Channel state information prediction method and related equipment

Abstract

The invention provides a channel state information prediction method and related equipment. The channel state information prediction method comprises the following steps: acquiring position information of a transmitting end and a receiving end of a signal and geometric information of a scatterer in an environment where the transmitting end and the receiving end are located; determining electromagnetic environment knowledge from the transmitting end to the receiving end according to the position information and the geometric information, wherein the electromagnetic environment knowledge is used for representing a relationship between channel information and environment information between the transmitting end and the receiving end; coding the electromagnetic environment knowledge into a channel knowledge vector based on a pre-trained scene embedding model; searching an embedded vector of which the similarity with the channel knowledge vector is higher than a threshold value from a pre-constructed channel knowledge base; obtaining one or more channel responses corresponding to the searched embedded vector; and fusing the searched channel responses corresponding to the one or more embedded vectors to obtain a channel state information prediction result of the sending end and the receiving end. According to the method, the service delay is greatly reduced, and the harsh requirements of scenes such as low-delay high-reliability communication and the like are met.

Description

Technical Field

[0001] This disclosure relates to the field of communication technology, and in particular to a method and related equipment for predicting channel state information. Background Technology

[0002] Currently, the deep integration of artificial intelligence and mobile communication systems has enabled numerous new intelligent application scenarios characterized by the Internet of Things, ultra-high reliability, and real-time interaction. To realize this forward-looking vision, there is an urgent need for 6G (6th generation mobile networks) wireless systems with self-sustaining capabilities and active online learning functions. Whether it's channel state information prediction and modulation / coding in small-scale fading channels or network planning and optimization in large-scale fading channels, channel models remain a crucial foundation for system design, theoretical analysis, performance evaluation, system optimization, and deployment. Therefore, scientifically conducting theoretical research on the analysis and modeling of wireless communication network channel characteristics can take into account the future development direction of core wireless communication technologies and is expected to become an important component in building 6G wireless communication systems.

[0003] In the field of artificial intelligence, knowledge bases, as core components for storing, managing, and utilizing knowledge, have long been crucial for achieving machine cognition and reasoning. Early knowledge bases relied on expert systems, representing knowledge through manually defined symbols and rules. While interpretable, these were costly to build and difficult to scale. With the advent of the big data era, data-driven knowledge bases, represented by vector databases and embedding technologies, have developed. However, existing vector knowledge bases struggle to meet the specific modeling needs of wireless channels—high-dimensional, rapidly changing, and strictly constrained by physical laws. Therefore, they cannot currently be used for channel state information prediction. Current deep learning-based channel modeling methods mostly construct a mapping between environmental perception data and channel characteristics. While this approach can learn the complex mapping relationship between the environment and the channel from massive amounts of data, its "black box" nature leads to a lack of physical interpretability in model decisions, making it unsuitable for debugging and trust building in critical tasks. Furthermore, such models typically construct a direct mapping between environmental data and the channel, heavily relying on training data specific to certain scenarios, resulting in limited generalization ability. The complex network structure also leads to high online inference latency, making it difficult to meet the stringent requirements of low-latency, high-reliability communication scenarios. Summary of the Invention

[0004] This disclosure proposes a channel state information prediction method and related equipment to solve or partially solve the above-mentioned problems.

[0005] This disclosure provides a channel state information prediction method, comprising: acquiring location information of a signal transmitter and receiver, and geometric information of scatterers in the environment in which the transmitter and receiver are located; determining electromagnetic environment knowledge from the transmitter to the receiver based on the location information and the geometric information, wherein the electromagnetic environment knowledge is used to characterize the relationship between channel information and environmental information between the transmitter and the receiver; encoding the electromagnetic environment knowledge into a channel knowledge vector based on a pre-trained scene embedding model; searching for embedding vectors with a similarity higher than a threshold to the channel knowledge vectors from a pre-constructed channel knowledge base, wherein the channel knowledge base includes multiple embedding vectors corresponding to multiple different communication scenarios; acquiring one or more channel responses corresponding to the found embedding vectors; and fusing the one or more channel responses to obtain a channel state information prediction result for the transmitter and the receiver.

[0006] A second aspect of this disclosure provides a computer device including one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and executed by the one or more processors, and the one or more programs include instructions for performing the method of the first aspect.

[0007] A third aspect of this disclosure provides a non-volatile computer-readable storage medium comprising a computer program that, when executed by one or more processors, causes the one or more processors to perform the method described in the first aspect.

[0008] This disclosure provides a fourth aspect of a computer program product, including one or more computer programs that, when executed by one or more processors, implement the method described in the first aspect.

[0009] The channel state information prediction method and related device of this disclosure utilize a pre-trained scenario embedding model to encode electromagnetic environment knowledge constructed based on the current communication scenario into a channel knowledge vector. This vector is used as a query vector to search for a matching embedding vector from a pre-constructed channel knowledge base containing multiple embedding vectors corresponding to various communication scenarios. Thus, the channel state information prediction result of the current communication environment can be obtained by fusing the channel response corresponding to the embedding vector. This method completes the most time-consuming process of precise simulation and forward propagation of the embedding model in the offline stage. During online prediction, only one fast embedding calculation, one efficient search, and one lightweight inference operation are required to complete the current channel state information prediction with the help of historical cases, thereby greatly reducing service latency and meeting the stringent requirements of scenarios such as low latency and high reliability communication. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in this disclosure or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A schematic diagram of an exemplary system provided by an embodiment of this disclosure is shown.

[0012] Figure 2 A flowchart illustrating an exemplary channel state information prediction method provided in an embodiment of this disclosure is shown.

[0013] Figures 3A-3C A schematic diagram of an effective scatterer and other scatterers in an exemplary scenario provided by embodiments of this disclosure is shown.

[0014] Figure 4 A schematic diagram of an exemplary directed graph constructed by a signal-based transmitter and receiver, as provided in an embodiment of this disclosure, is shown.

[0015] Figure 5 A schematic diagram illustrating the construction process of an exemplary single-hop path contribution provided in an embodiment of this disclosure is shown.

[0016] Figure 6 This illustration shows an exemplary method for finding a multi-hop propagation path from the transmitter through sampling points to the receiver, as provided in an embodiment of this disclosure.

[0017] Figure 7 A schematic diagram of an exemplary scatterer classification model provided in an embodiment of this disclosure is shown.

[0018] Figures 8A-8C This diagram illustrates a performance comparison between the channel state information prediction method provided in this embodiment and the channel parameter prediction method based on graph neural networks.

[0019] Figure 9 A schematic diagram of an exemplary computer device provided in an embodiment of this disclosure is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0021] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0023] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0024] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0025] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0026] Figure 1 A schematic diagram of an exemplary system provided by an embodiment of this disclosure is shown.

[0027] like Figure 1As shown, system 100 may include terminal device 102, terminal device 104, server 106, and database server 108. A medium (e.g., a network) may be included between terminal device 102, terminal device 104, server 106, and database server 108 to provide a communication link. This network may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0028] Various software or applications (APPs) may be installed on terminal devices 102 and 104, such as image processing software or applications, video conferencing software or applications, reading software or applications, video software or applications, social networking software or applications, payment software or applications, web browsers, and instant messaging tools. In some embodiments, these software or applications can all be used to predict channel state information.

[0029] The terminal devices 102 and 104 here can be hardware or software. When terminal devices 102 and 104 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, e-book readers, MP3 players, laptops, and desktop computers (PCs). When terminal devices 102 and 104 are software, they can be installed in the electronic devices listed above. They can be implemented as multiple software programs or software modules (e.g., to provide distributed services) or as a single software program or software module. No specific limitations are made here.

[0030] Server 106 can be a server that provides various services, such as a backend server that supports various applications displayed on terminal devices 102 and 104. Database server 108 can also be a database server that provides various services. It is understood that if server 106 can implement the relevant functions of database server 108, database server 108 may not need to be set up in system 100.

[0031] The server 106 and database server 108 here can be either hardware or software. When they are hardware, they can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When they are software, they can be implemented as multiple software programs or software modules (for example, to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0032] It should be noted that the channel state information prediction method provided in this embodiment can be executed by terminal device 102 and / or terminal device 104, or interactively executed by the various devices in system 100. It should be understood that... Figure 1The number of terminal devices, users, servers, and database servers shown is merely illustrative. Depending on implementation needs, there can be any number of terminal devices, users, servers, and database servers.

[0033] Figure 2 A flowchart illustrating an exemplary channel state information prediction method provided in an embodiment of this disclosure is shown. This method 200 can be used for channel state information prediction. Optionally, this method 200 can be... Figure 1 Terminal devices 102 and 104 can be implemented separately, or they can be... Figure 1 The server 106 can be used to implement this, or it can be implemented by... Figure 1 The interaction between devices in System 100 is implemented.

[0034] like Figure 2 As shown, the method 200 may further include the following steps.

[0035] In step 202, the location information of the signal transmitter and receiver, as well as the geometric information of the scatterers in the environment where the transmitter and receiver are located, are obtained.

[0036] In one or more embodiments of this disclosure, the location information of the signal transmitter and receiver may include real-time location information of the signal transmitter and receiver, such as the three-dimensional coordinates of the real-time location of the transmitter and receiver. The signal transmitter and receiver may be equipped with several sensing devices, such as cameras and radar, to achieve collaborative perception of the communication environment. The information acquired by these sensing devices may include environmental data of static objects (such as buildings, trees, etc., as an example of the aforementioned scattering body), which may be presented in image or numerical form, including information such as its location, size, and material (as an example of the geometric information of the aforementioned scattering body).

[0037] In communication systems, objects that cause signals to scatter, reflect, or attenuate are called scatterers. In outdoor urban areas, scatterers are densely distributed and vary in shape. Considering the correlation between signal propagation and scatterers, as well as the interrelationship between channel quality and information such as the shape and location of scatterers, embodiments of this disclosure employ a more refined representation of scatterers to better reflect the physical nature of electromagnetic wave propagation; that is, each scatterer is approximated by a set of rectangular planes. Therefore, the scatterer information in a scene can be represented as: ; in, Representing the indivual The first of the scatterer indivual rectangular plane Representing the The set of all rectangular planes of a scatterer. This represents the set of all rectangular planes representing all scatterers in the scene. Each rectangular plane... It can be uniquely determined by the following parametric equation: ; in, Let these be the parameters and coefficients of the above parametric equation. , , and Together, they define the position, size, and shape of the rectangular plane in three-dimensional space. The interaction between electromagnetic waves and scatterers (such as reflection and diffraction) essentially occurs on the physical surfaces and edges of the scatterer. Therefore, representing the scatterer using a set of planes more accurately describes the geometric relationships of the propagation path than using volume abstraction. This parameterized representation, while ensuring controllable computational complexity, significantly improves the modeling accuracy of the electromagnetic response of scatterers with complex shapes and provides a foundation for subsequent channel state information prediction.

[0038] In the method of this disclosure embodiment, scatterers in the communication scenario can be divided into effective scatterers and other scatterers, such as Figures 3A-3C As shown. An effective scatterer refers to the scatterer through which a wireless signal propagates from Tx to Rx via reflection, diffraction, or other means. Figures 3A-3C In the middle, the scattering objects through which the wireless signal does not pass (i.e., Figures 3A-3C The scatterers that the ray originating from Tx has not passed through are other scatterers. The same scatterer may be classified as a different type of scatterer under different Tx-Rx coordinates, which depends on the geometric position between Tx, Rx and the scatterer.

[0039] In step 204, the EEK (Electromagnetic Environment Knowledge) from the transmitter to the receiver is determined based on the position information of the transmitter and receiver and the geometric information of the scatterer.

[0040] In one or more embodiments of this disclosure, EEK is used to characterize the essential relationship between channel information and environmental information, thereby improving the physical interpretability of the mapping relationship between the environment and the channel, and realizing scene-adaptive mapping of environmental information and channel information. Channel information includes, but is not limited to, channel size / scale parameters, channel state information, channel propagation characteristics, communication tasks affected by channel characteristics, and other related parameters. Environmental information is a comprehensive description of a communication scenario. Environmental information can be environmental geometric data obtained by feature extraction from multimodal environmental data sensed by sensing devices at the transmitting and receiving ends, and may include scatterer geometry information and the location information of the transmitting and receiving ends. For example, the electromagnetic environment knowledge from the transmitting end to the receiving end of the signal in embodiments of this disclosure may include: ground reflection contribution, single-hop scatterer surface reflection contribution, single-hop scatterer edge diffraction contribution, and multi-hop path contribution.

[0041] In step 206, electromagnetic environment knowledge is converted into channel knowledge vectors based on a pre-trained scene embedding model.

[0042] For example, given the location information of the transmitter and receiver and the environmental geometry information (an example of the geometry of scatterers between the transmitter and receiver in their respective environments) input during the online testing phase, the pre-trained scene embedding model first calculates the electromagnetic environment knowledge of the current communication scene. Then, it efficiently encodes this knowledge into a low-dimensional, dense channel knowledge vector using the pre-trained scene embedding model. This vector serves as a feature of the current communication scene, representing its core physical characteristics. This process can be formally represented as: ; In step 208, embedding vectors with a similarity higher than a threshold between the embedding vector and the channel knowledge vector are searched from the pre-built channel knowledge base (CKB). The pre-built channel knowledge base includes various embedding vectors corresponding to different communication scenarios. Embedding vectors with a similarity higher than a threshold to the currently searched channel knowledge vector (hereinafter referred to as the query vector, as this vector is used to find embedding vectors from the channel knowledge base) are the most relevant historical cases for the current communication scenario. Optionally, a high-speed similarity comparison can be performed between the query vector and the embedding vectors of all archived communication scenarios in the pre-built channel knowledge base. This can be achieved by calculating the similarity between the query vector and all archived embedding vectors in the channel knowledge base.

[0043] In step 210, one or more channel responses corresponding to the found embedding vector are obtained.

[0044] In step 212, the channel responses corresponding to one or more found embedding vectors are fused to obtain the channel state information prediction results for the transmitter and receiver.

[0045] Optionally, the actual channel responses corresponding to the found embedding vectors can be obtained, and these actual channel responses can be intelligently fused to generate channel state information prediction results. Optionally, the fusion strategy can adopt a weighted average based on similarity weights: ; The weights in the above formula can be calculated using a gating network. A gating network can dynamically calculate a set of weights based on the characteristics of the current scene. The data is assigned to K historical channel responses (i.e., an example of the channel responses corresponding to the embedded vectors found above), indicating the degree of contribution of each historical channel response to the final prediction result. This avoids simple averaging and achieves more refined on-demand allocation.

[0046] The above-mentioned gating network It is a lightweight neural network (which may include 2-3 fully connected layers):

[0047] Wherein, the input x is the interactive representation of the query vector and the found embedding vector: ; The workflow can be formalized as follows: ; in, These are the parameters of the gating network; Softmax ensures... The weights can be interpreted as the probabilities of selecting corresponding channel knowledge. The gating network can be trained end-to-end, co-optimizing with the entire prediction model. The training objective is to minimize the composite loss function: ; in, The main task loss (such as mean square error) is addressed by introducing [a method / mechanism] to resolve the uneven utilization of historical channel responses. To balance load loss, encourage the even use of all historical channel responses: ; Where CV is the coefficient of variation. Indicates the first The load of historical channel responses in batch X.

[0048] The aforementioned lightweight gated network model has very few parameters, ensuring extremely low computational overhead. Compared to ordinary linear mappings, it offers greater flexibility. Furthermore, compared to the end-to-end nonlinear mappings of classic neural networks, this mechanism endows the model with inherent interpretability, allowing for the traceability of the prediction results' origins.

[0049] The channel state information prediction method of this disclosure utilizes a pre-trained scenario embedding model to encode electromagnetic environment knowledge constructed based on the current communication scenario into a channel knowledge vector. This vector is used as a query vector to search for a matching embedding vector from a pre-constructed channel knowledge base containing multiple embedding vectors corresponding to various communication scenarios. Thus, the channel state information prediction result of the current communication environment can be obtained by fusing the channel response corresponding to the embedding vector. This method completes the most time-consuming precise simulation and embedding model forward propagation process in the channel state information prediction process offline. During online prediction, only one fast embedding calculation, one efficient search, and one lightweight inference operation are required to complete the current channel state information prediction with the help of historical cases, thereby greatly reducing service latency and meeting the stringent requirements of scenarios such as low latency and high reliability communication.

[0050] To support low-latency retrieval at both the endpoint and edge, this embodiment of the disclosure can use LanceDB as a vector database to implement the aforementioned channel knowledge base, and establish vector indexes on the embedded columns. When constructing the channel knowledge base, the vector database table structure shown in Table 1 below can be used: Table 1

[0051] To ensure the efficiency of cosine matching, an IVF (Inverted File) vector index can be built on the embedding column, utilizing a mechanism of coarse quantization combined with local fine-grained search to achieve efficient matching. In practice, the vector space can first be divided using the K-means clustering algorithm. Each cluster performs fine-grained calculations only within the most relevant few clusters during queries. The IVF index is continuously updated as the channel knowledge base is developed. Querying historical cases matching the current communication scenario in the channel knowledge base can include the following processing: Training the coarse quantizer: on the training set Run the K-means algorithm to obtain the centroid. (Center of each group). Coarse quantification. .

[0052] Create an inverted index: for each vector in the database Computation cluster ID: And append its ID and the corresponding vector to the inverted index. middle.

[0053] Query phase: For query vectors First find the matching Recent A set of centroids Only in the corresponding inverted index Fine-grained matching is performed.

[0054] Fine-grained matching: Within the selected inverted bucket, cosine / Euclidean distance is directly calculated and local sorting is performed. Then, it is merged with other bucket candidates to obtain the Top-ranked candidate. result.

[0055] In the above process, coarse quantization first divides the high-dimensional space into several regions, ensuring that the query first falls into the most relevant local area. Fine matching is then performed only within these local areas, significantly reducing the number of candidate options that need to be compared and avoiding the performance overhead of matching one by one. Even with millions of data points, matching can be completed in milliseconds. This achieves low-latency retrieval while ensuring high recall.

[0056] During online queries, the candidate cluster set can be determined first based on the distance between the query vector and the centroids of each cluster. Then, the vectors in the corresponding inverted index are read, and in-bucket ranking is performed based on the cosine metric. Finally, all candidate scores are merged and a Top-ranking algorithm is applied. Select, and it will return the ID and raw metadata.

[0057] In one or more embodiments of this disclosure, the channel state information prediction method may further include: training a scenario embedding model, specifically including: Physical perturbation is applied to the anchor point sample data in the anchor point scene dataset based on perturbation parameters to generate semantically invariant positive sample data. The perturbation parameters include the position of the scatterer and the reflection contribution value, or the perturbation parameters include the position of the scatterer and the diffraction contribution value.

[0058] Select anchor point sample data from the anchor point scene dataset whose scattering distribution difference is greater than a preset value, and generate negative sample data.

[0059] The anchor sample data, positive sample data, and negative sample data are merged to obtain the training dataset.

[0060] For example, in the anchor point scenario By applying tiny physical perturbations to the properties of the scatterer, semantically invariant positive samples are obtained. This forces the model to learn essential scene features that are not robust to noise. The perturbation amplitude can be strictly constrained by physics, and the Rx parameter remains unchanged. The expression for applying physical perturbation to the scatterer properties is as follows:

[0061] in, Indicates truncation to interval , and Control the intensity of the disturbance. and This represents the upper limit of the disturbance amplitude.

[0062] Negative samples can be selected from the entire scene based on the largest differences in scattering distribution. Let the set of scattering points between the two scenes be denoted as . and The difference in scattering distribution is defined by adding the difference in variance to the distance between the centroids. : ; in, and For the mean and variance, The relative weights of the centroid and the distribution dispersion are weighed. From The largest Negative samples are obtained by sampling from the candidates.

[0063] Optionally, the scene embedding model in this embodiment of the disclosure can be designed based on the RAG (Retrieval-Augmented Generation) principle. RAG requires the embedding model to distinguish different environments and cluster minor perturbations within the same environment together. Contrastive learning uses each data point in the dataset as an anchor point, shaping the embedding space geometry by maximizing the similarity between "anchor point and positive samples" and minimizing the similarity between "anchor point and negative samples," making the nearest neighbor retrieval results more physically intuitive.

[0064] The sample data (including anchor sample data, positive sample data, and negative sample data) in the training dataset are encoded based on the embedding function to obtain the anchor embedding result, the positive sample embedding result, and the negative sample embedding result. The information noise contrast estimation loss is calculated based on the anchor point embedding results, positive sample embedding results, and negative sample embedding results. Backpropagation is performed based on the loss estimated by information noise comparison to update the model parameters, resulting in a scene embedding model.

[0065] For example, let the temperature parameter be... The number of positive samples is The number of negative samples is The anchor point embedding result is The positive sample embedding results are The negative sample embedding result is Similarity If the cosine is used, then the InfoNCE loss for a single sample is: ; The InfoNCE loss function described above has both positive and negative sample terms in the denominator, while the numerator contains only positive sample terms. Training follows a closed loop of "sampling-boosting-encoding-contrast-update". For a small batch, sampling is performed first. Anchor point scene Then generate for each anchor point Positive samples And based on the above distribution metrics, pick negative samples Then, the anchor point samples, positive samples, and negative samples of the B anchor point scenes are encoded using an embedding function to obtain... Calculate the InfoNCE loss and backpropagate to update the parameters. In this process, to facilitate subsequent use of cosine similarity matching, all embedding results are performed before similarity calculation. Normalization is performed, and all similarities are subtracted from the maximum value before entering the exponential operation, and the logarithm of the result is taken to avoid an excessively large loss function affecting subsequent learning.

[0066] The channel state information prediction method of this disclosure applies minute physical perturbations to the scatterer properties in a scene based on physical constraints to construct positive samples. Based on the spatial distribution differences of the scatterers, negative samples are actively selected from the dataset to increase the discriminative power of different scenes in the embedding space. An improved loss function is used to optimize and train the scene embedding model, enabling the model to autonomously learn the essential characteristics of the channel scene without manual annotation. By combining the established correlation between channel knowledge vectors and parameters such as channel response, efficient prediction of channel characteristics can be achieved. Furthermore, by using a scene embedding model to represent different communication scenarios, the communication system only needs to perform lightweight embedding model calculations to predict the channel state information of the current scenario based on historical data, greatly improving the efficiency of the communication system.

[0067] In one or more embodiments of this disclosure, the anchor point sample data described above may include: For each pair of signal transmitters and receivers, the directed graph constructed has a vertex set including transmitter nodes, receiver nodes, and active nodes. The active nodes include ground reflection points, scatterer surface reflection points, and scatterer edge diffraction points. The feature vector of each node in the vertex set includes coordinates (e.g., three-dimensional coordinates), type encoding, and contribution value. The edge set of the directed graph includes directed edges from the transmitter to the receiver and directed edges from the transmitter to each active node. The feature vector of each edge in the edge set includes the Euclidean distance between the two nodes, occlusion status, and contribution value.

[0068] For example, for each pair of signal transmitters and receivers (denoted as Tx-Rx), a directed graph is constructed based on the single-hop path and multi-hop path contribution information calculated in the above process. ,like Figure 4 As shown. Among them, A set of nodes, containing Tx nodes. Rx nodes and the point of action node ( ), can be represented as: The active node represents the reflection or diffraction point in a single hop path, which may include ground reflection points, scatterer surface reflection points, and scatterer edge diffraction points. This represents the total number of these points. The feature vectors of all nodes in the edge set can contain three-dimensional coordinates, type encoding, and contribution values.

[0069] Action Node The feature vector can be represented as: ,in, For nodes The three-dimensional coordinates For nodes The type encoding has possible values ​​of 0, 1, and 2, representing ground reflection, scatterer surface reflection, and scatterer edge diffraction, respectively.

[0070] For nodes The contribution value, when When the value is 0, 1, or 2, The values ​​are respectively , and .node The feature vector can be represented as: The type encoding Possible values ​​are 3 and 4, representing Tx and Rx points respectively. Since Tx and Rx points contribute no value, the last dimension of the feature vector is fixed at 0. It's worth noting that the number of active nodes is variable for different Tx-Rx values, therefore the graph... The size of the node set is dynamic.

[0071] Let be the set of edges, containing from point to The directed edge, from Point to each The directed edges, and from each point to A directed edge can be represented as: The feature vectors of all edges can contain the three-dimensional coordinates of the Euclidean distance between two points, the occlusion status, and the contribution value. (Edge) The feature vector can be represented as: ,in, The Euclidean distance between Tx and Rx This refers to the occlusion situation between Tx and Rx. This represents the contribution value of the multi-hop path between Tx and Rx. Among these, occlusion conditions are considered. Possible values ​​are 0 and 1, where 0 represents that the direct path between Tx and Rx is not blocked, and 1 represents that the direct path between Tx and Rx is blocked. (Edge) and The feature vector can be represented as: ,in, Let Tx / Rx be the Euclidean distance between the point of application and the point of application. Since Tx / Rx is always visible to the point of application and there are no multi-hop path contributions, the last two dimensions of the eigenvector are fixed at 0.

[0072] In one or more embodiments of this disclosure, encoding sample data in the training dataset based on an embedding function may include: The feature vectors of the transmitting and receiving nodes of the signal in the training dataset are transformed into the first dimension through a fully connected neural network. Based on the feature vectors of the sending node, the receiving node, the action node n, the directed edge between the sending node and the action node n, and the directed edge between the action node n and the receiving node in the training dataset, determine the feature vectors of the direct path and the scattering path between the sending node and the receiving node. The feature vector of the direct trajectory includes: [ ]; in, The coordinates of the sending node are... The coordinates of the receiving node are... = The distance between the sending node and the receiving node is the Euclidean distance. This refers to the occlusion situation between the sending node and the receiving node. The contribution value of the multi-hop path between the sending node and the receiving node; ; ; In the above formula, N is the number of active nodes in the training dataset. Let n be the distance from the active node n to the sending node. The distance from the active node n to the receiving node; The eigenvectors of the scattering path include: ; In the above formula, Indicates the type of the active node n. Let n be the coordinates of node n. For one-hot functions; The feature vector of the direct trajectory is transformed to the first dimension using a fully connected neural network; The feature vectors of the sending node and the receiving node after dimensionality transformation, as well as the direct path feature vector, are merged to obtain the joint feature vector. The joint feature vector is transformed to the second dimension through a fully connected neural network to obtain the first semantics; Each scattering path vector is transformed to the first dimension using MLP to obtain the scattering path matrix. The scattering path matrix is ​​then subjected to mean pooling and transformed to the second dimension to obtain the second semantic. The third semantics are obtained by multi-head attention aggregation of the scattering path matrix; The first semantic, second semantic, and third semantic are concatenated to obtain a concatenated vector. The concatenated vector is then transformed to the third dimension, and the concatenated vector in the third dimension is normalized to obtain the embedding vector.

[0073] For example, the embedding model is used to embed the above graph Encoded as a fixed-length vector with the unit norm:

[0074] In the above formula, Embed vectors for the final scene, and ; For the reason The encoding function for the fixed-length vector obtained by encoding (an example of the embedding function mentioned above); A fixed output dimension is used for cosine similarity retrieval. The model ensures accuracy in... While remaining insensitive to arrangement, it covers key physical factors. The following example illustrates the training process of a scene embedding model.

[0075] First, collect the sending information. Taking the sending end Tx as an example, the three-dimensional coordinates of Tx can be used as its feature vector. By using a fully connected neural network, it can be mapped from 3 dimensions to 32 dimensions (an example of the first dimension mentioned above).

[0076] Since the energy relationship between the direct and indirect paths has a significant impact on the final signal in real-world scenarios, path weighting and contribution weighting methods can be used to capture the internal relationship between the direct and indirect paths, as shown below: ; Using this as one of the feature vector elements of the direct trajectory, we can obtain the following direct trajectory feature vector:

[0077] By mapping using a fully connected neural network, it can be mapped from 8 dimensions to 32 dimensions. Therefore, the encoding of the joint direct path at the transmitting and receiving ends can be represented as: Then, a fully connected neural network is used for mapping, which... Mapping from 96 dimensions to 192 dimensions (an example of the second dimension mentioned above), i.e. .

[0078] Scattering path encoding may include the following processing: Each The path encoding consists of the following components: ; because Since these are enumerated values, and the different points of application can only be ground reflection, scatterer surface reflection, or scatterer edge diffraction, a one-hot function can be used for processing. Each scattering path vector is mapped to a 32-dimensional vector using an MLP. Therefore, the scattering path encoding is as follows: ; ; In practical work, due to the number of intermediate nodes Uncertain, so we can proceed with... Perform mean pooling mapping as Then, it is mapped to a fully connected neural network. ,Right now .

[0079] Multi-head attention aggregation in scene embedding models may include the following processing: set up Construct a global scattering path encoding matrix before dimension transformation. : ; in, This is a global query token. This indicates LayerNorm normalization. Let be a linear transformation matrix, Mapped to Sixteen attention heads can be used, each with a dimension of 24. ( () is the attention head index. , , These are the projections of the global query, path key, and path value onto a 24-dimensional matrix.

[0080] Under this premise, the first The head's Q / K / V and attention values ​​are calculated as follows: ; in, For the first Header query / key / value; To scale dot product similarity, it can be done along the path dimension. Scaling to stabilize gradients These are the attention weights after the softmax activation function; This is the weighted aggregation result for this head. The heads are concatenated and linearly mapped to the aggregation dimension, then the aggregated representation is obtained through a fully connected neural network:

[0081] in, The splicing result is the output of 16 heads; For linear mapping weights; For activation functions; This forms a two-layer feedforward network.

[0082] Finally, the three semantic paths are concatenated and projected onto a fixed dimension. (As an example of the third dimension mentioned above), then unit norm is applied to serve cosine retrieval: ; in, It is a concatenated vector of three semantics; for Normalization function; , These are the two layers of weights projected into the final fully connected neural network.

[0083] In one or more embodiments of this disclosure, searching for multiple embedding vectors from a pre-built channel knowledge base that have a similarity higher than a threshold with channel knowledge vectors may include: The K nearest neighbor samples with the highest similarity to the low-dimensional channel knowledge vector are determined by Top-K nearest neighbor retrieval; The embedding vectors of K nearest neighbor samples are divided into multiple clusters based on the clustering algorithm. A cluster identifier is calculated for each embedding vector, and the cluster identifier and the corresponding embedding vector are appended to the inverted index. Find the clusters corresponding to the centroids that are closest to the channel knowledge vector in the inverted index to obtain the candidate cluster set; Query multiple embedding vectors from the candidate cluster set whose similarity to the channel knowledge vector is higher than a threshold.

[0084] The goal of knowledge retrieval is to efficiently and accurately find the most relevant historical cases to the current query scenario from a vast channel knowledge base. The system performs a high-speed similarity comparison between the query vector and the embedding vectors of all archived scenarios in the pre-built channel knowledge base. Its core is calculating the query vector... With all archived vectors in the channel knowledge base The similarity between them.

[0085] For example, Top-K nearest neighbor retrieval returns the K nearest neighbor samples with the highest similarity. The goal of the retrieval is to select the Top-K nearest neighbor samples that are most similar to the current query vector, as shown in the following formula: ; An ANN (Approximate Nearest Neighbor) algorithm, such as IVF (Inverted File Index), can be used to divide the vector space into multiple units through clustering, searching only the units closest to the query vector. To optimize retrieval efficiency, an IVF index structure can be used to accelerate the retrieval process and meet the low latency requirements of online prediction. For a dataset containing the top K 1024-dimensional vectors... ,in It was divided into groups using K-Means clustering (an example of the clustering algorithm described above). There are n clusters (i.e., nlist). The clustering process aims to minimize the following objective function: ; in, Indicates the j-th cluster. It is the centroid of the cluster, through Calculated. Quantizer Map each vector to the index of its nearest centroid: ; For query vectors The bucket-finding phase involves finding the closest one. Clusters corresponding to each centroid:

[0086] After determining the set of clusters to be searched Then, the algorithm performs a fine-grained search on the vectors contained in these clusters. Query vector The distance between any vector y can be measured using Euclidean distance (L2 distance) or cosine similarity, etc.

[0087] Euclidean distance (L2 distance): ; Cosine similarity:

[0088] The IVF algorithm ultimately returns the top K vectors in the target cluster that have the smallest distance (or the highest similarity) to the query vector. Its complexity is as follows: ; in, For the size of the knowledge base, This represents the number of searches.

[0089] The channel state information prediction method of this disclosure directly derives its prediction results from physically verified real-world cases in the channel knowledge base, rather than abstract outputs from a "black box" model. This ensures the traceability of the prediction results, thereby enhancing their reliability and debuggability. Furthermore, by continuously collecting new scenarios and their corresponding precise channel responses (obtained through measurement or simulation) as new "cases" to supplement the channel knowledge base, the system can continuously expand its knowledge boundaries and adaptively improve its prediction performance for future unknown scenarios without requiring complex retraining of the entire model.

[0090] In one or more embodiments of this disclosure, due to the limitations of the scatterer height and the obstruction caused by the scatterer layout, the propagation process of electromagnetic waves can be considered as the signal originating from the transmitter (Tx), passing through the scatterer facade or edge, and arriving at the receiver (Rx). This process involves three electromagnetic wave propagation modes: (1) when Rx is within the line-of-sight propagation range of Tx in free space, it propagates directly to Rx; (2) it propagates to Rx via reflection from the scatterer surface or ground; and (3) it propagates to Rx via diffraction from the scatterer edge. Therefore, the channel impulse response can be expressed as: ; in, , and Let represent the channel impulse responses for direct, reflected, and diffracted light, respectively. All channel impulse responses can be expressed as:

[0091] in, It is the type of transmission. It is by Channel multipath number generated by type propagation , and These represent time-varying path loss, phase, and Doppler shift, respectively.

[0092] In one or more embodiments of this disclosure, determining the electromagnetic environment knowledge from the transmitter to the receiver based on the location information of the transmitter and receiver and the geometric information of the scatterer may include: The feature vectors extracted based on the position information of the signal transmitter and receiver and the geometric information of the scatterer are provided to the pre-trained scatterer classification model to obtain effective scatterer prediction results. For example, the geometric information of the scatterer and the coordinates of Tx and Rx can be input into a deep learning-based scatterer classification model to predict the effective scatterer corresponding to each pair of Tx-Rx.

[0093] The effective scatterers for the transmitter and receiver are determined based on the effective scatterer prediction results. Traverse all surfaces of the effective scattering body to determine the occlusion relationship between the transmitter and receiver; Optionally, an algorithm for determining whether a line intersects with a surface can be used to traverse all surfaces of all scatterers to determine the occlusion relationship between Tx and Rx. And save it.

[0094] If there is no obstruction between the transmitter and receiver, determine the ground reflection point and calculate the ground reflection contribution based on the ground reflection point; Furthermore, for unobstructed cases, the possible ground reflection points can be determined based on the light reflection theorem in geometric optics theory, and the corresponding ground reflection contribution can be calculated.

[0095] Determine the single-hop reflection point on the surface of the effective scatterer, and calculate the surface reflection contribution of the single-hop scatterer based on the single-hop reflection point; Optionally, for each pair of Tx-Rx, for the surface of the corresponding effective scatterer, determine the possible single-hop reflection points on the surface of the effective scatterer and calculate the corresponding single-hop reflection contribution.

[0096] Determine the actual edges that exist on each scatterer; Determine the single-hop diffraction point on the actual edge of each scatterer, and calculate the edge diffraction contribution of the single-hop scatterer based on the single-hop diffraction point; Optionally, based on the parametric equations of all scatterer surfaces, a geometric algorithm is used to determine the actual edges existing on each scatterer. For each pair of Tx and Rx, based on Fresnel's principle and uniform diffraction theory, the possible single-hop diffraction points on the edges of the effective scatterer are determined for the actual edges of the corresponding effective scatterer, and the corresponding single-hop diffraction contributions are calculated.

[0097] Determine the multi-hop propagation path from the transmitter through the sampling points to the receiver, and calculate the multi-hop path contribution of the path; A certain number of points on all scattering surfaces are selected as sampling points, and a visibility map is constructed based on the visibility relationships between the sampling points. For each pair of Tx-Rx, in Based on this, a visibility map is constructed according to the visibility relationship between the sampling points of the Tx-Rx pair and its corresponding effective scatterer. And use the constraint graph search algorithm to Perform pathfinding to find all paths that meet the conditions starting from Tx, passing through multiple sampling points, and reach Rx, and calculate the corresponding multi-hop path contribution.

[0098] By combining the contributions from ground reflection, single-hop scatterer surface reflection, single-hop scatterer edge diffraction, and multi-hop path, we obtain electromagnetic environment knowledge from the transmitter to the receiver.

[0099] The obtained ground reflection contribution, single-hop scatterer surface reflection contribution, single-hop scatterer edge diffraction contribution, and multi-hop path contribution are combined to obtain the EEK from a Tx location to a Rx location.

[0100] It should be noted that, according to the electromagnetic wave reflection theorem, when a plane wave is incident on the surface of an ideal medium, the propagation direction of its reflected wave satisfies the geometric relationship that the angle of incidence equals the angle of reflection. Depending on the surface where the reflection occurs, single-hop reflection contributions can be further distinguished into two types of paths with different characteristics: ground reflection contributions and scatterer surface reflection contributions. In wireless channels, the propagation path of electromagnetic waves from the transmitter to the receiver can be classified into single-hop paths and multi-hop paths based on the number of interactions with scatterers. A single-hop path refers to a path in which the electromagnetic wave interacts with only one scatterer (reflection or diffraction) during its journey from the transmitter to the receiver. A multi-hop path refers to a path in which the electromagnetic wave interacts with one or more scatterers multiple times.

[0101] Before constructing EEK, an explicit mapping relationship from the three-dimensional spatial structure to the multipath propagation gain can be established, which can be formalized as follows:

[0102] in, Contributing to location-based dissemination EEK represents the coupling function. This represents the three-dimensional position coordinates of the transmitter / receiver. The above formula accurately and quantitatively characterizes the influence weights of single-hop reflection and scattering paths, multi-hop paths, and the presence of obstructions on the overall radio propagation process between the transmitter and receiver in the current communication scenario. This process constructs an interpretable relationship between the environment and the channel. After quantifying the interpretable relationship through a large amount of environmental information, it can exhibit strong scenario generalization ability.

[0103] In one or more embodiments of this disclosure, traversing all surfaces of the effective scatterer to determine the occlusion relationship between the transmitter and receiver may include: traversing all surfaces of the scatterer and determining whether the line connecting the coordinates of the receiver and the coordinates of the transmitter intersects with all surfaces of the scatterer; if an intersection exists, it is determined that there is occlusion between the transmitter and receiver; otherwise, it is determined that there is no occlusion between the transmitter and receiver.

[0104] In the method of this disclosure embodiment, to finely quantify the impact of the environment on the channel, the contribution of the propagation path is decomposed hierarchically. For the contribution of a single-hop path, Figure 5 A schematic diagram illustrating the construction process of an exemplary single-hop path contribution provided in this disclosure embodiment is shown. Based on the fundamental physical principle of the interaction between electromagnetic waves and scatterers, the single-hop path contribution is further divided into single-hop reflection contribution and single-hop diffraction contribution.

[0105] The algorithm first iterates through all surfaces of all scatterers in the current communication scenario to determine whether the direct path between the transmitter and receiver is blocked. This problem can be further specified as determining whether the direct path between points... and Is the defined line segment related to the parametric equation (where...)? ) defined face Intersecting. For this problem, a numerical approach based on ray casting and Newton's method can be used. This approach aims to transform the problem into a root-finding problem in a parameter domain. Its core idea is: line segments... ( ) and surface The condition for intersection is the existence of a set of parameters. This makes the equation This condition holds true. This condition is equivalent to the vector... and Parallel, meaning their cross product is zero, therefore we obtain a condition regarding... and A system of two nonlinear equations: ; exist and unit square parameter domain Inside, Newton's iterative method is used to solve the system of equations. The root. The initial value for Newton's iteration is chosen as the center point of the parameter domain. Alternatively, a more advanced sampling strategy can be employed to avoid local optima. For the solution obtained through iterative convergence... : Condition 1: Check if it is within the valid parameter range: and .

[0106] Calculate the corresponding line segment parameters .

[0107] Condition 2: Verification Is it located in Interval.

[0108] like If both conditions one and two are satisfied, then the line segment and the surface intersect. If the iteration does not converge or all solutions do not satisfy the constraints, then they are determined to be non-intersecting.

[0109] The above processing procedure ensures accuracy by initiating iterations from multiple different initial points within the parameter domain to find all possible solutions. In practical processing, the ray intersection query function in mature computer graphics libraries (such as CGAL (Computational Geometry Algorithms Library)) can be used to improve the robustness and efficiency of the computation.

[0110] like Figure 5 As shown in process flow B, determining the ground reflection contribution may include the following steps: Determine the coordinates of the sending end Regarding the mirror point on the ground The coordinates of the receiving end and The intersection of the line and the plane containing the ground is determined as the ground reflection point. ; Iterate through all surfaces of all scatterers in the environment where the transmitter and receiver are located, and determine... and Connect or and Is the connection blocked? and Connect or and If the line is blocked, the ground reflection point Does not exist, if and Connect or and If the connecting lines are not obstructed, then the ground reflection point exist.

[0111] For example, if the direct path between the transmitting and receiving ends is not obstructed, the coordinates of the ground reflection point are determined first. Regarding the ground mirror point At this time, the ground reflection point That is and Connecting and The analytical solution for the intersection of the planes is: ; Then, iterate through all faces of all scatterers in the scene again and determine... and Connect or and Whether the connection is blocked is determined in the same way as the method described above for judging whether the direct path between the transmitting and receiving ends is blocked, and will not be repeated here. If and Connect or and If the line is blocked, the ground reflection point It does not exist. If and Connect or and If the line connecting the points is not obstructed by any scattering surface, then the ground reflection point calculated above... Yes, it exists. In the default scenario of this disclosure, there are no scatterers that are not in contact with the ground, that is, no scatterers "floating" in the air. Therefore, if the direct path between the transmitter and receiver is blocked, the ground reflection point... It does not exist.

[0112] Furthermore, for a specific pair of transmitter and receiver coordinates, if the ground reflection point... If it exists, then the corresponding ground reflection contribution It can be represented as: .in, The Fresnel reflection coefficient for vertical polarization can be expressed as: ; in, For refractive index ( (where is the relative permittivity of the ground). Let be the angle of incidence, and the calculation formula is: ; in, This is the normalized normal vector of the ground. The distance coefficient, used to measure the difference in distance between the reflected radius and the direct radius, can be expressed as: .

[0113] In one or more embodiments of this disclosure, determining a single-hop reflection point on the surface of an effective scatterer may include: Iterate through all surfaces of all effective scatterers in the environment where the transmitter and receiver are located. ,Sure Regarding the effective scattering surface mirror point ; Since only the scattering body through which the ray passes is called the effective scattering body, meaning that ray propagation behaviors such as reflection and diffraction only occur on the effective scattering body, the algorithm first only targets the predicted effective scattering body and traverses all surfaces of the effective scattering body.

[0114] Among them, the transmitting end Regarding the effective scattering surface The mirror point can be represented as: ; in, Representation surface any point on, Representation surface The normalized normal vector.

[0115] Judgment point and Do the connected line segments intersect the surface? intersect; If they do not intersect, then there is no single-hop reflection point on the scatterer surface being traversed. The judgment point in the above ground reflection point calculation algorithm and Do the connected line segments intersect the surface? The method of intersection is the same, so it will not be repeated here.

[0116] If they intersect, according to and The parameters of the connected line segments are used to calculate the single-hop reflection point on the scatterer surface. ; Line segment parameters that can be calculated Further calculations yielded the single-hop reflection point on the scattering surface: ; Iterate through all surfaces of all scatterers in the environment where the transmitter and receiver are located, and determine... and Connect or and Is the connection obstructed? like and Connect or and If the connection is blocked, then the single-hop reflection point on the currently traversed scattering surface... Does not exist, if and Connect or and If none of the connecting lines are obstructed, then the single-hop reflection point on the currently traversed scattering surface is... exist; If point and The line does not connect to the surface If they intersect, then there is no single-hop reflection point on the scatterer surface being traversed. The calculated single-hop reflection point Add to collection In this process, the set of single-hop reflection points is obtained; like Figure 5 As shown in Processing Flow A, this disclosure provides an algorithm for calculating the reflection point of a scattering surface based on geometric optics, for the contribution of the scattering surface to reflection.

[0117] Furthermore, for a specific pair of transmitter and receiver coordinates, if a single-hop reflection point on a certain scattering surface... If it exists, then the corresponding single-hop surface reflection contribution It can also be expressed as: The formulas for calculating the Fresnel reflection coefficient and the distance coefficient are both related to the above-mentioned ground reflection contribution. The calculation formula is the same, the difference being that the relative permittivity of the ground is used. Replace with the relative permittivity of each scattering surface , ground reflection point Replace with a single-hop reflection point on the scattering surface. Normalize the ground normal vector Replace with the normalized normal vector of the scattering surface .

[0118] When the propagation path of an electromagnetic wave is blocked by an obstacle, the wavefront will diffract at the edge of the obstacle. In the scatterer modeling based on surface parameter equations, this disclosure uses a set of parameterized rectangular planes to represent complex-shaped scatterers. Based on this, each scatterer is composed of multiple rectangular planes, each uniquely determined by its parameter equation. However, directly extracting the four edges of all faces as the edge set of the scatterer introduces two types of invalid edges: one is the inner edges at the junctions of multiple faces (i.e., edges shared by two or more faces), and the other is edges in contact with the ground (these edges do not contribute effectively to the actual electromagnetic wave propagation). Therefore, to accurately find the single-hop diffraction points on the edges of the scatterer and calculate the single-hop diffraction contribution, this disclosure also involves an AEE (Actual Edge Extraction) algorithm. This algorithm is used to determine the actual edges existing on each scatterer to efficiently handle the edge recognition problem of complex-shaped scatterers and eliminate the influence of redundant inner edges and ground edges. Based on this, in one or more embodiments of this disclosure, determining the actual edges existing on each scatterer may include: Iterate through all scatterers in the environment where the transmitter and receiver are located, and add the four edges of each face to the set. ; Among them, the surface The four sides are denoted as , , and Each edge represents the parametric equation of a line segment defined by two vertices.

[0119] Iterate through every two distinct faces Kneading noodles ,judge and Are they coplanar? For example, by judging the face unit normal vector With noodles unit normal vector cross product norm and With noodles any point on to noodles any point on Whether the dot product of all vectors is 0 can determine the surface. Kneading noodles Do they share the same surface?

[0120] if and If they share the same face, then traverse them. The four sides and Calculate the four edges. arrive Given a vector V, determine The l-th edge and The k-th edge Are they parallel and collinear? Among them, judgment and Whether the cross product norm is 0 can determine whether two sides are parallel. (This is related to the determination of V and...) Whether the cross product norm is 0 can determine whether two edges are collinear; If the edge and edge If they are parallel and collinear, then calculate the edges. The starting point and the end point On the side Projection parameters on and ; Representation surface The l-th edge, and Representing edges respectively The starting point and the ending point, Representing an edge directional vector, Representing an edge The direction vector; by The line containing it is the parametric axis. The parameter range is , The parameter range is ,judge The parameter range and Do the parameter ranges overlap (i.e.) and Does it have an intersection? If it does, then it means there is an edge. and edge (There are overlapping parts), if there is overlap, will as well as Add to collection middle; Let set With sets Same, then iterate. For each edge in the array, if the currently traversed edge is in... In or located in On a plane, from Remove the edges currently being traversed from the set to obtain the set. This refers to the set of actual edges that exist on the scatterer. like Figure 5As shown in process flow C, determining the single-hop diffraction point on the actual edge of each scatterer may include: Traverse the actual edges of all effective scatterers at both the signal transmitter and receiver, and calculate... and At the edge Projection point parameters and , ,in, Indicates edge The starting point ( Indicates edge (the end point) Indicates edge The direction vector; calculate and At the edge Projection point on and , ; calculate and The distance to their respective projection points, ; Calculate the edge Single-hop diffraction point parameters , ; If parameters Then calculate the edge Single-hop diffraction point , ; Iterate through all surfaces of all scatterers in the environment where the transmitter and receiver are located, and determine... and Connect or and Is the connection blocked? and Connect or and If the line is occluded, there is no single-hop diffraction point on the edge being traversed. If parameter If no single-hop diffraction point exists on the edge of the current traversal, then no single-hop diffraction point exists on the edge of the current traversal. Each calculated single-hop diffraction point Add to collection In this process, we obtain the set of single-hop diffraction points.

[0121] Furthermore, for a specific pair of transmitter and receiver coordinates, if a single-hop diffraction point on the edge of a scatterer... If it exists, then the corresponding single-hop diffraction contribution... It can be represented as: Among them, distance coefficient The calculation formula is the same as the single-hop reflection contribution of the scatterer surface mentioned above. The calculation formula is the same, the difference lies in the single-hop reflection point of the scattering surface. Replace with single-hop diffraction point . The uniform diffraction coefficient can be expressed as: ; in, , The wedge angle is the edge where the diffraction point is located. , λ is the wavelength. Indicates the angle of incidence. Indicates the angle of diffraction.

[0122] In real-world data propagation environments, besides single-hop paths, numerous multi-hop propagation paths consisting of multiple reflections and diffractions exist between the transmitter and receiver. Accurately identifying the interaction points in each multi-hop path is computationally complex and difficult to meet the real-time requirements of communication. Given that the objective of this disclosure is to evaluate the overall contribution of multi-hop paths to channel characteristics, rather than precisely reconstructing each multi-hop path, this disclosure proposes a graph-based efficient process for determining multi-hop propagation paths (hereinafter also referred to as multi-hop paths). This process aims to abstract the propagation environment into a visibility graph and identify representative multi-hop propagation paths through a constrained graph search. Based on this, in one or more embodiments of this disclosure, determining the multi-hop propagation path from the transmitter's sampling point to the receiver may include: Discrete sampling is performed on the rectangular surface of each scatterer, and the four vertices and the center point of each rectangular surface are selected as sampling points; Remove the sampling points located on the ground from the above sampling points ( The sampling points are used to obtain valid sampling points; Construct an undirected graph based on the valid sampling points. , where the node set Includes sampling points on each scatterer. include All edges in; Determine whether the line connecting any two sampling points belonging to different scatterers in the valid sampling points is blocked by any scatterer surface; If there is no obstruction, establish an undirected edge between the two sampling points currently being judged; For example, the first indivual( The first scatterer indivual( The sampling point can be represented as , For each node, its attributes can be represented as a feature vector. ,in, Sampling points The three-dimensional coordinates Sampling points The scatterer number is used to determine whether the line connecting the two sampling points is obscured by any scatterer surface, using the method for determining line-surface intersection in the single-hop path contribution calculation. If the two sampling points are visible, an undirected edge (also called a bidirectional edge) is established between them. It should be noted that visibility is not checked between sampling points on the same scatterer, and no connection is established, thus eliminating invalid paths within the scatterer.

[0123] For each pair of senders and receivers, in the undirected graph Add sender and receiver nodes to form a complete visibility graph. ; Determine whether there is any obstruction between the transmitter and receiver and the sampling point on the corresponding valid scatterer. If there is no obstruction, establish directed edges from the transmitter node to the valid scatterer sampling point and from the valid scatterer sampling point to the receiver node. Optionally, the visibility between Tx and Rx and the sampling points on the corresponding effective scatterers can be determined by judging whether the lines and surfaces intersect. If Tx and Rx are visible to the sampling points on the corresponding effective scatterers, directed edges are established from the Tx node to the effective scatterer sampling point, and from the effective scatterer sampling point to the Rx node. Based on this, the global environment graph can be focused on the local topology related to the current Tx-Rx link. It should be noted that this process narrows the scope of judgment to the effective scatterers. Since the effective scatterers are the scatterers that electromagnetic waves pass through during propagation, multiple reflections and diffractions in multi-hop paths can only occur on the effective scatterers. Considering only the visibility between Tx and Rx and the sampling points on the effective scatterers can effectively reduce the amount of computation and meet the requirements of high-efficiency online computing. Figure 6 This illustration shows an exemplary method for finding a multi-hop propagation path from the transmitter through sampling points to the receiver, as provided in an embodiment of this disclosure. Figure 6 As shown, since scatterer 1 is not an effective scatterer corresponding to this pair of Tx-Rx, it can be... and The edges connected to Tx and Rx are removed, leaving only the sampling points on the effective scatterer. Directed edges are then connected to these points based on the visibility between Tx and Rx, forming a complete visibility graph. .

[0124] For complete visibility graph A breadth-first search algorithm is used to find a path that meets the preset conditions, thus obtaining a multi-hop propagation path from the transmitter to the receiver via sampling points. The preset conditions may include: The number of intermediate nodes traversed by the path is within a first numerical range, and the total path length does not exceed the number of nodes within a second numerical range; for example, the number of hops (i.e. the number of intermediate nodes traversed) of the path is limited to between 2 and 6, corresponding to a total path length of 4 to 8 nodes (plus Tx and Rx nodes), in order to exclude unrealistically long paths and single-hop paths.

[0125] Duplicate sampling points are not allowed in a path to avoid loops; The scatterers to which the sampling points along any two paths belong are not exactly the same. These sampling points may belong to the same scatterers. Therefore, in the final set of retained paths, the scatterers to which the sampling points along any two paths belong cannot be exactly the same, in order to avoid excessive similarity between the path samples.

[0126] The search yields all paths that satisfy the above preset conditions, resulting in a multi-hop propagation path from the sender, through sampling points, to the receiver.

[0127] like Figure 6 As shown, the final valid paths that meet the above preset conditions are: and Due to the path The length is 3 nodes, which does not meet the above preset conditions, so it will not be retained. Furthermore, due to the path... and path The scatterers at all the sampling points are identical, so only one path is retained. Similarly, the path... and path The scatterers of the sampling points are all exactly the same, so only one of them is retained.

[0128] For example, calculating the multi-hop path contribution based on the multi-hop propagation path from the sampling point at the transmitter to the receiver obtained through the above processing may include the following steps: For all paths that meet the search criteria, calculate their total geometric length based on the 3D coordinates of each point, and sort them in ascending order. Retain at most the top 20 shortest paths and calculate their average path length. Multi-hop path contribution of a pair of Tx-Rx It can be defined as the ratio of the direct path length of Tx and Rx to the average path length over multiple hops, and can be expressed by the following formula: ; The ratio in the above formula reflects the efficiency of multi-hop path propagation relative to free space propagation. The closer the ratio is to 1, the more direct the propagation of the dominant multi-hop path is, and the smaller the path loss; the smaller the ratio, the more circuitous the signal needs to travel, indicating greater propagation loss.

[0129] The channel state information prediction method of this disclosure can efficiently obtain a reasonable estimate of the contribution to multi-hop path propagation within a controllable computational overhead by using a graph search-based approximation method, achieving a good balance between model complexity and accuracy.

[0130] In the method of this disclosure, before constructing the EEK, it is necessary to accurately identify the effective scatterers in the scene. In actual large-scale communication scenarios (such as dense urban areas), the number of scatterers is enormous. If the effective scatterers in the scene are not identified in advance, all scatterers in the scene must be traversed in subsequent processing such as reflection point calculation and diffraction contribution analysis, which will introduce huge computational overhead and seriously affect the efficiency and real-time performance of EEK construction. Based on this, in one or more embodiments of this disclosure, feature vectors extracted based on the location information of the transmitting and receiving ends and the geometric information of the scatterers are provided to a pre-trained effective scatterer classification model to obtain the effective scatterer prediction result, which may include: Extract features from the location information of the transmitter and receiver, as well as the geometric information of the scatterer; Whether a scatterer is valid can be uniquely determined by the overall spatial relationship formed by the Tx coordinates, Rx coordinates, and the geometric properties of all scatterers in the scene. Based on this, the position information of the transmitter and receiver can include the three-dimensional coordinates of the transmitter and receiver, and the geometric information of the scatterer can include the set of all rectangular planes of the scatterer; The extracted features are then normalized. The normalized features are constructed into a sequence; Generate a learnable positional encoding vector for each position in the constructed sequence to obtain the input sequence; The input sequence is encoded based on the encoding layer to obtain the encoded sequence. The encoding layer includes four Transformer encoders, each of which contains a multi-head self-attention mechanism and a feedforward neural network. The feedforward neural network may include two MLP layers. The vectors representing the transmitter and receiver are removed from the sequence output by the coding layer to obtain the vector corresponding to the scatterer. Average pooling is performed on the vectors belonging to the same scatterer to obtain the global representation vector of each scatterer. The global representation vectors of each scatterer are processed through a linear layer and an activation function to output the probability that each scatterer is a valid scatterer, thus obtaining the prediction results of valid scatterers.

[0131] The following example illustrates the scatterer classification model in the channel state information prediction method of this disclosure.

[0132] Figure 7 A schematic diagram of an exemplary scatterer classification model provided in an embodiment of this disclosure is shown. The training process of this model may include the following steps: First, a model is constructed based on the problem of effective scatterer identification: Each scatterer surface From its parametric equation coefficients Definition. Given transmitter coordinates and receiver The task of determining effective scatterers can be formulated as the following optimization problem: ; in, It is a neural network based on Transformer. These are its trainable parameters. It is the actual effective scatterer label vector. It is the network's predicted output. This represents the binary cross-entropy loss function.

[0133] Secondly, data preprocessing and input sequence construction are performed in input layers 72A and 72B: The input to the aforementioned network can be heterogeneous. The coordinates of Tx and Rx are combined into a 6-dimensional vector, and each scatterer surface is described by 12-dimensional parametric equation coefficients. First, to eliminate dimensional differences, the input features can be preprocessed using Z-score normalization, as shown in the formula: ; in, The characteristic mean, The standard deviation is used to normalize the features to a distribution with a mean of 0 and a standard deviation of 1, thereby accelerating network training convergence and improving model training stability. Then, the standardized scene data is constructed into a sequence. Specifically, the coordinates of the 6-dimensional Tx and Rx can be merged into a vector and passed through a linear layer (such as...). Figure 7 The fully connected layer 74A shown is projected onto... Dimensionality. The coefficients of the 12-dimensional parametric equations for each scatterer surface are passed through another linear layer (such as...). Figure 7 The fully connected layer 74B shown is projected into 128 dimensions. The coordinates of Tx and Rx after dimensionality transformation, along with the coefficients of the scatterer parameter equations, are connected at connection layer 76 to obtain the network's input sequence. It consists of three parts pieced together: ,in The total number of scattering surfaces in the scene; this serialization construction enables the model to handle communication scenarios of different scales.

[0134] Considering that the Transformer architecture is insensitive to sequence order, to endow the model with sequence position awareness, positional encoding 78 is added to connection layer 76 to introduce learnable positional encoding. A learnable embedding layer can be used to generate a unique, learnable 128-dimensional positional encoding vector for each position in the sequence. Final embedding representation It can be obtained by adding element by element, and the expression is: ; The model adopts Layered Transformer encoders are stacked, with each layer potentially containing a multi-head self-attention layer (710) and an FFN (Feed-Forward Network, 714). For multi-head self-attention, the following can be used: Each head has an attention mechanism. First, the 128-dimensional input is linearly projected onto each head. The attention mechanism can be computed using scaled dot product attention, as shown in the following expression: ; This mechanism enables the model to dynamically capture the global spatial dependencies between Tx, Rx and all scattering surfaces.

[0135] For a feedforward neural network, a two-layer MLP can be used to perform a non-linear transformation on the representation at each location. Its hidden layer dimension is... Using the ReLU activation function, the formula is as follows: ; In the above formula, and This is the weight matrix. and This is the bias vector. Each layer is followed by a residual connection layer and normalization, by... Figure 7 The addition and normalization layers 712 and 716 shown are used to stabilize the training process. After layer Transformer encoding, a sequence representation containing rich contextual information can be obtained. .

[0136] Since the goal of the model is to determine whether each scatterer is effective, we can start with the output sequence of the last Transformer encoder layer. The output vectors representing the first two positions of Tx and Rx are excluded, retaining only the output vectors corresponding to all scatterer surfaces. Then, average pooling is performed on the representations of all surfaces belonging to the same scatterer to obtain the global representation vector for each scatterer. : ; Then, the global representation of each scatterer is... Through a linear layer (used to map 128-dimensional features to 1 dimension, which can be obtained from...) Figure 7 The connection layer 718 shown is implemented with a sigmoid activation function 719, which outputs the probability that the scatterer is determined to be a valid scatterer. : ; In the above formula, This is the weight matrix. This is the bias vector.

[0137] Ultimately, the network's predicted output is a vector. .

[0138] The model can be trained end-to-end under supervised supervision using a labeled dataset generated by high-fidelity ray tracing software. The dataset labels are binary valid scatterer identifiers. The training strategy is as follows: The loss function can be the binary cross-entropy loss, which is formally defined as: ; This loss function is well-suited to the multi-label binary classification characteristics of the effective scatterer identification task in this embodiment. Furthermore, the Adam optimizer can be used for model parameter updates, and its adaptive learning rate characteristic contributes to efficient convergence.

[0139] After the Transformer encoding is completed, the embedded representations corresponding to the scatterers are extracted and mapped into a one-dimensional feature space by inputting them into the linear layer. Then, the effective probability score for each scatterer is obtained through the Sigmoid activation function. To prevent overfitting and improve generalization ability in different scenarios, dropout regularization can be applied to the attention module and feedforward network.

[0140] The method in this embodiment pre-trains a Transformer-based neural network model to identify effective scatterers. By learning the complex spatial relationships between Tx, Rx, and the scatterer, it achieves more intelligent and accurate identification of effective scatterers. Furthermore, by pre-screening effective scatterers through the scatterer classification model, subsequent complex geometric calculations can be focused on a very small, highly correlated subset, thereby significantly reducing the overall computational complexity.

[0141] Figures 8A-8C A performance comparison diagram is shown between the channel state information prediction method provided in the embodiments of this disclosure and the channel parameter prediction method based on graph neural networks, as follows: Figure 8A As shown, with sample sizes of 10, 50, and 100, the time taken for channel state information prediction using existing graph neural networks is 8.5s, 35.0s, and 70.0s, respectively, while the time taken for channel state prediction using the channel state information prediction method of this disclosure is 3.0s, 16.4s, and 35.7s, respectively. It is evident that compared to the graph neural network-based channel parameter prediction method, the channel state information prediction method of this disclosure performs superiorly in terms of total time and throughput. Figure 8B As shown, when the sample data is 10, 50, and 100, the throughput of existing channel state information prediction based on graph neural networks is 2.0 samples / second, 25.0 samples / second, and 28.0 samples / second, respectively. However, the throughput of channel state information prediction using the method of this embodiment is 2.6 samples / second, 44.0 samples / second, and 57.0 samples / second, respectively. Therefore, the channel state information prediction method of this embodiment can achieve better data throughput. Figure 8C As shown, with sample data of 10, 50, and 100, the peak memory usage for channel state information prediction using existing graph neural networks is 16.00MB, 17.00MB, and 18.00MB, respectively. In contrast, the peak memory usage for channel state estimation using the channel state information prediction method of this embodiment is 19.00MB, 22.00MB, and 23.00MB, respectively. This demonstrates that the channel state information prediction method of this embodiment significantly reduces computational overhead with a slight increase in peak memory usage. Therefore, the channel state information prediction method of this embodiment achieves efficient utilization of storage and computing resources while meeting real-time requirements.

[0142] One or more embodiments of this disclosure also provide a computer device, including one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and executed by one or more processors, and the one or more programs include instructions for a channel state information prediction method according to one or more embodiments of this disclosure.

[0143] One or more embodiments of this disclosure also provide a non-volatile computer-readable storage medium containing a computer program that, when executed by one or more processors, causes the one or more processors to perform a channel state information prediction method according to one or more embodiments of this disclosure.

[0144] This disclosure also provides a computer program product, including one or more computer programs, which, when executed by one or more processors, implement the channel state information prediction method of one or more embodiments of this disclosure.

[0145] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0146] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0147] For ease of description, the above computer devices are described in terms of function, divided into various modules. Of course, in implementing this disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0148] The computer device described in the above embodiments is used to implement the corresponding channel state information prediction method in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0149] This disclosure also provides a computer device for implementing the channel state information prediction method described above. Figure 9 A schematic diagram of the hardware structure of an exemplary computer device 900 provided in an embodiment of this disclosure is shown. The computer device 900 can be used to implement... Figure 1 Server 106 can also be used to implement Figure 1 Terminal devices 102 and 104. In some scenarios, this computer device 900 can also be used to implement... Figure 1 Database server 108.

[0150] like Figure 9As shown, the computer device 900 may include: a processor 902, a memory 904, a network interface 906, a peripheral interface 908, and a bus 910. The processor 902, memory 904, network interface 906, and peripheral interface 908 are interconnected within the computer device 900 via the bus 910.

[0151] Processor 902 may be a central processing unit (CPU), image processor, neural network processor (NPU), microcontroller (MCU), programmable logic device, digital signal processor (DSP), application-specific integrated circuit (ASIC), or one or more integrated circuits. Processor 902 can be used to perform functions related to the techniques described in this disclosure. In some embodiments, processor 902 may also include multiple processors integrated as a single logic component. For example, such as... Figure 9 As shown, processor 902 may include multiple processors 902a, 902b and 902c.

[0152] Memory 904 can be configured to store data (e.g., instructions, computer code, etc.). Figure 9 As shown, the data stored in memory 904 may include program instructions (e.g., one or more programs for implementing the channel state information prediction method of embodiments of this disclosure) and data to be processed (e.g., the memory may store configuration files of other modules, etc.). Processor 902 may also access the program instructions and data stored in memory 904 and execute the program instructions to operate on the data to be processed. Memory 904 may include volatile or non-volatile storage devices. In some embodiments, memory 904 may include random access memory (RAM), read-only memory (ROM), optical disk, magnetic disk, hard disk, solid-state drive (SSD), flash memory, memory stick, etc.

[0153] Network interface 906 can be configured to provide communication with other external devices to computer device 900 via a network. This network can be any wired or wireless network capable of transmitting and receiving data. For example, the network can be a wired network, a local wireless network (e.g., Bluetooth, WiFi, Near Field Communication (NFC), etc.), a cellular network, the Internet, or a combination thereof. It is understood that the type of network is not limited to the specific examples described above.

[0154] The peripheral interface 908 can be configured to connect the computer device 900 to one or more peripheral devices to enable information input and output. For example, peripheral devices may include input devices such as keyboards, mice, touchpads, touch screens, microphones, and various sensors, as well as output devices such as displays, speakers, vibrators, and indicator lights.

[0155] Bus 910 can be configured to transfer information between various components of computer device 900 (such as processor 902, memory 904, network interface 906, and peripheral interface 908), such as internal buses (e.g., processor-memory bus), external buses (USB port, PCI-E bus), etc.

[0156] It should be noted that although the architecture of the computer device 900 described above only shows the processor 902, memory 904, network interface 906, peripheral interface 908, and bus 910, in specific implementations, the architecture of the computer device 900 may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the architecture of the computer device 900 described above may only include the components necessary for implementing the embodiments of this disclosure, and does not necessarily include all the components shown in the figures.

[0157] Based on the same inventive concept, corresponding to any of the above embodiments, this disclosure also provides a non-volatile computer-readable storage medium containing a computer program, which, when executed by one or more processors, causes the one or more processors to perform the channel state information prediction method.

[0158] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0159] The computer program stored in the storage medium of the above embodiments is used to cause the one or more processors to execute the channel state information prediction method as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0160] Based on the same inventive concept, corresponding to the channel state information prediction method in any of the above embodiments, this disclosure also provides a computer program product, which includes one or more computer programs. In some embodiments, the one or more computer programs are executable by one or more processors to cause the one or more processors to perform the channel state information prediction method. Corresponding to the execution entity for each step in each embodiment of the channel state information prediction method, the processor executing the corresponding step may belong to the corresponding execution entity.

[0161] The computer program product of the above embodiments is used to cause the processor to execute the channel state information prediction method as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0162] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0163] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this disclosure, the provided drawings may or may not show well-known power / ground connections to integrated circuit (IC) chips and other components. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this disclosure, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this disclosure will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this disclosure, it will be apparent to those skilled in the art that the embodiments of this disclosure can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0164] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0165] This disclosure is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A channel state information prediction method, characterized in that, include: Acquire the location information of the signal transmitter and receiver, as well as the geometric information of the scatterers in the environment where the transmitter and receiver are located; The electromagnetic environment knowledge from the transmitting end to the receiving end is determined based on the location information and the geometric information. The electromagnetic environment knowledge is used to characterize the relationship between channel information and environmental information between the transmitting end and the receiving end. The electromagnetic environment knowledge is encoded into a channel knowledge vector based on a pre-trained scene embedding model; Find embedding vectors from a pre-built channel knowledge base that have a similarity higher than a threshold with the channel knowledge vectors, wherein the channel knowledge base includes multiple embedding vectors corresponding to multiple different communication scenarios; Obtain one or more channel responses corresponding to the found embedding vector; The one or more channel responses are fused to obtain the channel state information prediction results for the transmitting end and the receiving end.

2. The method according to claim 1, characterized in that, The method further includes: training the scene embedding model, specifically including: Physical perturbation is applied to the scatterer features of anchor point sample data in the anchor point scene dataset based on perturbation parameters to generate semantically invariant positive sample data. The perturbation parameters include the scatterer position and reflection contribution value, or the perturbation parameters include the scatterer position and diffraction contribution value. From the anchor point scene dataset, select anchor point sample data whose scattering distribution difference is greater than a preset value, and generate negative sample data; The anchor sample data, the positive sample data, and the negative sample data are merged to obtain the training dataset; The sample data in the training dataset are encoded based on the embedding function to obtain anchor embedding results, positive sample embedding results, and negative sample embedding results. The information noise contrast estimation loss is calculated based on the anchor point embedding results, the positive sample embedding results, and the negative sample embedding results. The scene embedding model is obtained by updating the model parameters based on the information noise contrast estimation loss.

3. The method according to claim 2, characterized in that, The anchor point sample data includes: For each pair of signals, a directed graph is constructed from the transmitting and receiving ends. The vertex set of the directed graph includes transmitting node, receiving node, and active node. The active node types include: ground reflection point, scatterer surface reflection point, and scatterer edge diffraction point. The feature vector of each node in the vertex set includes coordinates, type encoding, and contribution value. The edge set of the directed graph includes directed edges from the transmitting end to the receiving end and directed edges from the transmitting end to each active node. The feature vector of each edge in the edge set includes the Euclidean distance between the two nodes, occlusion status, and contribution value.

4. The method according to claim 3, characterized in that, Encoding the sample data in the training dataset based on the embedding function includes: The feature vectors of the sending node and the receiving node in the training dataset are transformed to the first dimension through a fully connected neural network; The feature vectors of the direct path and the scattering path between the transmitting node and the receiving node are determined based on the feature vectors of the transmitting node, the receiving node, the action node n, the directed edge between the transmitting node and the action node n, and the directed edge between the action node n and the receiving node. The feature vector of the direct trajectory is transformed to the first dimension through the fully connected neural network; The feature vectors of the transmitting and receiving nodes after dimensionality transformation, along with the feature vector of the direct path, are merged to obtain a joint feature vector. The joint feature vector is transformed to a second dimension through the fully connected neural network to obtain the first semantic; The feature vector of the scattering path is transformed to the first dimension through a multilayer perceptron (MLP) to obtain the scattering path matrix. The scattering path matrix is ​​then subjected to mean pooling and transformed to the second dimension to obtain the second semantic. The scattering path matrix is ​​then subjected to multi-head attention aggregation to obtain the third semantics; The first semantic, the second semantic, and the third semantic are concatenated to obtain a concatenated vector. The concatenated vector is then transformed to a third dimension, and the concatenated vector in the third dimension is normalized to obtain an embedding vector.

5. The method according to claim 1, characterized in that, From a pre-built channel knowledge base, find embedding vectors with a similarity higher than a threshold to the channel knowledge vectors, including: The K nearest neighbor samples with the highest similarity to the channel knowledge vector are determined by Top-K nearest neighbor retrieval; The embedding vectors of the K nearest neighbor samples are divided into multiple clusters based on the clustering algorithm. A cluster identifier is calculated for each embedding vector, and the cluster identifier and the corresponding embedding vector are appended to the inverted index. Search the inverted table corresponding to the channel knowledge vector for the clusters corresponding to the centroids that are closest to the channel knowledge vector to obtain a set of candidate clusters; From the candidate cluster set, query the plurality of embedding vectors that have a similarity to the channel knowledge vector that is higher than a threshold.

6. The method according to claim 1, characterized in that, Determining the electromagnetic environment knowledge from the transmitting end to the receiving end based on the location information and the geometric information includes: The feature vectors extracted based on the location information and the geometric information are provided to the pre-trained scatterer classification model to obtain effective scatterer prediction results; The effective scatterers corresponding to the transmitting end and the receiving end are determined based on the effective scatterer prediction results; Traverse all surfaces of the effective scatterer to determine the occlusion relationship between the transmitter and the receiver; If there is no obstruction between the transmitting end and the receiving end, determine the ground reflection point and calculate the ground reflection contribution based on the ground reflection point; Determine the single-hop reflection point on the surface of the effective scatterer, and calculate the surface reflection contribution of the single-hop scatterer based on the single-hop reflection point; Determine the actual edges that exist on each scatterer; Determine the single-hop diffraction point on the actual edge of each scatterer, and calculate the edge diffraction contribution of the single-hop scatterer based on the single-hop diffraction point; Determine the multi-hop propagation path from the transmitting end through the sampling point to the receiving end, and calculate the multi-hop path contribution of the multi-hop propagation path; The electromagnetic environment knowledge is obtained by combining the ground reflection contribution, the single-hop scatterer surface reflection contribution, the single-hop scatterer edge diffraction contribution, and the multiple hop path contributions.

7. The method according to claim 6, characterized in that, Determining the ground reflection point includes: Determine the coordinates of the sending end Regarding the mirror point on the ground The coordinates of the receiving end and The intersection of the line and the plane containing the ground is determined as the ground reflection point. ; Traverse all surfaces of all scattering bodies in the environment where the transmitting end and the receiving end are located, and determine... and Connect or and Is the connection blocked? and Connect or and If the connecting line is blocked, then the ground reflection point Does not exist, if and Connect or and If none of the lines are blocked, then exist; Determining the single-hop reflection point on the surface of the effective scatterer includes: Iterate through all surfaces of all effective scatterers in the environment where the transmitter and receiver are located. ,Sure Regarding the effective scattering surface mirror point ; Judgment point and Do the connected line segments intersect the surface? intersect; If they do not intersect, then there is no single-hop reflection point on the scatterer surface being traversed. If they intersect, the single-hop reflection point of the scattering surface is calculated based on the parameters of the line segments. ; Iterate through all surfaces of all scatterers in the environment where the transmitter and receiver are located again to determine... and Connect or and Is the connection obstructed? like and Connect or and If the connection is blocked, then the single-hop reflection point on the currently traversed scattering surface... Does not exist, if and Connect or and If none of the connecting lines are obstructed, then the single-hop reflection point on the currently traversed scattering surface is... exist; Each of the calculations Add to collection In this process, the set of the single-hop reflection points is obtained; Determine the actual edges that exist on each scatterer, including: Iterate through all scatterers in the environment where the transmitter and receiver are located, and add the four edges of each face of the scatterer to the set. ; Iterate through every two distinct faces Kneading noodles ,judge and They are coplanar; if and If they share the same face, then traverse them. The four sides and Determine the four edges of the given condition. The l-th edge and The k-th edge Are they parallel and collinear? If the edge and edge If they are parallel and collinear, then calculate the edges. The starting point and the end point On the side Projection parameters on and ; by The line containing it is the parametric axis. The parameter range is , The parameter range is ,judge The parameter range and Does the parameter range overlap? If it overlaps, then... as well as Add to collection middle; Let set With sets Same, then iterate. For each edge in the array, if the currently traversed edge is in... In or located in On a plane, from Remove the edges currently being traversed from the set to obtain the set. This refers to the set of actual edges that exist on the scatterer. Determining the single-hop diffraction point on the actual edge of each scatterer includes: Traverse the actual edges of the effective scatterer and calculate and At the edge Parameters of the projection points on; calculate and At the edge Projection points on; calculate and The distance to their respective projection points; The edge is calculated based on the projection point parameters and the distance. Single-hop diffraction point parameters ; If parameters Then calculate the edge Single-hop diffraction point ; Traverse all surfaces of all scattering bodies in the environment where the transmitting end and the receiving end are located, and determine... and Connect or and Is the connection blocked? and Connect or and If the line is occluded, there is no single-hop diffraction point on the edge being traversed. If parameter If no single-hop diffraction point exists on the edge of the current traversal, then no single-hop diffraction point exists on the edge of the current traversal. The calculated single-hop diffraction point Add to collection In this process, the set of single-hop diffraction points is obtained.

8. The method according to claim 6, characterized in that, Determining the multi-hop propagation path from the transmitting end through the sampling point to the receiving end includes: Discrete sampling is performed on the rectangular surface of each scatterer, and the four vertices and the center point of each rectangular surface are selected as sampling points; The sampling points located on the ground are removed from the sampling points to obtain the valid sampling points; Construct an undirected graph based on the valid sampling points. , where the node set Includes sampling points on each of the scatterers. include All edges in; Determine whether the line connecting any two sampling points belonging to different scatterers in the valid sampling points is blocked by any scatterer surface; If there is no obstruction, establish an undirected edge between the two sampling points currently being judged; For each pair of transmitters and receivers, in the undirected graph Add sender and receiver nodes to form a complete visibility graph. ; Determine whether there is any obstruction between the transmitter and receiver and the sampling point on the corresponding valid scatterer. If there is no obstruction, establish directed edges from the transmitter node to the valid scatterer sampling point and from the valid scatterer sampling point to the receiver node. For the complete visibility map A breadth-first search algorithm is used to find a path that meets preset conditions, and the path from the sending end through the sampling point to the receiving end is obtained. The preset conditions include: The number of intermediate nodes traversed by the path is within the first numerical range, and the total path length does not exceed the number of nodes within the second numerical range; Duplicate sampling points are not allowed in a single path; The scatterers to which the sampling points traversed by any two paths belong are not exactly the same; The search yields all paths that satisfy the preset conditions, thus obtaining the multi-hop propagation path.

9. The method according to claim 6, characterized in that, The feature vector extracted based on the location information and the geometric information is provided to the pre-trained effective scatterer prediction model to obtain the effective scatterer prediction result, including: Features of the location information and the geometric information are extracted, wherein the location information includes the three-dimensional coordinates of the transmitting end and the receiving end, and the geometric information includes the set of all rectangular planes of the scatterer; The extracted features are then normalized. The normalized features are constructed into a sequence; Generate a learnable position encoding vector for each position in the sequence to obtain the input sequence; The input sequence is encoded based on the encoding layer to obtain the encoded sequence. The encoding layer includes four Transformer encoders, each of which contains a multi-head self-attention mechanism and a feedforward neural network. The feedforward neural network includes two multilayer perceptron (MLP) layers. The vectors representing the transmitter and receiver are removed from the sequence output by the coding layer to obtain the vector corresponding to the scatterer. The vectors belonging to the same scatterer among the vectors corresponding to the scatterer are averaged and pooled to obtain the global representation vector of each scatterer. The global representation vectors of each scatterer are processed through a linear layer and an activation function to output the probability that each scatterer is a valid scatterer, thus obtaining the prediction results of valid scatterers.

10. The method according to any one of claims 6 to 9, characterized in that, Traversing all surfaces of the effective scatterer to determine the occlusion relationship between the transmitter and the receiver includes: Traverse all surfaces of the scatterer and determine whether the line connecting the coordinates of the receiver and the coordinates of the transmitter intersects with any surface of the scatterer. If an intersection point exists, it is determined that there is an obstruction between the transmitting end and the receiving end; otherwise, it is determined that there is no obstruction between the transmitting end and the receiving end.

11. A computer device comprising one or more processors, a memory; and one or more programs, wherein the one or more programs are stored in the memory and executed by the one or more processors, the one or more programs comprising instructions for performing the method of any one of claims 1 to 10.

12. A non-volatile computer-readable storage medium comprising a computer program, which, when executed by one or more processors, causes the one or more processors to perform the method of any one of claims 1 to 10.

13. A computer program product comprising one or more computer programs that, when executed by one or more processors, implement the method as described in any one of claims 1 to 10.

Images