A radar-assisted beam map learning internet of vehicles near-field beam prediction method

Abstract

The application discloses a radar-assisted beam map learning Internet of Vehicles near-field beam prediction method, comprising the following steps: constructing a three-dimensional near-field beam codebook; constructing a radar data cube; constructing a spatial snapshot matrix; generating a radar angle domain map; obtaining a corresponding optimal near-field beam index according to the three-dimensional near-field beam codebook; constructing a Gaussian soft heat map label; learning a nonlinear mapping relationship between the radar angle domain map and a communication angle domain beam map; using the Gaussian soft heat map label to supervise training of the communication angle domain beam heat map; in the prediction stage, performing peak value detection on the communication angle domain beam heat map; combining radar ranging to obtain a predicted near-field beam index; selecting a corresponding beam vector from the three-dimensional near-field beam codebook, and performing downlink beamforming. The application overcomes the problems of huge exhaustive scanning overhead of a traditional three-dimensional near-field codebook, easy influence of radar communication mismatch on direct radar geometry mapping, and difficulty of hard beam index classification in representing beam space continuity.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication, and relates to vehicle-to-everything (V2X) communication, ultra-large-scale MIMO near-field beamforming, radar-assisted communication, and intelligent beam management technology. Specifically, it relates to a near-field beam prediction method for V2X based on radar-assisted beam pattern learning. Background Technology

[0002] With the rapid development of intelligent connected vehicles, autonomous driving, and vehicle-to-everything (V2X) technologies, vehicle-to-everything (V2X) communication systems need to provide highly reliable, low-latency, and high-throughput wireless connections for high-speed moving vehicles. To improve the coverage and beamforming gain of high-frequency V2X downlinks, roadside base stations are increasingly adopting large-scale or even very large-scale antenna arrays to form highly directional narrow beams and enhance the quality of vehicle communication links. However, while very large-scale antenna arrays can significantly improve array gain, they also increase the complexity of beam training and beam tracking.

[0003] In ultra-large-scale uniform planar array vehicle-to-everything (V2X) communication scenarios, as the array aperture continuously increases, vehicles within the actual road distance are likely to be located in the array's near-field radiation region. Under near-field propagation conditions, the communication beam is no longer determined solely by azimuth and elevation angles, but is simultaneously influenced by range, azimuth, and elevation. Therefore, near-field beam selection requires searching within a three-dimensional codebook of range, azimuth, and elevation. For high-speed moving vehicles, the communication channel changes rapidly over time. If traditional exhaustive beam scanning or codeword-by-codeword measurement feedback methods are still used, it will generate huge signaling overhead and processing delays, easily causing beam training results to lag behind vehicle position and channel changes, thus leading to problems such as beam inaccuracy, link quality degradation, and even communication interruption.

[0004] To address the aforementioned beam management challenges, existing research attempts to utilize sensing information to assist beam selection, such as inferring candidate communication beams using vehicle position, visual images, LiDAR point clouds, or other environmental perception data. These methods can reduce communication beam scanning overhead to some extent, but their performance often depends on the accuracy of geometric position or line-of-sight path characteristics. In scenarios with dynamic occlusion, complex scattering, and non-line-of-sight propagation, beam prediction methods based solely on geometric position or visual perception struggle to accurately reflect the effective propagation path in the communication channel, resulting in insufficient prediction robustness.

[0005] Vehicle-mounted frequency-modulated continuous wave radar can provide range and angle information for vehicle targets and road scatterers, thus possessing the potential to assist in beam management for vehicle-to-everything (V2X) networks. Existing radar-assisted communication methods primarily map radar observations to communication beam indices using target detection results, radar covariance matrices, angle spectra, or deep learning models to reduce beam search overhead. However, most existing methods are based on far-field plane wave assumptions, uniform linear arrays, or low-dimensional angle codebooks, making them difficult to apply to near-field beam selection with range-dependent focusing characteristics under ultra-large-scale uniform plane arrays. Furthermore, differences between radar and communication systems in carrier frequency, array aperture, installation location, and propagation response mean that the radar angle peak and the optimal communication beam direction are not always strictly consistent. Directly selecting the communication beam based on the radar geometric peak is susceptible to radar-communication mismatch.

[0006] On the other hand, existing deep learning-based beam prediction methods typically treat beam selection as a hard beam indexing classification problem, directly outputting an optimal beam number. This approach ignores the continuity in the communication beam space: adjacent angular domain beams usually correspond to similar propagation directions and may have similar beamforming gains. If only hard label classification is used, it is not only difficult to characterize the physical correlation between adjacent beams, but also prone to large beam prediction errors in dynamic occlusion and non-line-of-sight scenarios.

[0007] In summary, how to fully utilize the spatial correlation between vehicle radar signals and communication channels in ultra-large-scale uniform planar array vehicle-to-everything (V2X) near-field communication scenarios, avoid real-time exhaustive search of the complete three-dimensional near-field codebook, and overcome problems such as the susceptibility of direct radar geometry mapping to radar-communication mismatch and the difficulty of hard beam index classification in representing beam spatial continuity, has become a key technical problem that urgently needs to be solved to achieve low-overhead, high-reliability V2X near-field beam management. Summary of the Invention

[0008] Purpose of the invention: In order to overcome the shortcomings of the existing technology, a near-field beam prediction method for vehicle-to-everything (V2X) based on radar-assisted beam pattern learning is provided. This method is designed for near-field communication scenarios of V2X using ultra-large-scale uniform planar arrays and solves the problems of high overhead of exhaustive search of three-dimensional near-field codebook, easy inaccuracy of high-dynamic vehicle beams, and susceptibility of direct radar geometric mapping to radar communication mismatch.

[0009] Technical Solution: To achieve the above objectives, this invention provides a near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beam pattern learning, comprising the following steps:

[0010] S1: System Configuration and Near-Field Codebook Construction: The roadside base station is configured with a large-scale uniform planar array for communication and a radar receiving array for passively receiving vehicle-mounted radar signals, and a three-dimensional near-field beamcodebook is constructed based on the distance, azimuth, and elevation dimensions. ;

[0011] S2: Passive Radar Signal Reception and Radar Data Construction: The roadside base station passively receives radar signals emitted by the vehicle-mounted frequency-modulated continuous wave radar of the target vehicle, demodulates, samples, and organizes the array dimensions of the received signals to construct a radar data cube. The radar data cube contains a slow time dimension, an antenna dimension, and a fast time dimension.

[0012] S3: Radar corner area map and range index acquisition: based on radar data cube A spatial snapshot matrix is ​​constructed, and the radar covariance matrix is ​​estimated from the spatial snapshot matrix; Bartlett angular domain spectrum estimation is performed based on the radar array steering vector to generate the radar angular domain map. Simultaneously utilizing radar data cubes Determine the distance index using fast time-dimensional ranging information. ;

[0013] S4: Near-field beam soft tag construction: Based on the near-field communication channel and a predefined 3D near-field beam codebook The optimal near-field beam index corresponding to the training samples is obtained by traversing the search. And extract the optimal azimuth index from it. and optimal pitch angle index ;by Create a Gaussian soft heatmap label around the center. It is used to characterize the spatial continuity between adjacent communication beams;

[0014] S5: Radar Beam Pattern Learning and Supervised Training: Radar Corner Domain Pattern The input is a lightweight encoder-decoder convolutional neural network, which learns the nonlinear mapping between radar angular domain maps and communication angular domain beam maps, and outputs a communication angular domain beam heatmap. ,in This indicates the confidence level that the beams at different azimuth and elevation angles are the optimal communication beams.

[0015] During the training phase, Gaussian soft heatmap labels were used. Communication angular domain beam heatmap Conduct supervised training;

[0016] S6: Near-field beam indexing and downlink beamforming: During the prediction phase, the communication angular domain beam heatmap is analyzed. Peak detection is performed to obtain the predicted azimuth index. And predicted pitch angle index Predict the azimuth index Predicted pitch angle index Distance index obtained in step S3 Combined, the predicted near-field beam index is obtained. Roadside base stations are based on predicted near-field beam indices. From the three-dimensional near-field beamcodebook Select the corresponding beam vector and perform downlink beamforming.

[0017] Furthermore, the ultra-large-scale uniform planar array in step S1 includes... There are 1 antenna element, of which and Represents the number of antenna elements in the vertical and horizontal directions, respectively; three-dimensional near-field beamcodebook The code is discretized based on three dimensions: distance, azimuth, and elevation. The codebook size is [size missing]. ,in , and These represent the number of range grids, azimuth grids, and elevation grids, respectively. Each codeword corresponds to a range-azimuth-elevation grid point, used to adapt to the range-dependent beam focusing characteristics caused by spherical wave propagation under ultra-large-scale uniform planar arrays.

[0018] Further, in step S2, the radar receiving array includes There are 1 receiving antenna element, among which and These represent the number of antenna elements in the vertical and horizontal directions of the radar receiving array, respectively. The radar receiving array is used to passively intercept signals emitted by the vehicle-mounted frequency-modulated continuous wave radar to obtain spatial perception information about the vehicle target and environmental scatterers. The roadside base station demodulates and samples the received vehicle-mounted radar signals and processes the data according to the linear frequency-modulated pulse dimension, antenna dimension, and fast time dimension to obtain a radar data cube. Radar data cube The dimensions include , and ,in Indicates the number of linear frequency modulated pulses in the radar. Indicates the number of radar receiving antenna elements. This represents the number of fast-time sampling points corresponding to each linear frequency modulation pulse; the radar data cube is used to simultaneously characterize the slow-time, spatial, and range-related information of the radar echo.

[0019] Furthermore, in step S3, the roadside base station uses radar data cubes... A spatial snapshot matrix is ​​constructed from the antenna-dimensional received data corresponding to different linear frequency modulated pulses and fast time sampling points. Based on the statistical estimation of multiple spatial snapshots, the radar covariance matrix is ​​used to characterize the spatial correlation of the radar receiving array in different angular directions. Radar angular domain map. The range index is obtained through the Bartlett spectral estimation method, which matches the radar covariance matrix with the radar array steering vectors corresponding to different azimuth and elevation angles to obtain a two-dimensional angular domain spectral distribution. Radar data cube The fast time-dimensional ranging results are obtained, and the ranging results are compared with the three-dimensional near-field beamcodebook. It is obtained by matching the distance grid in the data.

[0020] Furthermore, in step S4, during the training phase, the optimal near-field beam index... By analyzing the three-dimensional near-field beamcodebook The near-field beam label is obtained by performing an traversal search. The traversal search is based on the average beamforming gain of each candidate near-field beam on the communication channel. The codeword with the largest gain is selected as the optimal near-field beam label for the training sample.

[0021] Furthermore, in step S4, the Gaussian soft heatmap labels Indexed by optimal azimuth and optimal pitch angle index Centered on the communication angular domain beammap, a two-dimensional soft tag is formed with the largest central response and the responses of adjacent beams gradually decreasing; Gaussian soft heatmap tag. The angular domain diffusion range is determined by the soft tag variance parameter. control.

[0022] Furthermore, the lightweight encoder-decoder convolutional neural network in step S5 includes an encoder, a bottleneck layer, a decoder, skip connections, and an output mapping layer; wherein, the encoder is used to extract the radar angular domain map. Multi-scale spatial features are used, with the bottleneck layer for extracting high-level angular domain representations and the decoder for recovering the communication angular domain beam heatmap. Spatial resolution, skip connections are used to fuse shallow angular domain detail features and deep semantic features, and the output mapping layer is used to convert the decoder output into a communication angular domain beam heatmap. .

[0023] Furthermore, in step S5, the neighborhood-weighted mean square error loss function is used to enhance the training weights of the optimal beam neighborhood, enabling the network to improve the prediction accuracy near the optimal beam while maintaining the smoothness of the angular domain; the neighborhood-weighted mean square error loss function is based on Gaussian soft heatmap labels. Different training weights are assigned to different angular domain locations, giving higher weights to the optimal beam and its neighboring beam regions, and lower weights to non-neighboring beam regions; the training weights are determined by neighborhood weighting coefficients. adjust.

[0024] Furthermore, in step S6, the azimuth index is predicted. And predicted pitch angle index Communication angular domain beam heatmap The location of the maximum response in the distance index is determined. The predicted near-field beam index is obtained by combining the radar ranging results, the three factors being combined. .

[0025] Furthermore, in step S6, the roadside base station determines the near-field beam index based on the predicted near-field beam index. Directly from the 3D near-field beamcodebook The corresponding near-field beam vector is selected for downlink beamforming, eliminating the need for real-time exhaustive scanning of the complete three-dimensional near-field codebook of range-azimuth-pitch angle, thereby reducing the beam training overhead in high-dynamic scenarios of vehicle-to-everything (V2X) networks.

[0026] This invention passively receives vehicle-mounted frequency-modulated continuous wave radar signals from roadside base stations, constructs a radar data cube, and generates a radar angular domain map. It further utilizes a lightweight encoder-decoder convolutional neural network to learn the mapping relationship between the radar angular domain map and the communication angular domain beam heatmap, and combines Gaussian soft labels and a neighborhood-weighted mean square error loss function to characterize the spatial continuity of adjacent beams. In the prediction stage, azimuth and elevation indices are obtained from the communication angular domain beam heatmap, and combined with the distance index obtained from radar ranging, near-field beams are directly selected from a three-dimensional near-field beamcodebook, thereby achieving low-overhead, high-reliability vehicle-to-everything (V2X) near-field beam prediction and downlink beamforming.

[0027] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0028] 1. The method of this invention utilizes roadside base stations to passively receive the target vehicle's onboard frequency-modulated continuous wave radar signal, constructs a radar data cube, and generates a radar angular domain map. It then directly predicts the communication angular domain beam heatmap through a learning process. Compared to traditional exhaustive search schemes based on a three-dimensional near-field codebook of range-azimuth-elevation angle, this invention eliminates the need for real-time codeword scanning of the complete near-field codebook, effectively reducing beam training overhead and processing latency in near-field communication of ultra-large-scale uniform planar array vehicle-to-everything (V2X) networks.

[0029] 2. The method of this invention fully utilizes the spatial correlation between vehicle-mounted radar signals and communication channels on vehicle targets, road environments, and scattering bodies to map the radar angular domain spectrum into a communication angular domain beam spectrum. Compared to methods that directly perform geometric mapping based on radar angular peak values, the method of this invention can learn the nonlinear correspondence between the radar and communication systems through neural networks, alleviating mismatch problems in carrier frequency, array aperture, installation location, and propagation response, and improving the robustness of beam prediction in complex road environments.

[0030] 3. This invention addresses near-field communication scenarios using ultra-large-scale uniform planar arrays. It employs a three-dimensional near-field beamcodebook jointly described by range, azimuth, and elevation angles, and decomposes near-field beam selection into two parts: a range index determined by radar ranging and an angle index predicted by beam heatmap. Compared to methods applicable only to far-field angle beam selection, this invention can adapt to the range-dependent beam focusing characteristics under spherical wave propagation, making it more suitable for the high-gain beamforming requirements in vehicle-to-everything (V2X) near-field communication.

[0031] 4. The method of this invention employs Gaussian soft heatmap labels and a neighborhood-weighted mean square error loss function to transform traditional hard beam index classification into structured angular domain heatmap regression. This design can characterize the spatial continuity between adjacent communication beams and enhance the training weights of the optimal beam neighborhood, thereby improving the accuracy and stability of communication angular domain beammap prediction.

[0032] 5. The method of this invention employs a lightweight encoder-decoder convolutional neural network to achieve radar beam pattern learning. In the prediction stage, only the radar angular domain map needs to be input to output the communication angular domain beam heatmap, and the near-field beam index is quickly determined through a combination of peak detection and distance indexing. This method has a simple inference process and is suitable for low-latency beam management in highly dynamic vehicle-to-everything (V2X) scenarios. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of a near-field communication system model for vehicle-to-everything (V2X) networks.

[0034] Figure 2 This is a flowchart of the method of the present invention.

[0035] Figure 3 A schematic diagram of the network structure for learning beam patterns for radar. Detailed Implementation

[0036] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0037] Example 1:

[0038] like Figure 1As shown, in a vehicle-to-everything (V2X) near-field communication system, roadside base stations are fixedly deployed on the side of the road, equipped with a large-scale uniform planar array for communication and a radar receiver array for passively receiving vehicle-mounted radar signals. The target vehicle is equipped with a single-antenna communication receiver and a frequency-modulated continuous wave (FM-CW) radar mounted at the front of the vehicle. The roadside base station uses the passively intercepted FM-CW radar signals to acquire spatial perception information of the vehicle target and road environment scatterers. Simultaneously, it uses the communication channel and a three-dimensional near-field beamcodebook to construct a supervisory label, thereby learning the mapping relationship between the radar angular domain spectrum and the communication angular domain beamcodebook.

[0039] Based on the aforementioned vehicle-to-everything (V2X) near-field communication system model, this embodiment provides a V2X near-field beam prediction method based on radar-assisted beammap learning, such as... Figure 2 As shown, the specific steps include the following:

[0040] Step 1: As Figure 1 As shown, the roadside base station is configured with a very large-scale uniform planar array for communication. This array is arranged in a vertical plane, and the total number of antenna elements is [number missing]. ,in, and These represent the number of antenna elements in the vertical and horizontal directions, respectively. Due to the large aperture of the ultra-large-scale uniform planar array, the target vehicle may be located in the near-field radiation region of the roadside base station array within the actual vehicle-to-everything (V2X) communication distance. Therefore, the communication beam response is not only related to the azimuth and elevation angles but also closely related to the distance from the target vehicle to the array center.

[0041] Let the position of the target vehicle relative to the center of the array be determined by the distance. Azimuth and pitch angle Description, its direction vector It can be represented as Let the communication carrier wavelength be... In the ultra-large-scale uniform planar array of roadside base stations, the first The position vectors of the antenna elements relative to the array center are: Then the corresponding spherical wave near-field steering vector is the first... Each element can be represented as:

[0042]

[0043] For including One orthogonal frequency division multiplexing (OFDM) subcarrier, the subcarrier spacing is... The broadband downlink communication system, the first Near-field communication channel on each subcarrier It can be represented as:

[0044]

[0045] in, , , , and They represent the first The complex gain, time delay, physical distance, azimuth angle, and elevation angle of each propagation path.

[0046] Based on the above near-field channel model, the roadside base station constructs a three-dimensional near-field beamcodebook according to three dimensions: range, azimuth, and elevation. The codebook can be represented as:

[0047]

[0048] in, , and They represent the first The distance grid, the first The azimuth grid and the first A pitch angle grid; This represents the near-field focusing beam vector focused on the corresponding range-azimuth-elevation grid point. The total size of this near-field codebook is... ,in , and These represent the number of grid divisions in the distance, azimuth, and elevation dimensions, respectively.

[0049] Step Two: The roadside base station is equipped with a radar receiving array to passively intercept radar signals emitted by the target vehicle's onboard frequency-modulated continuous wave radar. The radar receiving array includes... There are 1 receiving antenna element, among which and These represent the number of radar receiving antenna elements in the vertical and horizontal directions, respectively. The roadside base station demodulates, samples, and organizes the array dimensions of the received vehicle-mounted radar signals to obtain a radar data cube. Its dimension can be represented as

[0050]

[0051] in, Indicates the number of linear frequency modulated pulses in the radar. This indicates the total number of radar receiving antenna elements. This represents the number of fast-time sampling points corresponding to each linear frequency modulated pulse. (Radar data cube) It also includes a slow time dimension, an antenna dimension, and a fast time dimension, which can be used to characterize the motion-related information, spatial angle information, and distance-related information of radar echoes, respectively.

[0052] Step 3: Roadside base stations use radar data cubes Constructing a spatial snapshot matrix The radar covariance matrix is ​​estimated statistically from multiple spatial snapshots. ,Right now

[0053]

[0054] in, Indicates the number of spatial snapshots. This represents the conjugate transpose. Subsequently, based on the steering vectors of the radar receiving array in different azimuth and elevation directions, the Bartlett spectral estimation method is used to generate the radar angular domain map. Its angular domain spectral value can be expressed as

[0055]

[0056] in, Indicates the radar receiving array in azimuth angle and pitch angle The steering vector in the direction. The resulting radar angular domain map. It can characterize the spatial distribution of vehicle targets and road environment scatterers in the angular domain. Meanwhile, roadside base stations utilize radar data cubes. The distance to the target vehicle is obtained from the fast time-dimensional ranging information, and this distance is compared with the three-dimensional near-field beamcodebook. The distance index is obtained by matching the distance grid in the data. .

[0057] Step 4: During the training phase, the roadside base station uses the near-field communication channel and a predefined three-dimensional near-field beamcodebook. Constructing a supervisory label. By traversing the candidate codewords in the 3D near-field beamcodebook, the codeword with the largest average beamforming gain across all OFDM subcarriers is selected as the optimal near-field beam label, and its optimal near-field beam index is determined. Represented as

[0058]

[0059] in, , and These represent the optimal distance index, optimal azimuth index, and optimal elevation index, respectively. This indicates the number of OFDM subcarriers. Since the neural network in this invention primarily predicts the communication angular domain beam heatmap, the optimal azimuth index is extracted from the optimal near-field beam index. and optimal pitch angle index and with Construct Gaussian soft heatmap labels around the center , its first Line 1 Column elements can be represented as

[0060]

[0061] in, This represents the soft tag variance parameter, used to control the angular domain diffusion range. The Gaussian soft heatmap achieves maximum response at the optimal angular index and gradually decays at neighboring beam positions, characterizing the spatial continuity between adjacent communication beams.

[0062] Step 5: As Figure 3 As shown, this embodiment constructs a lightweight encoder-decoder convolutional neural network for learning radar angular domain maps. Communication angular domain beam heatmap The network employs a nonlinear mapping relationship. It comprises an encoder, a bottleneck layer, a decoder, skip connections, and an output mapping layer. The encoder extracts multi-scale spatial features from the radar angular domain map; the bottleneck layer extracts high-level angular domain representations; the decoder recovers the spatial resolution of the communication angular domain beam heatmap; skip connections fuse shallow angular domain detail features and deep semantic features; and the output mapping layer outputs the communication angular domain beam heatmap. .

[0063] During the training phase, Gaussian soft heatmap labels constructed in step four are used. For network output Supervised training is performed. To enhance the training weights of the optimal beam neighborhood, this embodiment employs a neighborhood-weighted mean square error loss function, which takes the form of:

[0064] ,

[0065] in, and These represent the dimensions of the communication angular domain beam heatmap in the azimuth and elevation dimensions, respectively. This represents the neighborhood weighting coefficient. Using this loss function, the network can improve the prediction accuracy of the optimal beam and its neighboring regions while maintaining angular domain smoothness.

[0066] Step Six: During the prediction phase, the roadside base station will obtain the radar angular domain map in real time. Inputting a trained lightweight encoder-decoder convolutional neural network yields a beam heatmap in the communication angular domain. Subsequently, peak detection was performed on the communication angular domain beam heatmap to obtain the predicted azimuth index. And predicted pitch angle index ,Right now

[0067]

[0068] Predict azimuth index Predicted pitch angle index Distance index obtained in step 3 By combining the results, the predicted near-field beam index is obtained. The roadside base station, based on this predicted near-field beam index, directly retrieves the beam pattern from the three-dimensional near-field codebook. Select the corresponding near-field beam vector And use the beam vector to perform downlink beamforming.

[0069] Through the above process, this embodiment does not require real-time exhaustive scanning of the complete three-dimensional near-field codebook of range-azimuth-elevation angle. Instead, it uses radar ranging to determine the range index and uses radar to predict the azimuth and elevation indexes through a beam pattern learning network, thereby achieving low-overhead and high-reliability beam selection in near-field communication of ultra-large-scale uniform planar array vehicle networks.

Claims

1. A near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beammap learning, characterized in that, Includes the following steps: S1: The roadside base station is equipped with a large-scale uniform planar array for communication and a radar receiving array for passively receiving vehicle-mounted radar signals, and constructs a three-dimensional near-field beamcodebook based on the dimensions of range, azimuth, and elevation. ; S2: The roadside base station passively receives radar signals emitted by the vehicle-mounted frequency-modulated continuous wave radar of the target vehicle, demodulates, samples, and organizes the array dimensions of the received signals to construct a radar data cube. ; S3: Based on radar data cube A spatial snapshot matrix is ​​constructed, and the radar covariance matrix is ​​estimated from the spatial snapshot matrix; Bartlett angular domain spectrum estimation is performed based on the radar array steering vector to generate the radar angular domain map. Simultaneously utilizing radar data cubes Determine the distance index using fast time-dimensional ranging information. ; S4: Based on the near-field communication channel and a predefined three-dimensional near-field beamcodebook The optimal near-field beam index corresponding to the training samples is obtained by traversing the search. And extract the optimal azimuth index from it. and optimal pitch angle index ;by Construct Gaussian soft heatmap labels around the center It is used to characterize the spatial continuity between adjacent communication beams; S5: Radar corner area map The input is a lightweight encoder-decoder convolutional neural network, which learns the nonlinear mapping between radar angular domain beammaps and communication angular domain beammaps, and outputs a communication angular domain beam heatmap. ,in This indicates the confidence level that the beams at different azimuth and elevation angles are the optimal communication beams. During the training phase, Gaussian soft heatmap labels were used. Communication angular domain beam heatmap Conduct supervised training; S6: During the prediction phase, the communication angular domain beam heatmap Peak detection is performed to obtain the predicted azimuth index. And predicted pitch angle index Predict the azimuth index Predicted pitch angle index Distance index obtained in step S3 Combined, the predicted near-field beam index is obtained. Roadside base stations are based on predicted near-field beam indices. From the three-dimensional near-field beamcodebook Select the corresponding beam vector and perform downlink beamforming.

2. The near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beam pattern learning according to claim 1, characterized in that, The ultra-large-scale uniform planar array in step S1 includes There are 1 antenna element, of which and Represents the number of antenna elements in the vertical and horizontal directions, respectively; three-dimensional near-field beamcodebook The code is discretized based on three dimensions: distance, azimuth, and elevation. The codebook size is [size missing]. ,in , and These represent the number of range grids, azimuth grids, and elevation grids, respectively. Each codeword corresponds to a range-azimuth-elevation grid point, used to adapt to the range-dependent beam focusing characteristics caused by spherical wave propagation under ultra-large-scale uniform planar arrays.

3. The near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beam pattern learning according to claim 2, characterized in that, The radar receiving array in step S2 includes There are 1 receiving antenna element, among which and These represent the number of antenna elements in the vertical and horizontal directions of the radar receiving array, respectively. The radar receiving array is used to passively intercept the signals emitted by the vehicle-mounted frequency-modulated continuous wave radar of the target vehicle in order to obtain spatial perception information of the vehicle target and environmental scatterers. The roadside base station demodulates and samples the received vehicle-mounted radar signals, and processes the data according to the linear frequency modulated pulse dimension, antenna dimension, and fast time dimension to obtain a radar data cube. Radar data cube The dimensions include , and ,in Indicates the number of linear frequency modulated pulses in the radar. Indicates the number of radar receiving antenna elements. This indicates the number of fast-time sampling points corresponding to each linear frequency modulation pulse; Radar data cubes are used to simultaneously characterize the slow time, spatial, and range-related information of radar echoes.

4. The near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beammap learning according to claim 3, characterized in that, In step S3, the roadside base station uses radar data cubes. A spatial snapshot matrix is ​​constructed from the antenna-dimensional received data corresponding to different linear frequency modulated pulses and fast time sampling points. The radar covariance matrix is ​​estimated based on multiple spatial snapshots to characterize the spatial correlation of the radar receiving array in different angular directions. Radar angular domain map The range index is obtained through the Bartlett spectral estimation method, which matches the radar covariance matrix with the radar array steering vectors corresponding to different azimuth and elevation angles to obtain a two-dimensional angular domain spectral distribution. Radar data cube The fast time-dimensional ranging results are obtained, and the ranging results are compared with the three-dimensional near-field beamcodebook. It is obtained by matching the distance grid in the data.

5. The near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beam pattern learning according to claim 4, characterized in that, In step S4, during the training phase, the optimal near-field beam index is determined. By analyzing the three-dimensional near-field beamcodebook The near-field beam label is obtained by performing an traversal search. The traversal search is based on the average beamforming gain of each candidate near-field beam on the communication channel. The codeword with the largest gain is selected as the optimal near-field beam label for the training sample.

6. The near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beammap learning according to claim 5, characterized in that, Gaussian soft heatmap labels in step S4 Indexed by optimal azimuth and optimal pitch angle index Centered on the communication angular domain beam pattern, a two-dimensional soft tag is formed with the largest central response and the responses of adjacent beams gradually decreasing. Gaussian soft heatmap labels The angular domain diffusion range is determined by the soft tag variance parameter. control.

7. The near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beammap learning according to claim 6, characterized in that, The lightweight encoder-decoder convolutional neural network in step S5 includes an encoder, a bottleneck layer, a decoder, skip connections, and an output mapping layer; wherein, the encoder is used to extract radar angular domain maps. Multi-scale spatial features are used, with the bottleneck layer for extracting high-level angular domain representations and the decoder for recovering the communication angular domain beam heatmap. Spatial resolution, skip connections are used to fuse shallow angular domain detail features and deep semantic features, and the output mapping layer is used to convert the decoder output into a communication angular domain beam heatmap. .

8. The near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beam pattern learning according to claim 7, characterized in that, In step S5, the neighborhood-weighted mean square error loss function is used to enhance the training weights of the optimal beam neighborhood, enabling the network to improve the prediction accuracy near the optimal beam while maintaining the smoothness of the angular domain; the neighborhood-weighted mean square error loss function is based on Gaussian soft heatmap labels. Different training weights are assigned to different angular locations.

9. The near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beammap learning according to claim 8, characterized in that, In step S6, the azimuth index is predicted. And predicted pitch angle index Communication angular domain beam heatmap The location of the maximum response in the distance index is determined. The predicted near-field beam index is obtained by combining the radar ranging results, the three factors being combined. .

10. The near-field beam prediction method for vehicle-to-everything (V2X) networks based on radar-assisted beammap learning according to claim 9, characterized in that, In step S6, the roadside base station uses the predicted near-field beam index. Directly from the 3D near-field beamcodebook In this process, the corresponding near-field beam vector is selected for downlink beamforming, eliminating the need for real-time exhaustive scanning of the complete range-azimuth-elevation three-dimensional near-field codebook.

Images