Dynamic spectrum sharing method based on joint neural network

Abstract

The invention discloses a dynamic spectrum sharing method based on a joint neural network, and belongs to the field of spectrum prediction methods. Aiming at the problems of low utilization rate and lack of spectrum resources and how to obtain the spectrum resources, the invention provides a dynamic spectrum sharing method based on a joint neural network, and the method comprises the steps: improving an LSTM recurrent neural network, and extracting the features of time sequence data; designing a spectrum sensing model based on the combination of a long and short term memory network LSTM and a recurrent neural network RNN; designing a time sequence prediction model based on a BP neural network and a GRNN neural network; and the accuracy of spectrum prediction is improved by using data detected by spectrum sensing, so that the spectrum utilization rate is enhanced, and the dynamic spectrum access is better completed. According to the method, the long-standing dependency problem in the RNN model is solved, and the risk of historical information loss caused by long distance is weakened, so that the accuracy of a trend prediction result of a future spectrum is improved.

Description

Technical Field

[0001] The present invention relates to a dynamic spectrum sharing method, and in particular to a dynamic spectrum sharing method based on a joint neural network. Background Art

[0002] The Internet of Things has become an important trend in the development of world communications and information technology. The Internet of Things technology realizes the connection of various things in different networks. With the gradual popularization of the Internet of Things technology, spectrum scarcity has become one of the most significant challenges facing today's wireless communication networks. The main difficulty facing wireless networks is how to reasonably obtain spectrum resources while meeting the needs of network communication. In order to achieve low latency, it is necessary to reasonably use the spectrum to meet the needs of communication. In wireless communications, spectrum resources are key resources and are indispensable in many fields.

[0003] Since spectrum resources are limited and non-renewable, they have become very precious and a hot spot for competition among countries. Currently, static spectrum allocation is mainly used to manage spectrum resources, and spectrum resources are allocated according to fixed frequency bands. However, since the static spectrum allocation method is very fixed, it is obvious that this approach will result in a lot of idle spectrum not being used, resulting in a lot of spectrum waste. In addition, with the rapid growth of Internet demand in recent years. Many emerging wireless access services cannot obtain appropriate certified frequency bands and can only be crowded in unlicensed frequency bands. Under this trend, the static radio spectrum resource management method is no longer effective.

[0004] In order to ensure that each standard network can still guarantee its peak service capacity requirements for spectrum resources after spectrum refarming, spectrum sharing technology between different standards is the key. With the industry's gradual deepening of research on frequency sharing technology, dynamic spectrum sharing (Dynamic Spectrum Sharing) technology has become an important means for the industry to solve the contradiction between spectrum supply and demand. Dynamic spectrum sharing has gradually become a hot topic in industry research. Dynamic spectrum sharing enables networks of different standards to dynamically apply for and release spectrum resources according to their own business conditions, greatly improving the overall spectrum utilization rate. In order to improve spectrum utilization rate in order to waste spectrum resources, Dr. Joseph Mitola proposed the concept of cognitive radio (CR) in his paper. Cognitive radio technology is a concept based on software radio, which aims to realize opportunistic dynamic spectrum. Dynamic spectrum sharing technology is an important topic in cognitive wireless network research, which aims to improve the utilization efficiency of wireless spectrum resources. Dynamic spectrum sharing will dynamically allocate spectrum in real time according to the need for spectrum. Dynamic spectrum sharing technology adopts a resource pool method to realize spectrum sharing by dynamically monitoring channels, sensing wireless environment, determining parameters, channel allocation and management. The main technologies include spectrum access, frequency band overlapping sharing, dynamic spectrum allocation and public spectrum sharing. In view of the current unreasonable spectrum allocation and low spectrum utilization, this method of dynamic spectrum management is more effective than traditional methods. Spectrum sensing technology is of great significance for improving spectrum utilization and realizing dynamic spectrum access. The main problem is how to efficiently detect spectrum holes. Because spectrum sensing has outstanding advantages such as low construction cost and strong compatibility with existing main systems, it has been recognized by most researchers. Spectrum sensing refers to users obtaining spectrum usage information in wireless networks through various signal detection and processing methods. Spectrum prediction can reduce the perception time. By predicting the future state of the spectrum and then perceiving the spectrum predicted to be idle, it effectively alleviates the problems of energy loss and perception delay in broadband spectrum perception. Spectrum prediction can also reduce the workload of spectrum sharing and improve the overall response speed. Determining the spectrum sharing strategy in advance based on the prediction results can not only effectively improve the response speed of the system, but also improve resource utilization and system throughput. Combine spectrum prediction and spectrum sensing to obtain relevant information on spectrum usage. For example, transmission power and modulation method, and then according to the information of spectrum holes and its own requirements for the channel, select the frequency band with the best communication effect for data transmission to avoid interference between users and improve spectrum utilization efficiency.

[0005] In summary, dynamic spectrum sharing technology is the key to improving spectrum utilization and can effectively deal with the problem of spectrum resource shortage. It is very important to combine it with spectrum prediction and spectrum sensing to understand the spectrum usage status in a certain area and then improve the efficiency of spectrum use. This technology is the key to the next generation network. Therefore, how to improve spectrum utilization through spectrum sensing technology and spectrum sharing technology is the focus of current research.

[0006] At present, the research on spectrum prediction can be roughly divided into two categories. The first category is to predict the quality of the channel, and the second category is to predict from the perspective of channel occupancy. Common spectrum prediction technology methods are regression analysis models and Hidden Markov Model (HMM). However, these prediction technology algorithms are complex and will cause certain obstacles to the success of dynamic spectrum access. Therefore, deep learning is gradually applied to spectrum prediction technology.

[0007] Deep learning algorithms are usually exploratory, and in many cases they provide higher accuracy than traditional probability and statistical algorithms. With the increasing popularity of deep learning in the field of communications in recent years, how to use deep learning algorithms to more efficiently perform spectrum perception and spectrum prediction has become a major research direction. In addition, since any spectrum data is closely related to its adjacent data in time, space, and frequency, it can continue to extend to multiple dimensions of time, frequency, and space in the future. Using these intrinsic relationships to carry out cooperative spectrum work will help improve data computing efficiency and reduce hardware expenses, thereby obtaining high-precision and multi-dimensional spectrum data at a lower cost, thereby effectively improving spectrum utilization efficiency. Summary of the invention

[0008] The purpose of the present invention is to address the problem of low utilization and scarcity of spectrum resources, to promote the development of Internet of Everything communications, and to achieve intelligent utilization of spectrum space in order to obtain spectrum resources. A dynamic spectrum sharing method based on a joint neural network is proposed.

[0009] The above purpose is achieved through the following technical solutions:

[0010] A dynamic spectrum sharing method based on a joint neural network is implemented by the following steps:

[0011] Step 1: Improve the LSTM recurrent neural network to extract the features of time series data;

[0012] Step 2: Design a spectrum sensing model based on the combination of long short-term memory network LSTM and recurrent neural network RNN;

[0013] The state of the target frequency band is detected by accessing users, and then spectrum holes are found as detection signals for dynamic access to provide data information;

[0014] Step 3: Design a time series prediction model based on BP neural network and GRNN neural network;

[0015] After sensing the frequency band, the spectrum data with time correlation is used for prediction. By mining the intrinsic relationship between the spectrum states of historical time slots, the spectrum state of the next time slot is predicted, and the spectrum occupancy and future spectrum trend are predicted.

[0016] Step 4: Use the data detected by spectrum sensing to improve the accuracy of spectrum prediction, thereby enhancing spectrum utilization and enabling better dynamic spectrum access.

[0017] Furthermore, the steps of designing a spectrum sensing model based on a combination of a long short-term memory network LSTM and a recurrent neural network RNN ​​described in step 2 are specifically:

[0018] Step 21: Introduce the connection relationship between the hidden layers of the neural network, receive historical information and current time information, and use historical information to predict future spectrum trends;

[0019] Step 22: Use the improved RNN to receive signals from other neurons and its own feedback signals;

[0020] Step 2 and 3: Introduce the threshold mechanism, i.e., memory unit, and design the structure of the long short-term memory network;

[0021] The structure of the long short-term memory network LSTM includes memory cells, input gate i, forget gate f and output gate o, wherein the input gate i controls the amount of information entering the storage unit, the forget gate f determines to remember the state information of the previous units, and the output gate o controls the amount of output information by calculating the storage size of the memory unit. The memory cell enables the LSTM to freely select the content to be remembered in each time step;

[0022] The training of the long short-term memory network LSTM includes the following steps:

[0023] First, the forget gate is controlled by the sigmoid activation function, which generates a 0-1 value based on the output of the previous moment and the current input to decide whether to let the information learned at the previous moment pass or partially pass:

[0024] f t =σ(W f ·[h t-1 , x t ]+b f )

[0025] In the formula, xt ∈R d ,h t-1 ∈R m ,W f ∈R m×(m+d) , b f ∈R m ;

[0026] After that, the input gate uses the sigmoid activation function to determine the value that needs to be updated, and the tanh activation function is used to generate new candidate values:

[0027] i t =σ(W i ·[h t-1 , x t ]+b i )

[0028] C t =tanh(W C ·[h t-1 , x t ]+b C )

[0029] In the formula, x t ∈R d ,h t-1 ∈R m ,W i ∈R m×m+d) , b i , b C ∈R m ,C t ∈R m ;

[0030] Afterwards, during the cell update process, the old cell state is updated by multiplying the old cell state by the information to be forgotten, and then obtaining the candidate value according to a certain ratio:

[0031] C t =f t *C t-1 +i t *C t ;

[0032] After that, the output gate is used to determine the size of the model output value; the initial output will be obtained through the sigmoid activation function; the tanh activation function is used to reduce the cell update information to a value between -1 and 1, and then the reduced value is multiplied by each pair of outputs obtained through the sigmoid activation function to obtain the model output:

[0033] o t =σ(W0[h t-1 , x t]+b0)

[0034] h t =o t *tanh(C t )

[0035] Where W0∈R m×(m+d) ,h t-1 ∈R m ,x t ∈R d , b0∈R m ,h t ∈R m .

[0036] Furthermore, the mathematical model of the memory cell network is as follows:

[0037] i t =σ(W i x t +U i h t-1 +b i )

[0038] f t =σ(W f x t +U f h t-1 +b f )

[0039] o t =σ(W o x t +U o h t-1 +b o )

[0040]

[0041] Among them, x t is the input vector at time step t, and the hidden layer state vector is h t , c t is a memory cell, σ is a sigmoid function, tanh is a hyperbolic tangent activation function, W and U are weight matrices of the corresponding modules, and b is a bias vector of the corresponding template.

[0042] The beneficial effects of the present invention are:

[0043] Since the spectrum signal is also equivalent to a long time series, it is regarded as a prediction of the series. The present invention will use a neural network to predict the spectrum occupancy and future spectrum trends, study the effect of time series prediction under different networks, change the network conditions to improve the accuracy of spectrum prediction, and achieve the effect of improving spectrum utilization.

[0044] Therefore, the spectrum sensing and spectrum prediction technology of cognitive radio is proposed for spectrum sharing. By comparing the traditional energy detection method and the spectrum sensing algorithm based on deep learning in low signal-to-noise ratio scenarios, it is found that the performance of the deep learning network is better. For the spectrum equivalent to a long time series, the deep learning method can also be used to improve the prediction effect and enhance the prediction performance.

[0045] The present invention uses deep learning and neural networks in spectrum technology, combines perception and prediction with neural networks, and achieves better dynamic spectrum access than traditional methods. Specifically:

[0046] (1) In the feature extraction task of time series data, RNN depends on data in the past short period of time. If the time interval is long, the performance of RNN may be limited and the accuracy of prediction will also be reduced. The present invention proposes an improved LSTM recurrent neural network, which greatly solves the long-standing dependency problem in the RNN model and reduces the risk of historical information loss due to long distances, thereby greatly improving the accuracy of the results.

[0047] (2) For the spectrum prediction technology based on deep learning, a time series prediction based on BP neural network and GRNN neural network was proposed, and the positive impact on the spectrum perception prediction effect was obtained by comparing the errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of the method of the present invention;

[0049] Figure 2 This is a diagram of the RNN unit structure involved in the present invention;

[0050] Figure 3 It is a simple cyclic network structure diagram involved in the present invention;

[0051] Figure 4 is a network structure diagram involved in the present invention;

[0052] Figure 5 It is a structural diagram of the LSTM memory unit structure involved in the present invention;

[0053] Figure 6 It is the ROC curve involved in the simulation verification part of the present invention;

[0054] Figure 7 It is the detection probability-SNR variation curve involved in the simulation verification part of the present invention;

[0055] Figure 8 It is a BP prediction result diagram involved in the simulation verification part of the present invention;

[0056] Fig. 9 It is a GRNN prediction result diagram involved in the simulation verification part of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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. Specific implementation method one:

[0059] A dynamic spectrum sharing method based on a joint neural network in this embodiment is as follows: Figure 1-9 As shown, the method is implemented by the following steps:

[0060] Step 1: Improve the LSTM recurrent neural network to extract the features of time series data; solve the long-standing dependency problem in the RNN model and reduce the risk of historical information loss due to long distances, thereby greatly improving the accuracy of the results;

[0061] Step 2: Design a spectrum sensing model based on the combination of long short-term memory network LSTM and recurrent neural network RNN;

[0062] The state of the target frequency band is detected by accessing users, and then spectrum holes are found as detection signals for dynamic access to provide data information;

[0063] Step 3: Since the model structure of BP neural network is relatively simple, including: input layer, hidden layer and output layer, although BP neural network has a simple structure, it has strong nonlinear mapping ability, and LSTM neural network has good performance in processing time series and sequence problems, so a time series prediction model based on BP neural network and GRNN neural network is designed;

[0064] After sensing the frequency band, the spectrum data with time correlation is used for prediction. By mining the intrinsic relationship between the spectrum states of historical time slots, the spectrum state of the next time slot is predicted, and the spectrum occupancy and future spectrum trend are predicted.

[0065] Step 4: Reduce the time of spectrum sensing through spectrum prediction, and use the data detected by spectrum sensing to improve the accuracy of spectrum prediction, thereby enhancing spectrum utilization and making dynamic spectrum access better. Specific implementation method 2:

[0067] Different from the first specific implementation, in the present implementation, a dynamic spectrum sharing method based on a joint neural network, the step of designing a spectrum sensing model based on a combination of a long short-term memory network LSTM and a recurrent neural network RNN ​​described in step 2 is specifically:

[0068] Step 21: Introduce the connection relationship between the hidden layers of the neural network, receive historical information and current time information, this feature is closely related to the method of processing time series tasks, that is, using historical information to predict future spectrum trends;

[0069] Step 2: Use the improved RNN to receive signals from other neurons and its own feedback signals; the structure of RNN is as follows Figure 2 and Figure 3 As shown;

[0070] Step 2 and 3: Improved RNN is very advantageous for processing time series problems. In theory, it can learn sequence features of any length. However, in practice, it is found that RNN is prone to extreme nonlinear behavior after the error gradient is back-propagated for multiple time steps, which limits the use of RNN. In order to solve the problem of gradient disappearance in RNN, a threshold mechanism, namely memory unit, is introduced to design the structure of long short-term memory network;

[0071] Among them, Figure 4 and Figure 5 As shown, the structure of the long short-term memory network LSTM includes memory cells, input gate i, forget gate f and output gate o, wherein the input gate i controls the amount of information entering the storage unit, the forget gate f determines to remember the state information of the previous units, and the output gate o controls the amount of output information by calculating the storage size of the memory unit. The most critical memory cell enables LSTM to freely choose the content to be remembered in each time step.

[0072] The training of the long short-term memory network LSTM includes the following steps:

[0073] First, the forget gate is controlled by the sigmoid activation function, which generates a 0-1 value based on the output of the previous moment and the current input to decide whether to let the information learned at the previous moment pass or partially pass:

[0074] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0075] In the formula, x t ∈R d ,h t-1 ∈R m ,Wf ,∈R m×(m+d) ,b f ∈R m ;

[0076] After that, the input gate uses the sigmoid activation function to determine the value that needs to be updated, and the tanh activation function is used to generate new candidate values:

[0077] i t =σ(W i ·[h t-1 , x t ]+b i )

[0078] C t =tanh(W C ·[h t-1 , x t ]+b C )

[0079] In the formula, x t ∈R d ,h t-1 ∈R m ,W i ∈R m×(m+d) , b i , b C ∈R m ,C t ∈R m ;

[0080] Afterwards, during the cell update process, the old cell state is updated by multiplying the old cell state by the information to be forgotten, and then obtaining the candidate value according to a certain ratio:

[0081] C t =f t *C t-1 +i t *C t ;

[0082] After that, the output gate is used to determine the size of the model output value; the initial output will be obtained through the sigmoid activation function; the tanh activation function is used to reduce the cell update information to a value between -1 and 1, and then the reduced value is multiplied by each pair of outputs obtained through the sigmoid activation function to obtain the model output:

[0083] o t =σ(W0[h t-1 , x t ]+b0)

[0084] h t =o t*tanh(C t )

[0085] Where W0∈R m×(m+d) ,h t-1 ∈R m ,x t ∈R d ,b0∈R m ,h t ∈R m . Specific implementation method three:

[0087] Different from the second specific implementation mode, in the dynamic spectrum sharing method based on the joint neural network of this implementation mode, the mathematical model of the memory cell network is as follows:

[0088] i t =σ(W i x t +U i h t-1 +b i )

[0089] f t =σ(W f x t +U f h t-1 +b f )

[0090] o t =σ(W o x t +U o h t-1 +b o )

[0091]

[0092] Among them, x t is the input vector at time step t, and the hidden layer state vector is h t , c t is a memory cell, σ is a sigmoid function, tanh is a hyperbolic tangent activation function, W and U are weight matrices of the corresponding modules, and b is a bias vector of the corresponding template.

[0093] Scheme and verification of the method of the present invention:

[0094] For deep learning spectrum sensing algorithms, the focus is on studying the performance indicators of spectrum sensing, among which the false alarm probability P is the main focus. f and the detection probability P d The receiver operating characteristic (ROC) and P d-SNR curve for spectrum sensing performance analysis, such as Figure 6 and Figure 7 As shown in the figure, compared with the previous traditional spectrum sensing algorithms, it can be seen that the detection effect obtained by the method using the LSTM-RNN network is the best, followed by the method using the convolutional neural network, and finally the traditional energy detection-based algorithm. In order to more comprehensively evaluate the performance of various spectrum sensing methods under different signal-to-noise ratio conditions, we can first fix different false alarm probabilities and gradually change the signal-to-noise ratio. As the signal-to-noise ratio changes, we can observe the indicators of each detection probability. Figure 7 As shown in the figure, when the fixed false alarm probability is higher, the detection accuracy of all schemes is also enhanced accordingly. By comparing the algorithms, it can be seen that the spectrum sensing algorithm based on the LSTM-RNN network has obvious advantages in detection performance.

[0095] The results show that compared with the traditional spectrum sensing based on energy detection algorithm, the algorithm based on deep learning achieves better detection effect. In the results of time series prediction based on BP neural network and GRNN neural network, the simulation results are as follows Figure 8 and Fig. 9 shown.

[0096] After comparing the two methods, the results show that in practical applications, the effect of using GRNN neural network is better than that of BP neural network. The error of GRNN network is usually smaller than that of BP neural network, and it has better prediction effect on spectrum perception.

[0097] The embodiments of the present invention disclose preferred embodiments, but are not limited thereto. A person skilled in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not deviate from the spirit of the present invention, they are all within the protection scope of the present invention.

Claims

1. A dynamic spectrum sharing method based on a joint neural network, characterized in that: The method is implemented by the following steps: Step 1: Improve the LSTM recurrent neural network to extract the features of time series data; Step 2: Design a spectrum sensing model based on the combination of long short-term memory network LSTM and recurrent neural network RNN; The state of the target frequency band is detected by accessing users, and then spectrum holes are found as detection signals for dynamic access to provide data information; Step 3: Design a time series prediction model based on BP neural network and GRNN neural network; After sensing the frequency band, the spectrum data with time correlation is used for prediction. By mining the intrinsic relationship between the spectrum states of historical time slots, the spectrum state of the next time slot is predicted, and the spectrum occupancy and future spectrum trend are predicted. Step 4: Use the data detected by spectrum sensing to improve the accuracy of spectrum prediction, thereby enhancing spectrum utilization and enabling better dynamic spectrum access.

2. The method for dynamic spectrum sharing based on a joint neural network according to claim 1, characterized in that: The steps described in step 2 for designing a spectrum sensing model based on a combination of a long short-term memory network LSTM and a recurrent neural network RNN ​​are as follows: Step 21: Introduce the connection relationship between the hidden layers of the neural network, receive historical information and current time information, and use historical information to predict future spectrum trends; Step 22: Use the improved RNN to receive signals from other neurons and its own feedback signals; Step 2 and 3: Introduce the threshold mechanism, i.e., memory unit, and design the structure of the long short-term memory network; The structure of the long short-term memory network LSTM includes memory cells, input gate i, forget gate f and output gate o, wherein the input gate i controls the amount of information entering the storage unit, the forget gate f determines to remember the state information of the previous units, and the output gate o controls the amount of output information by calculating the storage size of the memory unit. The memory cell enables the LSTM to freely select the content to be remembered in each time step; The training of the long short-term memory network LSTM includes the following steps: First, the forget gate is controlled by the sigmoid activation function, which generates a 0-1 value based on the output of the previous moment and the current input to decide whether to let the information learned at the previous moment pass or partially pass: f t =σ(W f ·[h t-1 ,x t ]+b f ) In the formula, x t ∈R d ,h t-1 ∈R m , W f ∈R m×(m+d) , b f ∈R m ; After that, the input gate uses the sigmoid activation function to determine the value that needs to be updated, and the tanh activation function is used to generate new candidate values: i t =σ(W i ·[h t-1 ,x t ]+b i ) C t =tanh(W C ·[h t-1 ,x t ]+b C ) In the formula, x t ∈R d ,h t-1 ∈R m , W i ∈R m×(m+d) , b i , b C ∈R m , C t ∈R m ; Afterwards, during the cell update process, the old cell state is updated by multiplying the old cell state by the information to be forgotten, and then obtaining the candidate value according to a certain ratio: C t =f t *C t-1 +i t *C t ; After that, the output gate is used to determine the size of the model output value; the initial output will be obtained through the sigmoid activation function; the tanh activation function is used to reduce the cell update information to a value between -1 and 1, and then the reduced value is multiplied by each pair of outputs obtained through the sigmoid activation function to obtain the model output: about t =σ(W0[h t-1 ,x t ]+b0) h t =o t *tanh(C t ) Where W0∈R m×(m+d) ,h t-1 ∈R m , x t ∈R d , b0∈R m ,h t ∈R m .

3. The method for dynamic spectrum sharing based on a joint neural network according to claim 2, characterized in that: The mathematical model of the memory cell network is as follows: i t =σ(W i x t +U i h t-1 +b i ) f t =σ(W f x t +U f h t-1 +b f ) the t =σ(W o x t +U o h t-1 +b o ) Among them, x t is the input vector at time step t, and the hidden layer state vector is h t , c t is a memory cell, σ is a sigmoid function, tanh is a hyperbolic tangent activation function, W and U are weight matrices of the corresponding modules, and b is a bias vector of the corresponding template.

Images