A cooperative spectrum sensing method and system

Abstract

The application discloses a kind of cooperative spectrum sensing method and system, it is related to the technical field of spectrum sensing, constructs cooperative spectrum sensing model, and by the signal in the authorized frequency band in cooperative secondary user in cooperative spectrum sensing model is observed and reconstruction, further by fusion center the signal of reconstruction is combined into signal matrix, and its corresponding covariance matrix is calculated, further the covariance matrix is divided into training set and test set, wherein, the sample with a small amount of label in training set then carries out clustering training to training set, and then a large number of sample labels are acquired, and corresponding spectrum sensing classifier is acquired based on clustering training result, finally according to spectrum sensing classifier, test set is classified, according to the classification result, judge whether primary user exists in the authorized frequency band observed by cooperative secondary user, verify the reliability and practicality of cooperative spectrum sensing method, improve the spectrum sensing precision in low signal-to-noise ratio environment, and then improve the overall performance of spectrum sensing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of spectrum sensing, and in particular to a cooperative spectrum sensing method and system. BACKGROUND

[0002] With the rapid popularization of 5G networks and intelligent electronic devices, as a non-renewable and scarce resource in wireless communication, the demand for spectrum is increasing. However, spectrum resources are allocated and authorized by relevant state departments, and the static spectrum management mode causes the imbalance between supply and demand and low utilization of spectrum resources. The emergence of cognitive radio technology alleviates the contradiction between the increasing spectrum demand and the low utilization of spectrum resources, so that more unlicensed users can reasonably use spectrum resources and improve the utilization of spectrum. In 1999, Dr. Joseph Mitola published a paper on cognitive radio, proposed the concept of cognitive radio, and explained its advantages and applications. Cognitive radio technology improves the utilization of radio spectrum resources through functions such as spectrum sharing, spectrum management, and spectrum sensing, and spectrum sensing plays an important role in it. Spectrum sensing technology can monitor and analyze the use of radio spectrum resources in real time, and it can effectively identify spectrum holes, so that the cognitive radio system can further realize the sharing and management of spectrum.

[0003] Cooperative spectrum sensing is a reliable spectrum sensing technology that aims to achieve efficient sensing and utilization of spectrum resources through cooperation between multiple wireless devices, thereby improving the utilization of spectrum resources.

[0004] On the one hand, traditional single-user spectrum sensing methods such as energy detection, cyclostationary feature detection, and matched filter detection have the disadvantages of channel fading and shadowing effects in practical applications. On the other hand, these methods require signal prior information and fixed threshold settings, which makes them not adaptive in changing environments and have poor spectrum sensing performance in low signal-to-noise ratio environments. Existing technologies disclose some spectrum sensing algorithms based on random matrix theory and eigenvalues of covariance matrices. These methods use eigenvalues or elements of covariance matrices as signal information intermediaries and set threshold parameters based on random matrix theory, without the need for prior information about channel types and signal models to be detected, thereby improving spectrum sensing performance. However, spectrum algorithms based on covariance eigenvalues require a fixed decision threshold to be set artificially based on random matrix theory, which is easily affected by noise uncertainty, and in a complex and changing wireless communication environment, setting a fixed threshold greatly reduces the spectrum sensing performance of the proposed algorithm.

[0005] With the rapid development of science and technology and artificial intelligence, there are many technologies that use machine learning to realize spectrum sensing to solve the problem that setting a fixed decision threshold can greatly reduce the spectrum sensing performance of the proposed algorithm. Some researchers have proposed a cooperative spectrum sensing algorithm combining information geometry theory and clustering algorithm of unsupervised learning. These methods form a sample matrix from the data observed by the secondary user and convert it into a corresponding covariance matrix. The covariance matrix is mapped to the manifold space by combining information geometry knowledge. Riemann distance is used as a geometric feature of the signal in different ways. Finally, different clustering algorithms are used in Euclidean space to train these geometric features offline.

[0006] However, the above algorithms extract the corresponding Riemann distance as the signal feature vector through information geometry, need to evaluate the noise environment to construct the reference point, and in the process of converting the signal from the manifold space to the Euclidean space, some important feature information may be lost, which will lead to the decline of the overall performance of spectrum sensing. SUMMARY

[0007] The present application aims at the problem of high requirement for sample label quantity and poor spectrum sensing performance of existing cooperative spectrum sensing methods. A cooperative spectrum sensing method is proposed, which does not need to spend a lot of time and effort to manually label data labels, is more practical, and improves the spectrum detection accuracy in low signal-to-noise ratio environment, thereby improving the overall performance of spectrum sensing.

[0008] In order to achieve the purpose of the present application, the present application adopts the following technical solutions:

[0009] A cooperative spectrum sensing method, comprising the following steps:

[0010] S1: Construct a cooperative spectrum sensing model, the cooperative spectrum sensing model at least comprising: one primary user, one fusion center and M cooperative secondary users;

[0011] S2: Use the cooperative secondary user to observe the signal, and reconstruct the observed signal;

[0012] S3: Use the fusion center to combine the reconstructed signal into a signal matrix, and calculate the corresponding covariance matrix thereof;

[0013] S4: Divide the covariance matrix into a training set and a test set;

[0014] S5: Cluster training is performed on the training set, and based on the clustering training result, a spectrum sensing classifier is obtained;

[0015] S6: The test set is classified by using the spectrum sensing classifier, and according to the classification result, it is judged whether the primary user exists.

[0016] According to the technical solution, the cooperative spectrum sensing model is constructed, the signal is observed and reconstructed by the cooperative secondary users in the cooperative spectrum sensing model, the reconstructed signal is combined into a signal matrix by the fusion center, the corresponding covariance matrix is calculated, the covariance matrix is further divided into a training set and a test set, the training set is clustered and trained, a large number of sample labels are obtained, the corresponding spectrum sensing classifier is obtained based on the clustering training result, and finally the sample labels in the test set are classified according to the spectrum sensing classifier, so that whether the primary user exists in the licensed frequency band observed by the cooperative secondary user is determined according to the classification result, thereby improving the spectrum sensing accuracy in a low signal-to-noise ratio environment, and improving the overall performance of spectrum sensing.

[0017] Further, the cooperative spectrum sensing model in step S1 satisfies a binary hypothesis test expression, and the binary hypothesis test expression is:

[0018]

[0019] wherein H1 represents that the primary user exists in the licensed frequency band, H0 represents that the primary user does not exist in the licensed frequency band, n represents the number of sampling points of the observed signal, n = 1, 2, … S, i represents the number of cooperative secondary users, i = 1, 2, …, M, w i (n) represents an additive white Gaussian noise and satisfies h i (n) represents the gain coefficient of the Rayleigh fading channel, x(n) represents the primary user signal, y i (n) represents the signal observed by the i th cooperative secondary user in the licensed frequency band.

[0020] According to the technical solution, whether the primary user exists in the signal observation target frequency band is represented by assuming H1 and H0, which further simplifies the complexity of the cooperative spectrum sensing model and improves the practicability of the cooperative spectrum sensing model.

[0021] Further, the process of signal observation by the cooperative secondary user and signal reconstruction of the observed signal in step S2 is:

[0022] The VMD is used to reconstruct the signal y i observed by the i th cooperative secondary user to obtain the reconstructed signal The expression is:

[0023]

[0024] wherein VMD represents the variational mode decomposition.

[0025] According to the technical solution, the variational mode decomposition model is used to reconstruct the signal y iThe pre-processing is performed, and then a corresponding reconstructed signal expression is obtained The reconstructed signal expression can effectively express the signal information, and the signal reconstruction can better output and display the signal, thereby enhancing the signal quality.

[0026] Further, the process of combining the reconstructed signals into a signal matrix by the fusion center in step S3 is:

[0027] S31: The fusion center combines the reconstructed signals collected by the M cooperative secondary users into a signal matrix Y, and the process satisfies:

[0028]

[0029] S32: The corresponding covariance matrix of the signal matrix Y is calculated, and the expression is:

[0030]

[0031] wherein R represents an MxM matrix, denotes matrix transposition, E[Y] represents an MxS matrix, and the (i,j) element in E[Y] represents the average value of the i-th row in the signal matrix Y, i=1,...,M, and j=1,...,S.

[0032] According to the above technical solution, the reconstructed signals collected by the M cooperative secondary users are combined into a signal matrix Y by the fusion center The signal matrix Y can enhance the stability of the cooperative spectrum sensing model, and the corresponding covariance matrix of the signal matrix Y is further calculated, which can effectively process two-dimensional problems and further process the problem of binary hypothesis testing.

[0033] Further, the process of dividing the covariance matrix into a training set and a test set in step S4 is:

[0034] S41: The probability distribution function family is mapped to a statistical manifold by information geometry, and the process satisfies:

[0035]

[0036] wherein S represents a probability distribution function family, p(x|θ) represents a probability distribution, Ω represents a random variable, C n represents an n-dimensional sample space, Θ represents an m-dimensional parameter set composed of characteristic vectors, θ represents a parameter in the parameter set, C m represents an m-dimensional characteristic vector space, and θ represents a parameter in the parameter set.

[0037] ​S42: parameterize the parameter θ in the probability distribution p(x|θ) with a covariance matrix using information geometry, define the span of all covariance matrices as a matrix manifold, and map the covariance matrix to a point signal on the matrix manifold, wherein the matrix manifold includes a signal covariance matrix R s and a noise covariance matrix R w ;

[0038] S43: set a spectrum sensing period, collect N point signals in the spectrum sensing period, and group the point signals into a data set D, and the process satisfies:

[0039] D={R1,R2,…R N};

[0040] wherein R i ∈{R s ,R w}, i=1…N;

[0041] Divide the data set D into a training set D1 and a test set D2, and mark 10% to 15% of the training set D1 as a sample set in the H0 state, denoted as Φ1; the remaining 85% to 90% of the unlabeled sample set, denoted as Φ2, and the process satisfies:

[0042]

[0043] wherein Φ1+Φ2=D1, represents the noise covariance matrix R w collected when the primary user is in the H0 state. represents the sample label when the primary user is in the H0 state, i=1, 2,...k, represents the covariance matrix when the state of the primary user is unknown, n=1…u.

[0044] According to the technical solution, the probability distribution function family is mapped to a statistical manifold by information geometry, and the span of all covariance matrices is defined as a matrix manifold, and then the covariance matrix is mapped to a point signal on the matrix manifold. In a spectrum sensing period, N point signals are collected, the point signals are used as point data on the matrix manifold space, and the probability distribution of the point signals is further expressed, which facilitates theoretical analysis of the data, and the use of the matrix manifold is more concise and effective. The point signals are used as a data set, and the data set is divided into a training set and a test set, so that the experimental data of spectrum sensing can be obtained more accurately.

[0045] Further, in step S5, the IGSSC algorithm is used for clustering training of the training set, and a spectrum sensing classifier is obtained based on the clustering training result, and the specific process is as follows:

[0046] S51: Let any one point signal in sample set Φ1 be initial clustering center Γ1 of primary user state H0, calculate Log-Euclidean distance of all point signals to the initial clustering center Γ1, and let the point signal with the largest Log-Euclidean distance to the initial clustering center Γ1 be initial clustering center Γ2 of primary user state H1;

[0047] S52: Initialize empty sets C1 and C2, store all labeled samples in sample set Φ1 into empty set C1, and calculate Log-Euclidean distance of all unlabeled samples in Φ2 to initial clustering centers Γ1 and Γ2, obtain the sample with the smallest Log-Euclidean distance to the nearest clustering center, and store the unlabeled sample into set C1 or C2;

[0048] S53: Calculate Log-Euclidean mean of all samples in sets C1 and C2 respectively and judge the relationship between Log-Euclidean mean and initial clustering centers Γ1 and Γ2, if and then output clustering center otherwise let return to step S52;

[0049] S54: Obtain a spectrum sensing classifier based on the clustering center, and the expression of the spectrum sensing classifier is:

[0050]

[0051] wherein, IGSSC represents information geometry-based semi-supervised clustering algorithm, represents Log-Euclidean distance between covariance matrix R in the training set and clustering center point of primary user state H0, represents Log-Euclidean distance between covariance matrix R and clustering center point of primary user state H1.

[0052] Further, the process of calculating Log-Euclidean distance of all point signals to the initial clustering center Γ1 in step S51 satisfies:

[0053] d L (A1,Γ1)=||log(A1)-log(Γ1)|| F ;

[0054] wherein, ||·|| F ​​Frobenius norm of a matrix, log(·) represents logarithm operation, A1 represents any one of all point signals; Log-Euclidean represents logarithm Euclidean.

[0055] Further, the Log-Euclidean mean of all samples in the set C1 and C2 is respectively calculated in step S53 The expression satisfies:

[0056]

[0057] wherein, represents the Log-Euclidean mean between two point signals, m≥2, exp represents exponential operation, A i represents any one of all point signals.

[0058] According to the above technical scheme, in the process of clustering training of the training set by using the IGSSC algorithm, the Log-Euclidean distance between the point signal and the initial clustering center is calculated, and the sample closest clustering center is obtained according to the Log-Euclidean distance value, and the unlabeled sample is stored in the set C1 or C2, and then the clustering center is obtained through the set C1 and C2, and the spectrum sensing classifier is obtained based on the clustering center, and the point signal in the training set can be effectively utilized through the spectrum sensing classifier, and the accuracy of cooperative spectrum sensing is improved.

[0059] Further, the process of classifying the test set by using the spectrum sensing classifier in step S6 is as follows:

[0060] S61: classifying the test set by using the spectrum sensing classifier, and setting a parameter λ to control the false alarm probability, so as to obtain an ROC curve;

[0061] S62: judging the state of the primary user through the ROC curve, and the expression satisfies:

[0062]

[0063] wherein, ROC represents receiver operating characteristic, R * represents the covariance matrix in the test set;

[0064] If f(R * )≥1, it indicates that the state of the predicted primary user is H1, at this time, the covariance matrix R * in the test set is closer to the clustering center point of the primary user state H1 If f(R *If <1, it indicates that the state of the primary user is H0, at this time the covariance matrix R * The cluster center point closer to the state H0 of the primary user

[0065] According to the above technical solution, the test set is classified by using the spectrum sensing classifier, and the ROC curve is drawn according to the classification result, and the state of the primary user is judged through the ROC curve, and the practicality and reliability of the cooperative spectrum sensing model are verified according to the judgment result.

[0066] A cooperative spectrum sensing system for implementing the above cooperative spectrum sensing method, comprising:

[0067] A construction module for constructing a cooperative spectrum sensing model, the cooperative spectrum sensing model comprising at least one primary user, one fusion center and M cooperative secondary users;

[0068] A sensing processing module for signal observation and signal reconstruction;

[0069] A fusion module for combining the reconstructed signals into a signal matrix and calculating the corresponding covariance matrix;

[0070] A division module for dividing the covariance matrix into a training set and a test set;

[0071] A training module for clustering training of the training set, and obtaining a spectrum sensing classifier based on the clustering training result;

[0072] A classification module for classifying the test set according to the spectrum sensing classifier, and judging whether the primary user exists based on the classification result.

[0073] According to the above technical solution, through the mutual cooperation between the sensing module, the processing module, the fusion module, the division module and the classifier module in the cooperative spectrum sensing system, the above cooperative spectrum sensing method is realized, and the reliability and practicality of the cooperative spectrum sensing method are embodied.

[0074] Compared with the prior art, the beneficial effects of the present application are:

[0075] The application provides a cooperative spectrum sensing method, which comprises the following steps: constructing a cooperative spectrum sensing model, observing and reconstructing signals by a cooperative secondary user in the cooperative spectrum sensing model, combining the reconstructed signals into a signal matrix by a fusion center, calculating a corresponding covariance matrix, dividing the covariance matrix into a training set and a test set, wherein the training set contains a small amount of labeled samples, then performing clustering training on the training set, obtaining a large amount of sample labels, obtaining a corresponding spectrum sensing classifier based on the clustering training result, and finally classifying the sample labels in the test set according to the spectrum sensing classifier, so as to determine whether the primary user exists in the licensed frequency band observed by the cooperative secondary user according to the classification result, thereby improving the spectrum sensing accuracy in a low signal-to-noise ratio environment and improving the overall performance of spectrum sensing; the observed signals are reconstructed by using a variational mode decomposition method, so that the signal output and display are better and the signal quality is enhanced; meanwhile, the covariance matrix is processed by using an information geometry and a matrix manifold method, the covariance matrix is mapped into point data on the matrix manifold, the point data are divided into the training set and the test set, and then the state of the primary user is determined according to the training set and the test set, thereby improving the reliability and practicability of the cooperative spectrum sensing method. BRIEF DESCRIPTION OF DRAWINGS

[0076] Figure 1 A flowchart of a cooperative spectrum sensing method provided by the embodiment of the application;

[0077] Figure 2 A composition diagram of a cooperative spectrum sensing model provided by the embodiment of the application;

[0078] Figure 3 A specific working schematic diagram of a cooperative spectrum sensing method provided by the embodiment of the application;

[0079] Figure 4 A ROC curve diagram of different spectrum sensing methods in the same signal-to-noise ratio environment provided by the embodiment of the application;

[0080] Figure 5 A ROC curve diagram of different spectrum sensing methods in different signal-to-noise ratio environments provided by the embodiment of the application;

[0081] Figure 6 A schematic diagram of a cooperative spectrum sensing system provided by the embodiment of the application;

[0082] In the figure, PU represents a primary user, and SU represents a cooperative secondary user. DETAILED DESCRIPTION

[0083] For the purposes of this disclosure, reference will be made to the accompanying drawings which form a part of the disclosure. The drawings are not necessarily to scale of proportion and are exemplary of particular embodiments of the present application. Embodiments of the present application, however, are not limited to the exemplary

[0084] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0085] Embodiment One:

[0086] A cooperative spectrum sensing method, see Figure 1 , comprising the steps of:

[0087] S1: constructing a cooperative spectrum sensing model, the cooperative spectrum sensing model comprising at least one primary user, one fusion center and M cooperative secondary users;

[0088] S2: using the cooperative secondary users to perform signal observation and reconstructing the observed signals;

[0089] S3: using the fusion center to combine the reconstructed signals into a signal matrix and calculating the corresponding covariance matrix thereof;

[0090] S4: dividing the covariance matrix into a training set and a test set;

[0091] S5: performing clustering training on the training set and obtaining a spectrum sensing classifier based on the clustering training result;

[0092] S6: using the spectrum sensing classifier to classify the test set and determining whether the primary user exists according to the classification result.

[0093] It can be understood that by constructing a cooperative spectrum sensing model, and observing and reconstructing the signals in the licensed frequency band by the cooperative secondary users in the cooperative spectrum sensing model, and then combining the reconstructed signals into a signal matrix by the fusion center, and calculating the corresponding covariance matrix, further dividing the covariance matrix into a training set and a test set, wherein the training set has a small number of labeled samples, then clustering training the training set, and then obtaining a large number of sample labels based on the clustering training result, and obtaining the corresponding spectrum sensing classifier, and finally classifying the sample labels in the test set according to the spectrum sensing classifier, and determining whether the primary user exists in the licensed frequency band observed by the cooperative secondary user according to the classification result, thereby improving the spectrum sensing accuracy in a low signal-to-noise ratio environment, and further improving the overall performance of spectrum sensing.

[0094] Referring to Figure 2 , the cooperative spectrum sensing model described in step S1 satisfies a binary hypothesis test expression, and the binary hypothesis test expression is:

[0095]

[0096] wherein H1 represents that the licensed frequency band has a primary user, H0 represents that the licensed frequency band does not have a primary user, n represents the number of sampling points of the observed signal, n=1, 2, …, S, i represents the number of cooperative secondary users, i=1, 2, …, M, w i (n) represents an additive white Gaussian noise and satisfies h i (n) represents the gain coefficient of the Rayleigh fading channel, x(n) represents the primary user signal, y i (n) represents the signal observed by the i-th cooperative secondary user on the licensed frequency band.

[0097] It can be understood that by assuming H1 and H0 to represent whether the primary user exists in the signal observation target frequency band, the complexity of the cooperative spectrum sensing model is further simplified, and the practicability of the cooperative spectrum sensing model is improved.

[0098] Referring to Figure 3 , the process of observing the signal by the cooperative secondary user and reconstructing the observed signal described in step S2 is:

[0099] The VMD is used to reconstruct the signal y i observed by the i-th cooperative secondary user, and the reconstructed signal is obtained, and the expression is:

[0100]

[0101] wherein VMD represents variational mode decomposition.

[0102] Exemplarily, the variational modal decomposition constructs a constraint optimization problem according to the intrinsic modal function, forms a variational modal decomposition model with constraint, and the expression of the variational modal decomposition model is as follows:

[0103]

[0104] wherein, u k represents the intrinsic modal function, δ(t) represents the unit impulse function, j represents the imaginary unit, and * represents the convolution operation, represents the partial derivative of the function, {u k}={u1,u2,…,u K}、{ω k}={ω1,ω2,…,ω K},u K (t) represents the Kth intrinsic modal function obtained by the variational modal decomposition, ω K represents the center frequency of the Kth component, represents the square of the L2 norm, and f represents the original observation signal.

[0105] The above model is a variational modal decomposition model with constraint, so the Lagrange multiplier λ and the quadratic penalty term α are used to convert it into a non-constrained variational problem, and the augmented Lagrange equation of the following formula is introduced:

[0106]

[0107] Therefore, the solution of the original minimization problem is converted into solving the saddle point of the above formula;

[0108] A series of iterative optimization {u k} and {ω k} are solved by using the alternating direction multiplier algorithm, and then {u k} is reconstructed to realize the noise reduction processing of the original signal f.

[0109] The signal y i detected by the ith SU is subjected to the VMD reconstruction processing, and the reconstructed signal shown in the following formula is obtained

[0110]

[0111] wherein, the original signal f is one of the symbols in the VMD, and is equivalent to the signal y i .

[0112] Understandably, by constructing a variational mode decomposition model and solving it using the alternating direction multiplier algorithm, a series of optimized intrinsic mode functions are obtained and reconstructed to achieve noise reduction of the original signal f in the variational mode decomposition model. Subsequently, the variational mode decomposition model is used to process the signal y observed by the cooperative sub-user. i Preprocessing is performed to obtain the corresponding reconstructed signal expression. Reconstructing signal expressions can better and more effectively express signal information, and signal reconstruction can better output and display signals, thus enhancing signal quality.

[0113] Furthermore, the process of combining the reconstructed signals into a signal matrix using the fusion center in step S3 is as follows:

[0114] S31: The fusion center will collect the reconstruction signals from M collaborative sub-users. The signals are fused into a single signal matrix Y, and the process satisfies:

[0115]

[0116] S32: Calculate the covariance matrix corresponding to the signal matrix Y, expressed as follows:

[0117]

[0118] Where R represents an M×M matrix, Let E[Y] denote the matrix transpose, and let E[Y] denote the M×S matrix. The (i,j)th element in E[Y] represents the average value of the i-th row in the signal matrix Y, where i = 1,...,M and j = 1,...,S.

[0119] For example, the fusion center will collect reconstruction signals from M collaborative sub-users. The signals are fused into a single signal matrix Y, which is an M×S matrix. The corresponding covariance matrix R can be obtained from Y.

[0120] When the observation data is collected under the signal H0 state, this represents the noise covariance matrix, denoted as R. w Additionally, R s This represents the signal covariance matrix when the signal is in state H1.

[0121] Understandably, the reconstructed signals collected from M collaborative sub-users are processed through the fusion center. The signal matrix Y is fused together to enhance the stability of the cooperative spectrum sensing model. The covariance matrix corresponding to the signal matrix Y is then calculated. The covariance matrix can effectively handle two-dimensional problems and further address the problem of binary hypothesis testing.

[0122] Further, the process of dividing the covariance matrix into a training set and a test set in step S4 is:

[0123] S41: Map the probability distribution function family to a statistical manifold using information geometry, and the process satisfies:

[0124]

[0125] Where S represents the probability distribution function family, p(x|θ) represents the probability distribution, Ω represents the random variable, C n represents the n-dimensional sample space, Θ represents the m-dimensional parameter set composed of feature vectors, θ represents a parameter in the parameter set, C m represents the m-dimensional feature vector space, and θ represents a parameter in the parameter set.

[0126] S42: Parameterize the parameter θ in the probability distribution p(x|θ) using information geometry, and define the span of all covariance matrices as a matrix manifold, and map the covariance matrix to a point signal on the matrix manifold, wherein the matrix manifold includes the signal covariance matrix R s and the noise covariance matrix R w .

[0127] S43: Set a spectrum sensing period, collect N point signals in the spectrum sensing period, and form a data set D from the point signals, and the process satisfies:

[0128] D={R1,R2,…R N};

[0129] Where R i ∈{R s ,R w}, i=1…N.

[0130] Divide the data set D into a training set D1 and a test set D2, and mark 10% to 15% of the training set D1 as a sample set in the H0 state, denoted as Φ1; the remaining 85% to 90% of the unlabeled sample set, denoted as Φ2, and the process satisfies:

[0131]

[0132] Where Φ1+Φ2=D1, represents the noise covariance matrix R w collected when the primary user is in the H0 state. represents the sample label when the primary user is in the H0 state, i=1,2,...k, represents the covariance matrix when the primary user state is unknown, n=1…u.

[0133] For example, the connection between the covariance matrix, information geometry, and matrix manifold is established through the following process:

[0134] Information geometry maps a family of parameterized probability distribution functions to a statistical manifold, on which each point signal represents a probability distribution function;

[0135] The family of probability distribution functions S is defined as shown in the following equation:

[0136]

[0137] Using information geometry theory, the parameter θ in the probability distribution p(x|θ) is parameterized by the covariance matrix. Then, the space spanned by the covariance matrix can be defined as a matrix manifold.

[0138] On a matrix manifold, each covariance matrix can be regarded as a point signal in the manifold space. The distance between two point signals is defined according to different distance metrics, and different distance metrics correspond to different means.

[0139] For two distinct n-order covariance matrices A1 and A2, if the distance between two point signals on the matrix manifold is defined by the log-Euclidean metric, this distance is called the Log-Euclidean distance, and its calculation formula is as follows:

[0140] d L (A1,A2)=||log(A1)-log(A2)|| F ;

[0141] Among them, ||·|| F represents the Frobenius norm of the matrix, and log(·) represents the logarithmic operation.

[0142] The matrix manifold includes the signal covariance matrix R. s And noise covariance matrix R w These two elements, through information geometry theory, can be mapped to a point signal on a matrix manifold;

[0143] Within a certain sensing period, N signals from these points are collected to form a dataset D, as shown in the following formula:

[0144] D = {R1, R2, ..., R} N}

[0145] Where R i ∈{R s ,R w}, i = 1…N;

[0146] The samples in the dataset D are divided into a training set D1 and a test set D2 in a ratio of 8:2, and 10% of the samples in the training set D1 are marked as H0 state, denoted as Φ1, and the remaining samples without labels are denoted as Φ2, and the specific expressions are as follows:

[0147]

[0148] It can be understood that the probability distribution function family is mapped to a statistical manifold by information geometry, and the span of all covariance matrices is defined as a matrix manifold, and then the covariance matrix is mapped to a point signal on the matrix manifold. In a spectrum sensing period, N point signals are collected, and the point signal is used as a point data on the matrix manifold space, and the probability distribution of the point signal is further expressed, which facilitates theoretical analysis of the data, and the use of the matrix manifold is more simple and effective; the point signal is used as a data set, and the data set is divided into a training set and a test set, so that the experimental data of spectrum sensing can be obtained more accurately.

[0149] Further, in step S5, the IGSSC algorithm is used for clustering training of the training set, and a spectrum sensing classifier is obtained based on the clustering training result, and the specific process is as follows:

[0150] S51: taking any one point signal in the sample set Φ1 as an initial clustering center Γ1 of the primary user state H0, calculating the Log-Euclidean distance of all point signals to the initial clustering center Γ1; and taking the point signal with the largest Log-Euclidean distance to the initial clustering center Γ1 as an initial clustering center Γ2 of the primary user state H1;

[0151] S52: initializing empty sets C1 and C2, storing all labeled samples in the sample set Φ1 into the empty set C1, and calculating the Log-Euclidean distance of all unlabeled samples in Φ2 to the initial clustering centers Γ1 and Γ2, obtaining the sample with the smallest Log-Euclidean distance to the sample according to the Log-Euclidean distance value, and storing the unlabeled sample in the set C1 or C2;

[0152] S53: calculating the Log-Euclidean mean of all samples in the sets C1 and C2 respectively and judging the relationship between the Log-Euclidean mean and the initial clustering centers Γ1 and Γ2, if and then outputting the clustering center otherwise, let return to step S52;

[0153] S54: Based on cluster centers, obtain the spectrum-aware classifier. The expression for the spectrum-aware classifier is:

[0154]

[0155] Here, IGSSC represents a semi-supervised clustering algorithm based on information geometry. This represents the covariance matrix R in the training set and the cluster center points of the master user with state H0. Log-Euclidean distance between them This represents the covariance matrix R and the cluster center points of the primary user state H1. The Log-Euclidean distance between them.

[0156] Furthermore, the process of calculating the Log-Euclidean distance from all point signals to the initial cluster center Γ1 in step S51 satisfies:

[0157] d L (A1,Γ1)=||log(A1)-log(Γ1)|| F ;

[0158] Among them, ||·|| F Let Frobenius norm be the matrix, log(·) denote the logarithmic operation, A1 denote any one of the point signals, and Log-Euclidean denotes the logarithmic Euclidean.

[0159] Further, step S53 involves calculating the Log-Euclidean mean of all samples in sets C1 and C2 respectively. The expression satisfies:

[0160]

[0161] in, This represents the Log-Euclidean mean between two point signals, where m ≥ 2, exp represents the exponential operation, and A i It represents any one of all point signals.

[0162] For example, there is a small number of training sets labeled as state H0. Represents the noise covariance matrix. Where i = 1, 2, ..., k; unlabeled training set Let u represent the covariance matrix where the main user's state is unknown, and n = 1*u. Number of categories: 2;

[0163] Randomly select one covariance matrix from the training set Φ1 as the initial cluster center Γ1 of the primary user state H0, calculate the Log-Euclidean distance between all covariance matrices and the initial cluster center Γ1, and take the point signal with the largest Log-Euclidean distance from the initial cluster center Γ1 as the initial cluster center Γ2 of the primary user state H1;

[0164] Initialize two empty sets C1 and C2;

[0165] Divide all samples in Φ1 into the set C1, calculate the Log-Euclidean distance between all unlabeled samples in Φ2 and the cluster centers Γ1 and Γ2, find the cluster center with the smallest Log-Euclidean distance from the sample according to the Log-Euclidean distance value, and divide the unlabeled sample into the corresponding cluster, and construct two new sets C1 and C2;

[0166] Calculate the Log-Euclidean mean of all samples in the sets C1 and C2 respectively

[0167] If the two conditions and are met, output the cluster centers Otherwise, take the current two Log-Euclidean means as the new cluster centers Γ1 and Γ2 respectively, and return to the step of initializing two empty sets C1 and C2;

[0168] Output the cluster centers

[0169] After the IGSSC algorithm is trained offline, an expression of a spectrum sensing classifier f(R) can be obtained as follows:

[0170]

[0171] wherein, represents the Log-Euclidean distance between the covariance matrix R and the cluster center of the primary user state H0, represents the Log-Euclidean distance between the covariance matrix R and the cluster center of the primary user state H1.

[0172] It can be understood that, in the process of clustering training of the training set by using the IGSSC algorithm, the Log-Euclidean distance between the point signal and the initial cluster center is calculated, and the cluster center with the smallest Log-Euclidean distance from the sample the most recent cluster center and the unlabeled sample are stored in the set C1 or C2, and then the cluster centers are obtained through the sets C1 and C2, and the spectrum sensing classifier is obtained based on the cluster centers, through which the point signals in the training set can be effectively utilized, and the accuracy of the cooperative spectrum sensing is improved.

[0173] Further, the process of classifying the test set by using the spectrum sensing classifier according to the classification result to determine whether the primary user exists is described in step S6.

[0174] S61: classifying the test set by using the spectrum sensing classifier, and setting a parameter λ to control the false alarm probability, so as to obtain an ROC curve;

[0175] S62: determining the state of the primary user through the ROC curve, and the expression satisfies:

[0176]

[0177] wherein, ROC represents the receiver operating characteristic, R * represents the covariance matrix in the test set;

[0178] If f(R * )≥1, it indicates that the state of the primary user is predicted as H1, at this time, the covariance matrix R * in the test set is closer to the cluster center point of the primary user state H1 If f(R * )<1, it indicates that the state of the primary user is predicted as H0, at this time, the covariance matrix R * in the test set is closer to the cluster center point of the primary user state H0

[0179] For example, in the spectrum sensing stage, first, the observation signals received by each cooperative secondary user are preprocessed to obtain the covariance matrix R * ; then, R * is input into the classifier f(R), if f(R * )≥1, it indicates that the state of the primary user is predicted as H1, because at this time, the covariance matrix R * in the test set is closer to the cluster center point of the primary user state H1 Therefore, it is attributed to the category of the primary user state H1; on the contrary, if f(R * )<1, it indicates that the state of the primary user is predicted as H0.

[0180] The trained classifier is used to classify the test set and draw an ROC curve, and compared with other existing spectrum sensing methods, the false alarm probability P fFor 0.1, compare the detection probability P of different algorithms d , the detection probability P d The spectrum sensing method with high spectrum sensing performance represents that its spectrum sensing performance is good.

[0181] It can be understood that the test set is classified by using the spectrum sensing classifier, and the ROC curve is drawn according to the classification result, and the state of the primary user is judged through the ROC curve, and the practicability and reliability of the cooperative spectrum sensing model are verified according to the judgment result.

[0182] In this embodiment, the cooperative spectrum sensing model is constructed, and the cooperative secondary users in the cooperative spectrum sensing model observe and reconstruct the signals in the licensed frequency band, and then the reconstructed signals are combined into a signal matrix by the fusion center, and the corresponding covariance matrix is calculated, and the covariance matrix is further divided into a training set and a test set, wherein the training set has a small number of labeled samples, and then the training set is clustered and trained, and a large number of sample labels are obtained, and the corresponding spectrum sensing classifier is obtained based on the clustering training result, and finally the sample labels in the test set are classified according to the spectrum sensing classifier, and whether the primary user exists in the licensed frequency band observed by the cooperative secondary user is judged according to the classification result, thereby improving the spectrum sensing accuracy in the low signal-to-noise ratio environment, and further improving the overall performance of the spectrum sensing; and the observed signal is reconstructed by using the method of variational mode decomposition, so that the signal output and display are better, and the signal quality is enhanced; at the same time, the covariance matrix is processed by using the method of information geometry and matrix manifold, the covariance matrix is mapped to point data on the matrix manifold, the training set and the test set are divided according to the point data, and then the state of the primary user is judged according to the training set and the test set, thereby improving the reliability and practicability of the cooperative spectrum sensing method.

[0183] Embodiment two:

[0184] In this embodiment, the effectiveness of the cooperative spectrum sensing method proposed in the present application is verified, specifically:

[0185] Referring to Figure 4 and Figure 5, assuming that the simulation primary user signal is a multi-component signal and the noise is an additive white Gaussian noise with mean 0 and variance 1. 500 covariance matrices containing H0 and H1 states are collected as dataset D respectively. In addition, 80% of dataset D is used as training set D1 and the rest is used as test set D2, and the proportion of labeled samples in training set D1 is 10%. The comparative spectrum sensing methods used by the method of the present application (VMDIGSSC) are the cooperative spectrum sensing methods of genetic simulated annealing algorithm based on quadratic covariance matrix and information geometry (QCIGSA), the cooperative spectrum sensing method based on information geometry and deep learning (IGDNN), the cooperative spectrum sensing method based on information geometry and K-means clustering (IGKCSS), and the cooperative spectrum sensing method based on information geometry and fuzzy C-means clustering algorithm (IGFCM).

[0186] When all the cooperative secondary users are distributed in the same signal-to-noise ratio environment, under the conditions of signal-to-noise ratio -19dB, sampling point number 1500, and cooperative secondary user number 5, Figure 4 The ROC curves of VMDIGSSC and other spectrum sensing methods are shown. When P f is equal to 0.1, the P d of the VMDIGSSC method of the present application is 0.991, which is improved by 8.54%, 46.60%, 40.17%, and 31.26% respectively in detection accuracy performance compared with QCIGSA, IGDNN, IGKCSS, and IGFCM methods.

[0187] When all the cooperative secondary users are distributed in different signal-to-noise ratio environments, under the conditions of sampling point number 1500, cooperative secondary user number 5, and each cooperative secondary user being in -17dB, -18dB, -19dB, -20dB, and -21dB signal-to-noise ratio environments respectively. Figure 5 The ROC curves of VMDIGSSC method and other spectrum sensing methods are shown when each cooperative secondary user is in different signal-to-noise ratio environments. When P f is equal to 0.1, the P d of the VMDIGSSC method of the present application is 0.987, which is improved by 7.52%, 16.80%, 27.35%, and 15.17% respectively in detection accuracy performance compared with QCIGSA, IGDNN, IGKCSS, and IGFCM methods.

[0188] Therefore, it is known that in a low signal-to-noise ratio environment, regardless of whether each cooperative secondary user is distributed in the same signal-to-noise ratio environment, the cooperative spectrum sensing method has higher detection accuracy in all comparison methods. At the same time, it is proved that compared with the supervised learning method and the unsupervised learning method, the spectrum sensing classifier trained by using the variational modal decomposition for noise reduction preprocessing and the information geometry semi-supervised clustering is more practical.

[0189] Embodiment three:

[0190] A cooperative spectrum sensing system, see Figure 6 , comprising:

[0191] The construction module is used for constructing a cooperative spectrum sensing model, and the cooperative spectrum sensing model at least comprises one primary user, one fusion center and M cooperative secondary users.

[0192] The perception processing module is used for signal observation and signal reconstruction.

[0193] The fusion module is used for combining the reconstructed signal into a signal matrix and calculating a corresponding covariance matrix.

[0194] The division module is used for dividing the covariance matrix into a training set and a test set.

[0195] The training module is used for clustering training of the training set, and a spectrum sensing classifier is obtained based on the clustering training result.

[0196] The classification module is used for classifying the test set according to the spectrum sensing classifier, and judging whether the primary user exists based on the classification result.

[0197] It can be understood that through the mutual cooperation between the perception module, the processing module, the fusion module, the division module and the classifier module in the cooperative spectrum sensing system, the above-mentioned cooperative spectrum sensing method is realized, and the reliability and practicability of the cooperative spectrum sensing method are embodied.

[0198] The above-mentioned is only an embodiment of the present application, and does not limit the patent range of the present application, and any equivalent structure or equivalent flow transformation obtained by using the content of the specification and the drawings, or direct or indirect application in other related technical fields, are also included in the patent protection range of the present application.

Claims

1. A cooperative spectrum sensing method, characterized in that, The method comprises the following steps: S1: constructing a cooperative spectrum sensing model, the cooperative spectrum sensing model comprising at least one primary user, one fusion center and M cooperative secondary users; S2: performing signal observation by using the cooperative secondary users, and reconstructing the observed signals; S3: combining the reconstructed signals into a signal matrix by using the fusion center, and calculating a corresponding covariance matrix; S4: dividing the covariance matrix into a training set and a test set; S5: performing clustering training on the training set, and obtaining a spectrum sensing classifier based on the clustering training result; S6: classifying the test set by using the spectrum sensing classifier, and judging whether the primary user exists according to the classification result. The process of performing signal observation by using the cooperative secondary users and reconstructing the observed signals in step S2 is as follows: The signal observed by the first cooperative secondary user is reconstructed by VMD to obtain a reconstructed signal. The expression is:​​ ； Wherein, VMD represents variational modal decomposition. The process of dividing the covariance matrix into a training set and a test set in step S4 is as follows: S41: mapping a probability distribution function family into a statistical manifold by using information geometry, and the process satisfies: ； wherein, S denotes a family of probability distribution functions, denotes a probability distribution, , denotes a random variable, denotes a d-dimensional sample space, denotes a parameter set consisting of d-dimensional eigenvectors, denotes a parameter of the parameter set, denotes a d-dimensional eigenvector space, denotes a parameter of the parameter set; S42: parameterizing the probability distribution in information geometry with covariance matrices, and defining the span of all covariance matrices as a matrix manifold, mapping the covariance matrices as point signals on the matrix manifold, wherein the matrix manifold comprises a signal covariance matrix and a noise covariance matrix ; S43: set a spectrum sensing period, collect a number of the point signals in the spectrum sensing period, and form a data set with the point signals N , and the process satisfies: D S43: set a spectrum sensing period, collect a number of the point signals in the spectrum sensing period, and form a data set with the point signals N , and the process satisfies: D ； wherein ; The dataset D Divided into training set and test set and the training set 10% to 15% of the total is marked as The sample set of states, denoted as The remaining 85% to 90% of the unlabeled sample set is denoted as... The process satisfies: ； wherein , represents a noise covariance matrix collected in the primary user being in the state ; represents a sample label of the primary user being in the state , represents a covariance matrix of the primary user state being unknown, .

2. The cooperative spectrum sensing method of claim 1, wherein The cooperative spectrum sensing model in step S1 satisfies a binary hypothesis test expression, and the binary hypothesis test expression is as follows: ； wherein, represents that a primary user exists in the licensed frequency band, represents that a primary user does not exist in the licensed frequency band, represents the number of sampling points of the observed signal, , represents the number of cooperative secondary users, , represents an additive white Gaussian noise and satisfies , represents a gain coefficient of a Rayleigh fading channel, represents a primary user signal, represents the signal observed by the th cooperative secondary user on the licensed frequency band.

3. The cooperative spectrum sensing method of claim 2, wherein, The process of combining the reconstructed signals into a signal matrix by using the fusion center in step S3 is as follows: S31: The fusion center will reconstruct the signal collected by the cooperative secondary users into a signal matrix , the process satisfies: ； S32: By signal matrix The corresponding covariance matrix is calculated, and the expression is as follows: ； wherein represents a matrix, T represents matrix transposition, represents a matrix, the element in the average value of the row in the , .

4. The cooperative spectrum sensing method of claim 3, wherein, In step S5, the IGSSC algorithm is used to perform clustering training on the training set, and a spectrum sensing classifier is obtained based on the clustering training result, and the specific process is as follows: S51: let any one point signal in the sample set be the initial clustering center with the primary user state being , calculate the Log-Euclidean distance of all point signals to the initial clustering center , and let the point signal with the largest Log-Euclidean distance to the initial clustering center be the initial clustering center with the primary user state being ; S52: initialize empty set and store all labeled samples in the sample set into the empty set and calculate the Log-Euclidean distance of all unlabeled samples in the sample set to the initial cluster centers and according to the Log-Euclidean distance values, obtain the cluster center closest to the sample and store the unlabeled sample into the set or ; S53: Compute the set of with the Log-Euclidean mean of all samples in ; and judge the Log-Euclidean mean value and the initial clustering center and , if and , output the clustering center ; Otherwise, let , return to step S52; S54: obtaining the spectrum sensing classifier based on the clustering center, and the expression of the spectrum sensing classifier is as follows: ； where IGSSC denotes the information geometry based semi-supervised clustering algorithm, denotes the covariance matrix in the training set and the Log-Euclidean distance between the cluster center point with the primary user state is denoted by denotes the covariance matrix and the Log-Euclidean distance between the cluster center point with the primary user state is denoted by 5. The cooperative spectrum sensing method of claim 4, wherein, The process of calculating the Log-Euclidean distance of all point signals to the initial cluster center described in step S51 satisfies: The process of calculating the Log-Euclidean distance of all point signals to the initial cluster center described in step S51 satisfies: ； wherein, denotes the Frobenius norm of a matrix, denotes a logarithm operation, denotes any one of all point signals; Log-Euclidean denotes a logarithm Euclidean.

6. The cooperative spectrum sensing method of claim 4, wherein, The respective calculation set of step S53 With Log-Euclidean mean of all samples in , the expression satisfies: ； wherein, denotes the Log-Euclidean mean between two point signals, m ≥ 2, exp denotes the exponential operation, A i denotes any one of all point signals.

7. The cooperative spectrum sensing method of claim 5 or 6, characterized by, The process of classifying the test set by using the spectrum sensing classifier in step S6 and judging whether the primary user exists according to the classification result is as follows: S61: classifying the test set using the spectrum sensing classifier and setting parameters to control the false alarm probability, thereby obtaining a ROC curve; S62: judging the state of the primary user by using an ROC curve, and the expression satisfies: ； where ROC denotes the receiver operating characteristic, denotes the covariance matrix in the test set; like This indicates that the predicted state of the primary user is... At this point, the covariance matrix in the test set... Closer to the main user status Cluster center ;like This indicates that the predicted state of the primary user is... At this point, the covariance matrix in the test set... Closer to the main user status Cluster center .

8. A cooperative spectrum sensing system, characterized in that, The cooperative spectrum sensing system comprises the cooperative spectrum sensing method according to any one of claims 1-7, and comprises: A construction module configured to construct a cooperative spectrum sensing model, the cooperative spectrum sensing model comprising at least one primary user, one fusion center and M cooperative secondary users; A sensing processing module configured to perform signal observation and reconstruct the observed signals; A fusion module configured to combine the reconstructed signals into a signal matrix and calculate a corresponding covariance matrix; A division module configured to divide the covariance matrix into a training set and a test set; A training module configured to perform clustering training on the training set, and obtain a spectrum sensing classifier based on the clustering training result; A classification module configured to classify the test set according to the spectrum sensing classifier, and judge whether the primary user exists based on the classification result.

Images