A hierarchical analysis method of micro-Doppler features for small UAV identification

Abstract

The invention discloses a micro-Doppler feature hierarchical analysis method for small-sized unmanned aerial vehicle identification, comprising the following steps: pre-processing the small-sized unmanned aerial vehicle echo signal to remove the Doppler signal component of the fuselage of the unmanned aerial vehicle; separating and extracting the micro-Doppler signal component of the unmanned aerial vehicle, and constructing a micro-motion feature set of the unmanned aerial vehicle, which includes a training data set and a measured data set; constructing an unmanned aerial vehicle classification model, using the training data set to construct a model training data set to train the unmanned aerial vehicle classification model, and establishing an optimal combination of micro-motion features of the unmanned aerial vehicle and preserving the optimal classification model; guiding a radar system to select a data set from the measured data set according to the combination of micro-motion features of the unmanned aerial vehicle, and inputting the data set into the optimal classification model to identify the type of the unmanned aerial vehicle; using the unmanned aerial vehicle echo signal pre-processing to suppress the Doppler signal of the fuselage, reduce the influence of the shielding effect, extract effective micro-motion features according to prior knowledge and train the optimal lightweight unmanned aerial vehicle classification model, so as to improve the recognition accuracy of the target.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle classification, and in particular relates to a micro-Doppler feature classification analysis method for small unmanned aerial vehicle identification. Technical Background

[0002] Small drones have the advantages of small size, good safety performance, outstanding maneuverability, low price, and easy control, and have been widely used in various fields. The monitoring and control of small drones has become a new issue for governments and public safety. Usually, the flight altitude of small drones is below 100 meters and the flight speed is less than 20m / s. They are typical "low, slow and small" aerial targets. This feature makes it difficult to distinguish them from natural targets (such as birds) and some ground vehicles and buildings, which makes it difficult to detect and identify small drones. Therefore, it is urgent to improve the detection and identification technology of small drones.

[0003] At present, the detection technology for small drones is mainly divided into four categories, namely radio monitoring, audio detection, photoelectric detection and radar detection. However, radio monitoring technology requires multiple stations to be deployed to obtain high-precision positioning information of the target, and may fail for "silent" drones that turn off remote control and image transmission signals; due to the weak sound of the rotor of light and small drones, when the background noise interference is serious, the audio detection effect is not good; photoelectric detection is easily affected by weather and environment, and the detection distance is very limited. Therefore, photoelectric detection is often used as an auxiliary detection device. Unlike other drone detection technologies, radar detection technology has many advantages. On the one hand, radar detection technology is not affected by environmental changes and can still work normally in harsh environments such as dark light and rainy and foggy weather; on the other hand, through the multi-transmit and multi-receive technology, a single radar system can accurately locate the drone.

[0004] In recent years, methods using radar signal processing technology to distinguish between birds and small drones have been proposed. These methods use time-frequency diagrams or spectrum diagrams to extract features for classification. The paper "Experimental Study On Radar Micro-doppler Signatures Of Unmanned Aerial Vehicles" proposed using short-time Fourier transform to process the echo signals of small aerial targets to obtain the corresponding time-frequency diagrams, which confirmed that different aerial targets exhibit different micro-Doppler features in the time-frequency domain. However, manual intervention is required to complete the recognition task.

[0005] A Chinese invention patent discloses a method for classifying small drones using a generative adversarial network (GAN). However, the high number of parameters and computational complexity of the network model make this classification method difficult to deploy on a cognitive radar system. In addition, the classification method based on deep learning requires a large number of training data sets, which results in a high training cost for this method. The echo signal of a small drone is a time-varying non-stationary signal, which is composed of multiple signals superimposed. The fuselage Doppler signal will mask the drone rotor micro-Doppler signal, and the rotor micro-Doppler signal is weak and easily submerged in noise. Therefore, the above problems will cause difficulties in effective feature extraction and classification and recognition. Summary of the invention

[0006] The purpose of the present invention is to overcome the problems of insufficient robustness of existing radar systems for small UAV identification in low signal-to-noise ratio environments, large number of model parameters and need for manual intervention, and to provide a micro-Doppler feature hierarchical analysis method for small UAV identification.

[0007] A micro-Doppler feature hierarchical analysis method for small UAV identification, characterized in that it includes the following steps:

[0008] S1: Preprocess the echo signal of the small UAV to remove the Doppler signal component of the UAV body;

[0009] S2: Separate and extract the UAV micro-Doppler signal components and construct the UAV micro-motion feature set, which includes the training data set and the measured data set;

[0010] S3: Build a UAV classification model, use the training data set to build a model, train the UAV classification model, and establish the optimal combination of UAV micro-motion features and save the optimal classification model;

[0011] S4: Based on the combination of UAV micro-motion features, the radar system is guided to select a data set from the measured data set and input into the optimal classification model for UAV type identification.

[0012] In the above scheme, the echo signal preprocessing of small UAV is used to suppress the Doppler signal of the fuselage, reduce the influence of the shielding effect on the micro-motion feature extraction, and increase the stability of the radar system in identifying small UAVs at a long distance; the rotor micro-Doppler signal is separated and extracted from the noise through the BFC scoring mechanism to improve the signal-to-noise ratio of the rotor micro-Doppler signal; based on prior knowledge, effective micro-motion features are extracted and the optimal lightweight UAV classification model is trained, which not only improves the recognition accuracy of the target, but also makes the model easier to deploy in the air surveillance radar system.

[0013] In step S1, the preprocessing includes the following steps:

[0014] S11: Construct the embedding matrix H of the small UAV echo signal H = [S1, S2, …, S K ;

[0015] S12: Decompose the left singular value matrix HH T and the right singular value matrix H T H in the matrix;

[0016] S13: Group the above two eigenvalues and construct the subspace H r of the remaining signal s r ;

[0017] S14: Restore the subspace H r to a one-dimensional signal sequence s r (n).

[0018] Preferably, in step S11, decompose the small UAV echo signal s(n) to construct the forward sliding window sequence S k = [s(k), s(k + 1), …, s(k + L - 1)] T , where k = 1, 2, …, K, K = N - L + 1; L is the forward sliding window size;

[0019] Construct the embedding matrix:

[0020]

[0021] In step S12,

[0022] S121: Let the square matrix A = HH T , flag = 0;

[0023] S122: Set C, Z and initialize the vector Υ 0,1 , let C = 1, Z = 1;

[0024] S123: Calculate the matrix eigenvalues and eigenvectors:

[0025] First execute: Υ c,z = AΥ c-1,z , then execute: λ c,z = max(Υ c,z ), and finally execute: and make

[0026] S124: If c < C, let c = c + 1 and execute step S123; otherwise, execute step S125;

[0027] S125: Update the square matrix: A = A - λ C,z Υ C,z Υ C,zT

[0028] S126. If z < Z, then let z = z + 1 and execute step S123; otherwise, execute step S127.

[0029] S127. If flag ≠ 1, then save the eigenvalues and eigenvectors of matrix HH: σ = {σ1, σ2, …, σ T}, p = {p1, p2, …, p L}, let square matrix A = H L H, flag = 1, and execute step S122; otherwise, execute step S128. T T

[0030] S128. Save the eigenvalues and eigenvectors of matrix H T H: δ = {δ1, δ2, …, δ K}, q = {q1, q2, …, q K};

[0031] In step S13, group σ l (l = 1, 2, …, l): If then it is considered that this eigenvalue belongs to signal s b (n); otherwise, it is considered that this eigenvalue belongs to signal s r (n);

[0032] Construct the subspace of signal s r (n): where β is the eigenvalue index belonging to signal s r (n); σ β is the eigenvalue belonging to signal s r (n); p β and q β are the corresponding eigenvectors in P and Q respectively;

[0033] In step S14, the one - dimensional signal sequence s r (n) is specifically:

[0034]

[0035] Preferably, in step S2, the separation and extraction of the UAV micro - Doppler signal includes the following steps:

[0036] S21: Perform modal decomposition to obtain J - order independent components f j (n).

[0037] S22: Perform BFC scoring on each order of independent components;

[0038] S23: Group the independent components; group f j(n) are divided into the following two groups:

[0039] {f j_up (n)|g j_up ≥BF scorethreshold ,j_up∈{1,2,…,J}}

[0040] {f j_dowm (n)|g j_down <BF scorethreshold ,j_downe{1,2,…,J}}

[0041] S24: Reconstruction of micro-Doppler signal m_Ds (n) = ∑f j_up (n).

[0042] S25: Constructing micro-motion feature dataset

[0043] Preferably, in step S21, specifically:

[0044] S211: Find the signal s r (n) All maximum and minimum points are fitted to form upper and lower envelopes;

[0045] S212: The average value of the upper and lower envelopes is recorded as m(n), and then m(n) is subtracted to obtain a new signal and calculate: h(n) = s r (n)-m(n),n=1,2,…,N;

[0046] S213: If h(n) satisfies the following conditions: 1. The absolute value of the difference between the number of extreme points and zero-crossing points of h(n) is not greater than 1; 2. The average values ​​of the upper and lower envelopes are zero, then let f j (n) = h(n) and execute step S214; otherwise, execute step S211;

[0047] S214: f j (n) from s r (n) is separated and we get: r(n) = s r (n)-f j (n),n=1,2,…,N

[0048] S215: If r(n) is not a monotonic function, let s r (n) = r(n) and execute step S211, repeat the cycle J times to obtain the J-order independent component; otherwise, end the modal decomposition;

[0049] In step S21, the following steps are included:

[0050] S221: Let j = 1, input independent component f j(n);

[0051] S222: Decompose f j (n) into a vector set W = {w j,1 , w j,2 , …, w j,I};

[0052] S223: Calculate the BFC scores of each element in W

[0053] S224: Calculate the total BFC score g of f j (n); where j ;

[0054] S225: If j < J, then set j = j + 1 and execute step S221; otherwise, execute S226;

[0055] S226: Calculate the score threshold In step S25, construct the micro - motion feature data set including the following steps: S251: Decompose s m_Ds (n) into a vector set

[0056] S252: Extract the micro - motion feature vectors:

[0057]

[0058] S253: Extract the micro - motion feature vectors:

[0059]

[0060] S254: Extract the micro - motion feature vectors:

[0061]

[0062] S255: Extract the micro - motion feature vectors:

[0063]

[0064] S256: Extract the micro - motion feature values:

[0065]

[0066] S257: Extract the micro - motion feature values:

[0067]

[0068] S258: Constructing the UAV micro-motion feature set: Preferably, in step S223, the BFC score function of each element in W is calculated for:

[0069] In step S258, the drone micro-motion feature library is constructed as follows:

[0070]

[0071] Preferably, in step S3, the following steps are included:

[0072] S31: Build model training dataset

[0073] S32: Construct sub-dataset

[0074] S3021: Divide the data into Γ groups equally and assign each group of data a group number γ.

[0075] S3022: Take one of the data sets without duplication as the test set The remaining Γ-1 groups are used as training sets The test set and the training set constitute a sub-dataset

[0076] S33: Solve the optimization problem and obtain a classification model;

[0077] S34: If γ is less than Γ, set γ=γ+1 and execute S32; otherwise, execute S35;

[0078] S35: Calculate the average accuracy of the model

[0079] S36: If ∈ is less than E, set ∈=∈+1 and execute S31; otherwise, execute S37;

[0080] S37: Obtain the optimal model opt_M and the optimal feature combination opt_F.

[0081] Preferably, in step S32, a sub-dataset is constructed Specifically: Divide the data into Γ groups equally, and assign a group number γ to each group of data; take one of the groups of data without duplication as the test set The remaining Γ-1 groups are used as training sets The test set and the training set constitute a sub-dataset

[0082] Preferably, in step S4, the following steps are included:

[0083] S41: Input the dataset to be type-identified where represents the data composed of the optimal feature combination;

[0084] S42: Use opt_M to classify the samples χ opt in D num (num = 1, 2, …, NUM):

[0085]

[0086] In the formula, y num ∈{+1, -1} represents the model prediction type of the sample χ num ;

[0087] S43: If num < NUM, then num = num + 1 and execute S42; otherwise, execute S44.

[0088] S44: Output the UAV prediction type Y = {y1, y2, …, y NuM}}.

[0089] Compared with the prior art, the present invention has the following beneficial effects:

[0090] The present invention provides a micro-Doppler feature hierarchical analysis method for small UAV identification. By preprocessing the echo signal of the small UAV, the body Doppler signal is suppressed, the influence of the shielding effect on the extraction of micro-motion features is reduced, and the stability of the radar system in remotely identifying small UAVs is increased; through the BFC scoring mechanism, the rotor micro-Doppler signal is separated and extracted from the noise, and the signal-to-noise ratio of the rotor micro-Doppler signal is improved; according to prior knowledge, effective micro-motion features are extracted and the optimal lightweight UAV classification model is trained, which not only improves the recognition accuracy of the target but also makes the model easier to be deployed in the air surveillance radar system. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 is a flowchart of a micro-Doppler feature hierarchical analysis method for small UAV identification according to the present invention;

[0092] Figure 2 is a flowchart of the preprocessing of the small UAV echo signal according to the present invention;

[0093] Figure 3 is a flowchart of the eigenvalue decomposition of the matrices HH T and H T H according to the present invention;

[0094] Figure 4 This is a flow chart of extracting and separating the UAV micro-Doppler signal of the present invention;

[0095] Figure 5 This is a flow chart of the present invention for performing BFC scoring on independent components of each order;

[0096] Figure 6 The micro-motion feature data set D is constructed in the present invention. m_Ds Flowchart of

[0097] Figure 7 A flowchart for training the drone classification model of the present invention;

[0098] Figure 8 A flowchart for drone type identification of the present invention; DETAILED DESCRIPTION

[0099] The present invention is further described in detail below in conjunction with the embodiments and drawings.

[0100] Example 1

[0101] like Figure 1 A micro-Doppler feature classification analysis method for small UAV identification is shown, comprising the following specific steps:

[0102] S1: Preprocessing of small drone echo signals. Preprocess a small drone echo signal s(n) with a length of N to remove the Doppler signal component s of the drone body. b (n), where N=512;

[0103] S2: Separation and extraction of drone micro-Doppler signal. Separate and extract the drone micro-Doppler signal component s m_Ds (n), and construct the UAV micro-motion feature set D m_Ds ;

[0104] S3: Drone classification model training. Using training dataset Build a model training dataset Train the drone classification model, find the optimal drone micro-motion feature combination opt_F, and save the optimal classification model opt_M;

[0105] S4: UAV type identification. According to the UAV micro-motion characteristics D m_Ds The set guidance radar system selects the data set from the measured data set and inputs it into the optimal classification model opt_M for UAV type identification.

[0106] Example 2

[0107] like Figure 2As shown, in S1, the method for preprocessing the small UAV echo signal includes the following specific steps:

[0108] S11: Construct the embedding matrix H = [S1, S2, ..., S K ]. Specifically includes the following decomposition steps:

[0109] S111: Decompose the signal s(n) and construct a forward sliding window sequence S k =[s(k),s(k+1),…,s(k+L-1)] T , where k = 1, 2, …, K, K = N-L+1; L is the forward sliding window size.

[0110] For example, when N = 512, L = 128, K = 385, the forward sliding window sequence S k =[s(k),s(k+1),…,x(k+127)] T , where k = 1, 2,…, 385.

[0111] S112: Construct embedding matrix:

[0112]

[0113] For example, when N=512, L=128, K=385,

[0114]

[0115] S12: Pair matrix HH T and H T H Eigenvalue decomposition.

[0116] S12 pair matrix HH T and H T The method of H eigenvalue decomposition is as follows Figure 3 As shown, the specific steps include:

[0117] S121: Let square matrix A = HH T , flag = 0. For example, flag=0.

[0118] S122: Setting C, Z and initialization vector Y 0,1 , let C=1,Z=1.

[0119] Where C represents the maximum number of iterations; Z represents the dimension of the matrix A; Υ 0,1 represents the initial eigenvector; c represents the current iteration number; z represents the current eigenvalue index. For example, C = 50, Z = 2,

[0120] S123: Calculate the eigenvalues and eigenvectors of the matrix:

[0121] First, execute: Υ c,z = AΥ c-1,z , then execute: λ c,z = max(Υ c,z ), and finally execute: and make

[0122] For example, when c = 1, z = 1, , λ 1,1 = max v∈γ (Υ 1,1 ) = 2,

[0123] S124: If c < C, then set c = c + 1 and execute step S123; otherwise, execute step S125.

[0124] S125: Update the square matrix: A = A - λ C,z Υ C,z Υ C,z T

[0125] For example, when c = 50, z = 1, λ 50,1 = 4, ,

[0126]

[0127] S126: If z < Z, then set z = z + 1 and execute step S123; otherwise, execute step S127.

[0128] S127: If flag ≠ 1, then save the eigenvalues and eigenvectors of matrix HH T : σ = {σ1, σ2, …, σ L}, p = {p1, p2, …, p L}, set the square matrix A = H T H, flag = 1, and execute step S122; otherwise, execute step S128. For example, if flag = 0 ≠ 1, then save the eigenvalues and eigenvectors of matrix HH T : σ = {σ1 = 4, σ2 = -1}, set flag = 1 and execute step S122.

[0129] S128: Save matrix H TEigenvalues ​​and eigenvectors of H: δ={δ1,δ2,…,δ K}, q={q1,q2,…,q K}.For example, Save the matrix H T The eigenvalue and eigenvector of H: δ = {δ1 = 0, δ2 = 10},

[0130] S13: Eigenvalue grouping and constructing the residual signal s r The subspace H of (n) r The specific decomposition steps are as follows:

[0131] S131: For σ l (l=1,2,…,L) grouping: if Then the eigenvalue is considered to belong to signal s b (n); otherwise, the eigenvalue is considered to belong to signal s r (n).

[0132] For example, when L = 128, τ = 0.98, When l=3≤5, the eigenvalue σ3 is considered to belong to signal s b (n); if l=9>5, then the eigenvalue σ9 is considered to belong to signal s r (n).

[0133] S132: Construct signal s r Subspace of (n): In the formula, β is the r (n) eigenvalue index; σ β belongs to signal s r (b) The eigenvalue of p β and q β are the corresponding eigenvectors in P and Q respectively.

[0134] For example, when L = 128, ω = 5, the signal s r The subspace of (n) is:

[0135] S14: Subspace H r Restore to a one-dimensional signal sequence s r (n).

[0136]

[0137] For example, when L=128, K=385, N=512, hour,

[0138]

[0139] Example 3

[0140] S2 UAV micro-Doppler signal extraction and separation method Figure 4 As shown, the specific steps include:

[0141] S21: Modal decomposition to obtain H-order independent components f j (n). Specifically, the decomposition steps include:

[0142] S211: Find the signal s r (n) All maximum and minimum points are fitted to form upper and lower envelopes.

[0143] S212: The average value of the upper and lower envelopes is recorded as m(n), and then m(n) is subtracted to obtain a new signal and calculate: h(n) = s r (n)-m(n),n=1,2,…,N

[0144] S213: If h(n) satisfies the following conditions: 1. The absolute value of the difference between the number of extreme points and zero-crossing points of h(n) is not greater than 1; 2. The average value of the upper and lower envelopes is 0, then let f j (n)=h(n) and execute step S214; otherwise, execute step S211.

[0145] S214: f j (n) from s r (n) is separated and we get: r(n) = s r (n)-f j (n),n=1,2,…,N

[0146] S215: If r(n) is not a monotonic function, let s r (n)=r(n) and execute step S211, repeat the cycle J times to obtain J-order independent components; otherwise, end the modal decomposition.

[0147] S22 performs BFC scoring on each order of independent components. Figure 5 As shown, the specific steps include:

[0148] S221: Let j = 1, input independent component f j (n).

[0149] S222: f j (n) is decomposed into a vector set W = {w j,1 ,w j,2 ,…,w j,I}.

[0150] For signal f j (n) is decomposed to construct the forward sliding window sequence w j,i, and its vector form can be expressed as:

[0151] w j,i = [f j (1 + (i - 1)×step), f j (2 + (i - 1)×step), …, f j (wlen + (i - 1)×step)] T , where j = 1, 2, …, J; i = 1, 2, …, I, wlen is the size of the forward sliding window; step is the step size.

[0152] For example, when N = 512, wlen = 64, step = 2, J = 13, I = 225, w j,i = [f j (2×i - 1), f j (2×i), …, f j (2×i + 62)] T , where i = 1, 2, …, 225, j = 1, 2, …, 13.

[0153] S223: Calculate the BFC scores of each element in W

[0154]

[0155] In the formula, is the mean of the forward sliding window sequence w j,i ; ρ is the score penalty factor; Z(*) is the zero-crossing operation.

[0156] For example, when I = 225, J = 13, wlen = 64, j = 1, ρ = 0.5, w 1,i = [f1(2×i - 1), f1(2×i), …, f1(2×i + 62)] t ,

[0157] S224: Calculate the total BFC score g of f j (n) j . Among them,

[0158] For example, when I = 225, J = 13, j = 1,

[0159] S225: If j < J, then let j = j + 1 and execute step S221; otherwise, execute S226.

[0160] S226: Calculate the score threshold

[0161] For example, when J = 13,

[0162] S23: Group independent components. j (n) are divided into the following two groups:

[0163] {f j_up (n)|g j_up ≥BF scorethreshold ,j_up∈{1,2,…,J}}

[0164] {f j_dowm (n)|g j_down <BF scorethreshold ,j_down∈{1,2,…,J}}

[0165] For example, when J = 13, if j = 1, g1 ≥ BF scorethreshold , then f1(n)∈{f j_up (n)|g j_up ≥BF scorethreshold ,j_up∈{1,2,…,13}}; if j=2, g2 <BF scorethreshold , then f2(n)∈{f j_down (n)|g j_down <BF scorethreshold ,j_down∈{1,2,…,13}}.

[0166] S24: Reconstruction of micro-Doppler signal m_Ds (n) = ∑f j_up (n).

[0167] S25: Constructing micro-motion feature dataset

[0168] in, Indicates s obtained from offline m_Ds (n) The training data set consists of the micro-motion features extracted; Indicates that the real-time acquired m_Ds (n) is a measured data set consisting of the micro-motion features extracted.

[0169] S25 builds micro-motion feature dataset D m_Ds Methods such as Figure 6 As shown, the specific steps include:

[0170] S251: m_Ds (n) decomposed into a vector set

[0171] For signal s m_Ds (n) Decompose and construct a forward sliding window sequence Its vector form can be expressed as: Among them, i=1,2,…,I, wlen is the forward sliding window size; step is the step size.

[0172] For example, when N=512, wlen=64, step=2, I=225, Among them, i=1,2,…,225.

[0173] S252: Extract micro-motion feature vector:

[0174]

[0175] In the formula,

[0176] For example, when hour,

[0177] S253: Extract micro-motion feature vector:

[0178]

[0179] For example, when hour,

[0180] S254: Extract micro-motion feature vector:

[0181]

[0182] For example, when hour,

[0183] S255: Extract micro-motion feature vector:

[0184]

[0185] In the formula, Indicates return Maximum value operation.

[0186] For example, when hour,

[0187] S256: Extract micro motion feature values:

[0188]

[0189] For example, when hour,

[0190] S257: Extract micro motion feature values:

[0191]

[0192] For example, when s(n) = [1, 2, 1, 3, 2, 4], s r (n) = [-1, 0, 1, 2, 1, 3], s m_Ds When (n)=[0,1,1,2,0,1], ED=11-7=4.

[0193] S258: Build a UAV micro-motion feature library:

[0194] Example 4

[0195] S3 drone classification model training method Figure 7 As shown, the specific steps include:

[0196] S31: Build model training dataset

[0197] in, μ represents the The number of eigenvectors or eigenvalues ​​used to train the model is selected without duplication; num=1,2,…,NUM, NUM represents The total number of samples in; ∈=1,2,…,E, E represents the number of samples from The total number of all combinations of μ eigenvectors or eigenvalues ​​selected without repetition to form the combination.

[0198] For example, when μ = 4, When ∈=1,

[0199] S32: Construct sub-dataset The specific decomposition steps are as follows:

[0200] S321: Divide the data into Γ groups and assign a group number γ to each group of data. For example, Γ=5.

[0201] S322: Take the γth group of data as the test set without repetition The remaining Γ-1 groups are used as training sets The test set and the training set constitute a sub-dataset

[0202] For example, when γ=1,∈=1,

[0203] S33: Solve the optimization problem and obtain a classification model.

[0204] The optimization problem can be expressed as follows:

[0205]

[0206] 0≤α m_i ≤C,m_i=1,2,…,M

[0207] In the formula, α={α m_i ,α m_j} represents the Lagrange multiplier added under the constraint condition; y={y m_i ,y m_j}express The true labels of the training samples in ; χ={χ m_i ,x m_j}express The training samples in express The total number of training samples in; κ(*,*) represents the kernel function; C represents the model penalty factor.

[0208] The classification model can be expressed as follows:

[0209]

[0210] In the formula, model γ,∈ (x test ) represents the classification model; χ test express The test sample in ; b represents the offset term.

[0211] For example, when γ=1,∈=1,Γ=5,NUM=100, When C=10,b=3,M=80,y∈{-1,1},

[0212]

[0213] The optimization problem can be expressed as follows:

[0214]

[0215] 0≤α m_i ≤10,m_i=1,2,…,80

[0216] The classification model can be expressed as follows:

[0217]

[0218] S34: If γ is less than Γ, set γ=γ+1 and execute S302; otherwise, execute S35.

[0219] S35: Calculate the average accuracy of the model

[0220] in, In the formula, Indicates the model prediction label of the m_test test sample; Indicates the true label of the m_test test sample.

[0221] For example, when γ=1,∈=1,Γ=5,NUM=100. Then df=0; if Then df=1.

[0222] S36: If ∈ is less than E, set ∈=∈+1 and execute S31; otherwise, execute S37.

[0223] S37: Obtain the optimal model opt_M and the optimal feature combination opt_F.

[0224] Where opt_M = model ζ,ξ (x test ), ζ=argmax ∈ {A_acc ∈},ξ=argmax γ {acc γ,ζ}; Represents the features that make up the opt_M training set.

[0225] For example, when ζ=1,ξ=2, When opt_M = model 1,2 (x test ),

[0226] Example 5

[0227] S4 UAV type identification method Figure 8 As shown, the specific steps include:

[0228] S41: Input the data set to be type identified

[0229] in, Represents data consisting of the optimal combination of features.

[0230] For example, when NUM = 50,

[0231] S42: Use opt_M to classify the samples χ opt in D num (num = 1, 2,..., NUM):

[0232]

[0233] where y num ∈ {+1, -1} represents the model prediction type of the sample χ num .

[0234] For example, when M = 80, b = 3, χ num ∈ D opt , if then y num = +1; if then y num = -1.

[0235] S43: If num < NUM, then num = num + 1 and execute S42; otherwise, execute S44.

[0236] S44: Output the drone prediction type Y = {y1, y2,..., y NUM}}. For example, Y = {y1 = 1, y2 = -1,..., y NUM = 1}.

[0237] The present invention uses a preprocessing method for small drone echo signals to suppress the fuselage Doppler signal, reduce the influence of the shielding effect on the extraction of micro-motion features, and increase the stability of the radar system in remotely identifying small drones; the present invention separates and extracts the rotor micro-Doppler signal from noise through a BFC scoring mechanism, improving the signal-to-noise ratio of the rotor micro-Doppler signal; the present invention extracts effective micro-motion features based on prior knowledge and trains an optimal lightweight drone classification model, which not only improves the recognition accuracy of the target but also makes the model easier to deploy in an air surveillance radar system.

[0238] The above are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above content. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A micro-Doppler feature hierarchical analysis method for small unmanned aerial vehicle recognition, characterized by comprising the following steps: S1: Preprocess the echo signal of the small unmanned aerial vehicle to remove the body Doppler signal component of the unmanned aerial vehicle; S2: Separate and extract the micro-Doppler signal component of the unmanned aerial vehicle, and construct a micro-motion feature set of the unmanned aerial vehicle, which includes a training data set and a measured data set; In step S2, the separation and extraction of the micro-Doppler signal of the unmanned aerial vehicle includes the following steps: S21: Modal decomposition to obtain J-order independent components f j (n); S22: Perform BFC scoring on each order of independent components; S23: Independent component grouping; f j (n) are divided into the following two groups: {f j_dowm (n)|g j_down <BF score threshold ,j_down∈{1,2,…,J}} S24: Reconstruction of micro-Doppler signals S25: Constructing micro-motion feature dataset In step S21, specifically: S211: Find the signal s r (n) All maximum and minimum points are fitted to form upper and lower envelopes; S212: The average value of the upper and lower envelopes is recorded as m(n), and then m(n) is subtracted to obtain a new signal and calculate: h(n) = s r (n)-m(n),n=1,2,…,N; S213: If h(n) satisfies the following conditions:

1. The absolute value of the difference between the number of extreme points and zero-crossing points of h(n) is not greater than 1; 2. The average values ​​of the upper and lower envelopes are zero, then let f j (n) = h(n) and execute step S214; otherwise, execute step S211; S214: f j (n) from s r (n) is separated and we get: r(n) = s r (n)-f j (n),n=1,2,…,N S215: If r(n) is not a monotonic function, let s r (n) = r(n) and execute step S211, repeat the cycle J times to obtain the J-order independent component; otherwise, end the modal decomposition; In step S22, it includes the following steps: S221: Let j = 1, input independent component f j (n); S222: f j (n) is decomposed into a vector set W = {w j,1 ,w j,2 ,…,w j,I }; S223: Calculate the BFC score of each element in W S224: Calculate f j (n) total BFC score g j ;in, S225: If j < J, then let j = j + 1 and execute step S221, otherwise, execute S226; S226: Calculate score threshold In step S25, construct a micro-motion feature data set The following steps are involved: S251: m_Ds (n) decomposed into a vector set S252: Extract micro-motion feature vector: S253: Extract micro-motion feature vector: S254: Extract micro-motion feature vector: S255: Extract micro-motion feature vector: S256: Extract micro motion feature values: S257: Extract micro motion feature values: S258: Construct a micro-motion feature set of the unmanned aerial vehicle; S3: Construct an unmanned aerial vehicle classification model, use the training data set in the model training data set to train the unmanned aerial vehicle classification model, and establish the optimal micro-motion feature combination of the unmanned aerial vehicle and save the optimal classification model; S4: According to the micro-motion feature combination of the unmanned aerial vehicle, guide the radar system to screen out a data set from the measured data set and input it into the optimal classification model for unmanned aerial vehicle type recognition.

2. The micro-Doppler feature classification analysis method for small UAV identification according to claim 1 is characterized in that: In step S1, the preprocessing includes the following steps: S11: Construct the embedding matrix H = [S1, S2, ..., S K ]; S12: Left singular value matrix HH in the matrix T and the right singular value matrix H T H is decomposed; S13: Group the left and right singular values ​​and construct the residual signal s r The subspace H of (n) r ; S14: Subspace H r Restore to a one-dimensional signal sequence s r (n).

3. The micro-Doppler feature classification analysis method for small UAV identification according to claim 2 is characterized in that: In step S11, the small drone echo signal s(n) is decomposed to construct a forward sliding window sequence S k =[s(k),s(k+1),…,s(k+L-1)] T , where k = 1, 2, ..., K, K = N-L+1; L is the forward sliding window size; Construct an embedding matrix: In step S12, S121, let square matrix A = HH T , flag = 0; S122, set C, Z and initialization vector Y 0,1 , let C = 1, Z = 1; S123: Calculate the matrix eigenvalues and eigenvectors: First execute: c,z =A c-1,z , then execute: c,z =max(Υ c,z ), and finally execute: and make S124: If c < C, then let c = c + 1 and execute step S123; otherwise, execute step S125; S125, update matrix: A = A - λ C,z Υ C,z Υ C,z T S126: If z < Z, then let z = z + 1 and execute step S123; otherwise, execute step S127; S127, if flag≠1, save the matrix HH T Eigenvalues ​​and eigenvectors: σ={σ1,σ2,…,σ L }, p={p1,p2,…,p L }, let the square matrix A = H T H, flag = 1, and execute step S122; otherwise, execute step S128; 128. Save the matrix H T Eigenvalues ​​and eigenvectors of H: δ={δ1,δ2,…,δ K }, q={q1,q2,…,q K }; In step S13, l Grouping, l=1,2,…,L: If Then the eigenvalue is considered to belong to signal s b (n); otherwise, the eigenvalue is considered to belong to signal s r (n); Constructing signals r Subspace of (n): In the formula, β is the r (n) eigenvalue index; σ β belongs to signal s r (n) eigenvalue; P β and Q β are the corresponding eigenvectors in P and Q respectively; In step S14, the one-dimensional signal sequence s r (n) Specifically:

4. The micro-Doppler feature classification analysis method for small UAV identification according to claim 3 is characterized in that: In step S223, the BFC score function of each element in W is calculated for: In the formula, is the forward sliding window sequence w j,i ; ρ is the score penalty factor; Z(*) is the zero-crossing operation.

5. The micro-Doppler feature classification analysis method for small UAV identification according to claim 4 is characterized in that: In step S258, the micro-motion feature library of the unmanned aerial vehicle is:

6. The micro-Doppler feature classification analysis method for small UAV identification according to claim 5 is characterized in that: In step S3, it includes the following steps: S31: Build model training dataset in, μ represents the The number of eigenvectors or eigenvalues ​​selected for training the model without duplication; num=1,2,…,NUM, NUM represents The total number of samples in; ∈=1,2,…,E, E represents the number of samples from The total number of all combinations of μ eigenvectors or eigenvalues ​​selected without repetition to form a combination; S32: Construct sub-dataset S321: Divide into Γ groups, and assign group number γ to each group of data. S322: Take one of the groups of data without duplication as the test set The remaining Γ-1 groups are used as training sets The test set and the training set constitute a sub-dataset S33: Solve the optimization problem to obtain the classification model; S34: If γ is less than Γ, then let γ = γ + 1 and execute S32; otherwise, execute S35; S35: Calculate the average accuracy of the model S36: If ∈ is less than Ε, then let ∈ = ∈ + 1 and execute S31; otherwise, execute S37; S37: Obtain the optimal model opt_M and the optimal feature combination opt_F.

7. The micro-Doppler feature classification analysis method for small UAV identification according to claim 6 is characterized in that: In step S4, it includes the following steps: S41: Input the data set to be type identified in, Represents data consisting of the optimal combination of features; S42: Using opt_M to D opt The sample χ in num To classify, num=1,2,…,NUM: In the formula, y num ∈{+1,-1} represents sample X num Model prediction type; S43: If num < NUM, then num = num + 1 and execute S42; otherwise, execute S44, S44: Output drone prediction type Y = {y1, y2, ..., y NUM }.

Images