A communication beam prediction method and system based on environmental perception

Abstract

This invention belongs to the field of integrated machine learning, communication, and sensing technology, specifically relating to a communication beam prediction method and system based on environmental perception. The method's specific process is as follows: Step 1: Establish a joint optimization model and constraints with the objective of maximizing the system's achievable downlink communication rate; Step 2: Configure LPB-Net including a multi-target state perception module, a graph frequency domain attention prediction module, and a beamforming module; In the first stage, the multi-target state perception module extracts the sensing target location features from the echo signals received by the RIS sensing unit; In the second stage, the graph frequency domain attention prediction module obtains the spatiotemporal correlation between multiple targets based on the sensing target location features, predicting the future channel state; In the third stage, the beamforming module embeds the Maximum Ratio Transmission (MRT) into a deep learning framework, and based on the optimization model and constraints, realizes a joint beamforming design combining model-driven and data-driven approaches.

Description

Technical Field

[0001] This invention belongs to the field of integrated machine learning, communication and sensing technology, and specifically relates to a communication beam prediction method and system based on environmental perception. Background Technology

[0002] In the intelligent connectivity scenarios targeted by the Sixth Generation (6G) mobile communication system, a type of system, represented by collaborative communication between infrastructure and mobile terminals, demonstrates high application potential. Its goal is to provide communication support for information interaction, collaborative perception, and collaborative decision-making among various intelligent terminals through high-speed and reliable connections between infrastructure and terminals.

[0003] Integrated Sensing and Communication (ISAC) technology combines sensing and communication functions on the same spectrum resources and hardware platform, enabling the collaborative design and joint utilization of sensing information and communication signals. This effectively reduces pilot and signaling overhead, improves spectrum utilization efficiency, and reduces system hardware costs and information interaction latency. Therefore, introducing ISAC into infrastructure and terminal collaborative communication systems helps achieve deep integration of communication and environmental perception in complex and dynamic environments, providing key technical support for the collaborative operation and intelligent applications of multi-agent systems.

[0004] A Reconfigurable Intelligent Surface (RIS) is a two-dimensional metasurface structure composed of numerous low-power, programmable electromagnetic units. It reconstructs the wireless propagation environment by controllably modulating the phase, amplitude, or polarization characteristics of incident electromagnetic waves, thereby constructing an equivalent virtual line-of-sight (LoS) propagation path without active RF transmission. Based on these characteristics, combining RIS with ISAC systems can simultaneously improve communication link quality and environmental awareness coverage in complex, obstructed environments. Furthermore, with the development of semi-passive RIS architectures, integrating some receiving sensing units into the RIS surface allows it to collect echo and environmental information while performing reflection modulation, introducing a new spatial dimension of observation, reducing information loss caused by multi-hop propagation and obstruction, and thus significantly improving the sensing accuracy and system robustness of intelligent communication systems in complex, dynamic electromagnetic environments.

[0005] Graph Frequency-Domain Attention is a method for feature modeling and prediction by introducing an attention mechanism into the graph frequency domain after the Graph Fourier Transform (GFT). This model models the temporal evolution, spatial distribution, and interaction topology of a multi-agent system in a unified way, mapping the original node domain information to a graph frequency domain representation. Low-frequency components typically characterize smooth collaborative behavior at the group level, while high-frequency components reflect rapid changes and sudden behaviors of individuals or localities. Based on this, an attention mechanism is introduced to adaptively weight different graph frequency components and their cross-node relationships, thereby achieving high-precision prediction of the state of complex dynamic systems.

[0006] In integrated communication and sensing systems, this type of model is particularly suitable for processing multi-target state information acquired by sensing units and provides effective support for subsequent communication parameter prediction and resource allocation.

[0007] Within the aforementioned generalized infrastructure and terminal cooperative communication framework, vehicle-to-infrastructure (V2I) communication, as a typical application scenario, places more stringent demands on beamforming prediction and communication robustness due to the characteristics of high-speed terminal movement, frequent multi-target interactions, and rapidly changing wireless propagation environments. Regarding the V2I beamforming design problem supported by ISAC, existing research mainly presents two implementation schemes that are closest to this invention.

[0008] Option 1: An ISAC beamforming prediction framework based on Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks, where the Rear State Unit (RSU) utilizes historical echo observations to predict the beamforming vectors for multiple vehicles in the next time slot through an unsupervised learning strategy. This method directly infers future beamforming decisions based on past Channel State Information (CSI), thus avoiding explicit channel modeling and online optimization processes. However, this work only focuses on the Loss-of-Sight (LoS) channel between the base station and the user terminal, and employs highly simplified kinematic assumptions in vehicle motion modeling, assuming the target travels along a straight trajectory without considering the more complex motion behaviors the target exhibits based on its surrounding environment.

[0009] Option 2: A two-stage RIS-ISAC beamforming framework based on CNNs. This option proposes deploying semi-passive reconfigurable smart surfaces on the infrastructure surface, constructing cascaded line-of-sight channels between the base station and the vehicle to effectively extend the system coverage, and introducing sensors at the RIS to receive signals. The model's first stage predicts the user's orientation angle through supervised learning, and the second stage uses manifold optimization to solve for beamforming. Although this option solves the coverage problem, it struggles to capture the complex motion characteristics of real-world targets due to considering only single-vehicle travel and a relatively simple motion model. Furthermore, the manifold optimization model requires separate optimization for different prediction results, resulting in high computational cost.

[0010] Existing research has effectively addressed several key challenges in intelligent collaborative communication and sensing systems to some extent, including high sensing signal overhead, high computational complexity of joint optimization, and system performance degradation caused by CSI lag or incompleteness. However, most existing beamforming and resource optimization methods still heavily rely on accurate and real-time updated channel state information, making it difficult to adapt to the rapidly changing time-varying characteristics of wireless channels caused by the high-speed movement of terminals or agents. Furthermore, traditional joint beamforming designs based on convex optimization or iterative search typically have high computational complexity, making it difficult to meet the stringent requirements of low latency and real-time response in application scenarios. In addition, existing methods often focus on a single terminal as the optimization object or neglect the interaction relationships between multiple terminals or agents during modeling, failing to fully characterize the comprehensive impact of multi-agent coupling behavior on communication and sensing performance in complex dynamic environments. Summary of the Invention

[0011] The purpose of this invention is to provide a communication beam prediction method and system based on environmental perception, which can more accurately predict the state information of mobile terminals or intelligent agents when direct communication links are blocked or channel quality is limited, and use perception information to assist in the design of communication parameters, thereby improving the downlink communication performance of the system.

[0012] The technical solution for implementing the present invention is as follows:

[0013] In a first aspect, the present invention provides a three-stage beam prediction method based on a graph frequency domain attention mechanism, applicable to an integrated communication and sensing system for collaborative communication between infrastructure and complex mobile terminals. The specific process is as follows: Step 1: Establish a joint optimization model with the goal of maximizing the downlink communication rate achievable by the system, and set the following constraints: (1) Use CRLB as the sensing performance index and introduce worst-case target constraints at the system level so that the sensing accuracy of all targets meets the preset threshold; (2) The RIS phase shift matrix meets the unit modulus constraint; (3) Base station transmit power constraint. Step 2: Configure LPB-Net including a multi-target state perception module, a graph frequency domain attention prediction module, and a beamforming module. In the first stage, the multi-target state perception module extracts the location features of the sensed targets from the echo signals received by the RIS sensing unit. In the second stage, the graph frequency domain attention prediction module obtains the spatiotemporal correlation between multiple targets based on the location features of the sensed targets and predicts the future channel state. In the third stage, the beamforming module embeds the maximum ratio transmission (MRT) into a deep learning framework, optimizes the model and constraints based on Step 1, and realizes a joint beamforming design that combines model-driven and data-driven approaches.

[0014] Optionally, the joint optimization model and constraints described in this invention are as follows:

[0015] in, This represents the base station's transmitted beam vector. This is the RIS phase shift matrix. and These represent the base station to RIS channel and the RIS to user channel, respectively. Represents the set of three-dimensional parameters for each target. The corresponding threshold for the comprehensive sensing accuracy index CRLB. Indicates time slot Transmit power budget for internal base stations Representation matrix The One diagonal element.

[0016] Optionally, the specific execution process of the multi-target state perception module of the present invention is as follows: First, the multi-target echo tensor of the RIS sensing unit in each historical time slot is obtained, and the echoes are separated and aligned to map to real values ​​of two channels: amplitude and phase. Secondly, multiple convolutional layers and max pooling layers are used to obtain high-dimensional feature tensors for the phase tensor, and the target vicinity is selected based on the amplitude tensor. One goal, retain The high-dimensional feature tensors of each target are flattened and processed into the information tensor of the current time slot t; Finally, the information tensor of the current time slot t is regressed through a fully connected layer to obtain the three-dimensional position and state sequence of each node within the historical window. .

[0017] Optionally, the specific execution process of the graph frequency domain attention prediction module of the present invention is as follows: First, construct the vehicle interaction graph at the current moment. , This represents a set of nodes, containing the service vehicle and its k key neighbors. Denotes the set of edges. The adjacency matrix, constructed from the pairwise spatial relationships between vehicles, is used to characterize the influence of neighboring vehicles on the target vehicle and the motion correlation of each vehicle; the degree matrix is ​​defined. ,in, ,in,[ express The elements are obtained based on the degree matrix and the adjacency matrix. Then, perform eigenvalue decomposition on it to obtain the eigenvector matrix. ; Secondly, using the eigenvector matrix node domain signal for time slot t Perform a graphical Fourier transform on the graph to obtain spectral coefficients in the graph frequency domain By introducing a learnable filter response in the graph frequency domain and weighting it with different graph frequency components, we obtain... ; Again, An attention module is introduced to model cross-band and cross-node coupling relationships, and these coupling relationships are then input into a fully connected layer to predict the graph frequency domain features of each target in the next time slot. ; Finally, the inverse graphical Fourier transform is used to extract the frequency domain features of the graph. By transforming back to the node domain, the predicted location of the next time slot is obtained. This is further mapped to the channel state quantity of the next time slot, the quasi-static RIS-vehicle channel. With base station – RIS channel .

[0018] Optionally, the specific working process of the beamforming module of the present invention is as follows: First, define an initial RIS phase shift matrix that satisfies the unity modulus constraint. ; Secondly, the equivalent downlink channel is constructed using the stage two output channel state information as follows:

[0019] The closed-form solution of the current beam vector is calculated based on the maximum transmission ratio (MRT), and the base station power constraint is satisfied through a normalization layer. :

[0020] Finally, the loss function is set as follows:

[0021] in, As a penalty weight, To achieve the required accuracy threshold, the system is trained until convergence is reached; the final output is the predicted RIS phase shift matrix. and base station beamforming vector .

[0022] Optionally, the multi-target state perception module and the graph frequency domain attention prediction module of the present invention are trained separately through supervised learning.

[0023] Secondly, the present invention provides a three-stage beam prediction system based on a graph frequency domain attention mechanism, comprising: a multi-target state perception module, a graph frequency domain attention prediction module, and a beamforming module; wherein, The multi-target state perception module is used to extract the position features of the perceived target from the echo signal received by the RIS perception unit; The image frequency domain attention prediction module is used to obtain the spatiotemporal correlation between multiple targets based on the perceived target location characteristics, and predict the future channel state. The beamforming module embeds the Maximum Ratio Transmission (MRT) into a deep learning framework, and implements a joint beamforming design that combines model-driven and data-driven approaches based on optimization models and constraints. The optimization model aims to maximize the downlink communication rate achievable by the system.

[0024] Beneficial effects: First, under the condition that the direct link is unavailable, sensing-assisted electromagnetic propagation control is achieved based on information received by a reconfigurable smart surface. This invention addresses complex environments where the direct electromagnetic propagation path between the transmitter and mobile terminal is obstructed or unavailable. It proposes a method for electromagnetic propagation control and directional transmission that relies solely on echo information acquired by the receiving unit at the RIS (Radio Receiving System). By introducing a sensing and receiving function at the RIS side, the system can reconstruct the equivalent cascaded propagation path based on the sensed echo without relying on real-time channel state information of the transmitter-terminal direct connection channel. This scheme fully utilizes environmental echo and target scattering information to achieve electromagnetic radiation direction and phase control assisted by sensing results, thereby significantly improving the transmission reliability and robustness of the system in obstructed environments.

[0025] Second, a graph frequency domain attention mechanism is introduced into the ISAC scenario to achieve joint prediction of the states of multiple agents. To address the challenges of multiple agents or terminals coexisting in an integrated communication and sensing system, where their motion states and spatial distributions are coupled, this application introduces a graph frequency domain attention mechanism for joint prediction of multi-target state information. By constructing a graph structure reflecting the spatial relationships and interactions between targets, the stable evolution of the group and sudden local changes are mapped to low-frequency and high-frequency components in the graph frequency domain. The attention mechanism is then used to adaptively model different frequency components and their cross-node associations, thereby achieving high-precision prediction of the future states of multiple agents. This prediction provides reliable prior information for subsequent electromagnetic propagation control and resource allocation.

[0026] Third, the model-driven MRT mechanism is embedded into the learning framework to achieve joint optimization of transmitter beam control and RIS phase modulation. To reduce the high computational complexity caused by non-convex optimization in traditional joint electromagnetic propagation control problems, this application embeds the MRT criterion as a model-driven module into the learning framework to guide the generation of the transmitter's radiation direction. The phase control parameters of the RIS are predicted by the learning model, and the directional radiation vector of the transmitter is directly constructed using the closed-form expression of the MRT, thus avoiding a complex iterative solution process. This "model-driven and data-driven" design significantly improves the algorithm's real-time performance and engineering feasibility while maintaining physical interpretability.

[0027] Taking a V2I downlink communication scenario as an example, simulation experiments were conducted based on a real vehicle trajectory dataset to verify the effectiveness of the proposed algorithm in terms of perception and communication performance. Simulation results show that the proposed three-stage LPB-Net can meet the CRLB constraint requirement for worst-case target perception accuracy, while achieving communication rate performance close to the perfect upper bound of CSI. Compared with existing single-target beam prediction methods, the proposed framework can more effectively characterize the collective dynamic characteristics of multiple vehicles, significantly improving the communication performance of communication users. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the three-stage beam prediction process.

[0030] Figure 2 A comparison chart showing the proposed three-stage model with existing methods. Detailed Implementation

[0031] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0033] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0034] This application provides an embodiment of a communication beam prediction method based on environmental perception, applicable to an integrated communication perception system where infrastructure and complex mobile terminals communicate collaboratively. The specific process is as follows: Step 1: Construction of an integrated communication and sensing system and modeling of sensing-assisted communication. Construction of an Integrated Communication and Sensing System: This invention constructs a semi-passive, reconfigurable smart surface RIS (Resonance Array) integrated communication and sensing system under conditions of blocked direct links. The communication base station and the semi-passive RIS work collaboratively to provide downlink communication services for a single terminal and perform environmental sensing for multiple targets. The RIS includes a reflective element array for regulating electromagnetic wave propagation and a sensing element array for receiving echo signals. In specific implementations, the reconfigurable smart surface RIS can employ reflective element arrays of different sizes and arrangements. The number, arrangement, and inclusion of active structures in the sensing elements can also be configured according to specific application scenarios. For example, sensing elements can be arranged only along a single direction, or a hybrid architecture combining some active and some passive elements can be used. These structural changes do not affect the overall technical solution of this invention. Based on this, an ISAC (Interactive Information Channel Control) transmission protocol is designed to process and reconstruct the echo signals received by the sensing elements, obtaining state information of multiple terminals or environmental targets. This state information is then used to assist in the design of communication parameters to improve downlink communication performance for the target terminal.

[0035] Furthermore, for multi-target sensing scenarios, a corresponding Cramer-Rao Lower Bound (CRLB) is derived to characterize the sensing accuracy requirements. Under the condition of satisfying the sensing accuracy constraints, a joint optimization model is established with the objective of maximizing the system's achievable downlink communication rate.

[0036] in, This represents the base station's transmitted beam vector. This is the RIS phase shift matrix. and These represent the base station to RIS channel and the RIS to user channel, respectively. Represents the set of three-dimensional parameters for each target. The corresponding threshold for the comprehensive sensing accuracy index CRLB. Indicates time slot Transmit power budget for internal base stations Representation matrix The One diagonal element.

[0037] The derivation process of the above joint optimization model is explained in detail below: In this system, the base station adopts a multiple-input single-output (MISO) architecture, and the baseband signal received by the user equipment can be represented as:

[0038] in, This represents the base station's transmitted beam vector. This is the RIS phase shift matrix. and These represent the base station to RIS channel and the RIS to user channel, respectively. Indicates the information symbols sent by the base station. This represents the receiver noise. The RIS phase shift matrix should satisfy the unity modulus constraint:

[0039] Meanwhile, the semi-passive RIS-integrated sensing unit array collects echoes from multiple vehicle targets in the environment, obtaining parameters such as azimuth, pitch, and range of the targets through array reception and spatial processing. To quantitatively evaluate sensing accuracy, CRLB is used as a sensing performance indicator, and a worst-case target constraint is introduced at the system level to ensure that the sensing accuracy of all targets meets a preset threshold, i.e.

[0040] in, , and They represent the first The target in the time slot The azimuth, elevation, and range parameters, The threshold for sensing accuracy.

[0041] Based on the aforementioned communication and sensing model, this invention jointly optimizes the base station beam vector and the RIS phase shift matrix within each time slot to maximize downlink communication performance. The corresponding optimization objective can be equivalently expressed as maximizing the received signal power.

[0042] And simultaneously meet the base station transmit power constraints

[0043] And the aforementioned perception accuracy and RIS unit modulus constraint.

[0044] Because the optimization variables are bilinearly coupled and simultaneously subject to non-convex constraints, the joint optimization problem has a high solution complexity, thus providing a technical background for introducing learning-driven prediction and beam design methods.

[0045] Step 2: A three-stage predictive beamforming framework based entirely on machine learning is proposed, named Localization-Prediction-Beamforming Network (LPB-Net), which includes three stages: localization, prediction, and beamforming. It is used to predict the beamforming parameters of future time slots, thereby effectively alleviating the problem of CSI obsolescence.

[0046] The LPB-Net consists of three sequentially executed phases: The first stage is the echo-based multi-target state perception stage, which uses a convolutional neural network to extract perception features from the echo signals received by the RIS perception unit and trains them through supervised learning. The second stage is the Graph Frequency-Domain Attention (GFA-Net) stage, which uses the overall characteristics of multi-agents to characterize the spatiotemporal correlation between multiple targets and achieves high-precision prediction of future channel states. The third stage is the beamforming design stage, which embeds Maximum Ratio Transmission (MRT) into a deep learning framework to achieve a joint beamforming design that combines model-driven and data-driven approaches under CRLB constraints.

[0047] This invention proposes a three-stage model for RIS-assisted ISAC vehicle-road cooperative scenarios. The overall model structure is shown below. Figure 1 The three-stage learning framework follows the causal relationship of information evolving step-by-step from "perception to prediction to decision." By decomposing the complex joint optimization problem into several learnable sub-problems, it effectively reduces the overall modeling and solving difficulty. After training in each stage, the model only performs forward inference during actual system operation, without requiring online optimization. This ensures perception accuracy while improving the robustness and real-time performance of communication beam prediction. The specific implementation steps for each stage are as follows: Phase 1: Location based on historical echoes First, obtain the RIS sensing unit in each historical time slot (time slot) Time slot Multi-target echo tensor within )

[0048] in, The number of separable targets, This represents the total number of horizontal and vertical receiving units. To set parameters.

[0049] The echoes are separated and aligned to target, and "target user + others" are selected based on user relevance. The "key neighbors" are then processed further. The complex echo is mapped to a two-channel real-value representation of "amplitude-phase":

[0050] in, Operators that convert complex numbers to real numbers.

[0051] For phase tensor High-dimensional feature tensors are obtained by using multi-kernel convolutional layers and max pooling layers, and then the magnitude tensors are used to... Filter out the area near the target Only keep this as an objective. The high-dimensional feature tensor of each target is processed through a flattening layer into the information tensor of the current time slot t.

[0052] Finally, the information tensor of the current time slot t is regressed through a fully connected layer to obtain the three-dimensional position and state sequence of each node within the historical window. The output is

[0053] In this context, the last dimension, 3, represents the location of the communication user, i.e., the nearby sensed target. This output serves as the input for Phase Two.

[0054] Phase Two: Prediction using Graph Frequency Domain Attention Mechanism To characterize the interaction between the target vehicle and its neighboring vehicles, a vehicle interaction graph is constructed in time slot t. ,in, This represents a set of nodes, containing the service vehicle and its k key neighbors. Denotes the set of edges. The adjacency matrix, constructed from the pairwise spatial relationships between vehicles, is used to characterize the influence of neighboring vehicles on the target vehicle and the motion correlation of each vehicle itself. In this embodiment, the construction method of the vehicle interaction graph is not limited to adjacency relationships based on physical distance; it can also be constructed by combining vehicle speed, lane information, or historical motion correlation. Further, a degree matrix is ​​defined. ,in, Based on this, the graph Laplacian matrix can be obtained. This embodiment uses a symmetric normalized graph Laplacian matrix.

[0055]

[0056] Its eigenvalue decomposition is expressed as

[0057] in, The eigenvector matrix, It is a diagonal matrix of eigenvalues.

[0058] The node domain signal of time slot t output in phase one Performing a graph Fourier transform (GFT) yields:

[0059] in, express The spectral coefficients in the graph frequency domain. Then, a learnable filter response is introduced in the graph frequency domain to weight different graph frequency components:

[0060] in, The weights of the graph-frequency filter represent the learnable parameters. Represents graph frequency information. This represents the spectral coefficients after filtering.

[0061] Will An attention module is introduced to model cross-band and cross-node coupling relationships, and then these coupling relationships are input into a fully connected layer to predict the graph frequency domain features of each target in the next time slot. .

[0062] Finally, the inverse graph Fourier transform (IGFT) is used to transform the information in the graph frequency domain back into the node domain to obtain the predicted location of the next time slot. :

[0063] This will then be mapped to the channel state quantity of the next time slot, the quasi-static RIS-vehicle channel. With base station – RIS channel .

[0064] Phase 3: Unsupervised Model-Driven Beamforming First, the channel state information output from stage two is concatenated and input into a deep Neural Network (DNN). An initial RIS phase shift matrix is ​​set, and the unity modulus constraint is satisfied through a normalization layer.

[0065] Then, the embedded MRT formula is used to calculate the determination. Beamforming matrix below .

[0066] Using the channel state information output from stage two, the following equivalent downlink channels can be constructed:

[0067] The closed-form solution of the current beam vector is calculated based on the MRT, and the base station power constraint is satisfied through the normalization layer. :

[0068] Finally, unsupervised training is used to obtain beamforming results with maximizing received power as the main objective, and a penalty term is introduced for the worst-case CRLB constraint.

[0069] Design a loss function for the current scenario:

[0070] in, As a penalty weight, To achieve the required accuracy threshold, the algorithm is trained until convergence. The final output is the predicted RIS phase shift matrix. and base station beamforming vector .

[0071] By transforming the non-convex optimization problem in traditional beamforming into a trainable, unconstrained learning objective, this stage effectively avoids the high computational complexity and latency overhead caused by iterative optimization, enabling beam prediction to be completed in real time in high-speed moving scenarios, which is particularly suitable for vehicle-road cooperative communication systems.

[0072] This embodiment discloses a three-stage beam prediction system based on a graph frequency domain attention mechanism, comprising: a multi-target state perception module, a graph frequency domain attention prediction module, and a beamforming module; wherein, The multi-target state perception module is used to extract the position features of the perceived target from the echo signal received by the RIS perception unit; The image frequency domain attention prediction module is used to obtain the spatiotemporal correlation between multiple targets based on the perceived target location characteristics, and predict the future channel state. The beamforming module embeds the Maximum Ratio Transmission (MRT) into a deep learning framework, and implements a joint beamforming design that combines model-driven and data-driven approaches based on optimization models and constraints. The optimization model aims to maximize the downlink communication rate achievable by the system.

[0073] Example: A communication beam prediction method is proposed to accurately predict user state information in obstructed spaces and utilize sensing-assisted communication to improve downlink communication performance. In millimeter-wave V2I scenarios, since signal propagation is primarily line-of-sight, downlink communication performance will significantly degrade when the direct path between the base station and user equipment is blocked. To address this issue, this invention introduces a semi-passive RIS (Reference Path Array), constructing a cascaded line-of-sight propagation path of "base station-RIS-user equipment" to effectively enhance the communication link in congested scenarios.

[0074] Simulation Scenarios and Parameter Settings Based on this example scenario, a simulation scenario is constructed to verify the effectiveness of the proposed method. The three-dimensional coordinates of the base station are as follows: A uniform linear array parallel to the ground is used. The system carrier frequency is set to... , the speed of light takes The number of base station antennas is The spacing between adjacent antennas is half a wavelength. ,in Maximum transmission power is .

[0075] RIS coordinates are The spacing between each unit in both the horizontal and vertical directions is half a wavelength, and the reflection phase can be within the range of... Internal continuous adjustment. The noise power at the communication receiver is... The noise power at the echo receiver is .

[0076] To simulate real-world vehicle motion behavior, this example uses a publicly available vehicle trajectory dataset to model a multi-vehicle scenario. Set as a time slot, within each time slot Sampling is performed to obtain each time slot A quick observation snapshot was taken. Under the above conditions, the effectiveness of the method of the present invention in terms of communication and sensing performance was verified by comparing it with a variety of existing methods.

[0077] Simulation Result Analysis: Figure 2 The communication performance of various beam prediction schemes under different RIS unit numbers was compared, with achievable communication rate used as the evaluation metric. The base station transmit power was fixed in the simulation. The number of base station antennas is set to The system performance was examined by changing the RIS aperture. The results show that the achievable rate of each scheme generally increases with the increase in the number of RIS units. However, the scheme based on outdated CSI exhibits increasingly significant performance degradation as the RIS size increases. This is because a larger RIS aperture results in a narrower main lobe, increasing the sensitivity of beamforming to CSI accuracy and amplifying the performance loss caused by outdated CSI. In contrast, the proposed three-stage beam prediction model (LPB-Net) achieves communication performance close to the upper bound of the ideal CSI under different RIS sizes. A comparative scheme is derived from a beam prediction method based on single-target information (PSEBP-Net) proposed for a similar scenario. This comparative model (EPSEBP-Net) was adjusted to enhance its adaptability to this scenario. LPB-Net exhibits more stable performance characteristics and achieves higher communication rates under all RIS configurations. This indicates that the method of this invention fully utilizes multi-target information for beam prediction, effectively improving the communication performance of end users.

[0078] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-stage beam prediction method based on a graph frequency domain attention mechanism, characterized in that, The specific process is as follows: Step 1: Establish a joint optimization model with the goal of maximizing the downlink communication rate achievable by the system, and set the following constraints: (1) Use CRLB as the sensing performance index and introduce worst-case target constraints at the system level so that the sensing accuracy of all targets meets the preset threshold; (2) The RIS phase shift matrix meets the unit modulus constraint; (3) Base station transmit power constraint. Step 2: Configure LPB-Net to include a multi-target state perception module, a graph frequency domain attention prediction module, and a beamforming module. In the first stage, the multi-target state perception module is used to extract the location features of the sensed targets from the echo signals received by the RIS sensing unit. In the second stage, the graph frequency domain attention prediction module is used to obtain the spatiotemporal correlation between multiple targets based on the location features of the sensed targets and predict the future channel state. In the third stage, the beamforming module embeds the maximum ratio transmission MRT into a deep learning framework, and based on the optimization model and constraints in step one, realizes a joint beamforming design that combines model-driven and data-driven approaches.

2. The three-stage beam prediction method based on graph frequency domain attention mechanism according to claim 1, characterized in that, The joint optimization model and constraints are as follows: in, This represents the base station's transmitted beam vector. This is the RIS phase shift matrix. and These represent the base station to RIS channel and the RIS to user channel, respectively. Represents the set of three-dimensional parameters for each target. The corresponding threshold for the comprehensive sensing accuracy index CRLB. Indicates time slot Transmit power budget for internal base stations Representation matrix The One diagonal element.

3. The three-stage beam prediction method based on graph frequency domain attention mechanism according to claim 2, characterized in that, The specific execution process of the multi-target state perception module is as follows: First, the multi-target echo tensor of the RIS sensing unit in each historical time slot is obtained, and the echoes are separated and aligned to be mapped to real values ​​of two channels: amplitude and phase. Secondly, multiple convolutional layers and max pooling layers are used to obtain high-dimensional feature tensors for the phase tensor, and the target vicinity is selected based on the amplitude tensor. One goal, retain The high-dimensional feature tensors of each target are flattened and processed into the information tensor of the current time slot t; Finally, the information tensor of the current time slot t is regressed through a fully connected layer to obtain the three-dimensional position and state sequence of each node within the historical window. That is, the node domain signal.

4. The three-stage beam prediction method based on graph frequency domain attention mechanism according to claim 3, characterized in that, The specific execution process of the graph frequency domain attention prediction module is as follows: First, construct the vehicle interaction graph at the current moment. , This represents a set of nodes, containing the service vehicle and its k key neighbors. Denotes the set of edges. The adjacency matrix, constructed from the pairwise spatial relationships between vehicles, is used to characterize the influence of neighboring vehicles on the target vehicle and the motion correlation of each vehicle; the degree matrix is ​​defined. ,in, ,in,[ express The elements are obtained based on the degree matrix and the adjacency matrix. Then, perform eigenvalue decomposition on it to obtain the eigenvector matrix. ; Secondly, using the eigenvector matrix node domain signal for time slot t Perform a graphical Fourier transform on the graph to obtain spectral coefficients in the graph frequency domain By introducing a learnable filter response in the graph frequency domain and weighting it with different graph frequency components, we obtain... ; Again, An attention module is introduced to model cross-band and cross-node coupling relationships, and these coupling relationships are then input into a fully connected layer to predict the graph frequency domain features of each target in the next time slot. ; Finally, the inverse graphical Fourier transform is used to extract the frequency domain features of the graph. By transforming back to the node domain, the predicted location of the next time slot is obtained. This is further mapped to the channel state quantity of the next time slot, the quasi-static RIS-vehicle channel. With base station – RIS channel .

5. The three-stage beam prediction method based on graph frequency domain attention mechanism according to claim 4, characterized in that, The specific working process of the beamforming module is as follows: First, define an initial RIS phase shift matrix that satisfies the unity modulus constraint. ; Secondly, the equivalent downlink channel is constructed using the stage two output channel state information as follows: The closed-form solution of the current beam vector is calculated based on the maximum transmission ratio (MRT), and the base station power constraint is satisfied through a normalization layer. : Finally, the loss function is set as follows: in, As a penalty weight, To achieve the required accuracy threshold, the system is trained until convergence is reached; the final output is the predicted RIS phase shift matrix. and base station beamforming vector .

6. The three-stage beam prediction method based on graph frequency domain attention mechanism according to claim 1, characterized in that, The multi-target state perception module and the graph frequency domain attention prediction module are trained separately through supervised learning.

7. A three-stage beam prediction system based on a graph frequency domain attention mechanism, characterized in that, include: The module comprises a multi-target state perception module, a graph-frequency domain attention prediction module, and a beamforming module; among which, The multi-target state perception module is used to extract the position features of the perceived target from the echo signal received by the RIS perception unit; The image frequency domain attention prediction module is used to obtain the spatiotemporal correlation between multiple targets based on the perceived target location characteristics, and predict the future channel state. The beamforming module embeds the Maximum Ratio Transmission (MRT) into a deep learning framework, and implements a joint beamforming design that combines model-driven and data-driven approaches based on optimization models and constraints. The optimization model aims to maximize the downlink communication rate achievable by the system.

8. The three-stage beam prediction system based on graph frequency domain attention mechanism according to claim 7, characterized in that, The specific execution process of the multi-target state perception module is as follows: First, the multi-target echo tensor of the RIS sensing unit in each historical time slot is obtained, and the echoes are separated and aligned to be mapped to real values ​​of two channels: amplitude and phase. Secondly, multiple convolutional layers and max pooling layers are used to obtain high-dimensional feature tensors for the phase tensor, and the target vicinity is selected based on the amplitude tensor. One goal, retain The high-dimensional feature tensors of each target are flattened and processed into the information tensor of the current time slot t; Finally, the information tensor of the current time slot t is regressed through a fully connected layer to obtain the three-dimensional position and state sequence of each node within the historical window. That is, the node domain signal.

9. The three-stage beam prediction system based on graph frequency domain attention mechanism according to claim 8, characterized in that, The specific execution process of the graph frequency domain attention prediction module is as follows: First, construct the vehicle interaction graph at the current moment. , This represents a set of nodes, containing the service vehicle and its k key neighbors. Denotes the set of edges. It is an adjacency matrix, constructed from the pairwise spatial relationships between vehicles, used to characterize the influence of neighboring vehicles on the target vehicle and the motion correlation of each vehicle itself. Define degree matrix ,in, ,in,[ express The elements are obtained based on the degree matrix and the adjacency matrix. Then, perform eigenvalue decomposition on it to obtain the eigenvector matrix. ; Secondly, using the eigenvector matrix node domain signal for time slot t Perform a graphical Fourier transform on the graph to obtain spectral coefficients in the graph frequency domain By introducing a learnable filter response in the graph frequency domain and weighting it with different graph frequency components, we obtain... ; Again, An attention module is introduced to model cross-band and cross-node coupling relationships, and these coupling relationships are then input into a fully connected layer to predict the graph frequency domain features of each target in the next time slot. ; Finally, the inverse graphical Fourier transform is used to extract the frequency domain features of the graph. By transforming back to the node domain, the predicted location of the next time slot is obtained. This is further mapped to the channel state quantity of the next time slot, the quasi-static RIS-vehicle channel. With base station – RIS channel .

10. The three-stage beam prediction system based on graph frequency domain attention mechanism according to claim 9, characterized in that, The specific working process of the beamforming module is as follows: First, define an initial RIS phase shift matrix that satisfies the unity modulus constraint. ; Secondly, the equivalent downlink channel is constructed using the stage two output channel state information as follows: The closed-form solution of the current beam vector is calculated based on the maximum transmission ratio (MRT), and the base station power constraint is satisfied through a normalization layer. : Finally, the loss function is set as follows: in, As a penalty weight, To achieve the required accuracy threshold, the system is trained until convergence is reached; the final output is the predicted RIS phase shift matrix. and base station beamforming vector .

Images