Information processing method and apparatus

By introducing two nonlinear units and permutation into GNN, the problem of poor generalization performance in wireless communication is solved, and the optimal precoding strategy learning in hypergraphs and the effective use of digital and analog hybrid precoding in digital and analog solve.

WO2024243813A1PCT designated stage expired Publication Date: 2024-12-05HUAWEI TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2023/097203
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-30
Publication Date
2024-12-05

AI Technical Summary

Technical Problem

Existing graph neural networks (GNNs) have poor generalization performance in wireless communications, and cannot effectively learn the optimal precoding strategy, which is particularly difficult to solve the problem of mixed precoding of digital and analog, and cannot be applied to hypergraphs.

Method used

A graph neural network containing two nonlinear units is designed to calculate the correlation between target edges and nodes in graph data, and to realize nonlinear functions through neural networks, with permutation and other degeneration, which can adapt to multidimensional and Multi-node type, directly using edge features for solving, suitable for hypergraphs.

Benefits of technology

It realizes solving the optimal precoding strategy in wireless communication systems, improves the generalization ability of GNN, can effectively solve the digital and analog hybrid precoding problems, and is suitable for dynamically changing wireless environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2023097203_05122024_PF_FP_ABST
    Figure CN2023097203_05122024_PF_FP_ABST
Patent Text Reader

Abstract

An information processing method, comprising: acquiring channel state information between at least one network device and at least one terminal device; and processing the channel state information by using a GNN, so as to obtain precoding between the at least one network device and the at least one terminal device, wherein the GNN comprises a first nonlinear unit and a second nonlinear unit. The first nonlinear unit is used for calculating the correlation between an edge feature of a target edge in graph data used by the GNN and an edge feature of each edge on each target node, so as to obtain at least one correlation result, wherein the target node comprises a node connected to the target edge and a node of the same category as the node connected to the target edge. The second nonlinear unit is used for aggregating the edge feature of the target edge and a result obtained by summarizing the at least one correlation result, so as to update the edge feature of the target edge. By means of the method, an optimal precoding policy can be learnt, and problems in hybrid digital-analog precoding can be solved.
Need to check novelty before this filing date? Find Prior Art

Description

Information processing method and device Technical Field

[0001] The present application relates to the field of artificial intelligence (AI) technology, and in particular to an information processing method and device. Background Art

[0002] Wireless policy optimization is a critical issue in wireless communications. Taking precoding as an example, implementing it in wireless communication systems can effectively improve system energy and spectrum efficiency. The wireless environment and communication system configurations change dynamically over time, causing the input and output dimensions of the wireless policy optimization problem to also change over time. Therefore, dimensionally scalable graph neural networks (GNNs) can be used to learn precoding and other policies in wireless environments. However, current GNNs have poor generalization performance and are incapable of solving hypergraph-related problems. This makes it difficult to learn optimal precoding strategies using GNNs, and they are unable to solve problems such as mixed digital and analog precoding.

[0003] Summary of the Invention

[0004] The present application provides an information processing method, an information processing device, a processing device, a computer storage medium, and a computer program product, which can use GNN to learn the optimal precoding strategy and solve the problem of mixed digital and analog precoding.

[0005] In a first aspect, the present application provides an information processing method, comprising: obtaining channel state information between at least one network device and at least one terminal device; processing the channel state information using a graph neural network (GNN) to obtain precoding between the at least one network device and the at least one terminal device. The GNN comprises a first nonlinear unit and a second nonlinear unit. The first nonlinear unit is used to calculate the correlation between the edge features of a target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result, wherein the target node comprises a node connected to the target edge and a node of the same category as the node connected to the target edge. The second nonlinear unit is used to aggregate the edge features of the target edge and the result obtained by aggregating at least one correlation result to update the edge features of the target edge.

[0006] In this way, after obtaining the channel state information between at least one network device and at least one terminal device, the channel state information is processed using the aforementioned GNN to obtain the precoding of the at least one network device and the at least one terminal device. Since the GNN not only considers the mutual influence between different edges in the graph data during the solution process, but also directly uses the edge features of the target edge (i.e., the edge to be solved) for solution. Therefore, when using the GNN to solve the precoding, the optimal precoding strategy can be solved. At the same time, the input of the first nonlinear unit in the two nonlinear units is the edge feature, rather than a specific matrix, which makes the first nonlinear unit not restricted by the two-dimensional matrix. Therefore, the GNN can be applied to the hypergraph, and then the problem of digital and analog mixed precoding can be solved by using the GNN solution. Exemplarily, when the graph constructed by the graph data used by the GNN is a hypergraph, the edges on each node contained in the graph data can be called hyperedges.

[0007] In a possible implementation, the precoding is digital precoding. In this case, the graph data includes: first-type nodes corresponding to network devices, second-type nodes corresponding to terminal devices, and edges connecting the first-type nodes and the second-type nodes.

[0008] In one possible implementation, precoding is a hybrid of digital and analog precoding. In this case, the graph data includes: first-category nodes corresponding to antennas of network devices, second-category nodes corresponding to antennas of terminal devices, third-category nodes corresponding to radio frequency links of network devices, and edges connecting the first-category nodes, the second-category nodes, and the third-category nodes. The initial features of the edges in the graph data are derived based on the channel state information. For example, the initial features of the edges in the graph data can be derived by performing a dimensionality increase operation on the channel state information.

[0009] In one possible implementation, both the first nonlinear unit and the second nonlinear unit are implemented by the neural network in the GNN. This means that the functions implemented by the first nonlinear unit and the second nonlinear unit are not derived through fixed formulas, but are trained by the NN using training data. This allows the two nonlinear units to adapt to a wider range of scenarios and improves the generalization ability of the GNN.

[0010] In one possible implementation, both the first and second nonlinear units exhibit n-1 permutation equivariance, where n is the number of node types in the graph data. This allows for dimensionally scalable GNNs and reduced training complexity when solving wireless policy optimization problems in wireless communication systems.

[0011] In a second aspect, the present application provides an information processing device, comprising: a processing module. The processing module is used to obtain channel state information between at least one network device and at least one terminal device, and use a graph neural network GNN to process the channel state information to obtain precoding between the at least one network device and the at least one terminal device. The GNN includes a first nonlinear unit and a second nonlinear unit. The first nonlinear unit is used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result, and the target node includes a node connected to the target edge and a node of the same category as the node connected to the target edge. The second nonlinear unit is used to aggregate the edge features of the target edge and the result obtained by summarizing at least one correlation result to update the edge features of the target edge. Exemplarily, the processing module can be a processor or a processing circuit, etc. For example, when the information processing device is a chip, the processing module can be a processing circuit.

[0012] In a possible implementation, the precoding is digital precoding. The graph data includes: first-type nodes corresponding to network devices, second-type nodes corresponding to terminal devices, and edges connecting the first-type nodes and the second-type nodes.

[0013] In one possible implementation, precoding uses a hybrid of digital and analog precoding. The graph data includes: first-category nodes corresponding to network device antennas, second-category nodes corresponding to terminal device antennas, third-category nodes corresponding to network device radio frequency links, and edges connecting the first, second, and third-category nodes. The initial features of the edges in the graph data are derived based on channel state information.

[0014] In one possible implementation, both the first nonlinear unit and the second nonlinear unit are implemented by a neural network in a GNN.

[0015] In a possible implementation, both the first nonlinear unit and the second nonlinear unit have n-1-dimensional permutation equivariance, where n is the number of node types in the graph data.

[0016] In a third aspect, the present application provides an information processing method, comprising: obtaining graph data obtained based on a wireless communication system; processing the graph data using a GNN to obtain a processing result. The GNN performs information processing based on a first nonlinear unit and a second nonlinear unit. The first nonlinear unit is used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result, and the target node includes a node connected to the target edge and a node of the same category as the node connected to the target edge. The second nonlinear unit is used to aggregate the edge features of the target edge and the result obtained by summarizing at least one correlation result to update the edge features of the target edge.

[0017] In this way, after obtaining the graph data in the wireless communication system, the graph data is processed using the aforementioned GNN to obtain the desired processing results. Since the GNN not only considers the mutual influence between different edges in the graph data during the solution process, but also directly uses the edge features of the target edge (i.e., the edge to be solved) for solution. Therefore, when using the GNN to solve precoding, the optimal processing result can be obtained. At the same time, the input of the first nonlinear unit in the two nonlinear units is the edge feature, rather than a specific matrix, which makes the first nonlinear unit not limited by the two-dimensional matrix, so it can be applied to hypergraphs.

[0018] In one possible implementation, both the first nonlinear unit and the second nonlinear unit are implemented by a neural network in a GNN.

[0019] In a possible implementation, both the first nonlinear unit and the second nonlinear unit have n-1-dimensional permutation equivariance, where n is the number of node types in the graph data.

[0020] In a fourth aspect, the present application provides an information processing device, comprising: a processing module. The processing module can be used to obtain graph data obtained based on a wireless communication system, process the graph data using a GNN, and obtain a processing result. The GNN performs information processing based on a first nonlinear unit and a second nonlinear unit. The first nonlinear unit is used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node, and obtain at least one correlation result. The target node includes a node connected to the target edge and a node of the same category as the node connected to the target edge. The second nonlinear unit is used to aggregate the edge features of the target edge and the results obtained by summarizing at least one correlation result to update the edge features of the target edge. Exemplarily, the processing module can be a processor or a processing circuit, etc. For example, when the information processing device is a chip, the processing module can be a processing circuit.

[0021] In one possible implementation, both the first nonlinear unit and the second nonlinear unit are implemented by a neural network in a GNN.

[0022] In a possible implementation, both the first nonlinear unit and the second nonlinear unit have n-1-dimensional permutation equivariance, where n is the number of node types in the graph data.

[0023] In a fifth aspect, the present application provides a processing device, comprising: at least one processor, configured to execute a computer program so that the method described in the first aspect or the third aspect is implemented.

[0024] In a possible implementation, the processing device further includes at least one memory, which can be used to store computer programs executed by the processor.

[0025] In a possible implementation, the processing device may be a chip or a chip system.

[0026] In a sixth aspect, the present application provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method described in the first aspect or the third aspect.

[0027] In a seventh aspect, the present application provides a computer program product, which, when executed on a processor, enables the processor to execute the method described in the first aspect or the third aspect.

[0028] It can be understood that the beneficial effects of the second aspect and the fourth to seventh aspects mentioned above can be found in the relevant descriptions of the first aspect and the third aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] FIG1 is a schematic diagram of the architecture of a wireless communication system provided in an embodiment of the present application;

[0030] FIG2 is a schematic diagram of a graph structure obtained based on a wireless communication system according to an embodiment of the present application;

[0031] FIG3 is a schematic diagram of nodes in graph data obtained based on a wireless communication system according to an embodiment of the present application;

[0032] FIG4 is a schematic diagram comparing data processing effects between a GNN provided in an embodiment of the present application and a traditional GNN;

[0033] FIG5 is a flow chart of an information processing method provided in an embodiment of the present application;

[0034] FIG6 is a flow chart of another information processing method provided in an embodiment of the present application;

[0035] FIG7 is a schematic structural diagram of an information processing device provided in an embodiment of the present application;

[0036] FIG8 is a schematic structural diagram of a processing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0037] The term "and / or" as used herein describes an association between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. The symbol " / " as used herein indicates that the related objects are in an "or" relationship, for example, A / B means either A or B.

[0038] The terms "first" and "second" in this specification and claims are used to distinguish different objects rather than to describe a specific order of objects. For example, "first response message" and "second response message" are used to distinguish different response messages rather than to describe a specific order of response messages.

[0039] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0040] In the description of the embodiments of the present application, unless otherwise specified, "multiple" means two or more, for example, multiple processing units means two or more processing units, etc.; multiple elements means two or more elements, etc.

[0041] First, some technical terms involved in the embodiments of this application are introduced.

[0042] (1) Substitution and other denaturation

[0043] Permutation equivariance (PE) is a property where changing the permutation of a function's inputs results in a corresponding permutation change in its outputs. Functions solving wireless policy optimization problems often exhibit PE.

[0044] Next, the technical solutions provided in the embodiments of the present application are introduced.

[0045] For example, Figure 1 shows a schematic diagram of the architecture of a wireless communication system provided by an embodiment of the present application. As shown in Figure 1, the wireless communication system 100 may include: a network device 110 and at least one terminal device 120. A wireless communication connection may be established between the network device 110 and each terminal device 120. In addition, wireless communication connections may also be established between the terminal devices 120, which is not limited here.

[0046] The network device 110 may be a base station, a base station controller, or a router. This embodiment does not impose any particular restrictions on the specific type of the network device 110. In some embodiments, the network device 110 may support communication protocols of different standards. For example, the network device 110 may be a base transceiver station (BTS) in a global system for mobile communications (GSM) or a code division multiple access (CDMA) system, or a wireless transceiver (NodeB) in a wideband code division multiple access (WCDMA) system, or an evolutionary Node B (eNB or eNode B) in a long term evolution (LTE) system, or a network device in a fifth generation mobile communication technology (5G) network, a sixth generation mobile communication technology (6G) network, or a future network, etc. In another possible scenario, different network nodes may respectively implement some functions of the base station. For example, the network node may be a centralized unit (CU), a distributed unit (DU), a CU-control plane (CP), a CU-user plane (UP), or a radio unit (RU). The CU and DU may be set separately, or may be included in the same network element, such as a baseband unit (BBU). The RU may be included in a radio frequency device or a radio frequency unit, such as a remote radio unit (RRU), an active antenna unit (AAU), or a remote radio head (RRH). In different systems, CU (or CU-CP and CU-UP), DU, or RU may also have different names, but those skilled in the art may understand their meanings.For example, in the ORAN system, the CU may also be referred to as an open CU (O-CU), the DU may also be referred to as an open DU (O-DU), the CU-CP may also be referred to as an open CU-CP (O-CU-CP), the CU-UP may also be referred to as an open CU-UP (O-CU-UP), and the RU may also be referred to as an open RU (O-RU). Any of the CU (or CU-CP, CU-UP), DU, and RU may be implemented by a software module, a hardware module, or a combination of a software module and a hardware module.

[0047] The terminal device 120 may be a mobile phone, tablet computer, desktop computer, laptop computer, handheld computer, notebook computer, ultra-mobile personal computer (UMPC), netbook, cellular phone, personal digital assistant (PDA), augmented reality (AR) device, virtual reality (VR) device, artificial intelligence (AI) device, wearable device, vehicle-mounted device, smart home device and / or smart city device, etc. This embodiment does not impose any special restrictions on the specific type of the terminal device 120.

[0048] In the wireless communication system shown in FIG1 , taking the network device 110 having N antennas and the terminal device 120 having K antennas as an example, the channel state information between the nth antenna in the network device 110 and the kth (k=1, 2, ..., K) terminal device 120 is denoted as h nk , where n = 1, 2, ..., N, then the channel state information from the network device 110 to the kth terminal device 120 is h k =[h 1k ,h 2k ,…,h Nk ], and the channel state information between the network device 110 and all the terminal devices 120 is H = [h1, h2, ..., h N ]. The graph structure used when using GNN for precoding can be shown in FIG2. In this case, the input of GNN can be the channel state information H, and the output can be the precoding scheme. FIG2 includes two types of nodes, namely the nodes corresponding to the antennas in the network device 110 and the nodes corresponding to the terminal device 120. n is the node corresponding to the nth (1≤n≤N) antenna in the network device 110, U kis the node corresponding to the kth (1≤k≤K) terminal device. In Figure 2, the feature (or representation) of the edge between two nodes in the GNN layer l can be obtained by GNN processing (i.e., aggregation and integration) the features of each edge connected to these two nodes in the GNN layer l-1. n Node U corresponding to the kth terminal device k Taking the edge (n, k) between as an example, the output of the edge (n, k) in the lth layer of GNN is recorded as Then the output of edge (n,k) in the l+1 layer of GNN is:

[0049] Among them, PL(.) is the pooling function, For aggregation and node U k Characteristics of connected edges except edge (n,k); For aggregation with node A n The features of the edges other than the edge (n, k) are connected. CB(.) is the combining function used to integrate the output of the edge (n, k) in the lth layer. The results of the aggregation and The result of aggregation. q(.) is used to extract edge features. P, Q, and S are all parameters to be learned or trained in GNN.

[0050] When q(.) is a linear function, CB(.) is a function composed of a linear equation and an activation function σ(·), and PL(.) is a summation function, "Formula 1" can be written as:

[0051] Furthermore, from "Formula 2", we can get the vector composed of the output of the lth layer in the GNN of all edges connected to the kth terminal device: Combined with "Formula 2", we can get the vector composed of the output of the l+1th layer in the GNN of all edges connected to the kth terminal device:

[0052] in, P, Q, and S are all parameters to be learned or trained in GNN.

[0053] In "Formula 3", function f V The input is and As can be seen, "Formula 3" simply adds the features of all edges and then integrates the result with the edge features of the edge to be solved to complete the update of the edge features of the edge to be solved. In wireless communication systems, signals generated by different devices often influence each other. This causes different edges in the graph data constructed based on wireless communication systems to often affect each other. However, the precoding method of "Formula 3" does not consider the correlation between different edges. Therefore, it is difficult to solve the optimal precoding strategy using this method.

[0054] In addition to using "Formula 3" to obtain the output (i.e., feature) of the edge (n, k) in the l+1 layer of the GNN, it can also be obtained based on the inversion operation of the two-dimensional matrix and the pseudo-inverse of the channel state information H and H through the first-order Taylor formula. The output of the edge (n, k) in the l+1 layer of the GNN obtained in this way is:

[0055] in, is the channel state information h from the network device 110 to the jth terminal device 120 j Perform the transpose operation to obtain, Represents the vector consisting of the outputs of the lth layer in the GNN for all edges connected to the jth terminal device. It can be understood as being used to aggregate nodes A and n and node U k Function of the characteristics of connected edges except edge (n,k); It can be understood as being used to combine the output of edge (n, k) at layer l and A function that aggregates the results.

[0056] Furthermore, from "Formula 4", we can get the vector composed of the output of the lth layer in the GNN of all edges connected to the kth terminal device: Combined with "Formula 4", we can get the vector composed of the output of the l+1th layer in the GNN of all edges connected to the kth terminal device:

[0057] In "Formula 5", since It is derived from the matrix inversion operation and specific mathematical formulas, which makes this method only applicable to two-dimensional graphs. In hypergraphs, the edge features are higher-dimensional tensors. For hypergraphs composed of high-dimensional tensors, "Formula 5" is not applicable. At the same time, it is necessary to use right This results in the subsequent use of has been transformed, making it difficult to use the original Direct solution makes it difficult to find the optimal precoding strategy.

[0058] In view of this, an embodiment of the present application provides a dimensionally generalizable GNN, which includes two nonlinear units. Dimensional generalization refers to the ability to process multiple dimensions and / or the ability to process a variable number of nodes in graph data. The first of the two nonlinear units can be used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result. The target nodes include nodes connected to the target edge and nodes of the same category as the nodes connected to the target edge. For example, referring to Figure 2, when the target edge is edge (A1, U1), the target nodes are nodes U1, U2, ..., U K or nodes A1, A2, ..., A N The second nonlinear unit in the two nonlinear units can be used to aggregate the edge features of the target edge and the results of summarizing at least one correlation result to update the edge features of the target edge. In this way, when using the GNN to solve, it is possible to take into account the mutual influence between different edges and directly use the edge features of the target edge for solving, so that when using the GNN to solve precoding, the optimal precoding strategy can be solved. At the same time, the input of the first nonlinear unit in the two nonlinear units is the edge feature, without the "Formula 5" This means that the first nonlinear unit is not restricted by the two-dimensional matrix, so it can be applied to hypergraphs, thereby solving the problem of digital and analog hybrid precoding.

[0059] Next, the structure and design principles of the dimensionally generalizable GNN provided in the embodiments of this application are introduced through two scenarios for solving the precoding scheme.

[0060] (1) Solving the digital precoding scenario

[0061] In the wireless communication system shown in FIG1 , when the precoding problem is digital precoding, the graph data used by the GNN includes: a first type of node corresponding to the network device 110, a second type of node corresponding to the terminal device 120, and edges connecting the first type of node and the second type of node. Exemplarily, the first type of node corresponding to the network device 110 may be a node corresponding to the antenna of the network device 110, or a node corresponding to the radio frequency link on the network device 110. In addition, the graph structure constructed by the graph data may be as shown in FIG2 , in which nodes A1 to AN may be the first type of nodes corresponding to the network device 110, and nodes U1 to Uk may be the second type of nodes corresponding to the terminal device 120.

[0062] In this precoding problem, when the order of the antennas of the network device 110 is determined by the permutation matrix Π AN The order of the terminal devices 120 is changed by the permutation matrix Ώ UE When the channel state information H between the network device 110 and all terminal devices changes, At this time, the rows and columns of the optimal precoding matrix also change accordingly, but the values ​​remain unchanged. Therefore, the precoding strategy at this time satisfies the following permutation equivariance: is the optimal precoding, F d (.) is H to The mapping (i.e. function) of F has permutation equivariance. d (.) is fitted, which can make the obtained digital precoding close to the optimal precoding, so that the digital precoding strategy obtained by using GNN can optimize the system performance.

[0063] From the above analysis, we can see that the GNN used to solve the digital precoding strategy needs to satisfy the permutation equivariance. in, w k is the precoding vector of the kth terminal device. At the same time, the channel state information between the nth antenna and the kth terminal device is recorded as h nk , and use it as the initial feature of edge (n, k). After multiple rounds of processing of the initial feature of edge (n, k) by GNN, the output of edge (n, k) by GNN (i.e., the feature of edge (n, k)) is the element w in the digital precoder W nk By w nk The precoding vector of the kth terminal device can be obtained

[0064] When using GNN for digital precoding, the feature vector composed of the features output by all edges connected to the node corresponding to the kth terminal device at the lth layer is Then the feature vector composed of the features output by all edges connected to the node corresponding to the k-th terminal device at the l+1 layer is:

[0065] In "Formula 6," q(.) is the processing function, implemented by the first nonlinear unit described above. f(.) is the combining function, implemented by the second nonlinear unit described above. Both q(.) and f(.) are nonlinear functions. Both q(.) and f(.) can be implemented using the neural network (NN) in the GNN. Furthermore, both q(.) and f(.) exhibit one-dimensional permutation equivariance, ensuring that the learned precoding strategy also satisfies PE for antenna order permutations, i.e., they conform to the following form:

[0066] In "Formula 7", ψ(.) and ξ(.) are both nonlinear functions and can be learned by NN. "Formula 7" can be understood as a function from x=[x1,…,x N ] T Map to y=[y1,…,y N ] T One-dimensional PE (1D-PE) function.

[0067] When using GNN to calculate the initial features h of all edges connected to the node corresponding to the k-th terminal device using "Formula 6" k =[h 1k ,h 2k ,…,h Nk ] After L rounds of processing, the output of GNN is That is, the precoding vector of the kth terminal device is obtained. k Corresponding to the channel state information of all antennas and the k-th terminal device, after performing L rounds of processing on the initial features of all edges adjacent to the nodes corresponding to other terminal devices using "Formula 6" through GNN, the precoding vectors of other terminal devices can be obtained, and then the digital precoding can be obtained.

[0068] (2) Solving the scenario of mixed digital and analog precoding

[0069] In the wireless communication system shown in Figure 1, when the precoding problem is hybrid precoding, the graph data used by GNN includes: a first type of node corresponding to the antenna of the network device, a second type of node corresponding to the antenna of the terminal device, a third type of node corresponding to the radio frequency (RF) link of the network device, and edges connecting the first type of nodes, the second type of nodes, and the third type of nodes. At this time, the graph constructed by the graph data can be a hypergraph, and the edges connecting the first type of nodes, the second type of nodes, and the third type of nodes in the hypergraph can be called hyperedges. Exemplarily, the nodes contained in the graph data at this time can be as shown in Figure 3, in which the first type of nodes are AN 1 to N, the second type of nodes are UE 1 to K, and the third type of nodes are RF links 1 to M. In Figure 3, the information between the terminal device and the antenna can be channel state information H, the information between the antenna and the RF link can be analog precoding WRF, and the information between the terminal device and the RF link can be digital precoding W BB .

[0070] As in the case where the precoding problem is digital precoding, when the order of the antennas of the network device 110 is determined by the permutation matrix Π AN Change the order of the terminal devices by the permutation matrix ∏ UEchanges, and the order of the RF chains is determined by the matrix ∏ RF When it changes, the channel state information H becomes At this time, the rows and columns of the optimal digital precoding matrix and the analog precoding matrix also change accordingly, but the values ​​remain unchanged. Therefore, the precoding strategy at this time satisfies the following permutation equivariance: is the optimal digital precoding, is the optimal analog precoding, F h (.) is H to and The mapping (i.e. function) of F has permutation equivariance. h (.) is fitted, which can make the obtained digital and analog hybrid precoding close to the optimal digital and analog hybrid precoding, so that the digital and analog hybrid precoding strategy obtained by using GNN can optimize the system performance.

[0071] Furthermore, in the hypergraph corresponding to the digital and analog hybrid precoding problem, the hyperedge connecting antenna n, RF link m and terminal device k can be recorded as a hyperedge (n, m, k), where 1≤n≤N, N is the number of antennas, 1≤m≤M, M is the number of RF links, and 1≤k≤K, K is the number of terminal devices. When using GNN for hybrid precoding, the channel state information H of the network device 110 and all terminal devices 120 can be first promoted to an N×M×K dimensional tensor, for example, the N×K dimensional channel state information is copied M times in the new dimension. Then, the obtained N×M×K dimensional tensor is input into the GNN to process the tensor using the GNN. After the N×M×K dimensional tensor is input into the GNN, the output of all hyperedges in the lth layer of the GNN constitutes a three-dimensional tensor in, is the output of the hyperedge (n,m,k).

[0072] Will Denoted as the matrix composed of the outputs of all hyperedges connected to the kth terminal device, which is the output of the lth layer in the GNN (i.e., D (l) ) is a slice in the terminal device dimension. At this time, all hyperedges connected to the node corresponding to the k-th terminal device are output at the l+1 layer as:

[0073] In "Formula 8", is a processing function implemented by the aforementioned first nonlinear unit. It is a combination function, which is realized by the aforementioned second nonlinear unit. and All of these can be achieved through NN in GNN. In addition, and They all have 2D permutation equivariance, so that the learned strategy for the sequential permutation of antennas and RF links also satisfies PE, that is, they all conform to the following form:

[0074] In "Formula 9", and are all nonlinear functions and can be learned through NN. "Formula 9" can be understood as a Map to 2D-PE function. The input and output of "Formula 9" is a two-dimensional matrix. In "Formula 9", and Both satisfy 1D-PE for the sequential permutation of antennas, and both have the form of the aforementioned "Formula 7", so that the learned strategy also satisfies PE for the sequential permutation of antennas or RF links.

[0075] After GNN processes the channel state information using Formula 8 for L rounds, the result output by GNN is That is, a hybrid precoding scheme is obtained.

[0076] It can be seen from the above "Formula 8" and "Formula 9" that in order to design a GNN that satisfies equivariance for the permutations of the terminal device sequence, antenna sequence, and RF link sequence, each terminal device, all RF links, and all antennas can be regarded as a whole, and a function that satisfies equivariance for the permutations of the terminal device sequence can be designed (i.e., "Formula 8"). Then, the processing function and the combination function in "Formula 8" are designed so that they all satisfy equivariance for the permutations of the RF link sequence (i.e., "Formula 9"). When designing the processing function and the combination function in "Formula 8", each RF link and all antennas can be regarded as a whole to design a function that satisfies equivariance for the permutations of the RF link sequence (i.e., "Formula 9"). Finally, among the functions designed to satisfy equivariance for the permutations of the RF link sequence, further design can be performed so that they satisfy equivariance for the permutations of the antenna sequence (i.e., "Formula 7").

[0077] From the introduction of the above two scenarios, we can see that GNN that satisfies the n-dimensional PE property has a "recursive" idea, where n is a positive integer and has nothing to do with the n in the nth antenna mentioned above. Specifically, first, the n-dimensional tensor output by each layer (such as the output D of the lth layer) is (l)) select one of the dimensions (such as the terminal device dimension), and use the composite function of "processing function + pooling function (such as summation, etc.) + combining function" (that is, the form of "Formula 8") to obtain a function that satisfies the permutation equivariance property for this dimension. Among them, the processing function and the combining function should satisfy (n-1)D-PE for the remaining n-1 dimensions. In order to satisfy the permutation equivariance, select another dimension from the remaining dimensions (such as the RF link dimension), and continue to use the composite function of "processing function + pooling function + combining function" to obtain a function that satisfies the permutation equivariance property for this dimension. And so on, until all n dimensions have the PE property. At the same time, the functions designed in GNN are all nonlinear functions.

[0078] In addition, from the above two scenarios of solving the precoding scheme, it can be learned that the update function used by the GNN provided in the embodiment of the present application is:

[0079] In "Formula 10", A vector, matrix, or tensor consisting of the outputs of the l+1th layer of the GNN for all edges connecting the kth node in the graph used by the GNN. F(.) is a nonlinear function implemented by the aforementioned second nonlinear unit, which has an n-1-dimensional PE, where n is the number of node types in the GNN graph. Q(.) is a nonlinear function implemented by the aforementioned first nonlinear unit, which also has an n-1-dimensional PE.

[0080] Taking precoding as an example, the beneficial effects of the dimensionally generalizable GNN provided by this embodiment are illustrated through simulation.

[0081] For example, consider a scenario where a base station with N=16 antennas serves K single-antenna mobile phones, and the signal-to-noise ratio is P max / σ 2 =10. Among them, P max is the maximum power of the signal, σ 2 is the noise power. In order to learn the precoding strategy, an unsupervised approach is used to train the GNN. All simulation results are obtained on a computer equipped with a 14-core Intel i9-9940X CPU, an Nvidia RTX 3080Ti GPU and 64GB of memory. The inputs in the training and test samples are generated by independent Rayleigh distribution channels, that is, the elements in the channel state information H obey the complex Gaussian distribution CN(0,1). Since an unsupervised approach is used for training, there is no need to generate labels as the expected output of each sample. Among them, as shown in Table 1, the network structure of RGNN (i.e., the dimensionally generalizable GNN provided in this embodiment) is compared with the network structure of traditional GNN (i.e., the GNN used in the relevant technology).

[0082] Table 1

[0083] Based on the network structure shown in Table 1, simulations were conducted to compare the performance of the two GNNs when the number of mobile phones changed, as shown in Figure 4. In Figure 4, GNN-NL refers to the traditional GNN in Table 1. As shown in Figure 4, compared to the traditional GNN, the dimensionally generalizable GNN provided in this embodiment maintains performance exceeding 90% of the optimal solution when the number of users changes, while the performance of the traditional GNN degrades significantly.

[0084] The above is an introduction to the structure and design principles of the dimensionally generalizable GNN provided in the embodiment of this application. Next, based on the above content, an information processing method provided in the embodiment of this application is introduced. Among them, this information processing method mainly uses the aforementioned GNN for data processing.

[0085] For example, FIG5 shows a flow chart of an information processing method provided by an embodiment of the present application. It is understood that the method can be executed by any device, equipment, platform, or device cluster with computing and processing capabilities. As shown in FIG5, the information processing method may include the following steps:

[0086] S501: Acquire channel state information between at least one network device and at least one terminal device.

[0087] In this embodiment, the channel state information between at least one network device and at least one terminal device may be calculated using, but is not limited to, a channel estimation technology.

[0088] S502. Use GNN to process the channel state information to obtain precoding between the at least one network device and the at least one terminal device, and the GNN includes a first nonlinear unit and a second nonlinear unit; wherein the first nonlinear unit is used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result, and the target nodes include nodes connected to the target edge and nodes of the same category as the nodes connected to the target edge; the second nonlinear unit is used to aggregate the edge features of the target edge and the results obtained by summarizing at least one correlation result to update the edge features of the target edge.

[0089] In this embodiment, after obtaining the channel state information, the channel state information can be input into the GNN, so that the GNN processes the channel state information to obtain a precoding scheme between the at least one network device and the at least one terminal device. The GNN used in this information processing method can be the aforementioned GNN. Exemplarily, the first nonlinear unit can implement the nonlinear function Q(.) in the aforementioned "Formula 10", and the second nonlinear unit can implement the nonlinear function F(.) in the aforementioned "Formula 10".

[0090] As a possible implementation, the precoding to be solved is digital precoding. In this case, the graph data used by the GNN may include: first-type nodes corresponding to network devices, second-type nodes corresponding to terminal devices, and edges connecting the first-type nodes and the second-type nodes. The initial characteristics of the edges in this graph data may be channel state information between network devices and terminal devices in a wireless communication system. Exemplarily, the nodes corresponding to the network devices may include: nodes corresponding to the antennas of the network devices or nodes corresponding to the radio frequency links of the network devices.

[0091] As another possible implementation method, the precoding solved is a mixed digital and analog precoding. At this time, the graph data used by the GNN may include: a first type of node corresponding to the antenna of the network device, a second type of node corresponding to the antenna of the terminal device, a third type of node corresponding to the radio frequency link of the network device, and an edge connecting the first type of node, the second type of node and the third type of node (also referred to as a "hyperedge"). The initial features of the edge in the graph data can be obtained based on the channel state information between the network device and the terminal device in the wireless communication system, for example, the aforementioned dimensionality increase operation on the channel state information.

[0092] In this way, after obtaining the channel state information between at least one network device and at least one terminal device, the channel state information is processed using the aforementioned GNN to obtain the precoding between the at least one network device and the at least one terminal device. Since the GNN not only considers the mutual influence between different edges in the graph data during the solution process, but also directly uses the edge features of the target edge (i.e., the edge to be solved) for solution. Therefore, when using the GNN to solve the precoding, the optimal precoding strategy can be solved. At the same time, the input of the first nonlinear unit in the two nonlinear units is the edge feature, without the "Formula 5" That is, it is not a specific matrix, which means that the first nonlinear unit is not restricted by the two-dimensional matrix. Therefore, the GNN can be applied to the hypergraph, and then the GNN solution can solve the problem of digital and analog mixed precoding.

[0093] In some embodiments, in a GNN, both the first nonlinear unit and the second nonlinear unit are implemented by the neural network in the GNN. This means that both the first nonlinear unit and the second nonlinear unit are not derived through fixed formulas, but are implemented by the neural network, making the two nonlinear units adaptable to a wider range of scenarios and improving the generalization ability of the GNN.

[0094] Furthermore, in the GNN, both the first and second nonlinear units exhibit n-1-dimensional permutation equivariance, where n is the number of node types in the graph data used by the GNN. This allows the GNN to adapt to changes in the order of devices in wireless communication systems when solving problems in these systems.

[0095] For example, FIG6 shows a flow chart of another information processing method provided by an embodiment of the present application. It is understood that the method can be executed by any device, equipment, platform, or device cluster with computing and processing capabilities. As shown in FIG6, the information processing method may include the following steps:

[0096] S601: Acquire graph data based on a wireless communication system.

[0097] In this embodiment, the graph data may be pre-constructed based on a wireless communication system. The initial features of the edges in the graph data may include: one or more of: pilot information, channel state information, modulation information, coding information, scheduling information, link quality information, quality of service information (QoS information), spectrum management information, clock and synchronization information, access control information, power control information, network topology information, error control information, message and control signaling, latency information, energy efficiency information, media access control information, and resource allocation information.

[0098] S602: Process the graph data using a GNN to obtain a processing result, wherein the GNN includes a first nonlinear unit and a second nonlinear unit. The first nonlinear unit is configured to calculate a correlation between an edge feature of a target edge in the graph data used by the GNN and an edge feature of each edge on each target node to obtain at least one correlation result, wherein the target node includes a node connected to the target edge and a node of the same category as the node connected to the target edge; and the second nonlinear unit is configured to aggregate the edge feature of the target edge and a result obtained by aggregating the at least one correlation result to update the edge feature of the target edge. Exemplarily, when the initial characteristic of the edge in the graph data is pilot information, the processing result may be a channel estimation result; when the initial characteristic of the edge in the graph data is modulation, the processing result may be a modulated signal; when the initial characteristic of the edge in the graph data is coding information, the processing result may be a coded signal; when the initial characteristic of the edge in the graph data is scheduling information, the processing result may be a scheduling result or a resource allocation result; when the initial characteristic of the edge in the graph data is link quality information, QoS information, resource allocation information, delay information or energy efficiency information, the processing result may be a resource allocation result; when the initial characteristic of the edge in the graph data is spectrum management information, the processing result may be a spectrum allocation result.

[0099] In this way, after obtaining the graph data in the wireless communication system, the graph data can be processed using the aforementioned GNN to obtain the desired processing results. Since the GNN not only considers the mutual influence between different edges in the graph data during the solution process, but also directly uses the edge features of the target edge (i.e., the edge to be solved) for solution. Therefore, when using the GNN to solve the wireless strategy optimization problem, the optimal processing result can be obtained. At the same time, the input of the first nonlinear unit in the two nonlinear units is all edge features, without the "Formula 5" That is, it is not a specific matrix, which makes the first nonlinear unit not restricted to two-dimensional matrices, so it can be applied to hypergraphs.

[0100] It is understood that the order of execution of the steps in the above embodiments does not necessarily imply a specific order of execution. The order of execution of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of this application. In addition, all or part of any of the features of any of the above embodiments may be freely and arbitrarily combined, provided that there is no contradiction; the resulting technical solutions are also within the scope of this application.

[0101] Based on the method in the above embodiment, an embodiment of the present application provides an information processing device for executing the information processing method described in FIG. 5 .

[0102] For example, FIG7 is a schematic diagram of the structure of an information processing device provided in an embodiment of the present application. As shown in FIG7 , the information processing device 700 includes: an acquisition module 701 and a processing module 702. The acquisition module 701 is used to acquire channel state information between at least one network device and at least one terminal device. The processing module 702 is used to process the channel state information using a graph neural network (GNN) to obtain a precoding scheme between the at least one network device and the at least one terminal device. It is understood that the acquisition module 701 and the processing module 702 can be combined into one processing module. The GNN includes a first nonlinear unit and a second nonlinear unit. The first nonlinear unit is used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result, wherein the target node includes a node connected to the target edge and a node of the same category as the node connected to the target edge. The second nonlinear unit is used to aggregate the edge features of the target edge and the result obtained by summarizing at least one correlation result to update the edge features of the target edge.

[0103] In some embodiments, the precoding is digital precoding. The graph data includes: first-type nodes corresponding to network devices, second-type nodes corresponding to terminal devices, and edges connecting the first-type nodes and the second-type nodes.

[0104] In some embodiments, precoding is a hybrid of digital and analog precoding. The graph data includes: first-type nodes corresponding to antennas of network devices, second-type nodes corresponding to antennas of terminal devices, third-type nodes corresponding to radio frequency links of network devices, and edges connecting the first-type nodes, the second-type nodes, and the third-type nodes. The initial features of the edges in the graph data are derived based on channel state information.

[0105] In some embodiments, the first nonlinear unit and the second nonlinear unit are both implemented by a neural network in a GNN.

[0106] In some embodiments, the first nonlinear unit and the second nonlinear unit both have n-1 dimensional permutation equivariance, where n is the number of node types in the graph data.

[0107] It should be understood that the above-mentioned device is used to execute the method in the above-mentioned embodiment. The implementation principle and technical effect of the corresponding program module in the device are similar to those described in the above-mentioned method. The working process of the device can refer to the corresponding process in the above-mentioned method and will not be repeated here.

[0108] Based on the method in the above embodiment, an embodiment of the present application provides an information processing device for executing the information processing method described in FIG. 6 . The information processing device includes a processing module. The processing module is configured to obtain graph data obtained based on a wireless communication system and process the graph data using a GNN to obtain a processing result. The GNN performs information processing based on a first nonlinear unit and a second nonlinear unit.

[0109] The first nonlinear unit is used to calculate the correlation between the edge feature of the target edge in the graph data used by the GNN and the edge features of each edge on each target node, thereby obtaining at least one correlation result. The target nodes include nodes connected to the target edge and nodes of the same category as the node connected to the target edge. The second nonlinear unit is used to aggregate the edge feature of the target edge and the result obtained by aggregating the at least one correlation result to update the edge feature of the target edge.

[0110] In some embodiments, the first nonlinear unit and the second nonlinear unit are both implemented by a neural network in a GNN.

[0111] In some embodiments, the first nonlinear unit and the second nonlinear unit both have n-1 dimensional permutation equivariance, where n is the number of node types in the graph data.

[0112] It should be understood that the above-mentioned device is used to execute the method in the above-mentioned embodiment. The implementation principle and technical effect of the corresponding program module in the device are similar to those described in the above-mentioned method. The working process of the device can refer to the corresponding process in the above-mentioned method and will not be repeated here.

[0113] Based on the methods in the above embodiments, this application also provides a processing device 800 for executing the above methods. As shown in Figure 8, processing device 800 includes: a bus 802, a processor 804, a memory 806, and a communication interface 808. Processor 804, memory 806, and communication interface 808 communicate with each other via bus 802. Processing device 800 can be the aforementioned network device or terminal device, etc. It should be understood that this application does not limit the number of processors and memories in processing device 800.

[0114] Bus 802 may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, among others. Buses may be classified as address buses, data buses, control buses, and the like. For ease of illustration, FIG8 illustrates a single bus line, but this does not imply a single bus or type of bus. Bus 804 may include a path for transmitting information between various components of processing device 800 (e.g., memory 806, processor 804, and communication interface 808).

[0115] The processor 804 may include any one or more processors such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0116] The memory 806 may include volatile memory, such as random access memory (RAM). The processor 804 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0117] The memory 806 stores executable code, and the processor 804 executes the executable code to respectively implement the functions of the information processing device 700 in Figure 7, thereby implementing the information processing method in the above embodiment. In other words, the memory 806 stores instructions for executing the information processing method in the above embodiment.

[0118] The communication interface 803 uses a transceiver module such as, but not limited to, a network interface card or a transceiver to implement communication between the processing device 800 and other devices or a communication network.

[0119] It should be understood that the structure illustrated in the embodiment of the present invention does not constitute a specific limitation on the processing device 800. In other embodiments of the present application, the processing device 800 may include more or fewer components than shown, or may combine or separate certain components, or arrange the components differently. The components shown in the figure may be implemented in hardware, software, or a combination of software and hardware.

[0120] Based on the method in the above embodiment, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method in the above embodiment.

[0121] Based on the method in the above embodiment, an embodiment of the present application provides a computer program product. When the computer program product runs on a processor, the processor executes the method in the above embodiment.

[0122] It is understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0123] The method steps in the embodiments of the present application can be implemented by hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, mobile hard disks, CD-ROMs or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0124] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted via the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid state drive (SSD)).

[0125] It will be understood that the various numerical numbers involved in the embodiments of the present application are merely distinctions for the convenience of description and are not intended to limit the scope of the embodiments of the present application.

Claims

1. An information processing method, characterized in that: The method comprises: Acquiring channel state information between at least one network device and at least one terminal device; Processing the channel state information using a graph neural network GNN to obtain precoding between the at least one network device and the at least one terminal device, the GNN comprising a first nonlinear unit and a second nonlinear unit; The first nonlinear unit is used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result, and the target node includes a node connected to the target edge and a node of the same category as the node connected to the target edge; The second nonlinear unit is used to aggregate the edge feature of the target edge and a result obtained by summarizing the at least one correlation result to update the edge feature of the target edge.

2. The method according to claim 1, characterized in that The precoding is digital precoding; The graph data includes: first-type nodes corresponding to network devices, second-type nodes corresponding to terminal devices, and edges connecting the first-type nodes and the second-type nodes.

3. The method according to claim 1, characterized in that The precoding is a mixed digital and analog precoding; The graph data includes: a first type of node corresponding to an antenna of a network device, a second type of node corresponding to an antenna of a terminal device, a third type of node corresponding to a radio frequency link of the network device, and edges connecting the first type of node, the second type of node and the third type of node.

4. The method according to claim 3, characterized in that Initial features of edges in the graph data are obtained based on the channel state information.

5. The method according to any one of claims 1 to 4, characterized in that: The first nonlinear unit and the second nonlinear unit are both implemented by the neural network in the GNN.

6. The method according to any one of claims 1 to 5, characterized in that: The first nonlinear unit and the second nonlinear unit both have n-1-dimensional permutation equivariance, where n is the number of types of nodes in the graph data.

7. An information processing device, characterized in that: include: A processing module, used to obtain channel state information between at least one network device and at least one terminal device, and process the channel state information using a graph neural network GNN to obtain precoding between the at least one network device and the at least one terminal device, wherein the GNN includes a first nonlinear unit and a second nonlinear unit; The first nonlinear unit is used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result, and the target node includes a node connected to the target edge and a node of the same category as the node connected to the target edge; The second nonlinear unit is used to aggregate the edge feature of the target edge and a result obtained by summarizing the at least one correlation result to update the edge feature of the target edge.

8. The device according to claim 7, characterized in that The precoding is digital precoding; The graph data includes: first-type nodes corresponding to network devices, second-type nodes corresponding to terminal devices, and edges connecting the first-type nodes and the second-type nodes.

9. The device according to claim 7, characterized in that The precoding is a mixed digital and analog precoding; The graph data includes: a first type of node corresponding to an antenna of a network device, a second type of node corresponding to an antenna of a terminal device, a third type of node corresponding to a radio frequency link of the network device, and edges connecting the first type of node, the second type of node and the third type of node.

10. The device according to claim 9, characterized in that Initial features of edges in the graph data are obtained based on the channel state information.

11. The device according to any one of claims 7 to 10, characterized in that: The first nonlinear unit and the second nonlinear unit are both implemented by the neural network in the GNN.

12. The device according to any one of claims 7 to 11, characterized in that: The first nonlinear unit and the second nonlinear unit both have n-1-dimensional permutation equivariance, where n is the number of types of nodes in the graph data.

13. An information processing method, characterized in that: include: Acquire graph data obtained based on a wireless communication system; Processing the graph data using a GNN to obtain a processing result, wherein the GNN performs information processing based on a first nonlinear unit and a second nonlinear unit; The first nonlinear unit is used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result, and the target node includes a node connected to the target edge and a node of the same category as the node connected to the target edge; The second nonlinear unit is used to aggregate the edge feature of the target edge and a result obtained by summarizing the at least one correlation result to update the edge feature of the target edge.

14. The method according to claim 13, characterized in that The first nonlinear unit and the second nonlinear unit are both implemented by the neural network in the GNN.

15. The method according to claim 13 or 14, characterized in that The first nonlinear unit and the second nonlinear unit both have n-1-dimensional permutation equivariance, where n is the number of types of nodes in the graph data.

16. An information processing device, characterized in that: include: A processing module, used to obtain graph data obtained based on a wireless communication system, and process the graph data using a GNN to obtain a processing result, wherein the GNN performs information processing based on a first nonlinear unit and a second nonlinear unit; The first nonlinear unit is used to calculate the correlation between the edge features of the target edge in the graph data used by the GNN and the edge features of each edge on each target node to obtain at least one correlation result, and the target node includes a node connected to the target edge and a node of the same category as the node connected to the target edge; The second nonlinear unit is used to summarize the edge feature of the target edge and the at least one correlation result to obtain The results are aggregated to update the edge feature of the target edge.

17. The device according to claim 16, characterized in that The first nonlinear unit and the second nonlinear unit are both implemented by the neural network in the GNN.

18. The device according to claim 16 or 17, characterized in that The first nonlinear unit and the second nonlinear unit both have n-1-dimensional permutation equivariance, where n is the number of types of nodes in the graph data.

19. A processing device, characterized in that: include: At least one processor is used to execute a computer program so that the method according to any one of claims 1 to 6 is implemented, or the method according to any one of claims 13 to 15 is implemented.

20. The processing device according to claim 19, characterized in that The processing device further comprises at least one memory for storing the computer program.

21. The processing device according to claim 19 or 20, characterized in that The processing device is a chip or a chip system.

22. A computer-readable storage medium storing a computer program, wherein when the computer program is executed on a processor, the processor executes the method according to any one of claims 1 to 6, or executes the method according to any one of claims 13 to 15.

23. A computer program product, characterized in that When the computer program product runs on a processor, the processor is enabled to execute the method according to any one of claims 1 to 6, or to execute the method according to any one of claims 13 to 15.

Citation Information

Patent Citations

  • Method and apparatus for generating graph vector representation of graph data structure

    CN115238134A

  • Channel prediction for MIMO systems based on reciprocity

    WO2023018607A1