Wireless Communication Resource Allocation Method Based on Deep Unfolding Distributed Graph Neural Network

Abstract

A wireless communication resource allocation method based on a depth-unrolled distributed graph neural network provided by the present invention obtains a channel gain matrix of the current time slot from a D2D network and converts it into a graph data structure stored by an adjacency list; any transmitter determines its own user channel gain and the status information of other transmitters in the graph data structure; taking the user channel gain of the current time slot, the transmission power of the previous time slot, and the status information of other transmitters as inputs, and aiming at maximizing the sum of communication rates, through the operation mode of a multi-layer graph neural network simulated and trained in multiple time slots, the finally allocated transmission power is obtained to transmit signals. The present invention can enhance the interpretability and generalization of the graph neural network, reduce the dependence of the graph neural network on training samples, use distributed parallel processing for iterative processing of transmission power, reduce the task calculation delay, and meet the actual requirements of future D2D network communication.

Description

Technical Field

[0001] The present invention belongs to the field of communication technologies, and particularly relates to a wireless communication resource allocation method based on a deep unfolding distributed graph neural network. Background Art

[0002] With the continuous development of the wireless communication field, the future communication transmission rate is continuously increasing, there are various types of communication devices, the network environment is complex and changeable, and the communication tasks have an increasingly lower demand for latency. The rapidly growing ultra-low latency communication requirements in the communication network have brought two major challenges to wireless resource management: high performance and low complexity. Thus, a very important problem is highlighted. In the case of a large number of communication scenarios and complex network heterogeneity, it is impossible to perform good unified resource allocation and management for a large number of communication devices and tasks.

[0003] In the existing D2D network communication resource allocation algorithms, there are mainly the following several methods for the access and unified management of communication devices.

[0004] The first is through the iterative convex optimization algorithm. This method uses mathematical model tools to add auxiliary variables or auxiliary formulas in the non-convex communication resource allocation problem, convert the non-convex problem into a convex optimization problem, and solve it through an iterative search method. Such as the iterative water filling algorithm, weighted minimum mean square error algorithm, successive convex approximation algorithm, etc. However, the iterative convex optimization algorithm usually has extremely high computational complexity. To meet higher performance requirements, the algorithm running time is relatively long, and it often cannot meet the requirements of ultra-low latency communication tasks.

[0005] The second is through deep learning methods. This method uses neural networks to learn a large number of communication resource allocation samples, combines the backpropagation method, and converts the communication resource allocation problem into a matrix multiplication problem. By continuously trial and error correction in a large number of samples, the neural network has the ability of fuzzy approximate solution. Such as DNN, CNN, DQN, etc. However, neural networks rely on a large number of high-precision training sample data. However, it takes a relatively long time to collect sufficient training data in radio resource management. And the performance of neural networks decreases significantly with the change of data distribution, and there is no theoretical guarantee.

[0006] As can be seen from the above, in the existing D2D network communication resource allocation algorithms, it is particularly important to solve the unified resource allocation of a large number of communication devices and communication tasks. In the above two solutions, the timeliness, practical feasibility, and cost of resource allocation cannot well meet the actual needs of future D2D network communication. Summary of the Invention

[0007] To solve the above problems existing in the prior art, the present invention provides a wireless communication resource allocation method based on a deep unfolded distributed graph neural network. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0008] The present invention provides a wireless communication resource allocation method based on a deep unfolded distributed graph neural network, which is applied to a transmitter of a D2D network. The wireless communication resource allocation method based on a deep unfolded distributed graph neural network includes:

[0009] Step 1, obtaining a channel gain matrix of the current time slot from the D2D network;

[0010] Step 2, converting the channel gain matrix of the current time slot into a graph data structure stored by an adjacency list;

[0011] Wherein, the graph data structure is composed of multiple nodes and directed edges. Each node corresponds to a node attribute. A node represents a transmitter and its corresponding receiver. The node attributes include transmission channel gain and the power of the transmitter. The directed edges between nodes represent the interference channels of other transmitters to the receivers pointed by the directed edges, and the directed edge attributes include the channel gains of the interference channels;

[0012] Step 3, any transmitter determines its own user channel gain and the status information of other transmitters in the graph data structure;

[0013] Wherein, the status information includes the channel gain of the interference channel in the current time slot and the power allocation information in the previous time slot;

[0014] Step 4, taking the user channel gain in the current time slot, the transmission power in the previous time slot, and the status information of other transmitters as inputs, and through the operation mode of a multi-layer graph neural network trained by simulation through multiple time slots with the goal of maximizing the sum of communication rates, obtaining the finally allocated transmission power;

[0015] Wherein, each time slot simulates one layer in the graph neural network for iteration;

[0016] Step 5, transmitting signals according to the finally obtained transmission power.

[0017] Advantages of the present invention:

[0018] A wireless communication resource allocation method based on a deep unfolding distributed graph neural network provided by the present invention obtains a channel gain matrix of the current time slot from the D2D network; converts the channel gain matrix of the current time slot into a graph data structure stored by an adjacency list; any transmitter determines its own user channel gain and the status information of other transmitters in the graph data structure; takes the user channel gain of the current time slot, the transmission power of the previous time slot, and the status information of other transmitters as inputs, and through the operation mode of a multi-layer graph neural network simulated and trained through multiple time slots with the goal of maximizing the sum of communication rates, obtains the finally allocated transmission power; and transmits signals according to the finally obtained transmission power. Compared with the prior art, the present invention can enhance the interpretability and generalization of the graph neural network, reduce the dependence of the graph neural network on training samples, use distributed parallel processing for iterative processing of transmission power, and reduce the task calculation delay; in addition, the present invention can enhance the robustness of the graph neural network, so that the algorithm performance does not decrease with data changes, and reduce the time for collecting training data and retraining the network.

[0019] The following will further describe the present invention in detail with reference to the drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a schematic diagram of a D2D network scenario;

[0021] Figure 2 is a schematic flow chart of a wireless communication resource allocation method based on a deep unfolding distributed graph neural network of the present invention;

[0022] Figure 3 is an architecture diagram of the graph neural network of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The following further describes the present invention in detail with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0024] A wireless communication resource allocation method based on a deep unfolding distributed graph neural network provided by the present invention is applied to a transmitter in a D2D network. The D2D network scenario is as Figure 1 shown. A pair of D2D transceiver users form a node. When the transmitter in the D2D network sends a signal to the corresponding transmitter, in addition to the corresponding receiver receiving the signal of the corresponding transmitter, it also receives the signal interference of other transmitters, which is expressed as:

[0025]

[0026] where y i is the signal received by the i-th receiver, h ii s i is the signal of the corresponding transmitter. is the interference signal for other users, n i is the channel Gaussian white noise;

[0027] The communication rate of each transmitter can be obtained from the Shannon formula:

[0028]

[0029] where v i is the transmission power of the i-th transmitter, and h ji is the channel gain of the interference channel from the i-th transmitter to the j-th transmitter;

[0030] The objective of the present invention is to find the globally optimal transmitter power such that the sum of the communication rates of all user pairs is maximized. The sum of the communication rates of all nodes composed of transmitters and receivers is as follows:

[0031]

[0032] The constraint condition is

[0033]

[0034] where P max represents the maximum transmission power.

[0035] As Figure 2 shown, a wireless communication resource allocation method based on a deep unfolding distributed graph neural network provided by the present invention includes:

[0036] Step 1, obtaining the channel gain matrix of the current time slot from the D2D network;

[0037] Step 2, converting the channel gain matrix of the current time slot into a graph data structure stored by an adjacency list;

[0038] Referring to Figure 1 , the graph data structure is composed of multiple nodes and directed edges. Each node corresponds to a node attribute. A node represents a transmitter and its corresponding receiver. The node attribute includes the transmission channel gain and the power of the transmitter; the directed edge between nodes represents the interference channel of other transmitters to the receiver pointed by the directed edge, and the directed edge attribute contains the channel gain of the interference channel;

[0039] The present invention converts the channel gain matrix H of the D2D network into a graph data structure stored by an adjacency list. The node contains a pair of user pairs. The node attribute includes the transmission channel gain, the user transmitter power, etc. The user interference channel is used as the edge between two nodes, and the edge attribute contains the interference channel gain.

[0040] Step 3: Any transmitter determines its own user channel gain and the status information of other transmitters in the graph data structure;

[0041] Among them, the status information includes the channel gain of the interference channel in the current time slot and the power allocation information in the previous time slot;

[0042] Step 4: Use the user channel gain in the current time slot, the transmit power in the previous time slot, and the status information of other transmitters as inputs, and through the operation mode of the multi-layer graph neural network trained by simulating multiple time slots with the goal of maximizing the sum of communication rates, obtain the finally allocated transmit power;

[0043] Among them, each time slot simulates one layer in the graph neural network for iteration;

[0044] In a specific embodiment, Step 4 includes:

[0045] Step 41: Use the user channel gain in the current time slot, the transmit power in the previous time slot, and the status information of other transmitters as the input of the first layer of the trained graph neural network, simulate the function of the first layer to process the input in the current time slot, and obtain the transmit power output by the first layer in the current time slot;

[0046] Step 42: Use the user channel gain in the current time slot, the transmit power output by the first layer, and the status information of other transmitters as the input of the second layer of the trained graph neural network, simulate the function of the second layer to process the input in the current time slot, and obtain the transmit power output by the second layer in the next time slot;

[0047] Step 43: For the i-th iteration, use the user channel gain in the current time slot, the transmit power output by the (i - 1)-th layer, and the status information of other transmitters as the input of the i-th layer of the trained graph neural network, simulate the function of the i-th layer to process the input in the current time slot, and obtain the transmit power output by the i-th layer in the i-th time slot;

[0048] Step 44: Let i = i + 1, repeat Step 43 until the value of i is the same as the number of layers of the graph neural network, and obtain the finally allocated transmit power output by the last layer.

[0049] Reference Figure 3 As shown, the graph neural network is composed of cascaded multi-layer GNN networks, and each layer of GNN network is sequentially connected by GNN u,w sub-network and GNN v sub-network; The GNN u,w sub-network of each layer is responsible for calculating the utility information u and weight information w required for power allocation; The GNN v sub-network of each layer is responsible for calculating the transmit power v of the transmitter; GNN u,wSub-network and GNN v The sub-network is jointly constructed by 5 multi-layer perceptron LMP modules. The 5 multi-layer perceptron LMP modules are respectively: GNN u,w The sub-network consists of a distributed neighbor node F(mu i ) message passing module, an F(u) information conversion module, an F(w) information conversion module, a distributed power F(mvi) message passing module, and a power F(v) calculation module;

[0050] For any transmitter,

[0051] The distributed neighbor node F(m ui ) message aggregation module is used to input the status information of neighbor transmitters and aggregate the status information;

[0052] The F(u) information conversion module is used to process the aggregated status information, the transmission power of the previous time slot of the any transmitter, and the user channel gain of the current time slot in an algorithm equivalent to formula (4) to obtain the utility information u of the any transmitter, and transmit the utility information u to neighbor transmitters;

[0053] The F(w) information conversion module is used to process the utility information u of the any transmitter and the transmission power of the previous time slot of the any transmitter in an algorithm equivalent to formula (5) to obtain the weight information w of the any transmitter, and transmit the weight information w to neighbor transmitters;

[0054] The distributed power F(mvi) message aggregation module is used to collect the utility information u, weight information w of neighbor transmitters, and the channel gain of the interference channel to obtain pre-power allocation information;

[0055] The power F(v) calculation module is used to aggregate the pre-power allocation information, its own utility information u, and weight information w in an algorithm equivalent to formula (6) to obtain the transmission power allocated in the current time slot.

[0056] Combined communication resource allocation WMMSE algorithm, that is, formula (4), formula (5), and formula (6) are respectively:

[0057]

[0058]

[0059]

[0060] where h ii represents the channel gain between the i-th transmitter and the corresponding receiver, σ i represents the channel Gaussian white noise, v idenotes the transmission power of the i-th transmitter, w j denotes the weight of the j-th transmitter, h ij denotes the channel gain of the interference channel from the i-th receiver to the j-th transmitter, u i denotes the utility information of the i-th transmitter, k denotes the k-th iteration or the k-th layer of the graph neural network, and α denotes the priority weight of the i-th transmitter in the sum rate problem.

[0061] After deploying the graph neural network on D2D devices, the power allocation algorithm is an iterative algorithm architecture and requires multiple message passing iterations. Whenever an iteration cycle starts, all nodes perform the following process distributively. Taking a single node in a single cycle as an example, the present invention illustrates the algorithm operation process:

[0062] User node i first sends the power v of the previous iteration i (l-1) information and adjacent channel h ij The interference information is processed into m using the F(mui) message passing module u information and sent to neighbor node j. After receiving the information sent by all neighbor nodes m uj User node j uses the F(u), F(w) information conversion module and uses GNN uw to aggregate all neighbor node information and generate u i , w i information in combination with its own power of the previous iteration. User node i uses the distributed power F(mvi) message passing module and combines u i , w i information and interference channel information h ij to generate the pre-power allocation information m vi , and then sends the generated pre-power allocation information m vi to neighbor node j. After receiving the pre-power allocation information m sent by all neighbors vj , user node j combines its own u i , w i information and uses the power F(v) calculation module to generate the power allocation result v of this iteration i (1) .

[0063] In a specific embodiment, the way to train the multi-layer graph neural network is:

[0064] Adopt an unsupervised method, use formula (7) as the loss function of the graph neural network, and use the backpropagation (BP) algorithm to optimize the network parameters of the multi-layer graph neural network:

[0065]

[0066] ​​Among them, represents the loss function, Θ represents the learnable parameters of the neural network, represents the mean function.

[0067] Step 5, transmit the signal according to the final transmit power.

[0068] The distributed graph neural network algorithm of the present invention combines the knowledge of the communication resource allocation WMMSE algorithm, requires a lower number of training samples, is not sensitive to changes in data distribution, and can achieve algorithm reuse without additional training in various communication scenarios, reducing the deployment delay of the communication network. And the distributed architecture can reduce the computational complexity of the algorithm and the computational delay, enabling the resource allocation task to be completed in only milliseconds.

[0069] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0070] Although the present application has been described herein in connection with various embodiments, nevertheless, during the implementation of the claimed present application, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the accompanying drawings, the disclosed content, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality of cases.

[0071] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A wireless communication resource allocation method based on a deep unfolding distributed graph neural network, applied to the transmitter of a D2D network, characterized in that, The wireless communication resource allocation method based on the deep unfolded distributed graph neural network includes: Step 1: Obtain the channel gain matrix of the current time slot from the D2D network; Step 2: Convert the channel gain matrix of the current time slot into a graph data structure stored by an adjacency list; Among them, the graph data structure consists of multiple nodes and directed edges. Each node corresponds to a node attribute. A node represents a transmitter and its corresponding receiver. The node attributes include the transmission channel gain and the power of the transmitter. The directed edges between nodes represent the interference channels of other transmitters to the receivers pointed by the directed edges, and the directed edge attributes contain the channel gains of the interference channels; Step 3: Any transmitter determines its own user channel gain and the status information of other transmitters in the graph data structure; Among them, the status information includes the channel gain of the interference channel in the current time slot and the power allocation information in the previous time slot; Step 4: Take the user channel gain in the current time slot, the transmission power in the previous time slot, and the status information of other transmitters as inputs, and through the operation method of the multi-layer graph neural network trained by simulation in multiple time slots with the goal of maximizing the sum of communication rates, obtain the finally allocated transmission power; Among them, each time slot simulates an iteration of one layer in the graph neural network; Step 5: Transmit signals according to the finally obtained transmission power; The graph neural network is composed of cascaded multiple layers of GNN networks. Each layer of GNN network is sequentially connected by a GNN u,w sub-network and a GNN v sub-network; the GNN u,w sub-network of each layer is responsible for calculating the utility information u and weight information w required for power allocation; the GNN v sub-network of each layer is responsible for calculating the transmission power v of the transmitter; the GNN u,w sub-network and the GNN v sub-network are jointly constructed by 5 multi-layer perceptron LMP modules; the 5 multi-layer perceptron LMP modules are respectively: the GNN u,w sub-network is composed of a distributed neighbor node F(m ui ) message passing module, an F(u) information conversion module, an F(w) information conversion module, a distributed power F(mvi) message passing module, and a power F(v) calculation module; For any transmitter, The distributed neighbor node F(m ui ) message aggregation module is used to input the status information of neighbor transmitters and aggregate the status information; The F(u) information conversion module is used to process the aggregated status information, the transmission power of the previous time slot of the any transmitter, and the user channel gain in the current time slot in an algorithm equivalent to formula (4) to obtain the utility information u of the any transmitter, and transmit the utility information u to the neighboring transmitters; The F(w) information conversion module is used to process the utility information u of the any transmitter and the transmission power of the previous time slot of the any transmitter in an algorithm equivalent to formula (5) to obtain the weight information w of the any transmitter, and transmit the weight information w to the neighboring transmitters; The distributed power F(mvi) message aggregation module is used to aggregate the utility information u, weight information w of the neighboring transmitters, and the channel gain of the interference channel to obtain the pre-power allocation information; The power F(v) calculation module is used to aggregate the pre-power allocation information, its own utility information u, and weight information w in an algorithm equivalent to formula (6) to obtain the transmission power allocated in the current time slot; The formula (4), formula (5), and formula (6) are respectively: where h ii represents the channel gain between the i-th transmitter and the corresponding receiver, σ i represents the channel Gaussian white noise, v i represents the transmission power of the i-th transmitter, w j represents the weight of the j-th transmitter, h ij represents the channel gain of the interference channel from the j-th transmitter to the i-th receiver, u i represents the utility information of the i-th transmitter, k represents the k-th iteration or the k-th layer of the graph neural network, and α represents the priority weight of the i-th transmitter in the sum-rate problem.

2. The wireless communication resource allocation method based on a depth-unrolled distributed graph neural network according to claim 1, wherein When the transmitter in the D2D network sends a signal to the corresponding transmitter and the corresponding receiver receives the signal of the corresponding transmitter, it also receives the signal interference of other transmitters, which is expressed as: where y i is the signal received by the i-th receiver, h ii s i is the corresponding transmitter signal, is the interference signal from other users, n i is the channel Gaussian white noise; The communication rate of each transmitter can be obtained from the Shannon formula: where v i is the transmission power of the i-th transmitter, and h ji is the channel gain of the interference channel from the i-th transmitter to the j-th transmitter; The sum of the communication rates of all nodes composed of transmitters and receivers is as follows: The constraint condition is Among them, P max represents the maximum transmit power.

3. A wireless communication resource allocation method based on a depth-unrolled distributed graph neural network according to claim 1, wherein Step 4 includes: Step 41: Use the user channel gain of the current time slot, the transmission power of the previous time slot, and the status information of other transmitters as the input to the first layer of the trained graph neural network, simulate the function of the first layer to process the input of the current time slot, and obtain the transmission power output by the first layer at the current time slot; Step 42: Use the user channel gain of the current time slot, the transmission power output by the first layer, and the status information of other transmitters as the input to the second layer of the trained graph neural network, simulate the function of the second layer to process the input of the current time slot, and obtain the transmission power output by the second layer at the next time slot; Step 43: For the i-th iteration, use the user channel gain of the current time slot, the transmission power output by the (i - 1)-th layer, and the status information of other transmitters as the input to the i-th layer of the trained graph neural network, simulate the function of the i-th layer to process the input of the current time slot, and obtain the transmission power output by the i-th layer at the i-th time slot; Step 44: Let i = i + 1, repeat Step 43 until the value of i is the same as the number of layers of the graph neural network, and obtain the finally allocated transmission power output by the last layer.

4. A wireless communication resource allocation method based on a depth-unrolled distributed graph neural network according to claim 1, characterized in that, The method for training the multi-layer graph neural network is as follows: Adopt an unsupervised method, use formula (7) as the loss function of the graph neural network, and use the backpropagation (BP) algorithm to optimize the network parameters of the multi-layer graph neural network: Among them, \(l(\Theta)\) represents the loss function, and \(\Theta\) represents the learnable parameters of the neural network. represents the mean function.

Images