Digital twin channel prediction method, device and equipment based on channel large model

Abstract

This application provides a digital twin channel prediction method, apparatus, and device based on a large channel model. The method includes: constructing a three-dimensional wireless environment corresponding to a target area using a digital twin channel architecture and multimodal sensing data for that area; extracting electromagnetic propagation parameter information for the three-dimensional wireless environment using the architecture; the electromagnetic propagation parameter information includes: transceiver location information, scatterer distribution information, and penetration characteristic information corresponding to the three-dimensional wireless environment; and predicting channel information between transceivers using the large channel model in the architecture, based on the transceiver location information, scatterer distribution information, and penetration characteristic information. This solution can address at least one of the problems in existing channel information prediction schemes, such as poor environmental adaptability, insufficient real-time performance, and weak multi-task joint prediction capability.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a digital twin channel prediction method, apparatus and device based on a large channel model. Background Technology

[0002] Digital twin channels represent a significant paradigm shift in channel information acquisition. However, existing digital twin channel implementations suffer from several shortcomings, including poor environmental adaptability, insufficient real-time performance, and weak multi-task joint prediction capabilities. Specifically: (1) Regarding poor environmental adaptability, traditional statistical modeling relies on offline CIR (Channel Impulse Response) data from limited typical scenarios to fit the channel parameter distribution; small-scale AI (Artificial Intelligence) models are only applicable to fixed environments; and mainstream large-scale AI models focus on directly processing channel data, leading to the submergence of key features and weakening cross-scenario generalization capabilities. (2) Regarding insufficient real-time performance, traditional statistical modeling focuses on offline CIR measurements and parameter fitting. (3) Regarding weak multi-task joint prediction capabilities, traditional statistical modeling methods and small-scale AI algorithms lack a unified multi-task collaboration mechanism. These shortcomings are mutually coupled and restrictive, making it difficult to implement the digital twin channel paradigm and severely hindering the development of 6G air interface intelligence.

[0003] Based on the above, existing channel information prediction schemes also suffer from problems such as poor environmental adaptability, insufficient real-time performance, and weak multi-task joint prediction capability. Summary of the Invention

[0004] The purpose of this application is to provide a digital twin channel prediction method, apparatus and device based on a large channel model, so as to solve at least one of the problems of poor environmental adaptability, insufficient real-time performance and weak multi-task joint prediction capability in the existing channel information prediction schemes.

[0005] To address the aforementioned technical problems, embodiments of this application provide a digital twin channel prediction method based on a large channel model, comprising: By utilizing a digital twin channel architecture, a three-dimensional wireless environment corresponding to the target area is constructed based on multimodal sensing data for the target area. Using the aforementioned digital twin channel architecture, electromagnetic propagation parameter information is extracted for the three-dimensional wireless environment. The electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment. The scatterers are other objects in the three-dimensional wireless environment besides the transceiver devices. Using the large-scale channel model in the digital twin channel architecture, channel information between transceivers is predicted based on the transceiver location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

[0006] Optionally, the multimodal sensing data includes: LiDAR point cloud data and RGB-D images; the construction of a three-dimensional wireless environment corresponding to the target area based on the multimodal sensing data for the target area includes: Based on the LiDAR point cloud data, a static scatterer in the target region is constructed; Based on the LiDAR point cloud data and RGB-D image, a dynamic scatterer in the target region is constructed; Based on the static and dynamic scatterers, a three-dimensional wireless environment corresponding to the target area is constructed.

[0007] Optionally, constructing the dynamic scatterer in the target region based on the LiDAR point cloud data and the RGB-D image includes: Based on the RGB-D image, determine the reference coordinates of the dynamic scatterer in the target region; Based on the reference coordinates, cluster analysis is performed on the LiDAR point cloud data in the neighborhood of the dynamic scatterer to determine the center coordinates and boundary information of the dynamic scatterer. Based on the center coordinates and boundary information, a dynamic scatterer is constructed in the target region.

[0008] Optionally, extracting electromagnetic propagation parameter information for the three-dimensional wireless environment includes: The target area is rasterized in the horizontal direction to obtain a two-dimensional raster coordinate system; Based on the two-dimensional grid coordinate system, electromagnetic propagation parameter information for the three-dimensional wireless environment is obtained.

[0009] Optionally, based on the two-dimensional grid coordinate system, the transmittance feature information is obtained, including: Determine whether a line-of-sight link exists between the receiving device and the transmitting device; the number of receiving devices is at least one. If it exists, the grid position corresponding to the receiving device in the two-dimensional grid coordinate system shall be the first value; If not, the transmittance corresponding to the grid position is determined based on the three-dimensional coordinates of the receiving device and the three-dimensional coordinates of the first and second positions on the first connecting line; the first connecting line refers to the connecting line between the receiving device and the transmitting device, the first position refers to the intersection point of the first connecting line and the scatterer that is closest to the transmitting device, and the second position refers to the intersection point of the first connecting line and the scatterer that is closest to the receiving device. The penetration characteristic information is obtained based on the first value and / or penetration rate.

[0010] Optionally, predicting the channel information between the transceiver devices based on the transceiver device location information, scatterer distribution information, and penetration characteristic information includes: Based on the location information of the transceiver device, the distribution information of the scatterer, and the penetration characteristic information, the global environmental characteristics are obtained; Based on the global environmental features, mapping processing is performed to predict the channel information between the transceiver devices.

[0011] Optionally, the small-scale channel information includes channel state information (CSI); the method further includes: Based on the number of detection signals sent by the transmitting device that is less than a first number, a portion of the CSI corresponding to the receiving device is obtained; The mapping process based on the global environmental features, which predicts the CSI between transceiver devices, includes: Based on the aforementioned CSI portion, obtain the CSI reference features; The local environmental features corresponding to the receiving device are obtained by extracting the global environmental features. The CSI reference features are fused with the local environment features and then mapped to predict the CSI between the transceiver devices.

[0012] Optional, also includes: For the sample region used for model training, the original wireless channel information corresponding to each receiving position is obtained; the sample region includes at least one area of ​​a scene. Based on the original wireless channel information, the channel information corresponding to each receiving location in the sample area is extracted as sample channel information; Based on the sample channel information, training data is constructed; The channel model is trained using the training data.

[0013] This application also provides a digital twin channel prediction device based on a large channel model, comprising: The first construction module is used to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area using a digital twin channel architecture. The first extraction module is used to extract electromagnetic propagation parameter information for the three-dimensional wireless environment using the digital twin channel architecture; the electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment; the scatterer is other objects in the three-dimensional wireless environment besides the transceiver. The first prediction module is used to predict the channel information between the transceiver devices based on the large channel model in the digital twin channel architecture, according to the transceiver device location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to the channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to the channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

[0014] Optionally, the multimodal sensing data includes: LiDAR point cloud data and RGB-D images; the construction of a three-dimensional wireless environment corresponding to the target area based on the multimodal sensing data for the target area includes: Based on the LiDAR point cloud data, a static scatterer in the target region is constructed; Based on the LiDAR point cloud data and RGB-D image, a dynamic scatterer in the target region is constructed; Based on the static and dynamic scatterers, a three-dimensional wireless environment corresponding to the target area is constructed.

[0015] Optionally, constructing the dynamic scatterer in the target region based on the LiDAR point cloud data and the RGB-D image includes: Based on the RGB-D image, determine the reference coordinates of the dynamic scatterer in the target region; Based on the reference coordinates, cluster analysis is performed on the LiDAR point cloud data in the neighborhood of the dynamic scatterer to determine the center coordinates and boundary information of the dynamic scatterer. Based on the center coordinates and boundary information, a dynamic scatterer is constructed in the target region.

[0016] Optionally, extracting electromagnetic propagation parameter information for the three-dimensional wireless environment includes: The target area is rasterized in the horizontal direction to obtain a two-dimensional raster coordinate system; Based on the two-dimensional grid coordinate system, electromagnetic propagation parameter information for the three-dimensional wireless environment is obtained.

[0017] Optionally, based on the two-dimensional grid coordinate system, the transmittance feature information is obtained, including: Determine whether a line-of-sight link exists between the receiving device and the transmitting device; the number of receiving devices is at least one. If it exists, the grid position corresponding to the receiving device in the two-dimensional grid coordinate system shall be the first value; If not, the transmittance corresponding to the grid position is determined based on the three-dimensional coordinates of the receiving device and the three-dimensional coordinates of the first and second positions on the first connecting line; the first connecting line refers to the connecting line between the receiving device and the transmitting device, the first position refers to the intersection point of the first connecting line and the scatterer that is closest to the transmitting device, and the second position refers to the intersection point of the first connecting line and the scatterer that is closest to the receiving device. The penetration characteristic information is obtained based on the first value and / or penetration rate.

[0018] Optionally, predicting the channel information between the transceiver devices based on the transceiver device location information, scatterer distribution information, and penetration characteristic information includes: Based on the location information of the transceiver device, the distribution information of the scatterer, and the penetration characteristic information, the global environmental characteristics are obtained; Based on the global environmental features, mapping processing is performed to predict the channel information between the transceiver devices.

[0019] Optionally, the small-scale channel information includes channel state information (CSI); the apparatus further includes: The first acquisition module is used to acquire a portion of the CSI corresponding to the receiving device based on the detection signals sent by the transmitting device, which are less than a first number. The mapping process based on the global environmental features, which predicts the CSI between transceiver devices, includes: Based on the aforementioned CSI portion, obtain the CSI reference features; The local environmental features corresponding to the receiving device are obtained by extracting the global environmental features. The CSI reference features are fused with the local environment features and then mapped to predict the CSI between the transceiver devices.

[0020] Optional, also includes: The second acquisition module is used to acquire the original wireless channel information corresponding to each receiving position for the sample area used for model training; the sample area includes at least one area of ​​a scene. The second extraction module is used to extract the channel information corresponding to each receiving position in the sample area based on the original wireless channel information, and use it as sample channel information. The second construction module is used to construct training data based on the sample channel information; The first training module is used to train the large channel model using the training data.

[0021] This application also provides a digital twin channel prediction device based on a large channel model, including: a processor; The processor is used to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area using a digital twin channel architecture. Using the aforementioned digital twin channel architecture, electromagnetic propagation parameter information is extracted for the three-dimensional wireless environment. The electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment. The scatterers are other objects in the three-dimensional wireless environment besides the transceiver devices. Using the large-scale channel model in the digital twin channel architecture, channel information between transceivers is predicted based on the transceiver location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

[0022] Optionally, the multimodal sensing data includes: LiDAR point cloud data and RGB-D images; the construction of a three-dimensional wireless environment corresponding to the target area based on the multimodal sensing data for the target area includes: Based on the LiDAR point cloud data, a static scatterer in the target region is constructed; Based on the LiDAR point cloud data and RGB-D image, a dynamic scatterer in the target region is constructed; Based on the static and dynamic scatterers, a three-dimensional wireless environment corresponding to the target area is constructed.

[0023] Optionally, constructing the dynamic scatterer in the target region based on the LiDAR point cloud data and the RGB-D image includes: Based on the RGB-D image, determine the reference coordinates of the dynamic scatterer in the target region; Based on the reference coordinates, cluster analysis is performed on the LiDAR point cloud data in the neighborhood of the dynamic scatterer to determine the center coordinates and boundary information of the dynamic scatterer. Based on the center coordinates and boundary information, a dynamic scatterer is constructed in the target region.

[0024] Optionally, extracting electromagnetic propagation parameter information for the three-dimensional wireless environment includes: The target area is rasterized in the horizontal direction to obtain a two-dimensional raster coordinate system; Based on the two-dimensional grid coordinate system, electromagnetic propagation parameter information for the three-dimensional wireless environment is obtained.

[0025] Optionally, based on the two-dimensional grid coordinate system, the transmittance feature information is obtained, including: Determine whether a line-of-sight link exists between the receiving device and the transmitting device; the number of receiving devices is at least one. If it exists, the grid position corresponding to the receiving device in the two-dimensional grid coordinate system shall be the first value; If not, the transmittance corresponding to the grid position is determined based on the three-dimensional coordinates of the receiving device and the three-dimensional coordinates of the first and second positions on the first connecting line; the first connecting line refers to the connecting line between the receiving device and the transmitting device, the first position refers to the intersection point of the first connecting line and the scatterer that is closest to the transmitting device, and the second position refers to the intersection point of the first connecting line and the scatterer that is closest to the receiving device. The penetration characteristic information is obtained based on the first value and / or penetration rate.

[0026] Optionally, predicting the channel information between the transceiver devices based on the transceiver device location information, scatterer distribution information, and penetration characteristic information includes: Based on the location information of the transceiver device, the distribution information of the scatterer, and the penetration characteristic information, the global environmental characteristics are obtained; Based on the global environmental features, mapping processing is performed to predict the channel information between the transceiver devices.

[0027] Optionally, the small-scale channel information includes channel state information (CSI); the processor is further configured to: Based on the number of detection signals sent by the transmitting device that is less than a first number, a portion of the CSI corresponding to the receiving device is obtained; The mapping process based on the global environmental features, which predicts the CSI between transceiver devices, includes: Based on the aforementioned CSI portion, obtain the CSI reference features; The local environmental features corresponding to the receiving device are obtained by extracting the global environmental features. The CSI reference features are fused with the local environment features and then mapped to predict the CSI between the transceiver devices.

[0028] Optionally, the processor is further configured to: For the sample region used for model training, the original wireless channel information corresponding to each receiving position is obtained; the sample region includes at least one area of ​​a scene. Based on the original wireless channel information, the channel information corresponding to each receiving location in the sample area is extracted as sample channel information; Based on the sample channel information, training data is constructed; The channel model is trained using the training data.

[0029] This application also provides a digital twin channel prediction device based on a large channel model, including a memory, a processor, and a program stored in the memory and executable on the processor; when the processor executes the program, it implements the above-described digital twin channel prediction method based on a large channel model.

[0030] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the steps in the above-described digital twin channel prediction method based on a large channel model.

[0031] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the above-described digital twin channel prediction method based on a large channel model.

[0032] The beneficial effects of the above technical solution in this application are as follows: In the above scheme, the digital twin channel prediction method based on a large channel model utilizes a digital twin channel architecture to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area. Using the digital twin channel architecture, electromagnetic propagation parameter information is extracted from the three-dimensional wireless environment. This electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment. The scatterers are other objects in the three-dimensional wireless environment besides the transceivers. Using the large channel model in the digital twin channel architecture, channel information between the transceivers is predicted based on the transceiver location information, scatterer distribution information, and penetration characteristic information. This channel information includes large-scale channel information and small-scale channel information. The information refers to channel information whose sensitivity to environmental changes is less than a first threshold. The small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold. It can support the realization of real-time channel prediction based on environmental adaptation, improve the poor environmental adaptability and real-time performance of the scheme, and the prediction of channel information based on electromagnetic propagation parameter information can suppress the interference of irrelevant environmental changes on channel prediction, avoid the problem of key feature overload and weak cross-scenario generalization caused by the model directly processing the original data. Moreover, based on the location information of the transceiver device, the distribution information of the scatterer, and the penetration characteristic information, it can collaboratively predict multiple types of channel information, support the realization of real-time joint prediction of multiple channel tasks to improve the joint prediction capability of multiple tasks, and thus solve at least one of the problems of poor environmental adaptability, insufficient real-time performance, and weak joint prediction capability of existing channel information prediction schemes. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the digital twin channel prediction method based on a large channel model according to an embodiment of this application; Figure 2 This is a schematic diagram of a real-time digital twin channel implementation method based on a large channel model according to an embodiment of this application; Figure 3 This is a schematic diagram of environment reconstruction based on multimodal perception, as described in an embodiment of this application. Figure 4 This is a schematic diagram illustrating the location of the transceiver device, the distribution of scatterers, and the extraction of penetration features in an embodiment of this application. Figure 5 This is a schematic diagram of a channel multi-task joint prediction network architecture based on a large channel model, according to an embodiment of this application. Figure 6 This is a schematic diagram of the path loss PL graph prediction branch network architecture in an embodiment of this application; Figure 7This is a schematic diagram of the CSI matrix prediction branch network architecture according to an embodiment of this application; Figure 8 This is a schematic diagram comparing different PL prediction algorithms in two test scenarios of an embodiment of this application; Figure 9 This is a schematic diagram of the structure of a digital twin channel prediction device based on a large channel model according to an embodiment of this application; Figure 10 This is a schematic diagram of the structure of a digital twin channel prediction device based on a large channel model, according to an embodiment of this application. Detailed Implementation

[0034] To make the technical problems, technical solutions and advantages of this application clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0035] The following is a brief introduction to the relevant content of this plan.

[0036] Currently, the International Telecommunication Union Radiocommunication Sector (ITU-R) has defined typical scenarios for Sixth Generation (6G) mobile communication, including ubiquitous connectivity and the integration of communication with Artificial Intelligence (AI), and has proposed stringent air interface performance indicators such as peak data rates of terabit-per-second (Tbps). This poses a significant challenge to accurate real-time wireless channel acquisition. However, channel acquisition methods based on traditional statistical modeling essentially rely on offline acquisition of Channel Impulse Response (CIR) data in limited, predefined typical scenarios such as urban macrocells and indoor hotspots, and then use this data to fit the statistical distribution of random parameters such as delay and angle. Its core limitation lies in its insufficient environmental generalization ability, making it difficult to cope with scene switching problems caused by dynamic environmental changes.

[0037] Digital Twin Channel (DTC), as a novel channel acquisition paradigm, collects, processes, and extracts features from environmentally sensed data, continuously interacting with the communication environment to achieve real-time prediction and dynamic evolution of channel fading states. This enables communication systems to proactively adapt to environmental changes. Compared to traditional statistical channel modeling, the DTC paradigm can overcome the constraints of typical scenarios such as urban macrocells and indoor hotspots, and can adapt to more complex and diverse real-world communication environments. Furthermore, its online update mechanism based on real-time environmental perception overcomes the inherent limitations of traditional offline statistical fitting methods.

[0038] High-fidelity environmental information acquisition and reconstruction can be achieved by scheduling multimodal sensing devices such as red-green-blue (RGB) cameras, depth cameras, and LiDAR (Light Detection and Ranging), as well as technologies such as Integrated Sensing and Communication (ISAC). Meanwhile, small-scale AI models such as U-shaped networks (U-Nets) and large-scale AI models such as Large Language Models (LLMs) can also be applied to the realization of digital twin channel paradigms. However, channel prediction methods based on small-scale AI models are typically only applicable to fixed environments or single prediction tasks, with limited generalization capabilities. Mainstream research based on large-scale AI models still focuses on the direct processing of channel data, leading to the overloading of key features and weak cross-scenario generalization capabilities.

[0039] Based on the above, this application addresses at least one of the problems existing in current channel information prediction schemes, namely poor environmental adaptability, insufficient real-time performance, and weak multi-task joint prediction capability. It provides a digital twin channel prediction method (a method for digital twin channel prediction information) based on a large channel model. Figure 1 As shown, it includes: Step 11: Using a digital twin channel architecture, construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area; The multimodal sensing data can be multimodal sensing data of the physical environment within the target area, such as LiDAR point cloud data and RGB-D images; the target area can be the target communication area, but is not limited to this. Furthermore, the digital twin channel architecture can include... Figure 5 The prediction network architecture shown may also include related architectures that implement steps 11 and 12, which are not limited here.

[0040] Step 12: Using the digital twin channel architecture, extract electromagnetic propagation parameter information for the three-dimensional wireless environment; the electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment; the scatterer is other objects in the three-dimensional wireless environment besides the transceiver. Step 13: Using the large-scale channel model in the digital twin channel architecture, predict the channel information between the transceiver devices based on the transceiver device location information, scatterer distribution information, and penetration characteristic information; the channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to the channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to the channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

[0041] Small-scale channel information may include channel information at the location of the receiving device. The first threshold can be obtained based on experience or relevant parameter information and is not limited here. Furthermore, a large-scale channel model can refer to a large-scale parameterized neural network model based on a pre-trained large-scale model architecture, specifically designed and optimized for wireless channel characteristics (such as time-varying nature, frequency selectivity, spatial correlation, etc.) through transfer learning or domain adaptation techniques to achieve channel state information prediction, estimation, or generation, but it is not limited to this. The large-scale channel model can achieve, for example... Figure 5 The large model backbone network shown may include, for example, a network containing GPT (Generative Pre-trained Transformer)-2, or GPT-2, without limitation. Additionally, the large channel model may also be referred to as a prediction model, or may be included in the prediction model within the digital twin channel architecture, without limitation.

[0042] The digital twin channel prediction method based on a large channel model provided in this application utilizes a digital twin channel architecture to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data of the target area. Using the digital twin channel architecture, electromagnetic propagation parameter information is extracted from the three-dimensional wireless environment. This electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment. The scatterers are objects in the three-dimensional wireless environment other than the transceivers. Using the large channel model in the digital twin channel architecture, channel information between the transceivers is predicted based on the transceiver location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. Channel information refers to channel information whose sensitivity to environmental changes is less than a first threshold, while small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold. This approach supports real-time channel prediction based on environmental adaptation, improving the scheme's poor environmental adaptability and real-time performance. Furthermore, predicting channel information based on electromagnetic propagation parameters can suppress interference from irrelevant environmental changes, avoid key feature overload and weak cross-scenario generalization caused by the model directly processing raw data. Moreover, by collaboratively predicting multiple types of channel information based on transceiver location information, scatterer distribution information, and penetration characteristic information, it can support real-time joint prediction of multiple channel tasks, improving multi-task joint prediction capabilities. This addresses at least one of the problems in existing channel information prediction schemes: poor environmental adaptability, insufficient real-time performance, and weak multi-task joint prediction capabilities.

[0043] The multimodal sensing data includes LiDAR point cloud data and RGB-D images. The construction of a 3D wireless environment corresponding to the target area based on the multimodal sensing data includes: constructing a static scatterer in the target area based on the LiDAR point cloud data; constructing a dynamic scatterer in the target area based on the LiDAR point cloud data and the RGB-D image; and constructing the 3D wireless environment corresponding to the target area based on the static and dynamic scatterers. This supports a separate reconstruction strategy for static and dynamic scatterers, fusing multimodal sensing data such as LiDAR point clouds and RGB-D images to construct a high-precision 3D wireless environment, enabling real-time online updates for complex dynamic scenes. Furthermore, the update of dynamic scatterers based on multimodal sensing wireless environment reconstruction effectively improves the real-time performance of the solution. Static scatterers can include scatterers that change only slightly over time, such as buildings, roads, and vegetation; dynamic scatterers can include scatterers that change significantly over time, such as pedestrians and vehicles. The phrase "constructing a three-dimensional wireless environment corresponding to the target area based on the static and dynamic scatterers" can include: using the positioning information of each dynamic scatterer to align its coordinates with the static scatterers, thus constructing a three-dimensional wireless environment corresponding to the target area, achieving real-time reconstruction and dynamic updating of complete three-dimensional wireless environment information, but is not limited to this. It should be noted that in this scheme, a single channel information prediction can use dynamic scatterer-related information at a certain moment, and channel information predictions at different moments can be performed based on dynamic scatterer-related information at the corresponding moment, but is not limited to this. For example, channel information prediction can be performed based on the average or weighted value of dynamic scatterer-related information from at least two moments (e.g., two consecutive moments).

[0044] In this embodiment, constructing a dynamic scatterer in the target region based on the LiDAR point cloud data and the RGB-D image includes: determining the reference coordinates of the dynamic scatterer in the target region based on the RGB-D image; performing cluster analysis on the LiDAR point cloud data in the neighborhood of the dynamic scatterer based on the reference coordinates to determine the center coordinates and boundary information of the dynamic scatterer; and constructing the dynamic scatterer in the target region based on the center coordinates and boundary information. This allows for the implementation of a hierarchical fusion localization mechanism between the RGB-D image and the LiDAR point cloud for the dynamic scatterer, accurately constructing the dynamic scatterer.

[0045] The step of extracting electromagnetic propagation parameter information for the three-dimensional wireless environment includes: rasterizing the target area in the horizontal direction to obtain a two-dimensional raster coordinate system; and obtaining electromagnetic propagation parameter information for the three-dimensional wireless environment based on the two-dimensional raster coordinate system. This allows for simple and accurate extraction of electromagnetic propagation parameter information.

[0046] In this embodiment, obtaining the transmittance feature information based on the two-dimensional grid coordinate system includes: determining whether a line-of-sight link exists between the receiving device and the transmitting device; the number of receiving devices is at least one; if it exists, the grid position corresponding to the receiving device in the two-dimensional grid coordinate system is taken as a first value; if it does not exist, the transmittance corresponding to the grid position is determined according to the three-dimensional coordinates of the receiving device and the three-dimensional coordinates of the first position and the second position on the first connecting line; the first connecting line refers to the line connecting the receiving device and the transmitting device, the first position refers to the intersection point of the first connecting line and the scatterer that is closest to the transmitting device, and the second position refers to the intersection point of the first connecting line and the scatterer that is closest to the receiving device; the transmittance feature information is obtained according to the first value and / or the transmittance. This allows for a more specific and accurate acquisition of the transmittance feature information.

[0047] The step of predicting channel information between transceiver devices based on the transceiver device location information, scatterer distribution information, and penetration characteristic information includes: obtaining global environmental features based on the transceiver device location information, scatterer distribution information, and penetration characteristic information; and performing mapping processing based on the global environmental features to predict the channel information between the transceiver devices. This allows for the precise and accurate acquisition of channel information.

[0048] Furthermore, the small-scale channel information includes Channel State Information (CSI); the method further includes: acquiring a portion of the CSI corresponding to the receiving device based on a number of probe signals transmitted by the transmitting device (less than a first number); wherein, predicting the CSI between the transmitting and receiving devices by performing mapping processing based on the global environmental features includes: acquiring CSI reference features based on the portion of the CSI; extracting local environmental features corresponding to the receiving device from the global environmental features; fusing the CSI reference features and the local environmental features and then performing mapping processing to predict the CSI between the transmitting and receiving devices. This allows for accurate acquisition of the CSI. The CSI reference features may include, but are not limited to, the characteristic linear modulation (FiLM) features of the CSI.

[0049] In this embodiment of the application, the digital twin channel prediction method further includes: acquiring original wireless channel information corresponding to each receiving location for a sample region used for model training; the sample region includes at least one area of ​​a scene; extracting channel information corresponding to each receiving location in the sample region based on the original wireless channel information, as sample channel information; constructing training data based on the sample channel information; and training the large-scale channel model using the training data. This allows for the training of a large-scale channel model with high prediction accuracy. The "receiving location" can be the location of the receiving device in the sample region, but is not limited to this.

[0050] The following example illustrates the digital twin channel prediction method based on a large channel model provided in this application, wherein the method can be implemented based on a digital twin channel architecture.

[0051] To address the aforementioned technical issues, this application provides a digital twin channel prediction method based on a Channel Large Model (ChannelLM). Specifically, it can be implemented as a real-time digital twin channel prediction method based on a Channel Large Model (ChannelLM). This solution mainly involves three core modules: a multimodal perception-based environment reconstruction module, an electromagnetic propagation-guided environment feature extraction module (which can extract electromagnetic propagation characteristic parameters), and a channel large model-driven digital twin prediction module. The multimodal perception-based environment reconstruction module employs a hierarchical reconstruction strategy, decoupling the wireless environment into static and dynamic components. It achieves real-time updates of sparse dynamic objects by fusing Red-green-blue-depth (RGB-D) images with LiDAR (Light Detection and Ranging) point cloud data, thus avoiding problems such as poor environmental adaptability and insufficient real-time reconstruction across all scenarios. The electromagnetic propagation-guided environment feature extraction module specifically extracts transceiver location, scatterer distribution, and penetration characteristics to suppress interference from irrelevant environmental changes on channel prediction, avoiding the problems of key feature overload and weak cross-scenario generalization caused by large-scale AI models directly processing raw data. The channel-scale model-driven digital twin prediction module can capture the global structure and local changes of the environment using a multi-scale block design. It learns stable representations of interpretable features (such as the information extracted by the environmental feature extraction module) using a pre-trained large model as the core. Furthermore, it designs a collaborative prediction branch to address the lack of multi-task collaboration mechanisms in existing methods, making it difficult to jointly predict large-scale fading information (or large-scale channel information; i.e., channel information insensitive to environmental changes, such as path loss (PL)) and small-scale channel information (i.e., channel information sensitive to environmental changes, such as channel state information (CSI)). Therefore, this application can utilize a unified architecture to achieve real-time joint prediction of channel multi-tasks in dynamic scenarios, breaking through the core bottleneck of implementing the digital twin channel paradigm. Alternatively, this can be understood as follows: the multi-modal perception-based environment reconstruction module can acquire a high-precision three-dimensional wireless environment; the electromagnetic propagation-guided environment feature extraction module can extract the location of transceivers, scatterer distribution, and penetration characteristics based on the aforementioned three-dimensional wireless environment; and finally, the channel-scale model-driven digital twin prediction module can achieve real-time joint prediction of channel multi-tasks. These three components work together to fully realize the entire process of "environmental perception reconstruction - environmental feature extraction - multi-task joint prediction of wireless channels".

[0052] The following is an example illustrating the solution provided in the embodiments of this application. It should be noted that the following description includes the complete process of model training. When using this solution for information prediction in practice, it may not include the model training-related content such as dataset construction, loss function, and performance evaluation. It may include: environment reconstruction, extraction of the location of the transceiver device, the distribution of scatterers, and the penetration characteristics, and information prediction based on the extracted content. Of course, the model can also be retrained to update it to ensure the accuracy of model prediction when the conditions are met. That is, it is not a restriction that this solution will not update the relevant model after the actual prediction is performed.

[0053] like Figure 2 As shown, the implementation process of the real-time digital twin channel implementation method based on the large channel model may include the following steps: Step 1: Based on multimodal sensing devices such as cameras and LiDAR, perform collaborative sensing and data fusion of the physical environment (of the target area) to construct a high-precision 3D wireless environment. That is, achieve environment reconstruction based on multimodal sensing. This step corresponds to the above-mentioned construction of the 3D wireless environment corresponding to the target area based on multimodal sensing data for the target area.

[0054] This step may include: 1) Selecting a target communication area (i.e., the target region) and using LiDAR point clouds to offline construct a high-precision 3D static scattering model of buildings, roads, vegetation, etc., which changes only slightly over time. This part of the operation corresponds to the above-mentioned construction of static scattering bodies in the target region based on the LiDAR point cloud data.

[0055] 2) In real time, the dynamic scattering bodies (models) of vehicles, pedestrians, etc., that change rapidly over time are perceived and reconstructed by fusing RGB-D images and LiDAR point cloud data. This part of the operation corresponds to the above-mentioned construction of dynamic scattering bodies in the target area based on the LiDAR point cloud data and RGB-D images.

[0056] 3) Utilizing the precise positioning information of each dynamic scatterer, their coordinates are aligned with those of the static scatterer to achieve real-time reconstruction and dynamic updating of the complete 3D wireless environment information. Specifically, this enables the construction of an environment for a specific scene at a given moment. This operation corresponds to the above-mentioned construction of the 3D wireless environment corresponding to the target area based on the static and dynamic scatterers.

[0057] For example, in this step, such as Figure 3 As shown, a high-precision static scatterer model (i.e., a static environment model) can first be constructed offline using LiDAR point cloud data. Then, RGB-D images and LiDAR point cloud data are fused in real time to perceive and construct dynamic scatterers. Finally, the two types of scatterers are precisely stitched together according to their spatial positions to form a complete three-dimensional wireless environment. Details are as follows: 1) For static scatterers such as buildings, roads, and vegetation that change only slightly over time, the target communication area can be selected first, and LiDAR can be used to collect the three-dimensional geometry and spatial layout information of all static scatterers in the area; then, the open-source 3D modeling software Blender can be used to reconstruct a high-precision digital model of the static environment.

[0058] 2) For dynamic scatterers such as vehicles and pedestrians that change rapidly over time, a layered progressive localization and reconstruction strategy can be adopted: In the coarse localization stage, based on the RGB images acquired by the RGB-D camera, the YOLOv11 algorithm is used to identify the scatterer category and mark the two-dimensional pixel area with a rectangular bounding box; using the mapping relationship between RGB pixels and three-dimensional coordinates, the approximate three-dimensional center coordinates and general orientation of the target within each bounding box are estimated to achieve preliminary localization based on the RGB-D camera (which can correspond to the above-mentioned determination of the reference coordinates of the dynamic scatterers in the target area based on the RGB-D image). In the fine localization stage, using the aforementioned three-dimensional center coordinates as reference points, a density-based spatial clustering algorithm (DBSCAN) is employed to select high-precision LiDAR point clouds within the target's (corresponding) neighborhood and perform cluster analysis to extract the precise three-dimensional center position and spatial boundary of the dynamic object (i.e., the target within the aforementioned box). Further, after the fine localization stage, a geometric simplification stage can be entered. Based on the clustered point cloud data, the center points and normal vectors of each surface of the bounding box (formed by the spatial boundaries) are calculated to estimate the orientation of the dynamic scatterer. Based on the orientation, the dynamic scatterer is abstracted into a cuboid geometric model (the center of the cuboid geometric model can be used as the fine localization center), achieving a balance between computational efficiency and model accuracy. Corresponding to the aforementioned cluster analysis performed on the LiDAR point cloud data within the neighborhood of the dynamic scatterer based on the reference coordinates, the center coordinates and boundary information of the dynamic scatterer are determined.

[0059] 3) Utilizing the precise positioning information of each dynamic scatterer (such as the center and boundaries of the aforementioned cuboid geometric model), align its coordinates with the static scatterer (perform dynamic target insertion) to achieve real-time reconstruction and dynamic updating of the complete three-dimensional wireless environment information, thus obtaining the reconstructed environment. The construction of the dynamic scatterer in this part corresponds to the construction of the dynamic scatterer in the target region based on the aforementioned center coordinates and boundary information.

[0060] Step 2: Using the 3D wireless environment reconstructed in Step 1, obtain the corresponding wireless channel database through simulation or actual measurement to provide label data for subsequent network training (which can correspond to training the large channel model using the training data mentioned above). That is, to realize the construction of a wireless channel dataset based on the reconstructed environment.

[0061] This step may include: 1) Channel data acquisition: Deploying transceiver equipment in the high-precision 3D wireless environment reconstructed in step 1, and acquiring raw wireless channel information at different receiving locations under various scenarios (such as multiple scenarios at multiple times (e.g., teaching buildings, basketball courts, etc.)) through simulation or actual measurement. During model training, the region in step 1 can be used as a sample region for model training. This part of operation 1) can correspond to the above-mentioned acquisition of raw wireless channel information corresponding to each receiving location for the sample region used for model training; the sample region includes at least one scene region.

[0062] 2) Channel Feature Extraction: Based on the aforementioned raw data (i.e., raw wireless channel information), large-scale fading features (i.e., large-scale channel information, such as path loss, beam index, etc.) and small-scale fading features (i.e., small-scale channel information, such as phase, complete channel state information matrix, etc.) are extracted respectively, constructing a database covering multi-scale channel characteristics. This part of the operation corresponds to the above-mentioned extraction of channel information corresponding to each receiving location in the sample area based on the raw wireless channel information, as sample channel information; and the construction of training data based on the sample channel information.

[0063] For example, in this step, the 3D wireless environment model reconstructed in step 1 is imported into the commercial ray tracing software Wireless Insite. Then, the transceiver parameters are configured to obtain multipath channel parameters. Based on these parameters, a complete CSI matrix and PL diagram are obtained (the complete CSI matrix and PL diagram are an example of the aforementioned channel information). Specifically, as follows: 1) The 3D wireless environment model reconstructed in step 1 can be imported into the commercial ray tracing software Wireless Insite, and simulation parameters can be configured, such as: deploying at a height of 2 meters above the ground at the center of the environment (i.e., the center of the target area). A uniform linear array of 64 antennas is used as the transmitting device. Single-antenna receiving devices are uniformly deployed at 0.1-meter intervals at a height of 1 meter above the ground (across the entire target area). The communication carrier frequency is set to 7.5 GHz, and the subcarrier spacing is... kHz, number of subcarriers The transmit power is 0 dBm, and the receive power threshold is -250 dBm. The ray tracing engine can be X3D, with a maximum multipath number of 25, a maximum reflection order of 6, and a maximum diffraction order of 1. In this way, multipath channel parameters for each receiving (device) location in various scenarios can be obtained through ray tracing simulation. These parameters may include at least one of the following: multipath gain, transmit angle, time delay, and phase.

[0064] 2) Based on the above multipath channel parameters, the classic Saleh-Valenzuela channel model can be used to construct a complete Multiple-input Single-output Orthogonal Frequency-division Multiplexing (MISO-OFDM) CSI matrix for each receiver (such as a single-antenna receiver). ,as follows: ; ; ; in, Indicates the first The MISO (Multiple-Input Single-Output) channel vectors corresponding to each subcarrier, k=1···K, where K represents the total number of subcarriers and L represents the total number of paths; , as well as They represent the first The complex gain, time delay, and emission angle of each path; the complex gain can be obtained by combining the multipath gain and phase; e represents the base of a natural number. Indicates the subcarrier spacing. This indicates the first subcarrier of the k-th subcarrier. The steering vector of the antenna array with a single path, where j represents the imaginary unit. It represents the antenna array spacing (i.e., the distance between two adjacent antenna elements in a uniform linear array, which can be half the wavelength of the communication carrier frequency signal). Indicates the first The wavelengths corresponding to each subcarrier. This indicates the number of antennas, and T indicates transpose.

[0065] Based on the obtained CSI matrix Calculate the PL value at each sampling location (i.e., the location of the receiving device): ; in, The square of the F-norm of a matrix is ​​given by its expression. The PL value is expressed in dB (decibels). Then, a complete PL map of the entire region is generated using the PL values ​​at each sampling location. Furthermore, a data augmentation strategy combining cropping and rotation can be used to expand the dataset. This involves randomly cropping and rotating the original PL map at multiple angles (e.g., 90°, 180°, 270°) to improve data diversity while maintaining environmental structural consistency, thus constructing a large-scale PL map training sample set.

[0066] Step 3: Extract the location of the transceiver device (i.e., the transceiver end), the distribution of scatterers, and the penetration feature map from the three-dimensional wireless environment in Step 1. That is, extract the location of the transceiver end, the distribution of scatterers, and the penetration feature. This step corresponds to the extraction of electromagnetic propagation parameter information for the three-dimensional wireless environment mentioned above; the electromagnetic propagation parameter information includes: the location information of the transceiver device in the three-dimensional wireless environment, the distribution information of scatterers in the three-dimensional wireless environment, and the penetration feature information corresponding to the three-dimensional wireless environment; the scatterers are other objects in the three-dimensional wireless environment besides the transceiver device.

[0067] This step may include: 1) Spatial rasterization: Dividing the target communication area selected in step 1 into uniform grids with a specified horizontal spatial resolution to construct a two-dimensional pixel coordinate system (i.e., a two-dimensional raster coordinate system) as a unified spatial representation framework for various feature maps. This operation corresponds to the above-mentioned horizontal rasterization of the target area to obtain a two-dimensional raster coordinate system. Based on the two-dimensional raster coordinate system, obtaining electromagnetic propagation parameter information for the three-dimensional wireless environment may include the following operations 2) to 4).

[0068] 2) Extraction of transceiver location map: Mark the pixel positions of the transmitter and receiver (of the probe signal) as 1, and set the remaining background pixels to 0 to generate a binarized transceiver location distribution map.

[0069] 3) Scatter distribution map extraction: Project each scatterer (such as all objects in the environment except for the transceiver) onto a two-dimensional pixel coordinate system, set the pixel value of the scatterer to the physical height of the scatterer, and set the region without scatterers to 0, thus constructing a scatterer distribution map.

[0070] 4) Transmittance Map Extraction: For each received pixel location, determine whether there is a line-of-sight link between it and the transmitter (i.e., whether there are no other objects on the transceiver connection line). If it exists, the pixel value is set to 0 (an example of the first value mentioned above); if it does not exist, calculate the cumulative length of the transceiver connection line penetrating the building (e.g., based on the three-dimensional position of the signal entering the building (closest to the receiving device) and the three-dimensional position of the signal exiting the building (closest to the transmitting device), and calculate the penetration loss coefficient (i.e., transmittance) accordingly to generate a transmittance map. This part of the operation corresponds to determining whether a line-of-sight link exists between the receiving device and the transmitting device; the number of receiving devices is at least one; if it exists, the grid position corresponding to the receiving device in the two-dimensional grid coordinate system is taken as a first value; if it does not exist, the transmittance corresponding to the grid position is determined according to the three-dimensional coordinates of the receiving device and the three-dimensional coordinates of the first position and the second position on the first connecting line; the first connecting line refers to the connecting line between the receiving device and the transmitting device, the first position refers to the intersection point of the first connecting line and the scatterer that is closest to the transmitting device, and the second position refers to the intersection point of the first connecting line and the scatterer that is closest to the receiving device; the transmittance feature information is obtained according to the first value and / or the transmittance.

[0071] For example, in this step, such as Figure 4 As shown, extracting transceiver locations, scatterer distributions, and transmittance feature maps (such as transceiver location maps, scatterer distribution maps, and transmittance maps) closely related to the electromagnetic signal propagation characteristics from the reconstructed 3D wireless environment in step 1 can also be understood as performing environmental feature extraction. The specific steps are as follows: 1) Spatial rasterization: The target communication area (i.e. the area corresponding to the three-dimensional wireless environment) selected in step 1 is uniformly divided into raster with a horizontal spatial resolution of 0.1 meters to construct a two-dimensional pixel coordinate system as a unified spatial representation framework for various feature maps.

[0072] 2) Extraction of transceiver device location map (i.e., transceiver end location map): Mark the pixel positions of the transmitting and receiving ends as 1, and set the remaining background pixels to 0 to generate a binarized transceiver device location distribution map: ; in Represents any position in a two-dimensional pixel coordinate system. This indicates the position of the transmitter or receiver in a two-dimensional pixel coordinate system. The location diagram of the transceiver equipment is shown in the location. The value at that location. Figure 4 The diagram showing the locations of the transmitter and receiver ends only indicates the location of the transmitter.

[0073] 3) Scatterer distribution map extraction: Project each scatterer onto a two-dimensional pixel coordinate system, set the pixel value of the scatterer to its physical height, and set the value of regions without scatterers to 0, thus constructing a scatterer distribution map:

[0074] in, Represents position in a two-dimensional pixel coordinate system The values ​​of the scatterer distribution map at the location, Represents position in a two-dimensional pixel coordinate system The value of the scatterer height (i.e., the maximum height) ), This represents the pixel region in a two-dimensional pixel coordinate system where the scatterer exists.

[0075] 4) Penetration Map Extraction: For each receiving pixel location, determine whether a line-of-sight link (i.e., a direct link) exists between it and the transmitting end; this can also be understood as determining whether there are any obstructions between the receiving device and the transmitting device. If such an obstruction exists, the pixel value is set to 0; otherwise, the cumulative length of the line of communication between the transceiver devices penetrating the building is calculated. Based on this, the penetration loss coefficient (i.e., penetration rate) is calculated, and a penetration rate map is generated:

[0076] in, Represents position in a two-dimensional pixel coordinate system The values ​​of the transmittance map at that location, and Representing the connection between the transmitter and the position, respectively. The three-dimensional coordinates of the intersection point of the scatterer closest to the transmitter on the straight line (i.e., the coordinates of the intersection point of the transceiver connection line with the scatterer that is closest to the transmitter, which can correspond to the first position mentioned above), and the distance position. The three-dimensional coordinates of the nearest scatterer intersection point (i.e., the coordinates of the intersection point between the transceiver connection and the scatterer that is closest to the receiver, which can correspond to the second position mentioned above) can be found in the reference section. Figure 4 .also, Represents the 2-norm operation for vectors. Position. The receiver position can be set at any position in the two-dimensional pixel coordinate system.

[0077] Step 4: Construct a multi-task joint prediction network based on a large channel model. Use the wireless channel database built in Step 2 as supervised labels to train the network, learning the mapping relationship from environmental features to channel parameters. That is, construct and train a multi-task joint prediction network based on a large channel model. The channel information described below uses PL and CSI as examples, but is not limited to these.

[0078] This step may include: 1) Inputting the transmitter location map, scatterer distribution map, and penetration map extracted in step 3 into a pre-trained large model (such as a large language model, a large vision model, etc.), and generating global environmental features characterizing the wireless propagation environment through deep feature extraction. This part of the operation corresponds to the above-mentioned acquisition of global environmental features based on the transceiver location information, scatterer distribution information, and penetration feature information.

[0079] 2) Using the extracted global environment features as input, corresponding prediction taps are constructed for different channel prediction tasks to achieve multi-task joint prediction of the channel. This part of the operation corresponds to the mapping process based on the global environment features described above, which predicts the channel information between the transceiver devices.

[0080] 3) For different channel prediction tasks, select appropriate loss functions to minimize the difference between the prediction results and the true labels during the training process.

[0081] 4) Divide the wireless channel database constructed in step 2 into training set, validation set and test set according to an appropriate ratio; perform iterative optimization of network parameters based on the training set, use the validation set to monitor overfitting and select the optimal model parameters, and finally obtain a converged channel large model multi-task joint prediction network.

[0082] For example, in this step, such as Figure 5 As shown, a pre-trained large model (i.e. Figure 5 The core network is a large-scale model backbone network (such as a large language model or a large vision model). Subsequently, PL map prediction taps (i.e., large-scale PL prediction taps) and CSI matrix prediction taps (i.e., small-scale CSI prediction taps) are connected to construct a multi-task joint prediction network architecture. Then, specific loss functions are selected for the PL map and CSI matrix prediction tasks, and the network is trained using the wireless channel database constructed in step 2. Before connecting the CSI matrix prediction taps, local features can be extracted from the global environment features output by the pre-trained large model based on the receiver location map, and fused with CSI features extracted from an incomplete CSI matrix obtained from a small number of pilots. The CSI matrix prediction taps are then connected. The specific steps are as follows: 1) Global environment feature extraction based on pre-trained large models: such as Figure 6 As shown, this scheme uses GPT (Generative Pretrained Transformer)-2 as the core architecture for pretraining large models (i.e., Figure 5The backbone network of the Zhongda model (taking GPT-2 as an example) is used to extract global environmental features of the scene. Specifically, firstly, a three-channel environmental feature map (representing the emitter position, scatterer distribution, and penetration rate, respectively) is input into the improved GPT-2 network. Then, the network sequentially performs convolution operations and layer normalization on the input feature map, and uses sliding windows of 4×4 and 8×8 sizes for multi-scale segmentation to obtain local feature blocks of corresponding scales (i.e., 4×4 and 8×8 feature blocks). Next, the multi-scale feature blocks are linearly mapped, and positional and segmental encoding information is superimposed before being input into the GPT-2 model for deep feature learning. Finally, the output of the GPT-2 model is linearly transformed and normalized, and multi-scale information is fused to obtain the final global environmental features. The use of sliding windows of 4×4 and 8×8 sizes for multi-scale segmentation allows for the simultaneous capture of fine-grained geometric details and global spatial context. The 4×4 size preserves subtle local changes, while the 8×8 size encodes a wider range of spatial context information. The "improved GPT-2 network" refers to the network structure formed by adding other network structures before the input and after the output of the GPT-2 network, for example... Figure 6 The network structure shown.

[0083] 2) Construction of PL diagram prediction taps: such as Figure 6 As shown, this scheme designs a two-layer convolutional structure to map the aforementioned global environment features to the final PL map. The first layer of the two-layer convolutional structure uses a 5×5 convolutional kernel for spatial smoothing; the second layer uses a 3×3 convolutional kernel to map the output of the first layer to a single-channel output. Furthermore, both convolutional layers maintain constant spatial resolution through zero padding, ensuring that the decoding result is strictly aligned with the input environment features at the pixel level.

[0084] CSI prediction branch network construction: such as Figure 7 As shown, the CSI prediction branch network can be divided into four parts: a CSI feature extraction network, a receiver local environment feature extraction network, a feature fusion network, and a CSI prediction tap. Details are as follows: a) CSI feature extraction network: such as Figure 7 As shown, the incomplete CSI matrix obtained from a small number of pilots is first processed. The real and imaginary parts are separated to construct a dual-channel input feature map; then, a proximal gradient unrolling algorithm based on a convolutional neural network is used to iteratively extract CSI latent features (i.e., perform CSI feature extraction). This could be achieved by transmitting a small number of probe signals to obtain a small number of pilot signals (corresponding to the above-mentioned acquisition of a portion of the CSI corresponding to the receiving device based on less than a first number of probe signals sent by the transmitting device); regarding CSI feature extraction, the first... The iteration process can be represented as: ; ; in, and They represent the first and the CSI latent features obtained in the next iteration; This is the first step in the proximal gradient iteration process. The intermediate value obtained from the next iteration; and They are the first The near-end mapping and stride at each iteration are modeled using two-layer and single-layer convolutional neural networks, respectively. The binary sampling mask matrix and the incomplete CSI matrix The value at the position of a non-zero element is 1, otherwise it is 0; This represents the Hadamard product. For example, in this scheme, the number of iterations can be... Set to 3, corresponding for .

[0085] b) Receiver-side local environment feature extraction network: used to characterize the local propagation characteristics around a specific receiver, improving CSI prediction accuracy. For example... Figure 5 and Figure 7 As shown, Figure 6 Improved GPT-2 network ( Figure 5 (An example of a large-scale model backbone network) The global environment features extracted in step 3 are combined with the receiver location map extracted in step 3 to form a multi-channel two-dimensional feature map, and then a deep convolutional neural network is used to extract the local environment features of the receiver; this can be included in the above extraction of the global environment features to obtain the local environment features corresponding to the receiving device.

[0086] c) Feature Fusion Network: Input the multi-channel two-dimensional feature map into the linear layer to generate antennas (dimensions). ) and subcarrier (dimension) Feature-wise Linear Modulation (FiLM) parameters: Antenna-FiLM, Subcarrier-FiLM. CSI latent features obtained based on the CSI feature extraction network. Conditional proximal gradient iteration based on convolutional neural networks is performed on antenna-FiLM and subcarrier-FiLM, where the CSI latent features are obtained by the CSI feature extraction network. Conditional modulation can be expressed as: ; ; in, and They represent the first The second iteration and the first The FiLM features of CSI obtained in the next iteration; It is the intermediate value in the conditional proximal gradient iteration process; and They are the first The near-end mapping and stride at each iteration are modeled using two-layer and single-layer convolutional neural networks, respectively. and They represent the first During the subconditional near-end gradient iteration process, features at different antenna-subcarrier positions are controlled. The relative importance weights and the bias parameters corresponding to the local environmental features of the receiver, The binary sampling mask matrix and the incomplete CSI matrix The value at the position of a non-zero element is 1, otherwise it is 0; This represents the Hadamard product. All convolutional layers in the feature fusion network maintain the same feature map spatial dimensions: environment-related convolutional layers use 5×5 kernels with padding of 2, and proximal gradient iteration-related convolutional layers use 3×3 kernels with padding of 1. The approximation condition is the number of proximal gradient iterations. It can be set to 2, which will give you This corresponds to obtaining CSI reference features based on the aforementioned partial CSI. Ultimately, this can be... The output of the feature fusion network is formed by superimposing the local environmental features of the receiving end with those of the receiving end. Specifically, the feature fusion network can be formed by combining the local environmental features of the receiving end of the same receiving device with... To integrate.

[0087] d) CSI prediction taps: A two-layer convolutional neural network can be used to map the output of the feature fusion network to the real and imaginary parts of the final CSI matrix, thereby obtaining the complete CSI matrix. This corresponds to the above-mentioned fusing of the CSI reference features and the local environment features, followed by mapping processing, to predict the CSI between the transceiver devices (such as the CSI matrix at the location of the receiving device). All convolutional layers can use 3×3 kernels padded with 1s to maintain a constant spatial dimension during the decoding process, but this is not a limitation.

[0088] 3) Regarding the loss for PL map prediction, PL map prediction can be regarded as a two-dimensional image reconstruction task, and a hybrid loss function with joint spatial and frequency domain supervision can be designed. :

[0089] in, Characterizing spatial loss, Characterizing frequency domain loss, This represents the total number of PL graphs in the training set; and They represent the first A prediction Pixel PL image and its corresponding real Pixel PL image; Represents the Charbonnier penalty function (x= ), and the constants in this scheme It can be set to 0.001; Represent the two-dimensional Fourier transform equation. The F-norm of a matrix.

[0090] Regarding the CSI matrix prediction loss, the classic mean squared error loss function can be used. : ; in This represents the total number of CSI matrices in the training set; and They represent the first A predicted CSI matrix and its corresponding true CSI matrix.

[0091] 4) The wireless channel database constructed in step 2 can be divided into training, validation, and test sets in a 7:1:2 ratio. To avoid data leakage, the data partitioning can be performed using the original samples as the basic unit: all augmented data obtained by rotating the same sample can be uniformly assigned to the same subset to ensure that there are as few duplicate or highly correlated samples from the same original sample as possible among the training, validation, and test sets. Then, network parameters can be iteratively optimized based on the training set, and the validation loss can be monitored using the validation set to prevent overfitting. The model parameters with the minimum validation loss are selected as the optimal weights, ultimately obtaining a converged multi-task joint prediction network for the large channel model. The hyperparameter configuration during the training process is shown in Table 1 below: Table 1. Hyperparameter configuration for training the multi-task joint prediction network of the large channel model.

[0092] Step 5: Utilize the multi-task joint prediction network based on the large channel model trained in Step 4 to achieve multi-task joint prediction of the channel. This step corresponds to the above-mentioned use of the large channel model in the digital twin channel architecture to predict the channel information between transceivers based on the transceiver location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

[0093] This step may include: 1) obtaining the input data required by the multi-task joint prediction network based on the channel large model trained in step 4 in a test scenario (such as testing based on the test set constructed in step 2).

[0094] 2) Implement multi-task joint prediction using a trained multi-task joint prediction network based on a large channel model.

[0095] For example, in this step, the incomplete CSI matrix obtained from a small number of pilot signals, along with the transceiver location map, scatterer distribution map, and transmittance map extracted based on the 3D wireless environment model, are used as inputs to the large channel model network in step 4 to achieve joint prediction of the PL map and CSI matrix. The specific steps are as follows: 1) In the test scenario, sparse pilot signals are deployed at the target receiver location (e.g., using a 1 / 2 uniform sampling mode in the antenna dimension and a 1 / 4 uniform sampling mode in the subcarrier dimension). The partial observation CSI matrix is ​​estimated based on the received pilot signals. The initial observations are used as channel state information. Then, based on the real-time reconstructed 3D wireless environment model of the test scenario, three types of environmental feature maps are extracted using the method described in step 3: transceiver location map, scatterer distribution map, and penetration map, to construct a complete representation of the wireless propagation environment.

[0096] 2) Input the above incomplete CSI matrix and the three types of environmental feature maps into the channel large model multi-task joint prediction network trained in step 4. Through the forward propagation calculation of the network, the global PL map of the test scenario and the CSI matrix at the target receiver location are jointly output.

[0097] Step 6: Evaluate the prediction performance.

[0098] This step may include: 1) Selecting appropriate quantitative evaluation indicators for different channel prediction tasks and establishing a comprehensive prediction performance evaluation system.

[0099] 2) Compare the prediction results generated in step 5 with the real labels of the test set obtained in step 4, calculate the specific values ​​of each performance evaluation index, and quantitatively verify the prediction accuracy and generalization ability of the proposed method.

[0100] For example: 1) For the PL diagram prediction task, this solution uses the root mean square error (RMSE) as the evaluation metric, where any predicted PL diagram... Compared to real PL diagrams RMSE (i.e.) ) can be represented as: ; in, This represents the total number of discrete locations in the non-scattering volume region of the selected scene that need to be predicted for the PL; and Representing positions respectively The predicted PL results and the actual PL results are compared. Then, the average RMSE is obtained by statistically averaging the RMSE of all PL graphs in the test set.

[0101] For the CSI matrix prediction task, this scheme uses Normalized Mean Square Error (NMSE) and Squared Generalized Cosine Similarity (SGCS) as evaluation metrics: ; ; in, This indicates the number of samples in the CSI matrix of the test set; and These represent the predicted CSI matrix and the actual CSI matrix, respectively. and They represent and The The CSI vectors corresponding to each subcarrier. K represents the total number of subcarriers. Indicated Conjugate transpose.

[0102] 2) Performance Evaluation of PL Graph Prediction: In the test scenarios, the PL graph prediction method based on the large channel model provided in this application achieved the best performance compared to existing baseline solutions. Quantitative analysis shows that compared with the classic PMNet (PathlossMap Prediction Network), the average RMSE of the large channel model in this application is reduced by 1.43 dB (multi-scenario); compared with the ablation version of the large channel model (W / o GPT-2 network with GPT-2 core removed), the average RMSE is reduced by 1.87 dB (multi-scenario), verifying the effectiveness of the pre-trained large model in extracting features from complex environments. Figure 8 The visualization shows a performance comparison of different PL prediction algorithms in two typical test scenarios. Figure 8 (The image shown is only a partial illustration.)

[0103] CSI matrix prediction performance evaluation: In the test scenario, the channel large model network provided in this application also achieves significant performance improvement compared with the small model that does not introduce global environment feature extraction. The specific quantitative comparison results are shown in Table 2 below.

[0104] Table 2 Comparison of different CSI matrix prediction algorithms in the test scenario

[0105] Furthermore, in accordance with this solution, the system-level real-time performance verification shows that in dynamic scenarios driven by pedestrian movement, the end-to-end prediction latency of the entire process, including environmental perception reconstruction, environmental feature extraction, and digital twin prediction, can be controlled within approximately 70 milliseconds.

[0106] Therefore, this application provides a Channel Large Model architecture for digital twin channels, comprising three core cascaded modules: a multimodal sensing-based environment reconstruction module that can reconstruct a high-precision three-dimensional wireless environment by fusing multiple environmental sensing data; an electromagnetic propagation-guided environment feature extraction module that can extract transceiver location features, scatterer distribution features, and penetration features from the reconstructed three-dimensional wireless environment; and a channel large model-driven digital twin prediction module that can use the above three types of features and a small amount of incomplete measured channel information as input to achieve multi-task joint prediction of the channel. This solution, through the cascaded collaboration of the three modules, constructs a complete theoretical scheme of "environmental sensing reconstruction—environmental feature extraction—multi-task joint prediction of the channel," overcoming the technical bottlenecks of existing schemes in achieving both real-time performance and multi-task joint prediction. Specifically, it involves: (1) Environment reconstruction module based on multimodal perception: adopting a static and dynamic scatterer separation reconstruction strategy, integrating multimodal perception data such as LiDAR point cloud and RGB-D image, constructing a high-precision three-dimensional wireless environment, and realizing real-time online updates of complex dynamic scenes.

[0107] (2) Environmental feature extraction module guided by electromagnetic propagation: Based on the physical mechanism of electromagnetic wave propagation, three key features are extracted from the three-dimensional wireless environment: the location of the transceiver, the distribution of scatterers and the penetration rate. This can effectively establish an interpretable mapping relationship between environmental geometry and channel fading.

[0108] (3) Channel large model driven digital twin prediction module: a deep feature extraction network is constructed with the pre-trained large model as the core, and the aforementioned three types of environmental features and a small amount of incomplete measured channel information are integrated to realize multi-task joint real-time prediction of the channel under a unified network architecture.

[0109] (4) Design of an environment reconstruction method based on multimodal perception: For static scatterers (such as buildings, roads, vegetation, etc.), a high-precision geometric model is reconstructed offline using LiDAR point clouds; for dynamic scatterers (such as vehicles, pedestrians, etc.), a hierarchical fusion localization mechanism of RGB-D images and LiDAR point clouds is implemented—in the coarse localization stage, two-dimensional target detection and three-dimensional coordinate estimation can be achieved based on the YOLOv11 algorithm; in the fine localization stage, the DBSCAN algorithm can be used to cluster LiDAR point clouds to obtain accurate spatial boundaries; and in the geometric simplification stage, dynamic objects can be abstracted into cuboid bounding box models. This strategy ensures the accuracy of environment reconstruction while significantly reducing system latency and improving environmental adaptability by updating only dynamic components.

[0110] (5) An electromagnetic propagation-guided environmental feature extraction method is specifically designed based on the physical mechanism of electromagnetic wave propagation, using three types of interpretable environmental feature extraction algorithms: transceiver location features are characterized by binary mapping to represent the spatial distribution of communication link endpoints; scatterer distribution features are encoded by height projection to encode the environmental geometry; and penetration rate features are used to quantify non-line-of-sight link loss by calculating the cumulative length of the transceiver connection penetrating obstacles. These features can collectively suppress irrelevant environmental noise interference and enhance the model's cross-scenario generalization ability.

[0111] (6) Channel large model driven digital twin prediction network: a dual-scale patch encoder is constructed with a pre-trained large model as the core. Local geometric details are captured by small-scale branches and global spatial context is encoded by large-scale branches to generate high-dimensional environmental semantic features. On this basis, corresponding taps are constructed for different channel prediction tasks, and finally, multi-task joint real-time prediction of the channel is realized under a unified architecture.

[0112] In summary, compared with existing technologies, this solution has the following advantages: 1. Significantly improves the real-time performance and environmental adaptability of wireless environment reconstruction. This solution adopts a hierarchical reconstruction strategy to decouple the wireless environment into static and dynamic components. By fusing RGB-D images and LiDAR point cloud data, it achieves real-time updates of sparse dynamic objects, effectively solving the problems of poor environmental adaptability and insufficient real-time performance of full-scene reconstruction in existing methods, and can quickly respond to changes in dynamic scenes.

[0113] 2. Enhance the effectiveness and cross-scenario generalization ability of channel feature extraction. This scheme extracts transceiver location, scatterer distribution, and penetration features through electromagnetic propagation guidance. This effectively suppresses the interference of irrelevant environmental changes on channel prediction, avoids the problem of key feature overload caused by large-scale AI models directly processing raw data, and significantly improves the model's generalization performance in different scenarios.

[0114] 3. Achieve high-precision joint prediction of large-scale fading and small-scale channel information. This scheme adopts a multi-scale block design to simultaneously capture the global structure and local changes of the environment. It uses a pre-trained large-scale channel model as the core to learn stable representations of interpretable features, and designs a collaborative prediction branch to build a multi-task collaborative mechanism. Finally, it uses a unified architecture to achieve real-time joint prediction of PL graph and CSI matrix in dynamic scenarios, breaking through the core bottleneck of the implementation of the digital twin communication paradigm.

[0115] This application also provides a digital twin channel prediction device based on a large channel model, such as... Figure 9 As shown, it includes: The first construction module 91 is used to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area using a digital twin channel architecture. The first extraction module 92 is used to extract electromagnetic propagation parameter information for the three-dimensional wireless environment using the digital twin channel architecture; the electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment; the scatterer is other objects in the three-dimensional wireless environment besides the transceiver. The first prediction module 93 is used to predict the channel information between the transceiver devices based on the large channel model in the digital twin channel architecture, according to the transceiver device location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to the channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to the channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

[0116] The digital twin channel prediction device based on a large channel model provided in this application utilizes a digital twin channel architecture to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area. Using the digital twin channel architecture, electromagnetic propagation parameter information is extracted for the three-dimensional wireless environment. This electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment. The scatterers are other objects in the three-dimensional wireless environment besides the transceivers. Using the large channel model in the digital twin channel architecture, channel information between the transceivers is predicted based on the transceiver location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. Channel information refers to channel information whose sensitivity to environmental changes is less than a first threshold, while small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold. This approach supports real-time channel prediction based on environmental adaptation, improving the scheme's poor environmental adaptability and real-time performance. Furthermore, predicting channel information based on electromagnetic propagation parameters can suppress interference from irrelevant environmental changes, avoid key feature overload and weak cross-scenario generalization caused by the model directly processing raw data. Moreover, by collaboratively predicting multiple types of channel information based on transceiver location information, scatterer distribution information, and penetration characteristic information, it can support real-time joint prediction of multiple channel tasks, improving multi-task joint prediction capabilities. This addresses at least one of the problems in existing channel information prediction schemes: poor environmental adaptability, insufficient real-time performance, and weak multi-task joint prediction capabilities.

[0117] The multimodal sensing data includes: LiDAR point cloud data and RGB-D images (red, green, blue, and depth). The construction of a three-dimensional wireless environment corresponding to the target area based on the multimodal sensing data includes: constructing a static scatterer in the target area based on the LiDAR point cloud data; constructing a dynamic scatterer in the target area based on the LiDAR point cloud data and the RGB-D image; and constructing the three-dimensional wireless environment corresponding to the target area based on the static and dynamic scatterers.

[0118] In this embodiment of the application, the step of constructing a dynamic scatterer in the target region based on the LiDAR point cloud data and the RGB-D image includes: determining the reference coordinates of the dynamic scatterer in the target region based on the RGB-D image; performing cluster analysis on the LiDAR point cloud data in the neighborhood of the dynamic scatterer based on the reference coordinates to determine the center coordinates and boundary information of the dynamic scatterer; and constructing the dynamic scatterer in the target region based on the center coordinates and boundary information.

[0119] The step of extracting electromagnetic propagation parameter information for the three-dimensional wireless environment includes: rasterizing the target area in the horizontal direction to obtain a two-dimensional raster coordinate system; and obtaining electromagnetic propagation parameter information for the three-dimensional wireless environment based on the two-dimensional raster coordinate system.

[0120] In this embodiment of the application, obtaining the transmittance feature information based on the two-dimensional grid coordinate system includes: determining whether a line-of-sight link exists between the receiving device and the transmitting device; the number of receiving devices is at least one; if it exists, the grid position corresponding to the receiving device in the two-dimensional grid coordinate system is taken as a first value; if it does not exist, the transmittance corresponding to the grid position is determined according to the three-dimensional coordinates of the receiving device and the three-dimensional coordinates of the first position and the second position on the first connecting line; the first connecting line refers to the connecting line between the receiving device and the transmitting device, the first position refers to the intersection point of the first connecting line and the scatterer that is closest to the transmitting device, and the second position refers to the intersection point of the first connecting line and the scatterer that is closest to the receiving device; the transmittance feature information is obtained according to the first value and / or the transmittance.

[0121] The step of predicting the channel information between transceivers based on the transceiver location information, scatterer distribution information, and penetration characteristic information includes: obtaining global environmental features based on the transceiver location information, scatterer distribution information, and penetration characteristic information; and performing mapping processing based on the global environmental features to predict the channel information between transceivers.

[0122] In this embodiment of the application, the small-scale channel information includes channel state information (CSI); the device further includes: a first acquisition module, used to acquire a portion of the CSI corresponding to the receiving device based on a number of probe signals sent by the transmitting device that is less than a first number; wherein, the mapping processing based on the global environmental features to predict the CSI between the transmitting and receiving devices includes: acquiring CSI reference features based on the portion of the CSI; extracting local environmental features corresponding to the receiving device from the global environmental features; and performing mapping processing after fusing the CSI reference features and the local environmental features to predict the CSI between the transmitting and receiving devices.

[0123] Furthermore, the digital twin channel prediction device based on the large channel model further includes: a second acquisition module, used to acquire original wireless channel information corresponding to each receiving position for a sample area used for model training; the sample area includes at least one area of ​​a scene; a second extraction module, used to extract channel information corresponding to each receiving position in the sample area based on the original wireless channel information, as sample channel information; a second construction module, used to construct training data based on the sample channel information; and a first training module, used to train the large channel model using the training data.

[0124] The implementation embodiments of the digital twin channel prediction method based on the large channel model described above are all applicable to the embodiments of the digital twin channel prediction device based on the large channel model, and can achieve the same technical effect.

[0125] This application also provides a digital twin channel prediction device based on a large channel model, such as... Figure 10 As shown, it includes: processor 101; The processor 101 is used to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area using a digital twin channel architecture. Using the aforementioned digital twin channel architecture, electromagnetic propagation parameter information is extracted for the three-dimensional wireless environment. The electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment. The scatterers are other objects in the three-dimensional wireless environment besides the transceiver devices. Using the large-scale channel model in the digital twin channel architecture, channel information between transceivers is predicted based on the transceiver location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

[0126] The digital twin channel prediction device based on a large channel model provided in this application embodiment utilizes a digital twin channel architecture to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area. Using the digital twin channel architecture, electromagnetic propagation parameter information is extracted for the three-dimensional wireless environment. This electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment. The scatterers are other objects in the three-dimensional wireless environment besides the transceivers. Using the large channel model in the digital twin channel architecture, channel information between the transceivers is predicted based on the transceiver location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. Channel information refers to channel information whose sensitivity to environmental changes is less than a first threshold, while small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold. This approach supports real-time channel prediction based on environmental adaptation, improving the scheme's poor environmental adaptability and real-time performance. Furthermore, predicting channel information based on electromagnetic propagation parameters can suppress interference from irrelevant environmental changes, avoid key feature overload and weak cross-scenario generalization caused by the model directly processing raw data. Moreover, by collaboratively predicting multiple types of channel information based on transceiver location information, scatterer distribution information, and penetration characteristic information, it can support real-time joint prediction of multiple channel tasks, improving multi-task joint prediction capabilities. This addresses at least one of the problems in existing channel information prediction schemes: poor environmental adaptability, insufficient real-time performance, and weak multi-task joint prediction capabilities.

[0127] The multimodal sensing data includes: LiDAR point cloud data and RGB-D images (red, green, blue, and depth). The construction of a three-dimensional wireless environment corresponding to the target area based on the multimodal sensing data includes: constructing a static scatterer in the target area based on the LiDAR point cloud data; constructing a dynamic scatterer in the target area based on the LiDAR point cloud data and the RGB-D image; and constructing the three-dimensional wireless environment corresponding to the target area based on the static and dynamic scatterers.

[0128] In this embodiment of the application, the step of constructing a dynamic scatterer in the target region based on the LiDAR point cloud data and the RGB-D image includes: determining the reference coordinates of the dynamic scatterer in the target region based on the RGB-D image; performing cluster analysis on the LiDAR point cloud data in the neighborhood of the dynamic scatterer based on the reference coordinates to determine the center coordinates and boundary information of the dynamic scatterer; and constructing the dynamic scatterer in the target region based on the center coordinates and boundary information.

[0129] The step of extracting electromagnetic propagation parameter information for the three-dimensional wireless environment includes: rasterizing the target area in the horizontal direction to obtain a two-dimensional raster coordinate system; and obtaining electromagnetic propagation parameter information for the three-dimensional wireless environment based on the two-dimensional raster coordinate system.

[0130] In this embodiment of the application, obtaining the transmittance feature information based on the two-dimensional grid coordinate system includes: determining whether a line-of-sight link exists between the receiving device and the transmitting device; the number of receiving devices is at least one; if it exists, the grid position corresponding to the receiving device in the two-dimensional grid coordinate system is taken as a first value; if it does not exist, the transmittance corresponding to the grid position is determined according to the three-dimensional coordinates of the receiving device and the three-dimensional coordinates of the first position and the second position on the first connecting line; the first connecting line refers to the connecting line between the receiving device and the transmitting device, the first position refers to the intersection point of the first connecting line and the scatterer that is closest to the transmitting device, and the second position refers to the intersection point of the first connecting line and the scatterer that is closest to the receiving device; the transmittance feature information is obtained according to the first value and / or the transmittance.

[0131] The step of predicting the channel information between transceivers based on the transceiver location information, scatterer distribution information, and penetration characteristic information includes: obtaining global environmental features based on the transceiver location information, scatterer distribution information, and penetration characteristic information; and performing mapping processing based on the global environmental features to predict the channel information between transceivers.

[0132] In this embodiment of the application, the small-scale channel information includes channel state information (CSI); the processor is further configured to: obtain a portion of the CSI corresponding to the receiving device based on a number of probe signals sent by the transmitting device that is less than a first number; wherein, the mapping process based on the global environmental features to predict the CSI between the transmitting and receiving devices includes: obtaining CSI reference features based on the portion of the CSI; extracting local environmental features corresponding to the receiving device from the global environmental features; and performing mapping processing after fusing the CSI reference features and the local environmental features to predict the CSI between the transmitting and receiving devices.

[0133] Furthermore, the processor is also configured to: acquire original wireless channel information corresponding to each receiving position for a sample region used for model training; the sample region includes at least one area of ​​a scene; extract channel information corresponding to each receiving position in the sample region based on the original wireless channel information, as sample channel information; construct training data based on the sample channel information; and train the large channel model using the training data.

[0134] The implementation embodiments of the digital twin channel prediction method based on the large channel model described above are all applicable to the embodiments of the digital twin channel prediction device based on the large channel model, and can achieve the same technical effect.

[0135] This application also provides a digital twin channel prediction device based on a large channel model, including a memory, a processor, and a program stored in the memory and executable on the processor; when the processor executes the program, it implements the above-described digital twin channel prediction method based on a large channel model.

[0136] The implementation embodiments of the digital twin channel prediction method based on the large channel model described above are all applicable to the embodiments of the digital twin channel prediction device based on the large channel model, and can achieve the same technical effect.

[0137] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the steps in the above-described digital twin channel prediction method based on a large channel model.

[0138] The implementation embodiments of the digital twin channel prediction method based on the large channel model described above are all applicable to the embodiments of this readable storage medium and can achieve the same technical effect.

[0139] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described method embodiment of the digital twin channel prediction method based on a large channel model, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0140] It should be noted that many of the functional components described in this specification are referred to as modules in order to more specifically emphasize the independence of their implementation.

[0141] In this embodiment, the module can be implemented in software so that it can be executed by various types of processors. For example, an identified executable code module may include one or more physical or logical blocks of computer instructions, which may be constructed as objects, procedures, or functions. Nevertheless, the executable code of the identified module does not need to be physically located together, but may include different instructions stored in different locations, which, when logically combined, constitute the module and achieve the module's intended purpose.

[0142] In practice, an executable code module can be a single instruction or many instructions, and can even be distributed across multiple different code segments, different programs, and across multiple memory devices. Similarly, operational data can be identified within the module and can be implemented in any suitable form and organized within any suitable data structure. This operational data can be collected as a single dataset or distributed across different locations (including different storage devices), and can exist, at least in part, solely as electronic signals within the system or network.

[0143] When a module can be implemented using software, considering the current level of hardware technology, modules that can be implemented in software can be implemented using hardware circuits by those skilled in the art to achieve the corresponding functions, without considering cost. These hardware circuits include conventional very-large-scale integrated circuits (VLSI) or gate arrays, as well as existing semiconductors such as logic chips and transistors, or other discrete components. Modules can also be implemented using programmable hardware devices, such as field-programmable gate arrays, programmable array logic, and programmable logic devices.

[0144] The above describes the preferred embodiments of this application. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A digital twin channel prediction method based on a large channel model, characterized in that, include: By utilizing a digital twin channel architecture, a three-dimensional wireless environment corresponding to the target area is constructed based on multimodal sensing data for the target area. Using the aforementioned digital twin channel architecture, electromagnetic propagation parameter information is extracted for the three-dimensional wireless environment. The electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment. The scatterer is any object in the three-dimensional wireless environment other than the transceiver device; Using the large-scale channel model in the digital twin channel architecture, channel information between transceivers is predicted based on the transceiver location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

2. The digital twin channel prediction method according to claim 1, characterized in that, The multimodal sensing data includes: LiDAR point cloud data and RGB-D images; the construction of a three-dimensional wireless environment corresponding to the target area based on the multimodal sensing data for the target area includes: Based on the LiDAR point cloud data, a static scatterer in the target region is constructed; Based on the LiDAR point cloud data and RGB-D image, a dynamic scatterer in the target region is constructed; Based on the static and dynamic scatterers, a three-dimensional wireless environment corresponding to the target area is constructed.

3. The digital twin channel prediction method according to claim 2, characterized in that, The step of constructing a dynamic scatterer in the target region based on the LiDAR point cloud data and RGB-D image includes: Based on the RGB-D image, determine the reference coordinates of the dynamic scatterer in the target region; Based on the reference coordinates, cluster analysis is performed on the LiDAR point cloud data in the neighborhood of the dynamic scatterer to determine the center coordinates and boundary information of the dynamic scatterer. Based on the center coordinates and boundary information, a dynamic scatterer is constructed in the target region.

4. The digital twin channel prediction method according to claim 1, characterized in that, The extraction of electromagnetic propagation parameter information for the three-dimensional wireless environment includes: The target area is rasterized in the horizontal direction to obtain a two-dimensional raster coordinate system; Based on the two-dimensional grid coordinate system, electromagnetic propagation parameter information for the three-dimensional wireless environment is obtained.

5. The digital twin channel prediction method according to claim 4, characterized in that, Based on the two-dimensional grid coordinate system, the transmittance feature information is obtained, including: Determine whether a line-of-sight link exists between the receiving device and the transmitting device; the number of receiving devices is at least one. If it exists, the grid position corresponding to the receiving device in the two-dimensional grid coordinate system shall be the first value; If not, the transmittance corresponding to the grid position is determined based on the three-dimensional coordinates of the receiving device and the three-dimensional coordinates of the first and second positions on the first connecting line; the first connecting line refers to the connecting line between the receiving device and the transmitting device, the first position refers to the intersection point of the first connecting line and the scatterer that is closest to the transmitting device, and the second position refers to the intersection point of the first connecting line and the scatterer that is closest to the receiving device. The penetration characteristic information is obtained based on the first value and / or penetration rate.

6. The digital twin channel prediction method according to claim 1, characterized in that, The step of predicting channel information between transceivers based on the transceiver location information, scatterer distribution information, and penetration characteristic information includes: Based on the location information of the transceiver device, the distribution information of the scatterer, and the penetration characteristic information, the global environmental characteristics are obtained; Based on the global environmental features, mapping processing is performed to predict the channel information between the transceiver devices.

7. The digital twin channel prediction method according to claim 6, characterized in that, The small-scale channel information includes channel state information (CSI); the method further includes: Based on the number of detection signals sent by the transmitting device that is less than a first number, a portion of the CSI corresponding to the receiving device is obtained; The mapping process based on the global environmental features, which predicts the CSI between transceiver devices, includes: Based on the aforementioned CSI portion, obtain the CSI reference features; The local environmental features corresponding to the receiving device are obtained by extracting the global environmental features. The CSI reference features are fused with the local environment features and then mapped to predict the CSI between the transceiver devices.

8. The digital twin channel prediction method according to claim 1, characterized in that, Also includes: For the sample region used for model training, obtain the original wireless channel information corresponding to each receiving position; The sample area includes at least one area of ​​a scene; Based on the original wireless channel information, the channel information corresponding to each receiving location in the sample area is extracted as sample channel information; Based on the sample channel information, training data is constructed; The channel model is trained using the training data.

9. A digital twin channel prediction device based on a large channel model, characterized in that, include: The first construction module is used to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area using a digital twin channel architecture. The first extraction module is used to extract electromagnetic propagation parameter information for the three-dimensional wireless environment using the digital twin channel architecture. The electromagnetic propagation parameter information includes: the location information of the transceiver device in the three-dimensional wireless environment, the distribution information of the scatterer in the three-dimensional wireless environment, and the penetration characteristic information corresponding to the three-dimensional wireless environment; the scatterer is other objects in the three-dimensional wireless environment besides the transceiver device. The first prediction module is used to predict the channel information between the transceiver devices based on the large channel model in the digital twin channel architecture, according to the transceiver device location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to the channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to the channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

10. A digital twin channel prediction device based on a large channel model, characterized in that, include: processor; The processor is used to construct a three-dimensional wireless environment corresponding to the target area based on multimodal sensing data for the target area using a digital twin channel architecture. Using the aforementioned digital twin channel architecture, electromagnetic propagation parameter information is extracted for the three-dimensional wireless environment. The electromagnetic propagation parameter information includes: transceiver location information in the three-dimensional wireless environment, scatterer distribution information in the three-dimensional wireless environment, and penetration characteristic information corresponding to the three-dimensional wireless environment. The scatterer is any object in the three-dimensional wireless environment other than the transceiver device; Using the large-scale channel model in the digital twin channel architecture, channel information between transceivers is predicted based on the transceiver location information, scatterer distribution information, and penetration characteristic information. The channel information includes large-scale channel information and small-scale channel information. The large-scale channel information refers to channel information whose sensitivity to environmental changes is less than a first threshold, and the small-scale channel information refers to channel information whose sensitivity to environmental changes is greater than or equal to the first threshold.

11. A digital twin channel prediction device based on a large channel model, comprising a memory, a processor, and a program stored in the memory and executable on the processor; characterized in that, When the processor executes the program, it implements the digital twin channel prediction method based on a large channel model as described in any one of claims 1 to 8.

12. A readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the digital twin channel prediction method based on a large channel model as described in any one of claims 1 to 8.

13. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the digital twin channel prediction method based on a large channel model as described in any one of claims 1 to 8.

Images