A channel prediction method and device

Abstract

The present application provides a channel prediction method and device, which are applied to the field of wireless channel prediction technology, wherein the method includes: constructing a communication environment map centered on each scatterer according to the acquired propagation environment information; detecting the type of scatterers according to the target map neural network classification model; predicting the large-scale parameters of the channel for the fully connected propagation environment map constructed according to all the first target scatterers according to the first target map neural network prediction model; predicting the small-scale parameters of the channel for the single-connected propagation environment map constructed according to the second target scatterer according to the second target map neural network prediction model, and obtaining a second prediction result; combining the first prediction result and the second prediction result to obtain a channel prediction result. The present application uses environmental structure information to perform channel prediction, which alleviates the dependence on the type of propagation scenario, reduces data redundancy, and improves prediction accuracy.

Description

Technical Field

[0001] The present application relates to the technical field of wireless channel prediction, and in particular to a channel prediction method and device. Background Art

[0002] With the widespread commercial use of 5G networks, in order to cope with the growing technology and communication scenarios, research on 6G networks has been gradually launched. In the face of the needs of the future digital society, how to ensure the efficiency and flexibility of 6G networks in a changing environment faces huge challenges. The traditional system design method is to first obtain the channel model of typical scenarios through channel measurement, and then implement network design through offline simulation, which leads to the insufficient adaptability of 6G networks to dynamic environments. Therefore, how to efficiently achieve accurate channel prediction in a variety of environments, including the prediction of large-scale parameters (Large-Scale Parameters, LSP) and small-scale parameters (Small-Scale Parameters, SSP), has become the key to subsequent applications.

[0003] Traditional channel prediction methods use collected historical data to predict future channels, that is, by mining existing data information and evaluating potential channel correlation. However, this correlation is difficult to obtain in different environments, so it is difficult to apply. With the rapid development of machine learning (ML) technology, neural networks have been applied to the design of physical layers and algorithms to achieve more reliable prediction models. Different learning methods can be used to improve accuracy, such as long short-term memory (LSTM), convolutional neural network (CNN) and deep neural network (DNN). In order to obtain higher model accuracy while reducing complexity, methods that introduce additional environmental information have also received widespread attention. For example, the method of actively adding environmental information: using latitude and longitude, altitude, and distance as inputs to the artificial neural network (ANN) model can achieve accurate path loss (PL) prediction in high frequency bands. And the method of passively using the original data of the environment: based on deep learning (DL), satellite images are used to predict LSP; panoramic images and CNN are used to predict channel parameters including delay spread (DS), azimuth angle spread of arrival (ASA), azimuth angle spread of departure (ASD), zenith angle spread of arrival (ZSA), and zenith angle spread of departure (ZSD). However, the above two methods lack the analysis of the propagation environment, resulting in a large amount of redundancy in the environmental data, affecting the prediction efficiency, and unable to meet the future 6G network's requirements for accuracy and real-time performance. Summary of the invention

[0004] The technical purpose to be achieved by the embodiments of the present application is to provide a channel prediction method and device to solve the problem that the traditional channel prediction method lacks analysis of the propagation environment, resulting in a large amount of redundancy in environmental data and low prediction efficiency, and cannot meet the future network's requirements for channel prediction accuracy and real-time performance.

[0005] To solve the above technical problems, the present application provides a channel prediction method, including:

[0006] According to the acquired propagation environment information, a communication environment graph (Scatterer-Centered Communication Environment Graph, SC-CEG for short) centered on each scatterer in the propagation environment information is constructed respectively;

[0007] Perform scatterer type detection on the communication environment graph respectively according to the pre-obtained target graph neural network classification model, and obtain the scatterer type of each scatterer;

[0008] A Full Connected Propagation Environment Graph (FC-PEG) is constructed according to all first target scatterers, and a large-scale parameter prediction of a channel is performed on the FC-PEG according to a pre-obtained first target graph neural network prediction model to obtain a first prediction result, wherein the scatterer type of the first target scatterer includes an effective scatterer (ES).

[0009] A single connected propagation environment graph (SC-PEG) is constructed according to the second target scatterer respectively, and a small-scale parameter prediction of a channel is performed on the single connected propagation environment graph according to a pre-obtained second target graph neural network prediction model to obtain a second prediction result, wherein the scatterer type of the second target scatterer includes a primary scatterer (PS);

[0010] The first prediction result and the second prediction result are combined to obtain a channel prediction result.

[0011] Specifically, in the channel prediction method as described above, the scatterer type of each of the scatterers includes: a first subtype and a second subtype;

[0012] Wherein, the first subtype is a valid scatterer or an invalid scatterer;

[0013] The second subtype is a primary scatterer or a non-primary scatterer.

[0014] Specifically, in the channel prediction method as described above, the fully connected propagation environment graph includes:

[0015] The node feature vectors corresponding to the first target scatterer, the receiver (Receiver, Rx for short) and the transmitter (Transmitter, Tx for short), and the edge set of the edges connecting the first target scatterer, the receiver and the transmitter in pairs;

[0016] Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver;

[0017] The node feature vector corresponding to the transmitter is the position coordinate of the transmitter;

[0018] The node feature vector corresponding to the first target scatterer includes: the midpoint coordinates and size of the first target scatterer and the occlusion feature corresponding to the first target scatterer.

[0019] Preferably, in the channel prediction method as described above, the large-scale parameter prediction of the channel is performed on the fully connected propagation environment graph according to the first target graph neural network prediction model obtained in advance to obtain a first prediction result, including:

[0020] Substituting the fully connected propagation environment graph into each first sub-target prediction model predicted by the neural network of the first target graph respectively, to obtain a first sub-target prediction result, wherein the first sub-target prediction model includes: a path loss model, a delay spread model, a horizontal arrival angle spread model, a horizontal departure angle spread model, a vertical arrival angle spread model, and a vertical departure angle spread model;

[0021] The first prediction result is obtained based on all the first sub-target prediction results.

[0022] Specifically, in the channel prediction method as described above, the single connection propagation environment graph includes:

[0023] Obtain node feature vectors corresponding to the second target scatterer, the receiver, and the transmitter, and edge sets of edges after the second target scatterer is connected to the receiver and the transmitter respectively;

[0024] Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver;

[0025] The node feature vector corresponding to the transmitter is the position coordinate of the transmitter;

[0026] The node feature vector corresponding to the second target scatterer includes: the midpoint coordinates and size of the second target scatterer and the occlusion feature corresponding to the second target scatterer.

[0027] Preferably, in the channel prediction method as described above, the predicting of small-scale parameters of the channel on the single connection propagation environment graph according to the pre-obtained second target graph neural network prediction model to obtain a second prediction result includes:

[0028] Substituting the single connection propagation environment graph into each second sub-target prediction model predicted by the neural network of the second target graph respectively, to obtain a second sub-target prediction result, the second sub-target prediction model including: a power model, a delay model, a horizontal angle of arrival (Azimuth Angle of Arrival, referred to as AOA) model, a horizontal angle of departure (Azimuth Angle of Departure, referred to as AOD) model, a vertical angle of arrival (Zenith Angle of Arrival, referred to as ZOA) model and a vertical angle of departure (Zenith Angle of Departure, referred to as ZOD) model;

[0029] The second prediction result is obtained based on all the second sub-objective prediction results.

[0030] Specifically, the channel prediction method as described above further includes:

[0031] Generate a preset number of environmental samples based on the acquired propagation environment information;

[0032] Generate corresponding reference channel data according to the environment samples, and construct a reference communication environment map with each scatterer in each environment sample as the center;

[0033] According to the reference channel data corresponding to all the environment samples and the reference communication environment map, an environment map-channel data set is formed;

[0034] According to the first training set composed of the environment map-channel data set, the preset graph neural network classification model is trained to obtain the target graph neural network classification model and the classification results corresponding to each of the reference communication environment maps.

[0035] Furthermore, the channel prediction method as described above further includes:

[0036] According to the classification result, a reference fully connected propagation environment map is constructed based on a third target scatterer corresponding to each of the environmental samples, and a reference single connected propagation environment map is constructed based on each fourth target scatterer corresponding to each of the environmental samples, wherein the scatterer type of the third target scatterer includes the effective scatterer, and the scatterer type of the fourth target scatterer includes the main scatterer;

[0037] Training the first preset graph neural network prediction according to a second training set consisting of all the reference fully connected propagation environment graphs to obtain the first target graph neural network prediction;

[0038] According to a third training set consisting of all the environmental samples corresponding to the reference single-connection propagation environment graph, the second preset graph neural network prediction is trained to obtain the second target graph neural network prediction.

[0039] Another embodiment of the present application further provides a control device, including:

[0040] A first processing module is used to construct, according to the acquired propagation environment information, a communication environment graph centered on each scatterer in the propagation environment information;

[0041] A second processing module is used to perform scatterer type detection on the communication environment graph according to a pre-obtained target graph neural network classification model, and obtain the scatterer type of each scatterer;

[0042] A third processing module is used to construct a fully connected propagation environment map according to all the first target scatterers, and predict the large-scale parameters of the channel of the fully connected propagation environment map according to the first target map neural network prediction model obtained in advance to obtain a first prediction result, wherein the scatterer type of the first target scatterer includes a valid scatterer;

[0043] a fourth processing module, configured to construct a single-link propagation environment map according to the second target scatterer, and predict the small-scale parameters of the channel for the single-link propagation environment map according to the pre-obtained second target map neural network prediction model to obtain a second prediction result, wherein the scatterer type of the second target scatterer includes a main scatterer;

[0044] The fifth processing module is used to integrate the first prediction result and the second prediction result to obtain a channel prediction result.

[0045] Yet another embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the channel prediction method described above are implemented.

[0046] Compared with the prior art, the channel prediction method and device provided in the embodiments of the present application have at least the following beneficial effects:

[0047] This application uses the most basic environmental raw data to retain environmental structural information, which can alleviate dependence on propagation scenario types and support specific predictions for specific environments; and starting from the wireless propagation mechanism, only effective environmental information is retained through scatterer detection, effectively reducing data redundancy; by constructing a fully connected propagation environment graph and a single-connection propagation environment graph, environments in different states are represented using nonlinear graph structure data, alleviating the limitations of using linear data to represent the environment; and large-scale parameters and small-scale parameters are predicted based on different graph structure data, which can effectively reduce prediction errors and improve prediction accuracy compared to graph neural network models and artificial neural network models based on environmental raw data. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of the flow of the channel prediction method of the present application;

[0049] Figure 2 This is a schematic diagram of the propagation environment;

[0050] Figure 3 is a schematic diagram of a communication environment diagram centered on a scatterer;

[0051] Figure 4 is a schematic diagram of a fully connected propagation environment graph based on effective scatterers;

[0052] Figure 5 is a schematic diagram of a single-connection propagation environment map based on a main scatterer;

[0053] Figure 6 This is a diagram of the training process of the graph neural network prediction model;

[0054] Figure 7 The ROC curve and area characteristic diagram of effective scatterers and main scatterers;

[0055] Figure 8 It is a statistical diagram of the classification results of effective scatterers and main scatterers;

[0056] Fig. 9 This is a schematic diagram of the structure of the control device of the present application. DETAILED DESCRIPTION

[0057] In order to make the technical problems, technical solutions and advantages to be solved by the present application clearer, OICN2234778- will be described in detail below in conjunction with the accompanying drawings and specific embodiments. In the following description, specific details such as specific configurations and components are provided only to help fully understand the embodiments of the present application. Therefore, it should be clear to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, for clarity and brevity, the description of known functions and structures is omitted.

[0058] It should be understood that the references to "one embodiment" or "an embodiment" throughout the specification mean that the specific features, structures, or characteristics associated with the embodiment are included in at least one embodiment of the present application. Therefore, the references to "in one embodiment" or "in an embodiment" appearing throughout the specification do not necessarily refer to the same embodiment. In addition, these specific features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0059] In the various embodiments of the present application, it should be understood that the size of the serial numbers of the following processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0060] It should be understood that the term "and / or" in this article is only a description of the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0061] In the embodiments provided in the present application, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined according to A. However, it should also be understood that determining B according to A does not mean determining B only according to A, and B can also be determined according to A and / or other information.

[0062] See also Figure 1 , an embodiment of the present application provides a channel prediction method, comprising:

[0063] Step S101: construct a communication environment graph centered on each scatterer in the propagation environment information according to the acquired propagation environment information. Specifically, in the present application, the propagation environment is defined as an environment consisting of a transceiver and effective scatterers participating in the propagation, such as Figure 2 As shown in the figure, the rectangular blocks in the figure represent the scatterer 203, the receiver 201 and the transmitter 202 respectively, and a coordinate system is constructed with the lower left corner of the environment as the origin. The propagation environment information obtained in this step at least includes the scene size, the transmitter position, the receiver position, the scatterer position and the scatterer size; and a communication environment graph centered on each scatterer is constructed, that is, the edges between the central scatterer and the transmitter, the receiver and the remaining scatterers are established. Here, the communication graph is represented as G c (V c ,E c ), where V c represents a node, E c represents the edge. For each node, X is used to represent the node feature vector. For an environment containing n scatterers, Vc = {V t ,V r ,V1,V2,…,V n}, where V t ,V r denote the transmitter and receiver nodes, V1, V2, …, V n Represents a scatterer node. Figure 3 As shown in the figure, for the scatterer node V3 to be detected in the environment sample with 5 scatterers, by constructing V3 and the other scatterer nodes V1, V2, V4, V5 and Tx node V t , Rx node V r The corresponding communication environment graph SC-CEG is constructed by using the edges between them.

[0064] In a specific embodiment, the propagation environment information may be acquired through methods including but not limited to actual measurement, map-assisted acquisition, and the use of computer vision related technologies.

[0065] Step S102, according to the pre-obtained target graph neural network classification model, the communication environment graph is respectively subjected to scatterer type detection, and the scatterer type of each scatterer is obtained. In this step, each communication environment graph obtained as above will be substituted into the pre-obtained target graph neural network classification model for automatic classification, that is, scatterer type detection is performed to obtain the scatterer type of the central scatterer in each communication environment graph. In a specific embodiment, the scatterer type of each scatterer is set to include: a first subtype and a second subtype; wherein the first subtype is a valid scatterer or an invalid scatterer; and the second subtype is a main scatterer or a non-main scatterer. When performing scatterer type detection, the detection of the two subtypes can be performed simultaneously. Specifically, the first subtype and the second subtype can be represented respectively by setting two labels, or by setting two fields or bytes in the same label.

[0066] Step S103, construct a fully connected propagation environment map based on all the first target scatterers, and predict the large-scale parameters of the channel on the fully connected propagation environment map according to the first target map neural network prediction model obtained in advance, and obtain a first prediction result, wherein the scatterer type of the first target scatterer includes a valid scatterer. In this step, the scatterer determined to include a valid scatterer in the scatterer type detection in the previous step is recorded as the first target scatterer, and then a fully connected propagation environment map is constructed based on all the first target scatterers, and the fully connected propagation environment map is substituted into the first target map neural network prediction obtained in advance to predict the large-scale parameters of the channel, and the first prediction result can be obtained.

[0067] Step S104, construct a single connection propagation environment map according to the second target scatterer, and predict the small-scale parameters of the channel for the single connection propagation environment map according to the second target map neural network prediction model obtained in advance, to obtain a second prediction result, and the scatterer type of the second target scatterer includes a main scatterer. In this step, the scatterer determined to include the main scatterer in the scatterer type detection in the previous step is recorded as the second target scatterer, and then a single connection propagation environment map is constructed according to each second target scatterer, and the single connection propagation environment map is substituted into the second target map neural network prediction obtained in advance to predict the small-scale parameters of the channel, and the second prediction result can be obtained.

[0068] Step S105, combining the first prediction result and the second prediction result to obtain a channel prediction result. In this step, by combining the first prediction result about the large-scale parameter and the second prediction result about the small-scale parameter prediction obtained above, a complete channel prediction result can be obtained to achieve channel prediction.

[0069] In summary, this embodiment utilizes the most basic environmental raw data to retain environmental structural information, which can alleviate the dependence on the propagation scenario type and support specific predictions for specific environments; and starting from the wireless propagation mechanism, only effective environmental information is retained through scatterer detection, effectively reducing data redundancy; by constructing a fully connected propagation environment graph and a single-connection propagation environment graph, environments in different states are represented using nonlinear graph structure data, alleviating the limitations of using linear data to represent the environment; and large-scale parameters and small-scale parameters are predicted according to different graph structure data, which can effectively reduce prediction errors and improve prediction accuracy compared to graph neural network models and artificial neural network models based on environmental raw data.

[0070] Specifically, in the channel prediction method as described above, the fully connected propagation environment graph includes:

[0071] Node feature vectors corresponding to the first target scatterer, the receiver, and the transmitter, and edge sets of edges connecting the first target scatterer, the receiver, and the transmitter in pairs;

[0072] Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver;

[0073] The node feature vector corresponding to the transmitter is the position coordinate of the transmitter;

[0074] The node feature vector corresponding to the first target scatterer includes: the midpoint coordinates and size of the first target scatterer and the occlusion feature corresponding to the first target scatterer.

[0075] In a specific embodiment of the present application, the above-mentioned fully connected propagation environment diagram is exemplified, see Figure 4 (Invalid scatterers are not shown). The nodes in the figure include: the first target scatterer (V3, V5), the receiver and the transmitter. For the convenience of representation and subsequent calculation, the fully connected propagation environment graph can be represented by the node set of each node and the edge set between any two nodes, for example: G f (V f ,E f ), node set V f = {V t ,V r ,V1,V2,…,V m}, where m represents the number of the first target scatterers, V1, V2, …, V m represents the node feature vector of the first target scatterer, X t ,X r Represent the node feature vectors corresponding to the transmitter and receiver respectively. Specifically, the node feature vectors of the receiver and transmitter are their position coordinates, i.e., X t =(x t ,y t ,z t ),X r =(x r ,y r ,z r ), since the scatterer also has size features and occlusion features, the node feature vector corresponding to the first target scatterer includes: the midpoint coordinates and size of the first target scatterer and the occlusion features corresponding to the first target scatterer, that is, X i =(x i ,y i ,z i ,l i ,w i ,h i ,b i ), where X i represents the node feature vector of the i-th first target scatterer, l i ,w i ,h i Respectively represent the length, width, and height of the i-th first target scatterer, b i represents the occlusion feature of the i-th first target scatterer, which is specifically the degree of occlusion of the scatterer on the line-of-sight (LOS) path calculated based on the geometric relationship between the first target scatterer and the receiver and transmitter. f ={(V t ,V3),(V r ,V3),(V5,V3),(V t,V5),(V r ,V5)}.

[0076] Preferably, in the channel prediction method as described above, the large-scale parameter prediction of the channel is performed on the fully connected propagation environment graph according to the first target graph neural network prediction model obtained in advance to obtain a first prediction result, including:

[0077] Substituting the fully connected propagation environment graph into each first sub-target prediction model predicted by the neural network of the first target graph respectively, to obtain a first sub-target prediction result, wherein the first sub-target prediction model includes: a path loss model, a delay spread model, a horizontal arrival angle spread model, a horizontal departure angle spread model, a vertical arrival angle spread model, and a vertical departure angle spread model;

[0078] The first prediction result is obtained based on all the first sub-target prediction results.

[0079] In a specific embodiment of the present application, in order to accurately obtain each large-scale parameter, the first target graph neural network prediction is set as a multi-target prediction model, including multiple first sub-target prediction models, each first sub-target prediction model corresponds to a large-scale parameter. After substituting the fully connected propagation environment graph into the first target graph neural network prediction, the fully connected propagation environment graph will be predicted based on each first sub-target prediction model to obtain the first sub-target prediction result, so that all the first sub-target prediction results are integrated to obtain the first prediction result about the large-scale parameter.

[0080] Specifically, in the channel prediction method as described above, the single connection propagation environment graph includes:

[0081] Obtain node feature vectors corresponding to the second target scatterer, the receiver, and the transmitter, and edge sets of edges after the second target scatterer is connected to the receiver and the transmitter respectively;

[0082] Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver;

[0083] The node feature vector corresponding to the transmitter is the position coordinate of the transmitter;

[0084] The node feature vector corresponding to the second target scatterer includes: the midpoint coordinates and size of the second target scatterer and the occlusion feature corresponding to the second target scatterer.

[0085] In a specific embodiment of the present application, the above-mentioned single connection propagation environment diagram is exemplified, see Figure 5, the nodes in the graph include: the second target scatterer, the receiver and the transmitter. For the convenience of representation and subsequent calculation, the single-connection propagation environment graph can be represented by the node set of each node and the edge set of the second target scatterer connected to the receiver and the transmitter, for example: G f (V f ,E f ), node set V f = {V t ,V r ,V p}, where X p represents the node feature vector of the second target scatterer, X t ,X r Represent the node feature vectors corresponding to the transmitter and receiver respectively. Specifically, the node feature vectors of the receiver and transmitter are their position coordinates, i.e., X t =(x t ,y t ,z t ),X r =(x r ,y r ,z r ), since the scatterer also has size features and occlusion features, the node feature vector corresponding to the second target scatterer includes: the midpoint coordinates and size of the second target scatterer and the occlusion features corresponding to the second target scatterer, that is, X p =(x p ,y p ,z p ,l p ,w p ,h p ,b p ), where X p represents the node feature vector of the second target scatterer, l p ,w p ,h p Respectively represent the length, width and height of the second target scatterer, b p represents the occlusion feature of the second target scatterer, specifically the degree of occlusion of the scatterer on the line-of-sight path calculated based on the geometric relationship between the second target scatterer and the receiver and transmitter. f ={(V t ,V p ),(V r ,V p )}.

[0086] Preferably, in the channel prediction method as described above, the predicting of small-scale parameters of the channel on the single connection propagation environment graph according to the pre-obtained second target graph neural network prediction model to obtain a second prediction result includes:

[0087] Substituting the single connection propagation environment graph into each second sub-target prediction model predicted by the second target graph neural network respectively, to obtain a second sub-target prediction result, wherein the second sub-target prediction model includes: a power model, a delay model, a horizontal angle of arrival model, a horizontal angle of departure model, a vertical angle of arrival model, and a vertical angle of departure model;

[0088] The second prediction result is obtained based on all the second sub-objective prediction results.

[0089] In a specific embodiment of the present application, in order to accurately obtain each small-scale parameter, the second target graph neural network prediction is set as a multi-target prediction model, including multiple second sub-target prediction models, each second sub-target prediction model corresponds to a small-scale parameter. After the single-connection propagation environment graph is substituted into the second target graph neural network prediction, the single-connection propagation environment graph is predicted based on each second sub-target prediction model to obtain the second sub-target prediction result, so that all the second sub-target prediction results are integrated to obtain the second prediction result about the small-scale parameter.

[0090] Specifically, the channel prediction method as described above further includes:

[0091] Generate a preset number of environmental samples based on the acquired propagation environment information;

[0092] Generate corresponding reference channel data according to the environment samples, and construct a reference communication environment map with each scatterer in each environment sample as the center;

[0093] According to the reference channel data corresponding to all the environment samples and the reference communication environment map, an environment map-channel data set is formed;

[0094] According to the first training set composed of the environment map-channel data set, the preset graph neural network classification model is trained to obtain the target graph neural network classification model and the classification results corresponding to each of the reference communication environment maps.

[0095] In a specific embodiment of the present application, an example is given of how to obtain a target graph neural network classification model, wherein a preset number of environmental samples of the same size as the scene in the propagation environment information is first generated based on the acquired propagation environment information, wherein the number, position and size of the scatterers in each sample are different. In a preferred embodiment, the size (length, width, height) of the scatterers is randomly generated within a corresponding preset range. For ease of calculation, etc., the scatterers are preferably rectangular structures, and there is no overlap between the scatterers. Three-dimensional graphics and image software such as Blender software can be used for generation. Specifically, Figure 2In a specific embodiment, 4566 training data and 842 test data are generated for an effective propagation environment of 15m×10m×3m, wherein the training data set is an environment sample with 3, 5, 6, 7, 9, 10, and 11 random scatterers, and the test data is an environment sample with 4, 8, and 12 random scatterers.

[0096] Then, the generated environment samples are input into the channel simulation platform one by one, and the same communication parameter configuration, receiver and transmitter positions are set for the satellite spectrum to simulate and generate corresponding channel data, wherein the simulation platform preferably uses the Wireless Insite algorithm based on ray tracing (RT) to generate channel data. At the same time, based on the above-mentioned method of constructing a communication environment map with each scatterer as the center, a reference communication environment map can be constructed with each scatterer in each environment sample as the center; then, based on the generated channel data and the constructed reference communication environment map, an environment map-channel data set can be constructed; then, according to the first training set composed of the environment map-channel data set, the preset graph neural network classification model is trained to obtain the target graph neural network classification model and the classification results corresponding to each reference communication environment map. In a specific embodiment, the environment map-channel data set is used as the first training set, and the preset graph neural network classification model is trained to obtain the target graph neural network classification model and the classification results corresponding to each reference communication environment map, including: sequentially substituting the reference communication environment map in the environment map-channel data set into the preset graph neural network classification model to obtain a first pre-classification result, and the first pre-classification result includes the scatterer type of each scatterer. Then, according to the reference channel data corresponding to the reference communication environment map, the first pre-classification result is confirmed or corrected to obtain a second pre-classification result, wherein when there is a path corresponding to the scatterer in the channel data, the scatterer is determined to be the effective scatterer, otherwise, it is determined to be an invalid scatterer; when there is a main path corresponding to the scatterer in the channel data, the scatterer is determined to be the main scatterer, otherwise, the scatterer is determined to be a non-main scatterer, and the main path includes a primary reflection path and a primary scattering path. Thus, according to the second pre-classification result, the preset graph neural network classification model is corrected so that the corrected preset graph neural network classification model better meets the classification requirements and improves the classification accuracy; when the training is completed, the corrected preset graph neural network classification model can be determined as the target graph neural network classification model, wherein the classification result corresponding to each reference communication environment map is the second pre-classification result at the last correction.

[0097] It should be noted that the preset graph neural network classification model preferably uses multiple multi-layer perceptrons (Multi-Layer Perceptron, MLP for short), multiple hidden layers are used for each MLP, multiple neurons are set in each layer, and a rectified linear unit (Rectified Linear Unit, ReLU for short) is used as the activation function of the hidden layer, and the classification result is output after the communication environment graph is input. In a specific embodiment, the number of multi-layer perceptrons is 8, the number of hidden layers in each MLP is 3, and the number of neurons in each layer is 256.

[0098] Furthermore, the channel prediction method as described above further includes:

[0099] According to the classification result, a reference fully connected propagation environment map is constructed based on a third target scatterer corresponding to each of the environmental samples, and a reference single connected propagation environment map is constructed based on each fourth target scatterer corresponding to each of the environmental samples, wherein the scatterer type of the third target scatterer includes the effective scatterer, and the scatterer type of the fourth target scatterer includes the main scatterer;

[0100] Training the first preset graph neural network prediction according to a second training set consisting of all the reference fully connected propagation environment graphs to obtain the first target graph neural network prediction;

[0101] According to a third training set consisting of all the environmental samples corresponding to the reference single-connection propagation environment graph, the second preset graph neural network prediction is trained to obtain the second target graph neural network prediction.

[0102] In another specific embodiment of the present application, an example is given of how to obtain a target graph neural network classification model, wherein, based on the above-mentioned step of constructing a fully connected propagation environment map for the first target scatterer, a reference fully connected propagation environment map is constructed for the third target scatterer corresponding to each environmental sample, and a second training set is formed; based on the above-mentioned step of constructing a single-connection propagation environment map for the second target scatterer, a reference single-connection propagation environment map is constructed for each fourth target scatterer corresponding to each environmental sample, and a third training set is formed. The first target graph neural network prediction is trained according to the second training set to obtain the first target graph neural network prediction; the second target graph neural network prediction is trained according to the third training set to obtain the second target graph neural network prediction, wherein the first preset graph neural network prediction and the second preset graph neural network prediction are both multi-target models, and correspond to the large-scale parameters and small-scale parameters of the channel, respectively.

[0103] It should be noted that in order to better utilize the correlation between the large-scale parameters, the embedded features obtained by the preliminary training of multiple sub-prediction models in the first preset graph neural network prediction model will be connected through the concatenation operation and used as additional features in the subsequent training. Figure 6 Similarly, it can also be applied in the training of the second preset graph neural network prediction model.

[0104] In another embodiment of the present application, in order to ensure the accuracy of the target graph neural network classification model, the first target graph neural network prediction model and the second target graph neural network prediction model obtained above, the accuracy of the three models will be evaluated separately.

[0105] The accuracy evaluation of the target graph neural network classification model is preferably evaluated by using the accuracy and the receiver operating characteristic curve (ROC). The accuracy can be expressed as the ratio of the true samples classified as positive samples (True-Positive, TP) to the true samples classified as negative samples (True-Negative, TN), expressed as:

[0106]

[0107] In a specific embodiment, through calculation, the classification accuracies of effective scatterers and main scatterers are 0.9 and 0.92 respectively.

[0108] When evaluating through ROC, the ROC curve and the important area characteristics of the curve are drawn for evaluation. The horizontal axis of the ROC curve is the true positive rate (TPR) and the vertical axis is the false positive rate (FPR). The formulas for the true positive rate and false positive rate are:

[0109]

[0110] Among them, FP is a true negative sample (False-Positive) classified as a positive sample, and FN is a true negative sample (False-Negative) classified as a negative sample. The ROC curve and area characteristics (area) of the classification of effective scatterers and main scatterers are as follows: Figure 7 As shown in the figure, the area characteristics corresponding to the effective scatterer and the main scatterer are 0.9 and 0.91 respectively. The larger the value of the area characteristic, the stronger the classification ability, and 1 represents the strongest.

[0111] In order to intuitively evaluate the accuracy of scatterer classification in each environmental sample, the number of effective scatterers and main scatterers in each environment can also be counted according to the scatterer classification results. Figure 8 As shown, in a specific embodiment, more than 30% of the environments can correctly classify all scatterers. The classification accuracy of the vast majority of test samples is above 70%.

[0112] The accuracy evaluation of the first target graph neural network prediction model and the second target graph neural network prediction model is preferably evaluated by normalized root mean square error (NMSE), where NMSE can be expressed as:

[0113]

[0114] Since the large-scale parameter prediction based on effective scatterers (Effective Scatterer-Based LSPPrediction, referred to as ESLP) and the small-scale parameter prediction based on primary scatterers (Primary Scatterer-Based SSP Prediction, referred to as PSSP) are implemented based on the results of scatterer classification, the ESLP method (ESLP-gt) and PSSP method (PSSP-gt) implemented according to the real ES or PS are evaluated with the ESLP and PSSP methods implemented according to the predicted ES and PS. And the CNN and DNN methods with matrices and vectors as input are used for comparison. The CNN model consists of three convolutional layers, and 64, 32 and 16 2×2 convolutional filters are set for each convolutional layer. For LSP prediction, the input of CNN is a (m+2)×6 matrix, where m is the number of scatterers in the environment, and each scatterer has 6-dimensional basic data, namely, position coordinates and length, width and height. Similarly, for SSP prediction, the input is a 3×7 matrix. The DNN model architecture is the same as GNN. The input of DNN is the sequence form of CNN input matrix. The LSP prediction results based on ESLP are shown in Table 1.

[0115] Table 1

[0116] ESLP-gt ESLP CNN DNN PL 0.0019 0.0022 0.0083 0.0084 DS 0.16 0.76 0.19 0.23 ASA 0.07 0.08 0.43 0.43 ASD 0.06 0.09 0.19 0.21 ZSA 0.32 0.33 0.78 0.96 ZSD 0.23 0.37 0.81 0.92

[0117] From the results, we can see that the ESLP method based on the real ES has a lower NMSE than the CNN and DNN methods. This shows that using a graph to represent the environment can effectively retain the structural information of the environment, thereby improving the prediction accuracy. Although the accuracy of the ESLP based on the predicted ES is lower than that of ESLP-gt, it is still higher than the CNN and DNN models in most cases. Similarly, the LSP prediction results based on ESLP are shown in Table 2.

[0118] Table 2

[0119] PSSP-gt PSSP CNN DNN power 0.0012 0.0024 0.0043 0.0051 delay 0.00012 0.00017 0.00052 0.00058 AOA 0.0012 0.00022 0.0046 0.0039 AOD 0.0048 0.088 0.01 0.1 ZOA 0.00011 0.00016 0.00015 0.00043 ZOD 0.00012 0.00017 0.00018 0.0012

[0120] The results show that, similar to the LSP prediction results, PSSP-gt can achieve a smaller NMSE than the CNN and DNN models. Even the PSSP implemented based on the predicted PS performs better than the model learned from unstructured data.

[0121] In summary, it can be seen that the present application adopts the above-mentioned target graph neural network classification model, the first target graph neural network prediction model and the second target graph neural network prediction model to effectively reduce the prediction error and improve the prediction accuracy.

[0122] See also Fig. 9 Another embodiment of the present application further provides a control device, comprising:

[0123] The first processing module 901 is used to construct a communication environment map centered on each scatterer in the propagation environment information according to the acquired propagation environment information;

[0124] The second processing module 902 is used to perform scatterer type detection on the communication environment graph according to the pre-obtained target graph neural network classification model, and obtain the scatterer type of each scatterer;

[0125] The third processing module 903 is used to construct a fully connected propagation environment map according to all the first target scatterers, and predict the large-scale parameters of the channel on the fully connected propagation environment map according to the first target map neural network prediction model obtained in advance to obtain a first prediction result, wherein the scatterer type of the first target scatterer includes a valid scatterer;

[0126] The fourth processing module 904 is used to construct a single-link propagation environment map according to the second target scatterer, and predict the small-scale parameters of the channel for the single-link propagation environment map according to the pre-obtained second target map neural network prediction model to obtain a second prediction result, wherein the scatterer type of the second target scatterer includes a main scatterer;

[0127] The fifth processing module 905 is used to integrate the first prediction result and the second prediction result to obtain a channel prediction result.

[0128] Specifically, in the control device as described above, the scatterer type of each of the scatterers includes: a first subtype and a second subtype;

[0129] Wherein, the first subtype is a valid scatterer or an invalid scatterer;

[0130] The second subtype is a primary scatterer or a non-primary scatterer.

[0131] Specifically, in the control device as described above, the fully connected propagation environment graph includes:

[0132] Node feature vectors corresponding to the first target scatterer, the receiver, and the transmitter, and edge sets of edges connecting the first target scatterer, the receiver, and the transmitter in pairs;

[0133] Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver;

[0134] The node feature vector corresponding to the transmitter is the position coordinate of the transmitter;

[0135] The node feature vector corresponding to the first target scatterer includes: the midpoint coordinates and size of the first target scatterer and the occlusion feature corresponding to the first target scatterer.

[0136] Preferably, in the control device as described above, the third processing module comprises:

[0137] The first processing unit is used to substitute the fully connected propagation environment graph into each first sub-target prediction model predicted by the first target graph neural network, and obtain a first sub-target prediction result, wherein the first sub-target prediction model includes: a path loss model, a delay spread model, a horizontal arrival angle spread model, a horizontal departure angle spread model, a vertical arrival angle spread model, and a vertical departure angle spread model;

[0138] The second processing unit is used to obtain the first prediction result according to all the first sub-target prediction results.

[0139] Specifically, in the control device as described above, the single connection propagation environment graph includes:

[0140] Obtain node feature vectors corresponding to the second target scatterer, the receiver, and the transmitter, and edge sets of edges after the second target scatterer is connected to the receiver and the transmitter respectively;

[0141] Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver;

[0142] The node feature vector corresponding to the transmitter is the position coordinate of the transmitter;

[0143] The node feature vector corresponding to the second target scatterer includes: the midpoint coordinates and size of the second target scatterer and the occlusion feature corresponding to the second target scatterer.

[0144] Preferably, in the control device as described above, the fourth processing module comprises:

[0145] A third processing unit is used to substitute the single connection propagation environment graph into each second sub-target prediction model predicted by the second target graph neural network, to obtain a second sub-target prediction result, wherein the second sub-target prediction model includes: a power model, a delay model, a horizontal arrival angle model, a horizontal departure angle model, a vertical arrival angle model, and a vertical departure angle model;

[0146] The fourth processing unit is used to obtain the second prediction result according to all the second sub-target prediction results.

[0147] Specifically, the control device as described above further includes:

[0148] A fifth processing module, used to generate a preset number of environment samples according to the acquired propagation environment information;

[0149] A sixth processing module, configured to generate corresponding reference channel data according to the environmental samples, and construct a reference communication environment map with each of the scatterers in each of the environmental samples as the center;

[0150] A seventh processing module, configured to form an environment map-channel data set according to the reference channel data corresponding to all the environment samples and the reference communication environment map;

[0151] The eighth processing module is used to train the preset graph neural network classification model according to the first training set composed of the environment map-channel data set to obtain the target graph neural network classification model and the classification results corresponding to each of the reference communication environment maps.

[0152] Furthermore, the control device as described above further includes:

[0153] a ninth processing module, configured to construct, according to the classification result, a reference fully connected propagation environment map based on a third target scatterer corresponding to each of the environmental samples, and to construct a reference single connected propagation environment map based on each fourth target scatterer corresponding to each of the environmental samples, wherein the scatterer type of the third target scatterer includes the effective scatterer, and the scatterer type of the fourth target scatterer includes the main scatterer;

[0154] a tenth processing module, configured to train the first preset graph neural network prediction according to a second training set consisting of all the reference fully connected propagation environment graphs to obtain the first target graph neural network prediction;

[0155] An eleventh processing module is used to train the second preset graph neural network prediction according to a third training set consisting of all the environmental samples corresponding to the reference single-connection propagation environment graph, so as to obtain the second target graph neural network prediction.

[0156] The device embodiment of the present application is a device corresponding to the embodiment of the above-mentioned channel prediction method. All implementation means in the above-mentioned method embodiment are applicable to the embodiment of the device and can achieve the same technical effect.

[0157] Yet another embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the channel prediction method described above are implemented.

[0158] In addition, the present application may repeat reference numerals and / or letters in different examples. This repetition is for the purpose of simplicity and clarity, and does not in itself indicate the relationship between the various embodiments and / or settings discussed.

[0159] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusions.

[0160] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A channel prediction method, It is characterized in that include: According to the acquired propagation environment information, a communication environment graph centered on each scatterer in the propagation environment information is constructed respectively; Perform scatterer type detection on the communication environment graph respectively according to the pre-obtained target graph neural network classification model, and obtain the scatterer type of each scatterer; Constructing a fully connected propagation environment graph according to all first target scatterers, and predicting large-scale parameters of channels on the fully connected propagation environment graph according to a pre-obtained first target graph neural network prediction model to obtain a first prediction result, wherein the scatterer type of the first target scatterer includes a valid scatterer; Constructing a single connection propagation environment map according to the second target scatterer respectively, and predicting the small-scale parameters of the channel for the single connection propagation environment map according to the pre-obtained second target map neural network prediction model to obtain a second prediction result, wherein the scatterer type of the second target scatterer includes a main scatterer; Combining the first prediction result and the second prediction result to obtain a channel prediction result; The fully connected propagation environment graph includes: Node feature vectors corresponding to the first target scatterer, the receiver, and the transmitter, and edge sets of edges connecting the first target scatterer, the receiver, and the transmitter in pairs; Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver; The node feature vector corresponding to the transmitter is the position coordinate of the transmitter; The node feature vector corresponding to the first target scatterer includes: the midpoint coordinates and size of the first target scatterer and the occlusion feature corresponding to the first target scatterer; The single connection propagation environment graph includes: The node feature vectors corresponding to the second target scatterer, the receiver and the transmitter, and the edges connecting the second target scatterer with the receiver and the transmitter respectively; Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver; The node feature vector corresponding to the transmitter is the position coordinate of the transmitter; The node feature vector corresponding to the second target scatterer includes: the midpoint coordinates and size of the second target scatterer and the occlusion feature corresponding to the second target scatterer; The step of predicting the large-scale parameters of the channel on the fully connected propagation environment graph according to the first target graph neural network prediction model obtained in advance to obtain a first prediction result includes: Substituting the fully connected propagation environment graph into each first sub-target prediction model predicted by the neural network of the first target graph respectively, to obtain a first sub-target prediction result, wherein the first sub-target prediction model includes: a path loss model, a delay spread model, a horizontal arrival angle spread model, a horizontal departure angle spread model, a vertical arrival angle spread model, and a vertical departure angle spread model; Obtaining the first prediction result according to all the first sub-goal prediction results; The predicting of small-scale parameters of the channel on the single connection propagation environment graph according to the pre-obtained second target graph neural network prediction model to obtain a second prediction result includes: Substituting the single connection propagation environment graph into each second sub-target prediction model predicted by the second target graph neural network respectively, to obtain a second sub-target prediction result, wherein the second sub-target prediction model includes: a power model, a delay model, a horizontal angle of arrival model, a horizontal angle of departure model, a vertical angle of arrival model, and a vertical angle of departure model; The second prediction result is obtained based on all the second sub-objective prediction results.

2. The channel prediction method according to claim 1, It is characterized in that The scatterer type of each of the scatterers includes: a first subtype and a second subtype; Wherein, the first subtype is a valid scatterer or an invalid scatterer; The second subtype is a primary scatterer or a non-primary scatterer.

3. The channel prediction method according to claim 1, It is characterized in that Also includes: Generate a preset number of environmental samples based on the acquired propagation environment information; Generate corresponding reference channel data according to the environment samples, and construct a reference communication environment map with each scatterer in each environment sample as the center; According to the reference channel data corresponding to all the environment samples and the reference communication environment map, an environment map-channel data set is formed; According to the first training set composed of the environment map-channel data set, the preset graph neural network classification model is trained to obtain the target graph neural network classification model and the classification results corresponding to each of the reference communication environment maps.

4. The channel prediction method according to claim 3, It is characterized in that Also includes: According to the classification result, a reference fully connected propagation environment map is constructed based on a third target scatterer corresponding to each of the environmental samples, and a reference single connected propagation environment map is constructed based on each fourth target scatterer corresponding to each of the environmental samples, wherein the scatterer type of the third target scatterer includes the effective scatterer, and the scatterer type of the fourth target scatterer includes the main scatterer; According to a second training set consisting of all the reference fully connected propagation environment graphs, training the first preset graph neural network prediction to obtain the first target graph neural network prediction model; According to a third training set consisting of all the environmental samples corresponding to the reference single-connection propagation environment graph, the second preset graph neural network prediction is trained to obtain the second target graph neural network prediction model.

5. A control device, It is characterized in that include: A first processing module is used to construct, according to the acquired propagation environment information, a communication environment graph centered on each scatterer in the propagation environment information; A second processing module is used to perform scatterer type detection on the communication environment graph according to a pre-obtained target graph neural network classification model, and obtain the scatterer type of each scatterer; A third processing module is used to construct a fully connected propagation environment map according to all the first target scatterers, and predict the large-scale parameters of the channel of the fully connected propagation environment map according to the first target map neural network prediction model obtained in advance to obtain a first prediction result, wherein the scatterer type of the first target scatterer includes a valid scatterer; a fourth processing module, configured to construct a single connection propagation environment map according to the second target scatterers respectively, and predict the small-scale parameters of the channel for the single connection propagation environment map according to the pre-obtained second target map neural network prediction model to obtain a second prediction result, wherein the scatterer type of the second target scatterer includes a main scatterer; A fifth processing module, configured to combine the first prediction result and the second prediction result to obtain a channel prediction result; The fully connected propagation environment graph includes: Node feature vectors corresponding to the first target scatterer, the receiver, and the transmitter, and edge sets of edges connecting the first target scatterer, the receiver, and the transmitter in pairs; Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver; The node feature vector corresponding to the transmitter is the position coordinate of the transmitter; The node feature vector corresponding to the first target scatterer includes: the midpoint coordinates and size of the first target scatterer and the occlusion feature corresponding to the first target scatterer; The single connection propagation environment graph includes: The node feature vectors corresponding to the second target scatterer, the receiver and the transmitter, and the edges connecting the second target scatterer with the receiver and the transmitter respectively; Wherein, the node feature vector corresponding to the receiver is the position coordinate of the receiver; The node feature vector corresponding to the transmitter is the position coordinate of the transmitter; The node feature vector corresponding to the second target scatterer includes: the midpoint coordinates and size of the second target scatterer and the occlusion feature corresponding to the second target scatterer; The third processing module comprises: The first processing unit is used to substitute the fully connected propagation environment graph into each first sub-target prediction model predicted by the first target graph neural network, and obtain a first sub-target prediction result, wherein the first sub-target prediction model includes: a path loss model, a delay spread model, a horizontal arrival angle spread model, a horizontal departure angle spread model, a vertical arrival angle spread model, and a vertical departure angle spread model; A second processing unit, configured to obtain the first prediction result according to all the first sub-objective prediction results; The fourth processing module comprises: A third processing unit is used to substitute the single connection propagation environment graph into each second sub-target prediction model predicted by the second target graph neural network, to obtain a second sub-target prediction result, wherein the second sub-target prediction model includes: a power model, a delay model, a horizontal arrival angle model, a horizontal departure angle model, a vertical arrival angle model, and a vertical departure angle model; The fourth processing unit is used to obtain the second prediction result according to all the second sub-target prediction results.

6. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the channel prediction method according to any one of claims 1 to 4 are implemented.

Images