Robust recognition method of radar targets based on multi-domain feature fusion under short dwell conditions

Abstract

The present invention relates to a radar target robust identification method based on multi-domain feature fusion under short dwell conditions, comprising: step 1: obtaining a test data set; step 2: generating a test feature matrix according to the test data set; step 3: inputting the test feature matrix into a trained SVM classifier to obtain a target classification result; wherein the test data set includes multiple radar echo signals, and each X radar echo signals constitute a test sample, X≥3; using the mean feature and variance feature of the time domain entropy matrix, the time domain variance matrix, the Doppler domain entropy matrix and the Doppler domain variance matrix of each test sample in the test data set, the test feature matrix is ​​constructed. The method of the present invention overcomes the problem that the radar has a relatively short dwell time when observing a target, and the resolution of the micro-Doppler modulation spectrum of the target is reduced, resulting in a reduced target classification accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a radar target robust recognition method based on multi-domain feature fusion under short dwell conditions. Background Art

[0002] Radar target recognition refers to the use of the target's radar echo signal to determine the target type. In actual work, due to limited resources, the radar's dwell time when observing the target will be relatively short, even shorter than the time domain echo period of the target's rotating parts, and the target's micro-Doppler modulation spectrum resolution will be reduced. At this time, if the traditional classification method is used, the accuracy of target classification will be reduced.

[0003] To address this problem, the radar multiple short-dwell observation fusion classification method proposed in the existing literature mainly performs voting fusion on the classification results of radar single short-dwell observation samples. The voting method is a simple fusion method, and its basic principle is "the minority obeys the majority". However, due to the reduced resolution of the target's micro-Doppler modulation spectrum, the classification results of radar single short-dwell observation samples are not robust, and when the same number of votes appears, the voting results cannot be determined. Therefore, when the classification results of radar single short-dwell observation samples are voted and fused, the results are not robust.

[0004] In response to this problem, researchers have proposed a new robust target recognition method, which mainly solves the problem of the degradation of aircraft target classification performance when the radar carrier frequency changes in the existing technology, and solves the problem of radar target recognition when the pulse repetition frequency does not match in the training and testing stages. However, in the actual working process, due to limited resources, the radar's dwell time when observing the target will be relatively short, even less than the time domain echo period of the target's rotating parts, and the resolution of the target's micro-Doppler modulation spectrum will be reduced. The above classification method has not conducted relevant research experiments on this problem, and the use of this method will still have the problem of reduced target classification accuracy. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a radar target robust recognition method based on multi-domain feature fusion under short dwell conditions. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] The present invention provides a radar target robust recognition method based on multi-domain feature fusion under short dwell conditions, comprising:

[0007] Step 1: Get the test dataset;

[0008] Step 2: Generate a test feature matrix based on the test data set;

[0009] Step 3: Input the test feature matrix into the trained SVM classifier to obtain the target classification result;

[0010] The test data set includes multiple radar echo signals, and each X radar echo signals form a test sample, where X≥3; the test feature matrix is ​​constructed by using the mean characteristics and variance characteristics of the time domain entropy matrix, the time domain variance matrix, the Doppler domain entropy matrix, and the Doppler domain variance matrix of each test sample in the test data set.

[0011] In one embodiment of the present invention, before the target classification is performed on the test data set, the method further includes training the SVM classifier to obtain a trained SVM classifier, and the training method of the SVM classifier includes:

[0012] S1: Generate a training data set; the training data set includes a plurality of training samples, each of which is marked with a corresponding category label;

[0013] S2: Generate a training feature matrix according to the training data set;

[0014] S3: Train the SVM classifier using the training feature matrix and the corresponding category labels to obtain a trained SVM classifier.

[0015] In one embodiment of the present invention, the S1 includes:

[0016] S11: Acquire radar echo signals including Y categories, each category including Z radar echo signals, where Y≥3 and Z≥6000;

[0017] S12: performing clutter suppression and main component removal processing on each radar echo signal;

[0018] S13: performing modulo-2 norm normalization processing on the amplitude of each radar echo signal after removing the main component;

[0019] S14: In each category, combine every X normalized radar echo signals into a two-dimensional matrix as a training sample, X ≥ 3;

[0020] S15: Add a corresponding category label to each of the training samples to obtain the training data set.

[0021] In one embodiment of the present invention, the S2 includes:

[0022] S21: Calculate the waveform entropy and amplitude variance of each time domain echo signal in each training sample, and construct a time domain entropy matrix and a time domain variance matrix corresponding to each training sample;

[0023] S22: extracting L-dimensional mean features and variance features of each time-domain entropy matrix, and L-dimensional mean features and variance features of each time-domain variance matrix, where L=fix(X / 2), indicating the number of columns of the matrix;

[0024] S23: performing a fast Fourier transform on each time domain echo signal in each training sample to obtain a corresponding Doppler domain echo signal;

[0025] S24: Calculate the waveform entropy and amplitude variance of each Doppler domain echo signal in each training sample, and construct a Doppler domain entropy matrix and a Doppler domain variance matrix corresponding to each training sample;

[0026] S25: extracting L-dimensional mean features and variance features of each Doppler domain entropy matrix, and L-dimensional mean features and variance features of each Doppler domain variance matrix, where L=fix(X / 2), representing the number of columns of the matrix;

[0027] S26: construct an initial training feature matrix F according to the L-dimensional mean and variance features of the time domain entropy matrix of each training sample in the training data set, the L-dimensional mean and variance features of the time domain variance matrix, the L-dimensional mean and variance features of the Doppler domain entropy matrix, and the L-dimensional mean and variance features of the Doppler domain variance matrix train , and the initial training feature matrix F train Perform normalization to obtain the training feature matrix

[0028] In one embodiment of the present invention, the S21 includes:

[0029] S211: Calculate the waveform entropy and amplitude variance of each time domain echo signal in each training sample according to the following formula:

[0030]

[0031]

[0032] Where k = 1, 2, ..., P, P represents the number of points of the time domain echo signal, s i (k) represents the i-th time domain echo signal in a training sample, i = 1, 2, ..., X, X represents the number of echo signals contained in a training sample, a i Represents the time domain echo signal s i (k) waveform entropy, c i Represents the time domain echo signal s i (k) Amplitude variance;

[0033] S212: According to the waveform entropy and amplitude variance of each time domain echo signal in each training sample, a time domain entropy vector A and a time domain variance vector C corresponding to each training sample are constructed:

[0034] A=[a1,a2,...,a i ,...,a X ];

[0035] C=[c1,c2,...,c i ,...,c X ];

[0036] S213: According to the time domain entropy vector A and the time domain variance vector C corresponding to each training sample, a time domain entropy matrix S of each training sample is constructed. A And the time domain variance matrix S C :

[0037]

[0038]

[0039] Wherein, L=fix(X / 2), which indicates the number of columns of the matrix, and fix(·) indicates rounding toward zero.

[0040] In one embodiment of the present invention, the S24 includes:

[0041] S241: Calculate the waveform entropy and amplitude variance of each Doppler domain echo signal in each training sample according to the following formula:

[0042]

[0043]

[0044] Where k = 1, 2, ..., P, P represents the number of points of Doppler domain echo signal, S i (k) represents the i-th Doppler domain echo signal in a training sample, i = 1, 2, ..., X, X represents the number of echo signals contained in a training sample, bi Represents the Doppler domain echo signal S i (k) waveform entropy, di Represents the Doppler domain echo signal S i (k) Amplitude variance;

[0045] S242: According to the waveform entropy and amplitude variance of each Doppler domain echo signal in each training sample, a Doppler domain entropy vector B and a Doppler domain variance vector D corresponding to each training sample are constructed:

[0046] B=[b1,b2,...,bi ,...,b X ];

[0047] D=[d1,d2,...,d i ,...,d X ];

[0048] S243: Construct a Doppler domain entropy matrix S for each training sample based on the Doppler domain entropy vector B and the Doppler domain variance vector D corresponding to each training sample. B and the Doppler domain variance matrix S D :

[0049]

[0050]

[0051] Wherein, L=fix(X / 2), which indicates the number of columns of the matrix, and fix(·) indicates rounding toward zero.

[0052] In one embodiment of the present invention, the S26 includes:

[0053] S261: Construct a feature vector f corresponding to each training sample based on the L-dimensional mean and variance features of the time domain entropy matrix of each training sample, the L-dimensional mean and variance features of the time domain variance matrix, the L-dimensional mean and variance features of the Doppler domain entropy matrix, and the L-dimensional mean and variance features of the Doppler domain variance matrix. n :

[0054] f n =[μ A ,μ B ,μ C ,μ D ,σ A ,σ B ,σ C ,σ D ];

[0055] Where n = 1, 2, ..., N, represents the sequence number of the training sample, N represents the total number of training samples, μ A represents the L-dimensional mean characteristic of the time domain entropy matrix, σ A represents the L-dimensional variance characteristics of the time domain entropy matrix, μ B represents the L-dimensional mean characteristic of the Doppler domain entropy matrix, σ B represents the L-dimensional variance characteristic of the Doppler domain entropy matrix, μ C represents the L-dimensional mean characteristic of the time domain variance matrix, σ C represents the L-dimensional variance feature of the time domain variance matrix, μ D represents the L-dimensional mean characteristic of the Doppler domain variance matrix, σD Represents the L-dimensional variance characteristics of the Doppler domain variance matrix;

[0056] S262: According to the feature vector f corresponding to each training sample n , construct the initial training feature matrix F train :

[0057] F train =[f1;f2;...;f n ;...;f N ];

[0058] S263: Calculate the initial training feature matrix F train The mean and standard deviation of the eigenvalues ​​of each dimension in , and the mean vector μ and standard deviation vector σ are obtained:

[0059] μ=[μ1,μ2,μ3,...,μ L×8 ];

[0060] σ=[σ1,σ2,σ3,...,σ L×8 ];

[0061] S264: Initial training feature matrix F train In the example, the feature vector f corresponding to each training sample is n Perform normalization processing to obtain the corresponding normalized results

[0062]

[0063] S265: Normalizing the training sample according to the result Get the training feature matrix

[0064]

[0065] In one embodiment of the present invention, step 1 comprises:

[0066] Step 1.1: Acquire multiple radar echo signals received in real time;

[0067] Step 1.2: Perform clutter suppression and main component removal processing on each radar echo signal;

[0068] Step 1.3: Perform modulo-2 norm normalization on the amplitude of each radar echo signal after removing the main component;

[0069] Step 1.4: Combine every X normalized radar echo signals into a two-dimensional matrix as a test sample, X≥3, and a plurality of the test samples constitute the test data set.

[0070] In one embodiment of the present invention, step 2 comprises:

[0071] Step 2.1: Calculate the waveform entropy and amplitude variance of each time domain echo signal in each test sample, and construct a time domain entropy matrix and a time domain variance matrix corresponding to each test sample;

[0072] Step 2.2: Extract the L-dimensional mean feature and variance feature of each time-domain entropy matrix, and the L-dimensional mean feature and variance feature of each time-domain variance matrix, where L=fix(X / 2), indicating the number of columns of the matrix;

[0073] Step 2.3: Perform a fast Fourier transform on each time domain echo signal in each test sample to obtain the corresponding Doppler domain echo signal;

[0074] Step 2.4: Calculate the waveform entropy and amplitude variance of each Doppler domain echo signal in each test sample, and construct a Doppler domain entropy matrix and a Doppler domain variance matrix corresponding to each test sample;

[0075] Step 2.5: Extract the L-dimensional mean feature and variance feature of each Doppler domain entropy matrix, and the L-dimensional mean feature and variance feature of each Doppler domain variance matrix, where L=fix(X / 2), indicating the number of columns of the matrix;

[0076] Step 2.6: According to the L-dimensional mean and variance characteristics of the time domain entropy matrix of each test sample in the test data set, the L-dimensional mean and variance characteristics of the time domain variance matrix, the L-dimensional mean and variance characteristics of the Doppler domain entropy matrix, and the L-dimensional mean and variance characteristics of the Doppler domain variance matrix, construct an initial test feature matrix F test , and the initial test feature matrix F test Perform normalization to obtain the test feature matrix

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] 1. The robust identification method of radar targets based on multi-domain feature fusion under short-dwell conditions of the present invention performs multiple short-dwell observations on the target, and combines the samples of multiple short-dwell observations into a two-dimensional matrix as a sample, thereby overcoming the problem of instability of the results when voting fusion is performed on the classification results of a single short-dwell observation sample of the radar, and can significantly reduce the influence of the radar on the classification of micro-Doppler narrowband targets when the dwell time of the radar is relatively short when observing the target, and the resolution of the micro-Doppler modulation spectrum of the target will be reduced;

[0079] 2. The robust radar target recognition method based on multi-domain feature fusion under short-dwell conditions of the present invention overcomes the problem of reduced target classification accuracy due to reduced micro-Doppler modulation spectrum resolution of the target when the radar dwells for a short time when observing the target, by extracting the mean characteristics and variance characteristics of the time domain entropy matrix, time domain variance matrix, Doppler domain entropy matrix and Doppler domain variance matrix of the samples after multiple short-dwell observations of the radar. The method of the present invention can eliminate the influence of incomplete micro-Doppler information of the target under short dwell time by performing multiple short-dwell observations on the target, and still has good classification performance when the dwell time is short.

[0080] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the following specifically cites a preferred embodiment and describes it in detail with the accompanying drawings as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 is a schematic diagram of a radar target robust recognition method based on multi-domain feature fusion under short dwell conditions provided by an embodiment of the present invention;

[0082] Figure 2 It is a simulation experiment result diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0083] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description of a robust radar target recognition method based on multi-domain feature fusion under short dwell conditions proposed by the present invention in combination with the accompanying drawings and specific implementation methods.

[0084] The above and other technical contents, features and effects of the present invention are clearly presented in the following detailed description of the specific implementation modes in conjunction with the accompanying drawings. Through the description of the specific implementation modes, the technical means and effects adopted by the present invention to achieve the predetermined purpose can be more deeply and specifically understood. However, the attached drawings are only for reference and explanation purposes and are not used to limit the technical solutions of the present invention.

[0085] Embodiment 1

[0086] See also Figure 1 , Figure 1 : is a schematic diagram of a radar target robust identification method based on multi-domain feature fusion under short-dwell conditions provided by an embodiment of the present invention. As shown in the figure, the radar target robust identification method based on multi-domain feature fusion under short-dwell conditions of this embodiment includes:

[0087] Step 1: Get the test dataset;

[0088] Step 2: Generate a test feature matrix based on the test data set;

[0089] Step 3: Input the test feature matrix into the trained SVM classifier to obtain the target classification result;

[0090] The test data set includes multiple radar echo signals, and each X radar echo signals constitute a test sample, where X≥3; a test feature matrix is ​​constructed by using the mean characteristics and variance characteristics of the time domain entropy matrix, time domain variance matrix, Doppler domain entropy matrix, and Doppler domain variance matrix of each test sample in the test data set.

[0091] In this embodiment, before the target classification is performed on the test data set, the SVM classifier is trained to obtain a trained SVM classifier. The training method of the SVM classifier includes:

[0092] S1: Generate a training data set; the training data set includes multiple training samples, each training sample is marked with a corresponding category label;

[0093] Specifically, S1 includes:

[0094] S11: Acquire radar echo signals including Y categories, each category including Z radar echo signals, where Y≥3 and Z≥6000;

[0095] S12: performing clutter suppression and main component removal processing on each radar echo signal;

[0096] S13: performing modulo-2 norm normalization processing on the amplitude of each radar echo signal after removing the main component;

[0097] S14: In each category, combine every X normalized radar echo signals into a two-dimensional matrix as a training sample, X ≥ 3;

[0098] In this embodiment, through the above grouping, each category contains Z / X training samples, X≥3. The time domain form of a training sample is represented by u(k):

[0099] u(k)=[s1(k);s2(k);...;s i (k); ...; s X (k)] (1);

[0100] Among them, s i (k) represents the i-th time domain echo signal in a training sample, i=1, 2, ..., X, and X represents the number of echo signals contained in a training sample.

[0101] S15: Add a corresponding category label to each training sample to obtain a training data set.

[0102] S2: Generate a training feature matrix based on the training data set;

[0103] Specifically, S2 includes:

[0104] S21: Calculate each time domain echo signal s in each training sample i (k) waveform entropy a i and amplitude variance c i , construct the time domain entropy matrix S corresponding to each training sample A And the time domain variance matrix S C ;

[0105] In this embodiment, S21 includes:

[0106] S211: Calculate each time domain echo signal s in each training sample according to the following formula i (k) waveform entropy a i and amplitude variance c i :

[0107]

[0108]

[0109] Where k = 1, 2, ..., P, P represents the number of points of the time domain echo signal, s i (k) represents the i-th time domain echo signal in a training sample, i=1, 2, ..., X, X represents the number of echo signals contained in a training sample;

[0110] S212: According to each time domain echo signal s in each training sample i (k) waveform entropy a i and amplitude variance c i , construct the time domain entropy vector A and time domain variance vector C corresponding to each training sample:

[0111] A=[a1,a2,...,a i ,...,a X ] (4);

[0112] C=[c1,c2,...,c i ,...,c X ] (5);

[0113] S213: According to the time domain entropy vector A and the time domain variance vector C corresponding to each training sample, a time domain entropy matrix S of each training sample is constructed. A And the time domain variance matrix SC :

[0114]

[0115]

[0116] Wherein, L=fix(X / 2), represents the number of columns of the matrix, and fix(·) represents rounding toward zero, that is, sliding the time domain entropy vector A and the time domain variance vector C using a window length of fix(X / 2).

[0117] S22: Extract each time domain entropy matrix S A The L-dimensional mean feature μ A and variance characteristic σ A , and each time domain variance matrix S C The L-dimensional mean feature μ C and variance characteristic σ C , where L = fix(X / 2), represents the number of columns of the matrix;

[0118] In this embodiment, S22 includes:

[0119] S221: Calculate each time domain entropy matrix S according to the following formula A The L-dimensional mean feature μ A and variance characteristic σ A :

[0120]

[0121]

[0122] S222: Calculate each time domain variance matrix S C The L-dimensional mean feature μ C and variance characteristic σ C :

[0123]

[0124]

[0125] Among them, mean(·) means calculating the mean, and var(·) means calculating the variance.

[0126] S23: For each time domain echo signal s in each training sample i (k), perform fast Fourier transform to obtain the corresponding Doppler domain echo signal S i (k);

[0127] Specifically, for each time domain echo signal s in each training sample i(k), perform fast Fourier transform at point P to obtain its corresponding Doppler domain echo signal S i (k), the Doppler domain form of a training sample is expressed as U(k):

[0128] U(k)=[S1(k); S2(k);...;S i (k); ...; S X (k)] (12);

[0129] Wherein, i=1, 2, ..., X, and X represents the number of echo signals contained in a training sample.

[0130] S24: Calculate each Doppler domain echo signal S in each training sample i (k) waveform entropy b i and amplitude variance d i , construct the Doppler domain entropy matrix S corresponding to each training sample B and the Doppler domain variance matrix S D ;

[0131] In this embodiment, S24 includes:

[0132] S241: Calculate each Doppler domain echo signal S in each training sample according to the following formula i (k) waveform entropy b i and amplitude variance d i :

[0133]

[0134]

[0135] Where k = 1, 2, ..., P, P represents the number of points of Doppler domain echo signal, S i (k) represents the i-th Doppler domain echo signal in a training sample, i=1, 2, ..., X, X represents the number of echo signals contained in a training sample;

[0136] S242: According to each Doppler domain echo signal S in each training sample i (k) waveform entropy b i and amplitude variance d i , construct the Doppler domain entropy vector B and Doppler domain variance vector D corresponding to each training sample:

[0137] B=[b1,b2,...,b i ,...,b X ] (15);

[0138] D=[d1,d2,...,di ,...,d X ] (16);

[0139] S243: Construct a Doppler domain entropy matrix S for each training sample based on the Doppler domain entropy vector B and the Doppler domain variance vector D corresponding to each training sample. B and the Doppler domain variance matrix S D :

[0140]

[0141]

[0142] Wherein, L=fix(X / 2), represents the number of columns of the matrix, and fix(·) represents rounding toward zero, that is, sliding the Doppler domain entropy vector B and the Doppler domain variance vector D using a window length of fix(X / 2).

[0143] S25: Extract each Doppler domain entropy matrix S B The L-dimensional mean feature μ B and variance characteristic σ B , and each Doppler domain variance matrix S D The L-dimensional mean feature μ D and variance characteristic σ D , where L = fix(X / 2), represents the number of columns of the matrix;

[0144] In this embodiment, S25 includes:

[0145] S251: Calculate each Doppler domain entropy matrix S according to the following formula B The L-dimensional mean feature μ B and variance characteristic σ B :

[0146]

[0147]

[0148] S252: Calculate each Doppler domain variance matrix S D The L-dimensional mean feature μ D and variance characteristic σ D :

[0149]

[0150]

[0151] Among them, mean(·) means calculating the mean, and var(·) means calculating the variance.

[0152] S26: According to the time domain entropy matrix S of each training sample in the training data set A The L-dimensional mean feature μ A and variance characteristic σ A , time domain variance matrix S C The L-dimensional mean feature μ C and variance characteristic σ C , Doppler domain entropy matrix S B The L-dimensional mean feature μ B and variance characteristic σ B , and the Doppler domain variance matrix S D The L-dimensional mean feature μ D and variance characteristic σ D , construct the initial training feature matrix F train , and the initial training feature matrix F train Perform normalization to obtain the training feature matrix

[0153] In this embodiment, S26 includes:

[0154] S261: According to the time domain entropy matrix S of each training sample A The L-dimensional mean feature μ A and variance characteristic σ A , time domain variance matrix S C The L-dimensional mean feature μ C and variance characteristic σ C , Doppler domain entropy matrix S B The L-dimensional mean feature μ B and variance characteristic σ B , and the Doppler domain variance matrix S D The L-dimensional mean feature μ D and variance characteristic σ D , construct the feature vector f corresponding to each training sample n :

[0155] f n =[μ A ,μ B ,μ C ,μ D ,σ A ,σ B ,σ C ,σ D ] (twenty three);

[0156] Where n = 1, 2, ..., N, represents the sequence number of the training sample, and N represents the total number of training samples;

[0157] S622: According to the feature vector f corresponding to each training sample n, construct the initial training feature matrix F train :

[0158] F train =[f1;f2;...;f n ;...;f N ] (twenty four);

[0159] S263: Calculate the initial training feature matrix F train The mean and standard deviation of the eigenvalues ​​of each dimension in , and the mean vector μ and standard deviation vector σ are obtained:

[0160] μ=[μ1,μ2,μ3,...,μ L×8 ] (25);

[0161] σ=[σ1,σ2,σ3,...,σ L×8 ] (26);

[0162] S264: Initial training feature matrix F train In the example, the feature vector f corresponding to each training sample is n Perform normalization processing to obtain the corresponding normalized results

[0163]

[0164] S265: Normalization results corresponding to training samples Get the training feature matrix

[0165]

[0166] S3: Train the SVM classifier using the training feature matrix and the corresponding category labels to obtain a trained SVM classifier.

[0167] Specifically, select the SVM classifier and set its kernel function to the Gaussian kernel function, and set the training feature matrix and the corresponding category labels are input into the SVM classifier for training to obtain a trained SVM classifier.

[0168] Furthermore, in this embodiment, step 1 includes:

[0169] Step 1.1: Acquire multiple radar echo signals received in real time;

[0170] Step 1.2: Perform clutter suppression and main component removal processing on each radar echo signal;

[0171] Step 1.3: Perform modulo-2 norm normalization on the amplitude of each radar echo signal after removing the main component;

[0172] Step 1.4: Combine every X normalized radar echo signals into a two-dimensional matrix as a test sample, X ≥ 3, and several test samples constitute a test data set.

[0173] Furthermore, in this embodiment, step 2 includes:

[0174] Step 2.1: Calculate the waveform entropy and amplitude variance of each time domain echo signal in each test sample, and construct a time domain entropy matrix and a time domain variance matrix corresponding to each test sample;

[0175] Step 2.2: Extract the L-dimensional mean feature and variance feature of each time-domain entropy matrix, and the L-dimensional mean feature and variance feature of each time-domain variance matrix, where L=fix(X / 2), indicating the number of columns of the matrix;

[0176] Step 2.3: Perform a fast Fourier transform on each time domain echo signal in each test sample to obtain the corresponding Doppler domain echo signal;

[0177] Step 2.4: Calculate the waveform entropy and amplitude variance of each Doppler domain echo signal in each test sample, and construct a Doppler domain entropy matrix and a Doppler domain variance matrix corresponding to each test sample;

[0178] Step 2.5: Extract the L-dimensional mean feature and variance feature of each Doppler domain entropy matrix, and the L-dimensional mean feature and variance feature of each Doppler domain variance matrix, where L=fix(X / 2), indicating the number of columns of the matrix;

[0179] Step 2.6: According to the L-dimensional mean and variance characteristics of the time domain entropy matrix of each test sample in the test data set, the L-dimensional mean and variance characteristics of the time domain variance matrix, the L-dimensional mean and variance characteristics of the Doppler domain entropy matrix, and the L-dimensional mean and variance characteristics of the Doppler domain variance matrix, construct the initial test feature matrix F test , and the initial test feature matrix F test Perform normalization to obtain the test feature matrix

[0180] It should be noted that, in this embodiment, the specific steps of generating the test feature matrix are similar to the steps of generating the training feature matrix, and will not be described in detail here.

[0181] The robust radar target recognition method based on multi-domain feature fusion under short-dwell conditions of this embodiment performs multiple short-dwell observations on the target, and combines the samples of multiple short-dwell observations into a two-dimensional matrix as a sample, thereby overcoming the problem of instability of the results when voting and fusing the classification results of a single short-dwell observation sample of the radar, and can significantly reduce the impact on the classification of radar micro-Doppler narrowband targets caused by a relatively short dwell time of the radar when observing the target, which reduces the resolution of the micro-Doppler modulation spectrum of the target.

[0182] The robust radar target recognition method based on multi-domain feature fusion under short-dwell conditions of this embodiment overcomes the problem of reduced target classification accuracy due to reduced micro-Doppler modulation spectrum resolution of the target when the radar dwells for a short time when observing the target, by extracting the mean features and variance features of the time-domain entropy matrix, time-domain variance matrix, Doppler-domain entropy matrix and Doppler-domain variance matrix of the samples after multiple short-dwell observations of the radar. The method of the present invention can eliminate the influence of incomplete micro-Doppler information of the target under short-dwell time by performing multiple short-dwell observations on the target, and still has good classification performance when the dwell time is short.

[0183] Embodiment 2

[0184] This embodiment further illustrates the effect of the robust radar target recognition method based on multi-domain feature fusion under short dwell conditions of the first embodiment through simulation experiments.

[0185] 1. Simulation experiment conditions:

[0186] The hardware platform of the simulation experiment of this embodiment is: the processor is Intel(R) Core(TM) i7-7700 CPU@3.60GHZ 3.60GHZ, the main frequency is 2.00GHz, and the memory is 16GB.

[0187] The software platform for the simulation experiment of this embodiment is: Windows 10 operating system and Python 3.7.

[0188] The data used in the simulation experiment of this embodiment are 12 types of three types of aircraft generated by electromagnetic simulation software CST, namely helicopters, propellers and jets, with 4 models of each of the three types of aircraft. The specific rotor physical parameters are shown in the following table:

[0189] Table 1 Aircraft rotor physical parameters

[0190] Aircraft target model Number of leaves <![CDATA[L1(m)]]> <![CDATA[L2(m)]]> Speed ​​(rmin) Helicopter BK17 4 0 5.5 383 Helicopter Mi-17 5 0 10.645 185 Helicopter AS350 3 0 5.345 394 Helicopter Bell 212 2 0 7.315 324 Propeller SAAB2000 6 0.28 1.905 950 Propeller L-420 5 0.12 1.15 1650 Propeller L-610G 4 0.23 1.675 1150 Propeller F406 3 0.23 1.18 1690 Jet A 30 0.3 1.0 3000 Jet B 38 0.38 1.1 3520 Jet C 27 0.18 0.51 8615 Jet D 33 0.2 0.6 5000

[0191] The radar echo signal of the training data set used in the simulation experiment of this embodiment is generated under the radar operating frequency of 10GHz, the pulse repetition frequency of 5KHz, and the dwell time of 20ms. Each type of aircraft contains 6000 radar echo signals, that is, the training data set contains a total of 18000 radar echo signals. Assume that the number of radar multiple short-dwelling observation fusion classifications is 5, that is, the training data set contains a total of 3600 training samples. In the experiment, Gaussian white noise with signal-to-noise ratios of 0dB, 2dB, 4dB, 6dB, 8dB, and 10dB is added to the training data set respectively, and the influence of the signal-to-noise ratio on the target classification accuracy is analyzed, where the signal-to-noise ratio is defined as the signal-to-noise ratio of the micro-motion component to the noise.

[0192] The radar echo signal of the test data set used in the simulation experiment of this embodiment is generated under the radar operating frequency of 10GHz, the pulse repetition frequency of 5KHz, and the dwell time of 20ms. Each type of aircraft contains 2000 radar echo signals, that is, the test data set contains a total of 6000 radar echo signals. The number of radar multiple short-dwelling observation fusion classifications is the same as that in the training stage, that is, the test data set contains a total of 1200 test samples. In the experiment, Gaussian white noise with signal-to-noise ratios of 0dB, 2dB, 4dB, 6dB, 8dB, and 10dB was added to the test data set respectively, and the influence of the signal-to-noise ratio on the target classification accuracy was analyzed, where the signal-to-noise ratio is defined as the signal-to-noise ratio of the micro-motion component to the noise.

[0193] 2. Simulation content and result analysis

[0194] The simulation experiment of this embodiment adopts the method of the first embodiment and two existing technologies (non-fusion classification method and result fusion classification method) to classify the test data set respectively to obtain classification results.

[0195] In the simulation experiment, the non-fusion classification method is: the radar micro-Doppler narrow-band target classification method proposed by Xidian University in its patent application document "Ground target classification method based on robust time-frequency characteristics".

[0196] The test data set is classified using the method of Example 1. First, a training data set is generated. The number of radar multiple short-dwell observation fusion classifications is set to 5, that is, every 5 echo signals in the training data set are combined into a two-dimensional matrix as a training sample. Then, the SVM classifier is trained according to the SVM classifier training method described in Example 1 to obtain a trained SVM classifier to implement classification testing of the test data set, obtain the classification results of the test samples, and statistically analyze the classification accuracy of the method for the test data under different signal-to-noise ratio conditions.

[0197] The test data set is classified using a non-fusion classification method. According to the classification method described in the patent literature, the test samples are classified and tested to obtain the classification results of the test samples. The classification accuracy of the non-fusion classification method on the test data under different signal-to-noise ratio conditions is statistically analyzed.

[0198] The test data set is classified using the result fusion classification method, which is to vote and fuse the classification results based on the non-fusion classification method to obtain the classification results of the test samples, and then statistically analyze the classification accuracy of the test data under different signal-to-noise ratio conditions using the result fusion classification method.

[0199] The classification accuracy of the test data under different signal-to-noise ratio conditions is plotted by the method of Example 1, the non-fusion classification method, and the result fusion classification method to obtain a curve graph, as shown in FIG. Figure 2 The figure shows the simulation experiment results provided by the embodiment of the present invention.

[0200] Figure 2 The horizontal axis represents the signal-to-noise ratio of the test data, which are 0dB, 2dB, 4dB, 6dB, 8dB, and 10dB respectively. The vertical axis represents the classification accuracy of the test data under different signal-to-noise ratio conditions.

[0201] Depend on Figure 2 It can be seen that the classification accuracy of the method of the present invention for the test data under different signal-to-noise ratio conditions is higher than that of the non-fusion classification method and the result fusion classification method. In particular, when the signal-to-noise ratio is 0dB, the classification accuracy of the method of the present invention for the test data is nearly 14 percentage points higher than that of the non-fusion classification method. When the signal-to-noise ratio is 0dB-6dB, the classification accuracy of the method of the present invention for the test data is nearly 3 percentage points higher than that of the result fusion classification method. When the signal-to-noise ratio is 8dB-10dB, the classification accuracy of the method of the present invention for the test data is nearly 2 percentage points higher than that of the result fusion classification method. This shows that the method of the present invention significantly reduces the influence of the radar on the classification of micro-Doppler narrowband targets when the radar has a short dwell time when observing the target, and the resolution of the micro-Doppler modulation spectrum of the target will be reduced. By performing multiple short-dwelling observations on the target, the influence of incomplete micro-Doppler information of the target under short dwell time is eliminated. It still has good classification performance when the dwell time is short, and can have a higher target classification accuracy than the prior art under different signal-to-noise ratio conditions.

[0202] It should be noted that, in this article, the term "comprises", "comprising" or any other variant is intended to cover non-exclusive inclusion, so that an article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed. In the absence of more restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the article or device comprising the element.

[0203] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A robust radar target recognition method based on multi-domain feature fusion under short dwell conditions, characterized in that: include: Step 1: Get the test dataset; Step 2: Generate a test feature matrix based on the test data set; Step 3: Input the test feature matrix into the trained SVM classifier to obtain the target classification result; The test data set includes multiple radar echo signals, and each X radar echo signals constitute a test sample, where X≥3; the test feature matrix is ​​constructed by using the mean characteristics and variance characteristics of the time domain entropy matrix, the time domain variance matrix, the Doppler domain entropy matrix, and the Doppler domain variance matrix of each test sample in the test data set; Before the target classification is performed on the test data set, the method further includes training the SVM classifier to obtain a trained SVM classifier; the training method of the SVM classifier includes: S1: Generate a training data set; the training data set includes a plurality of training samples, each of which is marked with a corresponding category label; S2: Generate a training feature matrix according to the training data set; S3: training the SVM classifier using the training feature matrix and the corresponding category labels to obtain a trained SVM classifier; The S2 includes: S21: Calculate the waveform entropy and amplitude variance of each time domain echo signal in each training sample, and construct a time domain entropy matrix and a time domain variance matrix corresponding to each training sample; S22: extracting L-dimensional mean features and variance features of each time-domain entropy matrix, and L-dimensional mean features and variance features of each time-domain variance matrix, where L=fix(X / 2), indicating the number of columns of the matrix; S23: performing a fast Fourier transform on each time domain echo signal in each training sample to obtain a corresponding Doppler domain echo signal; S24: Calculate the waveform entropy and amplitude variance of each Doppler domain echo signal in each training sample, and construct a Doppler domain entropy matrix and a Doppler domain variance matrix corresponding to each training sample; S25: extracting L-dimensional mean features and variance features of each Doppler domain entropy matrix, and L-dimensional mean features and variance features of each Doppler domain variance matrix, where L=fix(X / 2), representing the number of columns of the matrix; S26: construct an initial training feature matrix F according to the L-dimensional mean and variance features of the time domain entropy matrix of each training sample in the training data set, the L-dimensional mean and variance features of the time domain variance matrix, the L-dimensional mean and variance features of the Doppler domain entropy matrix, and the L-dimensional mean and variance features of the Doppler domain variance matrix train , and the initial training feature matrix F train Perform normalization to obtain the training feature matrix 2. The robust radar target recognition method based on multi-domain feature fusion under short dwell conditions according to claim 1 is characterized in that: The S1 includes: S11: Acquire radar echo signals including Y categories, each category including Z radar echo signals, where Y≥3 and Z≥6000; S12: performing clutter suppression and main component removal processing on each radar echo signal; S13: performing modulo-2 norm normalization processing on the amplitude of each radar echo signal after removing the main component; S14: In each category, combine every X normalized radar echo signals into a two-dimensional matrix as a training sample, X ≥ 3; S15: Add a corresponding category label to each of the training samples to obtain the training data set.

3. The method for robust radar target recognition based on multi-domain feature fusion under short dwell conditions according to claim 1 is characterized in that: The S21 includes: S211: Calculate the waveform entropy and amplitude variance of each time domain echo signal in each training sample according to the following formula: Where k = 1, 2, ..., P, P represents the number of points of the time domain echo signal, s i (k) represents the i-th time domain echo signal in a training sample, i = 1, 2, ..., X, X represents the number of echo signals contained in a training sample, a i Represents the time domain echo signal s i (k) waveform entropy, c i Represents the time domain echo signal s i (k) Amplitude variance; S212: According to the waveform entropy and amplitude variance of each time domain echo signal in each training sample, a time domain entropy vector A and a time domain variance vector C corresponding to each training sample are constructed: A=[a1,a2,...,a i ,...,a X ]; C=[c1,c2,...,c i ,...,c X ]; S213: According to the time domain entropy vector A and the time domain variance vector C corresponding to each training sample, a time domain entropy matrix S of each training sample is constructed. A And the time domain variance matrix S C : Wherein, L=fix(X / 2), which indicates the number of columns of the matrix, and fix(·) indicates rounding toward zero.

4. The method for robust radar target recognition based on multi-domain feature fusion under short dwell conditions according to claim 1 is characterized in that: The S24 includes: S241: Calculate the waveform entropy and amplitude variance of each Doppler domain echo signal in each training sample according to the following formula: Where k = 1, 2, ..., P, P represents the number of points of Doppler domain echo signal, S i (k) represents the i-th Doppler domain echo signal in a training sample, i = 1, 2, ..., X, X represents the number of echo signals contained in a training sample, bi Represents the Doppler domain echo signal S i (k) waveform entropy, di Represents the Doppler domain echo signal S i (k) Amplitude variance; S242: According to the waveform entropy and amplitude variance of each Doppler domain echo signal in each training sample, a Doppler domain entropy vector B and a Doppler domain variance vector D corresponding to each training sample are constructed: B=[b1,b2,...,b i ,...,b X ]; D=[d1,d2,...,d i ,...,d X ]; S243: Construct a Doppler domain entropy matrix S for each training sample based on the Doppler domain entropy vector B and the Doppler domain variance vector D corresponding to each training sample. B and the Doppler domain variance matrix S D : Wherein, L=fix(X / 2), which indicates the number of columns of the matrix, and fix(·) indicates rounding toward zero.

5. The method for robust radar target recognition based on multi-domain feature fusion under short dwell conditions according to claim 1 is characterized in that: The S26 includes: S261: Construct a feature vector f corresponding to each training sample based on the L-dimensional mean and variance features of the time domain entropy matrix of each training sample, the L-dimensional mean and variance features of the time domain variance matrix, the L-dimensional mean and variance features of the Doppler domain entropy matrix, and the L-dimensional mean and variance features of the Doppler domain variance matrix. n : f n =[μ A ,m B ,m C ,m D ,s A ,s B ,s C ,s D ]; Where n = 1, 2, ..., N, represents the sequence number of the training sample, N represents the total number of training samples, μ A represents the L-dimensional mean characteristic of the time domain entropy matrix, σ A represents the L-dimensional variance feature of the time domain entropy matrix, μ B represents the L-dimensional mean characteristic of the Doppler domain entropy matrix, σ B represents the L-dimensional variance characteristic of the Doppler domain entropy matrix, μ C represents the L-dimensional mean characteristic of the time domain variance matrix, σ C represents the L-dimensional variance feature of the time domain variance matrix, μ D represents the L-dimensional mean characteristic of the Doppler domain variance matrix, σ D Represents the L-dimensional variance characteristics of the Doppler domain variance matrix; S262: According to the feature vector f corresponding to each training sample n , construct the initial training feature matrix F train : F train =[f1;f2;...;f n ;...;f N ]; S263: Calculate the initial training feature matrix F train The mean and standard deviation of the eigenvalues ​​of each dimension in , and the mean vector μ and standard deviation vector σ are obtained: μ=[μ1,μ2,μ3,...,μ L×8 ]; σ=[σ1,σ2,σ3,...,σ L×8 ]; S264: Initial training feature matrix F train In the example, the feature vector f corresponding to each training sample is n Perform normalization processing to obtain the corresponding normalized results S265: Normalizing the training sample according to the result Get the training feature matrix

Images