A modulation type recognition method based on adaptive noise reduction and texture feature extraction

Abstract

The application discloses a modulation type recognition method based on adaptive noise reduction and texture feature extraction, and belongs to the field of signal modulation type recognition.The application realizes the method as follows: a DR2D pretreatment method is used to calculate the two-dimensional joint frequency features of a radiation source signal; according to the noise level, a processing threshold of threshold filtering and a useful signal frequency band are adaptively determined in sequence, and the two-dimensional joint frequency features are subjected to threshold filtering and narrowband filtering; two-dimensional inverse fast Fourier transform is carried out to obtain a two-dimensional rearranged matrix of the radiation source signal after noise reduction, and further normalization is carried out to obtain a two-dimensional rearranged gray image; a gray level co-occurrence matrix of the image is calculated to extract a texture feature vector of the radiation source signal; a support vector machine is used for classification to realize sorting and recognition of the modulation type of the radiation source signal.The application can realize sorting and recognition of the radiation source signal under a low signal-to-noise ratio, and has the advantages of high recognition accuracy, small calculation complexity and easy hardware implementation.

Description

TECHNICAL FIELD

[0001] The present application relates to a modulation type recognition method based on adaptive noise reduction and texture feature extraction, belonging to the field of signal modulation type recognition. BACKGROUND

[0002] Traditional radiation source signal modulation type sorting recognition mainly extracts the time-frequency image of the signal, and classifies it by using convolutional neural network. The recognition effect of this method depends on the time-frequency analysis method, time-frequency resolution, and the structure and number of layers of convolutional neural network (CNN). In order to achieve good recognition effect under low signal-to-noise ratio, it is often necessary to increase the number of fast Fourier transform (FFT) calculations of time-frequency analysis and the number of layers of convolutional neural network, which increases the calculation complexity and the requirement for hardware.

[0003] Prior art [1] (see Yao Y, Wang Z H. Radar signal recognition based on time-frequency preprocessing convolution network [J]. Journal of Detection and Control, 2018, 40(06): 99-105.) gives a signal sorting recognition method based on convolutional neural network (CNN) under time-frequency image preprocessing. The signal is transformed by using the good time-frequency aggregation characteristics of the smoothed pseudo Wigner-Ville distribution, and preprocessed by main energy ridge extraction and binaryzation, etc. Then the processed time-frequency image is input into the pre-trained CNN for recognition. Compared with the method of manually extracting features, the recognition accuracy of this recognition method is improved under low signal-to-noise ratio, which proves the feasibility of using time-frequency image for feature extraction and convolutional neural network for classification in signal sorting recognition technology. The recognition accuracy of this method is about 90% under-10dB signal-to-noise ratio.

[0004] Prior art [2] (see Liu Y T, Yan X P, Hao X H, et al. Automatic modulation recognition of radiation source signals based on data rearrangement and the 2D FFT [J]. Remote Sensing, 2023, 15(2).) gives a signal two-dimensional joint frequency feature extraction method based on data rearrangement and two-dimensional fast Fourier transform (Data Rearrangement and the 2D FFT, DR2D). The frequency feature has frequency resolution in two time scales, and is more concentrated in energy than one-dimensional FFT spectrum, which is suitable for feature extraction and recognition. SUMMARY

[0005] The main purpose of the present application is to provide a modulation type recognition method based on adaptive noise reduction and texture feature extraction, calculate the two-dimensional joint frequency feature of the radiation source signal, adaptively determine the processing threshold of threshold filtering and the useful signal frequency band according to the noise level, and perform threshold filtering and narrowband filtering on the two-dimensional joint frequency feature, perform two-dimensional inverse fast Fourier transform to obtain the two-dimensional rearrangement matrix of the noise-reduced radiation source signal, and further normalize to obtain a two-dimensional rearrangement grayscale image, calculate the gray level co-occurrence matrix of the image, extract the texture feature vector of the radiation source signal, use a support vector machine for classification, and realize the sorting and recognition of the radiation source signal modulation type. The present application can realize the sorting and recognition of the radiation source signal under low signal-to-noise ratio, has the advantages of high recognition accuracy, small calculation complexity and easy hardware implementation.

[0006] The purpose of the present application is realized by the following technical solutions.

[0007] The modulation type recognition method based on adaptive noise reduction and texture feature extraction disclosed by the present application comprises the following steps:

[0008] Step one: calculate the two-dimensional joint frequency feature Y(m,n) of the radiation source signal using the DR2D preprocessing method.

[0009] The M*N point radiation source signal is divided into N segments of M point sub-signal segments, each sub-signal segment is taken as a row of a two-dimensional matrix, and column stacking is performed in order to obtain a two-dimensional data rearrangement matrix X of M*N size.

[0010] The frequency spectrum matrix obtained by performing two-dimensional fast Fourier transform on the matrix X is:

[0011]

[0012] In the formula, (k,l) and (m,n) are indexes of the original matrix X and the frequency spectrum matrix Y respectively, and k,m=0,1,···,M-1, l,n=0,1,···,N-1.

[0013] The two-dimensional joint frequency feature obtained by taking the modulus of the frequency spectrum matrix and normalizing is:

[0014]

[0015] Step two: adaptively determine the processing threshold of threshold filtering according to the noise level, and perform threshold filtering on the two-dimensional joint frequency feature.

[0016] A certain threshold plane cuts the two-dimensional joint frequency feature image obtained in step one to obtain a feature image slice under the threshold thr, if the feature image is cut on the plane where the threshold is located, the pixel value of the feature image slice at the corresponding position is recorded as 1, otherwise as 0, then the feature image slice is represented as:

[0017]

[0018] frequency feature proportion is expressed as:

[0019]

[0020] The frequency feature proportion is a ratio of a sum of pixel values of a feature image slice at any threshold value to a number of pixels of a two-dimensional joint frequency feature image.

[0021] A group of threshold values in a predetermined range at a certain step value are used to cut the two-dimensional joint frequency feature image with a threshold plane, to obtain a group of values of frequency feature proportions, and a suitable corresponding threshold value is selected as a reference threshold thr0 according to the values of the frequency feature proportions and values of differentials of the frequency feature proportions.

[0022] To prevent too much loss of useful signals, the reference threshold is contracted to obtain a processing threshold:

[0023]

[0024] In the formula, α is a convergence coefficient.

[0025] The processing threshold thr c The two-dimensional joint frequency feature is threshold filtered.

[0026] Step three: the processing threshold thr c The useful signal frequency band is determined and narrowband filtering is performed to obtain a two-dimensional joint frequency feature after noise reduction.

[0027] The filtering threshold thr c The two-dimensional joint frequency feature is binarized to obtain a feature image slice S c at the threshold value; considering that the useful signal frequency feature is longitudinally distributed, the ratio of a sum of values of each column of the binarized distribution image to the column length is calculated to form a frequency feature distribution vector of the signal, denoted as:

[0028]

[0029] When an element value of the frequency feature distribution vector is less than a preset threshold value of the frequency feature distribution vector, an index range of the element value is an out-of-band noise distribution; when the element value of the frequency feature distribution vector is greater than the preset threshold value of the frequency feature distribution vector, an index range of the element value is an in-band useful signal distribution; according to the preset threshold value of the frequency feature distribution vector, the frequency feature distribution vector F cThe binarization processing is performed, and the two-dimensional joint frequency features are copied according to the size of the two-dimensional joint frequency features and arranged by columns to further obtain a frequency feature distribution mask plate; the out-of-band noise components of the two-dimensional joint frequency feature image are subjected to a zero operation according to the frequency feature distribution mask plate, narrowband filtering is realized, and the two-dimensional joint frequency features after noise reduction are obtained;

[0030] Step four: the two-dimensional joint frequency features after noise reduction obtained in step three are subjected to two-dimensional inverse fast Fourier transform to obtain a two-dimensional rearranged matrix of the radiation source signal after noise reduction, and a two-dimensional rearranged gray image is further normalized to obtain a two-dimensional rearranged gray image of the radiation source signal.

[0031] Step five: a gray level co-occurrence matrix of the two-dimensional rearranged gray image obtained in step four is solved to extract a texture feature vector of the radiation source signal.

[0032] The gray level co-occurrence matrix of the two-dimensional rearranged gray image in the direction of θ = 0°, θ = 45°, θ = 90° and θ = 135° is calculated.

[0033] The element value of the gray level co-occurrence matrix at index (a, b) is p(a, b), and the contrast, correlation, energy, homogeneity and entropy of the two-dimensional rearranged gray image are respectively:

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] wherein μ a and μ b are the mean values of the image gray levels in the row and column directions respectively, σ a and σ b are the standard deviations of the image gray levels in the row and column directions respectively.

[0040] The statistical feature vector of the two-dimensional rearranged gray image in the direction of θ is:

[0041] s θ = [Contrast θ , Correlation θ , Energy θ , Homogeneity θ , Entropy θ ] (12)

[0042] The statistical feature vectors in four directions of θ=0°, θ=45°, θ=90° and θ=135° are spliced to obtain a texture feature vector of the radiation source signal:

[0043] S f =[s θ＝0° ,s θ＝45° ,s θ＝90° ,s θ＝135° ] (13)

[0044] Step six: according to steps one to five, the texture feature vectors of the training samples of the radiation source signals of different modulation types are trained to obtain a support vector machine classification model.

[0045] Step seven: the support vector machine classification model obtained in step six is used for sorting and identifying the modulation type of the radiation source signal.

[0046] According to steps one to five, the texture feature vectors of the received radiation source signals are calculated, and the support vector machine classification model obtained in step six is used for classification, and the classification result obtained corresponds to the modulation type of the received radiation source signal, so that the sorting and identification of the modulation type of the radiation source signal are realized.

[0047] The present application extracts the texture features of the two-dimensional rearranged gray image of the radiation source signal for sorting and identifying the modulation type while denoising the radiation source signal, and the identification accuracy is high.

[0048] Advantages:

[0049] 1. The modulation type identification method based on adaptive noise reduction and texture feature extraction disclosed in the present application calculates the two-dimensional joint frequency features of the radiation source signal; the processing threshold value of threshold filtering and the useful signal frequency band are adaptively determined in sequence according to the noise level, and the two-dimensional joint frequency features are subjected to threshold filtering and narrowband filtering; two-dimensional fast inverse Fourier transform is performed to obtain a two-dimensional rearranged matrix of the denoised radiation source signal, and further normalization is performed to obtain a two-dimensional rearranged gray image; the gray level co-occurrence matrix of the image is calculated, and the texture feature vector of the radiation source signal is extracted; the support vector machine is used for classification to realize sorting and identification of the modulation type of the radiation source signal.

[0050] 2. The modulation type identification method based on adaptive noise reduction and texture feature extraction disclosed in the present application performs DR2D adaptive noise reduction on the radiation source signal, and has better noise reduction effect under low signal-to-noise ratio compared with the traditional noise reduction algorithm; the gray level co-occurrence matrix of the two-dimensional rearranged gray image of the denoised radiation source signal is calculated, and the texture feature vector is extracted to train the support vector machine classification model, which can realize accurate identification of the modulation type of the radiation source signal, and the principle is simple, the calculation complexity is small, and the hardware implementation is easy.

[0051] 3. The modulation type recognition method based on adaptive noise reduction and texture feature extraction disclosed in the present application, the two-dimensional joint frequency feature extracted by the DR2D method has the following characteristics: the useful signal frequency components are more concentrated in a certain area, the noise frequency components are dispersed in the entire feature space, and the noise amplitude is significantly smaller than the signal amplitude, which is suitable for threshold filtering for noise reduction. The time domain arrangement of different modulation type signals after noise reduction has obvious texture features, and the extraction thereof can be used for sorting and identifying the modulation type of the radiation source signal. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 Flow chart of the radiation source signal modulation type sorting and identification method based on DR2D adaptive noise reduction and texture feature extraction;

[0053] Figure 2 Two-dimensional joint frequency feature map of the linear frequency modulation signal;

[0054] Figure 3 Curve graph of the frequency feature proportion of the linear frequency modulation signal changing with the set threshold;

[0055] Figure 4 Linear frequency modulation signal frequency feature distribution mask plate;

[0056] Figure 5 Two-dimensional rearranged gray scale image of the linear frequency modulation signal after noise reduction;

[0057] Figure 6 Gray level co-occurrence matrix graph of the two-dimensional rearranged gray scale image of the linear frequency modulation signal after noise reduction; wherein, Figure 6 a) is the gray level co-occurrence matrix of the two-dimensional rearranged gray scale image of the linear frequency modulation signal after noise reduction in the direction of θ=0°, Figure 6 b) is the gray level co-occurrence matrix of the two-dimensional rearranged gray scale image of the linear frequency modulation signal after noise reduction in the direction of θ=45°, Figure 6 c) is the gray level co-occurrence matrix of the two-dimensional rearranged gray scale image of the linear frequency modulation signal after noise reduction in the direction of θ=90°, Figure 6 d) is the gray level co-occurrence matrix of the two-dimensional rearranged gray scale image of the linear frequency modulation signal after noise reduction in the direction of θ=135°;

[0058] Figure 7 Recognition accuracy fold line graph of the linear frequency modulation signal under different signal-to-noise ratios.

[0059] Figure 8 Two-dimensional joint frequency feature map of the pseudo-random code and linear frequency modulation composite modulation signal;

[0060] Figure 9 Curve graph of the frequency feature proportion of the pseudo-random code and linear frequency modulation composite modulation signal changing with the set threshold;

[0061] Figure 10 Frequency feature distribution mask plot of pseudo-random code and linear frequency modulation composite modulation signal;

[0062] Figure 11 Two-dimensional rearranged gray scale plot of pseudo-random code and linear frequency modulation composite modulation signal after noise reduction;

[0063] Figure 12 Gray level co-occurrence matrix plot of two-dimensional rearranged gray scale image of pseudo-random code and linear frequency modulation composite modulation signal after noise reduction; wherein, Figure 12 a) is the gray level co-occurrence matrix of the two-dimensional rearranged gray scale image of pseudo-random code and linear frequency modulation composite modulation signal after noise reduction in the direction of θ = 0°, Figure 12 b) is the gray level co-occurrence matrix of the two-dimensional rearranged gray scale image of pseudo-random code and linear frequency modulation composite modulation signal after noise reduction in the direction of θ = 45°, Figure 12 c) is the gray level co-occurrence matrix of the two-dimensional rearranged gray scale image of pseudo-random code and linear frequency modulation composite modulation signal after noise reduction in the direction of θ = 90°, Figure 12 d) is the gray level co-occurrence matrix of the two-dimensional rearranged gray scale image of pseudo-random code and linear frequency modulation composite modulation signal after noise reduction in the direction of θ = 135°;

[0064] Figure 13 Bend line plot of recognition accuracy of pseudo-random code and linear frequency modulation composite modulation signal under different signal-to-noise ratios. DETAILED DESCRIPTION

[0065] In order to better illustrate the purpose and advantages of the present application, the content of the application will be further described below in conjunction with the drawings.

[0066] Example 1

[0067] In order to verify the feasibility of the method, a linear frequency modulation signal with a carrier frequency of 50MHz, a modulation frequency of 200KHz and a modulation bandwidth of 100MHz is selected, and 65536 points of data obtained by using a 1GHz sampling rate are used under the condition of -10dB signal-to-noise ratio.

[0068] As Figure 1 shown, the present example discloses a modulation type recognition method based on adaptive noise reduction and texture feature extraction, and the specific implementation steps are as follows:

[0069] Step one: calculate the two-dimensional joint frequency feature of the linear frequency modulation signal using the DR2D preprocessing method.

[0070] The 256x256 point linear frequency modulation signal is divided into 256 segments of 256 point sub-signal segments, each sub-signal segment is taken as a row of two-dimensional matrix, and column stacking is performed in order to obtain a 256x256 size linear frequency modulation signal two-dimensional data rearrangement matrix X.

[0071] The frequency spectrum matrix obtained by performing two-dimensional fast Fourier transform on the matrix X is:

[0072]

[0073] where (k, l) and (m, n) are indexes of the original matrix X and the frequency spectrum matrix Y respectively, and k, m = 0, 1,..., 255, l, n = 0, 1,..., 255.

[0074] The two-dimensional joint frequency feature of the linear frequency modulation signal is obtained by taking the modulus of the frequency spectrum matrix and normalizing it:

[0075]

[0076] The two-dimensional joint frequency feature of the linear frequency modulation signal is shown in FIG. 2. Figure 2

[0077] Step 2: The processing threshold of threshold filtering is adaptively determined according to the noise level, and the two-dimensional joint frequency feature of the linear frequency modulation signal is threshold filtered.

[0078] Any threshold plane is used to cut the two-dimensional joint frequency feature image obtained in step 1 to obtain a feature image slice under the threshold thr. If the plane where the threshold is located has a feature image cut, the pixel value of the feature image slice at the corresponding position is recorded as 1, otherwise as 0. The feature image slice is expressed as:

[0079]

[0080] The frequency feature ratio is expressed as:

[0081]

[0082] The frequency feature ratio is the ratio of the sum of pixel values of the feature image slice under any threshold to the number of pixels of the two-dimensional joint frequency feature image;

[0083] The threshold range is set to [0, 1] with a step of 0.01, and the frequency feature ratio under different thresholds is obtained, and the change curve is shown in FIG. 3. Figure 3 The curve has a clear change in slope before and after, representing whether the threshold plane is within the amplitude range of the noise component. The threshold corresponding to the position where the frequency feature ratio is less than 0.02 and the differential of the frequency feature ratio is greater than -0.002 is selected as the reference threshold thr0 of filtering.

[0084] To prevent too much loss of useful signals, the reference threshold is contracted to obtain the processing threshold:

[0085]

[0086] ​​In the formula, α is the convergence coefficient, and its value ranges from [1,2].

[0087] Taking α = 1.5, the filtering threshold thr for the linear frequency modulated signal c =0.07.

[0088] Use the processing threshold thr c Threshold filtering is applied to the two-dimensional joint frequency characteristics of a linear frequency modulated signal.

[0089] Step 3: Apply the processing threshold thr from Step 2 c The useful signal frequency band is determined and narrowband filtering is performed to obtain the two-dimensional joint frequency characteristics of the denoised linear frequency modulated signal.

[0090] Use the filter threshold thr obtained in step two c Binarizing the two-dimensional joint frequency features of the linear frequency modulated signal yields a feature image slice S at the specified threshold. c Considering that the frequency characteristics of the useful signal are distributed vertically, the ratio of the sum of the values ​​in each column of the binary distribution image to the column length is calculated to form the frequency characteristic distribution vector of the signal, denoted as:

[0091]

[0092] When the element value of the frequency characteristic distribution vector is less than a preset threshold for the frequency characteristic distribution vector, the index range of the element value is the out-of-band noise distribution; when the element value of the frequency characteristic distribution vector is greater than the preset threshold for the frequency characteristic distribution vector, the index range of the element value is the in-band useful signal distribution; the frequency characteristic distribution vector F is adjusted according to the preset threshold for the frequency characteristic distribution vector. c Binarization is performed, and the frequency features are copied and arranged column-wise according to their magnitude to obtain a frequency feature distribution mask. The threshold of the frequency feature distribution vector is set to 0.1 times the maximum value of the frequency feature distribution vector, resulting in the frequency feature distribution mask as shown below. Figure 4 As shown.

[0093] By zeroing out-of-band noise components in the two-dimensional joint frequency feature image using a frequency feature distribution mask, narrowband filtering is achieved, resulting in the two-dimensional joint frequency features of the denoised linear frequency modulated signal.

[0094] Step 4: Obtain a two-dimensional rearranged grayscale image of the linear frequency modulated signal.

[0095] A two-dimensional inverse fast Fourier transform is performed on the two-dimensional joint frequency features of the linear frequency modulated (LFM) signal after threshold filtering and narrowband filtering to obtain the two-dimensional rearranged matrix of the denoised LFM signal. Further normalization yields the two-dimensional rearranged grayscale image, as shown below. Figure 5 As shown.

[0096] Step 5: Solve for the gray-level co-occurrence matrix of the two-dimensional rearranged grayscale image obtained in Step 4, and extract the texture feature vector of the linear frequency modulated signal.

[0097] The gray-level co-occurrence matrices of the two-dimensional rearranged gray-level image obtained in step four are calculated in the directions θ = 0°, θ = 45°, θ = 90°, and θ = 135°, respectively, as shown below. Figure 6 As shown in 6a), 6b), 6c) and 6d).

[0098] If the element at index (a,b) of the gray-level co-occurrence matrix is ​​p(a,b), then the contrast, correlation, energy, homogeneity, and entropy of the two-dimensional rearranged gray-level image are as follows:

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] Where, μ a and μ b σ represents the mean of the image grayscale values ​​in the row and column directions, respectively. a and σ b These are the standard deviations of the image grayscale values ​​in the row and column directions, respectively.

[0105] The statistical feature vector of a two-dimensional rearranged grayscale image in the θ direction is:

[0106] s θ =[Contrast θ Correlation θ Energy θ Homogeneity θ Entropy θ (25)

[0107] By concatenating the statistical feature vectors in the four directions θ = 0°, θ = 45°, θ = 90°, and θ = 135°, the texture feature vector of the linear frequency modulated signal is obtained:

[0108] S f =[s θ＝0° ,s θ＝45° ,s θ＝90° ,s θ＝135° (26)

[0109] Step 6: Train the support vector machine classification model.

[0110] Based on steps one through five, the texture feature vectors of the linear frequency modulated signal training samples are calculated, and a support vector machine classification model is trained.

[0111] Step 7: Use the support vector machine classification model obtained in Step 6 to sort and identify the linear frequency modulated signal.

[0112] Based on the texture feature vectors of the received linear frequency modulated (LFM) signals calculated in steps one through five, the support vector machine classification model obtained in step six is ​​used for classification. The classification results correspond to the modulation type of the received LFM signals, thereby achieving the sorting and identification of LFM signals.

[0113] The recognition accuracy of linear frequency modulated signals at different signal-to-noise ratios is as follows: Figure 7 As shown, the accuracy of linear frequency modulation (LFM) signal recognition can reach 96% under a signal-to-noise ratio (SNR) of -12dB, and the accuracy of LFM signal recognition is greater than 96% and close to 100% under an SNR greater than -12dB.

[0114] Example 2

[0115] To verify the feasibility of the method, a pseudo-random code and linear frequency modulation composite modulation signal with a signal-to-noise ratio of -10dB, a carrier frequency of 50MHz, a modulation frequency of 200kHz, a modulation bandwidth of 100MHz, a pseudo-code level of 4, and a symbol width of 50ns were selected, and 65536 data points were obtained using a sampling rate of 1GHz.

[0116] like Figure 1 As shown in the figure, this example discloses a modulation type recognition method based on adaptive noise reduction and texture feature extraction. The specific implementation steps are as follows:

[0117] Step 1: Use the DR2D preprocessing method to calculate the two-dimensional joint frequency characteristics of the pseudo-random code and the linear frequency modulation composite modulation signal.

[0118] The 256×256 point radiation source signal is divided into 256 sub-signal segments of 256 points each. Each sub-signal segment is used as a row of a two-dimensional matrix, and the columns are stacked in order to obtain a two-dimensional data rearrangement matrix X of size 256×256 for the composite modulation signal of pseudo-random code and linear frequency modulation.

[0119] The spectrum matrix obtained by performing a two-dimensional fast Fourier transform on matrix X is:

[0120]

[0121] In the formula, (k,l) and (m,n) are the indices of the original matrix X and the spectrum matrix Y, respectively, and k,m=0,1,···,255, l,n=0,1,···,255.

[0122] Taking the modulus of the spectrum matrix and normalizing it, the two-dimensional joint frequency characteristics of the pseudo-random code and linear frequency modulation composite modulation signal are obtained as follows:

[0123]

[0124] Two-dimensional joint frequency characteristics of pseudo-random code and linear frequency modulation composite modulation signal, as follows Figure 8 As shown.

[0125] Step 2: Adaptively determine the processing threshold for threshold filtering based on the noise level, and perform threshold filtering on the two-dimensional joint frequency characteristics of the pseudo-random code and linear frequency modulation composite signal.

[0126] The two-dimensional joint frequency feature image obtained in step one is cut using any threshold plane to obtain a feature image slice under that threshold thr. If a feature image is cut in the plane containing the threshold, the pixel value of the feature image slice at the corresponding position is recorded as 1; otherwise, it is recorded as 0. The feature image slice is then represented as follows:

[0127]

[0128] Frequency feature proportion Represented as:

[0129]

[0130] The frequency feature ratio is the ratio of the sum of pixel values ​​of the feature image slices under any threshold to the number of pixels in the two-dimensional joint frequency feature image.

[0131] The threshold range is set to [0,1], with a step size of 0.01. The frequency characteristic proportions under different thresholds are obtained, and their variation curves are shown below. Figure 9 As shown, the slope of the curve changes significantly before and after, indicating whether the threshold plane is within the amplitude range of the noise component. The threshold corresponding to the position where the frequency feature proportion is less than 0.02 and the derivative of the frequency feature proportion is greater than -0.002 is selected as the reference threshold thr0 for filtering.

[0132] To prevent excessive loss of useful signals, the reference threshold is shrunk, resulting in the following processing threshold:

[0133]

[0134] In the formula, α is the convergence coefficient, and its value ranges from [1,2].

[0135] Taking α = 1.5, the filtering threshold thr for the composite modulation signal of pseudo-random code and linear frequency modulation under a signal-to-noise ratio of -10dB is obtained. c =0.24.

[0136] Use the processing threshold thr c Threshold filtering is applied to the two-dimensional joint frequency characteristics of the composite modulation signal of pseudo-random code and linear frequency modulation.

[0137] Step 3: Apply the processing threshold thr from Step 2 c The useful signal frequency band is determined and narrowband filtering is performed to obtain the two-dimensional joint frequency characteristics of the pseudo-random code and the linear frequency modulation composite signal after noise reduction.

[0138] Use the filter threshold thr obtained in step two c The two-dimensional joint frequency features of the pseudo-random code and linear frequency modulation composite signal are binarized to obtain the feature image slice S under the given threshold. c Considering that the frequency characteristics of the useful signal are distributed vertically, the ratio of the sum of the values ​​in each column of the binary distribution image to the column length is calculated to form the frequency characteristic distribution vector of the signal, denoted as:

[0139]

[0140] When the element value of the frequency characteristic distribution vector is less than a preset threshold for the frequency characteristic distribution vector, the index range of the element value is the out-of-band noise distribution; when the element value of the frequency characteristic distribution vector is greater than the preset threshold for the frequency characteristic distribution vector, the index range of the element value is the in-band useful signal distribution; the frequency characteristic distribution vector F is adjusted according to the preset threshold for the frequency characteristic distribution vector. c Binarization is performed, and the frequency features are copied and arranged column-wise according to their magnitude to obtain a frequency feature distribution mask. The threshold of the frequency feature distribution vector is set to 0.1 times the maximum value of the frequency feature distribution vector, resulting in the frequency feature distribution mask as shown below. Figure 10 As shown.

[0141] By zeroing out-of-band noise components of the two-dimensional joint frequency feature image using a frequency feature distribution mask, narrowband filtering is achieved, resulting in the two-dimensional joint frequency features of the denoised pseudo-random code and the linear frequency modulation composite signal.

[0142] Step 4: Obtain a two-dimensional rearranged grayscale image of the composite modulation signal of pseudo-random code and linear frequency modulation.

[0143] A two-dimensional inverse fast Fourier transform is performed on the two-dimensional joint frequency features of the pseudo-random code and the linear frequency modulated composite signal after threshold filtering and narrowband filtering. This yields a two-dimensional rearranged matrix of the denoised pseudo-random code and the linear frequency modulated composite signal. Further normalization yields the two-dimensional rearranged grayscale image, as shown below.Figure 11 As shown.

[0144] Step 5: Solve for the gray-level co-occurrence matrix of the two-dimensional rearranged grayscale image of the pseudo-random code and the linear frequency modulation composite signal obtained in Step 4, and extract the texture feature vector of the pseudo-random code and the linear frequency modulation composite signal.

[0145] The gray-level co-occurrence matrices of the two-dimensional rearranged gray-level image obtained in step four are calculated in the directions θ = 0°, θ = 45°, θ = 90°, and θ = 135°, respectively, as shown below. Figure 12 As shown in a), 12b), 12c) and 12d).

[0146] Assuming the element value at index (a,b) of the gray-level co-occurrence matrix is ​​p(a,b), then the contrast, correlation, energy, homogeneity, and entropy of the two-dimensional rearranged gray-level image are as follows:

[0147]

[0148]

[0149]

[0150]

[0151]

[0152] Where, μ a and μ b σ represents the mean of the image grayscale values ​​in the row and column directions, respectively. a and σ b These are the standard deviations of the image grayscale values ​​in the row and column directions, respectively.

[0153] The statistical feature vector of a two-dimensional rearranged grayscale image in the θ direction is:

[0154] s θ =[Contrast θ Correlation θ Energy θ Homogeneity θ Entropy θ (38)

[0155] By concatenating the statistical feature vectors in the four directions of θ = 0°, θ = 45°, θ = 90°, and θ = 135°, the texture feature vector of the pseudo-random code and linear frequency modulation composite modulation signal is obtained:

[0156] S f =[s θ＝0° ,s θ＝45° ,sθ＝90° ,s θ＝135° (39)

[0157] Step 6: Train the support vector machine classification model.

[0158] Based on steps one through five, the texture feature vectors of the training samples of the pseudo-random code and linear frequency modulation composite signal are calculated, and the support vector machine classification model is trained.

[0159] Step 7: Use the support vector machine classification model obtained in Step 6 to sort and identify the pseudo-random code and linear frequency modulation composite signal.

[0160] Based on the texture feature vectors of the received pseudo-random code and linear frequency modulated composite signal calculated in steps one through five, the support vector machine classification model obtained in step six is ​​used for classification. The recognition accuracy of the pseudo-random code and linear frequency modulated composite signal under different signal-to-noise ratios is obtained as follows: Figure 13 As shown, under a signal-to-noise ratio of -20dB to 0dB, the recognition accuracy of pseudo-random codes and linear frequency modulation composite modulation signals both reach over 98%, indicating that they also have excellent recognition performance for composite modulation type radiation source signals.

[0161] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A modulation type recognition method based on adaptive noise reduction and texture feature extraction, characterized in that: Includes the following steps, Step 1: Calculate the two-dimensional joint frequency characteristics Y(m,n) of the radiation source signal using the DR2D preprocessing method; Step 2: Adaptively determine the threshold for threshold filtering based on the noise level. c And threshold filtering is applied to the two-dimensional joint frequency features; The second step is implemented as follows: The two-dimensional joint frequency feature image obtained in step one is cut using a certain threshold plane to obtain a feature image slice under the threshold thr. If a feature image is cut in the plane where the threshold is located, the pixel value of the feature image slice at the corresponding position is recorded as 1, otherwise it is 0. The feature image slice is then represented as follows: Frequency feature proportion Represented as: The frequency feature ratio is the ratio of the sum of pixel values ​​of the feature image slices under any threshold to the number of pixels in the two-dimensional joint frequency feature image. The two-dimensional joint frequency feature image is cut by a threshold plane corresponding to a set of thresholds at a certain step value within a predetermined range, and a set of frequency feature proportion values ​​are obtained. Based on the frequency feature proportion values ​​and the differential value of the frequency feature proportion, an appropriate corresponding threshold is selected as the reference threshold thr0. To prevent excessive loss of useful signals, the reference threshold is shrunk, resulting in the following processing threshold: In the formula, α is the convergence coefficient; Use the processing threshold thr c Threshold filtering is applied to the two-dimensional joint frequency characteristics; Step 3: Use the processing threshold thr from Step 2 c The useful signal frequency band is determined and narrowband filtering is performed to obtain the two-dimensional joint frequency characteristics after noise reduction; Step 4: Perform a two-dimensional fast Fourier inverse transform on the denoised two-dimensional joint frequency features obtained in Step 3 to obtain the two-dimensional rearranged matrix of the denoised radiation source signal. Further normalize to obtain the two-dimensional rearranged grayscale image, that is, obtain the two-dimensional rearranged grayscale image of the radiation source signal. Step 5: Solve for the gray-level co-occurrence matrix of the two-dimensional rearranged gray-level image obtained in Step 4, and extract the texture feature vector of the radiation source signal; Step 6: Calculate the radiation source signals of different modulation types according to Steps 1 to 5, train the texture feature vectors of the training samples, and train the support vector machine classification model. Step 7: Use the support vector machine classification model obtained in Step 6 to sort and identify the modulation type of the radiation source signal.

2. The modulation type recognition method based on adaptive noise reduction and texture feature extraction as described in claim 1, characterized in that: The implementation method for step one is as follows: The M×N point radiation source signal is divided into N M-point sub-signal segments. Each sub-signal segment is used as a row of a two-dimensional matrix, and the columns are stacked in order to obtain a two-dimensional data rearrangement matrix X of size M×N. The spectrum matrix obtained by performing a two-dimensional fast Fourier transform on matrix X is: In the formula, (k,l) and (m,n) are the indices of the original matrix X and the spectrum matrix Y, respectively, and k,m=0,1,···,M-1, l,n=0,1,···,N-1; Taking the modulus of the spectrum matrix and normalizing it, we obtain the two-dimensional joint frequency characteristics as follows:

3. The modulation type recognition method based on adaptive noise reduction and texture feature extraction as described in claim 2, characterized in that: The method for implementing step three is as follows: Use the filter threshold thr obtained in step two c Binarizing the two-dimensional joint frequency features yields the feature image slice S at this threshold. c Considering that the frequency characteristics of the useful signal are distributed vertically, the ratio of the sum of the values ​​in each column of the binary distribution image to the column length is calculated to form the frequency characteristic distribution vector of the signal, denoted as: When the element value of the frequency characteristic distribution vector is less than a preset threshold for the frequency characteristic distribution vector, the index range of the element value is the out-of-band noise distribution; when the element value of the frequency characteristic distribution vector is greater than the preset threshold for the frequency characteristic distribution vector, the index range of the element value is the in-band useful signal distribution; the frequency characteristic distribution vector F is adjusted according to the preset threshold for the frequency characteristic distribution vector. c Binarization is performed, and the two-dimensional joint frequency features are copied and arranged in columns according to their magnitude to obtain a frequency feature distribution mask. The out-of-band noise components of the two-dimensional joint frequency feature image are zeroed according to the frequency feature distribution mask to achieve narrowband filtering and noise reduction of the two-dimensional joint frequency features.

4. The modulation type recognition method based on adaptive noise reduction and texture feature extraction as described in claim 3, characterized in that: Step five is implemented as follows: Calculate the gray-level co-occurrence matrix of a two-dimensional rearranged grayscale image in the directions θ = 0°, θ = 45°, θ = 90° and θ = 135°; If the element at index (a,b) of the gray-level co-occurrence matrix is ​​p(a,b), then the contrast, correlation, energy, homogeneity, and entropy of the two-dimensional rearranged gray-level image are as follows: Where, μ a and μ b σ represents the mean of the image grayscale values ​​in the row and column directions, respectively. a and σ b These are the standard deviations of the image grayscale values ​​in the row and column directions, respectively; The statistical feature vector of a two-dimensional rearranged grayscale image in the θ direction is: s θ =[Contrast θ ,Correlation θ ,Energy θ ,Homogeneity θ ,Entropy θ ] (12) By concatenating the statistical feature vectors in the four directions of θ = 0°, θ = 45°, θ = 90°, and θ = 135°, the texture feature vector of the radiation source signal is obtained: S f =[s θ＝0° ,s θ＝45° ,s θ＝90° ,s θ＝135° ] (13)。 5. The modulation type recognition method based on adaptive noise reduction and texture feature extraction as described in claim 4, characterized in that: Step seven is implemented as follows: Based on the texture feature vectors of the received radiation source signals calculated in steps one through five, the support vector machine classification model obtained in step six is ​​used for classification. The classification results correspond to the modulation type of the received radiation source signals, thereby achieving the sorting and identification of the modulation type of the radiation source signals.

Images