Method of radar recognition of class of ballistic targets

Abstract

FIELD: radar systems.SUBSTANCE: invention can be used to recognize a class of ballistic targets. The claimed method includes preliminary training of a neural network, reception of a radar signal reflected from a target by a radar system, primary processing of the received signal with target detection. After detection, the target is put into tracking, during which a multidimensional time series is formed from its primary parameters by repeatedly measuring and accumulating them. The generated multidimensional time series is then fed to the inputs of a pre-trained neural network, processed in the neural network, and a decision on the target class is formed at the output of the neural network. To train the neural network, mathematical models of the movement of each type of ballistic target are developed in turn, as well as a model of the radar system that receives signals from the target movement model. In the radar model, its detection, tracking and formation of a multidimensional time series of the primary parameters of the target are carried out, then the detection, tracking and formation of a multidimensional time series are repeated many times with other values of the primary parameters for the same target, thus forming a multitude of multidimensional time series with different trajectories of the ballistic target. The multidimensional time series generated during the training process for each type of ballistic target are fed to the inputs of the neural network, where a matrix of weight coefficients of the trained neural network is formed, used for recognizing the target class.EFFECT: increase in the probability of recognizing a class of ballistic targets while significantly simplifying the method of radar recognition of ballistic targets.1 cl, 3 dwg

Description

[0001] The invention relates to radar and can be used to recognize a class of ballistic targets.

[0002] The “Method of radar recognition of classes of aerospace objects for a multi-band diversity radar complex with phased antenna arrays” is known [RU 2741057 C1 published on 22.01.2020, MGZ G01S 13 / 52, G011S 13 / 02], in which the processing of radar information (RI) is carried out by starting the process of searching, detecting and tracking aerospace objects (ASO) in a given field of view, information about the ASO echo signals detected by radar modules (RLM) of the meter and decimeter wavelength range, containing the range, azimuth, elevation angle and amplitude of the echo signals at each probing frequency, as well as information on the nationality of the ASO, is sent to the processing unit of the radar complex RLC. In the radar processing unit, further recalculation of coordinates into a rectangular system is carried out, the value of the vertical component of the velocity (V) is determined. Hi ), the value of the VKO track speed (V) is calculatedTi ), based on previously obtained calculations, preliminary radar recognition of classes (RRC) of air defense systems is carried out based on trajectory characteristics, while information on the altitude of the air defense system, its vertical component of speed and track speed is compared with a priori specified information on the possible values ​​of these characteristics for each class of target. Moreover, if a final decision on the classification of the air defense system cannot be made based on trajectory characteristics or the probability of such recognition is insufficient, then recognition is carried out based on a signal characteristic: the effective scattering surface (ESR) of the air defense system, which is estimated based on data on the range and elevation angle of the target, the amplitude of its echo signals and the a priori dependence of the range of a target with an RCS of 1 m 2from the elevation angle. Next, they recognize the target based on its longitudinal size, which is calculated based on an analysis of the target echo signal amplitudes at each long-range radar frequency for a preliminary determination of the target size, and a precise determination of the size of a short-wavelength radar with an active phased array (PAR), which maintains prolonged contact with the target in the beam-stop mode and recognizes it with a higher probability. In the case of electronic countermeasures (ECM) and a reduction in the maximum radar reconnaissance range of a single-position radar, the use of multi-band, spatially distributed radar systems that can operate in both passive and active modes, or switch between each other according to a predetermined program, and recognition can be performed based on the processing of signals from both spaced long-wavelength radars and spaced short-wavelength radars.The integration of radars of different wavelengths at combined and spaced positions into a single radar system is achieved by mutually referencing all system assets and jointly processing the radar data received by each radar. Furthermore, the presence of a digital active phased array (PAA) operating in reception allows for the simultaneous formation of a fan-shaped directional pattern in the azimuth and elevation planes, covering the possible ranges of airborne targets illuminated by the active radar system. Due to its multi-position capability, it reduces the pulse volume, maintains the detection range and range of the air defense radar system in the face of electronic countermeasures (ECM) and active noise interference, provides additional recognition indicators, and increases the probability of accurate air defense radar system detection.

[0003] The disadvantages of this method are the complexity of its technical implementation and the large volume of calculations performed by the radar complex.

[0004] The closest in technical essence is the "Method of multi-feature recognition in a multifunctional radar station of an aircraft class based on the principle of "aircraft with a turbojet engine - aircraft with a turboprop engine - helicopter - missile - unmanned aerial vehicle" based on the combined use of Coleman filtering and a neural network" [RU 2832712 published on 28.12.2024, IPC G01S 7 / 41, G01S 13 / 52], which consists in the fact that the radar signal reflected from the aircraft, from the output of the receiver of the multifunctional radar station at an intermediate frequency, is subjected to narrowband Doppler filtering based on the fast Fourier transform procedure and is converted into an amplitude-frequency spectrum, the spectral components of which are caused by reflections of the radar signal from the airframe of the aircraft. By threshold processing the amplitude-frequency spectrum of the signal, only those F samples are formed q, where q=1, …, Q; Q is the number of aircraft classes and Doppler frequencies at which the spectral component amplitudes exceed the set threshold. In neural network training mode, in accordance with the q-th dynamic Doppler frequency models:

[0005]

[0006]

[0007] where F q and ΔF q - respectively, the deterministic and fluctuation components of the Doppler frequency for an aircraft of the q-th class;

[0008] F* q (t) is the derivative of the Doppler frequency for a q-class aircraft;

[0009] α q - a value inversely proportional to the Doppler frequency correlation time for a q-class aircraft;

[0010] β q - the square of the natural frequency of the autocorrelation function of the Doppler frequency for a q-class aircraft;

[0011] - the variance of the derivative of the Doppler frequency fluctuations for a q-class aircraft, defined as

[0012]

[0013] where - dispersion of Doppler frequency fluctuations for a q-class aircraft;

[0014] n q - mutually independent forming “white” Gaussian noises with zero mathematical expectations and unit intensities for an aircraft of the q-th class;

[0015] Doppler frequency samples are generated sequentially in discrete time in the form

[0016]

[0017] where r=1, …, R is the current clock cycle for generating Doppler frequencies;

[0018] R - total number of Doppler frequency generation cycles;

[0019] T - time discrete value,

[0020] by which, in R cycles, the values ​​of the corresponding g-th autocorrelation function of the Doppler frequency velocity fluctuations are calculated and the parameters of the autocorrelation function α are determined by it q , β q And In the aircraft class recognition mode, Kalman filtering of real Doppler frequency readings is performed in accordance with the expressions

[0021] P - (r+1)=Ф(r)Р(r)Ф T (r)+G(r);

[0022] Ψ(r+1)=Н(r)Р - (r+1)H T (r)+R (r);

[0023] S(r+1)=P - (r+1)H T (r)Ψ - (r+1);

[0024] Y(r+1)=H(r)X(r)+Y(r+1);

[0025]

[0026] P(r+1)=[I - S(r+1)H(r)]P - (r+1);

[0027] where P - (r+1) and P(r+1) are the covariance matrices of extrapolation and filtering errors, respectively, of dimension n×n, n is the dimension of the state vector X(r+1);

[0028] Ф(r) - state transition matrix of dimension n×n;

[0029] G(r) and R(r) are the covariance matrices of excitation and observation noise of dimensions n×n and m×m, respectively, m is the dimension of the observation vector Y(r+1);

[0030] Y(r) is a column vector of observation noises, which are Gaussian “white” sequences with zero mathematical expectations and spectral density matrices N r (r) of dimension m;

[0031] H(r) is an observation matrix of dimension m×n

[0032] Z(r+1) - matrix of measurement residuals of dimension m;

[0033] S(r+1) - matrix of weight coefficients of dimension n×m;

[0034] I is the identity matrix of dimension n×n;

[0035] «т» - transposition symbol;

[0036] «-1» - the symbol for finding the inverse matrix;

[0037] «∧» is the symbol for finding the assessment,

[0038] with the corresponding dynamic model, resulting in the corresponding estimate being formed at the output of each q-th Coleman filter fluctuations of the Doppler frequency, after R cycles of operation of all Coleman filters, based on the results of the neural network operation, a preliminary decision is made on the q-th class of the aircraft with the corresponding probability P q , which is compared with the threshold value Pth, when the condition is met for each probability value P q ≥P пop A final decision is made that the aircraft belongs to class q; otherwise, a decision is made that there is no aircraft of this class. In the neural network training mode, the corresponding values ​​of the coefficients K are calculated. q for an aircraft of class q in accordance with the expression

[0039]

[0040] which, together with additional recognition features

[0041] H q ,V q , ΔF Д , ΔF Л , ΔF B , H q - flight altitude of the q-class aircraft; V q - flight speed of the q-class aircraft; - Doppler frequency values ​​caused by the reflection of a radar signal from the rotating blades of a low-pressure compressor wheel of an aircraft of the “turbojet” and “turboprop” classes; ΔF Д - the range of Doppler frequencies caused by reflections of a radar signal from the engine of a helicopter-class aircraft; ΔF Л - the range of Doppler frequencies caused by reflections of the radar signal from the blades of a helicopter-class aircraft; ΔF B- the Doppler frequency range caused by radar signal reflections from the propellers of an aircraft of the “turboprop aircraft” class is fed to the corresponding input neurons of the neural network for its training for each aircraft of the q-th class, in the aircraft class recognition mode based on estimates fluctuations of the Doppler frequency, the corresponding estimate of the autocorrelation function is calculated and based on it, in accordance with the expression

[0042]

[0043] where estimation of the variance of the Doppler frequency fluctuation estimate for a q-class aircraft;

[0044] an estimate of the value inversely proportional to the correlation time of the Doppler frequency estimate for a q-class aircraft; estimate of the square of the natural frequency of the autocorrelation function of the Doppler frequency estimate of a q-class aircraft; corresponding estimates of the coefficients which have additional characteristics H q , V q , ΔF Д , ΔF Л , ΔF B for an aircraft of the q-th class, are fed to the corresponding inputs of the neural network to make a preliminary decision on the q-th class of the aircraft in R cycles of operation of all Coleman filters, while the architecture of the neural network is a multilayer direct propagation neural network consisting of four layers: the first input layer of the neural network, to the input of which the values ​​of the coefficients K with six additional features H are sequentially fed during training of the neural network q , V q , ΔF Д , ΔF Л , ΔF B and in the aircraft class recognition mode - assessments and six additional signs of H q , V q , ΔF Д , ΔF Л , ΔF B the second and third hidden layers of the neural network in the form of layers with nonlinear activation functions, and the fourth output layer of the network, at the qx outputs of which estimates of P are formed q probabilities of preliminary recognition of the q-th class of aircraft. This classification method contains three stages:

[0045] 1. Filtration.

[0046] 2. Selection of classification features.

[0047] 3. Direct classification using a neural network based on selected features.

[0048] The classification features are selected empirically and thus do not take into account all the information about the target's trajectory, which makes it impossible to apply the specified method for recognizing the class of ballistic targets, and Coleman filtering complicates the implementation of the method.

[0049] The technical problem solved by the proposed invention is the creation of a method for radar recognition of a class of ballistic targets with a high recognition probability, maximizing the information contained in the target's trajectory. The technical result achieved herein is an increase in the recognition probability of a class of ballistic targets while significantly simplifying the method for radar recognition of ballistic targets.

[0050] The essence of the proposed method for radar recognition of a class of ballistic targets is that it includes preliminary training of a neural network, reception of a radar signal reflected from a target by a radar system, primary processing of the received signal with target detection, formation of a multidimensional time series from primary parameters for the period of observation of the target, feeding the formed multidimensional time series to the inputs of a pre-trained neural network, processing it in the neural network and formation of a decision on the class of the target at the output of the neural network.

[0051] What is new in the claimed method is that after detection, the target is put on tracking, during which multiple measurements and accumulation of the primary parameters of the target are performed: the effective reflection area (ERA) σ Ц , range D Ц , azimuth ϕ Ц , elevation angle θ Ц and Doppler frequency F ДЦand form a multidimensional time series from the measured primary parameters σ Ц ., D Ц , ϕ Ц , θ Ц , F ДЦA pre-trained three-layer neural network is used, the number of inputs of which is determined by the number of measured primary target parameters. To train the neural network, mathematical models of the movement of each type of ballistic target are developed in turn: mortar shells, artillery shells, and multiple launch rocket systems (MLRS) shells. A radar system model is also developed that receives signals from the target's movement model. The radar model detects, tracks, and generates a multidimensional time series of the target's primary parameters. The detection, tracking, and generation of the multidimensional time series are then repeated multiple times for different values ​​of the primary parameters for the same target, thus generating multiple multidimensional time series with different ballistic target trajectory parameters, such as range, direction, and observation noise.The multidimensional time series generated during the training process for each type of ballistic target are fed to the inputs of the neural network, where a matrix of weight coefficients of the trained neural network is formed, used for recognizing the target class.

[0052] Fig. 1 shows a functional diagram of a variant of elements of a radar system that implements a method of radar recognition of a class of ballistic targets.

[0053] Fig. 2 shows the operating algorithm of the ballistic target class recognition system.

[0054] Fig. 3 shows the neural network training algorithm.

[0055] The method for radar recognition of a class of ballistic targets can be implemented in a radar system (RLS), the elements of which are shown in Figure 1, and including an antenna (1), a receiver (2), a target recognition system (3), which includes a computing system (4) and a neural network (5). The output of the antenna (1) is connected to the input of the receiver (2). The output of the receiver (2) is connected to the input of the computing system (4), which is part of the target recognition system (3). The output of the computing system (4) is connected via a data bus to the first, second, third, fourth and fifth inputs of the neural network (5), which is part of the target recognition system (3). The output of the neural network (5) is the output of the target recognition system (3).

[0056] The method of radar recognition of the class of ballistic targets is carried out as follows.

[0057] During operation (survey), the radar system receives a radar signal reflected from a ballistic target, for example, a multiple launch rocket system (MLRS) projectile, by an antenna (1). The signal from the antenna (1) is fed to the receiver (2). In the receiver (2), amplification, conversion to an intermediate frequency, and analog-to-digital conversion of the signal, and its decomposition into quadrature components are performed. From the output of the receiver (2), the signal in digital form is fed to the input of the computing system (4) as part of the target recognition system (3). The computing system (4) performs primary processing of the received signal with target detection. In this embodiment of the invention, the primary processing may include demodulation of the received signal, signal compression by range, and spectral analysis of the received signal sample. Target detection is performed when the signal exceeds a predetermined threshold.

[0058] Figure 2 shows the target recognition system (3) operating algorithm. In the computing system (4), after the initial target detection (6), the target is set for tracking (7). During target tracking (7), by processing the detected target signal, multiple measurements of the target's primary parameters are performed, for example, the effective reflection area (ERA) σ Ц , range D Ц , azimuth ϕ Ц , elevation angle θ Ц and Doppler frequency F ДЦ , and thus form a multidimensional time series from the primary parameters of the target (8) σ Ц , D Ц , ϕ Ц , θ Ц , F ДЦduring the target observation period. The generated multidimensional time series of the primary parameters from the output of the computing system (4) is transmitted via the data bus to the inputs of a pre-trained neural network (5), for example a three-layer neural network. The number of inputs of the neural network (5) is determined by the number of primary measurable target parameters; accordingly, in this embodiment of the invention, the neural network (5) has five inputs. The neural network (5) processes (9) the multidimensional time series of the primary target parameters, automatically extracting from it the features characteristic of each target class. The neural network (5) determines the features that best correspond to a certain target class and makes a decision on the target class, in this case, an MLRS projectile. From the output of the neural network (5), the data on the target class are sent for display to the operator of the radar system or to external systems.

[0059] Figure 3 shows the training algorithm of the neural network (5). Preliminary training of the neural network (5) is carried out using time series of primary target parameters obtained using mathematical modeling. For this purpose, mathematical models of target motion (10) of the target classes of interest, for example, MLRS, are developed, as well as a radar model (11), which receives signals from the target motion model (10). In the radar model (11), target detection (12), target tracking (13) and the formation of a multidimensional time series of primary target parameters (14) are carried out. Using these models, through multiple repetitions of detection, tracking and the formation of a multidimensional time series with different values ​​of the primary parameters for the same target, thus forming a multitude of multidimensional time series with different trajectories of a ballistic target (range, direction, observation noise, etc.).

[0060] Time series simulated during training (σ oЦ , DоЦ , ϕ оЦ , θ оЦ , F оДЦ ) are fed to the corresponding inputs of the neural network (5). During the training of the neural network (15), a matrix of weighting coefficients of the trained neural network (16) is formed, which is used in recognizing the target class in the target recognition system (3) as distinctive features of target classes.

[0061] The number of classes a neural network can recognize (5) is determined by the number of targets it has been pre-trained on. For example, the following ballistic target classes can be distinguished: mortar shell, artillery shell, MLRS shell, and their specific types depending on caliber or type.

[0062] A significant advantage of using a neural network is the absence of empirically selected recognition features. During training, the neural network independently identifies distinguishing features between all classes of ballistic targets based on the primary target parameters, and during operation, it recognizes the class of ballistic targets based on these parameters. This increases the probability of recognizing a class of ballistic targets and simplifies the implementation of the proposed recognition method.

Claims

A method for radar recognition of a class of ballistic targets, including preliminary training of a neural network, reception of a radar signal reflected from a target by a radar system (RLS), primary processing of the received signal with target detection, formation of a multidimensional time series from primary parameters over a period of target observation, feeding the formed multidimensional time series to the inputs of a pre-trained neural network, processing it in the neural network and formation of a decision on the class of target at the output of the neural network, characterized in that after detection the target is put on tracking, during which multiple measurements and accumulation of the primary parameters of the target are performed: the effective reflection area (ERA) σ Ц , range D Ц , azimuth ϕ Ц , elevation angle θ Ц and Doppler frequency F ДЦ and form a multidimensional time series from the measured primary parameters σ Ц , D Ц , ϕЦ , θ Ц , F ДЦ, use a pre-trained three-layer neural network, the number of inputs of which is determined by the number of measured primary parameters of the target, and to train the neural network, mathematical models of the movement of each type of ballistic target are developed in turn: a mortar shell, an artillery shell, a multiple launch rocket system (MLRS) shell, and a model of a radar system is also developed, which receives signals from the target movement model, in the radar model, its detection, tracking and the formation of a multidimensional time series of the primary parameters of the target are carried out, then the detection, tracking and the formation of a multidimensional time series are repeated many times with other values ​​of the primary parameters for the same target, thus forming a multitude of multidimensional time series with different parameters of the trajectories of the ballistic target: range, direction, observation noise,The multidimensional time series generated during the training process for each type of ballistic target are fed to the inputs of the neural network, where a matrix of weight coefficients of the trained neural network is formed, used for recognizing the class of the target.

Images