A method for detecting spread spectrum signals based on L-ResNet

Abstract

The application discloses a direct spread signal detection method based on L-ResNet and belongs to the wireless communication field. First, according to the characteristics of the direct spread signal, a to-be-detected signal matrix model is constructed, and a characteristic value vector is obtained by performing eigenvalue decomposition operation on the detection signal matrix. Second, the characteristic value vector is subjected to superposition processing according to the principal component analysis method, and the superposed characteristic value vector is drawn into a characteristic value column chart. Finally, an L-ResNet model is built, the local binary pattern (LBP) is introduced to extract the texture features of the picture, the similarity between the texture feature blocks is calculated by using the LBP, the loss function of the network model is constrained, and the direct spread signal detection is performed. The application has better anti-interference performance and better detection performance.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of wireless communication, and mainly relates to a direct spread signal detection method based on a local binary pattern residual network (L-ResNet). BACKGROUND

[0002] In modern wireless communication, in order to ensure the high confidentiality and strong anti-interference transmission requirements of data, spread spectrum communication emerges as the times require. Direct sequence spread spectrum communication refers to that the sender uses a spread spectrum code symbol rate higher than the baseband signal symbol rate to transmit information. After such processing, the signal bandwidth is increased while the signal power spectral density is reduced, and secret transmission in noise is realized. Code division multiple access can also be realized by using mutually orthogonal spread spectrum codes, which plays an important role in the third generation of wireless mobile communication, saves spectrum resources, and increases the information transmission rate of the system.

[0003] Under non-cooperative communication, when the signal-to-noise ratio decreases, the performance of the traditional direct spread signal detection method becomes poor, and it cannot meet the detection performance requirements. In recent years, the development of deep learning algorithms has been rapid, and the communication field has gradually combined with deep learning algorithms. Deep learning algorithms use the excellent convolutional computing power of deep learning networks to achieve better performance in solving problems in the field of communication. Then scholars proposed using deep learning methods to realize direct spread signal detection, and the detection performance is better than that of traditional detection methods. However, when the signal-to-noise ratio decreases, the input data set picture deteriorates, resulting in poor detection performance of the network model. Principal component analysis (PCA) is a commonly used dimensionality reduction method in signal processing and image recognition. The basic idea is to use a set of independent principal components to reconstruct the original information data through linear combination of these principal components, and minimize the error between the reconstructed information data and the original information data. Based on this, the direct spread signal detection is combined with the deep learning algorithm, and the strong mapping ability of the deep learning algorithm is used to achieve better detection effect, which becomes a breakthrough in signal detection algorithm. SUMMARY

[0004] The object of the present application is to optimize the detection performance of direct spread signal in fading channel with low signal-to-noise ratio, and a direct spread signal detection method based on L-ResNet is proposed. According to the characteristics of direct spread signal, a signal matrix model is constructed, and the eigenvalues of the signal covariance matrix are obtained by eigenvalue decomposition and drawn into eigenvalue pictures. The energy contrast and gradient change difference of eigenvalues are used to distinguish the presence of direct spread signal. By building an L-ResNet model, the local binary pattern (LBP) is introduced to extract the texture features of the picture, increase the distinguishability of the eigenvalue histogram of direct spread signal in the gradient, and calculate the similarity between texture feature blocks to constrain the loss function of the network model and improve the detection ability of the network model. Theoretical analysis and simulation experiments show that the detection performance of the direct spread signal detection method based on eigenvalue histogram and L-ResNet is better than that of E-D, FFT-DNN, CM-CNN and ResNet. Under different signal-to-noise ratios and different false alarm probabilities, the detection performance of the direct spread signal detection method based on eigenvalue histogram and L-ResNet is better than that of E-D, FFT-DNN, CM-CNN and ResNet without LBP constraint.

[0005] The technical solution adopted by the present application to solve its technical problems comprises the following steps:

[0006] Step 1, according to the generation mode of direct spread signal, a mathematical model of the detected signal is constructed, and the eigenvalue vector is obtained by eigenvalue decomposition operation on the detected signal matrix.

[0007] Step 2, according to the principal component analysis method, the eigenvalue vector is superimposed, and the superimposed eigenvalue vector is drawn into an eigenvalue histogram.

[0008] Step 3, build an L-ResNet model, extract the texture features of the eigenvalue histogram by introducing the local binary pattern (LBP), increase the distinguishability of the eigenvalue histogram of direct spread signal in the gradient, calculate the similarity between texture feature blocks to constrain the loss function of the network model, and improve the detection ability of the network model.

[0009] Step 4, construct eigenvalue histogram data set under different false alarm probabilities and different signal-to-noise ratios, take the eigenvalue histogram as the input of the network model, train the network model, test the direct spread signal detection performance of the network model after the loss of the network model converges, and obtain the probability of direct spread signal detection of the network model.

[0010] Step 1 is implemented as follows:

[0011] 1-1. For the presence detection of DSSS signals in wireless communication, consider the case where there is no Rayleigh fading. Assume that there is a DSSS signal in the signal to be detected. The mathematical model of the asynchronous multi-user signal to be detected containing K user DSSS signals is:

[0012]

[0013] where T1 is the sampling period, s k (nT1) is the DSSS signal component of the kth user, τ k is the signal delay of the kth user, ρ k is the signal amplitude of the kth user, b k (nT1) is the information sequence of the kth user signal, c k (nT1) is the spreading sequence of the kth user, and N(nT1) is a component of the additive Gaussian white noise with mean 0 and variance σ 2 .

[0014] M observation vectors are used for DSSS signal detection. The mth observation vector is:

[0015] r m = [r((m-1)L+1+τ k ), r((m-1)L+2+τ k ),..., r(mL+τ k )] T (m = 1, 2, 3,..., M) (2)

[0016] where L is the pseudo-code period. According to equation (1) and equation (2), the observation signal vector X m with a vector length of 2L is defined as:

[0017] X m = [r m Τ , r m+1 Τ ] Τ = Gb m + v m (3)

[0018] In the equation, G is a mixed signal matrix containing DSSS sequence modulation, b m is the mth information code vector of the user, and v m is the Gaussian white noise vector at the sampling time of the mth information code. Where G and b m can be expressed as:

[0019] G = [ρ1C 1,e , ρ1C 1,l ,..., ρ K CK,e ,ρ K C K,l ] 2L×2K (4)

[0020]

[0021] G and b of the kth user signal m are expanded to obtain:

[0022]

[0023] In formula (6) and formula (7), C k,e and C k,l are two channel vectors of the kth user in the detection signal respectively, b m,k,e and b m,k,l are two signal vectors of the mth data in the kth user respectively, each detection signal corresponds to two signal subspaces, when the detection signal does not contain a direct spread signal, two subspaces contained in the detection signal are both noise subspaces. The covariance matrix of the observation vector is calculated at this time, and eigenvalue decomposition is performed on the covariance matrix to obtain:

[0024]

[0025] In the formula, the column vectors of U S and U N correspond to the signal subspace and the noise subspace respectively, D S and D N correspond to the eigenvalues of the signal and noise subspaces respectively, O is a zero matrix, the number of eigenvalues is 2L, E k,e =||C k,e || / L and E k,l =||C k,l || / L are the energies of the pseudo code sequence, ||.|| is the 2-norm of the orientation vector, μ k,e =C k,e / ||C k,e || and μ k,l =C k,l / ||C k,l || are normalized pseudo code vectors, and I is a unit matrix. The signal space of each user is divided into two uncorrelated subspaces, and the sample covariance matrix of the signal is composed of 2K signal subspaces and noise subspaces, so the eigenvalue of the kth user can be approximately expressed as:

[0026]

[0027] wherein, for the kth user, λ k,e is the eigenvalue of the front part, λ k,l is the eigenvalue of the last part, and λN,n is the eigenvalue of the last noise subspace.

[0028] 1-2. In wireless communication, Rayleigh fading often exists. When considering the presence of Rayleigh fading, assuming that there is a direct spread signal in the detection signal, the mathematical model expression of the asynchronous multi-user containing K user direct spread signals to be detected is:

[0029]

[0030] where h k is the Rayleigh channel fading gain of the kth user, which is subject to Rayleigh distribution. According to the derivation result of equation (10), when there is Rayleigh fading, the eigenvalue decomposition of the detection signal covariance matrix is performed, and the result is:

[0031]

[0032] Step 2 is implemented as follows:

[0033] 2-1. When considering the absence of Rayleigh fading, according to the principal component analysis method, the sum of the first 2K eigenvalues of the detection signal accounts for most of the total eigenvalues, and at low signal-to-noise ratio, the signal components with small energy are easily affected by noise. Therefore, the first 2K principal component subspaces of the detection signal are selected, and other subspaces are equivalent to noise subspaces. According to equation (9), the first 2K eigenvalues are larger, and the last eigenvalue is smaller. The sum of the first 2K eigenvalues of the signal subspace is much larger than the eigenvalue of the noise subspace. Therefore, when constructing the column chart of the eigenvalues of K users, the superposition result of the first 2K eigenvalues is taken as a new first eigenvalue, and other eigenvalues remain unchanged. After the eigenvalue superposition, the mathematical expression of each eigenvalue in the eigenvalue column chart can be expressed as:

[0034]

[0035] where, for the kth user in the Rayleigh fading channel, λ1is the first user eigenvalue, λ k is the kth user eigenvalue, λ N,n is the last L-2K noise subspace eigenvalue.

[0036] When there is a direct spread signal, the first 2K eigenvalues are larger, and the contrast of the eigenvalues is larger. The first 2K eigenvalues are accumulated as the first new eigenvalue. After superimposing the eigenvalue vector, the superimposed eigenvalue vector is drawn into an eigenvalue column chart.

[0037] 2-2. When constructing the histogram of eigenvalues of K users' direct spread signals in the presence of Rayleigh fading, the superposition of the first 2K eigenvalues of the covariance matrix is taken as the new first eigenvalue, and the other eigenvalues remain unchanged. After the eigenvalues are added, the mathematical expression of each eigenvalue in the eigenvalue histogram can be expressed as:

[0038]

[0039] where, for the kth user in the Rayleigh fading channel, is the first user eigenvalue, is the kth user eigenvalue, is the eigenvalue of the last L-2K noise subspace. In the Rayleigh fading channel, when there is a direct spread signal, the first 2K eigenvalues are larger, and the first 2K eigenvalues are also superimposed into the new first eigenvalue. After superimposing the eigenvalue vector, the superimposed eigenvalue vector is drawn into an eigenvalue histogram.

[0040] Step 3 is implemented as follows:

[0041] In order to solve the problem that the feature extraction process of the residual network model does not pay enough attention to the texture features of the eigenvalue histogram, the L-ResNet model proposed by the application combines ResNet and LBP, adds LBP loss in the loss function calculation, uses local binary pattern LBP to extract the texture features of the feature map, increases the attention of the network model to the texture features of the eigenvalue histogram, and through the L s loss and LBP loss jointly constrain the network model. The loss function formula is shown in formula (14):

[0042] L total = L s + L lbp (14)

[0043] In the formula, L total is the total loss in the network model training process, L s is the cross-entropy loss, and L lbp is the texture loss. The calculation formula of L s is shown in formula (15):

[0044]

[0045] In the formula, N is the number of a batch of images in the training set, n is the number of classes of eigenvalue histogram images in the training set, f i is the feature vector of the ith eigenvalue histogram image, y i is the corresponding class of f i , is the corresponding class y i of f iThe weight matrix of the residual network is W j is the jth column of the weight matrix W, b j is a bias term.

[0046] In order to extract the fine texture features of the feature value column chart, the texture features of the residual network feature map are extracted using LBP, and the texture loss constraint of the network model is obtained by calculating the triplet loss of the texture features. lbp The calculation formula of is shown in formula (16):

[0047]

[0048] In the formula, is the ith residual network feature map for calculating the texture loss, is the ith residual network feature map containing the direct spread signal or not containing the direct spread signal, is the ith residual network feature map containing the direct spread signal or not containing the direct spread signal, and alpha is a difference constant between positive and negative, f lbp is an LBP texture feature extraction function, and it is assumed that the residual network feature map The LBP value calculation formula is shown in formula (17):

[0049]

[0050] In the formula, Omega c is the neighborhood of the target pixel value (p c , q c ), the neighborhood radius is 1, and s(·) is a sign function. The calculation formula f lbp (·) of any residual network feature map is the same. The calculation process of the LBP value is as follows: the target pixel value (p c , q c ) is compared with the surrounding pixel values, if the current pixel value is greater than the neighborhood target pixel value, the current pixel value is determined as 1, otherwise it is determined as 0, then the determination values of all pixel points in the neighborhood are obtained, the obtained binary determination values are converted into decimal numbers, and the LBP value of the neighborhood is obtained, wherein d is the number of binary determination values.

[0051] LBP has good ability to represent image texture features, and after the feature map of the residual network is subjected to LBP texture feature extraction, the entire network model is added by LBP loss constraint, and the classification detection ability of the network model for the direct spread signal is enhanced.

[0052] The beneficial effects of the present application are as follows:

[0053] The application provides a direct spread signal detection method based on a signal eigenvalue column chart and L-ResNet combination. The method adopts an LBP residual network model and reasonably designs a loss function. The method is mainly used for signal detection of asynchronous multi-users and asynchronous multi-users under Rayleigh fading. Through a main feature analysis method, eigenvalue vectors are superimposed and processed. The superimposed eigenvalue vectors are drawn into an eigenvalue column chart to highlight the eigenvalue size and eigenvalue gradient change. A residual network is trained to establish a precise mapping relationship between a direct spread signal and a direct spread signal-free signal, so that the direct spread signal is detected. The mapping relationship between a mixed signal containing a direct spread signal and a target signal without a direct spread signal is established to suppress the direct spread signal. Theoretical analysis and simulation experiments show that the network model method based on the eigenvalue column chart as input and the LBP constraint is introduced has better anti-interference performance than E-D, FFT-DNN, CM-CNN and ResNet algorithm without LBP constraint under low signal-to-noise ratio Rayleigh fading. Therefore, the residual network model with the LBP constraint introduced has a very wide application prospect in asynchronous two-user signal detection.

[0054] 1. The application provides a new idea for detecting a direct spread signal, converts a detection problem into a mapping problem of an image, and has better detection performance than E-D, FFT-DNN, CM-CNN and ResNet model without LBP constraint.

[0055] 2. Compared with E-D, FFT-DNN and CM-CNN model, under the condition of the same signal-to-noise ratio and different false alarm probabilities, the detection probability of the ResNet model without LBP constraint is improved by at least 33% for asynchronous two-user signal detection, and the detection probability of the L-ResNet model is improved by 5% compared with the detection probability of the ResNet model without LBP constraint. Under the Rayleigh fading channel, under the condition of the same signal-to-noise ratio and different false alarm probabilities, the detection probability of the ResNet model without LBP constraint is improved by at least 73% for asynchronous two-user signal detection, and the detection probability of the L-ResNet model is improved by 4% compared with the detection probability of the ResNet model without LBP constraint.

[0056] 3. Compared with the E-D, FFT-DNN and CM-CNN models, the detection probability of the ResNet model without LBP constraint is improved by at least 64% for asynchronous two-user signal detection under the same false alarm probability and different signal-to-noise ratios, and the detection probability of the L-ResNet model is improved by 5% compared with the detection probability of the ResNet model without LBP constraint. Under the Rayleigh fading channel, the detection probability of the ResNet model without LBP constraint is improved by at least 52% for asynchronous two-user signal detection under the same false alarm probability and different signal-to-noise ratios, and the detection probability of the L-ResNet model is improved by 4% compared with the detection probability of the ResNet model without LBP constraint. BRIEF DESCRIPTION OF DRAWINGS

[0057] Fig. 1(a) is a characteristic value column chart when there is an asynchronous multi-user signal;

[0058] Fig. 1(b) is a characteristic value column chart when there is no asynchronous multi-user signal;

[0059] Fig. 2(a) is a network feature visualization chart of an FFT spectrum chart when there is a direct spread signal;

[0060] Fig. 2(b) is a network feature visualization chart of an FFT spectrum chart when there is no direct spread signal;

[0061] Fig. 3(a) is a network feature visualization chart of a covariance matrix chart when there is a direct spread signal;

[0062] Fig. 3(b) is a network feature visualization chart of a covariance matrix chart when there is no direct spread signal;

[0063] Fig. 4(a) is a network feature visualization chart of a characteristic value column chart when there is a direct spread signal without LBP constraint;

[0064] Fig. 4(b) is a network feature visualization chart of a characteristic value column chart when there is no direct spread signal without LBP constraint;

[0065] Fig. 4(c) is a network feature visualization chart of a characteristic value column chart when there is a direct spread signal with LBP constraint;

[0066] Fig. 4(d) is a network feature visualization chart of a characteristic value column chart when there is no direct spread signal with LBP constraint;

[0067] Figure 5 Fig. 5 is a residual block structure diagram;

[0068] Figure 6 Fig. 6 is a LBP calculation process diagram;

[0069] Figure 7 Fig. 7 is an ROC curve for asynchronous two-user signal detection;

[0070] Figure 8ROC curve of asynchronous two-user signal detection under Rayleigh fading;

[0071] Figure 9 SNR-Pd curve of asynchronous two-user signal detection;

[0072] Figure 10 SNR-Pd curve of asynchronous two-user signal detection under Rayleigh fading. DETAILED DESCRIPTION

[0073] The specific embodiments of the present application will be further described below in conjunction with the accompanying drawings.

[0074] Step 1, the asynchronous multi-user signal containing multiple direct spread signals, construct mathematical models containing asynchronous multi-user signals and not containing asynchronous multi-user signals, and obtain the eigenvalue vector by performing eigenvalue decomposition operation on the detection signal matrix.

[0075] Step 2, according to the principal component analysis method, the eigenvalue vector of the obtained signal is stacked, and the stacked eigenvalue vector is drawn into a column chart, and the column chart is drawn as shown in Figures 1(a) and 1(b).

[0076] Step 3, the eigenvalue column chart drawn by the asynchronous multi-user signal and the asynchronous multi-user signal is taken as the input of the L-ResNet network, and the L-ResNet detection model is obtained by training according to the model structure shown in Table 1, until the loss function converges, and finally the test set data is input into the trained network model for testing.

[0077] Table 1

[0078]

[0079]

[0080] Step 4, construct eigenvalue column data sets under different false alarm probabilities and different signal-to-noise ratios, take the eigenvalue column chart as the input of the network model, train the network model, and test the network model after the network model loss converges. The performance of the direct spread signal detection of the network model is obtained, and the probability of the network model detecting the direct spread signal is obtained.

[0081] Steps 1 and 2 are implemented as follows:

[0082] The covariance matrix of the observation matrix containing asynchronous multi-user signals is calculated, the eigenvalue decomposition is performed on the covariance matrix, the eigenvalue vector of the detected signal is superimposed, and the superimposed eigenvalue vector is drawn into a column chart to make a picture as shown in Fig. 1(a), wherein the pseudo code rate is 127 kbps, the sampling frequency is 127 kHz, the information code symbol rate is 1 kbps, the pseudo code period is 127, and the information code symbol length is 100. Finally, a picture with a size of 224*224 is drawn. It can be seen that the first two eigenvalues are large, account for a large part of the entire eigenvalue vector, and the eigenvalue gradient changes obviously.

[0083] The covariance matrix of the observation matrix not containing asynchronous multi-user signals is calculated, the eigenvalue decomposition is performed on the covariance matrix, the eigenvalue vector of the detected signal is superimposed, and the superimposed eigenvalue vector is drawn into a column chart to make a picture as shown in Fig. 1(b), wherein the pseudo code rate is 127 kbps, the sampling frequency is 127 kHz, the information code symbol rate is 1 kbps, the pseudo code period is 127, and the information code symbol length is 100. Finally, a picture with a size of 224*224 is drawn. It can be seen that the first two eigenvalues are small, and the eigenvalue gradient does not change obviously.

[0084] According to the network model and the data set picture, the L-ResNet model is used to extract the features of the two types of pictures, and the visualized results of the last layer output feature mapping of the network model after convergence of the three methods of the application, FFT-DNN and CM-CNN are drawn. Fig. 2(a) is a network feature visualization chart of the FFT spectrum chart with signals, and Fig. 2(b) is a network feature visualization chart of the FFT spectrum chart without signals; Fig. 3(a) is a network feature visualization chart of the covariance matrix chart with signals, and Fig. 3(b) is a network feature visualization chart of the covariance matrix chart without signals.

[0085] Step 3 is specifically implemented as follows:

[0086] The network structure of the L-ResNet is shown in Table 1. Wherein [·] represents a residual block, and the internal network structure of the residual block is shown in Table 2. Figure 5 The L-ResNet model combines ResNet and LBP. ResNet avoids the network gradient disappearance problem caused by network layer stacking, uses local binary pattern LBP to extract the texture features of the feature map, and improves the feature extraction and classification recognition ability of the network model through LBP loss constraint. As shown in Table 3. Figure 6As shown, the calculation process of the LBP value is as follows: the value of a target pixel point is compared with the values of surrounding pixel points, if the value of the current pixel point is greater than the value of the target pixel point, the value of the current pixel point is set to 1, otherwise, it is set to 0, then the decision values of 8 pixel points are obtained, a binary number 11110010 is obtained, 8-bit binary number is converted into decimal number 242, and the LBP value of the neighborhood is obtained.

[0087] Step 4 is specifically implemented as follows:

[0088] 3-2. The eigenvalue histogram of the asynchronous multi-user signal shown in FIG. 1(a) and FIG. 1(b) is drawn as a detection model input, and the model is trained, and an L-ResNet detection model is obtained by training.

[0089] 3-3. The signal picture of the test set is input into the trained network model, and finally the eigenvalue histogram of the asynchronous multi-user signal is detected.

[0090] Embodiment:

[0091] In this experiment, the GPU is NVIDIA RTX 2080, and the deep learning framework used in the training process is Tensorflow 2.1. The signal used in the experiment is a direct spread signal with a pseudo code rate of 127 kbps, a sampling frequency of 127 kHz, an information code rate of 1 kbps, a pseudo code period of 127, and an information code length of 100. For asynchronous multi-user signal detection, there are 2 user signals superimposed, and the user signals have random time delay. The size of the covariance matrix is 100 information code lengths. The asynchronous multi-user signal and the non-asynchronous multi-user signal are used as training data, and the one-dimensional signal is converted into a 224x224 eigenvalue histogram on the Matlab 2019a simulation platform.

[0092] As can be seen from the feature maps of FIG. 2(a) and FIG. 2(b), the distributions of the two types of image features are very similar, and the network model extracts interference and similar features in the background area, which makes the DSSS signal detection performance of the network model poor. As can be seen from the feature maps of FIG. 3(a) and FIG. 3(b), the network model does not focus on the signal part on the diagonal line of the two images, but extracts similar interference features from the noise part on both sides of the diagonal line, which reduces the DSSS signal detection performance of the network model.

[0093] As can be seen from the feature maps of FIGS. 4(a) and 4(b), the feature distributions of the two images are different. The network model focuses on the signal area of the two types of feature value histograms, focuses on the feature value area, and extracts different features from the feature value area. As can be seen from the feature maps of FIGS. 4(c) and 4(d), the feature distributions of the two images are quite different. The network model pays more attention to the signal area features of the two types of feature value histograms. FIG. 4(c) focuses on the signal feature value area and the noise feature value area of the feature value histogram, while FIG. 4(d) only focuses on the noise feature value area of the feature value histogram. The features extracted from the signal feature value area are more obvious.

[0094] According to the network feature map visualization of the three methods, the ResNet model without LBP module constraint proposed in the present application improves the contrast of the two types of images, and is affected by the background area attention, while the L-ResNet model further improves the contrast of the two types of images, and is less affected by the background area attention.

[0095] The detection performance of the test model under the same signal-to-noise ratio and different false alarm probabilities, the network is trained under different false alarm probabilities with asynchronous two-user signals, after the network converges, the corresponding test set is tested, and the detection probability is counted. The training set and the test set are generated by MATLAB 2019a simulation to generate asynchronous two-user signals. The training set is 4000 feature value histograms drawn from asynchronous two-user signals, and the test set is 1000 feature value histograms drawn from asynchronous two-user signals. The asynchronous two-user signals and the asynchronous two-user signals under Rayleigh fading are used as the input and output of the model for training, and the test results are as shown in Figure 7 and Figure 8 Figure 7 and Figure 8 In the above, the abscissa is the false alarm probability, and the ordinate is the detection probability. As can be seen from the ROC curves of the five detection methods of Figure 7 Under the condition of signal-to-noise ratio -20dB and false alarm probability 1e-3, the detection probability of the feature value histogram of the present application without LBP module constraint is 87%, while the detection probability of the traditional E-D is 11%, the detection probability of CM-CNN is 41%, and the detection probability of FFT-DNN is 54%, which proves that the present application has better detection performance under low signal-to-noise ratio. The detection probability of the L-ResNet model is 92%, which is 5% higher than that of the ResNet model without LBP module constraint, which shows that the method of the present application improves the detection performance of the network model after introducing the LBP constraint network model. As can be seen from Figure 8 ​The ROC curves of the five detection methods of the application can know that, under the signal-to-noise ratio of -20 dB, when the false alarm probability is 1e-3, for asynchronous two-user signal detection under Rayleigh fading, the detection probability of the ResNet model adopting the eigenvalue histogram of the application without the LBP module constraint is 74%, while the detection probability of the traditional E-D is 0.01%, the detection probability of the CM-CNN is 0.03%, and the detection probability of the FFT-DNN is 0.04%, which indicates that the eigenvalue histogram of the application has better detection performance for asynchronous multi-user signal detection under low signal-to-noise ratio and Rayleigh fading. The detection probability of the L-ResNet model is 78%, which is 4% higher than that of the ResNet model without the LBP module constraint, which proves that the method of the application improves the detection performance of the network model for asynchronous multi-user signal detection under low signal-to-noise ratio and Rayleigh fading after introducing the LBP constraint network model.

[0096] The detection performance of the test model under the same false alarm probability and different signal-to-noise ratios, under different signal-to-noise ratios, the network is trained with asynchronous two-user signals, and after the network converges, the corresponding test set is used for testing, and the detection probability is counted. Among them, the false alarm probability is 1e-3, the signal-to-noise ratio is -22dB to -16dB with an interval of 1dB, the detection probability under each signal-to-noise ratio is counted, and compared with the E-D, FFT-DNN, CM-CNN and ResNet 4 detection methods, the test results are as shown in Figure 9 and Figure 10 , Figure 9 and Figure 10 , wherein the abscissa is the signal-to-noise ratio, and the ordinate is the detection probability. From the detection curves of the five methods under the false alarm probability of 1e-3 in Figure 9 , for asynchronous two-user signal detection, the detection probability of the L-ResNet model of the application is 68% when the signal-to-noise ratio is -22dB, which is better than 0.02% of the traditional E-D, 0.1% of the CM-CNN, and 4% of the FFT-DNN, and the detection rate of the L-ResNet model of the application is improved by 5% than the ResNet model without the LBP module constraint. From the detection curves of the five methods under the false alarm probability of 1e-3 in Figure 10 , for asynchronous two-user signal detection under Rayleigh fading, the detection probability of the L-ResNet model of the application is 53% when the signal-to-noise ratio is -22dB, which is better than 0.01% of the traditional E-D, 0.03% of the CM-CNN, and 0.07% of the FFT-DNN, and the detection probability of the L-ResNet model of the application is improved by 4% than the ResNet model without the LBP module constraint.

[0097] In summary, the algorithm is based on the L-ResNet model for asynchronous multi-user detection, which has better performance than the traditional E-D, FFT-DNN, CM-CNN and ResNet without LBP constraint, and still has good coping strategies under different signal-to-noise ratios and different false alarm probabilities.

[0098] Finally, it should be noted that the purpose of the embodiments disclosed is to help further understand the present application, but those skilled in the art can understand that various replacements and modifications are possible without departing from the spirit and scope of the present application and the appended claims. Therefore, the present application should not be limited to the disclosed embodiments, and the scope of the present application is defined by the scope of the claims.

Claims

1. A method for detecting a direct spread signal based on L-ResNet, characterized in that, The method comprises the following steps: Step 1, constructing a mathematical model of the detection signal according to the generation mode of the direct spread signal, and obtaining an eigenvalue vector by performing eigenvalue decomposition operation on the detection signal matrix; Step 2, superimposing the eigenvalue vector by the principal component analysis method, and drawing the superimposed eigenvalue vector into an eigenvalue histogram; Step 3, building a local binary pattern residual network L-ResNet model, introducing the local binary pattern LBP into the residual network ResNet model to extract the texture features of the eigenvalue histogram, calculating the similarity between the texture feature blocks by using the LBP, and constraining the loss function of the L-ResNet model; The L-ResNet model combines the ResNet and the LBP, and adds the LBP loss in the loss function calculation, and the loss function formula is shown as formula (14): L total = L s + L lbp (14) In the formula, L total is the total loss in the L-ResNet model training process, L s is the cross-entropy loss, L lbp is the texture loss; The texture loss L lbp The texture features of the residual network feature map are extracted using the LBP, and a ternary loss of the texture features is calculated as a texture loss constraint of the L-ResNet model; L lbp The calculation formula of L is shown as formula (16): In the formula, N is the number of a batch of images in the training set, is the i-th residual network feature map for calculating the texture loss at present, is the i-th residual network feature map for the same containing or not containing direct spread signal, is the i-th residual network feature map for the opposite containing or not containing direct spread signal, and α is a difference constant between positive and negative, f lbp (·) is an LBP texture feature extraction function, and the residual network feature map is selected as The calculation formula is shown in formula (17): where Ω c is the neighborhood of the target pixel value (p c , q c ), the neighborhood radius is 1, s(·) is the sign function, and d is the number of bits of the binary decision value; Step 4, constructing an eigenvalue histogram data set under different false alarm probabilities and different signal-to-noise ratios, taking the eigenvalue histogram as the input of the L-ResNet model, training the L-ResNet model, testing the direct spread signal detection performance of the L-ResNet model after the loss of the L-ResNet model converges, and obtaining the probability of the L-ResNet model in detecting the direct spread signal.

2. The L-ResNet-based direct spread signal detection method of claim 1, wherein, The step 1 is divided into the following two cases: 1-1, when there is no Rayleigh fading, the mathematical model expression of the asynchronous multi-user containing K user direct spread signals to be detected is: where T1 is the sampling period, s k (nT1) is the spread spectrum signal component of the kth user, τ k is the signal delay of the kth user, p k is the signal amplitude of the kth user, b k (nT1) is the information sequence of the kth user signal, c k (nT1) is the spreading sequence of the kth user, N(nT1) is a component of additive white Gaussian noise with mean 0 and variance σ 2 ​ M observation vectors are used for direct spread signal detection, and the mth observation vector is: r m = [r((m-1)L+1+τ k ),r((m-1)L+2+τ k ),...,r(mL+τ k )] T , m = 1, 2, 3,..., M (2) where L is the pseudo-code period, and according to equations (1) and (2), the observed signal vector X of length 2L is defined as m is: X m = [r m Τ ,r m+1 Τ ] Τ = Gb m + v m (3) where G is a mixed signal matrix containing direct spread sequence modulation, b m is the mth information code vector of the user, v m is the mth information code sampling time of the Gaussian white noise vector, where G and b m may be represented as: G=[ρ1C 1,e ,ρ1C 1,l ,...,r K C K,e ,r K C K,l ] 2L×2K (4) G and b for the kth user signal m Unfolding, we get: In formula (6) and formula (7), C k,e and C k,l are two channel vectors of the kth user in the detection signal respectively, b m,k,e and b m,k,l are two signal vectors of the mth data in the kth user, each detection signal corresponds to two signal subspaces, when the detection signal does not contain a direct spread signal, the two subspaces contained by the detection signal are both noise subspaces; The covariance matrix of the observation vector is calculated, and the eigenvalue decomposition of the covariance matrix is performed to obtain: where U S and U N are column vectors corresponding to the signal and noise subspaces, respectively, D S and D N are the eigenvalues of the signal and noise subspaces, respectively, O is a zero matrix, and the number of eigenvalues is 2L, E k,e = ||C k,e || / L and E k,l = ||C k,l || / L are the energies of the pseudo-code sequence, ||.|| is the 2-norm of a vector, μ k,e = C k,e / ||C k,e || and μ k,l = C k,l / ||C k,l || are the normalized pseudo-code vectors, and I is an identity matrix; the signal space of each user is divided into two uncorrelated subspaces, and the sample covariance matrix of the signal is composed of 2K signal subspaces and noise subspaces, so the eigenvalue of the kth user is expressed as: where, for the kth user, λ k,e is the eigenvalue of the first part, λ k,l is the eigenvalue of the last part, λ N,n is the eigenvalue of the last noise subspace; 1-2, when there is Rayleigh fading, the mathematical model expression of the asynchronous multi-user containing K user direct spread signals to be detected is: where h k is the Rayleigh channel fading gain for the kth user, which is subject to Rayleigh distribution. According to the derivation of equation (10), the eigenvalue decomposition of the detected signal covariance matrix is performed, and the result is:

3. The L-ResNet-based direct spread signal detection method of claim 2, wherein, The step 2 specifically comprises the following two cases: 2-1, when there is no Rayleigh fading, according to the principal component analysis method, the first 2K principal component subspaces of the detection signal are selected, other subspaces are equivalent to noise subspaces, the histogram of the signal eigenvalues of the K users is constructed, the superposition result of the first 2K eigenvalues is taken as a new first eigenvalue, and other eigenvalues remain unchanged; After the eigenvalues are superimposed, the mathematical expression of each eigenvalue in the eigenvalue histogram is: where λ1is the first eigenvalue for the kth user in a Rayleigh fading channel, λ k is the kth eigenvalue, and λ N,n is the last L-2K noise subspace eigenvalue. 2-2, when there is Rayleigh fading, when the histogram of the eigenvalues of the direct spread signals of the K users is constructed, the superposition of the first 2K eigenvalues of the covariance matrix is taken as a new first eigenvalue, and other eigenvalues remain unchanged; after the eigenvalues are superimposed, the mathematical expression of each eigenvalue in the eigenvalue histogram is: where, for the kth user in a Rayleigh fading channel, is the eigenvalue for the first user, is the eigenvalue for the kth user, is the eigenvalue for the last L-2K noise subspace.

4. The L-ResNet-based direct spread signal detection method of claim 3, wherein, The cross-entropy loss L s The calculation formula is shown as formula (15): where N is the number of batches of images in the training set, n is the number of classes of feature value histogram images in the training set, f i is the feature vector of the i-th feature value histogram image, y i is the i-th feature value histogram image, f i corresponding class, is the i-th feature value histogram image, f i corresponding class y i is the weight matrix of f j is the j-th column of the weight matrix W, b j is the bias term.

Images