Frequency spectrum conflict time ratio prediction method based on GC-2LSTM-NN model

By fusing the spatial and temporal characteristics of the sub-transmitter using the GC-2LSTM-NN model, the problem of existing models being unable to accurately predict the spectrum collision time ratio is solved, thereby improving prediction accuracy and spectrum utilization.

CN118944786AActive Publication Date: 2024-11-12ZHEJIANG UNIV OF TECH
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410762186.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-13
Publication Date
2024-11-12
Estimated Expiration
2044-06-13

AI Technical Summary

Technical Problem

Existing deep learning models cannot effectively integrate the spatial and temporal correlations between sub-transmitters, making it difficult to accurately predict the frequency of spectral collisions.

Method used

We employ a GC-2LSTM-NN model to construct an asynchronous cognitive radio network, extract spatial features using graph convolutional layers, and combine this with a two-layer LSTM to learn time series features, thereby predicting the spectral conflict time ratio.

Benefits of technology

It improves the prediction accuracy of spectrum conflict time ratio, provides reliable information for subsequent scheduling tasks, and enhances spectrum utilization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118944786A_ABST
    Figure CN118944786A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of cognitive radio, and particularly relates to a spectrum conflict time ratio prediction method based on a GC-2LSTM-NN model. Spatial-temporal data of secondary transmitters are obtained from an asynchronous cognitive radio network, historical information of the secondary transmitters of past xi time slots in a sensing matrix of a t-th time slot is extracted and input into a GC-2LSTM-NN model, the GC-2LSTM-NN model comprises a graph convolution layer and a double-layer LSTM, the double-layer LSTM comprises N individual LSTMs, a group LSTM and a linear layer, and the N individual LSTMs and the group LSTM are distributed according to the graph convolution layer. And finally outputting predicted conflict time ratio vectors of H time slices of the time slot t. A back propagation algorithm is used to train the GC-2LSTM-NN model; and utilizing the trained GC-2LSTM-NN model to predict a conflict time ratio vector of H time slices of each time slot. According to the invention, the accuracy of conflict time ratio prediction is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of cognitive radio, in particular to the field of spectrum prediction technology in cognitive radio, and specifically relates to a spectrum conflict time ratio prediction method based on a GC-2LSTM-NN (Graph Convolution and Two-tier Long Short-Term Memory Neural Network) model. Background Art

[0002] In order to cope with the shortage of spectrum resources, cognitive radio networks use dynamic spectrum access technology to allow unauthorized wireless communication devices to share the idle spectrum resources of authorized devices under certain conditions. At the same time, in order to cope with the challenge of insufficient energy supply of wireless communication devices, energy capture technology enables wireless communication devices to obtain energy from the environment or dedicated energy sources. In an asynchronous energy capture cognitive radio network, the primary user authorizes the primary transmitter to access a certain bandwidth of spectrum resources and enjoys priority use rights to perform the primary user's data transmission tasks. Through spectrum sensing, spectrum decision-making, spectrum sharing and spectrum switching, the secondary user enables the secondary transmitter to use dynamic access to opportunistically access the primary user's authorized spectrum and capture energy from the primary user's wireless RF signal.

[0003] Since the spectrum state of the asynchronous energy capture cognitive radio network may change at any time, and the severity of transmission conflicts between primary and secondary users varies, it is necessary not only to pay attention to the changes in the spectrum state, but also to understand the impact of transmission conflicts on the primary user's transmission. The conflict time ratio can reflect the spectrum state while measuring the impact of transmission conflicts between primary and secondary users. Therefore, the conflict time ratio becomes a key reference for the impact of subsequent scheduling tasks on the spectrum state and transmission conflicts.

[0004] Predicting spectrum status has become a key issue in wireless spectrum management and cognitive radio networks. Spectrum status prediction technology can effectively help the fusion center predict spectrum status and provide important support for subsequent scheduling tasks. In an asynchronous energy capture cognitive radio network, the secondary receiver, as the fusion center, records and fuses the historical information of the secondary transmitter using a deep learning model, including the actions of the secondary transmitter (such as energy capture or spectrum sensing, etc.), the local sensing results of the secondary transmitter, and the multi-dimensional spatiotemporal variation characteristics of the local energy state to predict the future spectrum conflict time ratio. Accurate prediction of spectrum status helps the correct implementation of subsequent dynamic scheduling strategies, thereby improving spectrum utilization. Therefore, a deep learning model that can effectively learn and fuse the above multi-dimensional spatiotemporal variation characteristics is crucial for the fusion center to predict spectrum status.

[0005] However, the learning and fusion of multi-dimensional spatiotemporal variation features in the historical information of secondary transmitters is very complex, making it a challenge to accurately predict the spectrum conflict time ratio. At present, traditional deep learning model solutions, such as the model of convolutional neural network (CNN) fused with LSTM, are mostly used only to predict spectrum sensing data such as the cumulative strength of the received signal of the secondary transmitter. At the same time, CNN is usually used to process Euclidean spatial data with regular shapes such as images, and LSTM is used to process the correlation between time series. Both cannot describe the spatial correlation between secondary transmitters. The graph data structure can well describe the spatial correlation of secondary transmitters. With the rapid development of graph neural networks in recent years, graph convolutional neural networks can effectively extract the spatial correlation of graph data and have been widely used in graph data prediction tasks in non-Euclidean spaces such as knowledge graphs and network traffic. However, graph convolutional neural networks cannot effectively extract the temporal correlation between the historical information of secondary transmitters. Therefore, there is an urgent need for a deep learning model that can effectively fuse the multi-dimensional features of graph data space and time series, and can predict the spectrum conflict time ratio as accurately as possible, so as to provide more accurate spectrum status information for subsequent scheduling tasks. Summary of the invention

[0006] The purpose of the present invention is to solve the problem that the model of convolutional neural network fused with LSTM cannot describe the spatial correlation between sub-transmitters, and the graph convolutional neural network cannot effectively extract the temporal correlation between the historical information of sub-transmitters, making it difficult to accurately predict the spectrum conflict time ratio, and proposes a spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model.

[0007] In order to achieve the above object, the technical solution provided by the present invention is:

[0008] The spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model includes:

[0009] Constructing an asynchronous cognitive radio network, the asynchronous cognitive radio network comprising a primary network and a secondary network, the primary network comprising a primary user and an authorized channel, the secondary network comprising N single-antenna secondary transmitters and an N-antenna secondary receiver, and the transmission phase of the secondary network being divided into H equal-length time slices;

[0010] The spatiotemporal data of the secondary transmitter is obtained from the asynchronous cognitive radio network, wherein the spatiotemporal data includes an adjacency matrix with the secondary transmitter as a node. and a perception matrix formed by historical information of the secondary transmitter received by the secondary receiver within a predetermined time in the past;

[0011] Construct a GC-2LSTM-NN model, extract the historical information of the sub-transmitters in the past ξ time slots from the perception matrix of the t-th time slot, and calculate the real conflict time ratio vector y' of the H time slices contained in the t-th time slot as the label of the extracted perception matrix. Construct a separate undirected graph for each time slot

[0012] An undirected graph The graph convolution layer input into the GC-2LSTM-NN model calculates a spatial enhanced graph feature matrix

[0013] According to ξ spatial enhancement graph feature matrix Get the spatiotemporal feature matrix f st , extract the spatiotemporal feature matrix f st N individual time series features in the time dimension corresponding to each secondary transmitter in ;

[0014] Input N individual time series features into the N individual LSTMs in the first layer of the double-layer LSTM respectively, and output N enhanced individual time series feature results;

[0015] The N enhanced individual time series feature results are stacked on the time series feature dimension to obtain the global time series feature, and the global time series feature is input into the second layer group LSTM in the double-layer LSTM to obtain the enhanced global time feature result;

[0016] The enhanced global time feature result is flattened into a one-dimensional feature vector and input into the linear layer of the double-layer LSTM for fusion. The predicted conflict time ratio vector y of H time slices in the tth time slot is output in combination with the Sigmoid function.

[0017] The mean square error between the real conflict time ratio vector y' and the predicted conflict time ratio vector y is calculated as the loss function, and the GC-2LSTM-NN model is trained using the back propagation algorithm;

[0018] The trained GC-2LSTM-NN model is used to predict the conflict time ratio vector of H time slices in each time slot.

[0019] Furthermore, the adjacency matrix It is obtained by calculating the physical two-dimensional coordinates of the secondary transmitter and the radial basis function, where the adjacency matrix The matrix components of The calculation formula is as follows:

[0020]

[0021] Among them, di represents the physical two-dimensional coordinates of the secondary transmitter i, i∈[1,N], d j represents the physical two-dimensional coordinates of the secondary transmitter j, j∈[1,N], and ρ represents the preset hyperparameter.

[0022] Further, the extracting historical information of the secondary transmitters in the past ξ time slots from the perception matrix of the t-th time slot to obtain the extracted perception matrix includes:

[0023] Each time slot includes a control phase and a transmission phase. The control phase is divided into a sensing sub-phase and a reporting sub-phase.

[0024] In the sensing sub-phase, each sub-transmitter performs sensing independently, and y i (k) represents the kth sample of the primary network signal received by the secondary transmitter i, which is expressed by the following formula:

[0025]

[0026] Among them, w i (k) represents the noise of the kth sample of the primary network signal received by the secondary transmitter i, v i (k) represents the primary transmitter signal of the kth sample of the primary network signal received by secondary transmitter i, and They represent two situations respectively assuming that the main network is not using spectrum resources and is using spectrum resources;

[0027] The specific perception data is the cumulative received signal strength of the secondary transmitter, which is expressed by the formula:

[0028]

[0029] Among them, R i represents the cumulative received signal strength of the secondary transmitter i, and K represents the number of samples of the primary network signal received by the secondary transmitter i;

[0030] In the reporting sub-phase, each secondary transmitter reports the cumulative received signal strength and the remaining energy to the secondary receiver. If the remaining energy of the secondary transmitter is less than the preset energy, it remains silent in the reporting sub-phase;

[0031] The energy indicator for energy capture during the transmission phase is defined as:

[0032]

[0033] Among them, μ i (t) represents the energy indicator of the energy captured during the transmission phase of transmitter i in the tth time slot, represents the remaining energy of transmitter i at the end of the control phase of the tth time slot, represents the remaining energy of transmitter i at the beginning of the control phase of the t+1th time slot. i When (t) = 1, it means that the transmitter i is charged in the transmission phase of the tth time slot. i When (t) = -1, it means that the transmitter i is inactive during the transmission phase of the tth time slot. i When (t) = 0, it means that the transmitter i transmits the update packet in the transmission phase of the tth time slot;

[0034] Let λ i (t) indicates that in the tth time slot, the transmission indicator corresponding to the update packet sent by the secondary transmitter i is:

[0035]

[0036] Among them, when λ i (t) = 1 means that the secondary transmitter i successfully transmits the update packet to the secondary receiver in the tth time slot without collision. i (t) = -1 indicates that the update packet transmission fails. i (t) = 0 for other cases;

[0037] The extracted perception matrix S(t) is expressed as follows:

[0038]

[0039] Among them, s i (t)=(μ i (t-1),λ i (t-1),R i (t)) represents the historical information of the secondary transmitter provided by the secondary transmitter i in the tth time slot, i∈[1,N], μ i (t-1) represents the energy indicator of the energy captured during the transmission phase of transmitter i in the t-1th time slot, λ i (t-1) represents the transmission indicator corresponding to the update packet sent by transmitter i in the t-1th time slot, R i (t) represents the cumulative received signal strength of transmitter i at the tth time slot, Where N represents the number of sub-transmitters, ξ represents the time dimension, |s i (t)| represents the dimension of the specific content of the historical information provided by the secondary transmitter i in the tth time slot.

[0040] Further, the calculation of the real conflict time ratio vector y' of H time slices contained in the t-th time slot includes:

[0041] For the situation The actual conflict time ratio of the hth time slice in the tth time slot It is given by the following formula:

[0042]

[0043] Among them, τ t represents the duration of the transmission phase, τ I (t) represents the time that the primary user has been idle, x represents the duration of the primary user's idle or busy state, z represents the differential variable,

[0044] For the situation The actual conflict time ratio of the hth time slice in the tth time slot It is given by the following formula:

[0045]

[0046] in, as well as τ B (t) indicates the time the primary user has been busy;

[0047] The real conflict time ratios of H time slices are combined to obtain the real conflict time ratio vector y' of the t-th time slot.

[0048] Furthermore, the extracted perception matrix and adjacency matrix Construct a separate undirected graph for each time slot Including: let all components of the same column in the extracted perception matrix form a shape of N×|s i (t)|The graph signal feature matrix X t , the graph signal feature matrix X t Combined adjacency matrix Modeled as an undirected graph in the spatial dimension

[0049] Furthermore, the undirected graph The graph convolution layer input into the GC-2LSTM-NN model calculates a spatial enhanced graph feature matrix The calculation formula is as follows:

[0050]

[0051] Among them, ReLU represents the activation function, and its calculation expression is Represents the adjacency matrix of the self-connected relationship of the added nodes, that is I represents the identity matrix, express The diagonal matrix of The diagonal components of the matrix express The matrix components, W GCN are the learnable weights of the graph convolutional layer.

[0052] Further, the feature matrix of the spatial enhancement graph is Get the spatiotemporal feature matrix f st ,include:

[0053] The feature matrix of the spatial enhancement graph is Sort and stack by time dimension to obtain the spatiotemporal feature matrix f st , which can be expressed as follows:

[0054]

[0055] Among them, Concatenate represents the stacking operation, F E It indicates the number of enhanced spatial features and represents the feature dimension.

[0056] Compared with the prior art, the present invention has the following significant advantages: the GC-2LSTM-NN model uses the graph convolution layer GCN to enhance the expression of spatial features, and uses the double-layer LSTM to learn the time series features from individual to global, thus improving the learning and integration capabilities of the correlation of multi-dimensional spatiotemporal variation features of non-Euclidean spatial data. The GC-2LSTM-NN model improves the accuracy of conflict time ratio prediction and provides reliable information for subsequent scheduling tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of the overall structure of the present invention;

[0058] Figure 2 This is a diagram of the spatial feature fusion structure of the GC-2LSTM-NN model of the present invention;

[0059] Figure 3 This is a diagram of the temporal feature fusion structure of the GC-2LSTM-NN model of the present invention;

[0060] Figure 4 It is a training loss evaluation diagram of the GC-2LSTM-NN model of the present invention and the comparison model. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and 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.

[0062] The present invention provides a spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model, such as Figure 1 As shown, the specific method is as follows:

[0063] (1) Construct an asynchronous cognitive radio network and create a data set. The constructed asynchronous cognitive radio network includes a primary network and a secondary network. The primary network contains a primary user and an authorized channel. The secondary network includes N single-antenna secondary transmitters and an N-antenna secondary receiver. The secondary transmitter is equipped with an energy capture module and a radio frequency link module, and the secondary receiver is equipped with a stable power supply. There is no synchronous time-sharing communication protocol between the primary network and the secondary network. The secondary network follows the time division multiple access protocol, namely the H-TDMA protocol. Each time slot includes a control phase and a transmission phase. The transmission phase of the secondary network is divided into H equal-length time slices. The idle state of the authorized channel of the primary user follows a two-parameter Weibull distribution with a scale parameter α and a shape parameter λ. Its idle duration is represented by T I The busy state of the channel follows a two-parameter Weibull distribution with scale parameter β and shape parameter λ, which is the same as the idle state of the channel. Its busy duration is denoted as T B . T I and T B The expectations can be expressed as The secondary network dynamically accesses the authorized channel of the primary network to transmit update packets.

[0064] Furthermore, the control phase model is given. The control phase is divided into a sensing sub-phase and a reporting sub-phase. In the sensing sub-phase, each sub-transmitter performs sensing independently. Let y i (k) represents the kth sample of the primary network signal received by the secondary transmitter i, which can be expressed as the following binary hypothesis test: and Assume that the primary network is not using spectrum resources and is using spectrum resources:

[0065]

[0066] Among them, w i (k) represents the noise of the kth sample of the primary network signal received by the secondary transmitter i, v i (k) represents the primary transmitter signal of the kth sample of the primary network signal received by the secondary transmitter i. The specific perception data is the cumulative received signal strength of the secondary transmitter, which is expressed as:

[0067]

[0068] Among them, R i represents the cumulative received signal strength of the secondary transmitter i, and K represents the number of samples of the primary network signal received by the secondary transmitter i.

[0069] In the reporting sub-phase, each secondary transmitter first reports the cumulative received signal strength and residual energy to the secondary receiver as the fusion center, specifically, represents the remaining energy of transmitter i at the beginning of the control phase of the tth time slot, represents the remaining energy of secondary transmitter i at the end of the control phase of the tth time slot. If the secondary transmitter does not have enough energy to report its cumulative received signal strength and remaining energy, it remains silent during the reporting sub-phase.

[0070] Furthermore, the process of generating the sensing matrix S(t) by the secondary receiver is given. After receiving the reports from all the secondary transmitters, the secondary receiver performs information fusion: First, the energy indicator of energy capture in the transmission phase is defined as:

[0071]

[0072] in, represents the remaining energy of transmitter i at the beginning of the control phase of the t+1th time slot. i When (t) = 1, it means that the transmitter i is charged in the transmission phase of the tth time slot. i When (t) = -1, it means that the transmitter i is inactive during the transmission phase of the tth time slot. i When (t)=0, it indicates that the secondary transmitter i transmits the update packet in the transmission phase of the tth time slot.

[0073] Secondly, the transmission results of the past update packets record some historical information of the past state of the primary user network. Let λ i (t) indicates that in the tth time slot, the transmission indicator corresponding to the update packet sent by the secondary transmitter i is:

[0074]

[0075] Among them, when λ i (t) = 1 means that the secondary transmitter i successfully transmits the update packet to the secondary receiver in the tth time slot without collision. i (t) = -1 indicates that the update packet transmission fails. i (t) = 0 indicates other situations. The busy state of the channel can be determined from whether the transmission is successful or not. It should be noted that from the perspective of the secondary receiver, it knows the transmission indicators of all secondary transmitters, so the secondary transmitter only needs to report its cumulative received signal strength and remaining energy to the secondary receiver in the reporting sub-phase.

[0076] The perception matrix S(t) constructed for time slot t in the embodiment of the present invention is composed of the cumulative received signal strength, energy indicator and transmission indicator of the past ξ time slots, such as:

[0077]

[0078] Among them, the component s of the perception matrix S(t) i (t)=(μ i (t-1),λ i (t-1),R i (t)) corresponds to the historical information of transmitter i in the tth time slot. i (t-1) represents the energy indicator of the energy captured during the transmission phase of transmitter i in the t-1th time slot, λ i (t-1) represents the transmission indicator corresponding to the update packet sent by transmitter i in the t-1th time slot, R i (t) represents the cumulative received signal strength of transmitter i at the tth time slot, Where N represents the individual dimension of the sub-transmitter, ξ represents the time dimension, |s i (t)| represents the dimension of the specific content of the historical information provided by the secondary transmitter i in the tth time slot. The sensing matrix S(t) constructed in the tth time slot reflects the spatial and temporal characteristics of the state of the primary user network.

[0079] Furthermore, a specific calculation method for the real conflict time ratio is given. The real conflict time ratio of each time slice is divided into two cases according to the active state of the primary user at the end of the control phase: and At the end of the control phase of the tth time slot, for the case τ I (t) indicates the time that the primary user has been idle. τ B (t) indicates the time the primary user has been busy.

[0080] For the situation The actual conflict time ratio of the hth time slice in the tth time slot It is given by the following formula:

[0081]

[0082] Among them, τ t represents the duration of the transmission phase, x represents the duration of the idle or busy state of the primary user, as well as

[0083] For the situation The actual conflict time ratio of the hth time slice in the tth time slot It is given by the following formula:

[0084]

[0085] in, as well as The actual conflict time ratio of each time slice can be calculated by the above formula.

[0086] Furthermore, the network model constructed in this embodiment is simulated and modeled, and a perception matrix S(t) is generated and saved as the constructed data set input into the GC-2LSTM-NN model. Numerical integration calculation is performed according to the above-mentioned time slice conflict time ratio formula, and the real conflict time ratio vector y' of the H time slices contained in the tth time slot is saved as the constructed data set label.

[0087] (2) The GC-2LSTM-NN model learns and infers the input data. The structure of the GC-2LSTM-NN model is as follows: Figure 1 First, the GC-2LSTM-NN model splits the perception matrix of the input data shape of N×F×ξ according to the time dimension, and obtains ξ N×F matrices as the graph signal feature matrix of the tth time slot Where F represents the number of specific historical information features of the secondary transmitter in each time slot. The adjacency matrix of size N×N is obtained based on the physical two-dimensional coordinates of the secondary transmitter and the radial basis function calculation Specifically, its matrix components The calculation is as follows:

[0088]

[0089] Among them, d i represents the physical two-dimensional coordinates of the secondary transmitter i, d j represents the physical two-dimensional coordinates of secondary transmitter j, and ρ is a preset hyperparameter. The purpose of designing this adjacency matrix calculation method is to make full use of the prior knowledge that secondary transmitters with close physical distances have closer physical wireless communication environments. It should be noted that the present invention only calculates the adjacency matrix once. According to the adjacency matrix With ξ graph signal feature matrix X t , using the conventional mathematical steps for modeling graphs to construct an undirected graph Where V and E represent the nodes formed by the secondary transmitter and the adjacency matrix And the signal feature matrix X t Describes the edge relationship.

[0090] Next, we transform an undirected graph Input to the graph convolution layer (represented by GCN) in the GC-2LSTM-NN model for calculation to obtain ξ spatial enhanced graph feature matrices It should be noted that the GC-2LSTM-NN model uses spectral graph convolution as the calculation method of the GCN layer. Specifically, the calculation expression of the graph convolution layer in the GC-2LSTM-NN model is as follows:

[0091]

[0092] Among them, ReLU represents the activation function, and its calculation expression is is the adjacency matrix with node self-connection added, that is I represents the identity matrix, express The diagonal matrix of The diagonal components of the matrix in for The matrix components, W GCN is the learnable weight of the GCN layer. The specific process of the spatial feature fusion part is as follows Figure 2 shown.

[0093] Furthermore, the internal structure of LSTM is explained. LSTM consists of three basic components, namely input gate, forget gate and output gate. The expressions and descriptions of each gate are as follows:

[0094] 1. The input gate controls how to update the cell state to store new information, which is represented by I t , using the Sigmoid activation function to generate an output in the range of [0,1] to control the weight assigned to the input data. The expression of the input gate is as follows:

[0095] I t =σ(W I ·[h t-1 ,x t ]+b I )

[0096] Among them, σ() represents the Sigmoid activation function, h t-1 represents the hidden state of the t-1th time slot, x t Represents the input data of the tth time slot. I and b I They correspond to the learnable weights and biases of the input gate respectively.

[0097] 2. The forget gate controls the retention or discarding of information in the cell state, which is represented by F t , using the Sigmoid activation function, the expression of the forget gate is as follows:

[0098] Ft =σ(W F ·[h t-1 ,x t ]+b F )

[0099] Among them, W F and b F They represent the learnable weight and bias of the forget gate respectively.

[0100] 3. The output gate controls the output of the cell state and the update of the hidden state. The output gate is represented by O t , use the Sigmoid activation function to generate output values ​​in the range of [0,1], and use the Tanh activation function to determine the value of the current cell state to be propagated to the t+1th time slot. The expressions for the output gate to control the output, cell state, and hidden state are as follows:

[0101] O t =σ(W O ·[h t-1 ,x t ]+b O )

[0102] C t =f t ⊙C t-1 +I t ⊙Tanh(W C ·[h t-1 ,x t ]+b C )

[0103] h t =O t ⊙Tanh(C t )

[0104] Among them, W O and W c denote the learnable weights of the output gate and the unit state, respectively, and b O and b C denote the bias of the output gate and the unit state respectively, and ⊙ denotes element-wise multiplication.

[0105] Further, the double-layer LSTM structure proposed in the present invention is explained. The double-layer LSTM hierarchical structure proposed in the present invention is specifically as follows: the first layer is composed of N individual LSTMs, namely LSTM-1, LSTM-2, ..., LSTM-N, which learn and enhance the independent timing features of the sub-transmitters; the second layer is composed of a group LSTM and a linear layer, wherein the group LSTM is responsible for stacking the timing features output after learning and enhancing the individual LSTM in the feature dimension, and then learning and enhancing the global timing features, and then handing over the linear layer for feature fusion to output the predicted conflict time ratio vector. The linear layer is a single-layer perceptron. It should be noted that the N individual LSTMs set in the GC-2LSTM-NN model do not share weights.

[0106] Further, the specific process of using the double-layer LSTM proposed in the present invention to learn and fuse the temporal features of data is described. The specific process of using the double-layer LSTM to learn and fuse the temporal features of data is as follows: Stacking is performed on the time dimension to obtain the spatiotemporal feature matrix f st The stacking operation (represented by Concatenate) refers to connecting multiple matrices along a specified dimension to form a new multidimensional matrix. In the current step, the spatiotemporal feature matrix The ξ feature vectors of the i-th row in the time dimension are stacked in order, that is, where F E It indicates the number of enhanced spatial features and represents the feature dimension.

[0107] Next, extract the above-mentioned spatiotemporal feature matrix f st In corresponds to the N individual timing characteristics of the same transmitter in the time dimension, that is, The N individual time series features are input into the corresponding individual LSTMs respectively; after the N individual LSTMs learn and enhance the individual time series features, a total of N enhanced individual time series feature results are output, namely Among them, F T Represents the enhanced temporal feature dimension.

[0108] Then, the proposed GC-2LSTM-NN model stacks the N enhanced individual time series feature results in the time series feature dimension, and the overall feature after stacking is the time series feature F of N sub-transmitters. T Global temporal features composed of The * operator represents scalar multiplication. Input to the group LSTM (represented by Cooperative LSTM), learn and enhance, so as to obtain the final enhanced global time feature result Right now Finally, the GC-2LSTM-NN model will Perform the flattening operation, that is, Flatten it into a one-dimensional feature vector, input it into the linear layer for fusion, and finally combine it with the Sigmoid function to output the predicted conflict time ratio vector y of the H time slices of the predicted time slot t, that is:

[0109] Among them, Flatten is the flattening operation, Linear is the linear layer, σ is the Sigmoid function, and the Sigmoid function can be specifically expressed as: The temporal feature fusion part is as follows Figure 3 shown.

[0110] (3) The proposed GC-2LSTM-NN model is trained using the back propagation algorithm. The embodiment of the present invention uses the mean square error (MSE) as the loss function to calculate the loss Loss. Specifically, the GC-2LSTM-NN model calculates the mean square error of the predicted conflict time ratio vector y obtained by the perception matrix S(t) of a certain time slot input and the corresponding real conflict time ratio vector y' of the perception matrix S(t) in the data set to obtain the training loss (expressed as Loss), which is the following expression:

[0111]

[0112] Among them, y h represents the predicted conflict time ratio of the hth time slice, y' h Represents the actual conflict time ratio of the hth time slice.

[0113] Afterwards, the embodiment of the present invention uses the back propagation algorithm to derive the parameters in the GC-2LSTM-NN model and executes the gradient descent algorithm to update the existing parameters in the GC-2LSTM-NN model to minimize the loss, where the parameters are represented by θ. The specific update formula of the parameters in the gradient descent algorithm is: Where α is the preset initial hyperparameter learning rate.

[0114] Furthermore, the embodiment of the present invention uses the root mean square gradient descent method (Root Mean Square Propagation, RMSProp) to dynamically adjust the learning rate. In each iteration, the RMSProp algorithm dynamically adjusts the learning rate according to the historical gradient information, which makes the learning rate smaller for parameters with larger gradients and larger for parameters with smaller gradients. RMSProp enables the GC-2LSTM-NN model to better adapt to the update requirements of different parameters during the training process. Specifically, RMSProp uses exponentially weighted average to calculate the moving average of the square of the gradient. For the gradient of each parameter Compute the moving average of the gradient Where β is a decay factor, which is used to control the weight of the influence of historical gradient on current gradient, and its value is generally in the range of [0,1]. Next, adjust the learning rate For training, ε is a small constant used to avoid division by zero. The parameters are then updated using the adjusted learning rate.

[0115] Furthermore, the four models of CNN, LSTM, GCN and CNN-LSTM are applied to the tasks described in the embodiment of the present invention respectively, and the same simulation environment parameter settings are maintained. They are evaluated and compared with the GC-2LSTM-NN model proposed in the present invention. Figure 4 As shown in the figure, the proposed GC-2LSTM-NN model has a lower mean square error and can improve the prediction accuracy, which is more advantageous than the other models mentioned above.

[0116] The above embodiments may be presented in whole or in part in the form of a computer program product, which includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions according to the embodiments of the present application are generated in whole or in part.

[0117] The present invention focuses on providing a novel end-to-end deep learning model. The proposed GC-2LSTM-NN uses GCN to strengthen the expression of spatial features, and uses a two-layer LSTM to learn time series features from individual to global, which improves the learning and integration capabilities of the correlation of multi-dimensional spatiotemporal variation features of non-Euclidean spatial data. The GC-2LSTM-NN model improves the accuracy of conflict time ratio prediction and provides reliable information for subsequent scheduling tasks. The GC-2LSTM-NN model also provides new ideas and methods for further research in the field of spectrum state prediction technology in cognitive radio technology.

[0118] The above-mentioned embodiments only express one or several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A spectrum conflict time ratio prediction method based on GC-2LSTM-NN model, characterized in that: The spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model includes: Constructing an asynchronous cognitive radio network, the asynchronous cognitive radio network comprising a primary network and a secondary network, the primary network comprising a primary user and an authorized channel, the secondary network comprising N single-antenna secondary transmitters and an N-antenna secondary receiver, and the transmission phase of the secondary network being divided into H equal-length time slices; The spatiotemporal data of the secondary transmitter is obtained from the asynchronous cognitive radio network, wherein the spatiotemporal data includes an adjacency matrix with the secondary transmitter as a node. and a perception matrix formed by historical information of the secondary transmitter received by the secondary receiver within a predetermined time in the past; Construct a GC-2LSTM-NN model, extract the historical information of the sub-transmitters in the past ξ time slots from the perception matrix of the t-th time slot, and calculate the real conflict time ratio vector y' of the H time slices contained in the t-th time slot as the label of the extracted perception matrix. Construct a separate undirected graph for each time slot An undirected graph The graph convolution layer input into the GC-2LSTM-NN model calculates a spatial enhanced graph feature matrix According to ξ spatial enhancement graph feature matrix Get the spatiotemporal feature matrix f st , extract the spatiotemporal feature matrix f st N individual time series features in the time dimension corresponding to each secondary transmitter in ; Input N individual time series features into the N individual LSTMs in the first layer of the double-layer LSTM respectively, and output N enhanced individual time series feature results; The N enhanced individual time series feature results are stacked on the time series feature dimension to obtain the global time series feature, and the global time series feature is input into the second layer group LSTM in the double-layer LSTM to obtain the enhanced global time feature result; The enhanced global time feature result is flattened into a one-dimensional feature vector and input into the linear layer of the double-layer LSTM for fusion. The predicted conflict time ratio vector y of H time slices in the tth time slot is output in combination with the Sigmoid function. The mean square error between the real conflict time ratio vector y' and the predicted conflict time ratio vector y is calculated as the loss function, and the GC-2LSTM-NN model is trained using the back propagation algorithm; The trained GC-2LSTM-NN model is used to predict the conflict time ratio vector of H time slices in each time slot.

2. The spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model according to claim 1 is characterized in that: The adjacency matrix It is obtained by calculating the physical two-dimensional coordinates of the secondary transmitter and the radial basis function, where the adjacency matrix The matrix components of The calculation formula is as follows: Among them, d i represents the physical two-dimensional coordinates of the secondary transmitter i, i∈[1,N], d j represents the physical two-dimensional coordinates of the secondary transmitter j, j∈[1,N], and ρ represents the preset hyperparameter.

3. The spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model according to claim 1 is characterized in that: The step of respectively extracting historical information of the secondary transmitters in the past ξ time slots from the perception matrix of the t-th time slot to obtain the extracted perception matrix includes: Each time slot includes a control phase and a transmission phase. The control phase is divided into a sensing sub-phase and a reporting sub-phase. In the sensing sub-phase, each sub-transmitter performs sensing independently, and y i (k) represents the kth sample of the primary network signal received by the secondary transmitter i, which is expressed by the following formula: Among them, w i (k) represents the noise of the kth sample of the primary network signal received by the secondary transmitter i, v i (k) represents the primary transmitter signal of the kth sample of the primary network signal received by secondary transmitter i, and They represent two situations respectively assuming that the main network is not using spectrum resources and is using spectrum resources; The specific perception data is the cumulative received signal strength of the secondary transmitter, which is expressed by the formula: Among them, R i represents the cumulative received signal strength of the secondary transmitter i, and K represents the number of samples of the primary network signal received by the secondary transmitter i; In the reporting sub-phase, each secondary transmitter reports the cumulative received signal strength and the remaining energy to the secondary receiver. If the remaining energy of the secondary transmitter is less than the preset energy, it remains silent in the reporting sub-phase; The energy indicator for energy capture during the transmission phase is defined as: Among them, μ i (t) represents the energy indicator of the energy captured during the transmission phase of transmitter i in the tth time slot, represents the remaining energy of transmitter i at the end of the control phase of the tth time slot, represents the remaining energy of transmitter i at the beginning of the control phase of the t+1th time slot. i When (t) = 1, it means that the transmitter i is charged in the transmission phase of the tth time slot. i When (t) = -1, it means that the transmitter i is inactive during the transmission phase of the tth time slot. i When (t) = 0, it means that the transmitter i transmits the update packet in the transmission phase of the tth time slot; Let λ i (t) indicates that in the tth time slot, the transmission indicator corresponding to the update packet sent by the secondary transmitter i is: Among them, when λ i (t) = 1 means that the secondary transmitter i successfully transmits the update packet to the secondary receiver in the tth time slot without collision. i (t) = -1 indicates that the update packet transmission fails. i (t) = 0 for other cases; The extracted perception matrix S(t) is expressed as follows: Among them, s i (t)=(μ i (t-1),λ i (t-1),R i (t)) represents the historical information of the secondary transmitter provided by the secondary transmitter i in the tth time slot, i∈[1,N], μ i (t-1) represents the energy indicator of the energy captured during the transmission phase of transmitter i in the t-1th time slot, λ i (t-1) represents the transmission indicator corresponding to the update packet sent by transmitter i in the t-1th time slot, R i (t) represents the cumulative received signal strength of transmitter i at the tth time slot, Where N represents the number of sub-transmitters, ξ represents the time dimension, |s i (t)| represents the dimension of the specific content of the historical information provided by the secondary transmitter i in the tth time slot.

4. The spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model according to claim 3 is characterized in that: The calculation of the real conflict time ratio vector y' of H time slices contained in the t-th time slot includes: For the situation The actual conflict time ratio of the hth time slice in the tth time slot It is given by the following formula: Among them, τ t represents the duration of the transmission phase, τ I (t) represents the time that the primary user has been idle, x represents the duration of the primary user's idle or busy state, z represents the differential variable, For the situation The actual conflict time ratio of the hth time slice in the tth time slot It is given by the following formula: in, as well as τ B (t) indicates the time the primary user has been busy; The real conflict time ratios of H time slices are combined to obtain the real conflict time ratio vector y' of the t-th time slot.

5. The spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model according to claim 3 is characterized in that: The extracted perception matrix and adjacency matrix Construct a separate undirected graph for each time slot Including: let all components of the same column in the extracted perception matrix form a shape of N×|s i (t)|The graph signal feature matrix X t , the graph signal feature matrix X t Combined adjacency matrix Modeled as an undirected graph in the spatial dimension 6. The spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model according to claim 5 is characterized in that: The undirected graph The graph convolution layer input into the GC-2LSTM-NN model calculates a spatial enhanced graph feature matrix The calculation formula is as follows: Among them, ReLU represents the activation function, and its calculation expression is Represents the adjacency matrix of the self-connected relationship of the added nodes, that is I represents the identity matrix, express The diagonal matrix of The diagonal components of the matrix express The matrix components, W GCN are the learnable weights of the graph convolutional layer.

7. The spectrum conflict time ratio prediction method based on the GC-2LSTM-NN model according to claim 1 is characterized in that: According to the ξ spatial enhancement map feature matrix Get the spatiotemporal feature matrix f st ,include: The feature matrix of the spatial enhancement graph is Sort and stack by time dimension to obtain the spatiotemporal feature matrix f st , which can be expressed as follows: Among them, Concatenate represents the stacking operation, F E It indicates the number of enhanced spatial features and represents the feature dimension.

Citation Information

Patent Citations

  • Multi-channel spectrum sensing and energy capturing method for cognitive radio energy supply network

    CN117692911A

  • Wireless spectrum sensing data prediction method based on graph convolutional network

    CN117834061A

  • Electronic device for predicting channel in MIMO communication system and method of predicting the same

    US20240048257A1