RU2763125C1 - Method for operational identification of marine targets by their information fields based on neuro-fuzzy models - 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. The expected result of the invention is the automation of the process of recognizing classes of marine targets detected by the signs of amplitude-phase modulation of low-frequency signals of pumping the marine environment with radiation and fields of objects, a comprehensive reduction in the dimension of data with automatic adjustment of the rule base due to the formation and reduction of the retrieval of reference samples of mathematical models of marine targets, carried out using the path of regulation of parameters of formation and reduction of samples, necessary for the implementation of the final classification process in the neural network recognition and classification path, which provides an increase in the probability of reliable operational classification of a marine target by 7-9% greater than when using a prototype.

EFFECT: automation of the process of recognizing classes of marine targets.

1 cl, 8 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 by the signs of amplitude-phase modulation of low-frequency signals of pumping the marine environment by radiation and fields of objects.

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 natural nonlinear properties of the marine environment.

Known methods and implementing them parametric antennas using high-frequency pumping of the marine environment (see Novikov B.K., Timoshenko V.I. Parametric antennas in sonar systems. - L .: Shipbuilding. - 1990. - S. 17-40, 203 -225; Mironenko M.V., Malashenko A.E., Karachun L.E., Vasilenko AM Low-frequency transmissive method of long-range sonar location of hydrophysical fields of the marine environment: monograph - Vladivostok: SKB SAMI FEB RAS, 2006. - 173 p .; Pyatakovich V.A., Vasilenko AM, Filippova A.V. The technology of creating an automated system for long-range reception and neural network classification of hydrophysical fields in the sea area. - Bulletin of the Tula State University. Technical sciences. Issue 7. Tula: Publishing house of Tula State University, 2017. - P. 247-258; Pyatakovich V.A., Vasilenko AM, Rychkova V.F. Intelligent system of neural network classification of sea targets. - Marine intelligent technologies. - SPb .: 2018. No. 2 (40). Volume 2. - P . 115-120; Pyatakovich V.A., Rychkova V.F., Filippov E.G. Neurons e networks as a variant of the computational structure of the naval target classification system. - Marine Intelligent Technologies. - SPb .: 2020. No. 1 (47) Volume 2. - P. 163-174; Pyatakovich V.A., Surov A.B., Rychkova V.F. Calculation of the efficiency of target classification by an intelligent system of the Navy, using a complex of computational operations of neural networks. - Marine Intelligent Technologies. - SPb .: 2020. No. 1 (47) Volume 2. - P. 175-185; Pyatakovich V.A., Nikolaev A.V., Kostikov E.A. Automation of information processing in an intelligent marine monitoring system. - Problems of mechanical engineering and automation. - M .: 2020. No. 4. - S. 72-79).

Known methods for receiving an elastic wave in a marine environment, increasing the value of the parameter of nonlinearity of the medium in the signal receiving area. For this, an additional electromagnetic signal, subjected to time-frequency modulation, or modulated hydrodynamic fluid flow, is introduced into the parametric signal reception zone, in addition to the elastic wave. Modulation of the hydrodynamic fluid flow is carried out by changing the density due to controlled gas saturation, or changing the temperature, or changing its chemical composition, (Pat. No. 2158029 RF. Method for receiving an elastic wave in a marine environment (options); Publ. 20.10.2000. Bull. No. 29 .; pat. No. 2452041 RF. Method of parametric reception of waves of various physical nature in the marine environment. Publ. 05/27/2012. Bull. No. 15.). The disadvantages of the methods are low sensitivity and, as a result, limited range (units of kilometers) of parametric reception of information waves of various physical nature in infrasonic and fractional (units-fractions of hertz) frequency ranges.

The closest in technical essence to the claimed invention is "Method of detection and neural network recognition of the features of fields of different physical nature, generated by sea targets" (US Pat. No. 2682088 C1 (RF, IPC G01S 15/02 (2006.01); Publ. 03/14/2019. Bulletin) No. 8), which consists in the fact that first a zone of nonlinear interaction and parametric conversion of pump waves with object signals is formed in the marine environment, for which the emitter and the receiving antenna are placed on opposite boundaries of the monitored area of the marine environment, then the pump waves modulated with the object signals , are received and amplified in the parametric transformation band, their time-frequency scale is transferred to the high-frequency region, a narrow-band spectral analysis is carried out, the parametric components of the total or difference frequency are isolated, according to which, taking into account the temporal and parametric transformation of waves, the characteristics of the object signals are restored, which are fed to the tr an act of neural network recognition and classification, consisting of a target class recognition unit based on amplitude-frequency characteristics, which implements computational operations of an artificial neural network and is covered by feedback with a learning unit, in the memory of which data of mathematically processed images of spectrograms of sea targets are recorded, and to the first input of the recognition unit of the target class in terms of amplitude-frequency characteristics, data is received from the output of the spectrum analyzer of the signal receiving, processing and recording path, and its second input receives data from the training unit of the neural network recognition and classification path, after which the artificial neural network is tuned according to the classification features of the fields generated by marine targets , start computing operations, according to the results of which its weight coefficients are corrected and a conclusion is drawn about the degree of belonging of the investigated spectral region 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 difference in frequencies (within the same order of magnitude) 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

where γ is the coefficient of nonlinearity of the marine environment; ω _n , ω _c - frequency of the pump wave and the useful signal, respectively; P _n , P _c - damping of the pump wave and the useful signal, respectively; V is the volume of the medium of nonlinear interaction and parametric transformation of waves; R is the distance from the point of radiation to the point of location of the object; ρ ₀ is the density, c ₀ is the speed of sound in the marine environment.

The parametric components of the total and difference frequencies formed as a result of the transformation of transmission waves during the processing of broadband signals are distinguished as signs of amplitude-phase modulation, which is substantiated by mathematical dependences and confirmed by the results of marine experiments (see Mironenko M.V., Malashenko A.E., Karachun L.E., Vasilenko AM Low-frequency transmissive method of long-range sonar location of hydrophysical fields of the marine environment: monograph .-- Vladivostok: SKB SAMI FEB RAS, 2006 .-- 173 p .; Pyatakovich V.A., Vasilenko AM, Mironenko M.V .; Technologies nonlinear transmissive hydroacoustics and neuro-fuzzy operations in problems of recognition of sea objects: - monograph. - Vladivostok: Far Eastern Federal University, 2016 190 p. ISBN 978-5-7444-3790-9 .; Mironenko M.V., Pyatakovich V.A., Vasilenko AM Results of experimental studies of the method for determining the profile of a marine object and the system that implements it. - Monitoring. Science and technology. 2017. - No. 2 (31) - P. 64-6 9.).

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

P - результирующее и мгновенное значения давления модулированной волны, соответственно; 2ω_c - удвоенная частота модулированной волны; Ω - волна, генерируемая объектом; t - время; J_n - функции Бесселя n-го порядка; A_n - амплитуда модулированной волны; m_p - коэффициент модуляции. Как видно из выражения, значения частот боковых составляющих отличаются от удвоенной центральной частоты 2ωо (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

P - the resulting and instantaneous values of the pressure of the modulated wave, respectively; 2ω _c - doubled frequency of the modulated wave; Ω - wave generated by the object; t is time; J _n - Bessel functions of the n-th order; A _n is the amplitude of the modulated wave; m _p - modulation factor. As can be seen from the expression, the values of the frequencies of the side components differ from the doubled central frequency 2ωо (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 factor

At low values of the modulation coefficient m _{p, 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 ensure the regulation of the parameters of the formation and reduction of samples by the training unit, a comprehensive reduction in the dimension of the data during auto-tuning of the rule base, based on a sample of the values of the parameters of the classification object, the generation of a signal with the number of the production rule and the type of the membership function to the target type for optimization computational processes performed in the neural network recognition and classification path, which provides the final classification solution for the detected sea targets (surface or underwater object), which limits the functionality of the prototype system.

The problem to be solved by the claimed invention is to expand the functionality of the prototype method, i.e. its implementation as a method of operational identification of sea targets by their information fields based on neuro-fuzzy models. The final classification solution for the detected sea targets (surface or underwater object) is made by the trained neural network according to the degree of belonging of the studied spectral region to the classification object.

The technical result of the proposed invention is the automation of the process of recognizing the classes of sea 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, complex reduction of the data dimension during automatic adjustment of the rule base due to the formation and reduction of a sample of reference samples of mathematical models of sea targets, carried out using a path for regulating the parameters of the formation and reduction of samples necessary for the implementation of the final classification process in the path of neural network recognition and classification, which provides an increase in the probability of a reliable operational classification of a sea target by 7-9% more than when using a prototype ...

The specified technical result is achieved by forming and reducing a sample of reference samples of mathematical models of sea targets by a path for regulating the parameters of the formation and reduction of samples, independently performing auto-tuning of its rule base and its neuro-fuzzy correction, using computational operations of a modified combined recognition network consisting of Kohonen networks - Grosberg and the adaptive neuro-fuzzy network ANFIS, which uses a hybrid learning algorithm with genetic tuning of parameters, for an operatively updated library of spectrograms of maritime targets of the training unit of the neural network recognition and classification path, which provides the final classification solution for the detected maritime targets.

To solve this problem, a method has been developed for the operational identification of sea targets by their information fields on the basis of neuro-fuzzy models, which consists in the fact that, first, a zone of nonlinear interaction and parametric conversion of pump waves with object signals is formed in the marine environment, for which the emitter and the receiving antenna are placed at the opposite boundaries of the controlled area of the marine environment, then the pump waves, modeled by the signals of the object, are received and amplified in the parametric conversion band, their time-frequency scale is transferred to the high-frequency region, a narrow-band spectral analysis is carried out, the parametric components of the total or difference frequency are isolated, according to which with taking into account the temporal and parametric transformation of waves, the characteristics of the object signals are restored, which are fed to the neural network recognition and classification path, which consists of a target class recognition unit by amplitude-frequency characteristics, which implements computational operations of an artificial neural network and is covered by feedback with a learning unit, the memory of which contains data of mathematically processed images of spectrograms of sea targets, and the first input of the target class recognition unit by amplitude-frequency characteristics receives data from the output of the spectrum analyzer of the receiving, processing and registration path signals, and its second input receives data from the training unit of the neural network recognition and classification path.

What is fundamentally different from the prototype is that the data is fed into an additionally introduced path for regulating the formation and reduction parameters of samples to the input, covered by feedback with the fuzzification block, the fuzzy rule block, where the signal of the number of the new production rule and the new type of function of membership in the target type are generated, and so on. is fed to the input of a logical device for the formation and reduction of samples, the function of which is performed by a modified combined recognition network consisting of Kohonen-Grosberg networks and an adaptive neuro-fuzzy network ANFIS using a hybrid learning algorithm with genetic tuning of parameters, in which the membership functions of a new type of rules are compensated , determine the number of the rule required for replacement in the main rule base, as well as a new type of membership function for this rule, and then to the fuzzification block, the output of which is connected to the input of the defuzzification and automatic regulation of the rights base sludge, in which auto-tuning of their rule base is carried out, based on a sample of mathematical models of sea targets, the formation and reduction of a sample of reference samples of mathematical models of sea targets and data correction of the online library of mathematically processed images of spectrograms of sea targets for the training unit of the neural network recognition and classification path, after which the artificial neural network is tuned and a conclusion is drawn about the degree of belonging of the investigated spectrum region to the classification object (surface or underwater object).

As you know, the extraction of useful information from hydroacoustic signals determines the fundamentals of data processing algorithms in the method of detecting and neural network recognition of the features of fields of different physical nature generated by sea targets. To form the feature vector, which is the input information array of the recognition network, the method of masks 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 target recognition (see Pyatakovich V.A., Bogdanov V.I., Nazarenko P .K. The principle of automatic recognition of the target image: materials of the International conference "Mathematical modeling of physical, economic, technical, social systems and processes." - Ulyanovsk: USU, 2003. - P. 31, 32; Pyatakovich V.A., Vasilenko AM, Khotinsky O.V., Recognition and classification of sources of formation of fields of various physical nature in the marine environment: monograph. - Vladivostok: Mor.Gos.University, 2017. - 255 p .; Pyatakovich V.A., Vasilenko AM, Khotinsky O. V. Neural network technologies in intelligent systems for the detection and operational identification of sea targets: monograph .-- Vladivostok: Marine State University, 2018 .-- 263 p .; V.A. Pyatakovich, AM Vasilenko, S.V. Pashkeev. An automated system for monitoring the marine environment for solving problems of detecting a technical object. - Dual technologies. - M: 2018. No. 4 (85). - S. 85-88; Pyatakovich V.A., Rychkova V.F. Parametric optimization of a neural network system for the classification of sea targets by the criterion of reliability. - Marine Intelligent Technologies. - SPb .: 2018. No. 4 (42) Volume 5. - P. 153-162; V.A. Pyatakovich An experimental model of the developed sample of the recognition module for the classification system of sea targets based on neural network technologies. Problems and methods of development and operation of weapons and military equipment of the Navy: Sat. articles. - Vladivostok: TOVVMU, 2018. - Issue. 98. - S. 176-181; V.A. Pyatakovich Solving the problem of classifying a marine target using neural fuzzy networks based on the Mamdani-Zade inference model. Problems and methods of development and operation of weapons and military equipment of the Navy: Sat. articles. - Vladivostok: TOVVMU, 2019. - Issue. 101. - S. 275-284; Pyatakovich V.A., Filippov E.G. The method of formation and reduction of samples for solving problems of classification of marine targets by a neuroclassifier. Problems and methods of development and operation of weapons and military equipment of the Navy: Sat. articles. - Vladivostok: TOVVMU, 2019. - Issue. 101. - S. 312-320; V.A. Pyatakovich Training of deep neural networks of the operational identification system for sea targets of the Russian Navy. Problems and methods of development and operation of weapons and military equipment of the Navy: Sat. articles. - Vladivostok: TOVVMU, 2020. - Issue. 104 - S. 177-186; Pyatakovich V.A., Rychkova V.F., Filippov E.G. Neural networks as a variant of the computational structure of the marine targets classification system. - Marine Intelligent Technologies. - SPb .: 2020. No. 1 (47) Volume 2. - P. 163-174; Pyatakovich V.A., Surov A.B., Rychkova V.F. Calculation of the efficiency of target classification by an intelligent system of the Navy, using a complex of computational operations of neural networks. - Marine Intelligent Technologies. - SPb .: 2020. No. 1 (47) Volume 2. - P. 175-185; Pyatakovich V.A., Nikolaev A.V., Kostikov E.A. Automation of information processing in an intelligent marine monitoring system. - Problems of mechanical engineering and automation. - M: 2020. No. 4. - S. 72-79).

The idea is that for each mask the maximum amplitude value is sought, which is the unit vector of the classification feature vector. To automate the process of searching for an extremum in the zone of one mask, the MAXNET maximum search network was used. The network iterations are completed after the output neurons of the network stop changing. The type of elements of the input signals - integers or real numbers, the type of elements of the 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 equal to N (N-1). The formation of synaptic weights occurs according to the formula

i=1, …, N; j=1, …, N.where W _ij is the 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 y _j (0) = x _j ,

i = 1, ..., N; j = 1, ..., N.

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

Normalization of the input feature vector obtained after analyzing the masks by the MAXNET network is performed according to the expression

известны и определяются моделью входного гидроакустического сигнала.Value range boundaries

are known and determined by the model of the input hydroacoustic signal.

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

- норма вектора (см. Пятакович В.А., Василенко A.M., Хотинский О.В. Распознавание и классификация источников формирования полей различной физической природы в морской среде: монография. - Владивосток: Морской гос. ун-т им. Г.И. Невельского, 2017. - 255 с.; Пятакович В.А., Василенко A.M. Перспективы и ограничения использования геометрических методов распознавания акустических образов морских объектов применительно к задаче управления нейросетевой экспертной системой. - Фундаментальные исследования. - М: 2017. - №7. - С. 65-70; Пятакович В.А., Василенко A.M. Мироненко М.В. Обучение нейронной сети как этап разработки экспертной системы для классификации источников физических полей при мониторинге акваторий. - Вестник Инженерной школы Дальневосточного федерального университета. - Владивосток: Дальневост. федерал, ун-т, 2017. №3 (32). - С. 138-149. DOI.org/10.5281/zenodo.897021; Пятакович В.А. Система классификации морских целей на базе нейросетевых технологий. - Морские интеллектуальные технологии. - СПб.: 2018. №4 (42) Том 5. - С. 169-176).The recognition network is trained on the basis of an error backpropagation algorithm that implements a gradient method for optimizing a functional of the form:

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

- vector norm (see Pyatakovich V.A., Vasilenko AM, Khotinsky O.V. Recognition and classification of sources of formation of fields of different physical nature in the marine environment: monograph. - Vladivostok: Marine State University named after G.I. Nevelskoy, 2017. - 255 p .; Pyatakovich V.A., Vasilenko AM Prospects and limitations of the use of geometric methods for recognizing acoustic images of sea objects in relation to the problem of controlling a neural network expert system. - Fundamental research. - М: 2017. - No. 7. - Pp. 65-70; Pyatakovich V.A., Vasilenko AM Mironenko M.V. Training of a neural network as a stage in the development of an expert system for the classification of sources of physical fields when monitoring water areas. - Bulletin of the Engineering School of the Far Eastern Federal University. - Vladivostok: Far Eastern Federal , un-t, 2017. No. 3 (32). - P. 138-149. DOI.org/10.5281/zenodo.897021; Pyatakovich V.A. Classification system of marine targets based on neural network technologies. - Marine intelligent technologist ui. - SPb .: 2018. No. 4 (42) Volume 5. - P. 169-176).

For autocorrection and regulation of the algorithm of back propagation of the error during training of the recognition network, a path for regulating the parameters of the formation and reduction of samples is used, which combines adaptive approaches of self-learning and the experience of an expert, which, in the presence of undefined parametric disturbances, makes it possible to promptly correct the values corresponding to the new conditions for the classification of a sea target ...

The invention is illustrated by drawings, where Fig. 1 shows a functional diagram of a system for operational identification of sea targets by their information fields based on neuro-fuzzy models, containing the following elements:

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

2. Receiving transducer.

3. Marine environment.

4. Working area of nonlinear interaction and parametric conversion of pump waves and information waves.

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

6. Path of emission of pump signals.

6.1. Stabilized frequency pump signal generator.

6.2. Amplifier.

6.3. Matching block.

7. The path for receiving, processing and registering 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 unit by amplitude-frequency characteristics.

8.2. Learning block.

9. The path for regulating the parameters of the formation and reduction of samples.

9.1. Block of fuzzy rules and functions.

9.2. Fuzzification block.

9.3. Logic device for the formation and reduction of samples.

9.4. Defuzzification and automatic regulation of the rule base.

The general structure of the modified combined recognition network consisting of Kohonen and Grosberg networks is shown in Fig. 2.

The network modification consists in adding the MAXNET network to the Kohonen and Grosberg network, which is very important for the problem being solved.

Число нейронов во втором (скрытом) слое определяется взаимным расположением и формой разделяемых множеств.For each neuron of the first layer through synapses with weights {T _ij ⁽¹⁾ }, i = 1, 2, 3; j = 1, 2, 3, all components of the input vector are supplied

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

For each neuron of the second layer through synapses with weights {T _ij ⁽²⁾ }, i = 1, 2, 3; j = 1, 2, 3, the outputs of the first layer are supplied. The number of neurons in the third (output) layer is determined by the number of considered classes to be recognized.

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

где x_2n ⁽ⁱ⁾, y_n ⁽ⁱ⁾ и I_n ⁽ⁱ⁾ - значения r-го входного сигнала, выходного сигнала и внешнего смещения n-го нейрона i-го слоя; N_i - число нейронов в i-м слое; i=1, 2, 3.For each neuron of the third layer through synapses with weights {T _ij ⁽³⁾ }, i = 1, 2, 3; j = 1, 2, 3, the output signals of the second layer are supplied. 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 a known type of activation function

where x _2n ⁽ⁱ⁾ , y _n ⁽ⁱ⁾ and I _n ⁽ⁱ⁾ are the values of the r-th input signal, the output signal and the external displacement of the n-th neuron of the i-th layer; N _i - the number of neurons in the i-th layer; i = 1, 2, 3.

Границы диапазона значений

известны и определяются моделью входного гидроакустического сигнала.Pre-processing of the input vectors is performed by normalizing the input feature vector obtained after analyzing the masks by the MAXNET network or after obtaining statistical estimates according to the expression

Value range boundaries

are known and determined by the model of the input hydroacoustic signal.

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 the weights (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 training vector of features and the vector of synapses (weight coefficients) of the "winner" neuron is determined.

4. If the condition β * <β is satisfied, 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 by 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.

l=1, 2. где

- l-й компонент желаемого выходного вектора классификатора;

- выходной сигнал j-го нейрона слоя Кохонена, при обучении S-му входному вектору признаков.Learning the Grosberg layer is a traditional supervised learning and can be performed both simultaneously with the architectural self-organization and tuning the Kohonen layer for each input sample vector, or automatically (after training the Kohonen layer). In all cases, the learning rule can be represented by the following algorithm:

l = 1, 2.where

- l-th component of the desired output vector of the classifier;

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

на нейроны слоя Кохонена. Входные потенциалы нейронов

i=1, …, L.Functioning of the considered recognition device. The preprocessed vector of input features of the observed object {xi} is fed to the input of the network and is distributed through the weighting coefficients of the connections (synapses)

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

i = 1, ..., L.

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

l=1, 2, где 10 ( ) - функция единичного скачка.After this, the Kohonen layer begins to function as a competing network with lateral connections. 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 enters the neurons of the Grosberg output layer, which function according to the following algorithm:

l = 1, 2, where 10 () is the unit jump function.

FIG. 3 presents a neural network interpretation of the multivariate classification algorithm.

FIG. 4 shows a table of interpretation of the three-dimensional output vector of recognition of hydroacoustic signals by the amplitude-frequency characteristic.

FIG. 5 shows the mask method used for frequency response recognition.

In each mask, according to the real characteristic, the maximum amplitude value of the signal A ₁ , A ₂ , ..., A _j ..., A _k is determined. The choice of the value of Δ, and, consequently, of the number of masks is determined by the capabilities of the recognition 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 (an increase in the number of neurons in the input layer) of the recognition device, 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.

FIG. 6 presents a comparative characteristic of the methods for classifying a sea target in terms of training and classification time.

The fastest method in terms of the quality of classification is the modified method of potential functions based on NN (Fig. 6. Table 2). Neural networks have stronger approximating abilities than other methods of target classification, and the multidimensional classification algorithm, in contrast to the methods of potential functions, takes into account the significance of features by taking into account the particular values of two-feature classifications.

FIG. 7 and FIG. 8 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. As can be seen from the figures, the recognition and classification of sea targets using computational operations of a modified combined recognition network consisting of Kohonen (competing network) and Grosberg networks and an adaptive neuro-fuzzy network ANFTS, using a hybrid learning algorithm, makes it possible to increase the likelihood of a reliable operational classification as surface. and underwater targets by 7-9%.

Method for operational identification of sea targets by their information fields based on neuro-fuzzy models

The emitting transducer 1 and the receiving transducer 2 are placed in the marine environment 3, taking into account the patterns of multipath propagation of waves in an extended hydroacoustic channel, which ensures the formation and effective use of a spatially developed working zone 4 of nonlinear interaction and parametric transformation of transmission waves and waves of various physical nature generated by objects 5 (see Certificate of state registration of the computer program "Calculation of ray pattern" No. 2016616822. 06/21/2016; Certificate of state registration of the computer program "Program for simulation of the propagation of hydroacoustic signals" No. 2017664296. 20.12.2017; Certificate of state registration computer programs "Software and computing complex for imitating the maritime information situation in identifying targets" No. 2018612944 03/01/2018 Bulletin No. 3; Certificate of state registration of the computer program "SP The formalized neural network complex for the classification of noisy signals of sea targets ”№2018619739 RF. 08/10/2018. Bul. # 8; Certificate of state registration of the computer program "Program for the design and training of artificial neural networks of perceptron type" №2019611559 RF. 01/29/2019. Bul. No. 2).

The pump signal of the 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 a submarine cable connecting the output of the pump signal emission path 6 and the input of the emitting converter 1. Emitter 1 sounds medium with pump signals of a stabilized frequency in the range of tens to hundreds of hertz. The pumping of the monitored marine environment with complex frequency- or phase-modulated signals provides an increase in the efficiency of parametric reception of waves of acoustically subtle objects.

In various modes of movement, objects 5 generate radiation, leading to a change in the value of the characteristics of the conducting liquid (density and (or) temperature and (or) heat capacity, etc.), which, depending on their physical nature, modulate low-frequency pumping 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 pumping wave by radiation and fields of objects.5 and registration of information signals 7.

Information waves are received by antenna 2, amplified in the parametric conversion band (block 7.1), transfer the time-frequency scale of signals to the high-frequency region (block 7.2), carry out a narrow-band spectral analysis (block 7.3) and extract from them the parametric components of the total and difference frequencies, according to to which the signs of manifestation of fields generated by objects are restored.

The operation of transforming the time-frequency scale of the signal in block 7.2 of the signal receiving, processing and recording path provides an increase in the energy concentration of nonlinearly converted signals and the efficiency of separating from them the features of the fields generated by sea targets.

Spectral analysis operations in block 7.3 allow to select discrete components of the total or difference frequency in the narrow-band spectra of nonlinearly transformed signals, which are used to restore the characteristics of the waves of objects 5.

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

The second input of the block 8.1 for recognizing the target class by the amplitude-frequency characteristics receives data from the training block 8.2 of the path of neural network recognition and classification 8.

Further, according to the amplitude-frequency characteristics, the class (surface or underwater object) of the detected sea target is recognized in the neural network recognition and classification path 8. The analysis of the low-frequency, mid-frequency and high-frequency components of the amplitude-frequency characteristics is carried out separately, since the general signs for different types of objects can be located in different frequency ranges.

The process of training a multilayer neural network (MLN) is iterative and, in the general case, rather lengthy, since it is impossible to determine in advance the number of iterations required to train a neural network.

For the neural network implementation of distance comparison and determination of the K _ij value, the following expression is used:

where ψ (x) is the logistic function.

If the function ψ (x) is discrete, for example, the threshold:

то K_ij будет принимать значение 0 или 1.

then K _ij will take on the value 0 or 1.

If the function ψ (x) is real, for example, sigmoid:

то K_ij 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.

then K _ij will take values on the interval [0, 1], the closer the value of this function will be to 0, the closer the instance will be to the class to which the value 0 is associated, and, accordingly, on the contrary, the closer the value of this function will be to 1, the closer the instance will be to the class that is mapped to 1.

Using a sigmoid function may be preferable in practice, since it allows you not only to 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:

и

вычисляются на основе формального нейрона, имеющего один вход, на который подается значение x_i или x_j, вес которого равен

или

соответственно.Expressions for

and

are calculated on the basis of a formal neuron with one input, to which the value x _i or x _j is supplied, the weight of which is

or

respectively.

или

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

or

respectively.

In this case, the rules for calculating the parameters of the multidimensional classification algorithm will remain unchanged, and the parameters and activation functions of the neural network (NN) must be determined on their basis according to the following rules. Activation function ψ ^{(μ, k) of the} kth neuron of the μth layer:

p-го входа k-го нейрона μ-го слоя:Weight coefficient

p-th input of the k-th neuron of the μ-th layer:

элементы которого представлены в табл. 1. на фиг. 4.The values of the output signals of the third layer form a decision vector

whose elements are presented in table. 1. in FIG. 4.

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

The weight coefficients of the recognition network of block 8.1, target class recognition by amplitude-frequency characteristics are corrected using the training unit 8.2, the neural network recognition and classification path 8. Next, the data is fed to the additionally introduced path for regulating the parameters of the formation and reduction of samples 9 to the input of the fuzzy rules and functions block 9.1, where the signal of the number of the new production rule and the new type of the function of belonging to the target type is generated and then fed to the input of the logical device for the formation and reduction of samples 9.3, the path for regulating the parameters of the formation and reduction of samples, the function of which is performed by a modified combined recognition network consisting of Kohonen networks - Grosberg and the adaptive neuro-fuzzy network ANFIS, which uses a hybrid learning algorithm that compensates for membership functions of a new type of rules, determine the number of the rule required for replacement in the main rule base, as well as a new type of function accessory for this rule, and then to the fuzzification block 9.2, the output of which is connected to the input of the defuzzification and automatic regulation block of the rule base 9.4, in which they automatically adjust their rule base based on a sample of mathematical models of sea targets, the formation and reduction of a sample of reference samples of mathematical models of sea targets and correction of the data of the operatively updated library of mathematically processed images of spectrograms of sea targets for training block 8.2 of the neural network recognition and classification path 8, after which the artificial neural network is tuned and a conclusion is drawn about the degree of belonging of the investigated spectrum region to the classification object (surface or underwater object).

The function of the logical device for the formation and reduction of samples 9.3 in the fuzzy control system of the path for regulating the parameters of the formation and reduction of samples 9 is performed by a modified combined recognition network consisting of Kohonen-Grosberg networks and an adaptive neuro-fuzzy network ANFIS using a hybrid learning algorithm. The initial sample, for the operation of the logical device for the formation and reduction of samples, being mapped into one-dimensional space, will allow to select on the one-dimensional axis the intervals of its values corresponding to clusters of different classes in the original N-dimensional space. Having determined the boundaries of the intervals on the one-dimensional axis, one can find the instances closest to them, which will make up the generated subsample.

Algorithm of the logic device for the formation and reduction of samples.

Initialization phase. Set the original data sample X = <x, y>.

и

- соответственно, минимальное и максимальное значения j-го признака, j=1, 2, …, N. Определить число интервалов для каждого признака: k=K 1n S, а также длины интервалов:

The stage of analyzing the characteristics of the sample. Define

and

- respectively, the minimum and maximum values of the j-th feature, j = 1, 2, ..., N. Determine the number of intervals for each feature: k = K 1n S, as well as the length of the intervals:

The stage of calculating generalized characteristics. For each s-th instance, s = 1, 2, ..., S: determine k _j - the number of the interval of values for each j-th feature, j = 1, 2, ..., N, into which the s-th instance falls

calculate the coordinate of the s-th instance along the generalized axis

This allows you to map the original sample on the one-dimensional generalized axis I (note that this will result in the loss of some information due to the implicit quantization of the feature space during transformation).

Generalized axis analysis stage. Form a set of tuples I = {<I^s, y^s, s>}. Order the set I in non-decreasing order of the values of I^s... Looking at the generalized axis in order of increasing its values, determine the boundary values of its intervals <l_q, r_q>, in which the class number y^s remains unchanged, where l_q, r_q - respectively, the left and right boundary values of the q-th interval of the generalized axis. We denote: K_q - class number corresponding to the q-th interval of the generalized axis; k_I - the number of intervals of the generalized axis.

The stage of analyzing the characteristics of the intervals. For each q-th interval of the generalized axis, k _q = 1, 2,…, k _I , determine S _q - the number of instances that fell into it, as well as the numbers of these instances.

The stage of forming a training sample. Among the instances of the q-th interval, include in the training set X ^* all instances of: its class that are on one of the boundaries of the interval

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

where α is a threshold coefficient that regulates the proximity of specimens to the boundaries of the interval (for example, you can set: α = 0, l (r _q -l _q ));

intervals with a small number of instances:

s=1, 2, …, S, q=1, 2, …, k_i,

s = 1, 2,…, S, q = 1, 2,…, k _i ,

- среднее число экземпляров в интервале обобщенной оси.where β is some threshold coefficient, 0 <β <1 (for example, you can set: β = 0.1);

is the average number of specimens in the interval of the generalized axis.

The stage of eliminating the redundancy of the training sample. Determine the distances between all instances included in the formed training sample by forming a distance matrix R (to simplify and speed up the calculations, we will operate with the squares of the distances):

Note that R (s, p) = R (p, s) and R (s, s) = 0.

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

, g ≠ р, perform the actions in a cycle: find in the distance matrix two instances with the smallest distance between themselves:

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

Correct the elements of the matrix R accordingly, setting: R (s, q) = R (q, s) = - 1; if the two closest instances belong to different classes, then proceed to the execution of the stage of supplementing (refining) the training sample. The stage of supplementing (refining) the training sample. Determine the difference between the original and formed samples X '= X / X ^* .

Sequentially for each s'th instance <x ^s' , y ^s' > of the sample X ', s' = 1, 2, ..., S' relative to the instances of the formed sample X ^* find the distance (square of the distance) from it to each instance of the sample X ^* :

if the closest to the s'th instance of the generated selection belongs to another class, then include it in the selection X ^* :

As a result of executing this method for the initial sample X, we obtain the formed training sample X ', which will have the basic topological properties of the original sample. At the same time, from the initial sample, it is also possible to obtain a test sample as the difference between the initial and formed training samples. Partitioning the feature space for a sample of empirical observations is necessary to determine fuzzy terms of features as projections of the corresponding blocks on the coordinate axes. It is proposed to form partitions by performing them in the sequence given below.

Step 1. Initialization. Set training set <x, y>.

.Step 2. Along the axis of each feature j = 1, 2,…, N, determine the one-dimensional distances between the instances:

...

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

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

and also determine the length of the interval of observed values of each feature: r _j = max (x _j ) -min (x _j ), j = 1, 2, ..., N.

Step 4. Divide the axis of the j-th feature into n _j intervals. Determine the coordinates of the left and right boundaries for each p-th interval of the j-th feature using the formulas:

Step 5. Form block-clusters and set their class numbers by performing steps 5.1-5.8.

в N-мерном пространстве признаков на пересечении соответствующих интервалов значений признаков. Занести в B_{q, j} номер интервала j-го признака, который соответствует q-му блоку.Step 5.1. Form rectangular blocks

in the N-dimensional space of features at the intersection of the corresponding intervals of feature values. Enter in B _{q, j the} number of the interval of the j-th feature, which corresponds to the q-th block.

Step 5.2. Determine the class numbers for rectangular blocks in the N-dimensional feature space:

Set the classification confidence factor for blocks:

Step 5.3. For those blocks for which K _q = -1, q = G + 1, G + 2, ...,

G + Q, install:

- количество экземпляров k-го класса, попавших в q-й блок-кластер.where

- the number of instances of the k-th class that got into the q-th block-cluster.

Step 5.4. For those blocks for which the class number K _q = 0, q = G + 1, G + 2,…, G + Q, determine the calculated class number, for which it is proposed to use a modified non-recurrent method of potential functions.

Step 5.5. Calculate the distance between q-th and p-th blocks, q = G + 1, G + 2,…, G + Q, p = q + 1,…, G + Q, as:

где

При этом R(B_q, B_p)=R(B_p, B_q).Or

where

Moreover, R (B _q , B _p ) = R (B _p , B _q ).

Step 5.6. Determine the potential induced by a set of blocks belonging to the kth class to the pth block with an unknown classification:

q = G + 1, G + 2, ..., G + Q, p = q + 1 ..... G + Q.

where L _k is the number of blocks belonging to the k-th class, S _q is the number of instances of the training sample that fell into the q-th block.

р=G+1, G+2, …, G+Q.Step 5.7. Set the class number for the p-th block with an unknown classification (K _p = 0) according to the formula:

p = G + 1, G + 2, ..., G + Q.

q =G+1, G+2, …, G+Q.Step 5.8. Modify the values of the confidence coefficients for the blocks:

q = G + 1, G + 2,…, G + Q.

Step 6. Perform the union of adjacent block-clusters.

∀i≠j:B_{q, i}=B_{p, i}, i=1, 2, …, N, j=1, 2, …, N; тогда объединить блоки q и p по j-му признаку:Perform the union of adjacent blocks belonging to the same class: for ∀q, p = G + 1, G + 2,…, G + Q; q ≠ р: if K _q > 0, K _q = K _p and

∀i ≠ j: B _{q, i} = B _{p, i} , i = 1, 2,…, N, j = 1, 2,…, N; then combine blocks q and p according to the j-th feature:

- install:

- delete the p-th block: K _p = 0, α _p = 0, B _{p, i} = 0, i = 1, 2, ..., N.

j=1, 2, …, N; s^*=1, 2, …, S^*; то необходимо из набора новых наблюдений исключить те наблюдения, которые попадают в блоки имеющегося разбиения и соответствуют им по номеру класса, скорректировав соответствующим образом S^*. 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 block-clusters, the class numbers of which do not coincide with the class numbers of the instances. Apply for the resulting partition and the selected subset the procedure for refining the partition and additional training of the model. Step 8. Stop. Refinement of the partition and additional training of the model. If there is a partition of the feature space that needs to be refined (retrained) based on new observations <x ^* , y ^* >,

j = 1, 2, ..., N; s ^* = 1, 2, ..., S ^* ; 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 the class number, correcting S ^* accordingly. For those observations that do not coincide with the block classes, it is advisable to form separate point clusters.

номера интервалов для каждого j-го признака, соответствующие новому кластеру, а также определить:For each new observation, form intervals by signs and enter into

interval numbers for each j-th feature, corresponding to the new cluster, and also determine:

where γ is some constant, γ∈.

To determine the feasibility of using 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 stages of the proposed method. For the initialization step, the computational complexity can be neglected, and the spatial complexity can be estimated as O (NS).

For the stage of analyzing the characteristics of the sample, the computational complexity will be O (2NS), and the spatial complexity will be O (4N). For the stage of calculating generalized features, the computational complexity can be estimated as O (9NS), and the spatial complexity - O (N + S). For the generalized axis analysis stage, the computational complexity can be estimated as O (2S ² + 2S), and the spatial complexity - O (3S + 3K1nS). For the stage of analyzing the characteristics of intervals, the computational complexity can be estimated as O (3KS1nS), and the spatial complexity - O (K1nS). For the stage of forming a training sample, the computational complexity can be estimated as O (20KS1nS), and the spatial complexity as O (0.2K1nS + 0.2S). For the stage of eliminating the redundancy of the OB, the computational and spatial complexity can be estimated, respectively, as O (0.0016S ⁴ + (0.44 + 0.08N) S ² ) and O (0.04S ² ).

For the stage of supplementing (refining) the training sample, the computational and spatial complexity are estimated as O ((0.48N + 0.32) S ² ) and O (0.8S), respectively. Thus, the overall complexity of the method can be estimated as: computational - O (0.0016S ⁴ + 0.56NS ² + 2.76S ² + 11NS + 2S ++ 23KS1nS); spatial - O (0.04S ² + NS + 5N + 5S + 4.2K1nS).

The algorithm for correcting the values of the settings of the regulator of the path for regulating the parameters of the formation and reduction of samples is implemented to find the number of a fuzzy rule of the form: (IF Tpi Condition _i AND G _j Condition _j TO N ^k "Conclusion")

k=1, 2, …, N,

k = 1, 2, ..., N,

in the presence of a training set ((D ¹ , N ¹ ),…, (D ^N , N ^N )), ((D ¹ , FP ¹ ),…, (D ^N , FP ^N )).

To simulate an unknown mapping f, a fuzzy inference algorithm is used, predicate rules are applied: P _i: IF e ₁ is A _i1 AND T _p1 is A _i2 AND G ₁ is A _in , THEN N = S _i, i = 1, 2, ..., m, where A _ij - fuzzy sets describing statements: "negative", "zero", "positive", "small", "average", "large", etc. S _i - real numbers (rule number). The degree of truth μ of rule i is determined using the Larsen multiplication operation:

Функция ошибки для i-го предъявленного значения вида: E_k=0.5(U^k-N^k) позволяет использовать градиентный метод для подстройки параметров заданных предикатных правил, а величина S_i корректируется по соотношению:to simulate the logical operator "AND", the output of the fuzzy system U ^k is determined by the center of gravity method:

The error function for the i-th presented value of the form: E _k = 0.5 (U ^k -N ^k ) allows you to use the gradient method to adjust the parameters of the given predicate rules, and the S _i value is corrected according to the ratio:

i=1, 2, …, m,

i = 1, 2, ..., m,

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

In a situation of uncertainty, i.e. the influence of random external and parametric disturbances, the block of defuzzification and automatic regulation of the rule base 9.4 performs their compensation in the rule base, as well as compensation of membership functions of a new kind (with other universe E and E '.

Thus, the path for regulating the parameters of the formation and reduction of samples determines the number of the rule (N) required for replacement in the main base of rules, as well as a new type of membership function for this rule. The block for defuzzification and automatic regulation of the rule base 9.4 of the path for regulating the parameters of the formation and reduction of samples independently performs the automatic adjustment of its rule base, based on a sample of mathematical models of sea targets, carrying out the formation and reduction of a sample of reference samples of mathematical models of sea targets and correction of the data of the operatively updated library of mathematically processed images of spectrograms of sea targets of the neural network recognition and classification path 8, which provides the final classification solution for the detected sea 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 base of rules of a smaller volume on the basis of an initial set of fuzzy rules, which would sufficiently represent the initial base and be would be less redundant.

To solve the problem of the formation and reduction of fuzzy rules, it is necessary to bring the stages of the method in accordance with the peculiarities of the problem being solved, which is presented in the form of a sequence of steps 1-17.

Step 1. Initialization. The static parameters of the method are set: the coefficients α, β, p. 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 values of the significance of a separate term are calculated.

- значение эвристической значимости лингвистического терма p для описания класса q; о - экземпляр входной выборки, содержащей N экземпляров; μ_p(o), μ_q(o) - значение функции принадлежности объекта о терму р и классу q, соответственно; Т - количество лингвистических термов; K - количество классов. В каждом пространстве поиска каждому узлу графа поиска ставится в соответствие начальное значение количества дискретных составляющих (ДС) τ_init

где

- значение количества ДС для р-го терма в пространстве поиска для q-го класса на первой итерации поиска.where

- the value of the heuristic significance of the linguistic term p for describing the class q; o - an instance of the input sample containing N instances; μ_p(o), μ_q(o) is the value of the object's membership function about the term p and class q, respectively; T is the number of linguistic terms; K is 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) τ_init

where

- the value of the number of DSs for the p-th term in the search space for the q-th class at the first iteration of the search.

Step 2. Set: t = 1. Step 3. Set: i = 1. Step 4. Set: j = 1. Step 5. Set: k = 1.

Step 6. Selecting a term to add to the rule of the j-th agent in the search space of the i-th class.

где

- вероятность добавления k-то терма в правило j-го агента в пространстве поиска для i-го класса; R^j - множество термов, которые могут быть добавлены в правило j-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 6.1. For the j-th agent, based on the random selection rule, the probability of including the k-th linguistic term in the rule describing the i-th class of the output value is calculated

where

- the probability of adding the k-th term to the rule of the j-th agent in the search space for the i-th class; R ^j is the set of terms that can be added to the rule of the j-th agent. The formation of this set determines the type of rules that can be formed during the search, that is, it is assumed that the rule can 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 excluded.

где rand (1) - случайное число из интервала [0; 1]. Если условие выполняется, тогда лингвистический терм k добавляется в правило j-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 rand (1) is a random number from the interval [0; one]. If the condition is met, then the linguistic term k is added to the rule of the j-th agent, removed from the set of possible terms for this agent, and the transition to step 7 is performed. Otherwise, the transition to step 6.3.

Step 6.3. Set k = k + 1.

Step 6.4. If all terms were considered, then set: k = 1. Go to step 6.1.

Step 7. Checking the completion of the movement of the j-th agent.

Step 7.1. If the set of terms that the j-th agent can add to the rule being formed is empty, then go to step 8.

Step 7.2. It is determined how many instances of the i-th class are covered by the rule of the j-th agent.

где match(R_j, o) - степень соответствия между правилом j-го агента R_j и экземпляромо;

- мера соответствия между р-м атрибутом в правиле R_j и соответствующим атрибутом экземпляра о. Данная мера рассчитывается следующим образом:Step 7.2.1. For an instance of o belonging to class i, the degree of compliance of the generated rule R _j and an instance of o

where match (R _j , o) is the degree of correspondence between the rule of the j-th agent R _j and the instance;

- the measure of correspondence between the p-th attribute in the rule R _j and the corresponding attribute of the instance o. This measure is calculated as follows:

where q is a separate term related to the description area of the p attribute; Q ^p - the number of terms related to the description area of the attribute p.

Step 7.2.2. Check the condition: match (R _j , о) ≥inMatchMin, where inMatchMin is a given parameter that determines what minimum match value is sufficient to consider that the rule R _j sufficiently describes the object о. If the condition is met, then it is considered that this object is covered by the rule R _j .

Steps 7.2.1 and 7.2.2 are performed for all instances of class i, and based on the received data, the counter cntMatch is incremented, which stores the number of instances covered by the rule R _j .

Step 7.3. Check the condition: cntMatch≥nCntMatchMini, where inCntMatchMini is the limiting minimum number of instances of the i-th class that should 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 j-th agent has completed its movement, go to step 8. Otherwise, go to step 5.

Step 8. If j <cntAgents, then set: j = j + 1 and go to step 5. Otherwise, go to step 9.

Step 9. If i <k, then set: i = i + 1 and go to step 4. Otherwise, go to step 10.

Step 10. Rule bases are generated randomly. This will create nBases of rule bases, selecting one rule from the corresponding search space to describe each class of output value.

где cntNotMatch - количество экземпляров, для которых класс был определен неверно с помощью заданной базы правил; Q - качество прогнозирования класса экземпляров на основе соответствующей базы правил.Step 11. Assessment of the quality of the generated rule bases. To assess the quality of the rulebases, an input training sample is used, for each instance of which the rule with the highest coincidence is selected, based on which the computational class to which the given instance belongs is determined based on the corresponding rule base. Based on the data obtained using the knowledge base, based on a given training sample, we calculate the quality score of the rule base:

where cntNotMatch is the number of instances for which the class was incorrectly defined using the specified rule base; Q is the quality of predicting the class of instances based on the corresponding rule base.

Step 12. Check the condition: Q _high ≥Q _threshold , where Q _high is the forecasting quality of the knowledge base, which is characterized by the best forecasting accuracy; Q _threshold - acceptable forecasting quality. If the specified condition is met, then go to step 17, otherwise, go to step 13.

∀p∈R, ∀R⊂Ω, где

- количество ДС для терма р в пространстве поиска для класса q, который определяется с помощью соответствующего правила.Step 13. Adding DS. The addition of DS is performed in order to increase the priority of those terms, the inclusion of which in the rules improves the quality of forecasting the resulting rule bases. In this regard, the amount of added priority coefficient is directly proportional to the forecasting quality of the rule base, which includes the given fuzzy rule. In this case, adding the DS offered is limited to those terms included in the rule base of fuzzy rules, for which the condition: Q _Ω ≥δ⋅Q _high, where δ ^_ a factor that determines how closely the quality of the base rules predicting Q _Ω should approach the best the quality of forecasting Q _high , so that the procedure for adding DS for the rules included in this rule base Ω can be applied. Thus, the addition of DS is performed for each term included in the rule, which, in turn, is included in the rule base Ω.

∀p∈R, ∀R⊂Ω, where

- the number of DSs for the term p in the search space for the class q, which is determined using the corresponding rule.

,

,

где ρ - коэффициент исключения, который задается при инициализации.Step 14. Elimination of DS. To exclude the worst terms, that is, those that, when included in the rules, lower the forecasting quality using the corresponding rule, the DS elimination procedure is applied, which is performed at the end of each iteration and is applied to all nodes in all search graphs. The elimination of the DS of the marine target (MT) is carried out in accordance with the formula:

,

,

where ρ is the exclusion factor, which is set during initialization.

Step 15. If t <t _max , then set: t = t + 1 and go to step 16, otherwise, it is considered that the maximum number of iterations has been completed, and go to step 17.

Step 16. Restarting 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 carried out, the bases of fuzzy rules were obtained, which were characterized by the accuracy of classification of the MC by this method in 85.6%. Consequently, the proposed method provides the formation of a reduced base of fuzzy rules, which is characterized by sufficiently accurate results, while reducing the time spent on the 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 predicting the recognition of different classes of MC.

During the operation of the method, the data of the operatively updated library of mathematically processed images of spectrograms of sea targets, statistics of the classification characteristics are constantly being corrected, the description of the classes of sea targets for the current hydrological-acoustic and noise-signaling conditions is refined, on the basis of which the actual threshold for classification is formed.

Thus, having detected a target based on the signs of amplitude-phase modulation of low-frequency pumping signals of the marine environment by radiation and fields of the object, and using adaptive neuro-fuzzy correction with a complex reduction of the data dimension to adjust the parameters of the formation and reduction of samples during automatic adjustment of the rule base due to the formation and reduction of the sample reference samples of the operatively updated library of mathematically processed images of spectrograms of sea targets, as well as the architecture of a recognition neural network in the form of a modified combined recognition network consisting of Kohonen (competing network) and Grosberg networks and an adaptive neural fuzzy network ANFIS, using a hybrid learning algorithm, can be in an automated mode, recognize the target class by the amplitude-frequency characteristics and draw a conclusion about the degree of belonging of the investigated spectral region to the classification object (surface or underwater object). According to the results of experimental data, the efficiency of the classification of sea targets in the claimed method increases by 7-9% relative to the prototype and other classification methods using modulation of low-frequency signals for pumping the marine environment by radiation and fields of objects.

The method is industrially applicable, since common components and products of the radio engineering industry and computers are used for its implementation.

Claims (1)

A method for the operational identification of sea targets by their information fields based on neuro-fuzzy models, which consists in the fact that first a zone of nonlinear interaction and parametric conversion of pump waves with object signals is formed in the marine environment, for which the emitter and the receiving antenna are placed on opposite boundaries of the controlled area marine environment, then the pump waves, modulated by the signals of the object, are received and amplified in the parametric transformation band, their time-frequency scale is transferred to the high-frequency region, a narrow-band spectral analysis is carried out, the parametric components of the total or difference frequency are isolated, according to which, taking into account the time and parametric transformation waves restore the characteristics of the object signals, which are fed into the neural network recognition and classification tract, which consists of a target class recognition unit based on the amplitude-frequency characteristics, which implements computational operations artificial neural network and covered by feedback with the learning unit, in the memory of which the data of mathematically processed images of spectrograms of sea targets are recorded, moreover, data from the output of the spectrum analyzer of the signal receiving, processing and registration path are received at the first input of the target class recognition unit based on the amplitude-frequency characteristics, and its second input receives data from the training unit of the neural network recognition and classification path, characterized in that the data is fed to the additionally introduced path for regulating the parameters of the formation and reduction of samples to the input of the fuzzy rules and functions block, covered by feedback with the fuzzification block, where the signal number is generated a new production rule and a new type of function of belonging to the type of goal, and then fed to the input of a logical device for the formation and reduction of samples, the function of which is performed by a modified combined recognition network consisting of Kohonen-Grosberg networks and an adaptive n the neuro-fuzzy network ANFIS, which uses a hybrid learning algorithm with genetic tuning of parameters, in which the membership functions of a new type of rules are compensated, the number of the rule required for replacement in the main rule base is determined, as well as a new type of membership function for this rule, and then by fuzzification block, the output of which is connected to the input of the defuzzification and automatic regulation of the rule base, in which they automatically adjust their rule base based on a sample of mathematical models of sea targets, form and reduce a sample of reference samples of mathematical models of sea targets and correct the data of an online updated library of mathematically processed images of spectrograms of sea targets for the training unit of the neural network recognition and classification tract, after which the artificial neural network is tuned up and a conclusion is drawn about the degree of belonging of the investigated spectrum region to the classification object (surface or underwater object).

Images