Intelligent decision support system under uncertainty

Abstract

FIELD: hydroacoustics.SUBSTANCE: invention can be used to construct systems for intelligent data analysis under conditions of uncertainty (the influence of random external disturbances) in order to automate the decision-making process for activating the executive system of military and special equipment complexes. An additional control path for the activation of the executive system is introduced, comprising a logical device, the input of which is connected to the output of the target class recognition unit according to the amplitude-frequency characteristics of the neural network recognition and classification path, covered by feedback with the pulse generator.EFFECT: automation of the decision-making process for activating the executive system of military and special equipment complexes through intelligent analysis of data on the degree of belonging of the studied spectral region to the classification object.1 cl, 8 dwg

Description

[0001] The invention relates to hydroacoustics and can be used to build systems for intelligent data analysis under conditions of uncertainty (the influence of random external disturbances) with the aim of automating the decision-making process for activating the executive system of military and special equipment complexes, for adjusting their operating modes to the diversity of manifestations of information wave sources, according to information on the degree of belonging of the studied spectral region to the classification object.

[0002] An intelligent system is known (patent No. 2681242 of the Russian Federation, IPC G01S 15 / 04 "Intelligent system for detecting and classifying sea targets", published on 05.03.2019, Bulletin No. 7), which includes a working zone of nonlinear interaction and parametric transformation of pump waves and information waves formed in the marine environment, wherein the length of the working zone is equal to the length of the controlled section of the marine environment, for which the emitting and receiving transducers are located on opposite boundaries of the section, while the input of the emitting transducer is connected by an underwater cable to the output of the pump signal radiation path, which contains a series-connected pump signal generator of a stabilized frequency, a power amplifier and a unit for matching its output with the underwater cable, and the output of the receiving transducer is connected by an underwater cable to the input of the path for receiving, processing and recording information signals, which contains a series-connected broadband amplifier,a frequency-time scale converter, a spectrum analyzer and a recorder functionally connected thereto, as well as a neural network recognition and classification path, containing a target class recognition unit based on the amplitude-frequency characteristics, covered by feedback with the training unit; wherein the output of the spectrum analyzer of the information wave reception, processing and recording path is connected to the input of the target class recognition unit based on the amplitude-frequency characteristics of the neural network recognition and classification path, at the output of which a signal is generated based on the type of target in accordance with the degree of belonging of the studied spectral region to the classification object.

[0003] The operating principle of parametric antennas is based on the use of natural nonlinear properties of the marine environment (see Novikov BK, Timoshenko VI Parametric antennas in sonar systems. - L.: Sudostroenie. - 1990. - Pp. 17-40, 203-225; Mironenko MB, Malashenko AE, Karachun LE, Vasilenko AM Low-frequency transmission method of long-range sonar of hydrophysical fields of the marine environment: monograph. - Vladivostok: SKB SAMI FEB RAS, 2006. - 173 p.; Pyatakovich VA, Vasilenko AM, Filippova AV Technology of creating an automated system for long-range reception and neural network classification of hydrophysical fields of marine waters / / Bulletin of the Tula State University. Technical sciences. Issue 7. Tula: Tula State University Publishing House, 2017. - P. 247-258; Pyatakovich V.A., Vasilenko A.M., Rychkova V.F. Intelligent system of neural network classification of sea targets / / Marine intelligent technologies. - 2018. - No. 2 (40), Vol. 2. - P. 115-120; Pyatakovich V.A., Rychkova V.F., Filippov E.G.Neural Networks as a Computational Structure Variant for a Sea Target Classification System / / Marine Intelligent Technologies. - 2020. - No. 1 (47), Vol. 2. - P. 163-174; Pyatakovich V.A., Surov A.B., Rychkova V.F. Calculating the Efficiency of Target Classification by an Intelligent Navy System Using a Complex of Neural Network Computing Operations / / Marine Intelligent Technologies. - 2020. - No. 1 (47), Vol. 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. - 2020. - No. 4. - P. 72-79).

[0004] When using multi-element parametric antennas towed behind sea vessels, in addition to the natural properties of the environment, the nonlinear properties of the wake are used.

[0005] Research and testing of parametric antennas using high-frequency pumping of the marine environment (tens to hundreds of kHz) have shown that their disadvantages as measuring systems are the short range of parametric wave reception (hundreds of meters and only in some cases 1-2 kilometers) and the limited ability to measure the spatio-temporal characteristics of signals, which is especially evident when receiving waves of various physical nature in the low, infrasound and fractional frequency ranges.

[0006] Parametric antennas, the operation of which is based on low-frequency illumination (pumping) of the environment with weakly attenuating signals with a frequency of tens to hundreds of hertz, are extended volumetric zones of nonlinear interaction and parametric signal transformation formed in the marine environment. Which leads to an increase in the range of parametric wave reception by tens to hundreds of times, relative to high-frequency parametric antennas (see Mironenko M.V., Malashenko A.E., Vasilenko A.M. et al. Nonlinear translucent hydroacoustics and means of marine instrumentation in the creation of the Far Eastern radiohydroacoustic system for illuminating the atmosphere, ocean and earth's crust, monitoring their fields of various physical nature: monograph. - Vladivostok: Publishing house of the Far Eastern University, 2014. - 404 p.; Malashenko A.E., Mironenko M.V., Chudakov A.I., Pyatakovich V.A. Long-range parametric reception of electromagnetic waves generated by technical sources in the marine environment / / Sensors and systems - 2016.- No. 8-9 (206). - P. 14-18; Mironenko MV, Vasilenko AM, Pyatakovich VA Long-range parametric reception and classification of sea targets / / Marine collection. - 2017. - Vol. 2048, No. 11. - P. 72-75; Pyatakovich VA, Vasilenko AM, Rychkova VF Method of classification of underwater technical objects by an expert intelligent system with a receiving parametric antenna / / Marine intellectual technologies. - 2018. - No. 2 (40), Vol. 2. - P. 121-126; Pyatakovich VA, Vasilenko AM, Rychkova VF Promising methods for solving the scientific problem of target classification by a neural network expert system when monitoring the maritime situation / / Marine intellectual technologies. - 2018. No. 4 (42), Vol. 5. - P. 139-148; Pyatakovich V.A., Rychkova V.F., Filippov E.G. Neural networks as a variant of the computational structure of the sea target classification system / / Marine intellectual technologies. - 2020. - No. 1 (47), Vol. 2. - P. 163-174; Pyatakovich V.A., Surov A.B., Rychkova V.F.Calculation of the efficiency of target classification by an intelligent naval system using a complex of neural network computational operations / / Marine intelligent technologies. - 2020. - No. 1 (47), Vol. 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. - 2020. No. 4. - P. 72-79).

[0007] Low-frequency spatially developed parametric antennas are formed and operate based on the laws of multipath propagation of transmissive acoustic waves (marine pumping signals with a stabilized frequency in the tens to hundreds of hertz range) in an extended hydroacoustic channel with variable environmental characteristics and its boundaries. Long-range parametric reception of information waves is based on the laws of nonlinear interaction and parametric transformation of emitted transmissive waves with waves generated by technical objects (marine targets) during their joint propagation in the marine environment. The frequency range of received waves is tens to single kilohertz, hundreds to tens to single or fractions of hertz, including ultra-low frequency (ULF) oscillations of moving objects.

[0008] It is known that the result of parametric transformation of interacting waves is their mutual amplitude-phase modulation. A small difference in frequency (within one order of magnitude) between the transmission waves and the waves generated by the object ensures their most intense interaction. The amplitude of the interacting waves and the phase modulation index can be represented as follows:

[0009]

[0010] where γ is the nonlinearity coefficient of the marine environment; ω n , ω с - frequency of the pump wave and the useful signal, respectively; P n ,P c - attenuation 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 emission point to the location of the object; ρ0 is the density, c0 is the speed of sound in the marine environment.

[0011] The parametric components of the sum 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 dependencies and confirmed by the results of marine experiments (see Mironenko MV, Malashenko AE, Karachun LE, Vasilenko AM Low-frequency transmission method of long-range sonar of hydrophysical fields of the marine environment: monograph. - Vladivostok: SKB SAMI FEB RAS, 2006. - 173 p.; Pyatakovich VA, Vasilenko AM, Mironenko MV Technologies of nonlinear transmission hydroacoustics and neuro-fuzzy operations in problems of recognition of marine objects: - monograph. - Vladivostok: Far Eastern Federal University. - 2016. - 190 p. ISBN 978-5-7444-3790-9; Mironenko M.V., Pyatakovich V.A., Vasilenko A.M. Results of experimental studies of the method for determining the profile of a marine object and the system implementing it / / Monitoring. Science and Technology. - 2017.- No. 2 (31) - P. 64-69).

[0012] 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

[0013]

[0014] where P is the resulting and instantaneous values ​​of the modulated wave pressure, respectively; 2ω с - double frequency of the modulated wave; Ω - wave generated by the object; t - time; J n - Bessel functions of the n-th order; A n - amplitude of the modulated wave; m p - modulation coefficient. 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

[0015]

[0016] For small 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.

[0017] The disadvantage of the specified system is the absence in the structural diagram of the control path for the parameters of sample formation and reduction, which should ensure the reduction of the data dimensionality during the automatic tuning of the rule bases, based on the sample of the values ​​of the classification object parameters, the formation of a signal with the number of the production rule and the type of the membership function for the type of target for the optimization of the computational processes performed in the neural network recognition and classification path, providing the final classification decision for the detected sea targets (surface or underwater object).

[0018] The closest in technical essence to the claimed invention is the System for the operational identification of sea targets (patent No. 2763384 of the Russian Federation, IPC G01S 15 / 04, G01S 13 / 12 "System for the operational identification of sea targets by their information fields based on neuro-fuzzy models", published on 28.12.21, Bulletin No. 1), which includes a working zone of nonlinear interaction and parametric transformation of pump waves and information waves formed in the marine environment, wherein the length of the working zone is equal to the length of the controlled section of the marine environment, for which the emitting and receiving transducers are located on opposite boundaries of the section, while the input of the emitting transducer is connected by an underwater cable to the output of the pump signal radiation path, which contains a series-connected pump signal generator of a stabilized frequency, a power amplifier and a unit for matching its output with the underwater cable,and the output of the receiving converter is connected by an underwater cable to the input of the data signal reception, processing, and recording path, which contains a series-connected broadband amplifier, a frequency-time scale converter, a spectrum analyzer, and a recorder functionally connected thereto. The output of the spectrum analyzer of the data signal reception, processing, and recording path is connected to the input of the target class recognition unit based on the amplitude-frequency characteristics of the neural network recognition and classification path, covered by feedback with the training unit. The system additionally includes a path for regulating the parameters of sample formation and reduction, containing a block of fuzzy rules and functions, the input of which is connected to the output of the training unit of the neural network recognition and classification path, and the output is connected to the input of the logical device for sample formation and reduction, the function of which is performed by a combined recognition network,consisting of Kohonen-Grosberg networks and an adaptive neuro-fuzzy network ANFIS, using a hybrid learning algorithm, and covered by feedback with a fuzzification block, wherein the output of the logical device for forming and reducing samples is connected to the input of the fuzzification block, the output of which is connected to the input of the defuzzification and automatic regulation block of the rule base, which independently carries out automatic adjustment of its rule base, based on a sample of mathematical models of sea targets, at the output of which a signal of the number of a new production rule is formed, as well as a new type of membership function for the type of target for the training block of the neural network recognition and classification path, then at the output of the target recognition and classification block based on the amplitude-frequency characteristics of the neural network recognition and classification path, providing the final classification decision for the detected sea targets,a signal is generated according to the target type according to the degree of belonging of the studied spectral region to the classification object.

[0019] As is well known, extracting useful information from hydroacoustic signals defines the foundations of data processing algorithms. The operating algorithm of the logical unit for sample generation and reduction includes the following steps:

[0020] Initialization step. Set the initial data sample X=<х,у>.

[0021] Sample characteristics analysis stage. Determine - the minimum and maximum values ​​of the j-th feature, j=1, 2, …, N. Determine the number of intervals for each feature k, as well as the lengths of the intervals:

[0022] Stage of calculating generalized features. 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

[0023] Calculate the coordinate of the s-th instance along the generalized axis

[0024]

[0025] This allows mapping the original sample onto a one-dimensional generalized I-axis (note that this will result in some information being lost due to the implicit quantization of the feature space during the transformation).

[0026] Generalized axis analysis step. Generate a set of tuples I={ s , y s , s>}. Arrange the set I in non-decreasing order of the values ​​of I s By viewing the generalized axis in order of increasing its values, determine the boundary values ​​of its intervals in which the class number is y s remains unchanged, where 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.

[0027] Interval characteristics analysis stage. For each q-th interval of the generalized axis, k q ​=1, 2, …, k I , determine S q - the number of copies included in it, as well as the numbers of these copies.

[0028] Training sample generation stage. Among the instances of the q-th interval, include all instances in the training sample X*:

[0029] - classes located on one of the interval boundaries

[0030]

[0031] - class closest to one of the interval boundaries

[0032]

[0033] where α is a threshold coefficient that regulates the proximity of instances to the interval boundaries (for example, you can set:

[0034] - intervals with a small number of instances:

[0035]

[0036] s=1, 2, …, S, q=1, 2, …,k I ,

[0037] where β is some threshold coefficient, 0<β<1; - the average number of copies in the interval of the generalized axis.

[0038] Training set redundancy elimination step. Determine the distances between all instances included in the generated training set by creating a distance matrix R (to simplify and speed up calculations, we will use squared distances):

[0039]

[0040] Note that R(s, p) = R(p, s), and R(s, s) = 0.

[0041] Until perform in a loop of actions:

[0042] Find two instances with the smallest distance between them in the distance matrix:

[0043]

[0044] If two closest instances belong to the same class, then leave in the training set only the one that is closer to the instances of other classes, and exclude the other from it:

[0045]

[0046] Adjust 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 stage of supplementing (refining) the training sample.

[0047] The stage of supplementing (refining) the training sample. Determine the difference between the original and generated samples.

[0048] X' = X / X*

[0049] Sequentially for each s'th instance <x s ', y s '> samples X', s'=1, 2, …, S' relative to the instances of the formed sample X*, find the distance (the square of the distance) from it to each instance of the sample X*:

[0050]

[0051] If the closest instance to the s-th instance of the formed sample belongs to another class, then include it in the sample X*:

[0052]

[0053] As a result of executing this algorithm for the original sample X, we obtain a sample X', which will have the basic topological properties of the original sample.

[0054] Then, the mask method is used to form the feature vector, which is the input information array of the recognition network. 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 training process of the recognition network, and the second for recognizing targets (see Pyatakovich V.A., Bogdanov V.I., Nazarenko P.K. The principle of automatic target image recognition: Proceedings of the International Conference "Mathematical Modeling of Physical, Economic, Technical, Social Systems and Processes". - Ulyanovsk: Ulyanovsk State University, 2003. - pp. 31-32; 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: Maritime State University, 2017. - 255 pp.; Pyatakovich V.A., Vasilenko AM, Khotinsky O.V.Neural network technologies in intelligent systems for detection and operational identification of sea targets: monograph. - Vladivostok: Maritime state University, 2018. - 263 p.; Pyatakovich V.A., Vasilenko A.M., Pashkeev S.V. Automated system for monitoring the marine environment for solving problems of detecting a technical object / / Dual technologies. - 2018. No. 4 (85). - P. 85-88; Pyatakovich V.A., Rychkova V.F. Parametric optimization of a neural network system for classifying sea targets based on the reliability criterion / / Marine intelligent technologies. - 2018. No. 4 (42), Vol. 5. - P. 153-162; Pyatakovich V.A. Pilot model of the developed sample of the recognition module of the sea target classification system based on neural network technologies / / Problems and methods of development and operation of weapons and military equipment of the Navy: Collection of articles. - Vladivostok: TOVVMU, 2018. - Issue. 98. - pp. 176-181; Pyatakovich V.A.Solution of the problem of classification of sea targets using neural fuzzy networks based on the Mamdani-Zadeh inference model / / Problems and methods of development and operation of weapons and military equipment of the Navy: Collection of articles. - Vladivostok: TOVVMU, 2019. - Issue 101. - P. 275-284; Pyatakovich V.A., Filippov E.G. Method of formation and reduction of samples for solving problems of classification of sea targets by a neural classifier / / Problems and methods of development and operation of weapons and military equipment of the Navy: Collection of articles. - Vladivostok: TOVVMU, 2019. - Issue 101. - P. 312-320; Pyatakovich V.A. Training of deep neural networks of the system of operational identification of sea targets of the Russian Navy / / Problems and methods of development and operation of weapons and military equipment of the Navy: Collection of articles. - Vladivostok: TOVVMU, 2020. - Issue 104. - P. 177-186; Pyatakovich V.A., Rychkova V.F., Filippov E.G. Neural networks as a variant of the computational structure of the sea target classification system / / Marine intellectual technologies. - 2020. - No. 1 (47), Vol. 2.- P. 163-174; Pyatakovich V.A., Surov A.B., Rychkova V.F. Calculation of the efficiency of target classification by an intelligent naval system using a complex of neural network computing operations / / Marine intelligent technologies. - 2020. - No. 1 (47), Vol. 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. - 2020. - No. 4. - P. 72-79).

[0055] The idea behind the mask method is to find the maximum amplitude value for each mask, which is the unit vector of the classification features. To automate the process of finding the extremum in the zone of a single mask, a MAXNET maximum search network was used. Network iterations are completed after the network's output neurons stop changing. The input signal elements are integers or real numbers, while the output signal elements are real numbers. The dimensions of the input and output signals are the same. The activation function is linear with saturation (a linear segment is used). The number of synapses in the network is N(N-1). Synaptic weights are formed according to the formula

[0056]

[0057] 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).

[0058] The network operation is given by the expression

[0059]

[0060] where x j - element of the network input signal; y i - output of the j-th neuron; i=1, …, N; j=1, …, N.

[0061] Normalization of the input feature vector obtained after mask analysis by the MAXNET network is performed according to the expression

[0062] Value range limits are known and determined by the model of the input hydroacoustic signal.

[0063] The training of the recognition network is carried out on the basis of the backpropagation algorithm, which implements the gradient optimization method of the functional of the form: F=||Y(T,X*)-Y*|| 2, where T is the vector of synaptic weights of the network; (X*,Y*) are training pairs; ||…|| - vector norm (see Pyatakovich V.A., Vasilenko A.M. Prospects and Limitations of Using Geometric Methods for Recognizing Acoustic Images of Marine Objects as Applied to the Problem of Controlling a Neural Network Expert System / / Fundamental Research. - 2017. - No. 7. - Pp. 65-70; Pyatakovich V.A., Vasilenko A.M. Mironenko M.V. Training a Neural Network as a Stage in Developing an Expert System for Classifying Sources of Physical Fields When Monitoring Water Areas / / Bulletin of the Engineering School of the Far Eastern Federal University. - 2017. - No. 3 (32). - Pp. 138-149; Pyatakovich V.A. Classification System for Marine Targets Based on Neural Network Technologies / / Marine Intellectual Technologies. - 2018. - No. 4 (42), Vol. 5. - Pp. 169-176).

[0064] A disadvantage of the prototype system is the absence in the structural diagram of the control path for the activation of the executive system, which should ensure the automation of the decision-making process for the use of military and special equipment complexes.

[0065] The problem that the claimed invention is aimed at solving is the further development of the structure of the prototype system for its implementation as an intelligent system for supporting decision-making under conditions of uncertainty.

[0066] The technical result of the proposed invention is the automation of the decision-making process for activating the executive system of military and special equipment complexes through intelligent analysis of data on the degree of belonging of the studied spectral region to the classification object.

[0067] To solve the stated problem, the intelligent decision support system under uncertainty conditions contains a working zone of nonlinear interaction and parametric transformation of pump waves and information waves formed in the marine environment, wherein the length of the working zone is equal to the length of the controlled section of the marine environment, for which the emitting and receiving converters are placed on the opposite boundaries of the section, while the input of the emitting converter is connected by an underwater cable to the output of the pump signal radiation path, which contains a series-connected pump signal generator of a stabilized frequency, a power amplifier and a unit for matching its output with the underwater cable, and the output of the receiving converter is connected by an underwater cable to the input of the path for receiving, processing and recording information signals, which contains a series-connected broadband amplifier, a frequency-time scale converter,a spectrum analyzer and a recorder functionally connected thereto, wherein the output of the spectrum analyzer of the information signal reception, processing and recording path is connected to the input of the target class recognition unit based on the amplitude-frequency characteristics of the neural network recognition and classification path covered by feedback with the training unit, the output of which is connected to the input of the fuzzy rules and functions unit of the sample formation and reduction parameters regulation path covered by feedback with the fuzzification unit, wherein the output of the fuzzy rules and functions unit is connected to the input of the sample formation and reduction logical device, the function of which is performed by a combined recognition network consisting of Kohonen-Grosberg networks and an adaptive neuro-fuzzy network ANFTS using a hybrid learning algorithm, the output of which is connected to the input of the fuzzification unit, the output of which is connected to the input of the defuzzification and automatic regulation unit of the rule base,independently performing automatic tuning of its rule base, based on a sample of mathematical models of sea targets, at the output of which a signal is formed for the number of a new production rule, as well as a new type of membership function for the target type for the training block of the neural network recognition and classification path, then at the output of the target recognition and classification block based on the amplitude-frequency characteristics of the neural network recognition and classification path, which provides the final classification decision for the detected sea targets, a signal is formed for the target type according to the degree of belonging of the studied spectral region to the classification object.

[0068] The fundamental difference from the prototype is that an additional control path for activating the executive system is introduced, containing a logical device, the input of which is connected to the output of the target class recognition unit based on the amplitude-frequency characteristics of the neural network recognition and classification path, covered by feedback with the pulse generator of the sending unit, while the first output of the logical device is connected to the input of the pump signal generator of the stabilized frequency of the pump signal emission path, and the second output is connected to the input of the executive system activation unit, at the output of which a signal is generated for the use of military and special equipment complexes.

[0069] The use of neural networks to solve the problem of classifying surface and underwater sources of hydroacoustic signals is advisable, since they have high adaptive capabilities and are capable of learning to approximate multidimensional functions, that is, they can extract, in a form implicit to the user, knowledge from the subject area being studied.

[0070] The process of training a multilayer neural network is iterative and, in general, quite lengthy, since the number of iterations required to train the neural network cannot be determined in advance.

[0071] For neural network implementation of distance comparison and K value determination ij you can use the following expression:

[0072]

[0073] where - logistic function.

[0074] If function will be discrete, for example, threshold:

[0075]

[0076] to K ij will take the value 0 or 1.

[0077] If function will be real, for example, sigmoid:

[0078]

[0079] to K ij will take values ​​on the interval [0,1].

[0080] The closer the function value will be to 0, the closer the instance will be to the class that is associated with the value 0, and, accordingly, vice versa, the closer the value of the function will be to 1, the closer the instance will be to the class that is associated with the value 1.

[0081] Using the sigmoid function may be preferable in practice, since it allows us to determine not only which class an instance is closer to, but also how much closer it is.

[0082] To calculate the difference in distances the corresponding expressions are substituted:

[0083]

[0084] After mathematical transformations we get:

[0085]

[0086] where

[0087]

[0088] Expressions for are calculated based on a neuron that has one input to which the value x is fed i or x j , the weight of which is equal to

[0089]

[0090] The neuron threshold (zero weight) in this case will be equal to

[0091]

[0092] 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 must be determined on their basis according to the following rules.

[0093] Activation function kth neuron of the μth layer:

[0094]

[0095] Weighting coefficient p-th input of the k-th neuron of the μ-th layer:

[0096]

[0097] The output signal values ​​of the third layer form a decision vector

[0098] For the ANFIS neuro-fuzzy network, it is proposed to partition the feature space by executing the algorithm in the sequence below.

[0099] Step 1. Initialization. Set the training sample <x, y>.

[0100] Step 2. On the axis of each feature j=1, 2, …, N, determine the one-dimensional distances between instances:

[0101] Among the obtained distances, find the minimum distance greater than zero:

[0102]

[0103] Step 3. For each feature, determine the number of intervals for dividing the range of its values:

[0104]

[0105] 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.

[0106] Step 4. Split 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:

[0107]

[0108] Step 5. Form cluster blocks and assign their class numbers by performing steps 5.1-5.8.

[0109] Step 5.1. Form rectangular blocks in the N-dimensional feature space at the intersection of the corresponding intervals of feature values. Enter in B q,j the interval number of the j-th feature that corresponds to the q-th block.

[0110] Step 5.2. Determine class numbers for rectangular blocks in N-dimensional feature space:

[0111]

[0112] Set classification confidence coefficient for blocks:

[0113]

[0114] Step 5.3. For those blocks that have install: Where - the number of instances of the k-th class that fall into the q-th block cluster.

[0115] Step 5.4. For those blocks with 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.

[0116] Step 5.5. Calculate the distance between the q-th and p-th blocks, q=G+1, G+2, …, G+Q, p=q+1, G+Q, as:

[0117] Or

[0118] In this case, R(B q , Bp) = R(B p , B q ).

[0119] Step 5.6. Determine the potential induced by the set of blocks belonging to the k-th class on the p-th block with an unknown classification:

[0120]

[0121] where L k - the number of blocks belonging to the k-th class, S q - the number of training sample instances included in the q-th block.

[0122] Step 5.7. Set the class number for the p-th block with unknown classification (K p =0) according to the formula: p=G+1, G+2, …, G+Q.

[0123] Step 5.8. Modify the values ​​of the confidence coefficients for the blocks: q=G+1, G+2, …, G+Q.

[0124] Step 6. Merge adjacent cluster blocks.

[0125] Execute merge of adjacent blocks belonging to the same class: for

[0126] then combine blocks q and p by the j-th feature:

[0127] - install:

[0128]

[0129] - delete the p-th block:

[0130]

[0131] Step 7. From the combined set, select a subset of instances belonging to cluster blocks whose class numbers do not match the instance class numbers. Apply the partition refinement and model retraining procedure to the resulting partition and the selected subset.

[0132] If there is a partition of the feature space that needs to be refined (trained further) based on new observations 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 class number, adjusting S* accordingly. For those observations that do not coincide with the block classes, it is advisable to form separate point clusters.

[0133] For each new observation, form intervals based on the characteristics and enter them into interval numbers for each j-th feature,

[0134] corresponding to the new cluster, and also determine:

[0135]

[0136] The essence of the invention is explained by the drawings:

[0137] Fig. 1 shows the functional diagram of an intelligent decision support system under uncertainty.

[0138] Fig. 2 shows the general structure of the modified combined recognition network.

[0139] Fig. 3 shows the neural network interpretation of the multidimensional classification algorithm.

[0140] Fig. 4 shows Table 1 of the interpretation of the three-dimensional output vector of hydroacoustic signal recognition based on the amplitude-frequency characteristic.

[0141] Fig. 5 shows the mask method used to form the feature vector for the recognition network.

[0142] Fig. 6 shows a comparative characteristic of the methods for classifying a sea target in terms of training and classification time.

[0143] Fig. 7 and Fig. 8 show the results of a computational experiment to determine the recognition (classification) coefficient for surface and underwater objects under signal noise conditions.

[0144] The functional diagram of the intelligent decision support system under uncertainty conditions shown in Fig. 1 contains the following elements:

[0145] 1. Transmitter (underwater sound beacon type PZM-400, emitting signals at a frequency of about 400 Hz).

[0146] 2. Receiving converter.

[0147] 3. Marine environment.

[0148] 4. Working area of ​​nonlinear interaction and parametric transformation of pump waves and information waves.

[0149] 5. Objects (sea targets generating acoustic and / or electromagnetic and / or hydrodynamic emissions).

[0150] 6. Pump signal emission path.

[0151] 6.1. Stabilized frequency pump signal generator.

[0152] 6.2. Power amplifier.

[0153] 6.3. Matching block.

[0154] 7. Path for receiving, processing and recording information signals.

[0155] 7.1 Broadband amplifier.

[0156] 7.2. Time-frequency converter.

[0157] 7.3. Spectrum analyzer.

[0158] 7.4. Registrar.

[0159] 8. Neural network recognition and classification path.

[0160] 8.1. Target class recognition unit based on amplitude-frequency characteristics.

[0161] 8.2. Training block.

[0162] 9. Path for regulating parameters of sample formation and reduction.

[0163] 9.1. Block of fuzzy rules and functions.

[0164] 9.2. Fuzzification block.

[0165] 9.3. Logical unit for sample formation and reduction.

[0166] 9.4. Block of defuzzification and automatic regulation of the rule base.

[0167] 10. Executive system activation control path.

[0168] 10.1 Logical device.

[0169] 10.2. Sending pulse generator.

[0170] 10.3. Executive system activation block.

[0171] The input of the emitting converter 1 is connected by a cable to the output of the matching unit 6.3 of the pumping signal emission path 6, the input of which is connected to the power amplifier 6.2, the input of which is connected to the generator of pumping signals of a stabilized frequency 6.1, the output of the receiving converter 2 is connected by a cable to the input of the broadband amplifier 7.1 of the path for receiving, processing and recording information signals 7, the output of which is connected to the input of the frequency-time scale converter 7.2, the output of which is connected to the input of the spectrum analyzer 7.3, covered by feedback with the recorder 7.4, wherein the output of the spectrum analyzer 7.3 is connected to the input of the unit for recognizing the class of the target by the amplitude-frequency characteristics 8.1 of the neural network recognition and classification path 8, covered by feedback with the training unit 8.2, the output of which is connected to the input of the fuzzy rules and functions unit 9.1 control path for the parameters of sample formation and reduction 9, the output of which is connected to the input of the logical device for sample formation and reduction 9.3, the function of which is performed by a combined recognition network consisting of Kohonen-Grosberg networks and an adaptive neuro-fuzzy network ANFIS, using a hybrid learning algorithm, also the block of fuzzy rules and functions 9.1 is covered by feedback with the fuzzification block 9.2, wherein the output of the logical device for sample formation and reduction 9.3 is connected to the input of the fuzzification block 9.2, the output of which is connected to the input of the block of defuzzification and automatic regulation of the rule base 9.4, which independently carries out automatic tuning of its rule base, based on a sample of mathematical models of sea targets, at the output of which a signal of the number of a new production rule is formed, as well as a new type of membership function for the type of target for the training block 8.2 neural network recognition and classification paths 8, then at the output of the target recognition and classification block based on the amplitude-frequency characteristics 8.1 of the neural network recognition and classification path 8, which provides the final classification decision for the detected sea targets, a signal is generated based on the target type in accordance with the degree of belonging of the studied spectrum region to the classification object, which is then fed to the input of the logical device 10.1 of the activation control path of the executive system 10, covered by feedback with the sending pulse former 10.2, wherein the first output of the logical device 10.1 is connected to the input of the stabilized frequency pumping signal generator 6.1 of the pumping signal radiation path 6, and the second output is connected to the input of the executive system activation block 10.3, at the output of which a signal is generated for the use of military and special equipment complexes.

[0172] As shown in Fig. 2, on 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 fed The number of neurons in the second (hidden) layer is determined by the mutual arrangement and shape of the separated sets.

[0173] For each neuron of the second layer through synapses with weights {T ij (2)}, i=1 2, 3; j=1 2, 3, the output signals of the first layer are fed. The number of neurons in the third (output) layer is determined by the number of classes to be recognized.

[0174] 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 fed. The values ​​of the output signals of the third layer form a vector solutions. The neurons that make up the network are identical and have an activation function of a known type

[0175]

[0176] where X 2n (i) , yn (i) and I n (i) - the values ​​of the r-th input signal, output signal and external bias 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.

[0177] Pre-processing of 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.

[0178] Data preprocessing, one of the most labor-intensive steps, is necessary to ensure that subsequent learning algorithms can extract maximum knowledge from the sample.

[0179] The procedure of architectural self-tuning and weight tuning of the Kohonen network includes the following steps.

[0180] 1. An arbitrary number of neurons L=L0 is introduced with randomly normalized synapse vectors uniformly distributed over the interval [-1, 1].

[0181] 2. One of the training input signal vectors (obviously pre-processed) is fed to the layer, the potentials at the outputs of all neurons and the number of the "winner" L*-neuron are determined.

[0182] 3. The angle β* between the training feature vector and the weight vector of the “winner” neuron is determined.

[0183] 4. If the condition β*<β is met, then the synapses of the “winner” neuron are tuned by averaging over all training steps in which it turned out to be the “winner” neuron and subsequent normalization.

[0184] If β*>β, then L+1 neurons are introduced into the layer in a directive order, the connection weights of which are assigned the values ​​of the components of the corresponding training vector.

[0185] 5. The next input vector of the training sample is selected and procedures 2, 3, 4 are repeated.

[0186] Grosberg layer training is supervised learning and can be performed either simultaneously with architectural self-organization and tuning of the Kohonen layer for each input vector, or after training the Kohonen layer. In all cases, the learning rule can be represented by the following algorithm:

[0187]

[0188] where - component of the desired output vector of the classifier;

[0189] - the output signal of the j-th neuron of the Kohonen layer, when training the S-th input feature vector.

[0190] Pre-processed vector of input features of the observed object {x i} is fed to the network input and distributed through the weighting coefficients of connections (synapses) to the neurons of the Kohonen layer. The input potentials of neurons are calculated as i=1, …, L.

[0191] After this, the Kohonen layer begins to function as a competitive network with lateral connections. As a result, the layer's output is a vector with a single unit component corresponding to the "winning" neuron and zero other coefficients. Through synapses The input vector is fed to the neurons of the Grosberg output layer, which operate according to the following algorithm: 1=1, 2, where - unit jump function.

[0192] Figure 3 shows a multidimensional classification algorithm in a neural network interpretation based on a three-layer perceptron, which is a special case of a multilayer neural network, with the aim of achieving high accuracy in classifying images characterized by a not very large number of features from 3 to 20 pieces.

[0193] In Fig. 4 in Table 1 the elements of the solution vector are given The problem of recognizing and classifying surface and underwater sources of hydroacoustic signals, solved using a three-layer neural network, recognizes seven objects and allows for the identification of one unknown class, which in the future will significantly expand the range of recognizable marine technical objects.

[0194] As shown in Fig. 5, for each mask, the maximum amplitude value of the signal A1, A2, …, A is determined j , …, A k The choice of the Δ value and the number of masks is determined by the capabilities of the recognition network (typically 10–100). Increasing the number of masks increases the reliability of the input information and increases the complexity (increasing the number of neurons in the input layer) of the recognition device, thereby creating a tradeoff between quality and complexity. It is also possible to examine the noise profile in sections, i.e., the low-frequency, mid-frequency, and high-frequency components separately.

[0195] Figure 6 in Table 2 presents a comparative analysis of the methods for classifying sea targets based on training and classification times. The fastest method in terms of classification quality is the modified potential function method. Unlike potential function methods, the multivariate classification algorithm takes into account the significance of features by considering the partial significances of two-feature classifications.

[0196] Figures 7 and 8 show 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 under signal noise conditions in the range from - 10 dB to 20 dB. As can be seen from the graphs, the recognition and classification of sea targets using the computational operations of a combined recognition network consisting of the Kohonen (competitive network) and Grosberg networks and the adaptive neuro-fuzzy network ANFIS, which uses a hybrid learning algorithm, can increase the classification probability of both surface and underwater targets by 7-9%.

[0197] The intelligent decision support system under uncertainty operates as follows:

[0198] The emitting transducer 1 and the receiving transducer 2 are placed in the marine environment 3 taking into account the laws 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 transmission waves and waves of various physical nature generated by objects 5 (see Certificate of State Registration of the computer program "Ray pattern calculation" No. 2016616822 dated 06 / 21 / 2016; Certificate of State Registration of the computer program "Program for simulation modeling of the propagation of hydroacoustic signals" No. 2017664296 dated 12 / 20 / 2017; Certificate of State Registration of the computer program "Software and computing complex for simulation modeling of the marine information situation in target identification" No. 2018612944 dated (01.03.2018).

[0199] The stabilized frequency pump signal generated by generator 6.1 is fed to the input of power amplifier 6.2 of pump signal emission path 6, then to the input of matching unit 6.3, the output of which is connected by an underwater cable to the input of emitting converter 1.

[0200] The emitting transducer 1 sounds the environment with pumping signals of a stabilized frequency in the range of tens to hundreds of hertz.

[0201] Under various motion modes, objects 5 generate radiation that changes the conductive fluid's characteristics (density and / or temperature and / or heat capacity, etc.), which, depending on their physical nature, modulate the low-frequency pumping signals of the marine environment. Low- and high-frequency components appear in the information wave spectrum as a result of the amplitude and phase modulation of the low-frequency pumping wave by the radiation and fields of objects 5. Being an inextricably linked component of the transmission wave, the modulation components are transmitted over long distances and are detected in the units of the information signal reception, processing, and recording path 7.

[0202] The signal from the receiving converter 2 is fed via a cable line to the input of the broadband amplifier 7.1 of the information signal reception, processing and recording path 7. The task of the blocks included in the information signal reception, processing and recording path 7 is to measure the signs of the manifestation of information waves from sources.

[0203] The signal from the output of the 7.1 wideband amplifier is fed to the input of the 7.2 frequency-to-time scale converter. The frequency-to-time scale converter increases the energy concentration of the transmission signals and the efficiency of extracting from them the features of the fields generated by objects.

[0204] The signal from the output of the time-frequency converter 7.2 is fed to the input of the spectrum analyzer 7.3. The purpose of spectral analysis is to isolate discrete components of the sum or difference frequency in the narrowband spectra of the converted information signals, which are used to reconstruct the wave characteristics of objects 5.

[0205] Then the signal from the output of the spectrum analyzer 7.3 is transmitted to the input of the recorder 7.4 and to the input of the target class recognition unit by amplitude-frequency characteristics 8.1, which is connected by feedback to the training unit 8.2 of the neural network recognition and classification path 8. The signal from the target class recognition unit by amplitude-frequency characteristics 8.1 is sent to the memory of the training unit 8.2, where the results are compared with the mathematical images of the spectrograms of marine objects and corrected based on the signals coming from the path for regulating the parameters of sample formation and reduction 9. For this purpose, the signal from the output of the training unit 8.2 of the neural network recognition and classification path 8 is sent to the block of fuzzy rules and functions 9.1 of the path for regulating the parameters of sample formation and reduction 9.

[0206] The signal from the output of the fuzzy rules and functions block 9.1, covered by the feedback with the fuzzification block 9.2, is fed to the input of the logical device for sample formation and reduction 9.3, and then to the fuzzification block 9.2, and then to the input of the block for defuzzification and automatic regulation of the rule base 9.4 of the path for regulating the parameters of sample formation and reduction 9, which independently carries out automatic tuning of its rule base, based on a sample of mathematical models of sea targets, at the output of which a signal of the number of a new production rule is formed, as well as a new type of membership function for the target type for the training block 8.2 of the path of neural network recognition and classification 8.

[0207] The function of the logical unit for sample generation and reduction 9.3 of the control path for the parameters of sample generation and reduction 9 is performed by a combined recognition network consisting of Kohonen-Grosberg networks and an adaptive neuro-fuzzy network ANFTS, which uses a hybrid learning algorithm. In a situation of uncertainty, i.e., the influence of random external disturbances, the defuzzification and automatic regulation unit of the rule base 9.4 compensates for them in the rule base, as well as for the membership functions of the new type, in accordance with the new conditions for classifying the sea target. Thus, the logical unit for sample generation and reduction 9.3 determines the rule number N required for replacement in the main rule base, as well as the new type of membership function for this rule.

[0208] Further, at the output of the target recognition and classification unit based on the amplitude-frequency characteristics 8.1 of the neural network recognition and classification path 8, a signal is formed based on the target type (surface or underwater object), according to the degree of belonging of the studied spectrum region to the classification object, which is fed to the input of the logical device 10.1 of the control path for activating the executive system 10, in which the values ​​of the control signals (commands) are calculated for the sending pulse generator 10.2, which in turn generates a sending pulse to adjust the operating modes of the proposed system to the changing conditions of the signal propagation environment and the diversity of manifestations of information wave sources, and through the logical device 10.1 feeds it to the input of the stabilized frequency pumping signal generator 6.1 of the pumping signal radiation path 6 or generates a sending pulse and through the logical device 10.1 feeds it to the input of the activation unit of the executive system 10.3 of the control path for activating the executive system 10 for the use of military and special equipment (see Certificate of State Registration of the computer program "Specialized neural network complex for classifying noisy signals of sea targets" No. 2018619739 of the Russian Federation dated 10.08.2018; Certificate of State Registration of the computer program "Software complex for modeling and training INS" No. 2019611455 dated 28.01.2019; Certificate of State Registration of the computer program "Program for designing and training artificial neural networks of the perceptron type" No. 2019611559 dated 29.01.2019; Pyatakovich V.A. Training of a fuzzy neural network of the sea target classification system with adjustment of weighting coefficients / / Problems and methods of development and operation of weapons and military equipment of the Navy: Collection of articles. - Vladivostok: TOVVMU, 2020. - Issue 104. - P. 187-192).

[0209] Thus, the intelligent decision support system in conditions of uncertainty, having detected a target based on the amplitude-phase modulation characteristics of low-frequency signals pumping the marine environment with the object's radiation and fields, allows, based on the intelligent analysis of data on the degree of belonging of the studied spectrum region to the classification object (due to the promptly updated library of spectrogram images of sea targets and a combined recognition network consisting of Kohonen networks (competing network), Grosberg and the adaptive neuro-fuzzy network ANFIS, using a hybrid learning algorithm), to automate the decision-making process for the use of military and special equipment, as well as to adjust their operating modes to the changing conditions of the signal propagation environment and the diversity of manifestations of information wave sources.

[0210] The intelligent decision support system under uncertainty is industrially applicable, as it is created using common components and products from the radio engineering industry and computing technology.

Claims

An intelligent decision support system under uncertainty conditions, comprising a working zone of nonlinear interaction and parametric transformation of pump waves and information waves formed in a marine environment, wherein the length of the working zone is equal to the length of the monitored section of the marine environment, for which purpose the emitting and receiving converters are placed on opposite boundaries of the section, wherein the input of the emitting converter is connected by an underwater cable to the output of the pump signal emission path, which contains a series-connected pump signal generator of a stabilized frequency, a power amplifier and a unit for matching its output with the underwater cable, and the output of the receiving converter is connected by an underwater cable to the input of the path for receiving, processing and recording information signals, which contains a series-connected broadband amplifier, a frequency-time scale converter,a spectrum analyzer and a recorder functionally connected thereto, wherein the output of the spectrum analyzer of the information signal reception, processing and recording path is connected to the input of the target class recognition unit based on the amplitude-frequency characteristics of the neural network recognition and classification path covered by feedback with the training unit, the output of which is connected to the input of the fuzzy rules and functions unit of the sample formation and reduction parameters regulation path covered by feedback with the fuzzification unit, wherein the output of the fuzzy rules and functions unit is connected to the input of the sample formation and reduction logical device, the function of which is performed by a combined recognition network consisting of Kohonen-Grosberg networks and an adaptive neuro-fuzzy network ANFIS, using a hybrid learning algorithm, the output of which is connected to the input of the fuzzification unit, the output of which is connected to the input of the defuzzification and automatic regulation unit of the rule base,independently performing automatic tuning of its rule base, based on a sample of mathematical models of sea targets, at the output of which a signal of the number of a new production rule is formed, as well as a new type of membership function for the type of target for the training unit of the neural network recognition and classification path, then at the output of the target recognition and classification block according to the amplitude-frequency characteristics of the neural network recognition and classification path, providing the final classification decision for the detected sea targets, a signal is formed according to the type of target according to the degree of belonging of the studied spectral region to the classification object, characterized in that an additional control path for the activation of the executive system is introduced, containing a logical device, the input of which is connected to the output of the target class recognition block according to the amplitude-frequency characteristics of the neural network recognition and classification path,covered by feedback with the pulse generator of the sending pulse, wherein the first output of the logical device is connected to the input of the pump signal generator of the stabilized frequency of the pump signal emission path, and the second output is connected to the input of the activation unit of the executive system, at the output of which a signal is generated for the use of military and special equipment complexes.

Images