Multi-antenna Cooperative Spectrum Intelligence Sensing Method, System, Device and Medium for Distributed Collaboration under Non-Gaussian Noise

Abstract

A multi-antenna cooperative spectrum intelligent sensing method, system, device and medium for distributed cooperation under non-Gaussian noise. The method includes: constructing a non-linear model based on fractional-order stochastic resonance to enhance the useful weak signals at the multi-antenna receiving end, and selecting the fractional lower-order covariance matrix as the intelligent representation of the preprocessed signals; constructing a distributed cooperative spectrum sensing system based on federated learning, and using a vision-based self-attention mechanism network model to complete the local training of the sub-sensing nodes, and using the dynamic weighted federated averaging method to weight and aggregate the network parameters transmitted to the fusion center to obtain the global optimal network parameters for the global network model of sensing. The detection statistic and detection threshold of spectrum sensing are constructed by using the ratio of the two-dimensional feature vectors output by the network, and the sizes of the obtained detection statistic and detection threshold are compared to realize the multi-antenna cooperative spectrum intelligent sensing for distributed cooperation under non-Gaussian noise, and it has good sensing performance under low zero-power conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of spectrum sensing in cognitive radio systems, and particularly relates to a multi-antenna cooperative spectrum intelligent sensing method, system, device and medium for distributed cooperation under non-Gaussian noise. Background Art

[0002] In recent years, with the rapid development of mobile communication technology, the contradiction between the huge demand for wireless spectrum resources and the extremely scarce spectrum resources has become increasingly apparent. Cognitive radio technology is the key to solving this contradiction and can effectively improve the utilization rate of spectrum resources. Among them, spectrum sensing technology, as an important part of cognitive radio, can achieve efficient utilization of spectrum resources through dynamic spectrum access. The accuracy of the sensing result will greatly affect the efficiency of secondary user random access and the stability of licensed users during communication. Therefore, it is extremely important to design efficient, intelligent and secure spectrum sensing methods.

[0003] Currently, there have been many methods for studying spectrum sensing. Energy detection methods (W. Wu et al., "IRS-Enhanced Energy Detection for Spectrum Sensing in Cognitive Radio Networks," in IEEE Wireless Communications Letters, vol. 10, no. 10, pp. 2254-2258, Oct. 2021, doi: 10.1109 / LWC.2021.3099121.), matched filter detection methods (A. Brito, P. and F.J. Velez, "Hybrid Matched Filter Detection Spectrum Sensing," in IEEE Access, vol. 9, pp. 165504-165516, 2021, doi: 10.1109 / ACCESS.2021.3134796.), methods based on cyclostationarity (M. Nouri, H. Behroozi, N.K. Mallat and S.A. Aghdam, "A Wideband 5G Cyclostationary Spectrum Sensing Method by Kernel Least Mean Square Algorithm for Cognitive Radio Networks," in IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 68, no. 7, pp. 2700-2704, July 2021, doi: 10.1109 / TCSII.2021.3051087.), methods based on the maximum-minimum eigenvalue (R.B. Chaurasiya and R. Shrestha, "Hardware-Efficient and Fast Sensing-Time Maximum-Minimum-Eigenvalue-Based Spectrum Sensor for Cognitive Radio Network," in IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 66, no. 11, pp. 4448-4461, Nov. 2019, doi: 10.1109 / TCSI.2019.2921831.). Utilizing the powerful data-driven capabilities of deep learning, it has unique advantages in dealing with sample classification problems with a large amount of data.Deep learning-based spectrum sensing methods include: cooperative spectrum sensing method based on deep reinforcement learning (R. Sarikhani and F. Keynia, "Cooperative Spectrum Sensing Meets Machine Learning: Deep Reinforcement Learning Approach," in IEEE Communications Letters, vol. 24, no. 7, pp. 1459-1462, July 2020, doi: 10.1109 / LCOMM.2020.2984430.), spectrum sensing method based on convolutional neural network and long short-term memory neural network (J. Xie, J. Fang, C. Liu and X. Li, "Deep Learning-Based Spectrum Sensing in Cognitive Radio: A CNN-LSTM Approach," in IEEE Communications Letters, vol. 24, no. 10, pp. 2196-2200, Oct. 2020, doi: 10.1109 / LCOMM.2020.3002073.), spectrum sensing method combining short-time Fourier transform and convolutional neural network (Z. Chen, Y.-Q. Xu, H. Wang and D. Guo, "Deep STFT-CNN for Spectrum Sensing in Cognitive Radio," in IEEE Communications Letters, vol. 25, no. 3, pp. 864-868, March 2021, doi: 10.1109 / LCOMM.2020.3037273.), unsupervised learning spectrum sensing method based on variational auto-encoder (J. Xie, J. Fang, C. Liu and L. Yang, "Unsupervised Deep Spectrum Sensing: A Variational Auto-Encoder Based Approach," in IEEE Transactions on Vehicular Technology, vol. 69, no. 5, pp. 5307-5319, May 2020, doi: 10.1109 / TVT.2020.2982203.).Most of the above-mentioned spectrum sensing methods have excellent sensing performance under Gaussian noise. However, in a real wireless environment scenario, the received signal at the receiving end will be interfered by non-Gaussian noise. If the Gaussian noise model is still used, the spectrum sensing performance will be greatly reduced.

[0004] For spectrum sensing under non-Gaussian noise assumptions, researchers have proposed some spectrum sensing methods. Li et al. proposed a spectrum sensing method for non-Gaussian multipath fading channels based on Rao detection (J. Li, Q. Chen, Z. Long, W. Wang, H. Zhu and L. Wang, "Spectrum Sensing With Non-Gaussian Noise Over Multi-Path Fading Channels Towards Smart Cities With IoT," in IEEE Access, vol. 9, pp. 11194-11202, 2021, doi: 10.1109 / ACCESS.2021.3051719.), Lee et al. proposed a non-linear combination scheme based on order analysis to cope with the heavy-tailed characteristics of impulse noise. In a Rayleigh fading impulse noise environment, the detection performance of the proposed method is better than that of traditional methods (S. Lee, S. R. Park, Y. H. Kim and I. Song, "Spectrum sensing for cognitive radio network with multiple receive antennas under impulsive noise environments," in Journal of Communications and Networks, vol. 23, no. 3, pp. 171-179, June 2021, doi: 10.23919 / JCN.2021.000016.), Bhavana et al. proposed a non-reconstruction-based wideband compressive spectrum sensing method under non-Gaussian noise by using the robustness of the maximum correlation entropy criterion to impulse noise (B. Bhavana, S. Namburu, T. Panigrahi and S. L. Sabat, "Robust Methods for Wideband Compressive Spectrum Sensing Under Non-Gaussian Noise," in IEEE Communications Letters, vol. 25, no. 10, pp. 3398-3402, Oct. 2021, doi: 10.1109 / LCOMM.2021.3098235.), Bkassiny et al. based on the locally optimal Neyman-Pearson detector can effectively reduce the impact of non-Gaussian noise on the detection performance (M. Bkassiny, A. Lima De Sousa and S. K.Jayaweera, "Wideband Spectrum Sensing for Cognitive Radios in Weakly Correlated Non-Gaussian Noise," in IEEE Communications Letters, vol. 19, no. 7, pp. 1137-1140, July 2015, doi: 10.1109 / LCOMM.2015.2434996.)。.

[0005] The above spectrum sensing method can achieve good sensing performance only under relatively high signal-to-noise ratio conditions. However, in the actual electromagnetic environment, especially in the background with non-Gaussian noise, the lack of enhancement of the useful weak received signal may lead to a decline in sensing performance. Therefore, this paper considers enhancing the weak received signal first and proposes a distributed cooperative multi-antenna collaborative spectrum intelligent sensing method under non-Gaussian noise to improve the detection performance of spectrum sensing.

[0006] Through the above analysis, the problems and defects existing in the prior art are as follows:

[0007] (1) The existing spectrum sensing technologies only consider sensing the received signal under relatively high signal-to-noise ratio, and the existence of the unprocessed received signal at low zero power has the defect of deteriorating the spectrum sensing performance.

[0008] (2) Most of the existing cooperative spectrum sensing methods consider directly sending the sensing information of each sub-node to the fusion center, but this will greatly occupy the transmission bandwidth of the channel. Summary of the Invention

[0009] In order to overcome the problems existing in the above prior art, the purpose of the present invention is to provide a distributed cooperative multi-antenna collaborative spectrum intelligent sensing method, system, device and medium under non-Gaussian noise, which can enhance the weak received signal under non-Gaussian noise. Based on the spectrum sensing model with parameter sharing, it can effectively save the network transmission bandwidth and solve the problem of excessive bandwidth occupation during data transmission; realizing the distributed cooperative multi-antenna collaborative spectrum intelligent sensing method under non-Gaussian noise can provide a technical basis for cognitive radio systems in complex communication environments, break through the existing static spectrum allocation mechanism, and improve the utilization rate of spectrum resources.

[0010] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0011] A distributed cooperative multi-antenna collaborative spectrum intelligent sensing method under non-Gaussian noise, comprising the following steps:

[0012] Step 1: Construct a non-linear model based on fractional-order stochastic resonance to enhance the useful weak signal y(n) at the multi-antenna receiving end, and select the fractional lower-order covariance matrix R y as an intelligent representation of the preprocessed signal;

[0013] Step 2: Construct a distributed cooperative spectrum sensing system based on federated learning. Under this system, use a vision-based self-attention mechanism network model to complete the local training of sub-sensing nodes;

[0014] Step 3: Meanwhile, under this system, use the dynamic weighted federated averaging method to weighted aggregate the network parameters transmitted to the fusion center to obtain the global optimal network parameters;

[0015] Step 4: Use the obtained global optimal network parameter w * to load the global network model for sensing, and use the ratio of the two-dimensional feature vectors output by the network to construct the detection statistic T and detection threshold γ of spectrum sensing. Compare the magnitudes of the obtained detection statistic T and detection threshold γ to achieve distributed cooperative multi-antenna cooperative spectrum intelligent sensing under non-Gaussian noise.

[0016] Furthermore, the specific process of Step 1 is as follows:

[0017] Based on the multi-antenna cooperative spectrum sensing scenario, when the sensing model consists of one primary user, L secondary users, and one fusion center, the primary user transmitter has M antennas, and the secondary user receivers have K antennas. At the nth sampling moment, the received signal at the dth secondary user's receiver can be expressed as:

[0018] H0: y d (n) = w d (n)

[0019] H1: y d (n) = h d (n)s(n) + w d (n)

[0020] where H1 indicates the presence of an authorized primary user signal, H0 indicates the absence of an authorized primary user signal, s(n) represents the transmitted signal of the primary user, h d (n) = [h d,1 (n),..., h d,m (n),..., h d,K (n)], h d,m (n) represents the channel response on the mth receiving antenna of the dth secondary user, and w d (n) represents non-Gaussian noise; the matrix representation of the received signal y(n) is:

[0021]

[0022] where y M (N) is the Nth sampling data of the Mth antenna;

[0023] The non-Gaussian noise is described by alpha-stable distribution noise, and the expression form of its characteristic function is as follows:

[0024]

[0025]

[0026] where α represents the characteristic exponent, with a value range of 0 < α ≤ 2, and the impulsiveness of the alpha-stable distribution is determined by this parameter; μ represents the symmetry parameter, with a value range of -1 ≤ μ ≤ 1, and the symmetry degree of the alpha-stable distribution is determined by this parameter; γ represents the scale parameter, with a value range of γ ≥ 0, and the dispersion degree of the sample relative to the mean can be represented by this parameter, which is equivalent to the definition of variance in the Gaussian distribution; β represents the location parameter, with a value range of -∞ < β < +∞, and the offset of the probability density function of the stable distribution on the x-axis is determined by this parameter, and sign(t) represents the sign function;

[0027] The matrix representation of the received signal y(n) is:

[0028]

[0029] where y M (N) is the Nth sampling data of the Mth antenna;

[0030] The fractional-order stochastic resonance system is used to enhance the useful weak signal at the multi-antenna receiving end, and its fractional-order Langevin equation is expressed as:

[0031]

[0032] where y(t) represents the input signal, ξ(t) represents the alpha-stable distribution noise, a and b represent the system structure parameters, and V'(x) represents the first-order derivative of V(x), The operator is used to represent non-integer differentiation and integration, where the order is extended from integers to real numbers, including the fractional part, and a and t represent the upper and lower critical values, The operator is defined as:

[0033]

[0034] where α represents the order. When α > 0, it represents the operation of taking the α-order derivative; when α < 0, it represents the integration operation, with the upper integration limit being a and the lower integration limit being t;

[0035] Solve for the solution \(x(t)\) of the fractional Langevin equation with respect to the input signal \(y(t)\) under alpha-stable distribution noise, and this solution set is the enhanced useful received signal.

[0036] Calculate the fractional lower-order covariance matrix \(R\) y As an intelligent representation of the enhanced useful received signal:

[0037]

[0038] where \(|y(n)|\) p \( = [|y_1(n)| p , \cdots, |y m (n)| p , \cdots, |y K (n)| p T , \(y m (n)\) is the signal representation after enhancement by the fractional-order stochastic resonance system of the signal received by the \(m\)-th antenna at the sampling time \(n\), \(|\cdot|\) represents the absolute value symbol, \(T\) represents the transpose operation, \(H\) represents the conjugate transpose, and \(|y m (n)|\) p represents the \(p\)-th order fractional lower-order moment operation on \(|y m (n)|\).

[0039] Furthermore, the specific process of the second step is as follows:

[0040] First, calculate the fractional lower-order covariance matrix for the received signal preprocessed in the first step, extract the real and imaginary parts of the covariance matrix for normalization processing and grayscale transformation, and form a two-channel grayscale image input;

[0041] Then, before inputting the grayscale image block with dimensions \(h\times w\times c\) into the vision-based self-attention network model, slicing processing is required, where \(h\) represents the length of the grayscale image, \(w\) represents the width of the grayscale image, \(c\) represents the number of channels, and \(\times\) represents the dot product operation; flatten the input grayscale image into \(N = hw / p\) 2 vectors \(x\) 2 of size \((p p \cdot c)\), where \(N\) represents the total number of flattened vectors \(x\) p , \(p\) represents the length and width of the vector, \(c\) represents the number of channels, \( / \) represents the division operation, and \(\cdot\) represents the dot product operation; and perform dimensional compression on the vectors, use a fully connected layer to map them to dimension size \(d\) through a linear transformation; add a learnable position vector to perform position encoding on the processed vectors, and introduce a learnable embedding vector \(x\) class ​Classified as the final output feature, the feature vector after slicing, dimensionality reduction, and embedding can be expressed as where is the projection matrix, x class represents the classification vector, represents the flattened input vector, E pos represents the position encoding vector;

[0042] Next, the extracted feature vector z0 is fed into the encoder module. In the encoder module, first, the feature vector is processed by layer normalization (LayerNormalization, LN). LN(z0) means that the mean and variance of the input of each layer of neurons can be guaranteed to be consistent, where z0 represents the extracted feature vector, and LN represents the layer normalization operation; the processed vector is transformed to obtain three feature vectors Q, K, and V, all of size (N + 1)×d v , where N + 1 represents the length of the feature vector, and d v represents the width of the feature vector; through the single-head attention unit SA, the attention weight output vector can be obtained as This formula is used to calculate the similarity between different values in the feature vector, where is a set fixed value, T represents the transpose operation, and · represents the dot product operation between vectors. The calculation formula of the softmax activation function is as follows:

[0043]

[0044] where, x i is the output value of the i-th node, C is the number of output nodes, that is, the number of classification categories. Through this function, the output value of multi-classification can be converted into a probability distribution in the range of [0,1];

[0045] The multi-head attention layer will concatenate the output vectors of N h single-head attention units, which is expressed as where represents the input vector of the i-th single-head attention unit, concat(·) represents the operation of concatenating multiple vectors, LN represents the layer normalization operation, N h represents the number of single-head attention units, and SA represents the single-head attention unit;

[0046] Finally, perform a residual connection on the concatenated output vector and z0, and after layer normalization, use it as the input to a multi-layer perceptron (MLP), denoted as z'0 = LN(MSA(z0) + z0), where z0 represents the extracted feature vector, MSA(z0) represents the output of the multi-head attention unit, LN represents the layer normalization operation, and z'0 represents the output vector after the residual connection.

[0047] Furthermore, the multi-layer perceptron described in step 2 consists of two fully connected layers. The activation function of the first fully connected layer is GeLu, and the activation function of the second fully connected layer is softmax. The calculation formula is where represents the output vector of the fully connected layer, z'0 represents the output vector after the residual connection, GeLu(z'0) represents the non-linear transformation of the vector z'0, and the GeLu activation function is expressed as:

[0048] GeLu(x) = x·Φ(x)

[0049] where x represents the neuron input value, Φ(x) is the probability density function of the normal distribution, and · represents the dot product operation.

[0050] Furthermore, the specific process of step 3 is as follows:

[0051] First, in the distributed cooperative spectrum sensing system based on federated learning, the fusion center will first initialize the parameters of the network model at all child nodes to w 0 , and send this parameter to all sensing nodes;

[0052] Then, during the local training process, randomly select sensing nodes from all nodes. The selected sensing nodes will use their local datasets for multiple rounds of training; during the (k + 1)-th round of training, the child nodes will perform a local update of the network parameters. The local update rule formula is as follows:

[0053]

[0054] where represents the global model parameters globally aggregated after the k-th communication, represents the local model parameters of the i-th child node during the (k + 1)-th communication, η represents the learning rate, and g i represents the local gradient update of the i-th child node;

[0055] Next, after several local network parameter updates, global aggregation of the parameters of each child node is performed at the central node, and the dynamic weighted federated averaging method is introduced to aggregate the network parameters at each child node; the similarity between the local model at the k-th child node and the current global model can be expressed as:

[0056]

[0057] where [x] represents the floor function, that is, the largest integer not exceeding x, and |·| represents the absolute value operation, w k+1,i,j represents the j-th network parameter at the i-th child node during the (k + 1)-th iterative training, w k,i,global represents the corresponding global network parameter during the k-th iterative training, and w represents the global aggregated network parameter;

[0058] The similarity weight between the local model and the global model at the child node during the k-th iteration is expressed as follows:

[0059]

[0060] where N represents the total number of all child nodes, and softmax([L] k ) is used to convert [L] k into a probability value between [0, 1].

[0061] The weight constraint condition is expressed as follows:

[0062]

[0063] α k is dynamically changing during model training. The larger the weight coefficient, the higher the similarity between the global model and the local model. After the k-th iterative training, the aggregated global model parameters can be expressed as follows:

[0064]

[0065] where α i is the similarity weight between the local model and the global model at the child node during the i-th iteration, and w k,i represents the local network parameters of the i-th child node during the k-th iteration;

[0066] Finally, the global loss function of federated learning is obtained by weighting the local loss functions of each child node, and the calculation formula of the global loss function can be expressed as follows:

[0067]

[0068] where w represents the global network parameters, D irepresents the size of the local dataset of the i-th sensing node, l i (w) represents the loss function of the i-th child node on the global network parameters and the local training set, N represents the total number of sub-sensing nodes, l global (w) represents the global loss function under the global network parameters, and × represents the dot product operation;

[0069] For federated learning to reach the training termination, a global aggregation parameter must be obtained Under this parameter, the global loss function l global (w) can be minimized, and we have:

[0070]

[0071] where w is the global aggregation parameter during the iterative process, represents the global aggregation parameter obtained after training stops, and argmin means that when l global (w) takes the minimum value, is the value of;

[0072] Use the Stochastic Gradient Descent (SGD) method to optimize the global aggregation parameter to minimize the global loss function; assume that after τ steps of local updates by the child nodes, a global aggregation is performed, and the model training terminates to obtain the global network model parameter w * .

[0073] Furthermore, the specific process of step four is as follows:

[0074] Utilize the obtained global network parameter w * , load it into the self-attention network of the central node to obtain the network model for final detection; the output feature vector z'0 can be expressed as:

[0075]

[0076] where, represents a two-dimensional output vector, z'0 represents the output feature vector obtained from the final detection network model, represents the non-linear expression of the network model for the hypothesis H i , represents the entire non-linear expression of the trained network model, R represents the input vector, H i represents the binary hypothesis about the presence or absence of the primary user signal, i = 1 indicates the presence of the primary user signal, and i = 0 indicates the absence of the primary user signal;

[0077] And the selection vector is expressed as:

[0078]

[0079] Among them, when i = 1, It represents a two-dimensional vector indicating that the hypothesis H1 holds; when i = 0, It represents a two-dimensional vector indicating that the hypothesis H0 holds;

[0080] The detection statistic is constructed and expressed as:

[0081]

[0082] The detection threshold γ is expressed as:

[0083]

[0084] Among them, R u is the noise sample vector;

[0085] The data set under the given noise sample vector Among them, L represents the number of noise samples, K represents the size of each noise sample, represents the Lth noise sample vector of the input, and then input into the network model, a set of threshold values γ i can be obtained, where i ∈ {1, 2,..., L}; by sorting these values in descending order, the data set The detection threshold with the desired false alarm probability value α can be expressed as:

[0086]

[0087] Among them, α represents the set false alarm probability value, L represents the number of noise sample vectors, represents the obtained threshold data set, represents the floor symbol;

[0088] If T > η, it indicates the presence of the primary user signal, that is, the primary user spectrum is occupied. If T < η, it indicates the absence of the primary user signal, that is, the primary user is in spectrum idle.

[0089] Furthermore, a distributed cooperative multi-antenna cooperative spectrum intelligent sensing system under non-Gaussian noise includes:

[0090] An intelligent characterization module that uses the fractional-order stochastic resonance model to enhance the useful weak signal y(n) at the multi-antenna receiving end and selects the fractional lower-order covariance matrix R y as the feature representation of the preprocessed signal;

[0091] The child node local training and global parameter aggregation module constructs a distributed cooperative spectrum sensing system based on federated learning. In this system, a vision-based self-attention mechanism network model is used to complete the local training of sub-sensing nodes, and the dynamic weighted federated averaging method is used to weight and aggregate the network parameters transmitted to the fusion center to obtain the globally optimal network parameter w * ;

[0092] The spectrum detection module uses the obtained globally optimal network parameter w * to load the global network model for sensing, and constructs the detection statistic T and detection threshold γ of spectrum sensing using the ratio of the two-dimensional feature vectors output by the network. By comparing the magnitudes of the obtained detection statistic T and detection threshold γ, the distributed cooperative multi-antenna cooperative spectrum intelligent sensing under non-Gaussian noise is realized.

[0093] Furthermore, a distributed cooperative multi-antenna cooperative spectrum intelligent sensing device under non-Gaussian noise includes:

[0094] A memory for storing computer programs;

[0095] A processor for implementing the method for distributed cooperative multi-antenna cooperative spectrum intelligent sensing under non-Gaussian noise described in any one of steps one to four when executing the computer program.

[0096] Furthermore, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can perform distributed cooperative multi-antenna cooperative spectrum intelligent sensing under non-Gaussian noise.

[0097] The beneficial effects of the present invention are as follows:

[0098] 1. The present invention proposes a data preprocessing method based on fractional-order stochastic resonance, which converts part of the noise energy in the electromagnetic environment into weak input signal energy to enhance the input signal under low zero power.

[0099] 2. The present invention proposes a spectrum sensing method under non-Gaussian noise, which overcomes the drawback that traditional spectrum sensing algorithms perform well only under Gaussian noise but degrade under alpha-stable distribution noise, and has a good suppression effect on alpha-stable distribution noise.

[0100] 3. The present invention is a spectrum sensing model based on parameter sharing, which can effectively save network transmission bandwidth. At the same time, considering the potential imbalance in the data distribution of different sensing nodes, the idea of dynamic weighting is used to dynamically adjust the weights of sensing nodes, thereby realizing the optimization of the global model and being more applicable in the actual environment. Description of the Drawings

[0101] To more clearly and effectively illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for use in the embodiments of the present invention. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other accompanying drawings can be obtained based on these drawings without creative efforts.

[0102] Figure 1 It is a flowchart of a multi-antenna cooperative spectrum intelligent sensing method for distributed cooperation under non-Gaussian noise provided by an embodiment of the present invention.

[0103] Figure 2 It is a schematic structural diagram of a multi-antenna cooperative spectrum intelligent sensing system for distributed cooperation under non-Gaussian noise provided by an embodiment of the present invention.

[0104] Figure 3 It is a schematic diagram of the performance of a multi-antenna cooperative spectrum intelligent sensing for distributed cooperation under non-Gaussian noise provided by an embodiment of the present invention. Specific embodiments

[0105] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the following further details the present invention in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0106] Aiming at the problems existing in the prior art, the present invention provides a multi-antenna cooperative spectrum intelligent sensing method and system for distributed cooperation under non-Gaussian noise. The following describes the present invention in detail with reference to the accompanying drawings.

[0107] S101, using a fractional-order stochastic resonance model to enhance the useful weak signals at the multi-antenna receiving end, and selecting the fractional lower-order covariance matrix as the feature representation of the preprocessed signals;

[0108] S102, constructing a distributed cooperative spectrum sensing system based on federated learning. Under this system, using a vision-based self-attention mechanism network model to complete the local training of sub-sensing nodes;

[0109] S103, using the dynamic weighted federated averaging method to weight and aggregate the network parameters transmitted to the fusion center to obtain the globally optimal network parameters;

[0110] S104, using the obtained globally optimal network parameters to load the global network model for sensing, and using the ratio of the two-dimensional feature vectors output by the network to construct the detection statistic and detection threshold of spectrum sensing, and comparing the sizes of the obtained detection statistic and detection threshold to achieve multi-antenna cooperative spectrum intelligent sensing for distributed cooperation under non-Gaussian noise.

[0111] As shown Figure 2 in the figure, the multi-antenna cooperative spectrum intelligent sensing system for distributed cooperation under non-Gaussian noise provided by the embodiment of the present invention includes:

[0112] Intelligent characterization module 1, which uses the fractional-order stochastic resonance model to enhance the useful weak signal y(n) at the multi-antenna receiving end, and selects the fractional lower-order covariance matrix R y as the feature representation of the preprocessed signal;

[0113] Sub-node local training and global parameter aggregation module 2, which constructs a distributed cooperative spectrum sensing system based on federated learning. Under this system, using a vision-based self-attention mechanism network model, local training of sub-sensing nodes is completed, and the dynamic weighted federated averaging method is used to weight and aggregate the network parameters transmitted to the fusion center to obtain the global optimal network parameter w * ;

[0114] Spectrum detection module 3, which uses the obtained global optimal network parameter w * to load the global network model for sensing, and uses the ratio of the two-dimensional feature vectors output by the network to construct the detection statistic T and detection threshold γ of spectrum sensing, and compares the magnitudes of the obtained detection statistic T and detection threshold γ to achieve multi-antenna cooperative spectrum intelligent sensing for distributed cooperation under non-Gaussian noise.

[0115] The present invention will be further described below with reference to the embodiments.

[0116] A multi-antenna cooperative spectrum intelligent sensing method under non-Gaussian noise provided by the embodiment of the present invention includes the following steps:

[0117] S101, constructing a non-linear model based on fractional-order stochastic resonance to enhance the useful weak signal y(n) at the multi-antenna receiving end, and selecting the fractional lower-order covariance matrix R y as the intelligent characterization of the preprocessed signal.

[0118] Based on the multi-antenna cooperative spectrum sensing scenario, it is assumed that the sensing model consists of one primary user, L secondary users, and one fusion center. The primary user transmitter has M antennas, and the secondary user receivers have K antennas. At the nth sampling moment, the received signal of the dth secondary user can be expressed as:

[0119] H0:y d (n) = w d (n)

[0120] H1:y d (n) = h d (n)s(n) + w d (n)

[0121] Among them, H1 indicates the existence of an authorized primary user signal, H0 indicates the non-existence of an authorized primary user signal, s(n) represents the transmitted signal of the primary user, and h d (n) = [h d,1 (n),..., h d,m (n),..., h d,K (n)], y d,m (n) represents the channel response on the m-th receiving antenna of the d-th secondary user, and w d (n) represents non-Gaussian noise. The matrix representation of the received signal y(n) is:

[0122]

[0123] Among them, y M (N) is the N-th sampling data of the M-th antenna.

[0124] Alpha-stable distribution noise is used to describe non-Gaussian noise, and the expression form of its characteristic function is as follows:

[0125]

[0126]

[0127] Among them, α represents the characteristic exponent, with a value of 0 < α ≤ 2, and the impulsiveness of the alpha-stable distribution is determined by this parameter. μ represents the symmetry parameter, with a value of -1 ≤ μ ≤ 1, and the symmetry degree of the alpha-stable distribution is determined by this parameter. γ represents the scale parameter, with a value of γ ≥ 0, and the dispersion degree of the sample relative to the mean can be represented by this parameter, which is equivalent to the definition of variance in the Gaussian distribution. β represents the location parameter, with a value of -∞ < β < +∞, and the offset of the probability density function of the stable distribution on the x-axis is determined by this parameter, and sign(t) represents the sign function.

[0128] In the first step, a non-linear model based on fractional-order stochastic resonance is constructed to enhance the useful weak signal y(n) at the multi-antenna receiving end, and the fractional lower-order covariance matrix R y The specific process of using it as an intelligent characterization of the preprocessed signal is as follows:

[0129] Based on the multi-antenna cooperative spectrum sensing scenario, assume that the sensing model consists of a primary user, L secondary users, and a fusion center. The primary user transmitter has M antennas, and the secondary user receivers have K antennas. At the sampling time n, the received signal of the d-th secondary user can be expressed as:

[0130] H0: y d (n) = w d (n)

[0131] H1: y d r(n) = h d (n)s(n) + w d (n)

[0132] where H1 indicates the presence of an authorized primary user signal, H0 indicates the absence of an authorized primary user signal, s(n) represents the transmitted signal of the primary user, h d (n) = [h d,1 (n),..., h d,m (n),..., h d,K (n)], h d,m (n) represents the channel response on the m-th receiving antenna of the d-th secondary user, w d (n) represents non-Gaussian noise. The matrix representation of the received signal y(n) is:

[0133]

[0134] where y M (N) is the N-th sampling data of the M-th antenna.

[0135] The non-Gaussian noise is described by using alpha-stable distribution noise, and the expression form of its characteristic function is as follows:

[0136]

[0137]

[0138] where α represents the characteristic exponent, with the value range of 0 < α ≤ 2, and the impulsive degree of the alpha-stable distribution is determined by this parameter. μ represents the symmetry parameter, with the value range of -1 ≤ μ ≤ 1, and the symmetry degree of the alpha-stable distribution is determined by this parameter. γ represents the scale parameter, with the value range of γ ≥ 0, and the dispersion degree of the sample relative to the mean can be represented by this parameter, which is equivalent to the definition of variance in the Gaussian distribution. β represents the location parameter, with the value range of -∞ < β < +∞, and the offset of the probability density function of the stable distribution on the x-axis is determined by this parameter, and sign(t) represents the sign function.

[0139] The matrix representation of the received signal y(n) is:

[0140]

[0141] where y M (N) is the N-th sampling data of the M-th antenna.

[0142] The fractional-order Langevin equation for enhancing the useful weak signal at the multi-antenna receiving end by using the fractional-order stochastic resonance system is:

[0143]

[0144] Among them, y(t) represents the input signal, and ξ(t) represents the alpha-stable distribution noise. a and b represent the system structure parameters, and V'(x) represents the first-order derivative of V(x). The operator is used to represent non-integer differentiation and integration, where the order is extended from an integer to a real number, including the fractional part. a and t represent the upper and lower critical values. The operator is defined as:

[0145]

[0146] Among them, α represents the order. When α > 0, it represents the operation of taking the α-order derivative. When α < 0, it represents the integration operation, with the upper integration limit being a and the lower integration limit being t.

[0147] Solve for the solution x(t) of the fractional-order Langevin equation for the input signal y(t) under alpha-stable distribution noise. This solution set is the enhanced useful received signal.

[0148] Calculate the fractional lower-order covariance matrix R y as an intelligent characterization of the enhanced useful received signal:

[0149]

[0150] Among them, |y(n)| p = [|y1(n)| p ,..., |y m (n)| p ,..., |y K (n)| p T , y m (n) is the signal representation after enhancement by the fractional-order stochastic resonance system for the signal received by the m-th antenna at the sampling time n. |·| represents the absolute value symbol, T represents the transpose operation, H represents the conjugate transpose, and |y m (n)| p represents the p-order fractional lower-order moment operation on |y m (n)|.

[0151] S102. The specific process of constructing a non-linear model based on fractional-order stochastic resonance to enhance the useful weak signal y(n) at the multi-antenna receiver and selecting the fractional lower-order covariance matrix R y as an intelligent characterization of the preprocessed signal is as follows:

[0152] ​Then, before inputting the input grayscale image patch with dimensions h×w×c into the vision-based self-attention network model, slicing processing is required, where h represents the length of the grayscale image, w represents the width of the grayscale image, c represents the number of channels, and × represents the dot product operation. The input grayscale image is flattened into N = hw / p 2 vectors x of size (p 2 ·c), where N represents the total number of flattened vectors x p , p represents the length and width of the vector, c represents the number of channels, and / represents the division operation. And the dimension of the vector is compressed, and a fully connected layer is used to map it to the dimension size d through a linear transformation. Add a learnable position vector p to perform position encoding on the processed vector, and introduce a learnable embedding vector x as the final output feature for classification. The feature vector after slicing, dimensionality reduction, and embedding can be expressed as class where is the projection matrix, x represents the classification vector, class represents the flattened input vector, and E represents the position encoding vector. pos

[0153] Next, the extracted feature vector z0 is fed into the encoder module. In the encoder module, first, the feature vector is processed by layer normalization (LayerNormalization, LN). LN(z0) means that the mean and variance of the input of each layer of neurons can be made consistent, where z0 represents the extracted feature vector, and LN represents the layer normalization operation. The processed vector is transformed to obtain three feature vectors Q, K, and V, all of size (N + 1)×d v , where N + 1 represents the length of the feature vector, and d v represents the width of the feature vector. Through the single-head attention unit SA, the attention weight output vector can be obtained as This formula is used to calculate the similarity between different values in the feature vector, where is a set fixed value, T represents the transpose operation, and · represents the dot product operation between vectors. The calculation formula of the softmax activation function is as follows:

[0154]

[0155] where x i is the output value of the i-th node, and C is the number of output nodes, that is, the number of classification categories. Through this function, the output value of multi-classification can be converted into a probability distribution in the range of [0,1]. ​

[0156] The multi-head attention layer concatenates the output vectors of N h single-head attention units, which can be expressed as where represents the input vector of the i-th single-head attention unit, concat(·) represents the operation of concatenating multiple vectors, LN represents the layer normalization operation, N h represents the number of single-head attention units, and SA represents the single-head attention unit.

[0157] Finally, a residual connection is made between the concatenated output vector and z0, and after layer normalization, it is used as the input to a multi-layer perceptron layer (MLP), which can be expressed as z'0 = LN(MSA(z0) + z0), where z0 represents the extracted feature vector, MSA(z0) represents the output of the multi-head attention unit, LN represents the layer normalization operation, and z'0 represents the output vector after the residual connection. The multi-layer perceptron consists of two fully connected layers. The activation function of the first fully connected layer is GeLu, and the activation function of the second fully connected layer is softmax. The calculation formula is where represents the output vector of the fully connected layer, z'0 represents the output vector after the residual connection, GeLu(z'0) represents the non-linear transformation of the vector z'0, and GeLu

[0158] The activation function represents:

[0159] GeLu(x) = x·Φ(x)

[0160] where x represents the neuron input value, Φ(x) is the probability density function of the normal distribution, and · represents the dot product operation.

[0161] S103. The specific process of using the dynamic weighted federated average method to weight-aggregate the network parameters transmitted to the fusion center is as follows:

[0162] First, in the distributed cooperative spectrum sensing system based on federated learning, the fusion center will first initialize the parameters of the network model at all child nodes to w 0 , and send this parameter to all sensing nodes.

[0163] Then, during the local training process, 0.12 of the sensing nodes are selected from all the nodes, and the selected sensing nodes will use their local datasets for multiple rounds of training. When performing the (k + 1)-th round of training, the child nodes will perform a local update of the network parameters. The local update rule formula is as follows:

[0164]

[0165] Among them, represents the global model parameters after the k-th communication for global aggregation, represents the local model parameters of the i-th child node at the (k + 1)-th communication, η represents the learning rate, and g i represents the local gradient update of the i-th child node.

[0166] Next, after several local network parameter updates, global aggregation of the parameters of each child node is to be performed at the central node, and the dynamic weighted federated averaging method is introduced to aggregate the network parameters at each child node. The similarity between the local model at the k-th child node and the current global model can be expressed as:

[0167]

[0168] Among them, [x] represents the floor function, that is, the largest integer not exceeding x, |·| represents the absolute value operation, and w k+1,i,j represents the j-th network parameter at the i-th child node during the (k + 1)-th iterative training, and w k,i,global represents the corresponding global network parameter during the k-th iterative training, and w represents the global aggregated network parameter.

[0169] The similarity weight between the local model and the global model at the child node during the k-th iteration is expressed as follows:

[0170]

[0171] Among them, N represents the total number of all child nodes, and softmax([L] k ) is used to convert [L] k into a probability value between [0, 1].

[0172] The weight constraint condition is expressed as follows:

[0173]

[0174] α k is dynamically changing during model training. The larger the weight coefficient, the higher the similarity between the global model and the local model. After the k-th iterative training, the aggregated global model parameters can be expressed as follows:

[0175]

[0176] Among them, α i is the similarity weight between the local model and the global model at the child node during the i-th iteration, and w k,i represents the local network parameters of the i-th child node during the k-th iteration.

[0177] Finally, the global loss function of federated learning is obtained by weighting the local loss functions of each child node. The calculation formula of the global loss function can be expressed as follows:

[0178]

[0179] where \(w\) represents the global network parameters, \(D_i\) i represents the size of the local dataset of the \(i\)-th sensing node, \(l_i\) i (w) represents the loss function of the \(i\)-th child node on the global network parameters and the local training set, \(N\) represents the total number of child sensing nodes, and \(l\) global (w) represents the global loss function under the global network parameters, and \(\times\) represents the dot product operation.

[0180] For federated learning to reach the training termination, a global aggregation parameter must be obtained Under this parameter, the global loss function \(l\) global (w) can reach the minimum, and we have:

[0181]

[0182] where \(w^t\) is the global aggregation parameter during the iteration process, represents the global aggregation parameter obtained after training stops, and \(\text{argmin}\) means when \(l\) global (w) takes the minimum value, the value of \(w\).

[0183] The stochastic gradient descent (SGD) method is used to optimize the global aggregation parameter to minimize the global loss function. Assume that the child nodes perform a global aggregation after \(\tau\) steps of local updates, and the model training terminates to obtain the global network model parameters \(w^*\) * .

[0184] S104. Using the obtained global optimal network parameters \(w^*\) * Load the global network model for sensing, and construct the detection statistic \(T\) and detection threshold \(\gamma\) of spectrum sensing using the ratio of the two-dimensional feature vectors output by the network. Compare the sizes of the obtained detection statistic \(T\) and detection threshold \(\gamma\). The specific process of realizing distributed cooperative multi-antenna collaborative spectrum intelligent sensing under non-Gaussian noise is as follows:

[0185] Using the obtained global network parameters \(w^*\) * , load it into the self-attention network of the central node to obtain the network model for final detection. The output feature vector \(z'_0\) can be expressed as:

[0186]

[0187] where, denotes a two-dimensional output vector, and z'0 denotes the output feature vector obtained from the final detection network model. represents the network model for hypothesis H i as a non-linear expression, represents the entire non-linear expression of the trained network model, R represents the input vector, and H i represents the binary hypothesis regarding the presence or absence of the primary user signal. i = 1 indicates the presence of the primary user signal, and i = 0 indicates the absence of the primary user signal.

[0188] And the selection vector is expressed as:

[0189]

[0190] where, when i = 1, represents the two-dimensional vector indicating the establishment of hypothesis H1; when i = 0, represents the two-dimensional vector indicating the establishment of hypothesis H0.

[0191] The detection statistic is constructed and expressed as:

[0192]

[0193] The detection threshold γ is expressed as:

[0194]

[0195] where, R u is the noise sample vector.

[0196] The data set under the given noise sample vector where L represents the number of noise samples, K represents the size of each noise sample, represents the L-th noise sample vector of the input, and then input into the network model, a set of threshold values γ i , i ∈ {1, 2,..., L} can be obtained. By sorting these values in descending order, the data set Therefore, the detection threshold with the desired false alarm probability value α can be expressed as:

[0197]

[0198] where α represents the set false alarm probability value, L represents the number of noise sample vectors, represents the obtained threshold data set, represents the floor symbol.

[0199] If T > η, it indicates the presence of the primary user signal, that is, the primary user spectrum is occupied. If T < η, it indicates the absence of the primary user signal, that is, the primary user is in spectrum idle.

[0200] The technical effects of the present invention will be described in detail below in combination with simulation experiments.

[0201] To evaluate the performance of the present invention, simulation verification is carried out. In the simulation experiment, Matlab software is used to simulate MIMO signals, whose baseband modulation method is QPSK, the non-Gaussian noise is alpha-stable distribution noise, and the channel is a Rayleigh flat fading channel. The signal-to-noise ratio range of the signal is set to -20dB to 5dB. 1000 pairs of data are generated at each signal-to-noise ratio, and the number of sampling points for each pair of data is 1024. These received signals are enhanced, their fractional lower-order covariance is extracted and the label [1,0] is set T , indicating the presence of the primary user. In addition, the generated pure noise data is processed and the label [0,1] is set T , indicating the absence of the primary user, and the processing method is the same as above. Finally, a training data set with a total sample size of 36000 and a test data set of 12000 are generated. The local sub-sensing node network model uses cross-entropy as the loss function, and the dynamic weighted federated averaging method is used to aggregate the global network parameters at the fusion center. After multiple training iterations, the optimal global network parameters are obtained. The stochastic gradient descent (SGD) method is used to optimize the loss function during the local training process of the sub-nodes, and finally a trained network model is obtained. The simulation diagram of the results of the present invention is as Figure 3 shown. The method proposed in the present invention (distributed cooperative multi-antenna cooperative spectrum sensing) is compared with the single-node spectrum sensing methods based on deep learning (Transformer, AlexNet, SVM) and the performance based on energy detection (ED) as Figure 3 shown. It can be seen that the overall performance of the distributed cooperative multi-antenna cooperative spectrum sensing method is better than that of the single-node spectrum sensing method, which proves the effectiveness of the distributed cooperative multi-antenna cooperative spectrum sensing method.

[0202] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention by those skilled in the art within the technical scope disclosed by the present invention shall be covered by the protection scope of the present invention.

Claims

1. A method for distributed cooperative multi - antenna collaborative spectrum intelligent sensing under non - Gaussian noise, characterized in that, It includes the following steps: Step 1: Construct a non-linear model based on fractional-order stochastic resonance to enhance the useful weak signals at the multi-antenna receiving end and select the fractional lower-order covariance matrix as the intelligent representation of the preprocessed signals; the specific process is as follows: Based on the multi-antenna cooperative spectrum sensing scenario, when the sensing model consists of one primary user, L secondary users, and one fusion center, the primary user's transmitter has M antennas, the secondary users' receivers have K antennas, and at the nth sampling moment, the received signal at the receiver of the dth secondary user can be expressed as: Among them, indicates the presence of an authorized primary user signal, indicates the absence of an authorized primary user signal, indicates the transmitted signal of the primary user, , represents the th secondary user's channel response on the th receiving antenna, represents non-Gaussian noise; the matrix representation of the received signal Among them, is the th sampling data of the th antenna; The alpha-stable distribution noise is used to describe the non-Gaussian noise, and the expression form of its characteristic function is as follows: Among them, represents the characteristic exponent, and its value range is , and the pulse degree of the alpha-stable distribution is determined by this parameter; represents the symmetry parameter, and its value range is , and the symmetry degree of the alpha-stable distribution is determined by this parameter; represents the scale parameter, and its value range is , and the dispersion degree of the sample relative to the mean can be represented by this parameter, which is equivalent to the definition of variance in the Gaussian distribution; represents the location parameter, and its value range is , and the offset of the probability density function of the stable distribution on the axis is determined by this parameter, represents the sign function; Received signal The matrix representation of is as follows: Among them, is the th sampling data of the th antenna; The fractional-order stochastic resonance system is used to enhance the useful weak signal at the multi-antenna receiver, and its fractional-order Langevin equation is expressed as: Among them, represents the input signal, represents alpha-stable distribution noise, , and represent the system structure parameters, represents the first-order derivative of , The operator is used to represent non-integer differentiation and integration, where the order is extended from integers to real numbers, including the fractional part, and represent the upper and lower critical values, The operator is defined as: Among them, represents the order, when it is, it means to perform the operation of taking the derivative of the order, when it is, it means to perform the integration operation, and the upper limit of the integration is and the lower limit of the integration is ; Solve the solution of the fractional Langevin equation for the input signal under alpha-stable distribution noise This solution is the enhanced useful received signal ; Calculate the fractional lower-order covariance matrix As an intelligent representation of the enhanced useful received signal: Among them, , is the signal representation after enhancement by the fractional-order stochastic resonance system for the signal received by the m-th antenna at the sampling moment n, represents the absolute value symbol, represents the transpose operation, represents the conjugate transpose, represents for to perform order fractional lower-order moment operation; Step 2: Construct a distributed cooperative spectrum sensing system based on federated learning. Under this system, use a vision-based self-attention mechanism network model to complete the local training of sub-sensing nodes; the specific process is as follows: First, for the received signal preprocessed in Step 1, calculate the fractional lower-order covariance matrix, extract the real and imaginary parts of the covariance matrix for normalization processing and grayscale transformation, and form a two-channel grayscale image input; Then, when the input dimension is Before the grayscale image block is input into the vision-based self-attention network model, it needs to be sliced. Represents the length of the grayscale image, Represents the width of the grayscale image, Indicates the number of channels, Represents a dot product operation; flattens the input grayscale image into The size is Vector ,in Represents the flattened vector The total number of represents the length and width of a vector, Indicates the number of channels, represents the division operation, Represents the dot multiplication operation; and compresses the vector dimension, using a fully connected layer to map it to the dimension size through linear transformation ; Add a learnable position vector To positionally encode the processed vector and introduce a learnable embedding vector As the final output feature for classification, the feature vector after slicing, dimensionality reduction and embedding can be expressed as ,in is the projection matrix, represents the classification vector, represents the flattened input vector, represents the position encoding vector; Next, the extracted feature vectors are fed into the encoder module. In the encoder module, first, the feature vectors are processed by layer normalization (LayerNormalization, LN), which can make the means and variances of the inputs of each layer of neurons consistent. Among them, represents the extracted feature vectors, represents the layer normalization operation; the processed vectors are transformed to obtain three feature vectors 、 and , all of size . Among them, represents the length of the feature vector, represents the width of the feature vector; through the single-head attention unit SA, the attention weight output vector is . This formula is used to calculate the similarity between different values in the feature vector. Among them, , is a set fixed value, represents the transpose operation, represents the dot product operation between vectors. The calculation formula of the softmax activation function is as follows: Among them, is the output value of the th node, is the number of output nodes, that is, the number of categories of classification. Through this function, the output values of multi-classification can be converted into a probability distribution within the range of ; The multi-head attention layer concatenates the output vectors of single-head attention units, denoted as , where represents the input vector of the -th single-head attention unit, represents the concatenation operation on multiple vectors, represents the layer normalization operation, represents the number of single-head attention units, represents a single-head attention unit; Finally, the concatenated output vector and are subjected to residual connection, and after layer normalization, they are used as the input of a multi-layer perceptron layer (MLP), denoted as , where represents the extracted feature vector, represents the output of the multi-head attention unit, represents the layer normalization operation, represents the output vector after residual connection; Step 3: At the same time, under this system, use the dynamic weighted federated average method to weight and aggregate the network parameters transmitted to the fusion center to obtain the globally optimal network parameters; Step 4: Using the obtained globally optimal network parameters Load the global network model for sensing, and construct the detection statistic of spectrum sensing by using the ratio of two-dimensional feature vectors output by the network and the detection threshold . Compare the obtained detection statistic with the detection threshold to achieve distributed cooperative multi-antenna collaborative spectrum intelligent sensing under non-Gaussian noise.

2. The multi-antenna cooperative spectrum intelligent sensing method for distributed cooperation under non-Gaussian noise according to claim 1, characterized in that The multi-layer perceptron layer (MLP) described in step 2 consists of two fully connected layers. The activation function of the first fully connected layer is GeLu, and the activation function of the second fully connected layer is softmax. The calculation formula is , where represents the output vector of the fully connected layer, represents the output vector after residual connection, represents the non-linear transformation of the vector , The activation function represents: Among them, represents the neuron input value, is the probability density function of the normal distribution, represents performing a dot product operation.

3. The multi-antenna cooperative spectrum intelligent sensing method for distributed cooperation under non-Gaussian noise as described in claim 1, wherein The specific process of Step 3 is as follows: First, in a distributed collaborative spectrum sensing system based on federated learning, the fusion center will first initialize the parameters of the network model at all child nodes to , and send this parameter to all sensing nodes; Then, during the local training process, sensing nodes are randomly selected from all nodes, and the selected sensing nodes will use their local datasets for multiple rounds of training; during the round of training, a local update of the network parameters will be performed at the child nodes, and the local update rule formula is shown as follows: Among them, represents the global model parameters after global aggregation after the -th communication, represents the local model parameters of the -th child node during the -th communication, represents the learning rate, represents the local gradient update of the -th child node; Next, after several local network parameter updates, global aggregation of the parameters of each child node is performed at the central node, and a dynamic weighted federated averaging method is introduced to aggregate the network parameters at each child node; the similarity between the local model at the th child node and the current global model can be expressed as: Among them, represents the rounding function, that is, the largest integer not exceeding ; represents the absolute value operation, represents the th iteration training, at the th child node, the th network parameter, represents the corresponding global network parameter at the th iteration training, represents the global aggregation network parameter; The similarity weights of the local model and the global model at the child node at the th iteration are represented as follows: Among them, represents the total number of all child nodes, is used to convert to a probability value between ; The weight constraint condition is expressed as follows: It changes dynamically during model training. The greater the weight coefficient, the higher the similarity between the global model and the local model; after the th iteration of training, the aggregated global model parameters can be expressed as follows: Among them, is the similarity weight between the local model and the global model at the child node during the th iteration, represents the local network parameters of the th child node during the Finally, the global loss function of federated learning is obtained by weighting the local loss functions of each sub-node, and the calculation formula of the global loss function can be expressed as follows: Among them, represents the global network parameters, represents the size of the local dataset of the th sensing node, represents the total number of sub-sensing nodes, represents the global loss function under the global network parameters, represents performing a dot product operation; For federated learning to reach the training termination, it is necessary to obtain a global aggregation parameter , under which the global loss function reaches the minimum, and we have: Among them, Global aggregation parameters during the iteration process, represent the global aggregation parameters obtained after training stops, indicating that when takes the minimum value, the value of; Optimize the global aggregation parameters using the Stochastic Gradient Descent (SGD) method to minimize the global loss function; assume that after steps of local updates, a global aggregation is performed once, and the model training terminates to obtain the global network model parameters .

4. The multi-antenna cooperative spectrum intelligent sensing method for distributed cooperation under non-Gaussian noise as described in claim 1, wherein The specific process of Step 4 is as follows: Using the obtained global network parameters , load them into the self-attention network of the central node to obtain the network model for final detection; the output feature vector can be expressed as: Among them, represents a two-dimensional output vector, represents the output feature vector obtained by the final detection network model, represents the network model for the hypothesis of the non-linear expression, represents the entire non-linear expression of the trained network model, represents the input vector, represents the binary hypothesis made on the presence or absence of the primary user signal, represents the presence of the primary user signal, represents the absence of the primary user signal; And the selection vector is expressed as: Among them, When represents an identifier two-dimensional vector for which the hypothesis holds; , represents an identifier two-dimensional vector for which the hypothesis holds; Construct the detection statistic and express it as: Detection threshold Expressed as: Among them, is the noise sample vector; Dataset under a given noise sample vector , where represents the number of noise samples, represents the size of each noise sample, represents the th noise sample vector of the input, and then input into the network model, a set of threshold values , can be obtained; by sorting these values in descending order, a data set is constructed; the detection threshold with the desired false alarm probability value can be expressed as: Among them, represents the set false alarm probability value,[ represents the number of noise sample vectors,[ represents the obtained threshold data set,[ represents the floor symbol; If When, it indicates that the primary user signal exists, that is, the primary user spectrum is occupied. If When, it indicates that the primary user signal does not exist, that is, the primary user is in spectrum idle state.

5. The sensing system for the multi-antenna cooperative spectrum intelligent sensing method for distributed cooperation under non-Gaussian noise according to any one of claims 1-4, characterized in that It includes: Intelligent representation module, using the fractional-order stochastic resonance model, for the useful weak signals at the multi-antenna receiving end to enhance them, and selecting the fractional lower-order covariance matrix as the feature representation of the preprocessed signals; The sub-node local training and global parameter aggregation module constructs a distributed cooperative spectrum sensing system based on federated learning. Under this system, a vision-based self-attention mechanism network model is used to complete the local training of sub-sensing nodes, and the dynamic weighted federated averaging method is used to weight and aggregate the network parameters transmitted to the fusion center to obtain the globally optimal network parameters ; Spectrum detection module, using the globally optimal network parameters obtained Load the global network model for sensing, and construct the detection statistic of spectrum sensing using the ratio of the two-dimensional feature vectors output by the network and the detection threshold , compare the obtained detection statistic with the detection threshold to realize distributed cooperative multi-antenna collaborative spectrum intelligent sensing under non-Gaussian noise.

6. A multi-antenna cooperative spectrum intelligent sensing device for distributed cooperation under non-Gaussian noise, characterized in that, It includes: A memory for storing a computer program; A processor for implementing the method for multi-antenna cooperative spectrum intelligent sensing with distributed cooperation under non-Gaussian noise according to any one of claims 1 to 4 when executing the computer program.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it can perform multi-antenna cooperative spectrum intelligent sensing with distributed cooperation under non-Gaussian noise as claimed in claims 1-4.

Images