CN114818770A - Radar signal identification based on time-frequency image feature fusion - Google Patents

Abstract

The invention discloses a radar signal identification method based on time-frequency image feature fusion, which mainly solves the problem that the radar signal identification rate is low in the low signal-to-noise ratio environment in the prior art. The invention has the following steps: (1) performing time-frequency transformation on the radar signal to obtain a time-frequency image; (2) preprocessing a time-frequency image; (3) extracting texture features and shape features of the image; (4) and identifying the radar signals by adopting a support vector machine. The method extracts the textural features and shape features of the radar signal based on the Choi-Williams time-frequency image and the contour map of the fuzzy function, fuses the textural features and the shape features with the contour map to form a new radar signal identification characteristic value, and solves the problem of low radar signal identification rate under low signal-to-noise ratio.

Description

Radar signal identification based on time-frequency image feature fusion

Technical Field

The invention belongs to the field of radar signal processing, and particularly relates to a radar signal identification method based on time-frequency image feature fusion in the technical field of signal identification.

Background

The radar signal identification is a key link of an electronic reconnaissance system, and aims to extract characteristic parameters of sorted radar signals and automatically identify radar radiation source signals. With the increasing complexity of modern battlefield electromagnetic environments and the wide application of novel radar systems, it becomes crucial to quickly and effectively identify radar signals. The traditional method based on five conventional characteristic parameters (carrier frequency, pulse width, pulse amplitude, arrival time and arrival angle) cannot effectively identify, and effective characteristics of signals need to be extracted for effective identification of modulation modes of radar radiation source signals with diversified modulation modes. The characteristic parameters in the radar signal pulse can effectively reflect the essential information of the signal and expand the parameter space of signal identification, so that the research on the characteristic parameters in the radar signal pulse becomes a hot spot in time. At present, under the condition of high signal-to-noise ratio, the existing methods aim at more methods for identifying the signal modulation type and have higher identification rate. However, under the environment with low signal-to-noise ratio, the signal identification rate is low and the methods are few, and the method for identifying the complex modulation type signal is few, the identification rate is low, and the algorithm robustness is poor.

In the existing technical scheme, the fuzzy function is an important mathematical tool for researching radar signals, and can be used for researching different radar waveforms and representing the difference of different radar signals. The time-frequency analysis method is an important method for processing non-stationary signals, and the common time-frequency analysis methods mainly include: short-time Fourier transform (STFT), Wigner-Ville Distribution (Wigner-Ville Distribution), CWD, and the like; the STFT has the defect of poor time-frequency aggregation, the WVD can generate cross terms which interfere with the characteristics of real signals, and the CWD time-frequency analysis method has good time-frequency resolution and can effectively inhibit the cross terms. In order to solve the defects of the existing scheme, the invention designs a radar signal identification method by simultaneously utilizing the advantages of a fuzzy function and CWD time frequency analysis and adopting an SVM classifier with lower complexity so as to more accurately identify radar signals under the condition of low signal-to-noise ratio.

Disclosure of Invention

The invention aims to design a radar signal identification method based on time-frequency image feature fusion aiming at the defects and shortcomings of the prior art and aims to solve the problem of low signal-to-noise ratio radar signal identification rate.

The object of the invention is achieved by the following steps:

(1) obtaining a CWD time-frequency image of the radar signal by using Choi-Williams time-frequency transformation, making a signal three-dimensional graph by using a fuzzy function, and then solving a corresponding fuzzy function contour map;

(2) preprocessing the CWD time-frequency image and the contour image of the fuzzy function obtained in the step (1), comprising the following steps:

(2-1) graying to obtain a grayscale map;

(2-2) filtering the gray level image to obtain a filter image, for example, using wiener filtering;

(2-3) performing bicubic differential scaling on the filtered graph to a set scale, such as 224 x 224 size;

(3) extracting a characteristic value of a signal, comprising:

extracting texture features of the CWD time-frequency image and the fuzzy function contour map by utilizing a gray gradient co-occurrence matrix; extracting shape characteristics of the CWD time-frequency image and the fuzzy function contour map by using the pseudo Zernike moment;

(4) fusing image texture features and shape features;

(5) and (3) carrying out recognition processing by using an SVM:

(5-1) adding labels to the corresponding feature vectors;

(5-2) determining the values of a penalty factor C and a kernel function parameter g of the SVM by adopting a grid parameter optimization method;

and (5-3) identifying the received signal and outputting a result.

The mathematical expression of Choi-Williams Distribution in the step (1) is as follows:

where t represents time, s (t) represents radar signal,. indicates complex conjugate operation,. e represents exponential operation with natural constant as base, j represents imaginary unit symbol, and σ represents attenuation coefficient, and is proportional to the amplitude of the cross phase.

The mathematical expression of the fuzzy function in the step (1) is as follows:

where τ denotes time delay, t denotes time, f _d Denotes the doppler frequency offset, # denotes the complex conjugate operation, e denotes the exponential operation with the natural constant as the base, and j denotes the imaginary unit symbol.

The gray level gradient co-occurrence matrix (GLGCM) in the step (3) is specifically as follows: recording the gray level image as f (M, N), wherein M and N are respectively the row number and the column number of the two-dimensional matrix corresponding to the gray level image; calculating the corresponding GLGCM by:

(1) calculating a normalized gradient matrix of f (M, N), and extracting a gradient matrix g (M, N) of f (M, N) by adopting a Sobel operator of a 3 x 3 window, wherein a gradient value calculation formula of the (k, l) th pixel point is as follows:

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)

wherein k is 1,2 …, M; 1,2 …, N. Normalized gradient matrix utilization formula

Wherein INT is a rounding operation;g _max is the largest gradient value in g (M, N); n is a radical of _g Is the maximum expected after normalizing the gradient; preferably 32.

(2) Calculating a normalized gray matrix of f (M, N); is given by the formula

Wherein f is _max Is the maximum gray value in f (M, N); n is a radical of _f Is the maximum value expected after the grey value is normalized; n is a radical of _f Preferably 32.

(3) The element value H (i, j) of the gray gradient co-occurrence matrix, i.e. F (k, l) is made to be i (i belongs to [1, N) simultaneously in the statistical normalization gray matrix and the gradient matrix _f ]) And G (k, l) ═ j (j. epsilon. [1, N) _g ]) The normalized GLGCM is obtained using the following formula:

the pseudo-Zernike moment in the step (3) is an orthogonal complex moment, the order is p, and the pseudo-Zernike moment with the repetition degree of q is defined as:

x ² +y ² ＝1

in the formula, p is a positive integer or zero, q is an integer, | q | < p, and f (x, y) is an image function; wherein the content of the first and second substances,

expressed in polar coordinates as

For kinematic rotation invariance and reducing the dynamic range of pseudo-Zernike moments

Extracting texture feature and shape feature of image, i.e. extracting texture feature T of image by using GLGCM _CWD 、T _AFCL Extracting shape feature Z of the image by using the pseudo Zernike moment _CWD 、Z _AFCL The two characteristic parameters are combined into a characteristic vector T _CWD ,T _AFCL ,Z _CWD ,Z _AFCL ]；

The invention has the following outstanding advantages: 1. the method comprises the steps of realizing effective identification of radar signals of 8 different modulation types based on time-frequency image feature fusion, wherein the modulation types comprise frequency modulation, phase modulation and composite modulation signals mixed by two modes; 2. the average correct recognition rate of 8 signals can reach more than 80% in the environment that the signal-to-noise ratio is as low as-8 dB; 3. the radar signal parameters used by the invention are all selected in a dynamic range, so that the radar signal is insensitive to parameter change, has better generalization capability and can meet the actual use requirement.

Drawings

FIG. 1 is a schematic diagram of a radar signal identification method based on time-frequency image feature fusion according to the present invention.

FIG. 2 is a diagram of an SVM training model used in the practice of the present invention to identify 8 radar signals.

Fig. 3 is a graph showing the change of the respective recognition rates of 8 radar signals in the implementation of the present invention.

Fig. 4 is a graph of the overall correct recognition rate of 8 radar signals as a function of the signal-to-noise ratio in the practice of the present invention.

Fig. 5 is a graph of a confusion matrix for 8 radar signals in an implementation of the present invention.

Detailed Description

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

As shown in fig. 1, the radar signal identification method based on time-frequency image feature fusion of the present invention mainly comprises the following steps: firstly, time-frequency transformation is carried out on radar signals to obtain time-frequency images of signals of different modulation types, then the images are preprocessed to inhibit the influence of noise and the like and reduce the calculated amount, then shape features and texture features of the images are extracted and fused, and finally an intra-pulse modulation mode of the radar signals is identified through an SVM and output.

The specific method for processing the radar signal comprises the following steps:

the radar signal time frequency transformation comprises CWD time frequency transformation and a fuzzy function. Wherein the mathematical expression of CWD time-frequency transformation is

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

The fuzzy function is another time-frequency distribution function and is an important mathematical tool for researching radar signals. The blur function can be used not only to study different radar waveforms, but also to characterize the differences of different radar signals. Mathematical expression of the fuzzy function is

Where τ denotes time delay, t denotes time, f _d Denotes the doppler shift, denotes the complex conjugate operation, e denotes the exponential operation with the natural constant as the base, and j denotes the imaginary unit symbol.

Due to the influence of noise and time-frequency cross terms, a signal time-frequency diagram has a large amount of interference information. Therefore, before image feature extraction, the image processing technology is adopted to preprocess the time-frequency graph, so that interference and redundant information can be effectively reduced, and the effectiveness of feature extraction is enhanced. In the preprocessing process, the time-frequency image is grayed to obtain a gray image, then the gray image is subjected to wiener filtering to obtain a filtering image, and finally the filtering image is subjected to bicubic difference scaling to 224 x 224.

The method adopts GLGCM and the pseudo Zernike moment to respectively extract the texture characteristic and the shape characteristic of the radar signal.

1. The texture feature extraction step comprises:

(1) a normalized gradient matrix of f (M, N) is calculated. Extracting a gradient matrix g (M, N) of f (M, N) by using a Sobel operator of a 3 x 3 window, wherein a gradient value calculation formula of a (k, l) th pixel point is

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)

Normalized gradient matrix utilization formula

Wherein k ═

1,2 …, M; 1,2 …, N; INT is rounding operation; g _max Is the largest gradient value in g (M, N);

N _g is the maximum value expected after normalizing the gradient, and the value of the embodiment is 32.

(2) A normalized grayscale matrix of f (M, N) is calculated. Is given by the formula

Wherein f is _max Is the maximum gray value in f (M, N); n is a radical of _f Is the maximum value expected after the gray value normalization, and this embodiment takes N _f Is 32.

(3) Element values H (i, j) of the gray gradient co-occurrence matrix, i.e. statistical normalized graySimultaneously making F (k, l) i (i ∈ [1, N) in the matrix and gradient matrix _f ]) And G (k, l) ═ j (j. epsilon. [1, N) _g ]) The number of pixel points. Normalized GLGCM was obtained using the following formula:

(4) in the embodiment, 15 characteristic parameters of the GLGCM are selected, specifically including small gradient dominance (T) ₁ ) Dominant large gradient (T) ₂ ) Gradient distribution inhomogeneity (T) ₃ ) Non-uniformity of gray distribution (T) ₄ ) Energy (T) ₅ ) Mean value of gray (T) ₆ ) Mean of gradient (T) ₇ ) Standard deviation of gray scale (T) ₈ ) Gradient standard deviation (T) ₉ ) Correlation (T) ₁₀ ) Entropy of gray scale (T) ₁₁ ) Gradient entropy (T) ₁₂ ) Mixed entropy (T) ₁₃ ) Differential moment (T) ₁₄ ) Inverse difference moment (T) ₁₅ )。

(5) The extracted texture feature vectors of the contour lines of the CWD image and the fuzzy function are respectively T _CWD And T _AFCL 。

T _CWD ＝[T ₁ ,T ₂ ,T ₃ ,T ₄ ,T ₅ ,T ₆ ,T ₇ ,T ₈ ,T ₉ ,T ₁₀ ,T ₁₁ ,T ₁₂ ,T ₁₃ ,T ₁₄ ,T ₁₅ ]

T _AFCL ＝[T ₁ ,T ₂ ,T ₃ ,T ₄ ,T ₅ ,T ₆ ,T ₇ ,T ₈ ,T ₉ ,T ₁₀ ,T ₁₁ ,T ₁₂ ,T ₁₃ ,T ₁₄ ,T ₁₅ ]

2. The shape features of the image employ pseudo-Zernike moments. The pseudo-Zernike moments with order p and degree of repetition q are defined as:

x ² +y ² ＝1

in the formula, p is a positive integer or zero, q is an integer, | q | ≦ p, and f (x, y) is an image function. Wherein the content of the first and second substances,

expressed in polar coordinates as

For kinematic rotation invariance and reducing the dynamic range of pseudo-Zernike moments

Selection of the invention

A set of 7-dimensional feature vectors is formed.

Therefore, the characteristic vectors of the CWD image and the contour image of the blur function are respectively Z _CWD ，Z _AFCL 。

The shape feature and texture feature of the combined image constitute the feature vector used in the invention

[T _CWD ,T _AFCL ,Z _CWD ,Z _AFCL ]

A specific implementation of the technical scheme of the invention is verified by experiments:

in the verification experiment, 8 typical radar radiation source signals are selected to establish a database: conventional radar signal (CW), a chirp signal (LFM), a binary coded signal (BPSK), a polyphase coded signal (MPSK), a four-phase frequency coded signal (4FSK), a chirp two-phase coded complex modulated signal (LFM/BPSK), a chirp four-term frequency coded complex modulated signal (LFM/4FSK) and a two-phase coded four-term frequency coded complex modulated signal (BPSK/4 FSK). The radar signal parameters are set as follows: pulse width T6 mus, sampling frequency f _s The remaining parameters are shown in the table below, where U (·) is based on the sampling frequency f _s Is uniformly distributed, e.g. U (1/8,1/4) indicates a parameter in the range of [ f ] _s /8,f _s /4]A random number in between.

Table-simulation signal parameter set

The database of radar signals is built according to the range of signal-to-noise ratio, the range of signal-to-noise ratio is set to-8 dB to 8dB, and the step is 2 dB. At each signal-to-noise ratio, the different types of signals each produce 250 sets of CWD and AFCL images, 200 of which are used in the training set and the rest in the test set.

The method adopts a support vector machine based on the kernel function of the radial basis function to identify the radar signals. The support vector machine is a learning algorithm with target output, belongs to one of supervised learning algorithms, and mainly solves the problem of data classification in the field of data mining or pattern recognition. In the SVM algorithm, different kinds of kernel functions have little influence on the result, and a penalty factor C and a kernel function parameter g play an important role.

The present invention uses a grid parameter optimization method to determine the values of penalty factor C and kernel function parameter g. Fig. 2 is a training model of the SVM. Firstly, setting the parameter range of C and g: c2 ⁱ ,g＝2 ^j And (i, j) is ∈ R. In any group (C) _i ,g _j ) Under the parameters, the training set is divided into 5 parts by default, one part of the training set is used as a test set to train the classifier, the other part of the training set is used for inspection, the method is traversed and circulated, a group of average recognition accuracy rates exist each time, and finally the optimal accuracy is obtainedThe C, g parameter corresponding to the accuracy is selected as the best choice. In the invention, the value range of the parameters C and g is set to be [ -10,10 [)]Setting the step length to be 0.2 in order to improve the model operation efficiency; to effectively terminate training, an error threshold of 10 is set ^-4 The cross-validation parameter is 5.

Simulation experiment 1:

the above 8 typical radar radiation source signals were identified, each signal yielding 250 sample data at each signal-to-noise ratio, 80% of which was used for training and the remaining 20% for testing. Therefore, 1600 sample data and 400 sample data of different radar signals under the same signal-to-noise ratio are respectively arranged in the training set and the test set. The range of the signal-to-noise ratio is-8 dB to 8dB, an experiment is performed every 2dB, and the identification rate of each radar modulation signal under different signal-to-noise ratios is shown in figure 3. As can be seen from fig. 3, the recognition accuracy of the recognition method of the present invention for 8 typical radar signals increases with the signal-to-noise ratio. Fig. 4 is a plot of the overall recognition rate of 8 exemplary radar signals. It can be seen from fig. 4 that the recognition method of the present invention has a high recognition accuracy at a low signal-to-noise ratio, the overall recognition rate of 8 signals can reach 81.5% in an environment with a signal-to-noise ratio of-8 dB, and the overall recognition rate of 8 signals can reach 100% in an environment with a signal-to-noise ratio of-2 dB, which indicates that the method of the present invention has a good recognition performance.

Simulation experiment 2:

robustness verification was performed on the above 8 typical radar radiation source signals. The snr range was set to-8 dB to 8dB, stepped by 2dB, with 8 signals at each snr generating 10 groups of data samples for training and 5 groups of data samples for testing. Thus, the training set and the test set have 720 combinations and 360 sets of sample data, respectively, and the confusion matrix of the test results is shown in FIG. 5. The diagonal values of fig. 5 indicate the probability of such signals being correctly identified, and it can be seen from fig. 5 that the identification rate of each signal is above 90%, where the identification rate of CW, BPSK/4FSK signals is 100%. The recognition rate of the 8 signals of the total confusion matrix is 96.39%, so that the method has better generalization capability and robustness under the condition of complex signal to noise ratio.

Claims (5)

1. The radar signal identification method based on time-frequency image feature fusion is characterized by comprising the following steps of:

(1) obtaining a CWD time-frequency image of the signal by using Choi-Williams time-frequency transformation on the received signal, making a three-dimensional graph of the signal by using a fuzzy function, and then solving a corresponding fuzzy function contour map;

(2) preprocessing the CWD time-frequency image and the contour image of the fuzzy function obtained in the step (1), comprising the following steps:

(2-1) graying to obtain a grayscale map;

(2-2) filtering the gray level image to obtain a filter image;

(2-3) performing bicubic difference value scaling on the filter graph to a set scale;

(3) extracting a characteristic value of a signal, comprising:

extracting texture features of the CWD time-frequency image and the fuzzy function contour map by utilizing the gray gradient co-occurrence matrix; extracting shape characteristics of the CWD time-frequency image and the fuzzy function contour map by using the pseudo Zernike moment;

(4) fusing image texture features and shape features;

(5) and (3) carrying out recognition processing by using an SVM:

(5-1) adding labels to the corresponding feature vectors;

(5-2) determining the values of a penalty factor C and a kernel function parameter g of the SVM by adopting a grid parameter optimization method;

and (5-3) identifying the received signal and outputting a result.

2. The radar signal identification method based on time-frequency image feature fusion as claimed in claim 1, wherein the mathematical expression of Choi-Williams Distribution in step (1) is as follows:

where t represents time, s (t) represents radar signal,. indicates complex conjugate operation,. e represents exponential operation with natural constant as base, j represents imaginary unit symbol, and σ represents attenuation coefficient, and is proportional to the amplitude of the cross phase.

3. The radar signal identification method based on time-frequency image feature fusion as claimed in claim 1, wherein the mathematical expression of the fuzzy function in the step (1) is as follows:

where τ denotes time delay, t denotes time, f _d Denotes the doppler shift, denotes the complex conjugate operation, e denotes the exponential operation with the natural constant as the base, and j denotes the imaginary unit symbol.

4. The radar signal identification method based on time-frequency image feature fusion according to claim 1, wherein the gray level gradient co-occurrence matrix (GLGCM) of step (3) is specifically: recording the gray level image as f (M, N), wherein M and N are respectively the row number and the column number of the two-dimensional matrix corresponding to the gray level image; calculating the corresponding GLGCM by:

(1) calculating a normalized gradient matrix of f (M, N), and extracting a gradient matrix g (M, N) of f (M, N) by adopting a Sobel operator of a 3 x 3 window, wherein a gradient value calculation formula of the (k, l) th pixel point is as follows:

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)

wherein k is 1,2 …, M; 1,2 …, N; normalized gradient matrix utilizing formula

Wherein INT is a rounding operation; g _max Is the largest gradient value in g (M, N); n is a radical of _g Is the maximum expected after normalizing the gradient;

(2) calculating a normalized gray matrix of f (M, N); is given by the formula

Wherein f is _max Is the maximum gray value in f (M, N); n is a radical of _f Is the maximum value expected after the grey value is normalized;

(3) the element value H (i, j) of the gray level gradient co-occurrence matrix, namely F (k, l) is made to be i (i belongs to [1, N) in the statistical normalization gray level matrix and the gradient matrix simultaneously _f ]) And G (k, l) ═ j (j. epsilon. [1, N) _g ]) The normalized GLGCM is obtained using the following formula:

5. the radar signal identification method based on time-frequency image feature fusion as claimed in claim 1, wherein the pseudo-Zernike moments in step (3) are orthogonal complex moments, the order is p, and the pseudo-Zernike moments with repetition degree q are defined as:

x ² +y ² ＝1

in the formula, p is a positive integer or zero, q is an integer, | q | < p, and f (x, y) is an image function; wherein the content of the first and second substances,

expressed in polar coordinates as

Get

Images