A radar active jamming recognition method based on feature extraction and combined classifier

Abstract

The present application belongs to the technical field of radar anti-jamming, and particularly relates to a radar active jamming recognition method based on feature extraction and combined classifier, comprising the following steps: step 1: adopting a CWD transformation method to process the jamming signal, and obtaining a time-frequency image of the jamming signal; step 2: extracting the contour feature of the time-frequency image, and obtaining the global distribution information of the active jamming signal; step 3: extracting the complexity, box dimension and information dimension features of the active jamming signal according to the time-frequency image of the active jamming; step 4: obtaining the instantaneous phase information of the active jamming signal, and extracting the instantaneous frequency feature of the active jamming signal according to the instantaneous phase information of the active jamming signal; and step 5: realizing the classification and recognition of the radar active jamming signal through the combined classifier according to the extracted feature vector of the active jamming signal; the present application realizes the feature extraction and dimension reduction processing of the active jamming signal, has small calculation amount, low realization difficulty, and can classify and recognize multiple active jamming signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar anti-interference, and in particular relates to a radar active interference identification method based on feature extraction and combined classifier. Background Art

[0002] In recent years, electronic countermeasure technology has developed rapidly, making radar face new challenges and difficulties, which are mainly manifested in the following three aspects: First, the interference power continues to increase, and the interference frequency coverage continues to expand. In system warfare, networked jammers can greatly increase the interference power density; second, due to the widespread application of the DRFM system, various coherent deception interferences have come out one after another; finally, the continuous innovation of interference algorithm theory has led to the continuous emergence of new interference means, and the simultaneous use of comprehensive interference methods has increased the anti-interference pressure faced by radar.

[0003] For the problem of radar active jammer identification, the main algorithms involved are the maximum likelihood criterion-based jammer identification algorithm and the feature extraction-based jammer identification algorithm.

[0004] The maximum likelihood criterion interference identification algorithm based on operator experience is based on the prior information obtained, uses probability knowledge, establishes a statistical model, and combines the radar operator's observation of the radar display interface; however, the interference identification method based on the maximum likelihood criterion establishes an interference statistical model and uses statistical hypothesis testing theory to identify active interference, which has a large amount of calculation and is not easy to implement in engineering.

[0005] Interference recognition algorithms based on feature extraction analyze the differences in characteristics of real target echoes and various types of interference in the time domain, frequency domain, time-frequency domain, polarization domain, and transform domain to construct a feature parameter set and build a recognition network to identify interference. Feature extraction-based interference recognition algorithms mainly include decision trees and BP (back propagation) neural networks. The decision tree classification method allows for intuitive visualization of the classification process, making the judgment criteria easy to understand and computationally simple. However, as the number of classification categories increases, similar judgment criteria may appear in other categories, requiring the complete redesign of the decision tree. Furthermore, the decision tree method is significantly affected by the interference-to-noise ratio (the ratio of the signal in the system to the sum of the interference and noise). At low interference-to-noise ratios, certain judgment criteria become invalid, affecting the accuracy of the judgment. The BP neural network has a simple structure and can automatically train the network parameters based on the input features of the interference signal, avoiding the drawback of decision tree recognition algorithms that require manual design of judgment criteria during the calculation process. However, the existing BP neural network training requires relatively fixed feature parameters for the interference signal feature extraction in the early stages, resulting in a low recognition rate for the BP neural network.

[0006] Therefore, a radar active interference identification method is needed, which has low computational complexity, low implementation difficulty, and can classify and identify multiple active interference signals with a high recognition rate. Summary of the Invention

[0007] The purpose of the present invention is to overcome the defects of the prior art and provide a radar active jammer identification method based on feature extraction and combined classifier.

[0008] A radar active jammer identification method based on feature extraction and combined classification includes the following steps:

[0009] Step 1: Use the CWD transform method to process the interference signal and obtain the time-frequency image of the interference signal;

[0010] Step 2: extracting the contour features of the time-frequency image to obtain global distribution information of the active interference signal;

[0011] Step 3: extracting the complexity, box dimension, and information dimension characteristics of the active interference signal based on the time-frequency image to obtain a change pattern of the active interference signal;

[0012] Step 4: Obtain the instantaneous phase information of the active interference signal, and extract the instantaneous frequency characteristics of the active interference signal based on the instantaneous phase information of the active interference signal;

[0013] Step 5: Based on the feature vectors of the active interference signal extracted from steps 2 to 4, the radar active interference signal is classified and identified through a combined classifier.

[0014] Furthermore, in step 1, the CWD transformation method uses the following expression for transformation:

[0015]

[0016] Where CWD(t,ω) is the time-frequency distribution result, t is the time variable, ω is the angular frequency variable, α is the scale factor, τ is the delay factor, and u is the translation factor. The symbol * represents conjugation.

[0017] Furthermore, in step 2, the specific steps of extracting the contour features of the time-frequency image are as follows:

[0018] S21: Mapping the time-frequency image matrix into a grayscale image;

[0019] S22: Segmenting the grayscale image and extracting an edge image using a Canny operator;

[0020] S23: Applying a singular value decomposition method to the image matrix of the edge image obtained in S22 to obtain the time-frequency image contour features of the active interference signal.

[0021] Furthermore, in S21, the maximum value of the time-frequency distribution in the time-frequency image is mapped to grayscale 255, and other time-frequency distribution values ​​are mapped proportionally;

[0022] Before segmenting the grayscale image, the method further includes denoising the grayscale image; including:

[0023] Performing median filtering on the grayscale image, and / or removing a no-signal area in the grayscale image after the median filtering;

[0024] The method for removing the no-signal area in the grayscale image includes: detecting the grayscale image from the starting time point and the ending time point of the interference signal, respectively; when the amplitudes of all frequencies at a certain moment are less than a fixed threshold, it is considered that no interference signal exists at that moment; and removing data information corresponding to the moment when no interference signal exists from the grayscale image;

[0025] S22 further includes: optimizing edge information of the edge image using a mathematical morphology method.

[0026] Furthermore, in S23, the image matrix of the edge image is subjected to singular value decomposition, and 10-bit singular values ​​are used as feature parameters. Since the first value of the singular value is always 1 and has no recognition ability, the last 9-bit singular values ​​are selected as the contour feature recognition parameters of the time-frequency image of the active interference signal, which are recorded as {X2, X3, ... X 10}.

[0027] Furthermore, in step 3, the complexity, box dimension, and information dimension are used to describe the variation law of the active interference signal, and extract the characteristic vectors of the complexity, box dimension, and information dimension of the active interference signal, specifically including:

[0028] 1) Extracting the LZ complexity feature vector of the active interference signal X(n), n=1, 2, ..., N, including the following steps:

[0029] 1.1) At the beginning, add an artificial interference signal point x(1) to the given N-point time series interference signal X(n), n=1,2,...,N generation pool S, and suppose that the interference signal sequence contained in the generation pool S at time l is S={x(1),x(2),...x(l)},l<N, a string of length l, and at this time x(l) is the interference signal point added by the addition operation; let Q=x(l+1), and judge whether Q is a substring of the sequence SQv, where the sequence SQv is the string obtained by concatenating S and Q and removing the last character; if Q is a substring of the sequence SQv If the interference signal sequence S is a substring of the sequence SQv, the interference signal sequence S remains unchanged, and Q is updated to {x(l+1), x(l+2)}, and Q is judged again; if Q is not a substring of the sequence SQv, x(l+1) is added to the generation pool, that is, S={x(1), x(2), ...x(l+1)}, at this time Q=x(l+2), and Q is judged again; this cycle is repeated until all interference signal points are in the generation pool, and the processed interference signal sequence x(n) is obtained. The number of addition operations c(N) is counted, which is the LZ complexity. The expression of normalized LZ complexity C(N) is:

[0030]

[0031] 1.2) When the signal sequence is not the original signal sequence x(n), a high-quality interference signal sequence is obtained by reconstructing the N-point signal sequence using a multi-level quantization method. The average value of the N-point sequence x(n) is taken and the N-point sequence is judged. Sequences greater than the average value are set to 1, and sequences less than or equal to the average value are set to 0. This results in a 0-1 sequence to solve the LZ complexity of the high-quality interference signal.

[0032] 2) Extracting the characteristic vectors of the active interference signal box dimension and information dimension, specifically including the following steps:

[0033] S31: Reconstruct the N-point signal sequence x(n) to reduce some in-band noise. The reconstructed sequence of the N-point signal is x'(n)=x(n+1)-x(n), where n=1, 2, ..., N-1:

[0034] S32: re-quantize the amplitude of the reconstructed active interference signal, with the quantization level being M;

[0035] S33: Expand the length N-1 of the quantized active interference signal to the nearest 2 greater than N-1 M , and resample the active interference signal; at this time the active interference signal has 2 M points, and there are 2 M Level amplitude, at this time the interference signal time-frequency image is a 2 M 2 MEach point of the graph is equivalent to a pixel, and the pixel where the active interference signal point is located has a value;

[0036] S34: Let {A i}(i=1,2,...,N) is a valid δ-cover of set X. In this case, set A(i) is a valid δ-cover grid. Change the size of the grid from large to small, and record the grid size δ and the corresponding number of grids N(δ) with non-empty coverage on the active interference signal graph: with lnN(δ) as the ordinate and ln(1 / δ) as the abscissa, the box dimension is obtained from the slope of the double logarithmic plot;

[0037] Box dimension, used to describe the collective scale of fractal active interference signals;

[0038] Let (X, d) be a metric space, H be a non-empty compact set of X, and ε be a non-negative real number; let B(x, ε) denote a closed ball with center at x and radius ε;

[0039] Let A be a non-empty compact set in X. For each positive number ε, N(A,ε) represents the number of the smallest closed balls covering A, and the radius of the closed ball is ε, that is,

[0040]

[0041] Among them, x1,x2,...x M is a different point of X; if it exists:

[0042]

[0043] D b is the box dimension of set A; lnN(A,ε) is the logarithm of the minimum number of closed balls covering set A; ln(1 / ε) is the logarithm of the inverse of the closed ball radius.

[0044] S35: When the grid size is δ, number each grid and record the number of points in the active interference signal fractal set that fall into the i-th grid, and obtain P i and S i (δ) The grid changes from large to small, so as to obtain the information entropy under different grid coverage, expressed as S I (δ) is the ordinate, ln(1 / δ) is the abscissa, and the information dimension is obtained from the slope of the double logarithmic graph;

[0045] Information dimension, used to describe the distribution density of active interference signals in the fractal set in the regional space;

[0046] Assume {A i}(i=1,2,...,N) is a valid δ-cover of set X, P iIt represents the probability that an element in set X falls into set A(i), and its probability value is:

[0047]

[0048] Among them, N(X) i and N(X∩A i ) respectively represent the number of elements;

[0049] The information dimension of the fractal set is expressed as:

[0050]

[0051] Among them S I is the information entropy, expressed as:

[0052]

[0053] 4) The extracted feature vectors of LZ complexity, box dimension and information dimension of active interference signal are recorded as {C LZ ,D b ,D i}.

[0054] Furthermore, the multi-level quantization method specifically includes:

[0055] Assume that the number of quantization levels is M, the maximum amplitude in x(n) is a, and the minimum amplitude is b. Then the active interference signal sequence g(k) after M-level quantization is:

[0056]

[0057] Substitute the M-level quantized active interference signal sequence g(k) into formula (2) to obtain the LZ complexity of the corresponding signal.

[0058] Furthermore, in step 4, the instantaneous frequency feature vector of the active interference signal is extracted, which specifically includes:

[0059] S41: Obtain the instantaneous phase information of the active jamming signal. Assume that the radar receives the active jamming signal input as X(k), X I (k) and X Q (k) are the real and imaginary parts of the active jammer signal received by the radar, respectively. The instantaneous phase information of the active jammer signal received by the radar is expressed as:

[0060]

[0061] S42: Obtain the instantaneous phase sequence p(k), p(k) = φ(k) + u(k) (10)

[0062] u(k) is the phase correction sequence:

[0063]

[0064] Where u(0) = 0, and the instantaneous phase sequence p(k) is obtained;

[0065] S43: Calculate the Nth order difference of the instantaneous phase sequence p(k) obtained in S42 to obtain the instantaneous frequency sequence f N (k):

[0066]

[0067] S44: Autoregressive modeling of the instantaneous frequency of the active interference signal is performed using the second-order Yule-Walker method:

[0068]

[0069]

[0070] Where r x (k)=E(f i * (n)·f i (n+k)) is the autocorrelation function of the sequence, f i * (n) is the conjugate of the frequency of the interference signal, f i (n+k) is the n+kth point of the instantaneous frequency of the interference signal, P e is the estimated mean square error; a set of instantaneous frequency feature vectors are obtained through AR modeling as {a1, a2, P e}.

[0071] Furthermore, in step 5, the combined classifier includes an SVM main classifier for correctly separating the two types of interference signals and maximizing the classification interval of the two types of interference signals;

[0072] When the samples are linearly inseparable, the decision result expression used by the main classifier is:

[0073]

[0074] Among them, K(x i ,x j ) is the inner product function, b * are the bias of the optimal classification surface, α i y i is the kernel function.

[0075] Furthermore, in step 5, the combined classifier includes a secondary classifier for identifying two types of active interference signal feature vectors not distinguished by the main classifier, specifically including:

[0076] S51: obtain the average frequency f in the signal duration through the average frequency f in the signal duration t The expression of the average frequency f in the signal duration is obtained by the instantaneous frequency sequence of two types of active jamming signals t The expression of the average frequency f in the signal duration is:

[0077]

[0078] Wherein, L is the number of points of the instantaneous frequency sequence; Indicates the sum of L instantaneous frequency sequence values;

[0079] S52: respectively, the number of points greater than the average frequency of the instantaneous frequency of two types of active jamming signals is l1 and l2, defined as t=l / L is a characteristic vector;

[0080] S53: the obtained characteristic vector is compared with the threshold value t th The type decision is made, and the greater than the threshold value is the intermittent sampling and forwarding active jamming signal, and the less than the threshold value is the sampling pulse and forwarding active jamming signal.

[0081] Compared with the prior art, the beneficial effects of the present application are:

[0082] The present application firstly adopts the CWD conversion method to obtain the time-frequency two-dimensional graph of the active jamming signal; then extracts the time-frequency image features, complexity, box dimension, information dimension and instantaneous frequency features of the active jamming signal, obtains the good performance interference signal through the multi-level quantization method, discards part of the interference signal with poor performance, realizes the feature extraction and dimension reduction processing of the active jamming signal, and the operation amount is small and the realization difficulty is low; finally, the characteristic vector of the extracted active jamming signal is sent into the combined classifier with the SVM as the main classifier and the threshold decision as the secondary classifier, the classification and identification of various active jamming signals of the radar are completed, and the identification rate is high. BRIEF DESCRIPTION OF DRAWINGS

[0083] The following drawings only schematically illustrate and explain the present application, and do not limit the scope of the present application, wherein:

[0084] Figure 1 It is the two-dimensional graph feature extraction process of the present application;

[0085] Figure 2 It is the design block diagram of the combined classifier of the present application;

[0086] Figure 3 It is the active jamming signal identification flow chart of the present application;

[0087] Figure 4 It is the CWD time-frequency distribution graph of five types of typical active jamming signals of the present application;

[0088] Figure 5It is a diagram for solving the drag interference box dimension of the present invention;

[0089] Figure 6 It is a dimension diagram for solving drag interference information of the present invention;

[0090] Figure 7 It is the instantaneous frequency diagram of five kinds of radar active jammers of the present invention;

[0091] Figure 8 It is the instantaneous frequency characteristic diagram of intermittent sampling and sample pulse interference under different SNRs of the present invention;

[0092] Figure 9 This is a simulation of the recognition success rate of various types of active interference signals of the present invention. DETAILED DESCRIPTION

[0093] In order to make the purpose, technical solution, design method and advantages of the present invention more clear, the present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0094] like Figure 1 - Figure 9 As shown, a radar active jamming identification method based on feature extraction and combined classifier is used for radar active jamming signal identification, with small computational complexity and low implementation difficulty, thereby improving the recognition rate of radar active jamming signals.

[0095] Figure 3 This is a flow chart of radar jamming signal identification of the present invention. In general, the radar active jamming identification method includes the following steps:

[0096] Step 1: Use the CWD transform method to process the interference signal and obtain the time-frequency image of the interference signal;

[0097] Step 2: extracting the contour features of the time-frequency image to obtain global distribution information of the active interference signal;

[0098] Step 3: extracting the complexity, box dimension, and information dimension characteristics of the active interference signal based on the time-frequency image to obtain a change pattern of the active interference signal;

[0099] Step 4: Obtain the instantaneous phase information of the active interference signal, and extract the instantaneous frequency characteristics of the active interference signal based on the instantaneous phase information of the active interference signal;

[0100] Step 5: Classify and identify the radar active jamming signal through a combined classifier based on the feature vectors of the active jamming signal extracted from steps 2 to 4;

[0101] The specific implementation of this method is further described in detail below with reference to the accompanying drawings;

[0102] In step 1: the CWD transform method is used to obtain the time-frequency two-dimensional diagram of the active interference signal for the five radar interference signals. The CWD transform is used to provide a time-frequency description of the active interference signal. To better describe the active interference signal on the time-frequency distribution diagram, good time-frequency local aggregation is required. The Choi-Williams distribution (CWD) is used as the description method of the active interference signal. This transformation suppresses the cross terms while maximally ensuring that the energy of the active interference signal term is not lost.

[0103] The kernel function used in the Choi-Williams distribution focuses on suppressing cross terms far from the origin. This characteristic is suitable for most intra-pulse modulated active jammers and is therefore also suitable for active jammers modulated and forwarded by radar intra-pulse modulated active jammers, because cross terms do not appear near the origin.

[0104] The CWD transformation method is performed using the following expression:

[0105]

[0106] Where CWD(t,ω) is the time-frequency distribution result, t is the time variable, ω is the angular frequency variable, α is the scale factor, τ is the delay factor, and u is the translation factor; the symbol * represents conjugation.

[0107] Generally speaking, radar active jamming signals mainly include noise jamming, dense false targets, intermittent sampling and forwarding, sample pulses, and drag jamming. The time-frequency diagrams generated by performing CWD transformation on the five radar active jamming signals are as follows: Figure 4 As shown. Figure 4 As can be seen from the figure, the time-frequency diagrams of the five jamming signals are clearly different. The time-frequency diagram of the densely packed false targets shows no obvious pattern, while the sample pulses exhibit a linear variation. Using CWD time-frequency diagrams to identify jamming signals is theoretically feasible.

[0108] Based on the above time-frequency analysis results, the subsequent step 2 can be performed to extract the time-frequency image features, complexity features, and instantaneous frequency features of the active interference signal for transforming the active interference signal from the active interference signal space to the feature space.

[0109] The time-frequency two-dimensional graph of active interference signal is an effective basis for classification. Therefore, the time-frequency two-dimensional graph of active interference signal is regarded as an image, and a feature that can preserve the image information to the greatest extent is found. The singular value decomposition method (SVD) is used to complete the feature extraction. For the feature extraction process, see Figure 1 .

[0110] In step 2, the feature extraction process of the active interference signal time-frequency two-dimensional graph is as follows:

[0111] S21: Map the active interference signal time-frequency two-dimensional image matrix into a grayscale image using The calculation method maps the maximum value of the time-frequency distribution to a grayscale of 255, and other time-frequency distribution values ​​are mapped proportionally;

[0112] Before segmenting the grayscale image, the method further includes denoising the grayscale image; including:

[0113] Performing median filtering on the grayscale image, and / or removing a no-signal area in the grayscale image after the median filtering;

[0114] The method for removing the no-signal area in the grayscale image includes: detecting the grayscale image from the starting time point and the ending time point of the interference signal, respectively; when the amplitudes of all frequencies at a certain moment are less than a fixed threshold, it is considered that no interference signal exists at that moment; and removing data information corresponding to the moment when no interference signal exists from the grayscale image;

[0115] S22: Perform maximum entropy segmentation on the image and use the Canny operator to obtain an edge image. Image segmentation can basically eliminate the influence of noise, but the edge image obtained by segmentation may be discontinuous.

[0116] S22 further includes: applying a mathematical morphological method to the edge image, selecting a closing algorithm of first dilation and then erosion based on the discontinuous characteristics of the edge image, so as to achieve optimal representation of the edge information;

[0117] In S23, the image matrix of the edge image obtained in S22 is subjected to singular value decomposition, and the nine singular values ​​from the second to the tenth digit are used as the feature parameters for recognition. The number of singular values ​​is minimized while ensuring a small energy loss. The first digit of the singular value is always 1, which has no recognition capability. The last 9 digits are selected and recorded as {X2, X3, ... X 10};

[0118] In step 3, the complexity, box dimension and information dimension features of the active interference signal are extracted to describe the change characteristics of the active interference signal.

[0119] The calculation process of LZ complexity can be roughly described as: given an N-point time series X(n) as x(1), x(2), ... x(N), continuously adding active interference signal sampling points to the generation pool, observing whether a specific active interference signal segment can be found in the generation pool, and recording the number of active interference signal points added.

[0120] 1.1) Start, to the active interference signal generation pool added x (1), without loss of generality, assume that the active interference signal sequence S = {x (1), x (2),... x (l)} in the pool at some time, l < N, length of the string, and at this time x (l) is added by the operation, let Q = x (l+1), judge Q is the substring of sequence SQv (sequence SQv is the string obtained by concatenating S, Q and removing the last character). If so, S remains unchanged, and Q becomes {x (l+1), x (l+2)}, and the judgment is made again; if not, x (l+1) is added to the pool, that is, S = {x (1), x (2),... x (l+1)}, and Q = x (l+2) is judged again. Such a cycle until all the active interference signal points are in the pool, count the number of addition operations c (N), that is, the L-Z complexity. The normalized L-Z complexity C (N) is defined as:

[0121]

[0122] 1.2) When the active interference signal sequence is not the original active interference signal sequence x (n), reconstruct the N-point signal sequence;

[0123] Take the average value of N-point sequence x (n), and determine the N-point sequence. If it is greater than the average value, it is 1, otherwise it is 0, to obtain a 0-1 sequence to solve the L-Z complexity. This is equivalent to quantizing the series to two levels. This method will lose a lot of information in the original active interference signal, making the extracted complexity classification performance poor. A multi-level quantization method is used to extract the complexity feature vector of the active interference signal with good performance, and some feature vectors with poor performance are discarded to realize the dimension reduction processing of the interference signal feature vector;

[0124] Let the quantization number be M, the maximum amplitude of x (n) be a, and the minimum amplitude be b. The M-level quantized active interference signal sequence is:

[0125]

[0126] Box dimension and information dimension are used to describe the complexity and irregularity of active interference signals in fractal sets;

[0127] Let (X, d) be a metric space, H be a non-empty closed set family of X, and ε be a non-negative real number. Let B (x, ε) represent a closed ball with center x and radius ε. Let A be a non-empty closed set in X. For each positive number ε, let N (A, ε) represent the minimum number of closed balls with radius ε covering A, that is,

[0128]

[0129] where x1, x2,... xM are different points of X. If they exist:

[0130]

[0131] D b is the box dimension of set A; lnN(A,ε) is the logarithm of the minimum number of closed spheres covering set A; ln(1 / ε) is the logarithm of the reciprocal of the closed sphere radius. The box dimension reflects the scale of the active interference signal fractal set and is related to the box covering the fractal, but it cannot describe the density of the fractal set in the regional space. Another fractal dimension, the information dimension, is used to reflect the distribution of the fractal set.

[0132] Assume {A i}(i=1,2,...,N) is a valid δ-cover of set X, let P i It represents the probability that an element in set X falls into set A(i), and its value is:

[0133]

[0134] Where N(X) i and N(X∩A i ) represent the number of elements respectively. The information dimension of a fractal set is defined as:

[0135]

[0136] Among them S I is the information entropy, expressed as:

[0137]

[0138] When the probability of the points of the active interference signal fractal set falling into each grid is equal, that is, P i =1 / N(δ), the information dimension is equal to the box dimension.

[0139] The feature extraction process of active interference signal box dimension and information dimension is as follows:

[0140] S31: Reconstruct the N-point signal sequence x(n) to obtain x'(n)=x(n+1)-x(n), n=1, 2, ..., N-1, which can reduce the influence of some in-band noise;

[0141] S32: re-quantize the amplitude of the reconstructed active interference signal, with the quantization level being M;

[0142] S33: Expand the length N-1 of the quantized active interference signal to the nearest 2 greater than N-1 M , and resample the active interference signal; at this time the active interference signal has 2M points, and there are 2 M Level amplitude, for a 2 M 2 M Each point of the graph is equivalent to a pixel, and the pixel where the active interference signal point is located has a value;

[0143] S34: Let {A i}(i=1,2,...,N) is a valid δ-cover of set X. In this case, set A(i) is a valid δ-cover grid. The grid size is changed from large to small. The grid size δ and the corresponding number of grids N(δ) with non-empty coverage on the active interference signal graph are recorded. With lnN(δ) as the ordinate and ln(1 / δ) as the abscissa, the box dimension is obtained from the slope of the double logarithmic plot.

[0144] Figure 5 The ln(1 / δ)-lnN(δ) relationship diagram solved for the dragged interference signal with a signal-to-noise ratio of 20dB;

[0145] S35: When the grid size is δ, number each grid and record the number of points in the active interference signal fractal set that fall into the i-th grid, and obtain P i and S i (δ); the grid changes from large to small, so as to obtain the information entropy under different grid coverage; S I (δ) is the ordinate, ln(1 / δ) is the abscissa, and the information dimension is obtained from the slope of the double logarithmic graph. Figure 6 The S solved for the drag interference signal with a signal-to-noise ratio of 20dB is I (δ)-lnN(δ) relationship diagram. Thus, the active interference signal waveform complexity feature vector is obtained, which is composed of the LZ complexity, box dimension and information dimension of the active interference signal and can be recorded as {C LZ ,D b ,D i};

[0146] In step 4, the instantaneous frequency sequence is used to supplement the detailed information of the active interference signal in the time-frequency domain;

[0147] S41: First, assume that the radar receives active interference signal input as X(k), X I (k) and X Q (k) are its real and imaginary parts respectively, then its instantaneous phase can be expressed as:

[0148]

[0149] S42: When the phase value exceeds [0, 2π], calculating the phase sequence according to 2π will cause phase ambiguity; therefore, a phase deambiguation algorithm must be used for the phase sequence modulo 2π. Select the phase deambiguation algorithm: add the following phase correction sequence to the phase sequence modulo 2π:

[0150]

[0151] Where u(0) = 0, so the instantaneous phase sequence after deblurring is:

[0152] p(k)=φ(k)+u(k) (11);

[0153] The instantaneous frequency of the active interference signal can be obtained by differentiating the instantaneous phase. The differential operation of the continuous-time active interference signal is equivalent to the differential operation of the discrete-time active interference signal, and the first-order differential operation is generally used. In order to fully utilize the phase jump information contained in the phase and improve the processing capability of the low signal-to-noise ratio active interference signal, the high-order differential phase method is used to obtain the instantaneous frequency sequence f by taking the N-order differential of the defuzzified instantaneous phase sequence p(k). N (k):

[0154]

[0155] The instantaneous frequency of the active interference signal is modeled by autoregression, and the second-order Yule-Walker equation is as follows:

[0156]

[0157] From formula (12), we can deduce:

[0158]

[0159] Among them, r x (k)=E(f i * (n)·f i (n+k)) is the autocorrelation function of the sequence, f i * (n) is the conjugate frequency of the interference signal, f i (n+k) is the n+kth point of the instantaneous frequency of the interference signal, P e is the estimated mean square error; a set of instantaneous frequency feature vectors are obtained through AR modeling as {a1, a2, P e}.

[0160] Figure 7 is the instantaneous frequency diagram of five typical radar active jammers. So far, the characteristic vector set of active jammer signals {X2,X3,...X 10 ,CLZ ,D b ,D i ,a1,a2,P e};

[0161] In step 5, the radar active jamming signal is identified by a combined classifier based on the feature vector of the instantaneous frequency feature of the active jamming signal; the combined classifier includes a main classifier and a slave classifier, the SVM is the main classifier, and the threshold decision device is the slave classifier; the feature vectors of the time-frequency image feature, the complexity feature, and the instantaneous frequency feature are sent to the hierarchical classifier of the main classifier and the slave classifier for identification;

[0162] The master-slave classifier is used to classify and identify the input interference signal feature vector. The master classifier selects the SVM classifier, and the slave classifier selects the threshold-based decision maker. Through a large number of simulation experiments, the front-end classifier, namely the SVM, has been obtained to identify various types of signals. The identification of intermittent sampling forwarding and sample pulse interference overlaps. When these two types of signals are sent to the slave classifier, the signal types that were originally poorly identified in the master classifier (SVM) can be well identified, and the results of stable and accurate classification in the master classifier are retained. The implementation block diagram is shown in Figure 2 ;

[0163] SVM is a specific implementation of the structural risk minimization criterion. The optimal hyperplane it constructs can not only correctly separate the two types of samples, but also maximize the classification interval between the two types of samples. When the sample points are separable, the following decision function is used to classify the sample points:

[0164] f(x)=sign(ω * Φ(x)+b * ) (15);

[0165] Among them, Φ(x) is the kernel function, ω * and b * are the weight vector and bias of the optimal classification surface respectively. When the samples are linearly inseparable, they can be transformed into a linear problem in a high-dimensional space through nonlinear transformation, and the optimal classification surface can be found in the transformed space. In the high-dimensional space, only the inner product operation is actually required, and this inner product operation can be realized using the function in the original space. According to the relevant theories of functional analysis, as long as a kernel function K(x i ,x j ) satisfies the Mercer condition, which corresponds to the inner product in a certain transformation space. In the optimal classification surface, an appropriate inner product function K(x i ,x j ) can realize linear classification after a nonlinear transformation without increasing the computational complexity. At this time, the decision function used becomes the inner product function K(x i ,xj ) Transformed linear combination:

[0166]

[0167] For the sample pulse forwarding interference signal, since it is generated by sampling and repeatedly forwarding a short segment of the signal before the radar transmits the signal, the instantaneous frequency of the first short segment of the interference signal is relatively large, and the instantaneous frequency of the forwarded segment is the same as the instantaneous frequency of the sampled segment. For most of the time in the instantaneous frequency sequence, the instantaneous frequency is lower than the average frequency value. For the intermittent sampling forwarding interference signal, since it is generated by sampling and forwarding, then sampling and forwarding, the instantaneous frequency of the interference signal continues to increase, and there is a frequency mutation point. The signal instantaneous frequency sequence is composed of several different values, and the center of gravity of the instantaneous frequency sequence is close to the average frequency value. The average frequency f during the signal duration t It can be found using the following formula:

[0168]

[0169] Where L is the number of points in the instantaneous frequency sequence; Represents the sum of L instantaneous frequency sequence values;

[0170] For intermittent sampling and forwarding interference signals, f t Roughly equal to the center of gravity of the instantaneous frequency sequence value, that is, the instantaneous frequency sequence value is greater than f t The number of points is less than f t The number of points is roughly the same. For the sample pulse forwarding interference signal, it is greater than f t The number of points is less than the instantaneous frequency value f t The points are much less.

[0171] The combined classifier includes a secondary classifier for identifying two types of active interference signal feature vectors not distinguished by the main classifier, specifically including:

[0172] S51: Obtain the average frequency of the instantaneous frequency sequences of the two types of active interference signals;

[0173] S52: Count the number of points in the instantaneous frequencies of the two types of active interference signals that are greater than their respective average frequencies, which are l1 and l2, and define t=l / L as the characteristic vector;

[0174] S53: Compare the obtained eigenvector with the threshold value t th The type of interference signal is determined by comparison. The signal with a value greater than the threshold is an intermittent sampling forwarding interference signal, and the signal with a value less than the threshold is a sample pulse forwarding interference signal.

[0175] Threshold t thDetermined by simulation. Here, for two types of interference with variable parameters, the SNR is evenly distributed between 5 and 15 dB. 1000 Monte Carlo simulations are performed to produce a scatter plot of the characteristic values ​​of the two types of interference signals when the SNR changes, as shown in the figure. Figure 8 As shown in the figure, it can be found that in the SNR range of 5 to 15dB, the two types of interference can be well distinguished by the characteristic values ​​of the two types of interference signals, and the threshold t th Select 0.4. When the SNR is close to 2dB, the characteristic values ​​of the two types of interference in the figure are not much different, and both are concentrated near the threshold. This shows that this classifier can work stably in an SNR environment above 2dB. Below this SNR value, it may cause a certain degree of misjudgment.

[0176] Example 1: The signal sampling rate is 60MHz, and the signal carrier frequency f c The frequency range varied uniformly from 3MHz to 8MHz, the distance drag speed varied uniformly from 600m / s to 900m / s, and the speed drag speed varied uniformly from 40kHz / s to 60kHz / s. The duration of a single dense false target varied uniformly from 0.5μs to 2.5μs, and the number of false targets was 32. The sample pulse sampling duration varied uniformly from 0.5μs to 2.5μs, and the intermittent sampling forwarding interference sampling time and forwarding time both varied uniformly from 0.5μs to 2.5μs. Noise interference selected narrowband interference mode and broadband interference mode, and the noise bandwidth varied uniformly from 10MHz to 100MHz. The interference signal-to-noise ratio was 5dB. The simulation was performed 240 times, of which the signal feature vectors from 40 experiments were used as training samples for the main classifier, and the results of the remaining 200 experiments were used to test the recognition success rate of the algorithm.

[0177] The midpoint of the signal-to-noise ratio is set to 15dB, and the signal-to-noise ratio variation ranges are 0dB, 5dB, 10dB and 15dB, which correspond to the SNR of the signal in the simulation being evenly distributed between 15dB, 12.5-17.5dB, 10-20dB and 7.5-22.5dB respectively. Figure 9 The recognition success rate of various types of signals shows that the recognition success rate for active interference types such as noise, dense false targets, intermittent sampling and forwarding, sample pulses, and dragging is between 90% and 98%.

[0178] In summary, the present application aims at identifying active interference types such as noise, dense false targets, intermittent sampling and forwarding, sample pulses, and dragging. The radar active interference signal is classified and identified through active interference signal feature extraction and master-slave classifiers. The active interference signal is first transformed from the active interference signal space to the feature space, which is called feature extraction. The three signal features of the interference signal, namely time-frequency image features, complexity features, and instantaneous frequency features, are extracted. The N-point signal sequence is reconstructed by a multi-level quantization method to obtain a good interference signal. Some feature vectors with poor performance are discarded to achieve dimensionality reduction of the interference signal feature vector, which has a small amount of computation. Then, the active interference signal is mapped from the feature space to the decision space through a certain criterion. This is achieved through a classifier and sent to a hierarchical classifier with SVM as the main classifier and a combined classifier as the slave classifier. The trained classifier is used to classify and identify the radar active interference signal. The recognition rate of various active interference signals is high. The recognition success rate of active interference types such as noise, dense false targets, intermittent sampling and forwarding, sample pulses, and dragging is 90%-98%.

[0179] The method for identifying typical radar active interference based on feature extraction and combined classifier design is described.

[0180] First, a CWD transform method is used to obtain a two-dimensional time-frequency image of the active interference signal. This image is then processed using an image contour extraction method, and the singular value decomposition method is used to extract the time-frequency image features and obtain the global distribution information of the active interference signal. The complexity, box dimension, and information dimension features of the active interference signal are then extracted to obtain the variation pattern of the active interference signal. Furthermore, the instantaneous frequency features are extracted to obtain the subtle features of the active interference signal. Finally, the feature vector is fed into a combined classifier with an SVM as the primary classifier and a threshold decision detector as the secondary classifier to output the active interference signal type identification result. Based on existing technologies, the classifier maps the active interference signal from the feature space to the decision space. When the sample points are separable, a decision function is used to classify the sample points. When the samples are linearly inseparable, a nonlinear transformation is used to transform them into a linear problem in a high-dimensional space, and the optimal classification surface is found in the transformed space. This method has good inter-class dispersion and intra-class aggregation, and can classify and identify a variety of active interference signals with high recognition rates. It is applicable to a wide range of radar platforms and interference types, and has high technological advancement and application adaptability.

[0181] While various embodiments of the present invention have been described above, the above descriptions are intended to be illustrative, non-exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or technological improvements in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A radar active jammer identification method based on feature extraction and combined classification, characterized by: The steps include: Step 1: Use the CWD transform method to process the interference signal and obtain the time-frequency image of the interference signal; Step 2: extracting the contour features of the time-frequency image to obtain global distribution information of the active interference signal; Step 3: extracting the complexity, box dimension, and information dimension characteristics of the active interference signal based on the time-frequency image to obtain the variation pattern of the active interference signal; specifically comprising: 1) Extracting the LZ complexity feature vector of the active interference signal X(n), n=1, 2, ..., N, including the following steps: 1.1) At the beginning, add an artificial interference signal point x(1) to the given N-point time series interference signal X(n), n=1,2,...,N generation pool S, and suppose that the interference signal sequence contained in the generation pool S at time l is S={x(1),x(2),...x(l)},l<N, a string of length l, and at this time x(l) is the interference signal point added by the addition operation; let Q=x(l+1), and judge whether Q is a substring of the sequence SQv, where the sequence SQv is the string obtained by concatenating S and Q and removing the last character; if Q is a substring of the sequence SQv If the interference signal sequence S is a substring of the sequence SQv, the interference signal sequence S remains unchanged, and Q is updated to {x(l+1), x(l+2)}, and Q is judged again; if Q is not a substring of the sequence SQv, x(l+1) is added to the generation pool, that is, S={x(1), x(2), ...x(l+1)}, at this time Q=x(l+2), and Q is judged again; this cycle is repeated until all interference signal points are in the generation pool, and the processed interference signal sequence x(n) is obtained. The number of addition operations c(N) is counted, which is the LZ complexity. The expression of normalized LZ complexity C(N) is: 1.2) When the signal sequence is not the original signal sequence x(n), a high-quality interference signal sequence is obtained by reconstructing the N-point signal sequence using a multi-level quantization method. The average value of the N-point sequence x(n) is taken and the N-point sequence is judged. Sequences greater than the average value are set to 1, and sequences less than or equal to the average value are set to 0. This results in a 0-1 sequence to solve the LZ complexity of the high-quality interference signal. 2) Extracting the characteristic vectors of the active interference signal box dimension and information dimension, specifically including the following steps: S31: Reconstruct the N-point signal sequence x(n) to reduce some in-band noise. The reconstructed sequence of the N-point signal is x'(n)=x(n+1)-x(n), where n=1, 2, ..., N-1: S32: re-quantize the amplitude of the reconstructed active interference signal, with the quantization level being M; S33: Expand the length N-1 of the quantized active interference signal to the nearest 2 greater than N-1 M , and resample the active interference signal; at this time the active interference signal has 2 M points, and there are 2 M Level amplitude, at this time the interference signal time-frequency image is a 2 M 2 M Each point of the graph is equivalent to a pixel, and the pixel where the active interference signal point is located has a value; S34: Let {A i }(i=1,2,...,N) is a valid δ-cover of set X. In this case, set A(i) is a valid δ-cover grid. Change the size of the grid from large to small, and record the grid size δ and the corresponding number of grids N(δ) with non-empty coverage on the active interference signal graph: with lnN(δ) as the ordinate and ln(1 / δ) as the abscissa, the box dimension is obtained from the slope of the double logarithmic plot; S35: When the grid size is δ, number each grid and record the number of points in the active interference signal fractal set that fall into the i-th grid, and obtain P i and S i (δ) The grid changes from large to small, so as to obtain the information entropy under different grid coverage, expressed as S I (δ) is the ordinate, ln(1 / δ) is the abscissa, and the information dimension is obtained from the slope of the double logarithmic graph; 3) The extracted feature vectors of LZ complexity, box dimension and information dimension of active interference signal are recorded as {C LZ ,D b ,D i }; Step 4: Obtain the instantaneous phase information of the active interference signal, and extract the instantaneous frequency characteristics of the active interference signal based on the instantaneous phase information of the active interference signal; Step 5: Based on the feature vectors of the active interference signal extracted from steps 2 to 4, the radar active interference signal is classified and identified through a combined classifier.

2. The radar active jammer identification method based on feature extraction and combined classification according to claim 1, characterized in that: In step 1, the CWD transformation method uses the following expression for transformation: Where CWD(t,ω) is the time-frequency distribution result, t is the time variable, ω is the angular frequency variable, α is the scale factor, τ is the delay factor, and u is the translation factor. The symbol * represents conjugation.

3. The radar active jammer identification method based on feature extraction and combined classification according to claim 1, characterized in that: In step 2, the specific steps of extracting the contour features of the time-frequency image are: S21: Mapping the time-frequency image matrix into a grayscale image; S22: Segmenting the grayscale image and extracting an edge image using a Canny operator; S23: Applying a singular value decomposition method to the image matrix of the edge image obtained in S22 to obtain the time-frequency image contour features of the active interference signal.

4. The radar active jammer identification method based on feature extraction and combined classification according to claim 3 is characterized in that: In S21, the maximum value of the time-frequency distribution in the time-frequency image is mapped to grayscale 255, and other time-frequency distribution values ​​are mapped proportionally; Before segmenting the grayscale image, the method further includes denoising the grayscale image; including: Performing median filtering on the grayscale image, and / or removing a no-signal area in the grayscale image after the median filtering; The method for removing the no-signal area in the grayscale image includes: detecting the grayscale image from the starting time point and the ending time point of the interference signal, respectively; when the amplitudes of all frequencies at a certain moment are less than a fixed threshold, it is considered that no interference signal exists at that moment; and removing data information corresponding to the moment when no interference signal exists from the grayscale image; S22 further includes: optimizing edge information of the edge image using a mathematical morphology method.

5. The radar active jammer identification method based on feature extraction and combined classification according to claim 3 is characterized in that: In S23, the image matrix of the edge image is subjected to singular value decomposition, and 10-bit singular values ​​are used as feature parameters. Since the first value of the singular value is always 1 and has no recognition ability, the last 9 singular values ​​are selected as the contour feature recognition parameters of the time-frequency image of the active interference signal, which are recorded as {X2, X3, ... X 10 }.

6. The radar active jammer identification method based on feature extraction and combined classification according to claim 1, characterized in that: In step 34: the box dimension is used to describe the collective scale of the fractal active interference signal; Let (X, d) be a metric space, H be a non-empty compact set of X, and ε be a non-negative real number; let B(x, ε) denote a closed ball with center at x and radius ε; Let A be a non-empty compact set in X. For each positive number ε, N(A,ε) represents the number of the smallest closed balls covering A, and the radius of the closed ball is ε, that is, Among them, x1,x2,...x M is a different point of X; if it exists: D b is the box dimension of set A; lnN(A,ε) is the logarithm of the minimum number of closed balls covering set A; ln(1 / ε) is the logarithm of the inverse of the closed ball radius; In step 35: the information dimension is used to describe the distribution density of the active interference signal in the fractal set in the regional space; Assume {A i }(i=1,2,...,N) is a valid δ-cover of set X, P i It represents the probability that an element in set X falls into set A(i), and its probability value is: Among them, N(X) i and N(X∩A i ) respectively represent the number of elements; The information dimension of the fractal set is expressed as: Among them S I is the information entropy, expressed as:

7. The radar active jammer identification method based on feature extraction and combined classification according to claim 6, characterized in that: The multi-level quantization method specifically includes: Assume that the number of quantization levels is M, the maximum amplitude in x(n) is a, and the minimum amplitude is b. Then the active interference signal sequence g(k) after M-level quantization is: Substitute the M-level quantized active interference signal sequence g(k) into formula (2) to obtain the LZ complexity of the corresponding signal.

8. The radar active jammer identification method based on feature extraction and combined classification according to claim 1, characterized in that: In step 4, the instantaneous frequency feature vector of the active interference signal is extracted, which specifically includes: S41: Obtain the instantaneous phase information of the active jamming signal. Assume that the radar receives the active jamming signal input as X(k), X I (k) and X Q (k) are the real and imaginary parts of the active jammer signal received by the radar, respectively. The instantaneous phase information of the active jammer signal received by the radar is expressed as: S42: Obtain the instantaneous phase sequence p(k), p(k) = φ(k) + u(k) (10); u(k) is the phase correction sequence: Where u(0) = 0, and the instantaneous phase sequence p(k) is obtained; S43: Calculate the Nth order difference of the instantaneous phase sequence p(k) obtained in S42 to obtain the instantaneous frequency sequence f N (k): S44: Autoregressive modeling of the instantaneous frequency of the active interference signal is performed using the second-order Yule-Walker method: Among them, r x (k)=E(f i * (n)·f i (n+k)) is the autocorrelation function of the sequence, f i * (n) is the conjugate frequency of the interference signal, f i (n+k) is the n+kth point of the instantaneous frequency of the interference signal, P e is the estimated mean square error; a set of instantaneous frequency feature vectors are obtained through AR modeling as {a1, a2, P e }.

9. The radar active jammer identification method based on feature extraction and combined classification according to claim 1, characterized in that: In step 5, the combined classifier includes an SVM main classifier, which is used to correctly separate the two types of interference signals and maximize the classification interval of the two types of interference signals; When the samples are linearly inseparable, the decision result expression used by the main classifier is: Among them, K(x i ,x j ) is the inner product function, b * are the bias of the optimal classification surface, α i y i is the kernel function.

10. The radar active jammer identification method based on feature extraction and combined classification according to claim 9, characterized in that: In step 5, the combined classifier includes a secondary classifier for identifying two types of active interference signal feature vectors not distinguished by the main classifier, specifically including: S51: average frequency f during the duration of the passing signal t The expression of the instantaneous frequency sequence of the two types of active interference signals is used to obtain the average frequency of the signal duration, the average frequency f t The expression is: Where, L is the number of points in the instantaneous frequency sequence; Represents the sum of L instantaneous frequency sequence values; S52: Count the number of points in the instantaneous frequencies of the two types of active interference signals that are greater than their respective average frequencies, which are l1 and l2, and define t=l / L as the characteristic vector; S53: Compare the obtained eigenvector with the threshold value t th The type of interference signal is determined by comparison. If the value is greater than the threshold, it is an intermittent sampling forwarding active interference signal; if the value is less than the threshold, it is a sample pulse forwarding active interference signal.

Images