Adaptive cooperative spectrum sensing method based on enhanced covariance residual matrix

Abstract

The present invention relates to an adaptive collaborative spectrum sensing method involving an enhanced covariance residual matrix, which is mainly completed by the following steps: Step 1, a random resonance system parameter optimization method: using the artificial fish swarm algorithm and a new optimization function introduced in the background technology to solve the optimal random resonance system parameters; Step 2, a dual-channel feature extraction step of an enhanced signal: using a random resonance system with set parameters to enhance the received signal, and using the orthogonal demodulation introduced in the background technology to extract I and Q path signals; Step 3, a multi-user data fusion method: calculating the covariance matrix of the I and Q path signals of each SU, the method of the present invention constructs a covariance residual matrix through random resonance preprocessing signals and matrix cancellation, reduces noise interference as much as possible, and greatly improves the spectrum sensing performance of the sensing algorithm in a low signal-to-noise ratio environment. And the calculation error caused by the progressiveness of the sample covariance is reduced by using multi-user data soft fusion and CNN classifier.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital communications, and in particular relates to an adaptive collaborative spectrum sensing method of an enhanced covariance residual matrix. Background Art

[0002] The traditional spectrum allocation method is static, which leads to the inability to fully utilize the spectrum, making spectrum resources increasingly scarce and limiting the development of wireless communications. With the emergence of cognitive radio (CR) technology, the primary user (PU) can intelligently access unoccupied idle spectrum, greatly improving spectrum utilization. Among them, spectrum sensing, as the key to CR, can accurately and intelligently identify and utilize idle spectrum to effectively improve spectrum utilization. Among them, traditional single-user spectrum sensing technology is not suitable for actual complex environments due to its limitations. For this reason, multi-user collaborative spectrum sensing technology has emerged. It can effectively improve the performance of spectrum sensing by fusing the perception results of multiple users. The method of the present invention constructs a covariance residual matrix through random resonance preprocessing signals and matrix cancellation, which reduces noise interference as much as possible and greatly improves the spectrum sensing performance of the perception algorithm in a low signal-to-noise ratio environment. And the calculation error caused by the progressiveness of sample covariance is reduced as much as possible by using multi-user data soft fusion and CNN classifier.

[0003] Some background technologies involved in the method of the present invention are as follows:

[0004] 1.QPSK modulation and demodulation

[0005] Since QPSK modulation is a very common modulation method in the current communication field, the main user signal in this paper uses QPSK modulation signal. The current common implementation method of QPSK modulation is to use IQ modulator to generate it. Its modulation principle is as follows Figure 1 As shown, the output signal after QPSK modulation is expressed as:

[0006] s(t)=I*cos(ωt)-Q*sin(ωt)=Acos(ωt+θ) (13)

[0007] c os(ωt) and s in(ωt) is the modulated carrier, ω is the modulation frequency, and θ is the phase of the QPSK modulated signal. Substitute (+1,+1), (-1,+1), (-1,-1), (+1,-1) as (I,Q) into equation (13) respectively, according to the trigonometric formula:

[0008] sinαcosβ±cosαsinβ=sin(α±β) (14)

[0009]

[0010] The corresponding output signal phases are: π / 4, 3π / 4, 5π / 4 and 7π / 4.

[0011] QPSK demodulation is the signal s( t) respectively multiply by c os(ωt) and s After in(ωt), the baseband signal is retained through a low-pass filter to obtain the I-path signal and the Q-path signal. Since spectrum sensing does not require specific signal information, the proposed sensing algorithm does not need to use a low-pass filter to completely restore the I and Q-path signals. Therefore, the I and Q-path signals obtained by orthogonal demodulation can be expressed as:

[0012] I=s(t)*cos(ωt) (16)

[0013] Q=s(t)*sin(ωt) (17)

[0014] 2. Stochastic resonance

[0015] In the study of stochastic resonance, the bistable stochastic resonance system driven by weak periodic signals and white noise is the most commonly used nonlinear system, which can be expressed by the Langevin equation:

[0016]

[0017] In the formula s(t) is the main user signal; n(t) The mean is 0 and the noise variance is Gaussian white noise, that is, E[n(t)] =0, U(x) is the potential function of the bistable system, and its expression is:

[0018]

[0019] In the formula, a and b are non-zero system parameters. Figure 3 As shown, the bistable state function has two stable states and an unsteady state x =0; barrier height ΔU =a 2 / 4b.

[0020] When there is no external input, the system is at the lowest point x of the potential well. ± , the potential energy is the smallest and the system is the most stable; when a weak signal is input to the system s(t) When the signal energy cannot overcome the potential barrier ΔU The system output state can only move in a potential well. If noise is added to the system n(t)When stochastic resonance is finally reached, part of the noise energy will be transferred to the signal, causing it to interact and overcome the system potential barrier, and transition between the two stable states at the signal frequency.

[0021] Initially, stochastic resonance could only be used for small signals, but as research progressed, it was found that after the system parameters were normalized, stochastic resonance could be used to detect large signals. The principle of normalization transformation is as follows:

[0022]

[0023] τ=at (21)

[0024] Substituting equations (20) and (21) into equation (18), we can obtain:

[0025]

[0026] After transformation, the signal and noise in equation (22) are s( τ / a) and n( τ / a), which is equivalent to performing a time domain analysis on the signal and noise. a times stretching, which is equivalent to the signal in the frequency domain s(t) and noise n(t) The frequency of the large signal is 1 / a times compressed. Therefore, after the large signal is normalized, the frequency becomes 1 / a of the original. When a is large enough, the large signal can be equivalent to a small signal, which meets the requirements of random resonance for the signal. As Gaussian white noise, n(t) is a constant component in all frequency ranges in the frequency domain, with the same power. Stretching or compression in the frequency domain does not change the power of the noise. Therefore, n(τ / a) still has a mean of zero and a variance of White noise:

[0027]

[0028] Substituting formula (23) into formula (22), we get:

[0029]

[0030] Formula (24) is the stochastic resonance system after normalization transformation. The output result of the stochastic resonance system is calculated by the fourth-order Runge-Kutta method. After passing through the stochastic resonance system, the QPSK signal does not change except for the amplitude, so its covariance matrix R s_sr The matrix characteristics of R s Keep consistent. The mean is 0 and the variance is After the Gaussian white noise passes through the stochastic resonance system, its Gaussian distribution characteristics become non-Gaussian distribution. Studies have shown that its probability density can be approximately expressed as:

[0031]

[0032] Since the probability density is symmetric, that is, p(x) = p(-x), the mean of the probability density function is:

[0033]

[0034] Since the mean is 0, its mean square value is equal to its variance:

[0035]

[0036] Among them is the integral result, which is a constant. The variance of this constant and Gaussian white noise is Related.

[0037] 3. Artificial fish swarm optimization algorithm

[0038] The variables in the artificial fish swarm algorithm include the total number of artificial fish Z, the number of generations GEN, the state X of the artificial fish individual (composed of the vector of the optimization variable), the maximum step length Step, the visual field Visual, the number of attempts Try_number, the crowding factor δ, and the distance d between the artificial fish individuals α and β. αβ =|X α -X β | and food concentration ψ ψ is the evaluation function in the optimization process and also the measurement index of stochastic resonance. The specific steps are as follows: First, initialize the parameters of the artificial fish school. If the signal frequency is 1×10 m Hz, then a The optimization range is set to [1×10 m-1 ,1×10 m+1 ], Visual is set to 0.5×10 m-1 , Step is set to 0.5×10 m-1 , the settings of the four parameters Try_number, δ, GEN, and Z are not based on the order of magnitude of the carrier frequency, and can be appropriately selected; second, the results of the initial generation of fish are calculated based on the stochastic resonance system and the initial parameter a, and the ψ The indicators are evaluated and the optimal parameters are selected; third, the artificial fish executes the behavior functions in sequence, continuously iterates, updates and records the global optimal parameters; fourth, when the number of iterations reaches the preset value, the current optimal parameter a is output.

[0039] 4. Convolutional Neural Network (CNN)

[0040] The convolutional neural network used in spectrum sensing is mainly composed of convolutional layers, pooling layers, and fully connected layers. The convolutional layer consists of a convolution kernel and a corresponding feature map. The convolution kernel performs a convolution operation with the feature map of the previous layer, and uses an activation function to generate the feature map of the current layer. The operation process of the entire convolutional layer is:

[0041]

[0042] Where L represents the layer number; X (L-1) Represents the feature map of the previous layer output; X (L) Represents the feature map output of this layer; W (L) represents the convolution kernel of this layer; b (L) represents the bias value of this layer; f(·) is the activation function, which can enhance the nonlinearity of the network. In this paper, relu is used as the activation function of each hidden layer, that is:

[0043]

[0044] The pooling layer can reduce the dimension of the feature map, reduce the training parameters, and relatively alleviate the overfitting of the training model. This patent uses mean pooling.

[0045] The fully connected layer is usually located at the end. It fully connects the feature maps of this layer and the previous layer, and converts the features obtained by multiple convolutions and pooling into output classification results.

[0046] In view of the above technical problems, improvements need to be made. Summary of the invention

[0047] In view of the shortcomings of the prior art methods, the present invention provides a highly efficient adaptive collaborative spectrum sensing method based on an enhanced covariance residual matrix and a neural network under low signal-to-noise ratio.

[0048] In order to achieve the above purpose, the technical solution adopted by the present invention is: an adaptive collaborative spectrum sensing method of an enhanced covariance residual matrix, which is completed by the following method or steps:

[0049] Step 1.1: Input the M×N dimensional received signal Y i , perform calculations according to the random resonance system parameter optimization method, and output the signal Y i Stochastic resonance system parameters that maximize the signal-to-noise ratio;

[0050] Step 1.2, the dual-channel feature extraction step of the enhanced signal, is completed in the following sub-steps:

[0051] Step 1.2.1, receive signal Y i and the optimal system parameters solved are used as input, and the stochastic resonance system preprocessing Y introduced in the background technology is used i, output M×N dimensional enhanced signal Y i_sr ;

[0052] Step 1.2.2, the enhanced signal Y i_sr As input parameters, the orthogonal demodulation introduced in the background technology is used to extract the I and Q path orthogonal signal matrices;

[0053] Step 1.3, taking the I and Q orthogonal signal matrices obtained in step 1.2 as input, performing calculations according to the multi-user data fusion method, and outputting the fused I and Q covariance matrices;

[0054] Step 1.4, taking the fused I and Q path covariances obtained in step 1.3 as input, performing calculations according to the covariance residual matrix construction method, and outputting a dual-channel covariance residual matrix;

[0055] Step 1.5, CNN construction and training step, is completed in the following sub-steps:

[0056] Step 1.5.1, build a CNN model based on the dimensions of the covariance residual matrix generated in step 1.4;

[0057] Step 1.5.2, according to the above covariance residual matrix generation method, generate a training set and a validation set, and train the CNN model built in step 1.5.1 by cross-training the training set and the validation set, and finally obtain the spectrum detector of this method.

[0058] As a preferred solution of the present invention, in step 1.1, the stochastic resonance system parameter optimization method is completed by the following steps:

[0059] Step 2.1, assuming that the received signal Y of the i-th SU is i is an M×N dimensional matrix, initializes the parameters of the stochastic resonance system, and uses the stochastic resonance system preprocessing Y introduced in the background technology i , and get the M×N dimensional output Y i_sr ;

[0060] Step 2.2: Calculate the output signal Y of the stochastic resonance system i_sr The covariance matrix R i_sr :

[0061]

[0062] Among them, Y i_sr is the output of the received signal of the i-th SU after passing through the stochastic resonance system, R s_sr is the covariance matrix of the main user signal after passing through the stochastic resonance system;

[0063] Step 2.3: According to R i_srThe normalized stochastic resonance parameter optimization problem can be equivalent to the following optimization problem:

[0064]

[0065] Among them, R i_sr (l,p) represents the covariance matrix R i_sr The element in the lth row and the pth column; the artificial fish swarm optimization algorithm in the background technology is used to solve the optimization problem, and the parameters obtained by the solution are used as the parameters of the stochastic resonance system.

[0066] As a preferred solution of the present invention, in step 1.3, the multi-user data fusion method is completed by the following steps:

[0067] Step 3.1: The output Y of the stochastic resonance system i_sr The orthogonal demodulation method introduced in the background technology is used for demodulation to obtain I and Q path signals:

[0068]

[0069] Among them, Y i_sr_I and Y i_sr_Q are the I and Q M×N dimensional signals output by the stochastic resonance system, y ij_sr (n) represents the output signal of the jth antenna in the ith use after passing through the random resonance system.

[0070] Step 3.2, calculate the dual-channel covariance matrix of the i-th SU based on the demodulated I and Q signals. Therefore, the covariance matrix of the I-channel signal after stochastic resonance is expressed as:

[0071]

[0072] in, is the variance of the noise after passing through the stochastic resonance system. The covariance matrix of the Q-path signal is expressed as:

[0073]

[0074] Step 3.3: Multiple SUs in a cognitive wireless network collaborate through the following soft fusion method:

[0075]

[0076] Where R fusion_I and R fusion_Q is the covariance matrix after multi-user fusion of I and Q channels, and ρ is the number of cooperating SUs. This method is theoretically equivalent to increasing the perception period of the SU receiving signal, thereby reducing the error caused by the progressiveness of the sample covariance matrix and does not increase the perception delay of the SU.

[0077] As a preferred solution of the present invention, in step 1.4, the covariance residual matrix construction method is completed by the following steps:

[0078] Step 4.1, R fusion_I and R fusion_Q Do Cholesky decomposition:

[0079] R=V(Λ s +Λ n )V H =U·U T (9)

[0080] Among them, Λ s =diag(χ1,…,χ p , 0, …, 0) M×M is the characteristic value of the main user signal,

[0081] is the eigenvalue of the noise, the matrix U is a lower triangular matrix, and the eigenvalue of the matrix R is λ i (i=1,2,…,M), then the diagonal elements of U are

[0082]

[0083] λ i The minimum value among Therefore, the noise variances of the I and Q paths are and

[0084] Step 4.2: Based on the estimated noise variance and Construct the covariance residual matrix:

[0085]

[0086] The beneficial effects of the present invention are as follows: the method of the present invention constructs a covariance residual matrix by preprocessing signals through stochastic resonance and matrix cancellation, thereby reducing noise interference as much as possible and greatly improving the spectrum perception performance of the perception algorithm in a low signal-to-noise ratio environment. In addition, the calculation error caused by the progressiveness of the sample covariance is reduced by using multi-user data soft fusion and CNN classifier. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 An efficient and adaptive collaborative spectrum sensing method based on enhanced covariance residual matrix and neural network;

[0088] Figure 2 To use the covariance residual matrix as a feature and use CNN for signal detection;

[0089] Figure 3 This is the schematic diagram of the IQ modulator for QPSK modulation;

[0090] Figure 4 Schematic diagram of the potential function of the bistable stochastic resonance system;

[0091] Figure 5 This is the principle diagram of the artificial fish swarm optimization algorithm;

[0092] Figure 6 It is a typical cognitive wireless network (CRN) system architecture; DETAILED DESCRIPTION

[0093] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0094] The efficient adaptive collaborative spectrum sensing method based on the enhanced covariance residual matrix and the neural network in this embodiment is implemented by the following steps in sequence:

[0095] Step 1, Stochastic Resonance System Parameter Optimization Method

[0096] Step 1.1, receiving signals, a typical cognitive radio network (CNR) consists of PU and SU. Usually, the wireless communication between PU and SU is physically separated, and SU cannot directly obtain the PU channel status. PU is also called an authorized user and has the priority right to use the authorized frequency band; SU is called a cognitive user and needs to perceive the spectrum occupancy status of PU users in real time to prevent communication conflicts with PU. Assuming that there are ρ SUs and T PUs in the entire cognitive network, and each SU device has M receiving antennas, the binary hypothesis of the jth antenna receiving signal of the i-th SU can be expressed as:

[0097]

[0098] H0 and H1 are the assumptions that the primary user signal does not exist and exists respectively. The kth sampled data sequence received by the i-th SU is:

[0099] y i (k) = [y i,1 (k),y i,2 (k),…,y i,M (k)] T (31)

[0100] The received signal matrix of the i-th SU is:

[0101]

[0102] Step 1.2, set the normalized stochastic resonance system parameters. The key to the stochastic resonance system is to make the signal and noise produce stochastic resonance. The stochastic resonance phenomenon is an abstract concept, so a specific measurement function is needed to describe whether the signal has produced stochastic resonance after passing through the stochastic resonance system. The system output signal-to-noise ratio SNR out It is the most commonly used measurement indicator, and its expression is as follows:

[0103]

[0104] Among them, s sr (n) is the output signal of PU after passing through the SR system, n sr (n) is the output signal after the noise passes through the SR system, and N is the number of sampling points. out When the maximum value is reached, it is considered that the signal and noise produce stochastic resonance, so the stochastic resonance parameter setting problem can be regarded as the following optimization problem:

[0105]

[0106] But using SNR out As a measurement function, it has the following disadvantages: 1. Since the signal and noise are mixed, it is difficult to accurately estimate the signal-to-noise ratio of the system output; 2. The conventional method of estimating the signal-to-noise ratio has a large computational complexity. out The value itself is not important, and the essence of the spectrum sensing algorithm of this patent is to use the difference between the covariance matrix of the noise and the covariance matrix of the primary user signal for sensing. Therefore, a new measurement function is proposed as the optimization objective function to solve the system parameters. First, the covariance matrix of the system output signal is calculated:

[0107]

[0108] Among them, Y i_sr is the output of the received signal of the i-th SU after passing through the stochastic resonance system, R s_sr is the covariance matrix of the main user signal after passing through the stochastic resonance system. From formula (35), we can see that theoretically when only noise exists, there are values ​​only on the diagonal of the matrix; when the main user signal exists, R s_sr The matrix has values ​​in each row and column, and the values ​​outside the diagonal are close in magnitude to the values ​​on the diagonal, because each antenna receives the same primary user signal, and the correlation is extremely high. Therefore, the new measurement function expression is as follows:

[0109]

[0110] Among them, R i_sr (l,p) represents the covariance matrix R i_srThe element in the lth row and pth column. When the received signal contains noise and the main user signal:

[0111]

[0112] According to the principle of stochastic resonance, when max(ψ), When the minimum is reached, part of the noise energy is transferred to the main user signal. At this time, it is considered that the signal and noise produce random resonance. Therefore, the optimization function of formula (34) is transformed into the following:

[0113]

[0114] According to the received signal Y i The artificial fish swarm algorithm in the background technology is used to solve the problem of Y i Matching stochastic resonance system parameters a opt and β opt .

[0115] Step 2: Dual-channel feature extraction step of enhanced signal

[0116] Step 2.1: Preprocess the received signal Y using the normalized stochastic resonance system with set parameters i , output signal Y i_sr

[0117] Y i_sr =[y i_sr (1),y i_sr (2),…,y i_sr (N)] (39)

[0118] Step 2.2: The output Y of the stochastic resonance system is i_sr The orthogonal demodulation method introduced in the background technology is used for demodulation to obtain I and Q path signals:

[0119]

[0120] Among them, Y i_sr_I and Y i_sr_Q are the I and Q M×N dimensional signals output by the stochastic resonance system, y ij_sr (n) represents the output signal of the jth antenna in the ith use after passing through the random resonance system.

[0121] Step 3: Multi-user data fusion method

[0122] Step 3.1, calculate the dual-channel covariance matrix of the i-th SU based on the demodulated I and Q signals.

[0123]

[0124] Among them, <·> represents the inner product operation, cos(ωn) is a cosine signal with a length of N and the same frequency as PU, and sin(ωn) is a sine signal with a length of N and the same frequency as PU. Therefore, the elements in the covariance matrix of the I and Q signals can be expressed as:

[0125]

[0126] Among them, R i_sr_I (l,p) and R i_sr_Q (l,p) are the elements in the lth row and pth column of the covariance matrix of the I and Q signals respectively. Therefore, the covariance matrix of the I signal after stochastic resonance is expressed as:

[0127]

[0128] The covariance matrix of the Q-path signal is expressed as:

[0129]

[0130] Step 3.2, SUs perform data soft fusion collaboration. Usually, there are multiple SUs in a cognitive wireless network, and the traditional collaboration method is through hard fusion such as AND criterion, OR criterion and K order fusion, which can appropriately improve the perception performance of the algorithm to a certain extent, but it mainly improves the algorithm performance by reducing the sudden detection error, and the entire collaboration calculation amount increases proportionally with the increase of SU. Therefore, the present invention collaborates through the following soft fusion method:

[0131]

[0132] Where R fusion_I and R fusion_Q It is the covariance matrix after multi-user fusion of I and Q paths. Since the sample covariance matrix is ​​progressive, this fusion method is equivalent to increasing the number of sampling points of the single-user received signal but not increasing the perception cycle of the perception algorithm. When the number of cooperative users increases, the sample covariance matrix approaches the statistical covariance matrix, which improves the accuracy of the feature statistics at the cost of a certain computational complexity, thereby improving the performance of the entire perception algorithm.

[0133] Step 4: Covariance residual matrix construction method

[0134] Step 4.1, use the covariance matrix and Cholesky decomposition to estimate the noise variance. According to equations (48) and (49), perform Cholesky decomposition on the covariance matrices of the I and Q signals respectively:

[0135] R=V(Λ s +Λ n )V H =U·UT (50)

[0136] Among them, Λ s =diag(χ1,…,χ p ,0,…,0) M×M is the characteristic value of the main user signal, is the eigenvalue of the noise, the matrix U is a lower triangular matrix, and the eigenvalue of the matrix R is λ i (i=1,2,…,M), then the diagonal elements of U are

[0137]

[0138] λ i The minimum value among Therefore, the noise variances of the I and Q paths are and

[0139] Step 4.2: Construct the covariance residual matrix. In order to further increase the difference between matrices under different assumptions and thus improve the perception performance of the perception algorithm under low signal-to-noise ratio, the covariance residual matrix is ​​constructed by the following formula:

[0140]

[0141] Step 5: Generate training samples using the above method of constructing the covariance residual matrix. Assume that there are n pairs of data for training. represents the nth two-channel covariance residual matrix, express The forward propagation starts from the input layer. After multiple layers of convolution, pooling and other operations, the final output mapping relationship is:

[0142]

[0143] Among them, W and b are the parameters that need to be trained in CNN, representing weights and bias values ​​respectively. express Output map after CNN.

[0144] The result obtained by forward propagation Label with the desired result There is an error. This paper uses cross entropy as the error loss function, which can be expressed as:

[0145]

[0146] In order to reduce the error loss function Loss value between the final predicted value and the actual value, the back propagation algorithm is used to adjust the parameter values ​​of W and b in CNN layer by layer.

[0147] The specific implementation modes of the present invention can be described in detail through the following embodiment drawings.

[0148] Figure 1 An efficient adaptive collaborative spectrum sensing method based on an enhanced covariance residual matrix and a neural network is described, which is characterized by sequentially performing the following steps or methods:

[0149] Step 1: Receive the signal, set the stochastic resonance system parameters according to the new measurement function, and solve the stochastic resonance system parameters that can maximize the signal-to-noise ratio of the received signal.

[0150] Step 2: Taking the received signal and the optimal system parameters as input parameters, an enhanced signal is obtained, and the orthogonal demodulation introduced in the background technology is used to extract the dual-channel signal.

[0151] Step 3: Calculate the covariance matrix of each channel signal of each SU, perform soft fusion of multi-user data according to the number of SUs in the cognitive network and the perception requirements, and output the fused covariance matrix.

[0152] Step 4: Take the fused covariance matrix as the input parameter, use Cholesky decomposition to estimate the noise variance, and output the covariance residual matrix through matrix cancellation.

[0153] Step 5: Generate training set and validation set according to the construction method of covariance residual matrix, input the training set and validation set into the built CNN network to train the model, and finally use the trained CNN model to determine whether the PU signal exists. If the PU signal exists, it is 1, otherwise it is 0, and the determination result vector is output.

[0154] In Figure 2, the signal time-frequency characteristics are extracted and the detection statistics are constructed. The specific steps are as follows:

[0155] Assume that there are ρ SUs and T PUs in the entire cognitive network, each SU device has M receiving antennas, and the binary hypothesis of the signal received by the jth antenna of the i-th SU is

[0156]

[0157] H0 and H1 are the assumptions that the primary user signal does not exist and exists respectively. The kth sampled data sequence received by the i-th SU is:

[0158] y i (k) = [y i,1 (k),y i,2 (k),…,yi,M (k)] T (57)

[0159] The received signal matrix of the i-th SU is:

[0160]

[0161] Step 1: Calculate the covariance matrix of the normalized stochastic resonance system output signal:

[0162]

[0163] Among them, Y i_sr is the output of the received signal of the i-th SU after passing through the stochastic resonance system, R s_sr The covariance matrix of the main user signal after passing through the stochastic resonance system. According to the following optimization objective function, the artificial fish swarm optimization algorithm is used to solve the optimal stochastic resonance system parameters that match the received signal:

[0164]

[0165] Among them, R i_sr (l,p) represents the covariance matrix R i_sr The element in row l and column p.

[0166] Step 2: The output Y of the stochastic resonance system i_sr The orthogonal demodulation method introduced in the background technology is used for demodulation to obtain I and Q path signals:

[0167]

[0168] Among them, Y i_sr_I and Y i_sr_Q are the I and Q M×N dimensional signals output by the stochastic resonance system, y ij_sr (n) represents the output signal of the jth antenna in the i-th SU after passing through the random resonance system.

[0169] Step 3: Calculate the dual-channel covariance matrix of the i-th SU based on the demodulated I and Q signals.

[0170]

[0171] Among them, <·> represents the inner product operation, cos(ωn) is a cosine signal with a length of N and the same frequency as PU, and sin(ωn) is a sine signal with a length of N and the same frequency as PU. Therefore, the elements in the covariance matrix of the I and Q signals can be expressed as:

[0172]

[0173]

[0174] Among them, R i_sr_I (l,p) and R i_sr_Q (l,p) are the elements in the lth row and pth column of the covariance matrix of the I and Q signals respectively. Therefore, the covariance matrix of the I signal after stochastic resonance is expressed as:

[0175]

[0176] The covariance matrix of the Q-path signal is expressed as:

[0177]

[0178] Multiple SUs in a cognitive wireless network collaborate through the following soft fusion methods:

[0179]

[0180] Where R fusion_I and R fusion_Q It is the covariance matrix after multi-user fusion of I and Q paths.

[0181] Step 4: R fusion_I and R fusion_Q Do Cholesky decomposition:

[0182] R=V(Λ s +Λ n )V H =U·U T (71)

[0183] Among them, Λ s =diag(χ1,…,χ p ,0,…,0) M×M is the characteristic value of the main user signal, is the eigenvalue of the noise, the matrix U is a lower triangular matrix, and the eigenvalue of the matrix R is λ i (i=1,2,…,M), then the diagonal elements of U are

[0184]

[0185] λ i The minimum value among Therefore, the noise variances of the I and Q paths are and

[0186] Construct the covariance residual matrix:

[0187]

[0188] The above method of constructing the covariance residual matrix is ​​used to generate the training set and the validation set, and the trained CNN is used to finally perform spectrum sensing.

[0189] Figure 3 QPSK modulation is realized by IQ modulator. The output signal after QPSK modulation is expressed as:

[0190] s(t)=I*cos(ωt)-Q*sin(ωt)=A cos(ωt+θ) (75)

[0191] cos(ωt) and sin(ωt) are modulated carriers, ω is the modulation frequency, and θ is the phase of the QPSK modulated signal. Substituting (+1,+1), (-1,+1), (-1,-1), and (+1,-1) as (I,Q) into equation (75) respectively, according to the trigonometric formula:

[0192] sinαcosβ±cosαsinβ=sin(α±β) (76)

[0193]

[0194] The corresponding output signal phases are: π / 4, 3π / 4, 5π / 4 and 7π / 4.

[0195] Figure 4 is the bistable system potential function U(x) of the stochastic resonance system, and its expression is:

[0196]

[0197] Where a and b are non-zero system parameters. The bistable state function has two stable states and a non-steady state x = 0; the barrier height ΔU = a 2 / 4b. When there is no external input, the system is at the lowest point x of the potential well. ± , the potential energy is the smallest and the system is the most stable; when a weak signal s(t) is input to the system, the signal energy cannot overcome the potential barrier ΔU The system output state can only move in a potential well. If noise is added to the system n(t) When stochastic resonance is finally reached, part of the noise energy will be transferred to the signal, causing it to interact and overcome the system potential barrier, and transition between the two stable states at the signal frequency.

[0198] Figure 5This is the overall process of the artificial fish swarm algorithm. This algorithm is a new optimization algorithm proposed based on the research on the intelligent behavior of animal groups. This algorithm simulates the foraging behavior of fish schools based on the characteristic that the place with the largest number of fish in the water area is the place with the most nutrients in the water area to achieve optimization. The algorithm mainly uses the three basic behaviors of fish: foraging, clustering and tail-chasing. It adopts a top-down optimization mode starting from the bottom-level behavior of the constructed individual, and through the local optimization of each individual in the fish school, the global optimal value is highlighted in the group.

[0199] Figure 6 This is a typical cognitive radio network (CRN) system architecture. A typical cognitive radio network (CNR) consists of primary users (PU) and secondary users (SU). It is usually assumed that the wireless network communications of PU and SU are physically separated, and SU cannot directly obtain the PU channel status. In this system, PU has priority to use the occupied channel. The cognitive base station (CBS) first determines the idle channel of the spectrum by detecting the PU signal in the channel. Then, the status of the PU receiver (PU-R) is sent and the idle spectrum is determined. SU can reuse the spectrum until PU no longer occupies the spectrum. If the spectrum being used by SU is accessed by PU, SU exits the spectrum and moves into the cache. The cognitive device detects other idle spectrum at the same time.

[0200] The method of the present invention mainly proposes an efficient adaptive collaborative spectrum sensing method based on an enhanced covariance residual matrix and a neural network under low signal-to-noise ratio (SNR), which includes random resonance enhanced signals, dual-channel orthogonal signal extraction, multi-user data fusion, Cholesky noise variance estimation, construction of a covariance residual matrix and CNN training. By constructing a covariance residual matrix through random resonance preprocessing signals and matrix cancellation, noise interference is reduced as much as possible, greatly improving the spectrum sensing performance of the perception algorithm in a low signal-to-noise ratio environment. And the calculation error caused by the progressiveness of the sample covariance is reduced as much as possible by using multi-user data soft fusion and CNN classifier, which has a high application value. The innovative points of the present invention are mainly the following:

[0201] 1. Propose a new stochastic resonance measurement function to reduce the complexity of system parameter optimization

[0202] 2. Use matrix fusion to achieve collaborative spectrum sensing among multiple users

[0203] 3. Use Cholesky to estimate noise and construct the covariance residual matrix as the perceptual feature

[0204] Finally, the above description is only a preferred embodiment of the method of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may also include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. An adaptive collaborative spectrum sensing method based on an enhanced covariance residual matrix, characterized in that: Complete it in the following ways or steps: Step 1.1: Input the M×N dimensional received signal Y i , perform calculations according to the random resonance system parameter optimization method, and output the signal Y i Stochastic resonance system parameters that maximize the signal-to-noise ratio; Step 1.2, the dual-channel feature extraction step of the enhanced signal, is completed in the following sub-steps: Step 1.2.1, receive signal Y i and the best system parameters solved are used as input, and the stochastic resonance system preprocessing Y i , output M×N dimensional enhanced signal Y i_sr ; Step 1.2.2, the enhanced signal Y i_sr As input parameters, orthogonal demodulation extracts the I and Q orthogonal signal matrices; Step 1.3, taking the I and Q orthogonal signal matrices obtained in step 1.2 as input, performing calculations according to the multi-user data fusion method, and outputting the fused I and Q covariance matrices; Step 1.4, taking the fused I and Q path covariances obtained in step 1.3 as input, performing calculations according to the covariance residual matrix construction method, and outputting a dual-channel covariance residual matrix; Step 1.5, CNN construction and training step, is completed in the following sub-steps: Step 1.5.1, build a CNN model based on the dimensions of the covariance residual matrix generated in step 1.4; Step 1.5.2, according to the above covariance residual matrix generation method, generate a training set and a validation set, and train the CNN model built in step 1.5.1 by cross-training the training set and the validation set, and finally obtain the spectrum detector of this method; In step 1.1, the stochastic resonance system parameter optimization method is completed by the following steps: Step 2.1, assuming that the received signal Y of the i-th SU is i is an M×N dimensional matrix, initializes the stochastic resonance system parameters, and preprocesses the stochastic resonance system Y i , and get the M×N dimensional output Y i_sr ; Step 2.2: Calculate the output signal Y of the stochastic resonance system i_sr The covariance matrix R i_sr : Among them, Y i_sr is the output of the received signal of the i-th SU after passing through the stochastic resonance system, R s_sr is the covariance matrix of the main user signal after passing through the stochastic resonance system; where, is the variance of the noise after passing through the stochastic resonance system; Step 2.3: According to R i_sr The normalized stochastic resonance parameter optimization problem can be equivalent to the following optimization problem: Among them, R i_sr (l,p) represents the covariance matrix R i_sr The element in the lth row and pth column; the artificial fish swarm optimization algorithm is used to solve the optimization problem, and the solved parameters are used as the parameters of the stochastic resonance system.

2. The adaptive collaborative spectrum sensing method of the enhanced covariance residual matrix according to claim 1, characterized in that: In step 1.3, the multi-user data fusion method is completed by the following steps: Step 3.1: The output Y of the stochastic resonance system i_sr Orthogonal demodulation is used to demodulate and obtain I and Q signals: Among them, Y i_sr_I and Y i_sr_Q are the I and Q M×N dimensional signals output by the stochastic resonance system, y ij_sr (n) represents the output signal of the jth antenna in the i-th use after passing through the stochastic resonance system; Step 3.2, calculate the dual-channel covariance matrix of the i-th SU based on the demodulated I and Q signals; therefore, the covariance matrix of the I-channel signal after stochastic resonance is expressed as: Among them, the covariance matrix of the Q-path signal is expressed as: Step 3.3: Multiple SUs in a cognitive wireless network collaborate through the following soft fusion method: Where R fusion_I and R fusion_Q is the covariance matrix after multi-user fusion of I and Q paths, and ρ is the number of cooperative SUs.

3. The adaptive collaborative spectrum sensing method of the enhanced covariance residual matrix according to claim 1, characterized in that: In step 1.4, the covariance residual matrix construction method is completed by the following steps: Step 4.1, R fusion_I and R fusion_Q Do Cholesky decomposition: R=V(Λ s +Λ n )V H =U·U T (9) Among them, Λ s =diag(χ1,…,χ p ,0,…,0) M×M is the characteristic value of the main user signal, is the eigenvalue of the noise, the matrix U is a lower triangular matrix, and the eigenvalue of the matrix R is λ i (i=1,2,…,M), then the diagonal elements of U are λ i The minimum value among Therefore, the noise variances of the I and Q paths are and Step 4.2: Based on the estimated noise variance and Construct the covariance residual matrix:

Images