A smart sensing-assisted end-to-end beamforming prediction method and system

Abstract

This invention belongs to the field of integrated machine learning, communication, and sensing technologies. Specifically, it relates to an intelligent sensing-assisted end-to-end beamforming prediction method and system. This method can more accurately predict the state information of mobile terminals or intelligent agents when direct communication links are blocked or channel quality is limited, thereby improving the beam prediction performance of the system's downlink communication. This invention constructs a complex tensor from historical multi-slot, multi-target echo observations as input to the end-to-end model. The echoes undergo structured preprocessing, and a dual-branch approach (amplitude / phase) is used to extract a more favorable representation for communication beam prediction. The amplitude branch provides stable intensity and correlation clues, while the phase branch carries spatial angle information and forms a multi-target embedding. Simultaneously, based on indicators such as amplitude or correlation, the method selects the k most relevant neighboring targets from the multi-sensing targets to participate in subsequent predictions, thereby improving the effective information density and prediction robustness from the source.

Description

Technical Field

[0001] This invention belongs to the field of integrated machine learning, communication and sensing technology, and specifically relates to an intelligent sensing-assisted end-to-end beamforming prediction method and system. Background Technology

[0002] As wireless communication evolves towards higher frequencies, large-scale arrays, and intelligence, high-frequency links such as millimeter waves are extremely sensitive to spatial pointing and environmental characteristics. System performance is heavily influenced by nonlinear and time-varying factors such as terminal location, obstruction, and motion. Traditional methods relying on precise physical modeling (such as channel parameter estimation and geometric derivation) face significant challenges in complex dynamic environments. They struggle to characterize irregular reflectors and random interference, and frequent parameter updates lead to excessive computational complexity and latency, making it difficult to meet real-time service requirements. To address this, Integrated Communication Sensing (ISAC) technology has emerged, aiming to acquire environmental and terminal information through integrated sensing functions to assist in beam selection and resource allocation, becoming a key path to improve communication performance in complex scenarios.

[0003] Despite the promising future of ISAC, existing perception-assisted communication schemes mostly employ a phased architecture of "first sensing and estimating intermediate parameters, then making communication decisions." These methods heavily rely on the accuracy of intermediate variables; once perception errors accumulate or information is incomplete, overall performance will significantly degrade, and multi-step inference increases system latency and implementation complexity. In contrast, the end-to-end architecture directly establishes a mapping relationship from "perception input" to "communication output," eliminating the dependence on explicit intermediate physical parameters. This architecture is not only structurally simple and inference-compact, but also implicitly integrates physical constraints into the learning process through model design and optimization objectives, automatically uncovering the intrinsic relationship between perception information and communication strategies, providing a new approach to achieving efficient, low-latency, robust communication.

[0004] Within the generalized end-to-end sensing-assisted communication framework, vehicle-to-infrastructure (V2I) communication, as a typical application scenario, presents more stringent requirements for beamforming prediction due to the characteristics of high-speed target movement, rapidly changing environment, and frequent multi-target interactions. Existing research for this scenario mainly focuses on two similar implementation schemes: one utilizes the multidimensional echo signals from base stations and reconfigurable smart surfaces (RIS) to directly predict transmitted and reflected beams via deep neural networks, avoiding explicit channel reconstruction; the other, based on historical channel observations, uses a coding-prediction-decoding structure to directly output beam parameters for future time slots, and introduces time-series modeling and the Cramer-Rao lower bound (CRLB) index to achieve joint optimization of sensing performance.

[0005] Scholars have proposed a perception-assisted end-to-end beamforming method utilizing ISAC echo information. This method statistically models multidimensional echo signals and inputs the models into a neural network to uniformly complete feature extraction and time-varying characteristic modeling, directly outputting beamforming results that satisfy constraints. This effectively reduces modeling complexity and improves real-time performance in high-speed scenarios. However, the main drawback of this approach is that its modeling object is limited to a single communication user and a single sensing target, failing to explicitly characterize the spatial coupling relationships between multiple targets. This results in insufficient information dimensions available to the system in complex traffic flows or scenarios with dense multi-target traffic, severely limiting the prediction accuracy of future time-slot beam configurations.

[0006] While some solutions introduce multi-target input and characterize motion correlation through temporal units, their processing of multi-target information typically relies on simple concatenation or uniform encoding, lacking structured modeling and feature decoupling mechanisms for the mutual influence between targets, making it difficult to fully explore the deep correlations of multi-source information. Furthermore, these solutions assume direct channels between multiple base stations and users, which often fails in typical scenarios such as urban highways or urban canyons where direct links are easily obstructed, thus affecting the actual performance of sensory collaboration. In summary, existing technologies have not yet simultaneously solved the problems of missing structured interaction modeling for multiple targets and poor adaptability to complex non-line-of-sight (NLOS) scenarios, which is precisely the technical challenge this invention aims to overcome. Currently, large models used in the industry (such as GPT-4 and Qwen2.5) have limited single-round maximum input text capabilities; for example, Qwen2.5 supports a maximum of 120,000 Chinese characters. Real-world business scenarios present a need for accurate field extraction from extremely long texts (such as large contracts of 500,000 characters), far exceeding the processing limits of large models. Currently, mainstream methods include RAG (Retrieval Enhanced Generation), long text summary variants such as MapReduce, pure fragmentation-summary concatenation (e.g., CN119783666A), cross-fragment summary generation, and long text question answering. Among these, RAG (Retrieval Enhanced Generation) involves retrieving first and then extracting, but long-term practice shows that its accuracy is often below 70%, and recall errors can lead to the omission of core fields. Long text summary variants such as MapReduce are only suitable for introductions or summaries and are not suitable for structured field extraction. Some related patent literature focuses on pure fragmentation-summary concatenation (e.g., CN119783666A), cross-fragment summary generation, and long text question answering, but these generally suffer from pain points such as field loss, fragmentation, and low efficiency, and lack an integrated extraction mechanism of "dynamic question grouping - fragmented topic matching - trusted global fusion".

[0007] Currently, there is a lack of a practical solution that finds the optimal technical balance between extraction accuracy and recall efficiency, can dynamically adapt to different document structures and field requests, and can overcome the maximum text input limit. Summary of the Invention

[0008] In view of this, the present invention provides an intelligent sensing-assisted end-to-end beamforming prediction method and system, which can more accurately predict the state information of mobile terminals or intelligent agents when the direct communication link is blocked or the channel quality is limited, and use the sensing information to assist in the design of communication parameters, thereby improving the beamforming prediction of the system's downlink communication performance.

[0009] To achieve the objectives of this invention, the following technical solutions are provided.

[0010] A smart perception-assisted end-to-end beamforming prediction method includes the following steps:

[0011] Step 1, Data Processing: Input the echo signals of all sensed targets received by the devices in the historical τ time slots, construct a complex input tensor, perform amplitude and phase decomposition on the complex input tensor to obtain a real tensor, select the k key targets most relevant to the communication user based on the amplitude part, input the selected target information into a convolutional neural network to extract features, and obtain a high-dimensional feature tensor of the historical echoes; Step 2, Graph Frequency Domain Prediction: The feature tensors of k+1 targets from the historical τ time slots are used as input. A fully connected graph is constructed based on all targets. The multi-target embedding features of each time slot are transformed to the graph frequency domain by graph Fourier transform. Based on the attention mechanism, the graph frequency domain features between multiple time slots are analyzed and predicted to obtain the features of multi-targets in the graph frequency domain of future time slots. The prediction results are restored to the time domain by inverse graph Fourier transform. After flattening and feature concatenation, the results are output. Step 3, Beam generation and optimization: The reconfigurable smart surface RIS phase shift matrix of the corresponding dimension is output by a deep neural network. The base station beamforming vector is calculated using the maximum ratio transmission as a known quantity using the RIS phase shift matrix. A loss function containing perception constraints is designed. The model parameters are updated synchronously through an end-to-end architecture for iterative training, and the beam prediction results for future time slots are output.

[0012] Specifically, constructing the complex input tensor in step 1 includes: If the receiving sensors consist of multiple sets, they are combined into a complex input tensor with a dimension of number of targets × number of sensors × 1 by tensor concatenation. The complex input tensor corresponding to the sensor is flipped and spliced ​​into a tensor with a dimension of target number × sensor number × 2, so that the echo characteristics between different sensors can be captured by using convolutional layers.

[0013] Specifically, step 1, which involves selecting the k key targets most relevant to the communication user, includes: By leveraging the relative stability of the magnitude component of the target, the correlation between each target and the user is calculated through pooling operations; Select the k targets that are closest to the user, and use the information of the k targets and the user as a total of k+1 targets as input for subsequent processing.

[0014] Specifically, step 2, which involves analyzing and predicting the graph frequency domain features across multiple time slots based on an attention mechanism, includes: The multi-target embedding features of each time slot are mapped to the graph frequency domain through graph Fourier transform, so that the features are decomposed into components with spatial variation patterns. Attention weighting is applied to cross-time slot features in the graph frequency domain to capture the spatiotemporal characteristics of actual vehicle motion and predict the feature representation of multiple targets in the future time slot graph frequency domain.

[0015] Specifically, step 3, which outputs the RIS phase shift matrix of the corresponding dimension, includes: The phase shift matrix is ​​output by a deep neural network and matched with the number of RIS units. The output phase shift matrix is ​​ensured to satisfy the unit modulus physical constraint, that is, the modulus of the diagonal elements of the matrix is ​​1.

[0016] Specifically, step 3, designing the loss function that includes perceptual constraints, includes: The loss function consists of a communication performance component and a perception constraint component; The communication performance section aims to maximize the received signal power. The perception constraint section uses the Cramer-Rao lower bound (CRLB) as the perception performance index and introduces the worst-case target constraint to ensure that the perception accuracy of all targets meets the preset threshold.

[0017] This method is applicable to urban canyon scenarios where direct links are completely blocked. By introducing a semi-passive RIS, it utilizes the echo received by its sensors to cancel signal attenuation, thus achieving a physical closed loop from sensing input to executable link configuration.

[0018] The present invention also provides an intelligent sensing-assisted end-to-end beamforming prediction system for implementing the method described in the present invention, comprising: The data processing module is used to receive and preprocess the sensing echo signals from τ historical time slots, including a tensor construction unit, an amplitude-phase decomposition unit, a target screening unit, and a feature extraction unit, to generate a high-dimensional feature tensor of the historical echoes. The graph frequency domain prediction module is used for time series prediction based on graph topology. It includes a graph topology building unit, a graph Fourier transform unit, an attention prediction unit, and an inverse transform recovery unit, and outputs the time domain prediction results of future time slots. The output module is used to generate executable communication link configuration parameters, including a RIS phase shift generation unit, a base station beamforming calculation unit, a joint optimization unit, and a result output unit, which outputs the base station beamforming vector and the RIS phase shift matrix.

[0019] Beneficial effects 1. This invention constructs a complex tensor from historical multi-slot, multi-target echo observations as the input to an end-to-end model. The echoes are preprocessed in a structured manner, and a dual-branch approach (amplitude / phase) is used to extract a characterization that is more conducive to communication beam prediction. The amplitude branch provides stable intensity and correlation clues, while the phase branch carries spatial angle information and forms a multi-target embedding. At the same time, based on indicators such as amplitude or correlation, the kkk neighbor targets most relevant to the target communication user are selected from the multi-sensor targets to participate in subsequent prediction, thereby improving the effective information density and prediction robustness from the source.

[0020] 2. This invention integrates a graph frequency domain attention mechanism into the end-to-end prediction link: a vehicle interaction graph reflecting spatial proximity or echo correlation is constructed using the target communication user and its kkk neighboring targets. The multi-target embedded features of each time slot are mapped to the graph frequency domain through graph Fourier transform, so that the features are decomposed into spatial change patterns with clear physical meaning. Then, attention prediction is performed on cross-time slot features in the graph frequency domain and the future time slot representation is recovered through inverse graph Fourier transform. Thus, the multi-target coupling relationship is explicitly characterized in the manner of "graph topology - graph frequency domain - attention", realizing interpretable spatiotemporal fusion prediction and improving adaptability to complex dynamic environments.

[0021] 3. This invention introduces a semi-passive RIS for scenarios with obstructed direct links. It uses the RIS to reflect the downlink signal of the base station and receive the echo of the sensing target to form a closed-loop configuration. The end-to-end model directly outputs the RIS phase shift matrix that meets the physical feasibility constraints such as unit modulus. It also combines the equivalent channel with the maximum ratio transmission to generate the base station beam for future time slot link configuration. At the same time, the sensing accuracy constraint is embedded in the loss function in the form of CRLB worst-case scenario for end-to-end training. Under the premise of ensuring the lower limit of sensing, it significantly improves the communication rate of the target user and the system coordination performance. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of a system according to an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram showing the simulation comparison results of the training convergence characteristics of the method in the embodiment of the present invention and the existing method. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0025] To address the shortcomings of existing technologies, such as lack of multi-target interaction modeling, insufficient feature fusion, and poor adaptability to strong occlusion scenarios, this invention proposes an intelligent perception-assisted end-to-end beamforming prediction method. Targeting "single-user + multi-target" scenarios, it constructs historical multi-timeslot echoes as tensors and extracts features using amplitude-phase dual-branch extraction, intelligently selecting key targets to maximize information utilization. It innovatively introduces a graph topology structure, explicitly fusing multi-target coupling relationships in the graph frequency domain using an attention mechanism and directly predicting high-dimensional features. Future timeslot representations are then recovered through inverse transformation, effectively overcoming the information dilution problem of traditional spatiotemporal methods. Simultaneously, for typical urban scenarios with completely blocked direct links, a semi-passive RIS is introduced to utilize sensor echoes to offset signal attenuation, directly outputting the RIS phase shift and the base station's maximum ratio transmission beam that satisfy the unit-mode physical constraint, achieving a physical closed loop from perception input to executable link configuration. The overall architecture deeply embeds the physical characteristics of target motion in a pure end-to-end process without generating any explicit intermediate variables, significantly reducing system latency and computational overhead, and achieving higher accuracy and stronger robustness in beamforming prediction performance under complex dynamic environments. The intelligent sensing-assisted end-to-end beamforming prediction method of the present invention includes the following steps: Step 1, Enter History Each time-slot device receives echoes from all sensed targets and processes the data.

[0026] If there are multiple sets of receiving sensors, they are combined into a complex input tensor with dimensions of target number × sensor number × 1 through tensor concatenation. Then, this tensor is concatenated by flipping the first and last sensors to form a tensor with dimensions of target number × sensor number × 2. This facilitates subsequent steps using convolutional layers to capture the echo characteristics between different sensors. Finally, the amplitude and phase of the tensor are decomposed to obtain two sets of real tensors with consistent dimensions for use in machine learning.

[0027] The amplitude of the target is relatively stable. After pooling, the k targets closest to the user are selected, and these targets have the main impact on the user's motion behavior. The phase part mainly contains spatial angle information. The information of the selected k+1 targets is input into a convolutional neural network to obtain features, and the high-dimensional feature tensor of the historical echo is obtained by using the embedding layer as the output.

[0028] Step Two: History The graph frequency domain prediction is performed using the feature tensors of k+1 targets in each time slot as input, where targets in different time slots have multi-dimensional temporal and spatial topological relationships. Specifically, a fully connected graph is constructed based on all targets, with the embedded feature tensor of each node serving as the node value. For the fusion dimension, a graph Fourier transform is performed on the embedded features of multiple targets in each time slot, and prediction is made using the graph frequency domain features across multiple time slots. Features of multiple targets in the graph frequency domain of future time slots are predicted based on an attention mechanism. Then, based on fixed topological information, the inverse graph frequency domain transformation is used to obtain the temporal prediction result, which is flattened and feature-concatenated as the output.

[0029] Step 3: Utilize a deep neural network to output the RIS phase shift matrix of the corresponding dimension. Then, using the beamforming vector as a known quantity, calculate the current optimal base station beamforming vector using maximum ratio transmission. Design a loss function under perceptual constraints, aiming to maximize user communication rate. The end-to-end architecture is characterized by the loss function synchronously updating all parameters in the model, directly training the optimal model through iteration, and finally outputting the beam prediction results for future time slots, the base station beamforming vector, and the RIS phase shift matrix.

[0030] This invention also proposes an intelligent sensing-assisted end-to-end beamforming prediction system to implement the aforementioned beamforming prediction method. The system includes a data processing module, a graph frequency domain prediction module, and an output module. These modules work collaboratively to complete the end-to-end prediction process from sensing echo input to beam configuration output. The system is as follows: Figure 1 As shown, a pure end-to-end architecture is adopted, eliminating the need to generate any explicit intermediate variables (such as channel matrices, position coordinates, etc.). Beam prediction can be completed using only historical echoes as input. The system deeply embeds the physical characteristics of target motion, achieving high-precision and robust beam prediction performance in complex dynamic environments while reducing system latency and computational overhead. It is particularly suitable for typical scenarios such as urban canyons where direct links are completely blocked.

[0031] The data processing module is responsible for receiving and preprocessing the sensing echo signals from the historical τ time slots, specifically including: Tensor building unit: When there are multiple receiving sensors, they are combined into a complex input tensor with a dimension of target number × number of sensors × 1 by tensor splicing. Then, the tensor corresponding to the sensor is flipped and spliced ​​into a tensor with a dimension of target number × number of sensors × 2, which makes it easier for subsequent convolutional layers to capture the echo features between different sensors. Amplitude-phase decomposition unit: Performs amplitude and phase decomposition on a complex tensor to obtain two sets of real tensors with consistent dimensions for machine learning use; Target filtering unit: Utilizing the relative stability of the amplitude component, the unit filters the k key targets closest to the user through pooling operations. These targets have a major impact on the user's motion behavior. Feature extraction unit: Input the selected k+1 target information into the convolutional neural network to extract features, and use the embedding layer to generate a high-dimensional feature tensor of historical echoes as the module output.

[0032] The graph frequency domain prediction module is responsible for performing time series prediction based on the graph topology, specifically including: Graph topology building unit: Construct a fully connected graph based on all targets, using the embedded feature tensor of each node as the node value, and establish multi-dimensional temporal and spatial topological relationships between targets; Graph Fourier Transform Unit: Performs graph Fourier transform on the multi-target embedding features of each time slot, transforming the spatial domain features to the graph frequency domain for fusion processing; Attention Prediction Unit: Based on the attention mechanism, it analyzes the frequency domain features of the graph across multiple time slots and predicts the feature representation of multiple targets in the frequency domain of the future time slot graph. Inverse Transform Recovery Unit: Based on fixed topological information, the inverse graph Fourier transform is used to recover the graph frequency domain prediction results to the time domain, and the results are flattened and feature-stitched as the module output.

[0033] The output module is responsible for generating executable communication link configuration parameters, specifically including: RIS phase shift generation unit: Utilizes a deep neural network to output a RIS phase shift matrix of the corresponding dimension, and ensures that the output satisfies physical constraints such as unity modulus; Base station beamforming unit: Using the RIS phase shift as a known quantity, calculate the current optimal base station beamforming vector using the maximum ratio transmission (MRT) method; Joint Optimization Unit: Design a loss function under perceptual constraints with the optimization objective of maximizing user communication rate. Update all parameters in the model synchronously through an end-to-end architecture and obtain the optimal model through iterative training. Result output unit: Outputs beam prediction results for future time slots, including base station beamforming vectors and RIS phase shift matrix, which can be directly used for actual link configuration.

[0034] Using a semi-passive RIS-assisted ISAC vehicle-to-everything (V2X) scenario as a typical example, this paper verifies the effectiveness of the model proposed in this invention. This example scenario involves urban road V2X communication, with the ISAC base station handling downlink communication for a single user. Because the direct channel from the base station to the user is obstructed, a semi-passive RIS is introduced, capable of both reflecting communication signals and receiving echo signals. Furthermore, the user is moving rapidly on the highway, and their behavior is influenced by several surrounding vehicles. The purpose of this example is to address the beam alignment problem in this scenario, where perception information is outdated, based on the end-to-end model proposed in this invention using graph frequency domain attention. While ensuring multi-vehicle target perception, the model fully utilizes information to assist in beam prediction, thereby optimizing the communication performance of the user's vehicle.

[0035] The specific scenario setting is as follows: the three-dimensional coordinates of the ISAC base station are... 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 The semi-passive 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 The 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.

[0036] Communication and sensing are modeled within the scenario. The baseband signal received by the user equipment can be represented as...

[0037] This represents the base station's transmitted beam vector. It is a RIS phase shift matrix that satisfies the unity modulus constraint.

[0038] and These represent the base station to RIS channel and the RIS to user channel, respectively.

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

[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] In addition to the aforementioned sensing accuracy and RIS unit modulus constraints, the optimization problem is difficult to solve due to its non-convex nature. Furthermore, the channel state from the RIS to the user terminal changes rapidly, and the environmental information in this example is unknown. Therefore, the end-to-end architecture proposed in this invention avoids explicit estimation and prediction of intermediate variables, directly solving for model parameters to achieve beam prediction. Thus, this invention is suitable for solving the problem in this example.

[0044] The specific implementation is as follows: The input to the end-to-end model is the RIS sensing unit in history. Multi-target echo tensor within a time slot

[0045] in The number of separable targets, This represents the total number of horizontal and vertical receiving units. The echoes are target-separated and aligned, and "target user + its..." is selected based on user relevance. The "key neighbors" are then processed. In the data processing module, a high-dimensional feature tensor is generated. In the graph frequency domain prediction module, a vehicle interaction graph is constructed using the target user and k neighbors as nodes. Multiple graph frequency components are fused in the graph frequency domain, and predictions are made based on an attention mechanism to capture the spatiotemporal characteristics of actual vehicle motion. For example, the low-frequency components of the feature tensor represent the commonalities among multiple targets, while the high-frequency components represent the specific behaviors of individual targets. The features from this module are input into the output module of the overall model, and a fully connected layer is used to obtain a RIS phase shift matrix that matches the number of RIS units and satisfies the modulus constraint. Based on the known RIS phase shift matrix... The base station beamforming vector is calculated using the maximum ratio transmission formula.

[0046] Design a loss function with perceptual constraints.

[0047] The first part describes the communication performance, and the second part describes the perception constraints. The smaller the overall loss function, the better it meets the requirements of this example. Unsupervised learning is used iteratively until convergence, ultimately yielding the optimal model parameters.

[0048] Simulation results analysis as follows Figure 2 As shown, the proposed EGFA-Net and two existing end-to-end learning models (i.e., PSEBP-Net and its enhanced version EPSEBP-Net) are compared in terms of base station transmit power. The training convergence characteristics under different conditions were compared and analyzed. Compared with PSEBP-Net, EGFA-Net explicitly incorporates the influence of surrounding targets on communication users in its network structure design and utilizes the characteristics of echoes in the graph frequency domain for beamforming prediction, thereby more fully exploring the correlation between perception information and communication performance. Meanwhile, EGFA-Net enhances the input processing module of the original PSEBP-Net, making it better suited to the perception-assisted communication application scenarios involved in this invention. Experimental results show that all three methods can achieve stable convergence under different parameter configurations. Among them, EGFA-Net not only has a higher system reachability rate than PSEBP-Net, but also exhibits faster and more stable convergence characteristics during training. Although PSEBP-Net has a faster convergence speed, its final reachability rate is significantly lower. In summary, the results show that EGFA-Net can effectively capture multi-target perception information and use it to enhance single-user communication performance, thereby significantly improving the system reachability rate.

[0049] This invention includes, but is not limited to, the above embodiments. Any equivalent substitutions or partial improvements made under the spirit and principles of this invention shall be considered within the scope of protection of this invention.

Claims

1. A smart sensing-assisted end-to-end beamforming prediction method, characterized in that, Includes the following steps: Step 1, Data Processing: Input the echo signals of all sensed targets received by the devices in the historical τ time slots, construct a complex input tensor, perform amplitude and phase decomposition on the complex input tensor to obtain a real tensor, select the k key targets most relevant to the communication user based on the amplitude part, input the selected target information into a convolutional neural network to extract features, and obtain a high-dimensional feature tensor of the historical echoes; Step 2, Graph Frequency Domain Prediction: The feature tensors of k+1 targets from the historical τ time slots are used as input. A fully connected graph is constructed based on all targets. The multi-target embedding features of each time slot are transformed to the graph frequency domain by graph Fourier transform. Based on the attention mechanism, the graph frequency domain features between multiple time slots are analyzed and predicted to obtain the features of multi-targets in the graph frequency domain of future time slots. The prediction results are restored to the time domain by inverse graph Fourier transform. After flattening and feature concatenation, the results are output. Step 3, Beam generation and optimization: The reconfigurable smart surface RIS phase shift matrix of the corresponding dimension is output by a deep neural network. The base station beamforming vector is calculated using the maximum ratio transmission as a known quantity using the RIS phase shift matrix. A loss function containing perception constraints is designed. The model parameters are updated synchronously through an end-to-end architecture for iterative training, and the beam prediction results for future time slots are output.

2. The method according to claim 1, characterized in that, The construction of the complex input tensor in step 1 specifically includes: If the receiving sensors consist of multiple sets, they are combined into a complex input tensor with a dimension of number of targets × number of sensors × 1 by tensor concatenation. The complex input tensor corresponding to the sensor is flipped and spliced ​​into a tensor with a dimension of target number × sensor number × 2, so that the echo characteristics between different sensors can be captured by using convolutional layers.

3. The method according to claim 1, characterized in that, Step 1, which involves selecting the k most relevant key targets to the communication user, specifically includes: By leveraging the relative stability of the magnitude component of the target, the correlation between each target and the user is calculated through pooling operations; Select the k targets that are closest to the user, and use the information of the k targets and the user as a total of k+1 targets as input for subsequent processing.

4. The method according to claim 1, characterized in that, Step 2, which involves analyzing and predicting the graph frequency domain features across multiple time slots based on an attention mechanism, specifically includes: The multi-target embedding features of each time slot are mapped to the graph frequency domain through graph Fourier transform, so that the features are decomposed into components with spatial variation patterns. Attention weighting is applied to cross-time slot features in the graph frequency domain to capture the spatiotemporal characteristics of actual vehicle motion and predict the feature representation of multiple targets in the future time slot graph frequency domain.

5. The method according to claim 1, characterized in that, Step 3, which outputs the corresponding dimension of the RIS phase shift matrix, specifically includes: The phase shift matrix is ​​output by a deep neural network and matched with the number of RIS units. The output phase shift matrix is ​​ensured to satisfy the unit modulus physical constraint, that is, the modulus of the diagonal elements of the matrix is ​​1.

6. The method according to claim 1, characterized in that, Step 3, designing the loss function that includes perceptual constraints, specifically includes: The loss function consists of a communication performance component and a perception constraint component; The communication performance section aims to maximize the received signal power. The perception constraint section uses the Cramer-Rao lower bound (CRLB) as the perception performance index and introduces the worst-case target constraint to ensure that the perception accuracy of all targets meets the preset threshold.

7. The method according to any one of claims 1 to 6, characterized in that, This method is applicable to urban canyon scenarios where direct links are completely blocked. By introducing a semi-passive RIS, it utilizes the echo received by its sensors to cancel signal attenuation, thus achieving a physical closed loop from sensing input to executable link configuration.

8. A smart perception-assisted end-to-end beamforming prediction system, used to implement the method as described in any one of claims 1 to 7, characterized in that, include: The data processing module is used to receive and preprocess the sensing echo signals from τ historical time slots, including a tensor construction unit, an amplitude-phase decomposition unit, a target screening unit, and a feature extraction unit, to generate a high-dimensional feature tensor of the historical echoes. The graph frequency domain prediction module is used for time series prediction based on graph topology. It includes a graph topology building unit, a graph Fourier transform unit, an attention prediction unit, and an inverse transform recovery unit, and outputs the time domain prediction results of future time slots. The output module is used to generate executable communication link configuration parameters, including a RIS phase shift generation unit, a base station beamforming calculation unit, a joint optimization unit, and a result output unit, which outputs the base station beamforming vector and the RIS phase shift matrix.

Images