.

The analysis of the low-frequency, mid-frequency and high-frequency components of the amplitude-frequency characteristic is carried out separately, with the help of a learning unit, the recognition network is adjusted according to the classification features of the fields generated by sea targets, and its weight coefficients are corrected, based on the results of computational operations, the recognition network forms a conclusion about the degree of belonging to the area under study spectrum to the classification object (surface or underwater object).

It is known that the result of parametric transformation of interacting waves is their mutual amplitude-phase modulation. A small frequency difference (within the same order) of transmissive waves and waves generated by the object ensures their most intense interaction. The amplitude of the interacting waves and the phase modulation index can be represented in the following form

;

.

;

.

,

- частота волны накачки и полезного сигнала, соответственно;

,

- затухание волны накачки и полезного сигнала, соответственно;

- объем среды нелинейного взаимодействия и параметрического преобразования волн;

- расстояние от точки излучения до точки расположения объекта;

- плотность,

- скорость звука в морской среде. where γ is the coefficient of nonlinearity of the marine environment;

,

are the frequency of the pump wave and the useful signal, respectively;

,

are the attenuation of the pump wave and the useful signal, respectively;

- the volume of the medium of nonlinear interaction and parametric transformation of waves;

- distance from the point of radiation to the point of location of the object;

- density,

is the speed of sound in the marine environment.

The parametric components of the sum and difference frequencies formed as a result of the conversion of transmission waves during processing of broadband signals are distinguished as signs of amplitude-phase modulation, which is justified by mathematical dependencies and confirmed by the results of marine experiments (see Nonlinear transmission hydroacoustics and marine instrumentation in the creation of the Far Eastern radio hydroacoustic lighting system of the atmosphere, the ocean and the earth's crust, monitoring their fields of various physical nature. : monograph / [M.V. Mironenko et al.] ; Russian Academy of Sciences, Far Eastern Branch, Special Design Bureau for Automation of Marine Research. - Vladivostok: Publishing House of the Far East un-ta, 2014. - 402 pp.: ill.; ISBN 978-5-906739-22-3 - Text: direct.; Low-frequency transmission method of long-range sonar of hydrophysical fields of the marine environment: monograph. / Mironenko M.V., Malashenko A.E., Karachun L.E. [and others] - Vladivostok: SKB SAMI FEB RAN, 2006. - 173 p. - Text : direct.; Technologies of non-linear transmission hydroacoustics and neuro-fuzzy operations in problems of recognition of marine objects: monograph / V.A. Pyatakovich, A.M. Vasilenko, M.V. Mironenko. - Vladivostok: Dalnevost. federal. un-t, 2016. - 190 p. - ISBN 978-5-7444-3790-9. - Text : direct.; Results of experimental studies of the method for determining the profile of a marine object and the system that implements it / M.V. Mironenko, V.A. Pyatakovich, A.M. Vasilenko. - Text : direct // Monitoring. Science and technology: 2017. No. 2(31). - S. 64-69.; Optimal and adaptive methods for classifying hydroacoustic signals for marine intelligent systems / V.A. Pyatakovich, A.B. Surov, V.F. Rychkov. - Text: direct // Marine intelligent technologies. - St. Petersburg: 2020. No. 1-2 (47). - S. 186-194.; Calculation of the efficiency of target classification by the intellectual system of the Navy, using a set of computational operations of neural networks. / V.A. Pyatakovich, V.F. Rychkova, N.V. Pyatakovich. - Text: direct // Marine intelligent technologies. - St. Petersburg: 2020. No. 1 (47) Volume 2. - S. 175-185 .; Automation of information processing in the intelligent system of marine monitoring / V.A. Pyatakovich, A.V. Nikolaev, E.A. Kostikov. - Text: direct // Problems of mechanical engineering and automation. - Moscow: 2020. No. 4. - S. 72-79 .; Observation results of disturbances in the marine environment generated by synoptic processes by the low-frequency transmission method. / V.A. Pyatakovich, A.M. Vasilenko, [i dr.]. - Text: direct // Materials of the International Scientific Readings "Primorskie Dawns - 2015". - Vladivostok: Dalnevost. federal. un-t, 2015. - S. 89-92.).

The spectrum of interacting waves consists of an infinite number of side components, the frequency and amplitude of which can be found from the well-known expression:

, - результирующее и мгновенное значения давления модулированной волны, соответственно;

- удвоенная частота модулированной волны;

- волна, генерируемая объектом;

- время;

- функции Бесселя n-го порядка;

- амплитуда модулированной волны;

- коэффициент модуляции. Как видно из выражения, значения частот боковых составляющих отличаются от удвоенной центральной частоты 2ω (равной сумме частот взаимодействующих волн) на величину ± n⋅Ω, где n - любое целое число. Амплитуды боковых составляющих для соответствующих частот (2ω ± n Ω) определяются величиной множителя

.where

, - resulting and instantaneous pressure values of the modulated wave, respectively;

- doubled frequency of the modulated wave;

- wave generated by the object;

- time;

- Bessel functions of the nth order;

- amplitude of the modulated wave;

- modulation factor. As can be seen from the expression, the values of the frequencies of the side components differ from the doubled central frequency 2ω (equal to the sum of the frequencies of the interacting waves) by ± n⋅Ω, where n is any integer. The amplitudes of the side components for the corresponding frequencies (2ω ± n Ω) are determined by the value of the multiplier

.

спектр взаимодействующих волн приближенно состоит из удвоенной центральной частоты 2ω и ее боковых частот 2ω+Ω и 2ω-Ω.For small values of the modulation coefficient

the spectrum of interacting waves approximately consists of the doubled central frequency 2ω and its side frequencies 2ω+Ω and 2ω-Ω.

The disadvantage of the prototype method is the lack of operations that independently perform auto-tuning of their rule base, based on a sample of mathematical models of marine targets, synthesizing neural network and neuro-fuzzy recognition models with feature grouping and correcting the data of the updated library of mathematically processed images of MC spectrograms for parameter control and optimization computational processes performed in the path of neural network recognition and classification, which provides the final classification solution for detected marine targets (surface or underwater object), which limits the functionality of the prototype system.

The task to be solved by the claimed invention is to expand the functionality of the prototype method, i.e. its implementation as a method for detecting and classifying marine targets based on neural network technologies and elements of artificial intelligence, which allows, based on training data samples, to create the structure of neural and neuro-fuzzy networks, as well as adjust their parameters without optimizing weights, synthesize neural and neuro-fuzzy models in a non-iterative mode with linearization, factor grouping and convolution of features, carried out using the synthesis path of neural network and neuro-fuzzy recognizing models with the grouping of features necessary to implement the final classification process in the path of neural network recognition and classification. The final classification decision on the detected sea targets (surface or underwater object) is made by the trained neural network according to the degree of belonging of the investigated spectrum region to the object of classification.

The technical result of the invention is the automation of the process of recognizing classes of marine targets (surface or underwater object) detected by the signs of amplitude-phase modulation of low-frequency pumping signals of the marine environment by radiation and fields of objects, using methods that allow, based on training data samples, to create a structure of neural and neuro-fuzzy networks, as well as adjust their parameters without optimizing weights, synthesize neural and neuro-fuzzy models in a non-iterative mode with linearization, factorial grouping and feature convolution, carried out using the synthesis path of neural network and neuro-fuzzy recognizing models with grouping of features necessary to implement the final classification process in path of neural network recognition and classification, which provides an increase in the probability of reliable classification of a marine target by 5-7% greater than when using the prototype.

The specified technical result is achieved by performing recognition operations by a path of neural network recognition and classification of a marine target using a modified combined recognition network consisting of Kohonen, Grosberg networks and a MAXNET network, the settings of which are regulated by the synthesis path of neural network and neuro-fuzzy recognition models with a grouping of features in the form of a six-layer neuro-fuzzy Wang-Mendel network, which generates a fuzzy logical inference based on the synthesis of neural and neuro-fuzzy models in a non-iterative mode with linearization, factorial grouping and convolution of marine target features for the operation of a logical device, which allows you to automatically recognize the target class by amplitude-frequency characteristics and draw a conclusion on the degree of belonging of the studied region of the spectrum to the object of classification (surface or underwater object).

To solve this problem, a method has been developed for detecting and classifying marine targets based on neural network technologies and elements of artificial intelligence, which consists in first forming in the marine environment a zone of nonlinear interaction and parametric conversion of pump waves with object signals, for which the emitter and receiving antenna are placed on opposite boundaries of the controlled area of the marine environment, then the pump waves, modulated by the object's signals, are received and amplified in the parametric conversion band, their frequency-time scale is transferred to the high-frequency region, a narrow-band spectral analysis is performed, the parametric components of the sum or difference frequency are isolated, which, taking into account temporal and parametric wave conversion restore the characteristics of object signals that are fed into the neural network recognition and classification path, consisting of a target class recognition unit by amplitude-frequency characteristics characteristics, which implements the computational operations of an artificial neural network and is covered by feedback with a learning unit, in whose memory data of mathematically processed images of spectrograms of sea targets are recorded, and the first input of the target class recognition unit by amplitude-frequency characteristics receives data from the output of the spectrum analyzer covered by feedback with a registrar of the path for receiving, processing and recording signals, and its second input receives data from the learning unit of the path of neural network recognition and classification.

The fundamental difference from the prototype is that the data is fed into the additionally introduced path for the synthesis of neural network and neuro-fuzzy recognition models with a grouping of features at the input, the block for analyzing information about the features and the topology of the training sample, and then fed to the input of the neural network synthesis block, covered by feedback from the block synthesis of neuro-fuzzy networks, the output of which is connected to the input of a logic device, the function of which is performed by a six-layer neuro-fuzzy Wang-Mendel network, where neural and neuro-fuzzy models are synthesized in a non-iterative mode with linearization, factorial grouping and convolution of sea target features and a fuzzy logical conclusion is formed for the learning block path of neural network recognition and classification, after which the artificial neural network is set up and a conclusion is formed about the degree of belonging of the studied region of the spectrum to the object of classification (surface or underwater object).

As is known, the extraction of useful information from hydroacoustic signals determines the basics of data processing algorithmization in the method of detecting and neural network recognition of signs of fields of various physical nature generated by marine targets. To form the feature vector, which is the input information array of the recognition network, the mask method is used. The process of forming information arrays is necessary to solve two problems, the first of which is the process of forming reference samples necessary for the implementation of the learning process of the recognition network, and the second for recognizing targets (see Recognition and classification of sources of formation of fields of various physical nature in the marine environment: monograph / V. A. Pyatakovich, A. M. Vasilenko, O. V. Khotinsky - Vladivostok: Marine State University named after G. I. Nevelsky, 2017. - 255 pp. - ISBN 978-5-8343- 1066-2. - Text: direct.; Neural network technologies in intelligent systems for detecting and operational identification of marine targets: monograph / V.A. Pyatakovich, A.M. Vasilenko, O.V. Khotinsky. - Vladivostok: Marine State University t named after G. I. Nevelsky, 2018. - 263 pp. - ISBN 978-5-8343-1083-9. - Text: direct.; Certificate of state registration of the database No. 2018621040. Russian Federation. Decision-making algorithms by a neuroclassifier in fuzzy conditions when recognizing a sea target. Application No. 2018620642: Appl. 05/21/2018: publ. (reg.) 10.07.2018. Bull. No. 7. / V.A. Pyatakovich; applicant: V.A. Pyatakovich; Processing of hydroacoustic signals in paths with a parametric receiving antenna for solving problems of classifying sources of field formation by a neural network expert system / V.A. Pyatakovich, O.V. Khotinsky. - Text: direct // Prospects for the development of modern science: Sat. articles based on the materials of the International Scientific and Practical Conference - Volgograd: TsPM, 2017. - P. 32-42 .; Certificate of state registration of the computer program No. 2019611559 Russian Federation. Program for designing and training artificial neural networks of perceptron type. Application No. 2019610135: Appl. 01/10/2019: publ. (reg.) 01/29/2019. Bull. No. 2. / A.M. Vasilenko, V.A. Pyatakovich., O.A. Alekseev.; applicant: A.M. Vasilenko, V.A. Pyatakovich; Certificate of state registration of the computer program No. 2019611455 Russian Federation. Software package for modeling and training ANN Application No. 2019610187 : Appl. 01/10/2019: publ. (reg.) 01/28/2019. Bull. No. 2 / V.A. Pyatakovich, A.M. Vasilenko, N.V. Pyatakovich; applicant: V.A. Pyatakovich, A.M. Vasilenko, N.V. Pyatakovich. ).

The idea of the method is that for each mask, the maximum amplitude value is searched, which is the orth of the vector of classification features. To automate the process of searching for an extremum in the zone of one mask, the maximum search network MAXNET was used. The iterations of the network are terminated after the output neurons of the network stop changing. Type of elements of input signals - integers or real numbers, type of elements of output signals - real numbers. The dimensions of the input and output signals are the same. The type of activation function is linear with saturation (a linear section is used). The number of synapses in the network is N (N - 1).

The formation of synaptic weights occurs according to the formula:

where W _ij - i-th synaptic weight of the j-th neuron; N is the number of elements of the input signal (the number of neurons in the network).

The functioning of the network is given by the expression:

where x _j - element (ort) of the input signal of the network; y _i - output of the j-th neuron.

The normalization of the input feature vector obtained after the analysis of the masks by the MAXNET network is performed according to the expression

известны и определяются моделью входного гидроакустического сигнала.Value Range Bounds

are known and determined by the input hydroacoustic signal model.

где Т - вектор синаптических весов сети; (Х^*,Y^*) - обучающие пары;

- норма вектора (см. Нелинейная гидроакустика в системах мониторинга гидрофизических и геофизических полей морских акваторий : монография / М.В. Мироненко, А.М. Василенко, В.А. Пятакович, [и др.]. - Владивосток : ВУНЦ ВМФ «ВМА» (филиал, г. Владивосток), 2013. - 324 с. - Текст : непосредственный.; Технологии нелинейной просветной гидроакустики и нейронечетких операций в задачах распознавания морских объектов : монография / В.А. Пятакович, А.М. Василенко, М.В. Мироненко. - Владивосток : Дальневост. федерал. ун-т, 2016. - 190 с. - ISBN 978-5-7444-3790-9. - Текст : непосредственный.; Распознавание и классификация источников формирования полей различной физической природы в морской среде : монография / В.А. Пятакович, А.М. Василенко, О.В. Хотинский. - Владивосток : Морской гос. ун-т им. Г. И. Невельского, 2017. - 255 с. - ISBN 978-5-8343-1066-2. - Текст : непосредственный.; Нейросетевые технологии в интеллектуальных системах обнаружения и оперативной идентификации морских целей : монография / В.А. Пятакович, А.М. Василенко, О.В. Хотинский. - Владивосток : Морской гос. ун-т им. Г. И. Невельского, 2018. - 263 с. - ISBN 978-5-8343-1083-9. - Текст : непосредственный.; Алгоритм автоматической классификации морских целей на основе анализа амплитудной огибающей их гидроакустических шумов. / В.А. Пятакович, А.М. Василенко. - Текст : непосредственный // Проблемы и методы разработки и эксплуатации вооружения и военной техники ВМФ : сб. статей. - Владивосток : ТОВВМУ, 2016. - Вып. 92. - С. 167-173.; Оптимальные и адаптивные методы классификации гидроакустических сигналов для морских интеллектуальных систем. / В.А. Пятакович, А.Б. Суров, В.Ф. Рычкова. - Текст : непосредственный // Морские интеллектуальные технологии. - Санкт-Петербург : 2020. - № 1-2 (47). - С. 186-194.).The recognition network is trained on the basis of the error backpropagation algorithm, which implements the gradient method for optimizing the functional of the form:

where T is the vector of synaptic weights of the network; (Х ^* ,Y ^* ) - training pairs;

- vector norm (see Nonlinear hydroacoustics in systems for monitoring hydrophysical and geophysical fields of marine areas: monograph / M.V. Mironenko, A.M. Vasilenko, V.A. Pyatakovich, [and others]. - Vladivostok: VUNTS VMF " VMA "(branch, Vladivostok), 2013. - 324 pp. - Text: direct .; Technologies of nonlinear transmissive hydroacoustics and neuro-fuzzy operations in the problems of recognizing marine objects: monograph / V.A. Pyatakovich, A.M. Vasilenko, M V. Mironenko, Vladivostok: Far East Federal University, 2016, 190 pp., ISBN 978-5-7444-3790-9, Text: direct, Recognition and classification of sources of formation of fields of various physical nature in marine environment: monograph / V. A. Pyatakovich, A. M. Vasilenko, O. V. Khotinsky - Vladivostok: Marine State University named after G. I. Nevelsky, 2017. - 255 pp. - ISBN 978- 5-8343-1066-2 - Text: direct .; Neural network technologies in intelligent systems for the detection and operational identification of marine targets: monograph / V.A. Pyatakovich, A.M. Vasilenko, O.V. Khotinsky. - Vladivostok: Marine State. un-t im. G. I. Nevelskoy, 2018. - 263 p. - ISBN 978-5-8343-1083-9. - Text : direct.; Algorithm for automatic classification of sea targets based on the analysis of the amplitude envelope of their hydroacoustic noise. / V.A. Pyatakovich, A.M. Vasilenko. - Text: direct // Problems and methods of development and operation of weapons and military equipment of the Navy: Sat. articles. - Vladivostok: TOVVMU, 2016. - Issue. 92. - S. 167-173.; Optimal and adaptive methods for classifying hydroacoustic signals for marine intelligent systems. / V.A. Pyatakovich, A.B. Surov, V.F. Rychkov. - Text: direct // Marine intelligent technologies. - St. Petersburg: 2020. - No. 1-2 (47). - S. 186-194.).

For auto-correction and regulation of the error backpropagation algorithm when training the recognition network, the synthesis path of neural network and neuro-fuzzy recognition models with feature grouping is used, which combines adaptive self-learning approaches and expert experience, which, in the presence of uncertain parametric disturbances, allows for the correction of values corresponding to new classification conditions sea target.

The invention is illustrated by drawings, where in Fig. 1 shows a functional diagram of a system for detecting and classifying marine targets based on neural network technologies and artificial intelligence elements:

1. Radiating transducer (underwater sound beacon brand PZM-400 emitting signals at a frequency of about 400 Hz).

2. Receiving transducer.

3. Marine environment.

4. Working zone of nonlinear interaction and parametric transformation of pump and information waves.

5. Objects (marine targets generating acoustic, electromagnetic and hydrodynamic radiation).

6. Emission path of pump signals.

6.1. Stabilized frequency pump signal generator.

6.2. Amplifier.

6.3. Consent block.

7. Path for receiving, processing and recording information signals.

7.1. Broadband amplifier.

7.2. Time-Frequency Converter.

7.3. Spectrum analyzer.

7.4. Registrar.

8. Path of neural network recognition and classification.

8.1. Target class recognition block by amplitude-frequency characteristics.

8.2. Learning block.

9. Synthesis path for neural network and neuro-fuzzy recognition models with feature grouping.

9.1. Block for analyzing information about the features and topology of the training sample.

9.2. Neural network synthesis block.

9.3. Synthesis block of neuro-fuzzy networks.

9.4. logical device.

The general structure of the modified combined recognition network, the path of neural network recognition and classification, consisting of the Kohonen and Grosberg networks is shown in Fig. 2, the neural network interpretation of the multivariate classification algorithm is shown in FIG. 3, the structure of the Wang-Mendel fuzzy neural network is shown in FIG. 4, the structure of the feed-forward neural network is shown in FIG. 5, and the structure of the six-layer neuro-fuzzy network in Fig. 6.

The mask method used for frequency response recognition is shown in FIG. 7, and the interpretation of the three-dimensional output vector of recognition of hydroacoustic signals from the amplitude-frequency characteristic in FIG. 8. In FIG. Figure 9 shows a comparative description of the methods for classifying a marine target by training and classification time. The results of the computational experiment to determine the recognition (classification) coefficient are shown in Fig. 10 and 11.

To synthesize logically transparent neuromodels in a non-iterative mode, it is necessary to analyze information about the features and topology of the training sample, for which the following sequence of steps is performed.

. Step 1. Define the training sample

.

,

. Set Threshold

,

.

:

.Calculate logarithms of feature values for all instances and expand feature set

:

.

и максимальные

значения признаков. Step 2. Find the minimum

and maximum

feature values.

:Normalize feature values for

:

. Например, коэффициенты парной корреляции:Step 3. Find the coefficients characterizing the stable relationships of features and their logarithms,

. For example, pair correlation coefficients:

Step 4. Group the features.

.Step 4.1. Set the number of feature groups:

.

элемент с максимальным по модулю значением:

. Определить знак

.Step 4.2. Find in matrix

the element with the maximum modulo value:

. Define sign

.

, то перейти на шаг 4.5.Step 4.3. If a

, then go to step 4.5.

, то выполнить шаги 4.4.1 - 4.4.4.Step 4.4. If a

, then follow steps 4.4.1 - 4.4.4.

.Step 4.4.1. Install:

.

: если

и

, то установить:

.Step 4.4.2. For all

: if

and

then install:

.

.Step 4.4.3. Install:

.

Step 4.4.4. Go to step 4.2.

, тогда принять:

, перейти на шаг 4.5, в противном случае - перейти на шаг 4.6. Step 4.5. If a

then accept:

, go to step 4.5, otherwise go to step 4.6.

, для всех

, для которых значение

не было установлено. Шаг 4.7. Для

, установить:Step 4.6. Install

, for all

, for which the value

has not been established. Step 4.7. For

, install:

, установить:

.Step 4.8. For

, install:

.

, характеризующие тесноту связи признаков и номера класса, такие, что

, и с увеличением значения

возрастает значимость признака для определения класса.Step 5. Find the coefficients

, characterizing the closeness of the relationship of features and class numbers, such that

, and with increasing value

the significance of the attribute for class definition increases.

, где

- количество кластеров, определить координаты центров кластеров

. Кластер-анализ можно выполнить, используя методы [2, 12]. Определить четкую принадлежность кластеров к каждому из

классов.Step 6. Perform a cluster analysis of the training sample: split it into clusters

, where

- the number of clusters, determine the coordinates of the centers of the clusters

. Cluster analysis can be performed using methods [2, 12]. Determine the clear belonging of clusters to each of

classes.

интервалов (термов) и определить их параметры для синтеза нейронных и нейронечетких сетей.Step 7. Split the value axis of each feature into

intervals (terms) and determine their parameters for the synthesis of neural and neuro-fuzzy networks.

.Step 7.1. Set current feature number:

.

, тогда перейти на шаг 7.6, в противном случае - перейти на шаг 7.3.Step 7.2. If a

, then go to step 7.6, otherwise go to step 7.3.

.Step 7.3. Set the number of the current interval of values of the current attribute

.

, тогда перейти на шаг 7.5, в противном случае - установить для текущего

:Step 7.4. If a

, then go to step 7.5, otherwise, set for the current

:

. Перейти на шаг 7.4.To accept:

. Go to step 7.4.

. Перейти на шаг 7.2.Step 7.5. To accept:

. Go to step 7.2.

и

определить

и

- номера класса и кластера для

-го терма,

:Step 7.6. Parameter Based

and

define

and

- class and cluster numbers for

th term,

:

Step 8. Set membership functions of the recognizable instance to fuzzy terms. To do this, it is proposed to use trapezoidal functions:

, где

- функция принадлежности распознаваемого экземпляра по признаку

к

-му терму

-го признака,

-

-образная функция, а

-

-образная функция:or U-shaped functions:

, where

- membership function of a recognizable instance by feature

to

term

-th sign,

-

-shaped function, and

-

-shaped function:

or triangular functions:

or Gaussian functions:

Based on the results of the analysis of the information about the features on the training sample, a neural network of direct propagation is synthesized (Fig. 5). The neurons of the first layer of the network normalize the input signals, mapping them to the range [0, 1], and also perform, where necessary, the logarithm of the normalized signal.

The neurons of the second layer group the signals transformed by the neurons of the first layer and find the weighted sum of the signals of the group, taking into account the estimates of their individual influence on the output feature (class number). The neurons of the third layer of the network determine the distances from the recognized instance to the cluster centers in the space of grouped signals and find the value of the Gaussian function, whose argument is a certain distance. The neurons of the fourth layer correspond to classes and output the value "1" if the recognized instance belongs to the corresponding class, and "0" otherwise.

The discriminant functions of the network neurons will be given by the formulas:

- дискриминантная (постсинаптическая, весовая) функция

-го нейрона

-го слоя,

- набор весовых коэффициентов

-го нейрона

-го слоя,

- весовой коэффициент

-го входа

-го нейрона

-го слоя,

- набор входов

-го нейрона

-го слоя,

- значение на

-ом входе

-го нейрона

-го слоя нейронной сети.where

- discriminant (postsynaptic, weight) function

-th neuron

th layer,

- set of weight coefficients

-th neuron

th layer,

- weight coefficient

-th entrance

-th neuron

th layer,

- a set of inputs

-th neuron

th layer,

- value on

-th input

-th neuron

th layer of the neural network.

The activation functions of neurons will be given by the formulas:

- функция активации

-го нейрона

-го слоя нейросети.where

- activation function

-th neuron

th layer of the neural network.

The weight coefficients of the network neurons are calculated by the formula:

Modification of the structure of the modified combined recognition network, neural network recognition and classification path shown in FIG. 2 consists in adding the MAXNET network to the Kohonen and Grosberg network, which is very important for the problem being solved.

,

;

подаются все компоненты входного вектора

Число нейронов во втором (скрытом) слое определяется взаимным расположением и формой разделяемых множеств.For each neuron of the first layer through synapses with weights

,

;

all components of the input vector are fed

The number of neurons in the second (hidden) layer is determined by the mutual arrangement and shape of the shared sets.

;

подаются выходные сигналы первого слоя. Число нейронов третьего (выходного) слоя определяется числом рассматриваемых классов, подлежащих распознаванию.For each neuron of the second layer through synapses with weights

;

the output signals of the first layer are applied. The number of neurons in the third (output) layer is determined by the number of considered classes to be recognized.

;

подаются выходные сигналы второго слоя. Значения выходных сигналов третьего слоя образуют вектор

решений. Нейроны, составляющие сеть, одинаковы и имеют функцию активации известного типа

где

,

и

- значения

-го входного сигнала, выходного сигнала и внешнего смещения

-го нейрона

-го слоя;

- число нейронов в

-м слое;

For each neuron of the third layer through synapses with weights

;

the output signals of the second layer are applied. The values of the output signals of the third layer form a vector

solutions. The neurons that make up the network are the same and have an activation function of a known type

where

,

and

- values

th input signal, output signal and external bias

-th neuron

-th layer;

- number of neurons in

-th layer;

Границы диапазона значений

известны и определяются моделью входного гидроакустического сигнала.The input vectors are preprocessed by normalizing the input feature vector obtained after mask analysis by the MAXNET network or after obtaining statistical estimates according to the expression

Value Range Bounds

are known and determined by the input hydroacoustic signal model.

Data preprocessing, one of the most labor-intensive steps, is necessary so that subsequent learning algorithms can extract the maximum knowledge from the sample.

The procedure for architectural self-tuning and tuning of weight coefficients (synapses) of the Kohonen network includes the following steps.

1. An arbitrary number of neurons L = L0 is introduced with randomly normalized synapse vectors uniformly distributed over the interval [-1, 1].

2. One of the training vectors of input signals (obviously, pre-processed) is fed to the layer, the potentials at the outputs of all neurons and the number of the L*-neuron of the “winner” are determined.

3. The angle β* between the learning vector of features and the vector of synapses (weight coefficients) of the “winner” neuron is determined.

4. If the condition β* < β is met, then the synapses of the “winner” neuron are tuned by averaging over all learning steps at which it turned out to be the “winner” neuron and subsequent normalization.

If β* > β, then L+1 neuron is introduced into the layer in a directive order, the synapse of which (connection weights) assigns the values of the components of the corresponding training vector.

5. The next input vector of the training sample is selected and procedures 2, 3, 4 are repeated.

Training of the Grosberg layer is traditional supervised learning and can be performed both simultaneously with architectural self-organization and adjustment of the Kohonen layer for each input vector of the sample, and automatically (after training the Kohonen layer). In all cases, the learning rule can be represented by the following algorithm:

-

-й компонент желаемого выходного вектора классификатора;

- выходной сигнал j-го нейрона слоя Кохонена, при обучении S-му входному вектору признаков. where

-

-th component of the desired classifier output vector;

- the output signal of the j-th neuron of the Kohonen layer, when learning the S-th input feature vector.

Functioning of the recognition network.

на нейроны слоя Кохонена. Входные потенциалы нейронов

После этого слой Кохонена начинает функционировать как конкурирующая сеть с латеральными связями. The pre-processed vector of input features of the observed object {x _i } is fed to the input of the network and distributed through the weight coefficients of connections (synapses)

on the neurons of the Kohonen layer. Input potentials of neurons

After that, the Kohonen layer begins to function as a competing network with lateral connections.

вектор входа поступает на нейроны выходного слоя Гросберга, которые функционируют согласно следующего алгоритма:

где 10 ( … ) - функция единичного скачка.As a result, a vector is formed at the output of the layer with one unit component corresponding to the “winner” neuron and with zero other coefficients. Through synapses

the input vector goes to the neurons of the Grosberg output layer, which function according to the following algorithm:

where 10 ( … ) is the unit jump function.

In FIG. Figure 3 shows a neural network interpretation of the multidimensional classification algorithm.

In FIG. Figure 4 shows the structure of a Wang-Mendel fuzzy neural network with several outputs, which forms a fuzzy logical conclusion. Fuzzy logical inference is formed in several steps: introduction of fuzziness: at this stage, membership functions are applied to the actual values of input variables; logical inference: a truth value is computed for the premises of each rule and applied to the conclusions of each rule. This results in one fuzzy subset to be assigned to each output variable for each rule; composition: fuzzy subsets assigned to each output variable are combined into one set for all output variables; hardening: used in cases where it is necessary to convert a fuzzy set of inferences to a hard number.

. Второй слой выполняет агрегирование значений активации условия, используя одну из алгебр:In the first layer of the network, fuzzification of the input vector is performed

. The second layer performs aggregation of condition activation values using one of the algebras:

- конъюнкция Геделяone)

- Gödel conjunction

- конъюнкция Гогена2)

- Gauguin conjunction

- конъюнкция Лукасевича (трехзначная логика Я. Лукасевича)3)

- conjunction of Lukasiewicz (three-valued logic of J. Lukasiewicz)

The third layer contains a fuzzy implication:

- нечеткая импликация Геделяone)

- fuzzy Gödel implication

- нечеткая импликация Гогена2)

- Gauguin's fuzzy implication

- нечеткая импликация Лукасевича3)

- Lukasiewicz's fuzzy implication

правил вывода и генерация нормализующего сигнала. Используемые операции:In the fourth layer, aggregation is carried out

inference rules and generation of a normalizing signal. Operations used:

- дизъюнкция Геделяone)

- Gödel disjunction

- дизъюнкция Гогена2)

- Gauguin disjunction

- дизъюнкция Лукасевича3)

- Lukasiewicz disjunction

и

. The fifth layer consists of two output neurons, the third neuron is used for normalization, generating output signals

and

.

The output signal in general terms is determined by the expression corresponding to the following relationship:

.

.

- количество выходов.where

- the number of exits.

Instead of the operations of sum (∑), product (Π), and ipplication (→), it is supposed to substitute the corresponding operations from the chosen fuzzy algebra.

In FIG. Figure 5 shows the structure of a feed-forward neural network. The neurons of the first layer of the network normalize the input signals, mapping them to the range [0, 1], and also perform, where necessary, the logarithm of the normalized signal.

The neurons of the second layer group the signals transformed by the neurons of the first layer and find the weighted sum of the signals of the group, taking into account the estimates of their individual influence on the output feature (class number).

The neurons of the third layer of the network determine the distances from the recognized instance to the cluster centers in the space of grouped signals and find the value of the Gaussian function, whose argument is a certain distance. The neurons of the fourth layer correspond to classes and output the value "1" if the recognized instance belongs to the corresponding class, and "0" otherwise.

In FIG. 6 shows the structure of a six-layer neuro-fuzzy network. The discriminant functions of neurons of a six-layer neuro-fuzzy network will be given by the formulas:

- количество нейронов в

-ом слое сети.where

- the number of neurons in

th layer of the network.

The activation functions of neurons of a six-layer neuro-fuzzy network will be given by the formulas:

The weight coefficients of neurons of a six-layer neuro-fuzzy network will be determined by the formula:

The discriminant functions of neurons in a five-layer neuro-fuzzy network will be given by the formulas:

- количество нейронов в

-ом слое сети.where

- the number of neurons in

th layer of the network.

The activation functions of neurons of a six-layer neuro-fuzzy network will be given by the formulas:

The weight coefficients of neurons of a six-layer neuro-fuzzy network will be determined by the formula:

Shown in FIG. 7 the mask method is used for recognition by the amplitude-frequency characteristic.

In each mask, according to the real characteristic, the maximum amplitude value of the signal А ₁ , А ₂ , … , A _j , … , A _k is determined. The choice of the value of Δ, and, consequently, the number of masks is determined by the capabilities of the recognizing network (actually 10 ÷ 100).

An increase in the number of masks leads to an increase in the reliability of the input information and to an increase in the complexity (increase in the number of input layer neurons) of the recognizer, that is, there is a classic conflict between quality and complexity. It is possible to study the noise portrait in parts, that is, the low-frequency, mid-frequency and high-frequency components separately.

In FIG. 8 shows the interpretation of the three-dimensional output vector of recognition of hydroacoustic signals by the amplitude-frequency characteristic. In FIG. Figure 9 presents a comparative description of the methods for classifying a marine target by training and classification time. The fastest method in terms of classification quality is the modified method of potential functions based on neural networks (NN) (Fig. 9). NNs have stronger approximation abilities than other target classification methods, and the multivariate classification algorithm, unlike potential function methods, takes into account the significance of features by taking into account the partial significances of two-feature classifications.

In FIG. 10 and FIG. Figure 11 shows the results of a computational experiment to determine the recognition (classification) coefficient, defined as the ratio of the number of recognized objects to the total number of tests in percent, for surface and underwater objects in conditions of signal noise in the range from -10 to 20 dB. According to the results of the computational experiment, the recognition and classification of MCs using the computational operations of a six-layer neuro-fuzzy network and a modified combined recognition network consisting of Kohonen networks (competing network) and Grosberg networks and the MAXNET network can increase the probability of reliable classification of both surface and underwater targets.

A method for detecting and classifying marine targets based on neural network technologies and elements of artificial intelligence

The radiating transducer 1 and the receiving transducer 2 are placed in the marine environment 3, taking into account the patterns of multipath wave propagation in an extended hydroacoustic channel, which ensures the formation and efficient use of a spatially developed working zone 4 of nonlinear interaction and parametric transformation of transmissive waves and waves of various physical nature generated by objects 5 (see Certificate of state registration of the computer program No. 2018611591. Russian Federation. Calculation of the effectiveness of the use of hydroacoustic means and control of the simulation model of a hydroacoustic experiment. Application No. 2017662812: application 05.12.2017 : publ. (reg.) 02.02.2018. Bull. No. 2. / A.M. Vasilenko, V.A. Pyatakovich.; Certificate of state registration of the computer program No. 2019611559. Russian Federation. Program for the design and training of artificial neural networks of the perceptron type. Application No. 2019610135: application 10.01.2019 : published (for reg.) 01/29/2019. Bull. No. 2. / A.M. Vasilenko, V.A. Pyatakovich, O.A. Alekseev.; Certificate of state registration of the computer program No. 2017664296. Russian Federation. The program for simulation modeling of the process of propagation of hydroacoustic signals. Application No. 2017661087 : Appl. 10/31/2017 : publ. (registered) 12/20/2017. / A.M. Vasilenko, V.A. Pyatakovich; Certificate of state registration of the computer program No. 2018612169. Russian Federation. Software-analytical complex for determining the direction to the sea target by the secondary hydroacoustic field. Application No. 2017663202 : Appl. 12/19/2017 : publ. (reg.) 13.02.2018. Bull. No. 2. / A.M. Vasilenko, V.A. Pyatakovich; Certificate of state registration of the computer program No. 2017619627. Russian Federation. Calculation of the directional characteristic of a hydroacoustic antenna to ensure the recognition and classification of signs of information fields of marine technical objects. Application No. 2017616544 : Appl. 07/04/2017 : publ. (reg.) 29.08.2017 / A.M. Vasilenko, V.A. Pyatakovich, O.A. Alekseev.; Certificate of state registration of the computer program No. 2018612414. Russian Federation. Automated positioning of a sea target. Application No. 2017663557: Appl. 12/19/2017 : publ. (reg.) 16.02.2018. Bull. No. 2 / A.M. Vasilenko, V.A. Pyatakovich; Certificate of state registration of the computer program No. 2018612944. Russian Federation. Software-computer complex for simulation modeling of the marine information situation in the identification of targets. Application No. 2018610256 : Appl. 01/10/2018 : publ. (reg.) 03/01/2018. Bull. No. 3 / A.M. Vasilenko, V.A. Pyatakovich; Certificate of state registration of the computer program No. 2018613558. Russian Federation. The program for simulating the signal of the noise emission of submarines in various modes of motion. Application No. 2018610751 : Appl. 01/31/2018 : publ. (reg.) 16.03.2018. Bull. No. 3 / A.M. Vasilenko, V.A. Pyatakovich).

The pump signal of a stabilized frequency generated by the generator 6.1 is fed to the input of the power amplifier 6.2, the pump signal emission path 6, then to the input of the matching unit 6.3, the output of which is connected to an underwater cable connecting the output of the emission path of the pump signals 6 and the input of the radiating transducer 1. Emitter 1 sounds environment with pump signals of a stabilized frequency in the range of tens to hundreds of hertz. Pumping the controlled marine environment with complex frequency- or phase-modulated signals improves the efficiency of parametric reception of waves from acoustically subtle objects.

In various modes of movement, objects 5 generate radiation, leading to a change in the magnitude of the characteristics of the conductive liquid (density and (or) temperature and (or) heat capacity, etc.), which, depending on their physical nature, modulate low-frequency pump signals of the marine environment. Low-frequency and high-frequency components appear in the spectrum of nonlinearly converted waves (information waves) as a result of modulation of the amplitude and phase of the low-frequency pump wave by radiation and fields of objects processing and registration of information signals 7.

Information waves are received by antenna 2, amplified in the parametric conversion band (block 7.1), the frequency-time scale of the signals is transferred to the high-frequency region (block 7.2), narrow-band spectral analysis is performed (block 7.3) and the parametric components of the sum and difference frequencies are extracted from them, according to which restore the signs of the manifestation of fields generated by objects.

The operation of converting the frequency-time scale of the signal in block 7.2 of the path for receiving, processing and recording signals ensures an increase in the energy concentration of non-linearly converted signals and the efficiency of extracting from them signs of fields generated by marine targets.

The operations of spectral analysis in block 7.3 allow you to select the discrete components of the sum or difference frequency in the narrow-band spectra of nonlinearly converted signals, from which the wave characteristics of objects 5 are restored.

The amplitude-frequency characteristics of the signals of an object (surface or underwater object) obtained using narrow-band spectral analysis in the path for receiving, processing and recording signals 7 are fed to the first input of the target class recognition unit 8.1 according to the amplitude-frequency characteristics of the path of neural network recognition and classification 8.

The second input of the target class recognition block 8.1 according to the amplitude-frequency characteristics receives data from the training block 8.2 of the neural network recognition and classification path 8, the data from which is fed into the additionally introduced path for the synthesis of neural network and neuro-fuzzy recognition models with a grouping of features 9 to the input of the block for analyzing information about features and topology of the 9.1 training sample, and then fed to the input of the 9.2 neural network synthesis unit, covered by feedback from the 9.3 neuro-fuzzy networks synthesis unit, the output of which is connected to the input of the 9.4 logic device, the data from which is fed to the 8.2 learning unit, covered by feedback from block 8.1 for recognizing the target class by amplitude-frequency characteristics, the path of neural network recognition and classification 8, where the recognition of the class of a marine target (surface or underwater object) is performed. The analysis of the low-frequency, mid-frequency and high-frequency components of the amplitude-frequency characteristic is carried out separately, since the general features for different types of objects can be in different frequency ranges.

The learning process of a multilayer neural network is iterative and, in general, quite long, since it is impossible to determine in advance the number of iterations required to train the neural network.

используется следующее выражение:For a neural network implementation of comparing distances and determining the value

the following expression is used:

- логистическая функция. Если функция

будет дискретной, например, пороговой:where

- logistic function. If the function

will be discrete, for example, threshold:

, то

будет принимать значение 0 или 1.

, then

will take the value 0 or 1.

будет вещественной, например, сигмоидной:

, то

будет принимать значения на интервале [0,1], чем ближе значение этой функции будет к 0, тем ближе экземпляр будет к классу, которому сопоставлено значение 0, и, соответственно, наоборот, чем ближе значение этой функции будет к 1, тем ближе экземпляр будет к классу, которому сопоставлено значение 1. Использование сигмоидной функции может быть более предпочтительным на практике, поскольку она позволяет не только определить к какому классу ближе экземпляр, но и на сколько ближе.If the function

will be real, for example, sigmoid:

, then

will take values in the interval [0,1], the closer the value of this function is to 0, the closer the instance will be to the class to which the value 0 is associated, and, accordingly, vice versa, the closer the value of this function is to 1, the closer the instance will be to the class to which the value 1 is associated. Using the sigmoid function may be more preferable in practice, since it allows you to not only determine which class the instance is closer to, but also how much closer.

подставляются соответствующие выражения:To calculate the distance difference

the corresponding expressions are substituted:

After mathematical transformations, we get:

где

where

и

вычисляются на основе формального нейрона, имеющего один вход, на который подается значение

или

, вес которого равен

или

, соответственно. Порог нейрона (нулевой вес) в этом случае будет равен

или

соответственно. Expressions for

and

calculated on the basis of a formal neuron that has one input, to which a value is supplied

or

, whose weight is

or

, respectively. The neuron threshold (zero weight) in this case will be equal to

or

respectively.

-го нейрона

-го слоя:The rules for calculating the parameters of the multidimensional classification algorithm in this case will remain unchanged, and the parameters and activation functions of the neural network (NN) must be determined based on them according to the following rules. Activation function

-th neuron

th layer:

-го входа

-го нейрона

-го слоя:Weight coefficient

-th entrance

-th neuron

th layer:

. Распознающую сеть блока 8.1 распознавания класса цели по амплитудно-частотным характеристикам тракта нейросетевого распознавания и классификации 8, настраивают по классификационным признакам полей, генерируемых морскими целями, с помощью блока обучения 8.2 тракта нейросетевого распознавания и классификации 8, в который входят априори полученные шумовые портреты вероятных объектов распознавания и устройство подготовки данных.The values of the output signals of the third layer form the decision vector

. The recognition network of the target class recognition block 8.1 according to the amplitude-frequency characteristics of the neural network recognition and classification path 8 is tuned according to the classification features of the fields generated by sea targets using the neural network recognition and classification path training block 8.2 8, which includes a priori obtained noise portraits of probable objects recognition and data preparation device.

The weight coefficients of the recognizing network of block 8.1, the recognition of the target class by amplitude-frequency characteristics are corrected using the learning block 8.2, the path of neural network recognition and classification 8. Next, the data is fed into the additionally introduced path for the synthesis of neural network and neuro-fuzzy models with grouping of features 9 at the input of the information analysis block about the features and topology of the training sample 9.1, and then fed to the input of the neural network synthesis block 9.2 covered by feedback with the neuro-fuzzy networks synthesis block 9.3, the output of which is connected to the input of the 9.4 logic device, the function of which is performed by a six-layer Wang-Mendel neural fuzzy network, where in automatic mode, on the basis of training data samples, the structure of neural and neuro-fuzzy networks is formed, and their parameters are adjusted, without optimization fitting of weights, neural and neuro-fuzzy models are synthesized in a non-iterative mode with linearization, factorial grouping and convolution signs of the sea target and a fuzzy logical conclusion is formed for the training block 8.2 of the path of neural network recognition and classification 8, after which the artificial neural network is tuned and a conclusion is formed about the degree of belonging of the studied region of the spectrum to the object of classification (surface or underwater object).

Algorithm for the operation of the logical device of the synthesis path of neural network and neuro-fuzzy recognizing models with feature grouping.

.Initialization stage. Set Initial Data Sample

.

- соответственно, минимальное и максимальное значения

-го признака,

. Определить число интервалов для каждого признака:

, а также длины интервалов:

.The stage of analysis of sample characteristics. Define

- respectively, the minimum and maximum values

-th sign,

. Determine the number of intervals for each feature:

, as well as the lengths of the intervals:

.

-го экземпляра,

: определить

- номер интервала значений по каждому

-му признаку,

, в который попадает

-й экземпляр

рассчитать координату

-го экземпляра по обобщенной осиStage of calculation of generalized features. For everybody

-th instance,

: define

- the number of the interval of values for each

-th sign,

, into which

-th copy

calculate coordinate

-th instance along the generalized axis

(заметим, что при этом произойдет потеря части информации вследствие неявного квантования пространства признаков при преобразовании).This allows you to display the original sample on a one-dimensional generalized axis

(Note that in this case, some information will be lost due to the implicit quantization of the feature space during the transformation).

. Упорядочить набор

в порядке неубывания значений

. Просматривая обобщенную ось в порядке увеличения ее значений, определить граничные значения ее интервалов

, в которых номер класса

остается неизменным, где

- соответственно левое и правое граничные значения

-го интервала обобщенной оси. Обозначим:

- номер класса, соответствующий

-му интервалу обобщенной оси;

- число интервалов обобщенной оси.Stage of analysis of the generalized axis. Form a set of tuples

. Arrange set

in non-decreasing order of values

. Looking through the generalized axis in increasing order of its values, determine the boundary values of its intervals

, in which the class number

remains unchanged, where

- respectively, the left and right boundary values

th interval of the generalized axis. Denote:

- class number corresponding to

-th interval of the generalized axis;

- number of intervals of the generalized axis.

-го интервала обобщенной оси,

, определить

- число попавших в него экземпляров, а также номера этих экземпляров.The stage of analysis of the characteristics of the intervals. For everybody

-th interval of the generalized axis,

, define

- the number of copies that fell into it, as well as the numbers of these copies.

-го интервала включить в обучающую выборку

все экземпляры:The stage of formation of the training sample. Among instances

-th interval to include in the training sample

all instances:

its class, located on one of the boundaries of the interval

of its class closest to one of the boundaries of the interval:

- пороговый коэффициент, регулирующий близость экземпляров к границам интервала (например, можно задать:

);where

- threshold coefficient that regulates the proximity of instances to the boundaries of the interval (for example, you can set:

);

intervals with a small number of instances:

- некоторый пороговый коэффициент,

(например, можно задать:

);

- среднее число экземпляров в интервале обобщенной оси.where

- some threshold coefficient,

(for example, you can set:

);

- the average number of instances in the interval of the generalized axis.

(для упрощения и ускорения вычислений будем оперировать квадратами расстояний):The stage of eliminating the redundancy of the training sample. Determine the distances between all instances included in the generated training sample by forming a distance matrix

(to simplify and speed up calculations, we will operate with squared distances):

Note that

, необходимо выполнять в цикле действия: найти в матрице расстояний два экземпляра с наименьшим расстоянием между собой:Until

, it is necessary to perform in the action loop: find two instances in the distance matrix with the smallest distance between them:

if the two closest instances belong to the same class, then leave in the training sample only one of them that is closer to instances of other classes, and exclude the other from it

, установив:

если два ближайших экземпляра принадлежат к разным классам, то перейти к выполнению этапа дополнения (уточнения) обучающей выборки. Этап дополнения (уточнения) обучающей выборки. Определить разность исходной и сформированной выборок

Adjust matrix elements accordingly

by setting:

if the two nearest instances belong to different classes, then proceed to the completion (refinement) stage of the training sample. The stage of supplementing (refinement) of the training sample. Determine the difference between the original and generated samples

-го экземпляра

выборки

,

относительно экземпляров сформированной выборки

найти расстояние (квадрат расстояния) от него до каждого экземпляра выборки

:consecutively for each

-th copy

samples

,

relative to the instances of the generated sample

find the distance (squared distance) from it to each instance of the sample

:

-му экземпляру экземпляр сформированной выборки принадлежит к другому классу, то включить его в выборку

:if closest to

-th instance, an instance of the generated selection belongs to another class, then include it in the selection

:

получаем сформированную обучающую выборку

, которая будет обладать основными топологическими свойствами исходной выборки. При этом из исходной выборки можно получить также тестовую выборку как разность исходной и сформированной обучающей выборок. Разбиение признакового пространства для выборки эмпирических наблюдений необходимо для определения нечетких термов признаков как проекций соответствующих блоков на координатные оси. Формирование разбиения признакового пространства необходимо осуществлять путем их выполнения в приведенной ниже последовательности.As a result of executing this method for the initial sample

we get the formed training sample

, which will have the main topological properties of the original sample. At the same time, a test sample can also be obtained from the original sample as the difference between the original and the formed training samples. Partitioning the feature space for a sample of empirical observations is necessary to determine the fuzzy terms of the features as projections of the corresponding blocks onto the coordinate axes. The formation of a partition of the attribute space must be carried out by performing them in the following sequence.

. Шаг 2. По оси каждого признака

определить одномерные расстояния между экземплярами:

Среди полученных расстояний найти минимальное расстояние, большее нуля:Step 1. Initialization. Set training sample

. Step 2. Along the axis of each feature

define one-dimensional distances between instances:

Among the obtained distances, find the minimum distance greater than zero:

Step 3. For each attribute, determine the number of intervals for dividing the range of its values:

Шаг 4. Разбить ось

-го признака на

интервалов. Определить координаты левых и правых границ для каждого

-го интервала

-го признака по формулам:and also determine the length of the interval of observed values of each feature:

Step 4. Split the axle

-th sign on

intervals. Determine the coordinates of the left and right borders for each

-th interval

-th sign according to the formulas:

, в

-мерном пространстве признаков на пересечении соответствующих интервалов значений признаков. Занести в

номер интервала

-го признака, который соответствует

-му блоку. Шаг 5.2. Определить номера классов для прямоугольных блоков в

-мерном пространстве признаков:Step 5. Form block-clusters and set the numbers of their classes by performing steps 5.1 - 5.8. Step 5.1. Form rectangular blocks

, in

-dimensional space of features at the intersection of the corresponding intervals of feature values. Bring in

interval number

-th feature that corresponds to

-th block. Step 5.2. Define class numbers for rectangular blocks in

-dimensional feature space:

Set classification confidence factor for blocks:

, установить:

Step 5.3. For those blocks that have

, install:

- количество экземпляров

-го класса, попавших в

-й блок-кластер.where

- number of copies

-th class, caught in

th block cluster.

,

, определить расчетный номер класса, для чего предлагается использовать модифицированный нерекуррентный метод потенциальных функций. Шаг 5.5. Вычислить расстояние между

-м и

-м блоками:Step 5.4. For those blocks that have a class number

,

, to determine the calculated class number, for which it is proposed to use a modified non-recurrent method of potential functions. Step 5.5. Calculate distance between

-m and

-m blocks:

,

,

.how:

.

где

. Or

where

.

Шаг 5.6. Определить потенциал, наводимый совокупностью блоков, принадлежащих к

-му классу, на

-й блок с неизвестной классификацией:Wherein

Step 5.6. Determine the potential induced by a set of blocks belonging to

-th class, on

-th block with unknown classification:

- количество блоков, принадлежащих к

-му классу,

- количество экземпляров обучающей выборки, попавших в

-й блок. Шаг 5.7. Установить номер класса для

-го блока с неизвестной классификацией

по формуле:where

- the number of blocks belonging to

-th class,

- the number of instances of the training sample that fell into

-th block. Step 5.7. Set class number for

th block with unknown classification

according to the formula:

Step 5.8. Modify values of confidence coefficients for blocks:

Step 6. Merge adjacent blocks-clusters. Merge adjacent blocks belonging to the same class:

: если

,

for

: if

,

and

и

по

-му признаку:then merge the blocks

and

on

-th sign:

- install:

-й блок:

.- delete

-th block:

.

,

,

,

,

;

; то необходимо из набора новых наблюдений исключить те наблюдения, которые попадают в блоки имеющегося разбиения и соответствуют им по номеру класса, скорректировав соответствующим образом

. Для тех наблюдений, которые не совпадают с классами блоков, целесообразно сформировать отдельные точечные кластеры. Step 7. From the combined set (OM), select a subset of instances belonging to cluster blocks, the class numbers of which do not match the class numbers of the instances. Apply for the resulting partition and the selected subset the procedure for refining the partition and retraining the model. Step 8. Stop. Refining the partition and retraining the model. If there is a partition of the feature space that needs to be refined (additionally trained) based on new observations

,

,

,

,

;

; then it is necessary to exclude from the set of new observations those observations that fall into the blocks of the existing partition and correspond to them by class number, adjusting accordingly

. For those observations that do not coincide with block classes, it is advisable to form separate point clusters.

, номера интервалов для каждого

-го признака, соответствующие новому кластеру, а также определить: For each new observation, form intervals according to features and enter them into

, interval numbers for each

-th feature corresponding to the new cluster, and also determine:

- некоторая константа,

Для определения целесообразности применения предложенного метода для конкретной задачи на практике используем нотацию Ландау в так называемом «мягком виде» и оценим сложность этапов предложенного метода. Для этапа инициализации вычислительной сложностью можно пренебречь, а пространственная сложность может быть оценена как

.where

- some constant,

To determine the expediency of applying the proposed method for a specific problem in practice, we use the Landau notation in the so-called “soft form” and estimate the complexity of the steps of the proposed method. For the initialization step, the computational complexity can be neglected and the space complexity can be estimated as

.

, а пространственная -

. Для этапа расчета обобщенных признаков вычислительная сложность может быть оценена как

, а пространственная -

. Для этапа анализа обобщенной оси вычислительная сложность может быть оценена как

, а пространственная -

. Для этапа анализа характеристик интервалов вычислительная сложность может быть оценена как

, а пространственная -

. Для этапа формирования обучающей выборки вычислительная сложность может быть оценена как

, а пространственная в виде

. Для этапа устранения избыточности ОВ вычислительная и пространственная сложность могут быть оценены соответственно как

. Для этапа дополнения (уточнения) обучающей выборки вычислительная и пространственная сложность оцениваются соответственно как

. For the stage of analyzing the characteristics of the sample, the computational complexity will be

, and the spatial

. For the stage of calculating generalized features, the computational complexity can be estimated as

, and the spatial

. For the generalized axis analysis step, the computational complexity can be estimated as

, and the spatial

. For the interval characteristics analysis stage, the computational complexity can be estimated as

, and the spatial

. For the training sample formation stage, the computational complexity can be estimated as

, and spatial in the form

. For the OB redundancy elimination stage, the computational and space complexity can be estimated, respectively, as

. For the stage of supplementing (refinement) of the training sample, the computational and spatial complexity are estimated, respectively, as

.

; пространственная -

. Thus, the overall complexity of the method can be estimated as: computational -

; spatial -

.

The algorithm for correcting the settings of the synthesis path of neural network and neuro-fuzzy recognizing models with feature grouping is implemented to find the number of a fuzzy rule of the form:

)(

)

.in the presence of a training set

.

используется алгоритм нечеткого вывода, применяются предикатные правила:To simulate an unknown mapping

fuzzy inference algorithm is used, predicate rules are applied:

,

,

- нечеткие множества описывающие высказывания: «отрицательная», «нулевая», «положительная», «малое», «среднее», «большое» и т.д.

- вещественные числа (номер правила). Степень истинности

правила

- определяется с помощью операции умножения Ларсена (Larsen):where

- fuzzy sets describing statements: "negative", "zero", "positive", "small", "medium", "large", etc.

- real numbers (rule number). Degree of truth

regulations

- is determined using the Larsen multiplication operation:

определяется методом центра тяжести:

for modeling the logical operator "AND", the output of a fuzzy system

determined by the center of gravity method:

позволяет использовать градиентный метод для подстройки параметров заданных предикатных правил, а величина

корректируется по соотношению:Error function for the i-th presented value of the form:

allows you to use the gradient method to adjust the parameters of the given predicate rules, and the value

corrected according to the ratio:

.

.

- константа, характеризующая скорость обучения сети. Аналогичным образом определяются параметры функции принадлежности.where

is a constant characterizing the network learning rate. The parameters of the membership function are defined in a similar way.

и

). In a situation of uncertainty, i.e. the influence of random external and parametric disturbances, the logical device 9.4 compensates for them in the rule base, as well as compensation for membership functions of a new type (with other universes

and

).

Thus, the path for the synthesis of neural network and neuro-fuzzy recognition models with feature grouping makes it possible to synthesize neural and neuro-fuzzy models in a non-iterative mode with linearization, factorial grouping, and feature convolution. Logic device 9.4 of the synthesis path of neural network and neuro-fuzzy recognizing models with grouping of features based on training data samples allows you to create the structure of neural and neuro-fuzzy networks, as well as adjust their parameters without optimizing weights and automatically correct the data of the updated library of mathematically processed images of spectrograms of sea targets path of neural network recognition and classification 8, which provides the final classification solution for detected marine targets (surface or underwater object).

Since the bases of fuzzy rules are often characterized by a large volume, the most urgent task is to combine and transform fuzzy rules, the essence of which is to form a new rule base of a smaller volume based on the initial set of fuzzy rules, which would sufficiently represent the initial base and be would be less redundant.

To solve the problem of forming fuzzy rules, it is necessary to bring the stages of the method in line with the features of the problem being solved, which is presented as a sequence of steps 1-17.

. Для каждого из возможных классов выходных значений создается свое пространство поиска, и, соответственно, свое отдельное множество агентов, отдельный граф поиска, представляющий собой лингвистические термы, которые могут быть включены в правила, а также для каждого пространства поиска рассчитываются эвристические значения значимости отдельного терма. Step 1. Initialization. The static parameters of the method operation are set: coefficients

. For each of the possible classes of output values, its own search space is created, and, accordingly, its own separate set of agents, a separate search graph representing linguistic terms that can be included in the rules, and for each search space, heuristic significance values of a separate term are calculated.

- значение эвристической значимости лингвистического терма

для описания класса

;

- экземпляр входной выборки, содержащей

экземпляров;

- значение функции принадлежности объекта

терму

и классу q, соответственно;

- количество лингвистических термов;

- количество классов. В каждом пространстве поиска каждому узлу графа поиска ставится в соответствие начальное значение количества дискретных составляющих (ДС)

где

- значение количества ДС для

-го терма в пространстве поиска для

-го класса на первой итерации поиска.where

- the value of the heuristic significance of a linguistic term

to describe a class

;

- an instance of the input sample containing

copies;

- the value of the object membership function

termu

and class q, respectively;

- number of linguistic terms;

- the number of classes. In each search space, each node of the search graph is associated with the initial value of the number of discrete components (DS)

where

- the value of the number of DS for

-th term in the search space for

th class at the first iteration of the search.

. Шаг 3. Установить:

. Шаг 4. Установить:

. Шаг 5. Установить:

. Шаг 6. Выбор терма для добавления в правило

-го агента в пространстве поиска

-го класса. Шаг 6.1. Для

-го агента на основе правила случайного выбора рассчитывается вероятность включения

-го лингвистического терма в правило, описывающего

-й класс выходного значения

где

- вероятность добавления

-го терма в правило

-го агента в пространстве поиска для

-го класса;

- множество термов, которые могут быть добавлены в правило

-го агента. Формирование данного множества определяет вид правил, которые могут составляться в процессе поиска, то есть предполагается, что правило может включать выражения типа ИЛИ. После добавления терма из данного множества исключается только данный терм, если же предполагается, что правило не может включать выражения типа ИЛИ, то кроме выбранного терма, исключаются и все термы, описывающие данный атрибут. Step 2 Install:

. Step 3 Install:

. Step 4 Install:

. Step 5 Install:

. Step 6. Selecting a term to add to the rule

-th agent in the search space

th class. Step 6.1. For

th agent, based on the random selection rule, the probability of including

-th linguistic term in the rule, describing

-th output value class

where

- probability of adding

-th term in the rule

-th agent in the search space for

-th class;

- many terms that can be added to the rule

th agent. The formation of this set determines the type of rules that can be compiled in the search process, that is, it is assumed that the rule may include expressions of the OR type. After adding a term, only this term is excluded from this set, but if it is assumed that the rule cannot include expressions of the OR type, then, in addition to the selected term, all terms describing this attribute are also excluded.

где

- случайное число из интервала [0; 1]. Если условие выполняется, тогда лингвистический терм

добавляется в правило

-го агента, удаляется из множества возможных термов для данного агента и выполняется переход к шагу 7. В противном случае - переход к шагу 6.3.Step 6.2. Check Condition

where

- random number from the interval [0; one]. If the condition is met, then the linguistic term

added to rule

-th agent is removed from the set of possible terms for this agent and go to step 7. Otherwise, go to step 6.3.

. Шаг 6.4. Если были рассмотрены все термы, то установить:

. Переход к шагу 6.1. Шаг 7. Проверка завершения перемещения

-го агента. Шаг 7.1. Если множество термов, которые

-й агент может добавить в формируемое правило, пусто, то выполняется переход к шагу 8. Шаг 7.2. Определяется, сколько экземпляров

-го класса покрывает правило

-го агента. Шаг 7.2.1. Для экземпляра o, относящегося к классу

, рассчитывается степень соответствия сформированного правила

Step 6.3. Install

. Step 6.4. If all terms have been considered, then set:

. Go to step 6.1. Step 7: Verify that the move is complete

th agent. Step 7.1. If the set of terms that

-th agent can add to the generated rule, is empty, then go to step 8. Step 7.2. Determine how many instances

-th class covers the rule

th agent. Step 7.2.1. For an instance o belonging to a class

, the degree of compliance of the formed rule is calculated

- степень соответствия между правилом

-го агента

и экземпляром

;

- мера соответствия между

-м атрибутом в правиле

и соответствующим атрибутом экземпляра

. Данная мера рассчитывается следующим образом:where

- the degree of correspondence between the rule

-th agent

and copy

;

- measure of correspondence between

-th attribute in the rule

and the corresponding instance attribute

. This measure is calculated as follows:

- отдельный терм, относящийся к области описания атрибута

; where

- a separate term related to the attribute description area

;

- количество термов, относящихся к области описания атрибута

.

- the number of terms related to the attribute description area

.

Step 7.2.2. Check condition:

- заданный параметр, который определяет, какое минимальное значения соответствия является достаточным, чтобы считать, что правило

в достаточной степени описывает объект

. Если условие выполняется, то считается, что данный объект o покрывается правилом

.where

is a given parameter that specifies what minimum match value is sufficient to consider that the rule

adequately describes the object

. If the condition is met, then it is considered that the given object o is covered by the rule

.

, и на основании получаемых данных увеличивается счетчик

, в котором хранится количество экземпляров, покрываемых правилом

. Steps 7.2.1 and 7.2.2 are performed for all instances belonging to the class

, and based on the received data, the counter is incremented

, which stores the number of instances covered by the rule

.

где

- предельное минимальное количество экземпляров

-го класса, которое должно покрываться правилом. Если указанное условие выполняется, то считается, что правило покрывает необходимое количество экземпляров, и

-й агент завершил свое перемещение, выполняется переход к шагу 8. Step 7.3. Check condition:

where

- maximum minimum number of instances

th class to be covered by the rule. If the specified condition is met, then the rule is considered to cover the required number of instances, and

The th agent has completed its move, go to step 8.

Otherwise, go to step 5.

, то установить:

и выполнить переход к шагу 5. В противном случае - переход к шагу 9. Step 8. If

then install:

and go to step 5. Otherwise, go to step 9.

, то установить:

и выполнить переход к шагу 4. В противном случае - переход к шагу 10. Шаг 10. Случайным образом формируются базы правил. Создается

баз правил, при этом для описания каждого класса выходного значения выбирается одно правило из соответствующего пространства поиска.Step 9. If

then install:

and go to step 4. Otherwise, go to step 10. Step 10. Rule bases are generated randomly. Created

rule bases, and one rule from the corresponding search space is selected to describe each class of the output value.

Step 11. Evaluation of the quality of the generated rule bases. To assess the quality of rule bases, an input training sample is used, for each instance of which the rule with the highest degree of coincidence is selected, on the basis of which the calculation class to which this instance belongs is determined based on the corresponding rule base.

где

- количество экземпляров, для которых класс был определен неверно с помощью заданной базы правил;

- качество прогнозирования класса экземпляров на основе соответствующей базы правил.Based on the data obtained using the knowledge base, based on the given training sample, we calculate the quality score of the rule base:

where

- the number of instances for which the class was determined incorrectly using the given rule base;

- the quality of predicting a class of instances based on the corresponding rule base.

где

- качество прогнозирования базы знаний, которая характеризуется наилучшей точностью прогнозирования;

- приемлемое качество прогнозирования. Если указанное условие выполняется, то производится переход к шагу 17, в противном случае - переход к шагу 13. Шаг 13. Добавление дискретных составляющих (ДС). Добавление ДС выполняется с целью повышения приоритетности тех термов, включение которых в правила способствует повышению качеству прогнозирования результирующих баз правил. Step 12. Check condition:

where

- the quality of forecasting the knowledge base, which is characterized by the best forecasting accuracy;

- acceptable quality of forecasting. If the specified condition is met, then go to step 17, otherwise go to step 13. Step 13. Adding discrete components (DS). The addition of DS is carried out in order to increase the priority of those terms, the inclusion of which in the rules improves the quality of prediction of the resulting rule bases.

где

- коэффициент, определяющий, насколько близко качество прогнозирования базы правил

должно приближаться к лучшему качеству прогнозирования

, чтобы можно было применять процедуру добавления ДС для правил, входящих в данную базу правил

. Таким образом, добавление ДС выполняется для каждого терма, входящего в правило, которое, в свою очередь, входит в базу правил

.

где

- количество ДС для терма

в пространстве поиска для класса

, который определяется с помощью соответствующего правила.In this regard, the number of added priority coefficient is directly proportional to the quality of prediction of the base of the rule, which includes the given fuzzy rule. At the same time, it is proposed to add DS only for those terms included in the rules of the fuzzy rule bases for which the following condition is satisfied:

where

- a coefficient that determines how close the quality of the rule base prediction is

should approach the best prediction quality

, so that you can apply the procedure for adding DS for the rules included in this rule base

. Thus, adding a DS is performed for each term included in the rule, which, in turn, is included in the rule base

.

where

- number of DS for the term

in the search space for the class

, which is determined using the corresponding rule.

где

- коэффициент исключения, который задается при инициализации.Step 14. Exclusion of DS. To eliminate the worst terms, that is, those that, when included in the rules, lower the quality of prediction using the corresponding rule, the DS elimination procedure is used, which is performed at the end of each iteration and is applied to all nodes in all search graphs. The exclusion of the DS of a sea target (MC) is carried out in accordance with the formula:

where

- exclusion factor, which is set during initialization.

, тогда установить:

и выполнить переход к шагу 16, в противном случае - считается, что выполнено максимально допустимое количество итераций, и выполняется переход к шагу 17. Шаг 16. Перезапуск агентов. Все данные обновляются, агенты размещаются в случайные точки пространств поиска. Переход к шагу 3. Шаг 17. Останов.Step 15 If

then install:

and go to step 16, otherwise, it is considered that the maximum allowable number of iterations has been completed, and go to step 17. Step 16. Restart agents. All data is updated, agents are placed at random points in search spaces. Go to step 3. Step 17. Stop.

On the basis of the experiments, fuzzy rule bases were obtained, which were characterized by an MC classification accuracy of 85.6%, which ensures the formation of a reduced fuzzy rule base, which is characterized by fairly accurate results, while reducing the time spent on calculation. An essential advantage of this method is that the search is performed simultaneously on several graphs, which allows you to control the influence of terms on the quality of prediction of recognition of different classes of MC. During the operation of the method, the data of the updated library of mathematically processed images of the MC spectrograms, the statistics of the classification characteristic are constantly corrected, the description of the MC classes for the current hydrological-acoustic and noise signal conditions is refined, on the basis of which the actual threshold for classification is formed.

Thus, by detecting the target by the signs of amplitude-phase modulation of low-frequency pumping signals of the marine environment by the radiations and fields of the object, and using the modified combined recognition network consisting of the Kohonen, Grosberg networks and the MAXNET network to perform recognition operations, the path of neural network recognition and classification of the marine target , the settings of which are regulated by the synthesis path of neural network and neuro-fuzzy recognizing models with feature grouping in the form of a six-layer neuro-fuzzy Wang-Mendel network, which allows, based on training data samples, to create a structure of logically transparent neural and neuro-fuzzy networks, as well as to tune their parameters without optimization fitting of weights, to synthesize neural and neuro-fuzzy models in a non-iterative mode with linearization, factorial grouping and convolution of features, for recognition of the MC class by the path of neural network recognition and classification of a sea target by amplitude-frequency character logists in an automated mode, making a decision on the degree of belonging of the studied region of the spectrum to the object of classification (surface or underwater object). According to the results of experimental data, the classification efficiency of marine targets in the claimed method is increased by 5-7% relative to the prototype and other classification methods using modulation of low-frequency pumping signals of the marine environment by radiation and object fields. The method is industrially applicable, since common components and products of the radio engineering industry and computer technology are used for its implementation.

Claims (1)

A method for detecting and classifying marine targets based on neural network technologies and elements of artificial intelligence, which consists in first forming in the marine environment a zone of nonlinear interaction and parametric conversion of pump waves with object signals, for which the emitter and receiving antenna are placed on opposite boundaries of the controlled area of the sea medium, then the pump waves, modulated by the signals of the object, are received and amplified in the parametric conversion band, their frequency-time scale is transferred to the high-frequency region, a narrow-band spectral analysis is performed, the parametric components of the sum or difference frequency are isolated, by which, taking into account the temporal and parametric conversion of the waves restore the characteristics of the object's signals, which are fed into the neural network recognition and classification path, consisting of a target class recognition unit based on amplitude-frequency characteristics, which implements computational operations of an artificial neural network and covered by feedback with a learning unit, in whose memory data of mathematically processed images of spectrograms of sea targets are recorded, and the first input of the target class recognition unit by amplitude-frequency characteristics receives data from the output of the spectrum analyzer, covered by feedback from the reception path recorder , processing and registration of signals, and its second input receives data from the training block of the path of neural network recognition and classification, characterized in that further data is fed to the input of the block for analyzing information about the features and topology of the training sample of the additionally introduced path for the synthesis of neural network and neuro-fuzzy recognition models with grouping of features that are fed to the input of the neural network synthesis unit, covered by feedback from the neuro-fuzzy networks synthesis unit, the output of which is connected to the input of a logical device, the function of which is performed by a six-layer Wang neuro-fuzzy network -Mendel, where neural and neuro-fuzzy models are synthesized in a non-iterative mode with linearization, factorial grouping and convolution of sea target features and a fuzzy logical conclusion is formed for the training block of the neural network recognition and classification path, after which the artificial neural network is tuned and a conclusion is formed about the degree of belonging to the studied area of the spectrum to the object of classification (surface or underwater object).

Images