Radar signal recognition based on time-frequency image feature fusion

Abstract

The application discloses a radar signal recognition method based on time-frequency image feature fusion, and mainly solves the problem of low radar signal recognition rate in the prior art under a low signal-to-noise ratio environment.The steps of the application are as follows: (1) time-frequency transformation is performed on a radar signal to obtain a time-frequency image; (2) time-frequency image pre-processing; (3) texture feature and shape feature extraction of the image; (4) support vector machine is used to recognize the radar signal.The application extracts the texture feature and shape feature of the radar signal based on the Choi-Williams time-frequency image and the contour map of the fuzzy function, and fuses the two to form a new radar signal recognition characteristic value, thereby solving the problem of low radar signal recognition rate under a low signal-to-noise ratio.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, and more specifically relates to a radar signal recognition method based on time-frequency image feature fusion in the field of signal recognition technology. Background Technology

[0002] Radar signal identification is a crucial component of electronic reconnaissance systems. Its purpose is to extract characteristic parameters from sorted radar signals and automatically identify radar radiation source signals. With the increasing complexity of the electromagnetic environment on modern battlefields and the widespread application of new radar systems, rapid and effective radar signal identification has become paramount. Traditional methods based on the five conventional characteristic parameters (carrier frequency, pulse width, pulse amplitude, time of arrival, and angle of arrival) are no longer effective. Faced with radar radiation source signals exhibiting diverse modulation schemes, effective identification requires extracting the signal's essential features. Intra-pulse characteristic parameters of radar signals can effectively reflect the signal's essential information and expand the parameter space for signal identification; therefore, researching intra-pulse characteristic parameters of radar signals has become a current hot topic. Currently, existing methods offer numerous and highly accurate methods for identifying signal modulation types under high signal-to-noise ratio (SNR) conditions. However, under low SNR conditions, the signal identification rate is low and the number of methods is limited. Furthermore, methods for identifying composite modulation types are scarce, resulting in low identification rates and poor algorithm robustness.

[0003] In existing technical solutions, fuzzy functions are an important mathematical tool for studying radar signals. They can be used not only to study different radar waveforms but also to characterize the differences between different radar signals. Time-frequency analysis is an important method for processing non-stationary signals. Common time-frequency analysis methods include Short-Time Fourier Transform (STFT), Wigner-Ville Distribution, and CWD. Among these, STFT suffers from poor time-frequency clustering, WVD generates cross-terms that interfere with the true signal characteristics, and CWD time-frequency analysis has better time-frequency resolution and can effectively suppress cross-terms. To address the shortcomings of existing solutions, this invention utilizes the advantages of both fuzzy functions and CWD time-frequency analysis, employing a low-complexity SVM classifier to design a radar signal recognition method that achieves more accurate radar signal recognition even at low signal-to-noise ratios. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings and deficiencies of existing technologies by designing a radar signal recognition method based on time-frequency image feature fusion, aiming to solve the problem of low recognition rate of low signal-to-noise ratio radar signals.

[0005] The objective of this invention is achieved by the following steps:

[0006] (1) The CWD time-frequency image of the radar signal is obtained by using Choi-Williams time-frequency transformation and the three-dimensional image of the signal is made by using fuzzy function. Then the corresponding fuzzy function contour map is obtained.

[0007] (2) Preprocessing of the CWD time-frequency image and the fuzzy function contour image obtained in step (1), including:

[0008] (2-1) Grayscale conversion yields a grayscale image;

[0009] (2-2) Filter the grayscale image to obtain a filtered image, for example, by using Wiener filtering;

[0010] (2-3) Scaling the filtered image to a set size using bicubic interpolation, such as 224*224;

[0011] (3) Extract the feature values ​​of the signal, including:

[0012] Texture features of CWD time-frequency images and fuzzy function contour maps are extracted using the gray-level gradient co-occurrence matrix; shape features of CWD time-frequency images and fuzzy function contour maps are extracted using pseudo-Zernike moments.

[0013] (4) Fuse image texture features and shape features;

[0014] (5) Use SVM for recognition processing:

[0015] Add labels to the feature vectors corresponding to (5-1);

[0016] (5-2) The values ​​of the penalty factor C and kernel function parameter g of SVM are determined by the grid parameter optimization method;

[0017] (5-3) Identify the received signal and output the result.

[0018] The mathematical expression for the Choi-Williams Distribution in step (1) is:

[0019]

[0020] Where t represents time, s(t) represents radar signal, * represents complex conjugate operation, e represents exponential operation with natural constant as base, j represents imaginary unit sign, and σ represents attenuation coefficient, which is proportional to the amplitude of cross phase.

[0021] The mathematical expression for the fuzzy function in step (1) is:

[0022]

[0023] Where τ represents the time delay, t represents the time, and fd denoted by Doppler frequency offset, * denotes complex conjugate operation, e denotes exponential operation with the natural constant as the base, and j denotes the imaginary unit symbol.

[0024] The gray-level gradient co-occurrence matrix (GLGCM) in step (3) is as follows: Let the gray-level image be f(M,N), where M and N are the number of rows and columns of the two-dimensional matrix corresponding to the gray-level image, respectively; calculate the corresponding GLGCM through the following steps:

[0025] (1) Calculate the normalized gradient matrix of f(M,N). Use the Sobel operator with a 3*3 window to extract the gradient matrix g(M,N) of f(M,N). The gradient value of the (k,l)th pixel is calculated using the following formula:

[0026]

[0027] g x =f(k+1,l-1)+2f(k+1,l)+f(k+1,l+1)-f(k-1,l-1)

[0028] -2f(k-1,l)-f(k-1,l+1)

[0029] g y =f(k-1,l+1)+2f(k,l+1)+f(k+1,l+1)-f(k-1,l+1)

[0030] -2f(k,l-1)-f(k+1,l-1)

[0031] Where k = 1, 2, ..., M; l = 1, 2, ..., N. The normalized gradient matrix is ​​obtained using the formula... Where INT represents the integer operation; g max It is the largest gradient value in g(M,N); N g It is the maximum expected value after gradient normalization; the preferred value is 32.

[0032] (2) Calculate the normalized gray matrix of f(M,N); the formula is: Where f max It is the maximum gray value in f(M,N); N f It is the expected maximum value after normalizing the grayscale values; N f The preferred value is 32.

[0033] (3) The element values ​​H(i,j) of the gray-level gradient co-occurrence matrix, that is, the elements in the statistically normalized gray-level matrix and gradient matrix that simultaneously make F(k,l)=i (i∈[1,N) f ]) and G(k,l)=j(j∈[1,N)g The normalized GLGCM is obtained by using the following formula to calculate the number of pixels in the given image.

[0034]

[0035] The pseudo-Zernike moment in step (3) is an orthogonal complex moment. The definition of a pseudo-Zernike moment with order p and repetition degree q is:

[0036]

[0037] x 2 +y 2 =1

[0038] In the formula, p is a positive integer or zero, q is an integer, and |q|≤p, f(x,y) is a graphical function; where,

[0039] In polar coordinates, it is represented as

[0040] To ensure rotational invariance and reduce the dynamic range of pseudo-Zernike moments, we take...

[0041] Extracting texture and shape features from an image, specifically using GLGCM to extract the image's texture features T. CWD T AFCL The shape feature Z of the image is extracted using pseudo-Zernike moments. CWD Z AFCL These two feature parameters are combined to form a feature vector [T] CWD ,T AFCL Z CWD Z AFCL ];

[0042] The key advantages of this invention are: 1. It achieves effective identification of radar signals with eight different modulation types based on time-frequency image feature fusion, including frequency modulation, phase modulation, and composite modulation signals that combine both; 2. The average correct identification rate of the eight signals can reach over 80% in environments with a signal-to-noise ratio as low as -8dB; 3. The radar signal parameters used in this invention are all selected within a dynamic range, thus they are not sensitive to parameter changes and have good generalization ability, which can meet the needs of practical use. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the radar signal recognition method based on time-frequency image feature fusion involved in the present invention.

[0044] Figure 2This is a diagram of the SVM training model used to identify eight types of radar signals in the implementation of this invention.

[0045] Figure 3 This is a graph showing the change in the recognition rate of each of the eight radar signals identified in the implementation of this invention.

[0046] Figure 4 This is a graph showing the overall correct recognition rate of eight radar signals as a function of signal-to-noise ratio in the implementation of this invention.

[0047] Figure 5 This is a confusion matrix diagram of eight radar signals in the implementation of this invention. Detailed Implementation

[0048] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0049] like Figure 1 As shown, the radar signal recognition method based on time-frequency image feature fusion of the present invention mainly includes the following steps: first, performing time-frequency transformation on the radar signal to obtain time-frequency images of signals with different modulation types; then, preprocessing the images to suppress the influence of noise and reduce the amount of computation; next, extracting the shape features and texture features of the images and fusing them; and finally, identifying the intra-pulse modulation mode of the radar signal through SVM and outputting it.

[0050] The specific procedures for processing radar signals are as follows:

[0051] Radar signal time-frequency transformation includes two types: CWD time-frequency transformation and ambiguity function. The mathematical expression for CWD time-frequency transformation is as follows:

[0052]

[0053] Where t represents time, s(t) represents radar signal, * represents complex conjugate operation, e represents exponential operation with natural constant as base, j represents imaginary unit sign, and σ represents attenuation coefficient, which is proportional to the amplitude of cross phase.

[0054] The fuzzy function is another time-frequency distribution function and an important mathematical tool for studying radar signals. It can be used not only to study different radar waveforms but also to characterize the differences between various radar signals. The mathematical expression for the fuzzy function is:

[0055]

[0056] Where τ represents the time delay, t represents the time, and f d denoted by Doppler frequency offset, * denotes complex conjugate operation, e denotes exponential operation with the natural constant as the base, and j denotes the imaginary unit symbol.

[0057] Due to the influence of noise and time-frequency cross terms, the signal time-frequency graph contains a large amount of interference information. Therefore, before performing image feature extraction, image processing techniques are used to preprocess the time-frequency graph, which can effectively reduce interference and redundant information and enhance the effectiveness of feature extraction. The preprocessing process first converts the time-frequency graph to grayscale to obtain a grayscale image, then performs Wiener filtering on the grayscale image to obtain a filtered image, and finally scales the filtered image to 224*224 using bicubic interpolation.

[0058] This invention uses GLGCM and pseudo-Zernike moments to extract the texture and shape features of radar signals, respectively.

[0059] 1. The steps for texture feature extraction are as follows:

[0060] (1) Calculate the normalized gradient matrix of f(M,N). The gradient matrix g(M,N) of f(M,N) is extracted using the Sobel operator with a 3*3 window. The gradient value of the (k,l)th pixel is calculated using the following formula:

[0061]

[0062] g x =f(k+1,l-1)+2f(k+1,l)+f(k+1,l+1)-f(k-1,l-1)

[0063] -2f(k-1,l)-f(k-1,l+1)

[0064] g y =f(k-1,l+1)+2f(k,l+1)+f(k+1,l+1)-f(k-1,l+1)

[0065] -2f(k,l-1)-f(k+1,l-1)

[0066] The normalized gradient matrix is ​​obtained using the formula Where k =

[0067] 1,2…,M; l = 1,2…,N; INT is the integer operation; g max It is the largest gradient value in g(M,N);

[0068] N g It is the expected maximum value after gradient normalization, and in this embodiment, the value is 32.

[0069] (2) Calculate the normalized gray matrix of f(M,N). The formula is as follows: Where f max It is the maximum gray value in f(M,N); N f It is the expected maximum value after normalizing the grayscale values. In this embodiment, N is taken as N.f It is 32.

[0070] (3) The element values ​​H(i,j) of the gray-level gradient co-occurrence matrix, that is, the elements in the statistically normalized gray-level matrix and gradient matrix that simultaneously make F(k,l)=i (i∈[1,N) f ]) and G(k,l)=j(j∈[1,N) g The number of pixels in the array. The normalized GLGCM is obtained using the following formula:

[0071]

[0072] (4) In this embodiment, 15 feature parameters of GLGCM were selected, specifically including small gradient dominance (T1), large gradient dominance (T2), gradient distribution non-uniformity (T3), gray-level distribution non-uniformity (T4), energy (T5), gray-level mean (T6), gradient mean (T7), gray-level standard deviation (T8), gradient standard deviation (T9), and correlation (T1). 10 ), grayscale entropy (T) 11 Gradient entropy (T) 12 ), Mixed entropy (T) 13 ), difference moment (T) 14 Inverse moment (T) 15 ).

[0073] (5) The extracted texture feature vectors of the CWD image and the contour lines of the blur function are T, respectively. CWD and T AFCL .

[0074] T CWD =[T1,T2,T3,T4,T5,T6,T7,T8,T9,T 10 ,T 11 ,T 12 ,T 13 ,T 14 ,T 15 ]

[0075] T AFCL =[T1,T2,T3,T4,T5,T6,T7,T8,T9,T 10 ,T 11 ,T 12 ,T 13 ,T 14 ,T 15 ]

[0076] 2. The shape features of the image are represented by pseudo-Zernike moments. A pseudo-Zernike moment of order p and repetition q is defined as follows:

[0077]

[0078] x 2 +y 2 =1

[0079] In the formula, p is a positive integer or zero, q is an integer and |q|≤p, and f(x,y) is a graphical function.

[0080] In polar coordinates, it is represented as

[0081]

[0082] To ensure rotational invariance and reduce the dynamic range of pseudo-Zernike moments, we take...

[0083] This invention selects This forms a set of 7-dimensional feature vectors.

[0084] Therefore, the feature vectors of the CWD image and the fuzzy function contour image are Z and Z, respectively. CWD Z AFCL .

[0085]

[0086]

[0087] The shape and texture features of the combined image are used to form the feature vector in this invention.

[0088] [T CWD ,T AFCL Z CWD Z AFCL ]

[0089] One specific embodiment of the technical solution of the present invention has been experimentally verified:

[0090] In the verification experiment, eight typical radar radiation source signals were selected to establish a database: conventional radar signal (CW), linear frequency modulated signal (LFM), binary phase coded signal (BPSK), polyphase coded signal (MPSK), four-phase frequency coded signal (4FSK), linear frequency modulated binary phase coded composite modulation signal (LFM / BPSK), linear frequency modulated four-phase frequency coded composite modulation signal (LFM / 4FSK), and binary phase coded four-phase frequency coded composite modulation signal (BPSK / 4FSK). The parameters for each radar signal were set as follows: pulse width T = 6 μs, sampling frequency f... s =200MHz, and the remaining parameters are shown in the table below, where U(·) represents the frequency based on the sampling frequency f. s A uniform distribution, for example, U(1 / 8, 1 / 4) represents a parameter range in [f s / 8,f s A random number between [ / 4].

[0091] Table 1 Simulation Signal Parameter Settings

[0092]

[0093] The radar signal database is established based on the signal-to-noise ratio (SNR) range, which is set from -8dB to 8dB in 2dB increments. For each SNR, 250 sets of CWD and AFCL images are generated for different types of signals, with 200 sets used as the training set and the remainder as the test set.

[0094] This invention employs a Support Vector Machine (SVM) based on radial basis function kernel functions to identify radar signals. SVM is a learning algorithm with a target output, belonging to the category of supervised learning algorithms, primarily used to solve data classification problems in data mining or pattern recognition fields. In the SVM algorithm, different types of kernel functions have little impact on the results; the penalty factor C and the kernel function parameter g are the most important.

[0095] This invention uses a grid parameter optimization method to determine the values ​​of the penalty factor C and the kernel function parameter g. Figure 2 This is the training model for SVM. First, set the parameter ranges for C and g: C = 2. i g = 2 j ,(i,j)∈R. In any set (C i ,g j Under the specified parameters, the training set is divided into 5 parts by default. One part is used as the test set to train the classifier, and the remaining parts are used for validation. This process is iterated repeatedly, and each iteration produces a set of average recognition accuracy. Finally, the parameters C and g corresponding to the optimal accuracy are selected as the best choices. In this invention, the values ​​of parameters C and g are set to the range of [-10, 10]. To improve the model's computational efficiency, the step size is set to 0.2; to effectively terminate training, the error threshold is set to 10. -4 The cross-validation parameter is 5.

[0096] Simulation Experiment 1:

[0097] The above eight typical radar emission source signals were identified. For each signal, 250 sample data points were generated at each signal-to-noise ratio (SNR). 80% of these were used for training, and the remaining 20% ​​for testing. Therefore, the training and testing sets contained 1600 and 400 sample data points of different radar signals at the same SNR, respectively. The SNR range was -8dB to 8dB, with experiments conducted every 2dB. The recognition rates of each radar modulation signal at different SNRs were as follows: Figure 3 As shown. From Figure 3As can be seen, the recognition accuracy of the identification method of the present invention for eight typical radar signals increases with the increase of signal-to-noise ratio. Figure 4 The overall recognition rate curves are for eight typical radar signals. From... Figure 4 As can be seen from the data, the recognition method of the present invention has a high recognition accuracy at low signal-to-noise ratios. In an environment with a signal-to-noise ratio of -8dB, the overall recognition rate of the eight signals can reach 81.5%, and in an environment with a signal-to-noise ratio of -2dB, the overall recognition rate of the eight signals can reach 100%. This demonstrates that the present invention has good recognition performance.

[0098] Simulation Experiment 2:

[0099] Robustness verification was performed on the above eight typical radar emission source signals. The signal-to-noise ratio (SNR) range was set to -8dB to 8dB, in 2dB increments. For each of the eight signals, 10 sets of data samples were generated for training and 5 sets for testing at each SNR. Therefore, the training set and the test set contained 720 and 360 sets of sample data, respectively. The confusion matrix of the test results is shown below. Figure 5 As shown. Figure 5 The values ​​on the diagonal represent the probability that this signal is correctly identified. Figure 5 As can be seen, the recognition rate of each signal is above 90%, with the recognition rate of CW and BPSK / 4FSK signals reaching 100%. The overall confusion matrix recognition rate of the eight signals is 96.39%, which demonstrates that the method of this invention has good generalization ability and robustness under complex signal-to-noise ratios.

Claims

1. A radar signal recognition method based on time-frequency image feature fusion, characterized in that, It comprises the following steps: (1) obtaining a CWD time-frequency image of a signal by using Choi-Williams time-frequency transform on the received signal and making a three-dimensional image of the signal by using a blur function, and then obtaining a corresponding blur function contour image; (2) pre-processing the CWD time-frequency image and the blur function contour image obtained in step (1), comprising: (2-1) obtaining a gray-scale image by gray-scale processing; (2-2) obtaining a filtered image by filtering the gray-scale image; (2-3) performing double cubic difference scaling on the filtered image to a set scale; (3) extracting characteristic values of the signal, comprising: extracting texture features of the CWD time-frequency image and the blur function contour image by using a gray-level gradient co-occurrence matrix, and extracting shape features of the CWD time-frequency image and the blur function contour image by using a pseudo-Zernike moment; (4) fusing image texture features and shape features; (5) using SVM for recognition processing: (5-1) adding labels to corresponding feature vectors; (5-2) determining values of a penalty factor C and a kernel function parameter g of SVM by using a grid parameter optimization method; (5-3) recognizing the received signal and outputting results; The gray-level gradient co-occurrence matrix (GLGCM) in step (3) is specifically: let the gray-scale image be f(M, N), and M and N are the row number and column number of the two-dimensional matrix corresponding to the gray-scale image, respectively; the corresponding GLGCM is calculated by the following steps: (1) calculating a normalized gradient matrix of f(M, N), and a gradient matrix g(M, N) of f(M, N) is extracted by using a 3*3 window Sobel operator, wherein the gradient value of the (k, l)th pixel point is calculated by the following formula: g x = f(k + 1, l - 1) + 2f(k + 1, l) + f(k + 1, l + 1) - f(k - 1, l - 1) -2f(k-1, l)-f(k-1, l+1) g y = f(k - 1, l + 1) + 2f(k, l + 1) + f(k + 1, l + 1) - f(k - 1, l + 1) -2f(k, l-1)-f(k+1, l-1) where k = 1, 2,..., M; l = 1, 2,..., N; the normalized gradient matrix is calculated using the formula where INT is an integer operation; g max is the maximum gradient value in g(M, N); N g is the expected maximum value after normalization of the gradient; (2) Calculate the normalized gray scale matrix of f(M,N); the formula is where f max is the maximum gray scale value in f(M,N); N f is the desired maximum value after normalizing the gray scale value; (3) the element value H(i,j) of the gray level gradient co-occurrence matrix, i.e. the number of pixel points simultaneously making F(k,l) = i (i ∈ [1,N f ]) and G(k,l) = j (j ∈ [1,N g ]) in the statistical normalized gray level matrix and the gradient matrix, the normalized GLGCM is obtained by using the following formula: The pseudo-Zernike moment in step (3) is a kind of orthogonal complex moment, and the definition of the pseudo-Zernike moment with order p and repetition q is: x 2 +y 2 =1 In the formula, p is a positive integer or zero, q is an integer, and |q|≤p, f(x, y) is an image function; wherein In polar coordinates, it is expressed as take In step (4), the shape features and the texture features of the image are combined to form a feature vector: [T CWD ,T AFCL ,Z CWD ,Z AFCL ].

2. The radar signal recognition method based on time-frequency image feature fusion according to claim 1, characterized in that, The mathematical expression of Choi-Williams Distribution in step (1) is: Where t represents time, s(t) represents a radar signal, * represents a complex conjugate operation, e represents an exponential operation with a natural constant as a base, j represents a virtual unit symbol, and σ represents an attenuation coefficient proportional to the amplitude of a cross-phase.

3. The radar signal recognition method based on time-frequency image feature fusion according to claim 1, characterized in that, The mathematical expression of the blur function in step (1) is: where τ denotes the time delay, t denotes time, f d denotes the Doppler frequency offset, * denotes the complex conjugate operation, e denotes the exponential operation with base of the natural constant, and j denotes the imaginary unit symbol.

Images