A spectrum availability prediction method and system based on wireless communication network
Through deep learning Bi-LSTM model and Bayesian optimization technology, the problems of high complexity and low accuracy of existing spectrum availability prediction methods are solved, and the rapid discovery of available spectrum gaps in complex electromagnetic environments is realized, and wireless communication is improved The anti-interference ability of the network.
Patent Information
- Application Number
- CN202111507860.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-10
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2041-12-10
AI Technical Summary
The existing spectrum availability prediction methods have high computational complexity and need to be improved. Especially in complex electromagnetic environments, it is difficult to quickly discover available spectrum gaps, and the anti-interference ability is insufficient.
Using the Bi-LSTM neural network model based on deep learning, combined with Bayesian optimization and gradient descent algorithm, the optimal model is obtained through offline training, predict the spectrum availability of future time slots online, and fine-tune hyperparameters to improve prediction using Bayesian optimization accuracy.
In the presence of malicious interference, accurately predicting future time slots can be used for spectral gaps, which improves the anti-interference ability of the wireless communication network, reduces prediction errors, and improves communication stability.
Smart Images

Figure CN114298166B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of wireless communications, and in particular relates to a spectrum availability prediction method and system based on a wireless communication network. Background Art
[0002] With the popularization of 5G communication technology and the advancement of Internet of Things technology, frequency-using devices have shown explosive growth, which has led to an increasing shortage of spectrum resources. In actual cognitive wireless communication networks, cognitive users conduct regular communications with base stations or other cognitive users, and malicious users interfere with communication links in wireless communication networks based on a certain attack task or purpose. Therefore, in order to avoid interference, users in cognitive wireless communication networks need to be able to avoid interfered frequency bands in complex and changeable electromagnetic environments and quickly discover available spectrum. The spectrum availability prediction method can predict future spectrum gaps, and cognitive users can make judgments before interference arrives and use idle frequency bands for communication, thereby improving the network's anti-interference ability.
[0003] At present, spectrum availability prediction mainly focuses on methods based on regression analysis, Markov chain, moving average and machine learning. The computational complexity of the first three methods is relatively high. For example, the update of regression coefficients and the increase of Markov order lead to exponential growth of estimated parameters. At the same time, they rely on prior information of the spectrum. As for machine learning-based methods, in recent years, data-driven deep learning methods have been the main focus, such as methods based on MLP, ANN and LSTM. Although these methods can effectively complete the prediction, the accuracy of the prediction needs to be improved. Summary of the invention
[0004] In view of the deficiencies in the prior art, the present invention provides a spectrum availability prediction method and system based on a wireless communication network.
[0005] In a first aspect, the present invention provides a spectrum availability prediction method based on a wireless communication network, the method being applied to a wireless communication network including a plurality of cognitive users and a malicious user, comprising:
[0006] Acquire a spectrum dataset for offline training of a deep learning model, wherein the spectrum dataset includes a spectrum training set, a spectrum verification set, and a spectrum test set;
[0007] Set the number of Bayesian optimization cycles, the search space for hyperparameters, and the number of deep learning model iterations;
[0008] Initialize the structure, weights, biases, and hyperparameters of the deep learning model;
[0009] Compute the loss between the input and output of a deep learning model for a spectral dataset;
[0010] Use gradient descent to update the weights and bias parameters of the deep learning model until the deep learning model training is completed;
[0011] Determine whether the loss curve and accuracy fitting curve of the deep learning model are in a convergence state;
[0012] If not, reset the number of deep learning model iterations until the loss curve and accuracy fitting curve of the deep learning model are in a convergence state;
[0013] If so, Bayesian optimization is used to fine-tune the hyperparameters of the deep learning model in the hyperparameter search space until the cycle ends and the optimal deep learning model is obtained; the obtained optimal deep learning model is used to predict the available spectrum online.
[0014] Furthermore, the deep learning model is a Bi-LSTM neural network.
[0015] Further, the calculating the loss of the spectrum dataset between the input and output of the deep learning model includes:
[0016] The loss between the input and output of the deep learning model for the spectral dataset is calculated according to the following formula:
[0017]
[0018] Where, χ is the loss of spectrum data between the input and output of the deep learning model; t is the time slot; T is the total number of time slots; p t is the observed spectral power; is the predicted spectral power; η is the regularization coefficient; L is the number of layers of the Bi-LSTM neural network; w l is the weight vector of the l-th layer Bi-LSTM neural network.
[0019] Furthermore, the method of updating the weights and bias parameters of the deep learning model by gradient descent until the deep learning model training is completed includes:
[0020] Constructing the objective function
[0021] The algorithm to obtain the optimal U is:
[0022]
[0023] Among them, W(s) and b(s) are the weight and bias of the deep learning model at the sth training time; α is the learning rate of the deep learning model.
[0024] Furthermore, the Bayesian optimization adopts a Gaussian process probability regression model and an ExpectedImprovement acquisition function.
[0025] In a second aspect, the present invention provides a spectrum availability prediction system based on a wireless communication network, wherein the system is applied to a wireless communication network including a plurality of cognitive users and a malicious user, and comprises:
[0026] An acquisition module is used to acquire a spectrum data set for offline training of a deep learning model, wherein the spectrum data set includes a spectrum training set, a spectrum verification set, and a spectrum test set;
[0027] The first setting module is used to set the number of Bayesian optimization cycles, the search space of hyperparameters, and the number of deep learning model iterations;
[0028] Initialization module, used to initialize the structure, weights, biases, and hyperparameters of the deep learning model;
[0029] A calculation module for calculating the loss between the input and output of the deep learning model for the spectral dataset;
[0030] An update module, which is used to update the weight and bias parameters of the deep learning model using gradient descent until the deep learning model training is completed;
[0031] A judgment module is used to judge whether the loss curve and accuracy fitting curve of the deep learning model are in a convergence state;
[0032] A second setting module is used to reset the number of iterations of the deep learning model when the judgment module determines that the loss curve and the accuracy fitting curve of the deep learning model are not in a convergence state, until the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state;
[0033] The fine-tuning module is used to fine-tune the hyperparameters of the deep learning model by using Bayesian optimization in the hyperparameter search space when the judgment module determines that the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state until the cycle ends and the optimal deep learning model is obtained; the obtained optimal deep learning model is used to predict the available spectrum online.
[0034] Furthermore, the calculation module includes:
[0035] A computation unit for calculating the loss between the input and output of the deep learning model for the spectral dataset according to the following formula:
[0036]
[0037] Where, χ is the loss of spectrum data between the input and output of the deep learning model; t is the time slot; T is the total number of time slots; p t is the observed spectral power; is the predicted spectral power; η is the regularization coefficient; L is the number of layers of the Bi-LSTM neural network; w l is the weight vector of the l-th layer Bi-LSTM neural network.
[0038] Furthermore, the update module includes:
[0039] Building blocks for constructing objective functions
[0040] The algorithm for obtaining the optimal U is:
[0041]
[0042] Among them, W(s) and b(s) are the weight and bias of the initialized deep learning model when training for the sth time; α is the learning rate of the deep learning model.
[0043] The present invention provides a spectrum availability prediction method and system based on a wireless communication network, wherein the method comprises obtaining a spectrum data set for offline training of a deep learning model, wherein the spectrum data set comprises a spectrum training set, a spectrum verification set and a spectrum test set; setting the number of cycles of Bayesian optimization, the search space of hyperparameters and the number of iterations of the deep learning model; initializing the structure, weights, biases and hyperparameters of the deep learning model; calculating the loss of the spectrum data set between the input and output of the deep learning model; updating the weights and bias parameters of the deep learning model by using gradient descent until the training of the deep learning model is completed; judging whether the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state; if not, resetting the number of iterations of the deep learning model until the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state; if yes, fine-tuning the hyperparameters of the deep learning model by using Bayesian optimization in the hyperparameter search space until the cycle is completed to obtain an optimal deep learning model; and using the obtained optimal deep learning model to predict the spectrum that can be used online. The present invention adopts the above scheme, and can accurately predict the spectrum gaps available in future time slots, so that cognitive users are protected from interference from malicious users, and normal communication between cognitive users and base stations and between cognitive users is guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0045] Figure 1 A flowchart of a spectrum availability prediction method based on a wireless communication network provided by an embodiment of the present invention;
[0046] Figure 2 A schematic diagram of the structure of a spectrum availability prediction system based on a wireless communication network provided by an embodiment of the present invention;
[0047] Figure 3 A scene diagram of a wireless communication network communication link being disturbed provided by an embodiment of the present invention;
[0048] Figure 4 A schematic diagram of the structure of a spectrum availability prediction method based on a wireless communication network provided by an embodiment of the present invention;
[0049] Figure 5 A comparison diagram of prediction errors between a spectrum availability prediction method based on a wireless communication network provided by an embodiment of the present invention and other methods;
[0050] Figure 6 A comparison chart of MAE and RMSE between a spectrum availability prediction method based on a wireless communication network provided by an embodiment of the present invention and other methods;
[0051] Figure 7 A Rank correlation comparison diagram of a spectrum availability prediction method based on a wireless communication network provided by an embodiment of the present invention and other methods;
[0052] Figure 8 A comparison chart of the prediction accuracy of a spectrum availability prediction method based on a wireless communication network provided by an embodiment of the present invention and other methods under different signal-to-noise ratios. DETAILED DESCRIPTION
[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0054] like Figure 1 As shown, the present invention provides a spectrum availability prediction method based on a wireless communication network, the method is applied to a wireless communication network including a plurality of cognitive users and a malicious user, the method comprising:
[0055] Step S101, obtaining a spectrum data set for offline training of a deep learning model, wherein the spectrum data set includes a spectrum training set, a spectrum verification set, and a spectrum test set.
[0056] In this step, the deep learning model is a Bi-LSTM neural network. We consider a wireless communication network consisting of malicious users and multiple cognitive users. Figure 3 Specifically, there are cognitive users communicate with the base station, all cognitive users communicate in the available frequency band, and each cognitive user does not interfere with each other. A malicious user interferes with the communication link between the jth cognitive user and the base station, and the interference behavior of the malicious user is observed. Then the spectrum data measured at the tth time slot of the nth channel at the cognitive user receiving end can be expressed as:
[0057]
[0058] in, is the signal energy (power level) of the jth cognitive user, is the signal energy (power level) of the malicious user, is the Gaussian white noise power. or is the indicator function, or Indicates that the signal does not exist. or Indicates that the signal exists. In the embodiment of the present invention, we consider a strong interference situation. For the convenience of theoretical analysis, we only consider the downlink of the interference communication. In addition, the cognitive user knows the communication frequency band of other users. Based on Figure 3 The time-frequency diagram in the figure can be used to make a power matrix of the spectrum situation (available or unavailable) that changes over time in different frequency bands. The rows of the matrix represent different channels (frequency bands) and the columns represent different time slots. Then the power vector of all time slots of the nth channel is expressed as:
[0059]
[0060] in, represents the power value of the tth time slot of the nth channel, t=1,2,3,…,T; n=1,2,3,…,N. The power matrix observed in N channels can be expressed as:
[0061] Γ=(p 1 ,p 2 ,p 3 ,…,p n ) T
[0062] Therefore, the spectrum availability prediction problem can be expressed as:
[0063]
[0064] in, represents the power value predicted by N channels in time slot t+1. M is the size of the observation time, and p(·) is the conditional probability, that is, predicting the spectrum data of the next time slot under the observed spectrum data.
[0065] The scenario of user communication link being interfered is simulated through a real experimental platform to generate spectrum data sets for deep learning model training. The jammer is implemented by USRP 2943R, using multiple interference modes (such as sweep frequency interference, fixed frequency interference, etc.), with a signal-to-interference ratio of 10dB and an interference signal strength (power) greater than -50dBm. Assume that the frequency band interfered by malicious users is between 1300MHz and 1700MHz, and each 20MHz interval is divided into a channel, with a total of 20 channels and a signal-to-noise ratio SNR = 20dB. The spectrum sensor (detector) collects power every 0.1 seconds. When the power exceeds -50dBm, it is considered to be interfered (i.e., communication is restricted). A total of 4000 time slots of data are collected. Cognitive users are aware of the working frequency bands of other cognitive users, and the power of cognitive users does not exceed -50dBm.
[0066] Step S102, setting the number of Bayesian optimization cycles, the search space of hyperparameters, and the number of deep learning model iterations.
[0067] Step S103, initializing the structure, weights, biases, and hyperparameters of the deep learning model.
[0068] In this step, the network configuration of the deep learning model is initialized: the hidden layer is 1, the number of LSTM units is 50, the learning rate is 0.01, and the L2 regularization coefficient is 10 -10 , the learning rate decrease factor is 0.5, MaxEpochs is 200 times, and the MiniBatch size is 100.
[0069] like Figure 4 As shown, first, the observed interference spectrum evolution (power value) matrix P passes through the forget gate of the Bi-LSTM neural network, which determines the output s at the previous moment. t-1 and the cell state c t-1 Keep some information until the current time c t . The forget gate formula is expressed as:
[0070] f t =σ(W f [s t-1 ,p t ]+b f )
[0071] Among them, σ represents the Sigmoid function, x t represents the input at the current moment, W f and b f Represent the weight and bias of the forget gate respectively. Then enter the input gate, which determines p t What information is retained by c t, and use Sigmoid and Tanh activation functions to implement state unit c t The update of . The formula is expressed as:
[0072] i t =σ(W i [s t-1 ,p t ]+b i )
[0073]
[0074] Among them, λ represents the Tanh function, W i and b i Represent the weight and bias of the input gate respectively. W c and b c Respectively represent the weight and bias of the selected update unit. The updated c t for:
[0075]
[0076] Among them, * represents the dot product between elements. After selective memory and update, the information finally enters the output gate. It completes two tasks: filtering information and obtaining output; state unit c t is selectively output to the next moment and output to the outside. The formula is:
[0077] o t =σ(W o [s t-1 ,p t ]+b o )
[0078] s t =o t *λ(c t )
[0079] Among them, W o and b o Represent the weight and bias of the output gate respectively. Then the data is expanded in the fully connected layer and predicted in the regression output layer. Finally, a hard decision is made on the predicted power to determine the availability of the channel. Specifically, it can be expressed as:
[0080]
[0081] Among them, "0" represents that the channel is available, and "1" represents that the channel is not available. λ is the decision threshold.
[0082] Step S104, calculating the loss of the spectrum dataset between the input and output of the deep learning model.
[0083] In this step, the loss between the input and output of the deep learning model is calculated according to the following formula:
[0084]
[0085] Where, χ is the loss of spectrum data between the input and output of the deep learning model; t is the time slot; T is the total number of time slots; p t is the observed spectral power; is the predicted spectral power; η is the regularization coefficient; L is the number of layers of the Bi-LSTM neural network; w l is the weight vector of the l-th layer Bi-LSTM neural network.
[0086] Step S105, using gradient descent to update the weights and bias parameters of the deep learning model until the deep learning model training is completed.
[0087] In this step, first let U = {W, b}, and the objective function is
[0088]
[0089] Then, the Adam optimization algorithm to obtain the optimal U is
[0090]
[0091] Among them, W(s) and b(s) are the weight and bias of the deep learning model at the sth training time; α is the learning rate of the deep learning model.
[0092] Step S106, determining whether the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state.
[0093] Step S107, if not, reset the number of deep learning model iterations until the loss curve and accuracy fitting curve of the deep learning model are in a convergence state.
[0094] Step S108, if yes, use Bayesian optimization to fine-tune the deep learning model hyperparameters in the hyperparameter search space until the loop ends and the optimal deep learning model is obtained; use the obtained optimal deep learning model to predict the available spectrum online.
[0095] In this step, first, assume that there are m hyperparameters that need to be optimized:
[0096]
[0097] in, Represents the value space of all hyperparameters. (m)Represents the value space of a certain hyperparameter. From each hyperparameter value space R (m) Select a value from the list to form a set of hyperparameter combinations:
[0098] r=(r 1 ,r 2 ,...,r m )
[0099] Among them, r m represents a random value of the mth hyperparameter, r m ∈R (m) . Then, define an objective function f: Here, the prediction error of the deep learning validation set is used as the objective function for optimization. The optimal set of hyperparameter combinations can be expressed as:
[0100]
[0101] Among them, the domain is a bounding box. Due to the high computational cost of hyperparameter adjustment, d is usually small. Here, the error of the deep learning model validation set is used as the objective function f(r) to obtain the optimal hyperparameter combination r * , you need to pre-set a known hyperparameter combination dataset:
[0102] Θ={(r 1 ,y 1 ),(r 2 ,y 2 ),...,(r i ,y i )},
[0103] Among them, r 1 is a random set of hyperparameter combinations, y i =f(r i ). Then, use the domain Initialize a probabilistic regression model with a small set of samples in After the initialization phase, by optimizing an acquisition function to sequentially select new locations in the domain, using a deep learning model as a cheap proxy for the expensive target f.
[0104] In practice, there is usually a quota for the total available time, so the number of loops g of the objective function f is set to 60. The specific implementation process of the algorithm is shown in Table 1. Here, Gaussian Process assumed μ is the mean function and K is the covariance kernel. The kernel defines the important parameters needed to calculate the predicted distribution, including:
[0105] k(r)=(K(r,r 1 ),...,K(r,r i )),K j,k =K(r j ,r k ),
[0106] Then,
[0107]
[0108] in,
[0109] y=(y 1 ,...,y i ) T
[0110]
[0111]
[0112] in, represents the estimated mean, represents the estimated variance.
[0113] Table 1 Bayesian optimization to select optimal neural network hyperparameters
[0114]
[0115] Collection function Use Expected Improvement. Assuming f' is the minimum value observed so far, Expected Improvement is to maximize the distance between the minimum value found in the future and f', which can be expressed as:
[0116] u(r)=max(0,f'-f(r))
[0117] If f(r) is less than f', you will get a reward, otherwise not. Then, the Expected Improvement acquisition function to solve r is expressed as
[0118] The trained target deep learning model (DL-BO model) is used for online prediction of spectrum availability. Figure 4 As can be seen from the figure, the power value of the current time slot is input into the trained DL-BO model to output the power value of the next time slot. Then, a hard decision is made on the power value to obtain the final spectrum state 0 or 1.
[0119] To verify the effectiveness of the proposed DL-BO prediction method, the following experiments are conducted: First, the ratio of the training set, validation set, and test set in the dataset is 4:1:1. The hyperparameter search space of Bayesian optimization is shown in Table 2. The optimal hyperparameter combination obtained by Bayesian optimization is: the number of hidden layers is 3, the number of Bi-LSTM units is 167, the learning rate is 0.046478, and L 2 The regularization coefficient is 3.2312 -10 The test data is input into the trained optimal deep learning model to obtain the predicted value. The prediction error comparison of DL-BO method and other neural network-based methods is shown in Figure 2. Figure 5 As shown. Figure 5 It can be seen that the prediction error of the proposed DL-BO method is significantly smaller than that of the MLP and LSTM methods in the selected 100 test data. From the prediction results, the trend of the prediction curve of DL-BO is roughly the same as that of the observation curve, while the prediction curve based on the MLP method fluctuates greatly, and the predicted value deviates from the observed value by a large distance. The prediction curve based on LSTM has certain fluctuations, but the overlap rate is low. From the perspective of prediction error, the error between the predicted value and the observed value is bounded by 20. There are 2 time slots where the error of the DL-BO method exceeds 20, while there are more than 15 time slots where the error of the MLP and LSTM methods exceeds 20.
[0120] Figure 6 The MAE and RMSE of the DL-BO method are compared with those of other neural network-based methods. It can be seen that the MAE of the DL-BO method is 11.2995 points lower than that of the MLP-based method and 3.6266 points lower than that of the LSTM-based method. The RMSE of the DL-BO method is 16.362 points lower than that of the MLP-based method and 4.4838 points lower than that of the LSTM-based method.
[0121] Table 2 Hyperparameter search space
[0122]
[0123] Figure 7 The rank correlation comparison between the DL-BO method and other neural network-based methods. It can be seen that under different numbers of time slots, the rank value of DL-BO is higher than the other two methods, 0.4 higher than the MLP-based method and 0.45 higher than the LSTM-based method. It is hereby stated that the rank value and the prediction accuracy are not sufficient and necessary conditions.
[0124] Figure 8The following is a comparison of the prediction RMSE of the DL-BO method and other neural network-based methods under different signal-to-noise ratios. It can be seen that the prediction accuracy of the DL-BO method is better than that of the MLP and LSTM-based methods. Taking SNR = 20dB as an example, the RMSE of the DL-BO method is 3.7dB lower than that of the MLP-based method and 2dB lower than that of the LSTM-based method.
[0125] like Figure 2 As shown, an embodiment of the present invention further provides a spectrum availability prediction system based on a wireless communication network, the system is applied to a wireless communication network including a plurality of cognitive users and a malicious user, the system comprising:
[0126] An acquisition module 10 is used to acquire a spectrum data set for offline training of a deep learning model, wherein the spectrum data set includes a spectrum training set, a spectrum verification set, and a spectrum test set;
[0127] A first setting module 20 is used to set the number of cycles of Bayesian optimization, the search space of hyperparameters and the number of iterations of the deep learning model;
[0128] Initialization module 30, used to initialize the structure, weights, biases, and hyperparameters of the deep learning model;
[0129] A calculation module 40, used to calculate the loss of the spectrum data set between the input and output of the deep learning model;
[0130] An updating module 50, used to update the weight and bias parameters of the deep learning model using gradient descent until the deep learning model training is completed;
[0131] A judgment module 60 is used to judge whether the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state;
[0132] A second setting module 70 is used to reset the number of iterations of the deep learning model when the judgment module determines that the loss curve and the accuracy fitting curve of the deep learning model are not in a convergence state, until the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state;
[0133] The fine-tuning module 80 is used to fine-tune the hyperparameters of the deep learning model by using Bayesian optimization in the hyperparameter search space when the judgment module determines that the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state until the cycle ends and the optimal deep learning model is obtained; and the obtained optimal deep learning model is used to predict the available spectrum online.
[0134] Optionally, the calculation module includes:
[0135] A computation unit for calculating the loss between the input and output of the deep learning model for the spectral dataset according to the following formula:
[0136]
[0137] Where, χ is the loss of spectrum data between the input and output of the deep learning model; t is the time slot; T is the total number of time slots; p t is the observed spectral power; is the predicted spectral power; η is the regularization coefficient; L is the number of layers of the Bi-LSTM neural network; w l is the weight vector of the l-th layer Bi-LSTM neural network.
[0138] Optionally, the update module includes:
[0139] Building blocks for constructing objective functions
[0140] The algorithm for obtaining the optimal U is:
[0141]
[0142] Among them, W(s) and b(s) are the weight and bias of the initialized deep learning model when training for the sth time; α is the learning rate of the deep learning model.
[0143] In this specification, the same or similar parts between the various embodiments can be referred to each other. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description in the method embodiment.
[0144] The present invention has been described in detail above in conjunction with specific implementations and exemplary examples, but these descriptions cannot be understood as limiting the present invention. Those skilled in the art understand that, without departing from the spirit and scope of the present invention, a variety of equivalent substitutions, modifications or improvements may be made to the technical solution of the present invention and its implementation methods, all of which fall within the scope of the present invention. The scope of protection of the present invention shall be subject to the attached claims.
Claims
1. A spectrum availability prediction method based on a wireless communication network, the method being applied to a wireless communication network comprising a plurality of cognitive users and one malicious user, characterized in that: include: Acquire a spectrum dataset for offline training of a deep learning model, wherein the spectrum dataset includes a spectrum training set, a spectrum verification set, and a spectrum test set; The deep learning model is a Bi-LSTM neural network; Set the number of Bayesian optimization cycles, the search space for hyperparameters, and the number of deep learning model iterations; Initialize the structure, weights, biases, and hyperparameters of the deep learning model; Calculate the loss between the input and output of the deep learning model for the spectral dataset as follows: Where, χ is the loss of spectrum data between the input and output of the deep learning model; t is the time slot; T is the total number of time slots; p t is the observed spectral power; is the predicted spectral power; η is the regularization coefficient; L is the number of layers of the Bi-LSTM neural network; w l is the weight vector of the first layer of Bi-LSTM neural network; Use gradient descent to update the weights and bias parameters of the deep learning model until the deep learning model training is completed, including: Constructing the objective function The algorithm to obtain the optimal U is: Where W(s) and b(s) are the weight and bias of the deep learning model at the sth training time, respectively; α is the learning rate of the deep learning model; Determine whether the loss curve and accuracy fitting curve of the deep learning model are in a convergence state; If not, reset the number of deep learning model iterations until the loss curve and accuracy fitting curve of the deep learning model are in a convergence state; If so, Bayesian optimization is used to fine-tune the hyperparameters of the deep learning model in the hyperparameter search space until the cycle ends and the optimal deep learning model is obtained; the obtained optimal deep learning model is used to predict the available spectrum online.
2. The spectrum availability prediction method according to claim 1, characterized in that: The Bayesian optimization adopts a probabilistic regression model of a Gaussian process and an Expected Improvement acquisition function.
3. A spectrum availability prediction system based on a wireless communication network, the system being applied to a wireless communication network comprising a plurality of cognitive users and one malicious user, characterized in that: include: An acquisition module is used to acquire a spectrum data set for offline training of a deep learning model, wherein the spectrum data set includes a spectrum training set, a spectrum verification set, and a spectrum test set; The first setting module is used to set the number of Bayesian optimization cycles, the search space of hyperparameters, and the number of deep learning model iterations; Initialization module, used to initialize the structure, weights, biases, and hyperparameters of the deep learning model; A calculation module for calculating the loss between the input and output of the deep learning model for the spectral dataset; An update module, which is used to update the weight and bias parameters of the deep learning model using gradient descent until the deep learning model training is completed; A judgment module is used to judge whether the loss curve and accuracy fitting curve of the deep learning model are in a convergence state; A second setting module is used to reset the number of iterations of the deep learning model when the judgment module determines that the loss curve and the accuracy fitting curve of the deep learning model are not in a convergence state, until the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state; A fine-tuning module is used to fine-tune the hyperparameters of the deep learning model by using Bayesian optimization in the hyperparameter search space when the judgment module determines that the loss curve and the accuracy fitting curve of the deep learning model are in a convergence state until the cycle ends and the optimal deep learning model is obtained; The obtained optimal deep learning model is used to predict the available spectrum online; The calculation module comprises: A computation unit for calculating the loss between the input and output of the deep learning model for the spectral dataset according to the following formula: Where, χ is the loss of spectrum data between the input and output of the deep learning model; t is the time slot; T is the total number of time slots; p t is the observed spectral power; is the predicted spectral power; η is the regularization coefficient; L is the number of layers of the Bi-LSTM neural network; w l is the weight vector of the Bi-LSTM neural network in layer l; The update module includes: Building blocks for constructing objective functions The algorithm for obtaining the optimal U is: Among them, W(s) and b(s) are the weight and bias of the initialized deep learning model when training for the sth time; α is the learning rate of the deep learning model.
