Wireless communication network optimization method and device based on deep learning

Abstract

The invention relates to a wireless communication network optimization method and device based on deep learning, and the method comprises the steps: obtaining complete channel information, inputting the channel information into a target deep learning model, obtaining a target power vector through the target deep learning model, and transmitting the target power vector to a wireless communication network. The complete channel information comprises channel information between the network base station and a plurality of users, the target power vector comprises transmitting power of the network base station acting on at least one target user, and the target user is a scheduled user in the plurality of users; determining a target channel vector corresponding to the target power vector, wherein the target channel vector comprises a channel vector between the network base station and at least one target user; and determining an optimal beam vector of the at least one target user according to the target channel vector. The joint optimization of the user scheduling and the beam forming design is realized, and the scheduling number constraint and the QoS constraint are considered in the optimization process, so that the finally obtained optimization scheme is more accurate and conforms to the reality.

Description

Technical Field

[0001] The present application relates to the field of wireless communications, and in particular to a method and device for optimizing a wireless communication network based on deep learning. Background Art

[0002] Since user scheduling and beamforming design problems are always coupled to each other, traditional research usually studies them separately, or divides the joint optimization problem into two sub-problems, and then designs an iterative optimization scheme based on convex optimization theory to solve them. However, the above optimization schemes are not only highly complex, but also have poor generalization capabilities. In order to break the limitations of traditional optimization schemes, many literatures have proposed joint optimization methods for beamforming design and user scheduling based on deep learning (DL) (wireless communication network optimization methods).

[0003] However, the above-mentioned wireless communication network optimization method based on deep learning does not consider comprehensive factors during optimization, resulting in the final optimization solution being unable to meet actual usage requirements. Summary of the invention

[0004] The present invention provides a wireless communication network optimization method and device based on deep learning to solve the problem that the wireless communication network optimization method based on deep learning cannot meet the actual use requirements.

[0005] In a first aspect, the present invention provides a wireless communication network optimization method based on deep learning, which is applied to a multi-user downlink MISO communication system, wherein the multi-user downlink MISO communication system includes a network base station and a plurality of users, wherein the network base station provides network services for the plurality of users, and the wireless communication network optimization method includes:

[0006] Acquire complete channel information, input the channel information into a target deep learning model, and obtain a target power vector through the target deep learning model, wherein the complete channel information includes channel information between the network base station and the plurality of users, and the target power vector includes a transmit power of the network base station acting on at least one target user, and the target user is a scheduled user among the plurality of users;

[0007] Determine a target channel vector corresponding to the target power vector, the target channel vector comprising a channel vector between the network base station and the at least one target user;

[0008] Determining an optimal beam vector for the at least one target user according to the target channel vector;

[0009] Among them, the target deep learning model adopts unsupervised training and the loss function L loss for:

[0010]

[0011] Among them, R k represents the network rate of the kth user, μ and λ k represents the non-negative Lagrange multiplier, ReLU represents the activation function, L represents the minimum threshold of the number of user scheduling, and f θ represents a continuous smooth function, K represents the total number of users, and p k It represents the transmission power of the network base station to the kth user, SINR k represents the downlink signal-to-noise ratio of the kth user, η k Indicates SINR k The minimum threshold value.

[0012] In a second aspect, the present invention provides a wireless communication network optimization device based on deep learning, which is applied to a multi-user downlink MISO communication system, wherein the multi-user downlink MISO communication system includes a network base station and a plurality of users, wherein the network base station provides network services for the plurality of users, and the wireless communication network optimization device includes:

[0013] a power determination module, configured to obtain complete channel information, input the channel information into a target deep learning model, and obtain a target power vector through the target deep learning model, wherein the complete channel information includes channel information between the network base station and the plurality of users, and the target power vector includes a transmit power of the network base station acting on at least one target user, and the target user is a scheduled user among the plurality of users;

[0014] A channel determination module, configured to determine a target channel vector corresponding to the target power vector, wherein the target channel vector includes a channel vector between the network base station and the at least one target user;

[0015] A beam determination module, configured to determine an optimal beam vector of the at least one target user according to the target channel vector;

[0016] Among them, the target deep learning model adopts unsupervised training and the loss function L loss for:

[0017]

[0018] Among them, R k represents the network rate of the kth user, μ and λ k represents the non-negative Lagrange multiplier, ReLU represents the activation function, L represents the minimum threshold of the number of user scheduling, and f θ represents a continuous smooth function, K represents the total number of users, and p krepresents the transmission power of the network base station acting on the kth user, SINR k represents the downlink signal-to-noise ratio of the kth user, η k Indicates SINR k The minimum threshold value.

[0019] Compared with the related art, the deep learning wireless communication network optimization method provided by the present invention determines the user scheduling scheme through the target deep learning model. The target deep learning model is trained through a given loss function, which can realize the joint optimization of user scheduling and beamforming design under the scheduling number constraint and QoS constraint. User scheduling optimization refers to the reasonable determination of the users who need to be scheduled, and beamforming design optimization refers to the reasonable determination of the channel vectors corresponding to each scheduled user. Since the factors considered in the above wireless communication network optimization method are relatively comprehensive, not only the joint optimization of user scheduling and beamforming design is realized, but also the scheduling number constraint and QoS constraint are considered in the optimization process, so that the final optimization scheme is more accurate and practical, which can meet the actual use needs, and solves the problem that the current wireless communication network optimization method based on deep learning cannot meet the actual use needs.

[0020] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flowchart of a wireless communication network optimization method based on deep learning provided in an embodiment of the present invention;

[0022] Figure 2 A schematic diagram of the structure of a target deep learning model in some embodiments of the present invention;

[0023] Figure 3 is a schematic diagram of constructing a wireless channel graph in some embodiments of the present invention;

[0024] Figure 4 This is a schematic diagram comparing the training convergence speed of different models;

[0025] Figure 5 is the average sum rate diagram under different signal-to-noise ratios in a multi-user MISO system;

[0026] Figure 6 Figure 2 is the average sum rate diagram for different numbers of users in a multi-user MISO system. DETAILED DESCRIPTION

[0027] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0028] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the", "these" and the like in this application do not represent quantitative restrictions, and they can be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusions; for example, a process, method and system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether directly or indirectly. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships. For example, "A and / or B" can mean: A exists alone, A and B exist at the same time, and B exists alone. Usually, the character " / " indicates that the objects associated with each other are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0029] The invention provides a wireless communication network optimization method in a multi-user downlink MISO communication system. Figure 1 Schematic diagram of a multi-user downlink MISO communication system used in the present invention. Figure 1 In a multi-user downlink MISO communication system, the transmitter is a base station (BS) with N antennas, and the receiver is a user with K single antennas. is the user set. There is a channel vector between the base station and the scheduled user. represents the beamforming vector (beam vector) between the base station and the user. Based on the above expression, the transmitted signal at the BS is expressed as:

[0030]

[0031] Among them, s k is the information-carrying symbol of user k, with normalized power. Therefore, the received signal of user k is:

[0032]

[0033] Among them, n k is a product with a mean of 0 and a variance of Based on the above formula, the downlink signal-to-noise ratio (SINR) of user k can be obtained as follows:

[0034]

[0035] According to the above description, the achievable network rate of user k is R k =log(1+SINR k ). In order to clearly indicate whether a user is scheduled, the present invention introduces As the status indicator of user k, if user k is scheduled, then If user k is not scheduled, then In order to ensure QoS and the total number of user scheduling, the present invention sets a minimum value γ of the user rate k , and the number of scheduled users cannot be less than L, with the goal of maximizing the system sum rate, the problem of joint optimization of user scheduling and beamforming is expressed as:

[0036]

[0037] In the above formula, P represents the total transmission power of the network base station, L represents the minimum constraint on the number of scheduled users, and γ k represents the minimum network rate requirement of user k. Obviously, constraint C1 is a power constraint, and constraint C2 is a 0-norm sparse beamforming design problem, that is, the minimum number of scheduled users constraint. Indicates that user k is not scheduled, and vice versa. Constraint C3 is the user minimum transmission rate constraint. By solving this problem, the beamforming vector and the maximum sum rate can be determined under the constraint of the minimum number of scheduled users. However, there is a coupling between the non-convex zero-norm constraint and the non-convex variable in the objective function. Still challenging.

[0038] In order to solve the non-convex zero norm problem of C2, the present invention adopts a continuous smooth function f θ (x) The discontinuous zero norm in the approximation constraint is defined as f θ (x) = 1-e -x / θ , where θ>0 is a parameter that controls the smoothness of the approximation. The larger θ is, the smoother the function is, but the worse the approximation is, and vice versa. For a smooth 0 norm, It can be approximated as:

[0039]

[0040] in Note the smooth function for It is concave in w k is not concave, and C3 is also non-convex, so, Still non-convex. Then the present invention uses a deep learning model to solve the problem equivalently For joint optimization of user scheduling and beamforming design. Further, in some embodiments, in order to solve While making full use of the user's historical CSI (channel information), reducing the computational complexity of the algorithm and improving the generalization ability of the algorithm, the deep learning framework is constructed by combining GCN and cross-attention mechanism, which will be explained in detail in the following text.

[0041] In order to avoid the real and imaginary part segmentation operation of the complex channel vector in the deep learning framework and reduce the complexity of the algorithm, the present invention first transforms the beamforming vector w k Divide the range and Phase Two parts, and derived For the convenience of expression, the present invention defines Then, in p k In the fixed case, constraint C2 has been satisfied, so we can solve It can be temporarily ignored. According to the minimum mean square error (MMSE), the downlink SINR of user k is k The optimal beamforming vector that is maximized is:

[0042]

[0043] Where v ≥ 0 and λ ≥ 0 are the Lagrange multipliers associated with the power constraint and QoS constraint, respectively. N is the N-dimensional identity matrix. There is an analytical solution, so the final optimization problem only contains one variable p k , can be rewritten as:

[0044]

[0045] question The solution can be approximated using the continuous convex approximation (SCA) and convex difference (DC) algorithms.

[0046] However, the solution based on traditional approximate iterative optimization algorithms (such as USBD algorithm) has the disadvantages of high complexity and poor generalization ability. In order to overcome these limitations, the present invention provides a wireless communication network optimization method based on deep learning to solve the problem. Used to achieve joint optimization of user scheduling and beamforming design.

[0047] The wireless communication network optimization method based on deep learning provided by the present invention is explained as follows.

[0048] In an embodiment of the present invention, a wireless communication network optimization method based on deep learning is provided, which is applied to a multi-user downlink MISO communication system. The multi-user downlink MISO communication system includes a network base station and several users. The network base station provides network services for the several users.

[0049] Figure 1 is a flow chart of a wireless communication network optimization method based on deep learning provided in an embodiment of the present invention. Figure 1 The wireless communication network optimization method includes step S110, step S120 and step S130.

[0050] Step S110, obtain complete channel information, input the channel information into the target deep learning model, and obtain the target power vector through the target deep learning model. The complete channel information includes the channel information between the network base station and several users. The target power vector includes the transmission power of the network base station on at least one target user. The target user is a scheduled user among several users.

[0051] Specifically, the complete channel information includes the existing channel information (historical channel information) between the network base station and all users in the multi-user downlink MISO communication system. For example, the channel information between the network base station and the user can be expressed as H, and the covariance matrix of H can be defined as in: h k Represents the channel vector between the network base station and the kth user. Using H as the input of the network avoids the step of splitting the complex channel vector into real and imaginary parts, thereby reducing the complexity of the algorithm.

[0052] The target deep learning model has the ability to determine the user scheduling scheme that maximizes the system and rate through training, that is, to determine the users that need to be scheduled at the moment. The scheduled users are defined as target users, and the network base station provides network services to them. The target deep learning model outputs the transmission power of the network base station acting on at least one target user. For non-target users, the corresponding transmission power is 0. The ability of the target deep learning model is mainly achieved by using the corresponding loss function.

[0053] Among them, the target deep learning model adopts unsupervised training and the loss function L loss for:

[0054]

[0055] Among them, R k represents the network rate of the kth user, μ and λ k represents the non-negative Lagrange multiplier, ReLU represents the activation function, L represents the minimum threshold of the number of user scheduling, and f θrepresents a continuous smooth function, K represents the total number of users, and p k represents the transmission power of the network base station acting on the kth user, SINR k represents the downlink signal-to-noise ratio of the kth user, η k Indicates SINR k The minimum threshold value.

[0056] The update rules of the two non-negative Lagrange multipliers are:

[0057]

[0058] Among them, ε v and ε λ Represent μ and λ respectively k The update step size, μ τ and μ τ+1 represent μ of the τth generation and the τ+1th generation respectively, and denote the λ of the τth generation and the τ+1th generation respectively k .

[0059] In the above loss function: the first is to sum the negative values ​​of the user network rate. When the sum of the network rates of all scheduled users is the largest, the sum of the sum is the smallest; the second is the scheduling quantity constraint penalty item, L represents the minimum threshold of the user scheduling quantity. When the user scheduling quantity is lower than this value, it is penalized and the loss increases; the third is the signal-to-noise ratio constraint (QoS constraint) penalty item. When the user's downlink signal-to-noise ratio is lower than the minimum threshold, it is penalized and the loss increases.

[0060] Through training with the above loss function, the target deep learning model has the ability to determine the user scheduling plan with the goal of maximizing the system and rate, and the user scheduling plan can meet the user scheduling number constraints and user QoS constraints. Finally, the network base station outputs the transmission power of each scheduled user (used to determine the beam), realizing the joint optimization of user scheduling and beamforming design.

[0061] Step S120: determining a target channel vector corresponding to the target power vector, where the target channel vector includes a channel vector between a network base station and at least one target user.

[0062] In this step, the target user (the user whose corresponding transmission power is greater than 0) can be determined by the target power vector, and then the channel vector between the network base station and the target user can be determined (the channel vectors between the network base station and each user are known).

[0063] Step S130: determining an optimal beam vector of at least one target user according to the target channel vector.

[0064] Specifically, the optimal beam vector of at least one target user is determined by a target formula, which is:

[0065]

[0066] in, represents the optimal beam vector for the kth user, express The phase of express The amplitude, represents the kth power value in the target power vector (the transmission power of the network base station acting on the kth user), v and λ represent the Lagrange multipliers associated with the power constraint and the quality of service constraint (which are hyperparameters in the target deep learning model and are determined according to the model training results), and I N represents the N-dimensional identity matrix, K represents the total number of users, and h k represents the channel vector between the network base station and the kth user, h k The variance of the Gaussian white noise,

[0067] Through the above target formula, the optimal beam vector (phase and amplitude) of each target user can be determined, and the corresponding beam vector of non-target users is 0. The optimal beam vector of the kth user refers to the optimal beam vector of the network signal sent by the network base station to the kth user. Determining the optimal beam vector of each target user is the key to beamforming design.

[0068] In summary, the present invention provides a deep learning wireless communication network optimization method, which determines the user scheduling scheme through a target deep learning model. The target deep learning model is trained through a given loss function, which can realize the joint optimization of user scheduling and beamforming design under the scheduling number constraint and QoS constraint. User scheduling optimization refers to the reasonable determination of the users who need to be scheduled, and beamforming design optimization refers to the reasonable determination of the channel vectors corresponding to each scheduled user. Since the factors considered in the above wireless communication network optimization method are relatively comprehensive, not only the joint optimization of user scheduling and beamforming design is realized, but also the scheduling number constraint and QoS constraint are considered in the optimization process, so that the final optimization scheme is more accurate and practical, which can meet the actual use needs, and solves the problem that the current wireless communication network optimization method based on deep learning cannot meet the actual use needs.

[0069] The model structure of the target deep learning model adopted by the present invention is described as follows.

[0070] Figure 2A schematic diagram of the structure of the target deep learning model in some embodiments of the present invention. Figure 2 In some embodiments, the target deep learning model includes a graph neural network module GCN and a cross-attention module CA (the deep learning model in this embodiment is defined as GCN-CA), the input of the graph neural network module is the input of the target deep learning model, the input of the cross-attention module is the input of the target deep learning model and the output of the graph neural network module, and the output of the cross-attention module is the output of the target deep learning model.

[0071] Figure 3 Schematic diagram of constructing a wireless channel graph in some embodiments of the present invention. Figure 3 ,In the target deep learning model, the channel information is represented by a wireless channel graph; in the wireless channel graph, nodes represent users and edges represent interference signals between users.

[0072] Among them, the feature of the i-th node is defined as The edge between the i-th node and the j-th node is defined as

[0073] The graph neural network module includes multiple graph convolutional layers connected in sequence. The feature processing flow of the lth graph convolutional layer includes:

[0074]

[0075] in, represents the embedding representation vector of the jth node output by the l-1th graph convolutional layer, represents the embedding representation vector of the i-th node output by the l-1-th graph convolutional layer, represents the embedding representation vector of the i-th node output by the l-th graph convolutional layer, It represents the aggregation result of neighborhood features of the ith node in the lth graph convolutional layer, the jth node is the neighborhood node of the ith node, MLP1 and MLP2 are both multi-layer perceptrons, and PE represents the permutation equivalent function.

[0076] Specifically, in the GCN module (graph neural network module), the downlink MISO communication system of K users is modeled as a complete graph. For ease of description, the wireless channel graph is represented as The i-th user is regarded as the i-th node in the graph. The nodes and edges in the graph have different feature information. The feature of the i-th node is defined as The edge between the i-th node and the j-th node is defined as After completing the graph representation of the wireless communication network, this embodiment focuses on designing a GCN module to optimize the power vector.

[0077] Each graph convolution layer in the GCN module includes two parts: node neighborhood feature aggregation and node embedding representation vector update. The input of the GCN module is the CSI (channel information) of the communication system. Since the communication system graph is usually fully connected, the nodes in the graph obtain their own embedded representation vectors in the first graph convolution layer and propagate them to the neighboring nodes of the next graph convolution layer. The embedding representation vector of the i-th node in the l-th graph convolution layer is Neighborhood feature aggregation function The update rules are as shown above. Among them, the permutation equivalent function can use the sum function or the mean function to aggregate and pool the neighborhood messages to obtain the aggregated features. Preferably, the graph neural network module uses one graph convolution layer, that is, the graph neural network module is a single-layer structure.

[0078] In addition to the above-described parts, the GCN module also includes a batch normalization layer (BN), a flat layer (FL), and an activation layer (AC). The use of the BN layer can avoid overfitting and accelerate convergence. The FL layer can also reshape the information of the feature vector. The AC layer can introduce nonlinearity, enhance the expression ability, and avoid gradient disappearance. The final node embedding representation vector obtained by the aggregation update of the L layer is the output of the GCN module, that is, the power vector p GCN .

[0079] The feature processing flow of the cross-attention module includes: performing three parallel linearization processes on the input of the target deep learning model and the graph neural network module to obtain the first feature, the second feature and the third feature respectively; performing a convolution operation on the first feature and the second feature to obtain an attention matrix, and then convolving the attention matrix with the third feature after activation to obtain a first combination result; performing a linear combination of the first combination result and the third feature to obtain a second combination result; processing the second combination result through a feedforward network to obtain a candidate power vector, and determining the target power vector based on the candidate power vector.

[0080] Specifically, in the cross attention module, the channel information H and the power vector p are first GCN After linearization, the first feature Q1, the second feature Q2, and the third feature Q3 are obtained. Then, Q1 and Q2 are convolved to generate an attention matrix. Subsequently, the attention matrix is ​​activated, and the activation result is convolved with Q3. Finally, the convolution result (the first combination result) is linearly combined with Q3 (the second combination result). Through a layer of feedback in the feedforward network, the final output result p is obtained. CA The power vector obtained by the cross-attention mechanism is sparse, and the expression of the entire cross-attention mechanism module is:

[0081]

[0082] Where d is the scale factor, p CA is the output of the cross attention module. A projection activation layer is designed at the end of each network layer to ensure the output power p CA The total power constraint is satisfied. The projection function is as follows:

[0083]

[0084] Among them, p k represents the candidate power vector (the output p of the cross-attention module CA ), represents the kth power value in the target power vector, which is used to characterize the power of the network base station on the kth user. ReLU(x)=max(x,0) represents the integer linear unit activation function, and P represents the total transmit power of the network base station.

[0085] A training example of a target deep learning model using the above framework is provided below.

[0086] Step 1: First, a real data set is established based on the user's historical complex channel vector. The total training samples are set to D = 5000, the batch size is 100, and the size of each sample is D i =30, number of epochs N e =50, the number of test samples is one tenth of the training samples.

[0087] Step 2: Collect training data samples. For each training sample D i Generate a user channel graph representation and initialize the feature vector of each node in the graph as the input of the GCN module. After the aggregation and combination of neighborhood node features and the update of node representation vector, the output of the GCN module is obtained, that is, the optimized transmission power.

[0088] Step 3: The node representation vector and channel vector output by the GCN module are used as the input of the cross attention mechanism module, and three parameter matrices Q1, Q2 and Q3 are extracted according to these two feature vectors. After convolution, linearization, activation and mapping operations are performed on these three matrices, a power vector with sparsity is finally output.

[0089] Step 4: Substitute the channel vector corresponding to the sparse power vector output by the GCN-CA framework into In the closed-form expression of This enables joint optimization of user scheduling and beamforming design.

[0090] Some comparative examples are provided below to illustrate the superiority of the target deep learning model using the above GCN-CA framework.

[0091] In the first comparative example, the change trend of the training loss function curve of the GCN-CA model in the present invention and the original GCN model is compared. From the formula of the loss function, it can be seen that the decrease in the loss function value indicates that the value of the system and rate is increasing. Figure 4 This is a schematic diagram comparing the training convergence speed of different models. Figure 4 It can be seen that the GCN-CA model converges faster and better. This is because the cross-attention mechanism can better capture the characteristics between the channel vector and the power vector, thereby improving the performance of the network. In addition, in this comparison, the GCN model in the GCN-CA framework uses a single-layer graph convolution layer, which greatly reduces the complexity of the training and testing phases compared to the three-layer GCN framework. Therefore, the deep learning framework based on GCN-CA has a certain performance improvement.

[0092] Figure 5 The figure shows the average sum rate under different signal-to-noise ratios in a multi-user MISO system. In the second comparative example, the number of BS antennas and the number of users are set to N = 32 and K = 30, and the maximum sum rate values ​​obtained by different schemes under different signal-to-noise ratio values ​​are compared. The results show that the GCN-CA algorithm is significantly better than the USBD algorithm, the 3-layer GCN algorithm and the CNN algorithm. This is because the GCN-CA algorithm takes advantage of the GCN's advantages in processing graph-structured data and the advantages of the cross-attention mechanism in extracting channel depth features.

[0093] Figure 6 The average sum rate diagram for different numbers of users in a multi-user MISO system. In the third comparative example, the number of BS antennas and the signal-to-noise ratio are set to N = 32 and SNR = 10, and the maximum sum rate values ​​obtained by different schemes under different numbers of users are compared. Figure 6 It can be seen that the maximum sum rate obtained by the GCN-CA algorithm is higher than that of the USBD algorithm and the three-layer GCN algorithm. This is because we combine GCN with the cross-attention mechanism to deepen the correlation between the power vector and CSI, so that the GCN-CA algorithm can select more and better users for data transmission faster and better, thereby maximizing the system sum rate.

[0094] It can be seen from the above comparative examples that the optimization method based on the GCN-CA framework model provided by the present invention has the following technical effects:

[0095] 1. The magnitude and speed of the decrease in the training loss value of the GCN-CA framework model are better than those of the single GNN model, which shows that the sum rate performance and network convergence performance of the improved framework have been improved to a certain extent. This is because the cross-attention mechanism can better capture the features between the channel vector and the power vector, thereby improving the performance of the network. In addition, the GCN-CA framework has only one layer of network, which greatly reduces the computational complexity of the training and testing phases compared to the three-layer GCN framework. Therefore, the GCN-CA deep learning framework designed by the present invention has a certain performance improvement.

[0096] 2. The present invention compares the maximum system sum rates that can be achieved by different schemes under different signal-to-noise ratios, where the number of BS antennas and the number of users in the system are set to 32 and 30 respectively. Experiments show that compared with other schemes, the GCN-CA framework we proposed has a higher system sum rate. This is because the GCN-CA framework can better extract the features between inter-user interference and its own channel gain, maximize its own gain, and reduce inter-user interference. In addition, the GCN-CA scheme takes advantage of the advantages of GCN in processing graph-structured data, and the advantages of the cross-attention mechanism in further extracting channel depth features. Therefore, the GCN-CA algorithm has great advantages in improving the resource allocation performance of wireless communications.

[0097] 3. The present invention compares the maximum system sum rate that can be achieved by different schemes under different numbers of users, where the number of antennas and signal-to-noise ratio of the BS are set to 32 and 10dB respectively. Experiments show that the maximum sum rate obtained by the algorithm designed by the present invention is better than other comparison schemes. This is because the algorithm designed by the present invention combines GCN with the cross-attention mechanism, deepens the correlation between the power vector and CSI, and enables the GCN-CA algorithm to select more and better users for data transmission faster, thereby maximizing the system sum rate.

[0098] In an embodiment of the present invention, a wireless communication network optimization device based on deep learning is also provided, which is applied to a multi-user downlink MISO communication system. The multi-user downlink MISO communication system includes a network base station and a number of users, and the network base station provides network services for the multiple users. The device is used to implement the above-mentioned embodiments and preferred implementation modes, and those that have been described will not be repeated. The terms "module", "unit", "subunit", etc. used below can implement a combination of software and / or hardware for predetermined functions. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceived.

[0099] The wireless communication network optimization device comprises:

[0100] The power determination module is used to obtain complete channel information, input the channel information into the target deep learning model, and obtain the target power vector through the target deep learning model. The complete channel information includes the channel information between the network base station and several users. The target power vector includes the transmission power of the network base station on at least one target user. The target user is a scheduled user among several users.

[0101] The channel determination module is used to determine a target channel vector corresponding to a target power vector, where the target channel vector includes a channel vector between a network base station and at least one target user.

[0102] The beam determination module is used to determine the best beam vector of at least one target user according to the target channel vector.

[0103] Among them, the target deep learning model adopts unsupervised training and the loss function L loss for:

[0104]

[0105] Among them, R k represents the network rate of the kth user, μ and λ k represents the non-negative Lagrange multiplier, ReLU represents the activation function, L represents the minimum threshold of the number of user scheduling, and f θ represents a continuous smooth function, K represents the total number of users, and p k It represents the transmission power of the network base station to the kth user, SINR k represents the downlink signal-to-noise ratio of the kth user, η k Indicates SINR k The minimum threshold value.

[0106] It should be noted that the above modules can be functional modules or program modules, and can be implemented by software or hardware. For modules implemented by hardware, the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0107] An embodiment of the present invention also includes an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the wireless communication network optimization method based on deep learning provided by the present invention.

[0108] An embodiment of the present invention also includes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the wireless communication network optimization method based on deep learning provided by the present invention are implemented.

[0109] It should be understood that the specific embodiments described herein are only used to explain the application, rather than to limit it. Based on the embodiments provided in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of this application.

[0110] Obviously, the drawings are only some examples or embodiments of the present application. For ordinary technicians in the field, the present application can also be applied to other similar situations based on these drawings without creative work. In addition, it is understandable that although the work done in this development process may be complicated and lengthy, for ordinary technicians in the field, certain changes in design, manufacturing or production based on the technical content disclosed in this application are only conventional technical means and should not be regarded as insufficient content disclosed in this application.

Claims

1. A wireless communication network optimization method based on deep learning, applied to a multi-user downlink MISO communication system, wherein the multi-user downlink MISO communication system comprises a network base station and a plurality of users, wherein the network base station provides network services for the plurality of users, wherein: The wireless communication network optimization method comprises: Acquire complete channel information, input the channel information into a target deep learning model, and obtain a target power vector through the target deep learning model, wherein the complete channel information includes channel information between the network base station and the plurality of users, and the target power vector includes a transmit power of the network base station acting on at least one target user, and the target user is a scheduled user among the plurality of users; Determine a target channel vector corresponding to the target power vector, the target channel vector comprising a channel vector between the network base station and the at least one target user; Determining an optimal beam vector for the at least one target user according to the target channel vector; Among them, the target deep learning model adopts unsupervised training and the loss function L loss for: Among them, R k represents the network rate of the kth user, μ and λ k represents the non-negative Lagrange multiplier, ReLU represents the activation function, L represents the minimum threshold of the number of user scheduling, and f θ represents a continuous smooth function, K represents the total number of users, and p k represents the transmission power of the network base station acting on the kth user, SINR k represents the downlink signal-to-noise ratio of the kth user, η k Indicates SINR k The minimum threshold value.

2. The wireless communication network optimization method based on deep learning according to claim 1, characterized in that: The update rules of the two non-negative Lagrange multipliers are: Among them, ε v and ε λ Represent μ and λ respectively k The update step size, μ τ and μ τ+1 represent μ of the τth generation and the τ+1th generation respectively, and denote the λ of the τth generation and the τ+1th generation respectively k .

3. The wireless communication network optimization method based on deep learning according to claim 1, characterized in that: The optimal beam vector of the at least one target user is determined by a target formula, which is: in, represents the optimal beam vector for the kth user, express The phase of express The amplitude, represents the kth power value in the target power vector, v and λ k represents the Lagrange multiplier associated with the power constraint and the quality of service constraint, I N represents the N-dimensional identity matrix, K represents the total number of users, and h k represents the channel vector between the network base station and the kth user, h k Variance of medium white Gaussian noise.

4. The method for optimizing a wireless communication network based on deep learning according to claim 2, characterized in that: The target deep learning model includes a graph neural network module and a cross attention module, the input of the graph neural network module is the input of the target deep learning model, the input of the cross attention module is the input of the target deep learning model and the output of the graph neural network module, and the output of the cross attention module is the output of the target deep learning model; In the target deep learning model, the channel information is represented by a wireless channel graph; in the wireless channel graph, a node represents a user, and an edge represents an interference signal between users; Among them, the feature of the i-th node is defined as The edge between the i-th node and the j-th node is defined as 5. The method for optimizing a wireless communication network based on deep learning according to claim 4, characterized in that: The graph neural network module includes a plurality of graph convolutional layers connected in sequence, and the feature processing flow of the lth graph convolutional layer includes: in, represents the embedding representation vector of the jth node output by the l-1th graph convolutional layer, represents the embedding representation vector of the i-th node output by the l-1-th graph convolutional layer, represents the embedding representation vector of the i-th node output by the l-th graph convolutional layer, It represents the aggregation result of neighborhood features of the ith node in the lth graph convolutional layer, the jth node is the neighborhood node of the ith node, MLP1 and MLP2 are both multi-layer perceptrons, and PE represents the permutation equivalent function.

6. The method for optimizing a wireless communication network based on deep learning according to claim 4, characterized in that: The feature processing flow of the cross-attention module includes: Performing three parallel linearization processes on the input of the target deep learning model and the graph neural network module to obtain a first feature, a second feature, and a third feature respectively; Performing a convolution operation on the first feature and the second feature to obtain an attention matrix; Activating the attention matrix and convolving it with the third feature to obtain a first combination result; Linearly combining the first combination result and the third feature to obtain a second combination result; The second combination result is processed through a feedforward network to obtain a candidate power vector, and the target power vector is determined according to the candidate power vector.

7. The method for optimizing a wireless communication network based on deep learning according to claim 6, characterized in that: Determining the target power vector according to the candidate power vector comprises: Each power value in the target power vector is obtained by projecting each power value in the candidate power vector through a projection function, and the projection function is: Among them, p k represents the kth power value in the candidate power vector, represents the kth power value in the target power vector, which is used to characterize the power of the network base station on the kth user. ReLU(x)=max(x,0) represents the integer linear unit activation function, and P represents the total transmit power of the network base station.

8. A wireless communication network optimization device based on deep learning, applied to a multi-user downlink MISO communication system, wherein the multi-user downlink MISO communication system comprises a network base station and a plurality of users, wherein the network base station provides network services for the plurality of users, characterized in that: The wireless communication network optimization device comprises: a power determination module, configured to obtain complete channel information, input the channel information into a target deep learning model, and obtain a target power vector through the target deep learning model, wherein the complete channel information includes channel information between the network base station and the plurality of users, and the target power vector includes a transmit power of the network base station acting on at least one target user, and the target user is a scheduled user among the plurality of users; A channel determination module, configured to determine a target channel vector corresponding to the target power vector, wherein the target channel vector includes a channel vector between the network base station and the at least one target user; A beam determination module, configured to determine an optimal beam vector of the at least one target user according to the target channel vector; Among them, the target deep learning model adopts unsupervised training and the loss function L loss for: Among them, R k represents the network rate of the kth user, μ and λ k represents the non-negative Lagrange multiplier, ReLU represents the activation function, L represents the minimum threshold of the number of user scheduling, and f θ represents a continuous smooth function, K represents the total number of users, and p k represents the transmission power of the network base station acting on the kth user, SINR k represents the downlink signal-to-noise ratio of the kth user, η k Indicates SINR k The minimum threshold value.

9. An electronic device, comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to execute the wireless communication network optimization method based on deep learning according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the wireless communication network optimization method based on deep learning described in any one of claims 1 to 7 are implemented.

Images