Hypersonic target pattern recognition method driven by wave physics insights

Abstract

The present invention relates to a hypersonic target pattern recognition method driven by wave physics insights, and relates to the field of classification and recognition of hypersonic aircraft. Six pattern recognition feature values ​​containing plasma wave attenuation and wave propagation properties are obtained, including the real and imaginary values ​​of scattered echoes in the three component directions of x, y, and z; the gaze rotation change rate of radar tracking hypersonic targets with different flight modes at any flight time is obtained as the seventh effective feature value of hypersonic target pattern recognition; the total area value of the target illuminated surface at different time points with a fixed percentage is obtained as the eighth effective feature value of hypersonic target pattern recognition; the eight effective feature values ​​are formed into a feature vector library, which is input into a recognition classifier for training; and the trained recognition classifier is used to perform high-precision classification of different flight modes of hypersonic targets. The present invention is superior to the existing traditional black box technology, and the recognition accuracy rate is nearly 50% higher than the latter.

Description

Technical Field

[0001] The present invention belongs to the field of classification and identification of hypersonic aircraft, and more specifically, relates to a high-precision classification and identification method for flight modes of hypersonic targets coated with plasma sheaths based on wave scattering and physical feature extraction. Background Art

[0002] Modern new-type space-based early warning and detection systems can implement navigation positioning, information defense, detection and tracking and other major defense strategies for dense high-speed and ultra-high-speed aircraft by configuring a group of satellite-borne radar arrays that are rapidly deployed, highly maneuverable and highly survivable. However, when an aircraft is flying at ultra-high speed in near-space, the gas molecules around the aircraft will be excited and ionized, and the outer surface of the fuselage will be covered with a thick layer of sheath-type plasma medium. This complex medium will absorb and attenuate the detection radar waves emitted by the satellite-borne radar, and even cause the electromagnetic waves to deviate from the propagation trajectory of the original scattering path, causing the signal echo intensity received by the radar to be greatly weakened and accompanied by irregular and chaotic changes, and the main electromagnetic information of the detected target will be lost, which brings great challenges to the construction of an electromagnetic defense equipment system enabled by hypersonic weapon early warning detection and tracking identification.

[0003] The detection frequency bands of the current satellite-borne space observation radars are mainly concentrated in the LX band range, and the main detection means is to emit multi-band continuous plane electromagnetic waves. Most of the electromagnetic detection and identification technology research based on the principle of electromagnetic scattering does not consider the use of the underlying formula mechanism of wave physics, but relies more on the stacking and brute force training of a large amount of detection data. At present, the research on the recognition of non-cooperative unknown flying targets by the space observation radars of the current satellite platforms is basically based on simple satellites, missiles or aircraft as research models. There are few related research reports on hypersonic aircraft flying in near space, and the complex physical phenomenon of plasma sheath is basically unable to be involved. In addition, compared with small targets such as hypersonic aircraft, most satellite-borne detection radars are low-resolution radars with limited detection capabilities and imaging accuracy, and cannot track hypersonic targets throughout the entire process. Therefore, it is of great practical significance to explore and study the dynamic electromagnetic scattering theory and feature extraction technology based on wave physics insights in the low-resolution satellite-borne radar system to solve the problem of rapid pattern recognition of hypersonic targets.

[0004] The problem of hypersonic target classification and recognition is essentially a typical inverse problem. Based on the electromagnetic scattering mechanism, the radar system detects the echo signals of complex targets and environments, extracts target features through signal processing algorithms, and determines the target type and specific attributes. Bian et al. established a hypersonic target flight trajectory model combined with the basic theory of dynamic electromagnetic scattering, extracted ten types of effective mathematical statistical eigenvalues ​​from the dynamic radar cross section (DRCS) time series using continuous wavelet transform (CWT) technology, and established a hypersonic target feature recognition database. Li et al. proposed a classification training model based on the WassersteinGAN (U-WGAN-GP) framework with gradient penalty, and successfully applied it to the effective classification of high / low resolution (HR / LR) image pairs covered with plasma sheath obtained by inverse synthetic aperture radar (ISAR). Ma et al. proposed a low-resolution radar space target recognition algorithm based on discrete wavelet transform for six types of space targets, including two types of three-axis stabilized satellites, two types of spin-stabilized satellites and two types of debris, under the low-resolution radar system.

[0005] At present, most of the research focuses on space targets. Dynamic electromagnetic scattering echoes or ISAR imaging technologies based on complete time links are used to obtain a large number of rough data sources for feature extraction and back-end recognition, which greatly increases the cost of calculation and time, and is not conducive to application in real-time aerospace over-the-horizon reconnaissance scenarios. In addition, it is difficult to form a one-to-one mapping relationship between the target's usual low-resolution microwave imaging or specific mathematical statistical feature-level rough data and the target state or category, which can be understood as a complex mapping relationship that is difficult to describe with a specific mathematical formula. Summary of the invention

[0006] The technical problems to be solved by the present invention are:

[0007] In view of the serious overlap of scattering eigenvalue distribution, complex calculation process and high resource consumption of hypersonic targets covered with plasma sheath in different flight modes, which leads to problems such as poor accuracy, slow response and large amount of training data in hypersonic target pattern recognition, the present invention provides a high-precision classification and recognition method for hypersonic targets in different flight modes inspired by wave physics.

[0008] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0009] A hypersonic target pattern recognition method driven by wave physics insights, characterized by comprising:

[0010] Acquire six pattern recognition feature values ​​containing plasma wave attenuation and wave propagation properties, wherein the pattern recognition feature values ​​include: real and imaginary values ​​of scattered echoes in three component directions of x, y, and z;

[0011] Obtain the gaze rotation change rate of the radar tracking hypersonic targets with different flight modes at any flight time as the seventh effective eigenvalue of hypersonic target pattern recognition;

[0012] Obtain the total area value of the target illuminated surface at different time points with a fixed percentage as the eighth effective eigenvalue of hypersonic target pattern recognition;

[0013] The above eight valid eigenvalues ​​are combined into a feature vector library and input into the recognition classifier for training;

[0014] The trained recognition classifier is used to perform high-precision classification of different flight modes of hypersonic targets.

[0015] A further technical solution of the present invention is as follows: the step of obtaining six pattern recognition feature values ​​containing plasma wave attenuation and wave propagation properties comprises:

[0016] Establish a multi-layer lossy medium model MLMM to express the static scattering characteristics of a hypersonic target covered with a plasma sheath;

[0017] The angle-time conversion algorithm is used to add the time dimension to the RCS data after the angle transformation quasi-static obtained by MLMM, and the DRCS time series of the target aircraft at different flight altitudes and flight speeds are calculated;

[0018] The DRCS time series of the target in different flight modes received by the radar sensor is separated into scattered echoes, and the scattered electric field E that varies with time is obtained by forward simulation through MLMM. s (r) Discrete data: Based on the PO scattered field integral formula, the vertical / horizontal polarization processing of the scattered echo is removed and the scattered echo E is obtained. s (r) The real and imaginary values ​​in the x, y, and z directions.

[0019] A further technical solution of the present invention is as follows: obtaining the gaze rotation change rate of the radar tracking a hypersonic target in different flight modes at any flight time is specifically as follows:

[0020] The angle-time conversion algorithm T(θ) is reversely deduced with the pitch angle θ of the radar line of sight to the flight direction as the independent variable and the cumulative flight time t as the dependent variable to obtain the functional relationship θ(t) of the pitch angle θ changing with the flight time t. Then, θ(t) is differentiated with respect to the time t to obtain the gaze rotation change rate dθ / dt of the radar tracking hypersonic targets in different flight modes at any flight time.

[0021] A further technical solution of the present invention is as follows: the total area value of the illuminated face of the target at a fixed percentage at different time points is obtained, specifically:

[0022] By inputting NASA's measured electron density data of different flight modes, the MLMM of hypersonic targets covered with plasma sheaths of different thicknesses is constructed. When the radar illuminates the target surface of the aircraft, the bubble sort and cumulative sum method are used to extract the total area value of the target illuminated surface at fixed percentages at different time points.

[0023] A further technical solution of the present invention is: the method of using bubble sort and cumulative summation to extract the total area value of the target illuminated surface at different time points with a fixed percentage is as follows:

[0024] The backscattering intensity of all illuminated triangular facets of the hypersonic target at random time points between 100s and 150s under a certain incident angle is calculated. Then, the facets are statistically numbered, bubble sorted according to their backscattering intensity, and the top 7% to 10.5% of the facets with stronger backscattering power increasing over time are selected and recorded, and the areas of the small triangular facets corresponding to the selected facets are accumulated and summed.

[0025] Keeping the same granularity, the total effective backscattering area eigenvalue S of targets in different flight modes is traversed and solved in the time interval of 100s-150s.

[0026] A further technical solution of the present invention is as follows: the recognition classifier adopts a single hidden layer MLP, and the MLP is in the form of a fully connected neural network, which includes, from left to right: an input layer, a hidden function layer, and an output layer.

[0027] A further technical solution of the present invention is that the feature vector library is trained by inputting data at a random ratio of 80% during a fixed time period.

[0028] A computer system, characterized in that it comprises: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0029] A computer-readable storage medium, characterized in that it stores computer-executable instructions, which are used to implement the above method when executed.

[0030] A computer program product, characterized by comprising computer executable instructions, wherein the instructions are used to implement the above method when executed.

[0031] The beneficial effects of the present invention are:

[0032] The present invention aims at the pattern recognition problem of hypersonic targets covered with plasma sheaths, takes the reduction of mutual aliasing between the eigenvalues ​​extracted from different flight mode categories as the design inspiration, and based on the maneuvering mode trajectory route of the hypersonic target side-station level flight, the present invention provides a method for capturing and constructing eigenvalues ​​related to high-frequency wave scattering, kinematics and geometric dimensions in a short fixed period of time in the early stage, and fits this complex function mapping relationship through machine learning, that is, a simple and reliable classification model or network can associate the input eigenvector with the undetermined output vector (category) with high quality, so as to realize the rapid and accurate recognition of RAM C-II targets with different flight modes. Compared with the traditional DRCS black box recognition method, the recognition accuracy of the present invention is improved by nearly 50%. It has the following advantages:

[0033] First, based on wave scattering theory, a deep mapping relationship between the plasma sheath echo attenuation, propagation mechanism and scattered electric field characteristic values ​​of targets in different flight modes is established.

[0034] Second, under the trajectory model of the RAM C-II aircraft target side-station level flight, two physical characteristic parameters of the kinematics and flow field geometry of a typical hypersonic target are described, including the scattering effective area and the gaze rotation change rate.

[0035] Third, a single hidden layer MLP with a simple structure is used as the classifier. Since the network input has only eight feature parameters and comes from wave physics insights, when using the same amount of training data, the classifier is more accurate than a neural network working in a black-box manner.

[0036] Fourthly, the feature extraction and classification method involved in the present invention has a clear and concise architecture. Compared with the traditional mathematical statistical feature extraction of the full time chain DRCS time series and the convolutional neural network classification of the ISAR image database based on large-scale data, it has the advantages of easy understanding, low computational cost, high real-time performance and good feasibility in terms of physical connotation and solution cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0038] Figure 1 A flowchart for implementing the present invention;

[0039] Figure 2 Schematic diagram of the geometric dimensions and structure of the RAM C-II target aircraft;

[0040] Figure 3 RAM C-II spacecraft re-entry electron density measured data;

[0041] Figure 4 Schematic diagram of TE / TM scattering in free space on the surface of multilayer lossy dielectrics;

[0042] Figure 5 Schematic diagram of the flight trajectory of a hypersonic vehicle and DRCS observations;

[0043] Figure 6 0s-350s, S-band, DRCS time series of RAM C-II aircraft in different flight modes;

[0044] Figure 7 100s-150s, S-band, DRCS time series of RAM C-II aircraft in different flight modes;

[0045] Figure 8 100s-150s, numerical distribution diagram of piecewise function relationship θ(t) under different flight modes;

[0046] Fig. 9 100s-150s, distribution diagram of characteristic values ​​of gaze rotation angle change rate dθ / dt under different flight modes;

[0047] Fig.10 Schematic diagram of longitudinal section of hypersonic target covered with plasma sheath in different flight modes modeled in FEKO;

[0048] Fig.11 100s-150s, distribution diagram of the eigenvalue of the total area S of effective backscatter contribution in different flight modes;

[0049] Fig.12 Schematic diagram of the fully connected neural network structure of MLP;

[0050] Fig.13 100s-150s, distribution of eight physical insight-level eigenvalues ​​of hypersonic targets in different flight modes;

[0051] Fig.14 100s-150s, DRCS characteristic value distribution of hypersonic targets in different flight modes;

[0052] Fig.15Confusion matrix recognition results of hypersonic targets based on wave physics eigenvalues ​​in eight flight modes;

[0053] Fig.16 Confusion matrix recognition results of hypersonic targets based on DRCS time series eigenvalues ​​under eight flight modes. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0055] The present invention provides a hypersonic target pattern recognition method based on wave scattering and physical feature extraction, which extracts eight physical features by combining the wave scattering principle and the kinematic and geometric features of the hypersonic target, and uses a multilayer perceptron (MLP) with a single hidden layer to achieve fast and high-precision classification. The eight feature values ​​derived from wave physics proposed by the present invention are selected as the input of the neural network, and the adopted classifier does not work in black box mode.

[0056] Embodiment 1:

[0057] like Figure 1 As shown, the specific steps include:

[0058] Step 1: Using the experimental data of plasma electron density measured by the reentry vehicle published by NASA as the model electromagnetic parameter input, a multilayer lossy media model (MLMM) for the hypersonic target covered with a plasma sheath expressing static scattering characteristics is established by solving the parallel physical optics (PO) scattering field integral formula based on triangular facet subdivision;

[0059] Step 2: Set the hypersonic target flight mode to the side-stand level maneuvering mode. The satellite-borne radar searches first and then tracks and adopts staring tracking for the unknown target aircraft. Based on the motion trajectory of the aircraft in horizontal flight, the angle-time conversion algorithm is used to add the time dimension to the RCS data after the angle-transformed quasi-static solution obtained by MLMM, and the DRCS time series of the target aircraft at different flight altitudes and flight speeds is calculated.

[0060] Step 3: The DRCS data of the target in different flight modes received by the radar sensor is separated into scattered echoes. Through the MLMM established in step 1, the time-varying scattered electric field E is obtained by forward simulation. s (r) Discrete data: Based on the PO scattered field integral formula, the vertical / horizontal polarization processing of the scattered echo is removed and the scattered echo E is obtained. s (r) The real and imaginary values ​​in the x, y, and z directions effectively construct six pattern recognition eigenvalues ​​of the hypersonic target containing plasma wave attenuation and wave propagation properties;

[0061] Step 4: In the hypersonic target dynamic flight trajectory model established in step 2, the angle-time conversion algorithm T(θ) is reversely deduced with the pitch angle θ of the radar line of sight to the flight direction as the independent variable and the flight cumulative time t as the dependent variable to obtain the functional relationship θ(t) of the pitch angle θ changing with the flight time t, and θ(t) is derived with respect to the time t to obtain the gaze rotation change rate dθ / dt of the radar tracking hypersonic targets with different flight modes at any flight time, which is recorded as the seventh effective eigenvalue of hypersonic target pattern recognition;

[0062] Step 5: By inputting NASA's measured electron density data of different flight modes, the MLMM of hypersonic targets covered with plasma sheaths of different thicknesses can be constructed. When the radar illuminates the target surface of the aircraft, the bubble sort and cumulative sum method are used to extract the total area value S of the target illuminated surface at different time points with a fixed percentage, which is recorded as the eighth effective eigenvalue of hypersonic target pattern recognition;

[0063] Step 6: Use a simple single hidden layer MLP as the recognition classifier, combine the above eight physical eigenvalues ​​into a feature vector library, and input training data at a random ratio of 80% in a fixed time period to achieve high-precision classification of different flight modes of hypersonic targets.

[0064] Embodiment 2:

[0065] Step 1: Aiming at the complex scattering characteristics of hypersonic targets covered with plasma sheaths, the electron density data measured by NASA is used as the input of plasma flow field modeling, and the high-frequency PO algorithm is used to output the scattering numerical results to establish the MLMM for solving the static backscattering problem;

[0066] Step 2: Based on the angle-time conversion algorithm, the backscatter echo data of the hypersonic target is deduced to the dynamic time domain, and the relative motion parameters and change laws between the satellite-borne radar and the horizontal trajectory flight target are added to construct the DRCS time series characteristics of the aircraft;

[0067] Step 3: Using the forward solution of the high-frequency PO fully analytical scattering integral formula, the target echo on the uniform motion path is depolarized and the real and imaginary parts are disassembled to accurately capture the deep mapping relationship between the flight mode state change and the scattering integral modulus attenuation, and effectively extract the pattern recognition eigenvalues ​​driven by the connotation of six wave scattering;

[0068] In step 3, the depolarization and real / imaginary part decomposition processing of the hypersonic target echo is realized by forward solving the high-frequency PO full-analysis scattering integral formula based on MLMM, deeply linking the intrinsic physical mapping relationship between the flight mode and the scattered echo, and performing absolute value and logarithm processing on the six types of eigenvalues ​​of the real / imaginary parts of the three-direction components of x, y, and z, so that the numerical distribution interval is aligned with the DRCS time series. The specific calculation formula and steps are as follows:

[0069]

[0070] Among them, C s C is the vector related to the propagation and loss characteristics in the integral sign of the electromagnetic scattering formula of hypersonic targets in different flight modes. s Expanding in the x, y, and z directions, it can be further written as:

[0071]

[0072] In the formula, C x ,C y ,C z is the core parameter of the eigenvalue related to the scattering loss characteristics of the plasma sheath after depolarization, and is therefore related to the scattered electric field E s (r) in the x direction The following formula is derived:

[0073]

[0074] Then there is a scattered echo in the x direction The real part of and the imaginary part Eigenvalue analytical expression:

[0075]

[0076] Similarly, the characteristic values ​​of the scattered echo in the y and z directions can be obtained and The analytical expression of is:

[0077]

[0078] To maintain the same distribution range of the target DRCS value as that calculated in step 2, we set:

[0079]

[0080] In the formula, and Recorded as six types of hypersonic target pattern recognition feature values ​​containing plasma wave attenuation and wave propagation properties.

[0081] Step 4: According to the maneuvering mode of the hypersonic target side-to-side level flight, the variable factors of the angle-time conversion algorithm T(θ) are reversed to obtain the functional relationship θ(t), and the first-order derivative is solved in the time dimension to design the gaze rotation change rate dθ / dt, which is recorded as the classification feature value under the motion physics parameter response mode;

[0082] The step 4 uses the motion trajectory model of the hypersonic target flying in the opposite direction of the satellite-borne radar under different flight conditions to reversely deduce the angle-time conversion algorithm T(θ) described in step 2 to obtain the piecewise function relationship θ(t). In order to further effectively distinguish the physical motion characteristics of the target under different flight modes, a gaze rotation angle change rate dθ / dt characteristic value with a larger span interval than the θ(t) numerical distribution is designed. The specific calculation formula and design steps are as follows. First, the angle-time conversion formula T(θ) of step 2 is reversely derived to obtain the following piecewise function relationship θ(t):

[0083]

[0084] Where, H is the flight altitude of the satellite-borne radar, H x and v x are the flight altitude variable of the hypersonic target in different flight modes and the speed variable of the flight relative to the satellite, respectively. θ0 is the angle between the irradiation direction of the electromagnetic wave of the satellite-borne radar and the axis of the target aircraft when the initial timing is 0s, and the value of the present invention is 2.5°. Further, the first-order derivative of θ(t) with respect to t can be obtained to obtain the expression of the characteristic value of the gaze rotation angle change rate dθ / dt:

[0085]

[0086] Wherein, the characteristic value of the radar gaze rotation angle change rate dθ / dt is recorded as the seventh type of hypersonic target pattern recognition characteristic value containing motion flight parameter attributes.

[0087] Step 5: Adopt the occlusion judgment in the PO algorithm, combine the trajectory strategy of the target aircraft and the satellite-borne radar moving towards each other from far to near, and reasonably calculate the total area value S of the triangular surface element of the illuminated area of ​​the outer surface of the hypersonic target under different flight modes, which is recorded as the classification feature value under the response of the geometric feature parameter. In particular, the percentage of the effective contribution total scattering intensity surface element value associated with S changes dynamically with the flight time;

[0088] In step 5, the hypersonic target under different flight modes is modeled, and the number of small triangular facets of its outer surface after segmentation and fitting is different, thereby constructing the total area value S of effective backscattering intensity contribution as a class of strong robust classification features. The characteristic value extraction method is to first calculate the backscattering intensity of all illuminated triangular facets of the hypersonic target within a random time point (determine the incident angle) in the interval of 100s-150s, then statistically number the facets, perform bubble sorting according to the size of their backscattering intensity, select the first 7%-10.5% of the facets with stronger backscattering power increasing over time, record the number, and accumulate and sum the small triangular facets corresponding to the selected facets. Finally, keep the same granularity, and traverse and solve the characteristic value S of the total effective backscattering area of ​​targets in different flight modes within the time interval of 100s-150s.

[0089] Step 6: A single hidden layer MLP with 1000 neurons is used as the classifier. The feature vector data set consisting of eight types of wave scattering and physical eigenvalues ​​is used as input, and eight types of hypersonic targets with different flight modes are used as output. Compared with the black box system, high-precision pattern recognition of hypersonic targets is achieved.

[0090] Embodiment 3:

[0091] Step 1: Establish an MLMM to express the static scattering characteristics of a plasma sheathed hypersonic target, where the geometric model of the RAMC-II target is a blunt cone structure with a head radius of R = 0.1524m, a total length of the aircraft of L = 1.295m, a semi-cone angle of α = 9°, and an incident angle of the electromagnetic wave irradiated by the radar of θ i , the specific structural details are expressed as Figure 2 The relative dielectric constant ε of the plasma sheath under different flight modes r The measured electron density data published by NASA is used as input for solution. The detailed data distribution is as follows: Figure 3 The present invention aims at RAM C-II aircraft with different flight modes. While ensuring the calculation accuracy, it prevents the calculation time from being out of control due to too many layers. The number of layers is divided according to the principle that the difference in electron density between two adjacent layers of uniform plasma does not exceed 10% of the maximum value of the plasma electron density. The number of layers of the non-uniform plasma sheath at each flight altitude, the thickness of each layer and the maximum plasma frequency data are published, and the summary is shown in Table 1:

[0092] Table 1 Plasma sheath stratification and electromagnetic parameter data at different reentry altitudes

[0093]

[0094] In order to solve the electromagnetic scattering characteristics of hypersonic targets through MLMM, the incident electromagnetic wave is often decomposed into TM wave and TE wave for analysis separately in the calculation. Here, the local coordinate system is introduced. Since the polarization mode of the final synthetic wave is related to the amplitude and phase relationship between the TM wave and the TE wave, we put the two in one coordinate system for analysis. Figure 4 As shown, it is a tiny triangular facet on the surface of the scatterer to be solved, which is a perfect plane in the limit approximation. Combining all the calculation parameters, the Stratton-Chu equation expression derived based on the divergence theorem and the metal-loss medium-free space boundary conditions are combined, and the Huygens equivalent principle and PO approximation conditions are applied to obtain the analytical expression of the far-field scattering integral of the frequency domain PO for hypersonic targets:

[0095]

[0096] The time harmonic factors derived from the formula of the present invention are all taken as e jωt , refer to Figure 4 , where the integral subscript S1 is the illumination area of ​​the scatterer target, |E0| is the electric field intensity of the incident wave, r is the distance from the scatterer target to the detection radar, θ i is the incident angle, θ s is the scattering angle, is the outward unit normal vector of a single triangular facet on the scatterer surface, and The polarization direction of the incident electric field and the scattered electric field is parallel to the incident plane -XOZ, that is, TM wave. and The incident electric field and the scattered electric field are perpendicular to the polarization direction of the incident plane -XOZ, that is, TE wave. The synthetic wave of the two is a linearly polarized wave. and are the incident wave unit vector and the scattered wave unit vector, respectively. In particular, the total reflection coefficient R of the plasma sheath in the MLMM to the attenuation value of the irradiated electromagnetic wave propagation is related to / / and R ⊥ The solution is obtained by using the continuous fraction (CF) formula or the step method, which is given by the following two recursive formula pairs. When the number of plasma sheath layers n = 2, 3, ···, N+1, we have:

[0097]

[0098] in,

[0099] A1=Z(n-1)R Y (n)-Z(n)R Y (n-1)

[0100] B1=Z(n-1)R Y (n)+Z(n)R Y (n-1)

[0101] A2=Z(n)R Y (n)-Z(n-1)R Y (n-1)

[0102] B2=Z(n)R Y (n)+Z(n-1)R Y (n-1)

[0103] In the above formula, θ i is the initial incident angle of the electromagnetic wave, and θ0 is the refraction angle of layer 0. When there is no plasma sheath on the outer surface of the target aircraft, that is, when there are only metal layers and vacuum layers, the relative dielectric constant of the ideal conductor can be regarded as infinite, then we have R ⊥ =-1, R / / =1, in which case there is no cross polarization. When n = 0, it is an ideal conductor-metal layer on the surface of the hypersonic target aircraft, which can be regarded as a semi-infinite space, with Z(0) = 0, and R ⊥ (0) = R / / (0)=0.

[0104] Step 2: Set the hypersonic aircraft to the sideways level flight maneuvering mode (that is, the aircraft lift is equal to the gravity, and the thrust is equal to the drag). During the sideways level flight, the radar line of sight to the target remains unchanged at an azimuth angle (the default angle is 0° during calculation), and the roll angle is also 0°. Only the pitch angle θ changes. The target aircraft maintains a uniform speed as it moves from far to near the satellite-borne radar, but the flight time varies non-uniformly with the unit angle. The specific design is as follows: Figure 5 As shown. H = 200000m is the height of the satellite-borne radar installed on the non-synchronous satellite from the ground, H1 and H2 are the different flight altitudes of the hypersonic aircraft, and H0 is the shortest vertical distance between the aircraft and the satellite-borne radar. Combining the law of universal gravitation and the circular motion formula, the flight speed v0 of the satellite-borne radar on the non-synchronous orbit satellite relative to the ground is calculated by the following formula:

[0105]

[0106] Wherein, G is the gravitational constant, M is the mass of the earth, and R is the radius of the earth. In particular, in the present invention, the target is set to fly along a horizontal trajectory, and the satellite-borne radar and the aircraft are set to fly towards each other, and the relative speed between the two is v total for:

[0107] v total =v+v0

[0108] In addition, the calculation model considers the cumulative time used by the hypersonic target aircraft at different positions when the pitch angle between the aircraft axis and the radar irradiation direction changes in the range of 2.5°≤θ≤177.5°, that is, when the angle between the satellite-borne radar electromagnetic wave irradiation direction and the target aircraft axis is 2.5°, and the initial time at this time is recorded as 0s. Here, the first half of the aircraft's flight trajectory is taken as the calculation interval, and the initial value θ=89.5°, the unit angle change Δθ=1°, and the number of angle changes in the first half is 0≤n≤87, then the angle change interval in the first half of the hypersonic aircraft flight is:

[0109] 2.5°≤θ-nΔθ≤89.5°

[0110] v is the average speed of the aircraft on the horizontal flight trajectory, then

[0111]

[0112] For the first half of the flight trajectory, the accumulated time for the aircraft to fly from infinity to each fixed angle is:

[0113]

[0114] Similarly, in the second half of the flight trajectory, the time taken by the aircraft to fly to each fixed angle is symmetrical with that in the first half. The aircraft flies from the point with the shortest vertical distance from the satellite-borne radar (θ = 90°) to infinity, and the accumulated time at each fixed angle is:

[0115]

[0116] Among them, m is the number of times the angle changes in the second half. The total time accumulated in the first and second halves is the total time taken by the hypersonic aircraft to fly along the horizontal flight trajectory at a fixed altitude. Therefore, the accumulated time taken by the aircraft to reach each radar illumination angle on the horizontal flight trajectory can be calculated. Finally, the present invention takes the hypersonic aircraft RAM C-II as the research object, adopts the above-mentioned angle-time conversion algorithm, and calculates the DRCS variation characteristics of the target aircraft flying along the horizontal route trajectory at different altitudes and speeds in the S band (f=3.25GHz) and the time interval of 0s-350s. The results are as follows Figure 6 It is worth mentioning that in real reconnaissance and tracking scenarios, the spaceborne radar always needs to predict the target state and detect the target type in advance, and ensure fast data reception and real-time calculation. Therefore, in the subsequent solution steps, the random extraction time range of the relevant eigenvalues ​​is uniformly intercepted to 100s-150s. Here, the DRCS value distribution of the RAM C-II target in this time period is given. The results are shown in Figure 7 shown.

[0117] Step 3: When the radar receiving sensor detects the scattered echo data of the hypersonic target in different flight modes, the difference in the amplitude attenuation and propagation loss information feedback of the echo is essentially due to the electromagnetic wave reflection coefficient R of the plasma sheath covering the outer surface. / / and R ⊥ This reflection characteristic depends on the relative dielectric constant ε of the non-magnetized plasma sheath. r and magnetic permeability μ r , ε r And the sheath electron density n e0 Determined plasma frequency f p and collision frequency υ e We try to magnify this difference and use the PO scattered field integral expression to reversely derive the eigenvalue, but the scattered electric field E of the hypersonic target s (r) is completely analytical and cannot be reasonably decomposed by reverse deduction. In general, the target RCS is calculated as:

[0118]

[0119] Among them, it corresponds to the polarization direction of the incident electric field is the polarization direction of the scattered electric field, r is the distance from the scatterer target to the detection radar, E s (r) is the scattered electric field of the target, and |E0| is the electric field strength of the incident wave. In combination with the above formula, in practical engineering applications, it can be considered to depolarize the received target scattered echo information, and decompose the real and imaginary values ​​according to the amplitude and phase information of the scattered echo components received in the three directions of x, y, and z. Here, the present invention adopts the forward decomposition method to solve the eigenvalues ​​related to the plasma sheath scattering loss characteristics. As shown in the following formula:

[0120]

[0121] Among them, C s C is the vector related to the propagation and loss characteristics in the integral sign of the electromagnetic scattering formula of hypersonic targets in different flight modes. s Expanding in the x, y, and z directions, it can be further written as:

[0122]

[0123] In the formula, C x ,C y ,C z is the core parameter of the eigenvalue related to the scattering loss characteristics of the plasma sheath after depolarization, and is therefore related to the scattered electric field Es (r) in the x direction E x s The following formula is derived:

[0124]

[0125] Then there is a scattered echo in the x direction The real part of and the imaginary part Eigenvalue analytical expression:

[0126]

[0127] Similarly, the characteristic values ​​of the scattered echo in the y and z directions can be obtained and The analytical expression of is:

[0128]

[0129] To maintain the same distribution range of the target DRCS value as that calculated in step 2, we set:

[0130]

[0131] So far, the present invention has successfully extracted six effective pattern recognition eigenvalues ​​containing plasma wave attenuation and wave propagation properties of hypersonic targets. and

[0132] Step 4: The flight altitude and flight speed of hypersonic targets are completely different in different flight modes. Therefore, in a fixed-altitude satellite-borne radar detection system, the gaze rotation angle change rate dθ / dt of the radar tracking target will also change with the flight mode. By reversely deducing the angle-time conversion algorithm T(θ) in step 2, the following piecewise function relationship θ(t) can be obtained:

[0133]

[0134] Where, H is the flight altitude of the satellite-borne radar, H x and v x are the flight altitude variable of the hypersonic target in different flight modes and the speed variable of the flight relative to the satellite, respectively. θ0 is the angle between the irradiation direction of the satellite-borne radar electromagnetic wave and the axis of the target aircraft when the initial timing is 0s, and the value of the present invention is 2.5°. Further, the expression of the characteristic value of the gaze rotation angle change rate dθ / dt can be obtained by taking the derivative of θ(t) with respect to t:

[0135]

[0136] Therefore, the piecewise function relationship θ(t) and the characteristic value distribution of the gaze rotation angle change rate dθ / dt in the time period of 100s-150s can be obtained as follows: Figure 8 and Fig. 9 As shown, it can be seen that the characteristic value dθ / dt of the target's gaze rotation angle change rate relative to the detection radar in different modes has a larger distribution span than the numerical distribution span of θ(t), which will further effectively distinguish the physical motion characteristics of the target in different modes and provide a good data basis for subsequent classifier training.

[0137] Step 5: Use the electron density data measured by NASA to construct the electromagnetic performance parameters of the plasma sheath. However, due to the limited amount of published data, the thickness of the dielectric layer containing the effective plasma electron density value on the outer surface of the hypersonic target under different flight modes is different when the commercial software FEKO models it. For example, Fig.10 As shown. When the high-frequency PO scattering algorithm is used to solve the electromagnetic scattering characteristics of the target, the model outer surface fitting is often divided into small triangular facets. Therefore, in a fixed flight path, the number of facets illuminated by the target at the same time under different flight states is not the same, making the total area value of the effective backscattering intensity contribution a strong robust classification feature. Let ΔS be an arbitrarily divided small triangular facet, and use FEKO for modeling output to obtain the three vertex parameters of the facets on the model, which are expressed as r1=(x1, y1, z1), r2=(x2, y2, z2), and r3=(x3, y3, z3). Let

[0138] a=r1-r2=(x1-x2,y1-y2,z1-z2)

[0139] b=r1-r3=(x1-x3,y1-y3,z1-z3)

[0140] Then we have the calculation formula for the area of ​​a small triangular face:

[0141]

[0142] in,

[0143]

[0144] The time period is uniformly intercepted from 100s to 150s, and the backscattering intensity of all illuminated patches of the hypersonic target model at each time point is solved. All patches are numbered and bubble sorted according to their backscattering intensity. The patch record number with stronger backscattering intensity is selected and included in the cumulative area. In the calculation time period of 100s-150s, the aircraft and the detection radar are moving towards each other, and the number of illuminated patches on its outer surface gradually increases with time. Therefore, here we stipulate that the scattering intensity patch count values ​​are counted according to the first 7%-10.5% of the total number that increases proportionally with time. Similarly, the above operation is repeated at other time points of 100s-150s. When the calculation of all time points is completed, the areas of the small triangular face elements corresponding to the counted patches are accumulated to form the total area eigenvalue S of the effective backscattering contribution. The calculation results are shown as follows: Fig.11 shown.

[0145] Step 6: The eight mathematical statistical eigenvalues ​​given in steps 3-5 can well describe and represent the geometric kinematics of hypersonic targets and the wave scattering physical characteristics of the plasma sheath, without the need to use complex neural networks to complete this classification task. Since the number of network inputs (eight eigenvalues) and outputs (eight types of targets) is small, a single hidden layer MLP can be used to achieve classification. The single hidden layer fully connected neural network structure used in the proposed method is as follows: Fig.12 As shown. The MLP neural network is a fully connected neural network, from left to right: input layer, hidden function layer and output layer. The red arrow represents the back propagation of the error value obtained after each training. In the backward propagation of the error, we can use optimization methods such as stochastic gradient descent (SGD) and adaptive moment estimation (Adam) to minimize the loss function, so as to automatically adjust and update the weights w1, w2 and biases b1, b2 at each neuron position. Here, the activation output of the jth neuron in the hidden layer is defined as:

[0146]

[0147] In the above formula, i∈(0,7), f i is the eigenvector composed of eight eigenvalues ​​of the input neuron, w 1ij is the value of weight w1 when the i-th input layer is connected to the j-th hidden layer, b 1j is the value of the bias b1 at the input of the jth hidden layer, f ReLU(·) is the nonlinear activation function connected to the output end of the hidden layer, and ReLU is selected as the activation function here. When training MLP, Pytorch is used as a deep learning framework, and the hidden layer consists of 1000 neurons. It is worth noting that the number of hidden layers and the number of neurons in the hidden layer are determined after comparing various combinations and combining the total training time. All algorithms and network architectures designed in the present invention are calculated, trained and tested on a PC equipped with an Intel (R) Core (TM) i7-10875H CPU (2.3GHz, 8 cores and 64GB RAM) and a single NVIDIA Geforce RTX 2070Super GPU. The specific parameter design of network training is shown in Table 2:

[0148] Table 2 Selection and design of MLP training parameters based on PyTorch framework

[0149] MLP basic structure and parameter settings Categories and Values Activation function for the connection after the hidden layer ReLU Loss Function Cross Entropy Loss Function MLP Optimization Method Adam Learning rate 0.0001 <![CDATA[β1 value of the Adam method]]> 0.9 <![CDATA[β2 value of Adam method]]> 0.999 Number of training rounds (epoch) 1500 The number of feature vectors loaded per batch 16

[0150] The input of MLP includes eight wave physics feature vector datasets of eight types of flight mode targets after data normalization, namely dθ / dt and S, each type of target contains 1001 sets of vector data in the S band and 100s-150s time period, of which the training data accounts for 80% of the total input data (800 sets of data for each type of target), and the test data accounts for about 20% of the total input data (201 sets of data for each type of target). The output is eight different flight modes of RAM C-II targets, including 21km13.29Ma, 25km17.16Ma, 30km21.66Ma, 47km23.01Ma, 53km23.30Ma, 61km24.55Ma, 71km25.91Ma and 76km26.56Ma. We plotted the histograms of 1001 sets of data in the S-band 100s-150s time period for eight wave physics characteristics of the RAM C-II target in eight flight modes and the histograms of 1001 sets of data in the S-band 100s-150s time period for DRCS time series, as shown in Figure 2. Fig.13 and Fig.14 As shown, this provides data visualization and regular guidance for feature selection and pattern classification of hypersonic targets.

[0151] The confusion matrix results of identifying eight types of targets based on wave physics eigenvalues ​​are as follows Fig.15 The confusion matrix results of DRCS single time series recognition of eight types of targets in black box mode are as follows Fig.16 The comparison of the two classification recognition results and the total training time required for hypersonic targets under different flight conditions is shown in Table 3:

[0152] Table 3 Comparison of eight-category comprehensive recognition results based on two feature extraction methods of MLP

[0153] Classification categories of MLP Black box recognition accuracy Wave physics identification accuracy 21km13.29Ma 93.53％ 100.00％ 25km17.16Ma 76.12％ 99.00％ 30km21.66Ma 39.80％ 100.00％ 47km23.01Ma 19.90％ 93.53％ 53km23.30Ma 44.78％ 98.00％ 61km24.55Ma 26.87％ 99.50％ 71km25.91Ma 0.00％ 78.10％ 76km26.56Ma 93.53％ 100.00％ Comprehensive recognition accuracy of neural network on test images 49.32％ 96.02％ Actual training time 480.08s 523.27s

[0154] By comparing the classification and recognition accuracy of the black box system of 49.32%, the present invention is based on wave scattering, motion and geometric physics feature extraction methods. The simple single hidden layer neural network only needs to be trained for about 500s to improve the comprehensive recognition effect of the final test set to 96%. In particular, it is very difficult for the black box to effectively classify the target flight mode types concentrated between 30km21.66Ma, 47km23.01Ma, 53km23.30Ma, 61km24.55Ma and 71km25.91Ma, which is caused by the similarity and aliasing of the DRCS characteristic amplitude distribution and time fluctuation of high-speed targets under five flight states, such as Figure 7 and Fig.14 As shown in the figure, even with more rounds of training, it is difficult to obtain ideal classification test results. Obviously, by establishing an effective mapping relationship between physical space and scattered echoes inspired by wave physics, eight physical parameter feature analytical expressions such as target backscatter power, kinematics and geometric features can be effectively captured and extracted without special requirements or assumptions. It can be applied to MLP with multi-parameter distribution, back-propagation error update and complex dynamic connection mode to better cope with this classification task. The final MLP classification and recognition results can increase the recognition rate of the target category at 30km21.66Ma from 39.80% to 100.00%, the recognition rate at 47km23.01Ma from 19.90% to 93.53%, the recognition rate at 53km23.30Ma from 44.78% to 98.00%, the recognition rate at 61km24.55Ma from 26.87% to 99.50%, and the recognition rate at 71km25.91Ma from 0.00% to 78.10%.

[0155] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should be included in the protection scope of the present invention.

Claims

1. A wave physics insight-driven hypersonic target pattern recognition method, characterized in that: include: Obtain six pattern recognition eigenvalues ​​containing plasma wave attenuation and wave propagation properties, wherein the pattern recognition eigenvalues ​​include: the real and imaginary values ​​of the scattered echo in the three component directions of x, y, and z; specifically: Establish a multi-layer lossy medium model MLMM to express the static scattering characteristics of a hypersonic target covered with a plasma sheath; The angle-time conversion algorithm is used to add the time dimension to the RCS data after the angle transformation quasi-static obtained by MLMM, and the DRCS time series of the target aircraft at different flight altitudes and flight speeds are calculated; The DRCS time series of the target in different flight modes received by the radar sensor is separated into scattered echoes, and the scattered electric field E that varies with time is obtained by forward simulation through MLMM. s (r) Discrete data: Based on the PO scattered field integral formula, the vertical / horizontal polarization processing of the scattered echo is removed and the scattered echo E is obtained. s (r) The real and imaginary values ​​in the x, y, and z directions; Obtain the gaze rotation change rate of the radar tracking hypersonic targets in different flight modes at any flight time as the seventh effective eigenvalue of hypersonic target pattern recognition; specifically: The angle-time conversion algorithm T(θ) is reversely deduced with the pitch angle θ of the radar line of sight to the flight direction as the independent variable and the flight cumulative time t as the dependent variable, and the functional relationship θ(t) of the pitch angle θ changing with the flight time t is obtained. Then θ(t) is differentiated with respect to the time t to obtain the gaze rotation change rate dθ / dt of the radar tracking hypersonic targets in different flight modes at any flight time. Obtain the total area value of the target illuminated surface at different time points with a fixed percentage as the eighth effective eigenvalue of hypersonic target pattern recognition; The above eight valid eigenvalues ​​are combined into a feature vector library and input into the recognition classifier for training; The trained recognition classifier is used to perform high-precision classification of different flight modes of hypersonic targets.

2. The method for hypersonic target pattern recognition driven by wave physics insights according to claim 1, characterized in that: The method of obtaining the total area value of the illuminated surface of the target at different fixed percentages at different time points is specifically: By inputting NASA's measured electron density data of different flight modes, the MLMM of hypersonic targets covered with plasma sheaths of different thicknesses is constructed. When the radar illuminates the target surface of the aircraft, the bubble sort and cumulative sum method are used to extract the total area value of the target illuminated surface at fixed percentages at different time points.

3. The method of hypersonic target pattern recognition driven by wave physics insights according to claim 2, characterized in that: The bubble sort and cumulative sum method is used to extract the total area value of the target illuminated surface at different time points with a fixed percentage, specifically: The backscattering intensity of all illuminated triangular facets of the hypersonic target at random time points between 100s and 150s under a certain incident angle is calculated. Then, the facets are statistically numbered, bubble sorted according to their backscattering intensity, and the top 7% to 10.5% of the facets with stronger backscattering power increasing over time are selected and recorded, and the areas of the small triangular facets corresponding to the selected facets are accumulated and summed. Keeping the same granularity, the total effective backscattering area eigenvalue S of targets in different flight modes is traversed and solved in the time interval of 100s-150s.

4. The method of hypersonic target pattern recognition driven by wave physics insights according to claim 1, characterized in that: The recognition classifier adopts a single hidden layer MLP, and the MLP is a fully connected neural network form, which includes from left to right: input layer, hidden function layer and output layer.

5. The method of hypersonic target pattern recognition driven by wave physics insights according to claim 1, characterized in that: The feature vector library is trained with 80% random proportion data input in a fixed time period.

6. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

7. A computer-readable storage medium, characterized in that Computer executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

8. A computer program product, characterized in that The invention comprises computer executable instructions, which are used to implement the method of claim 1 when being executed.

Images