Beam prediction method based on radar environment perception and RIS power distribution mode

Abstract

The invention discloses a beam prediction method based on radar environment perception and an RIS power distribution mode, and belongs to the technical field of communication, and the method comprises the steps: obtaining radar point cloud data of each time step, and carrying out the feature extraction of the radar point cloud data through a multi-layer perceptron and a maximum pooling layer, and obtaining radar point cloud features; obtaining an RIS power delay distribution data graph, and extracting RIS power delay distribution characteristics by using a CNN convolution layer and position coding; fusing the two features by using a cross-modal fusion module; a Transform encoder is used to encode the fusion feature, and an encoding feature containing space-time dependence is obtained; calculating the dynamic weight of each target, and performing weighted aggregation on the coding features by using the dynamic weights; and a Transform decoder is used to decode the weighted aggregation feature, and a beam prediction result is obtained. According to the method, the beam prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and more specifically to a beam prediction method based on radar environment perception and RIS power distribution pattern. Background Art

[0002] Beam prediction refers to predicting the optimal transmission direction and beam parameters of a signal through algorithms or models in a wireless communication system, so as to achieve more efficient signal coverage and improved communication quality.

[0003] Currently, beam prediction usually combines artificial intelligence (AI) and machine learning (ML) to achieve accurate prediction.

[0004] However, existing beam prediction technologies have many deficiencies, such as poor adaptability to dynamic environments and inability to solve the beam splitting problem caused by multipath effects, resulting in the beam prediction accuracy not meeting the requirements.

[0005] Therefore, how to improve the accuracy of beam prediction is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a beam prediction method based on radar environment perception and RIS power distribution pattern

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] Obtain the radar point cloud data at each time step in the radar time series data, and use a multi-layer perceptron and a max pooling layer to extract features from the radar point cloud data at each time step to obtain radar point cloud features;

[0009] Obtain the RIS power delay distribution data map, and use a CNN convolutional layer and position encoding to extract RIS power delay distribution features;

[0010] Use a cross-modal fusion module to fuse the extracted radar point cloud features and RIS power delay distribution features to obtain fused features;

[0011] Use a Transformer encoder to encode the obtained fused features to obtain encoded features containing spatio-temporal dependencies;

[0012] Calculate the dynamic weight of each target, and use the dynamic weight to perform weighted aggregation on the encoded features to obtain weighted aggregated features;

[0013] Use a Transformer decoder to decode the weighted aggregated features and obtain the beam prediction result.

[0014] Further, obtain the radar point cloud data at each time step in the radar time series data, including the following radar point cloud data sets:

[0015] P t ={p i |p i =(x i ,y i ,z i ,v y,i ,σ i ) ∈R 6}, where the radar point cloud data set P t each data point p i includes the position (x i ,y i ,z i ), velocity v y,i and radar cross-sectional area σ i ,R 6 represents 6 dimensions of the radar point cloud data.

[0016] Further, use a multi-layer perceptron and a max pooling layer to extract features from the radar point cloud data at each time step, specifically including:

[0017] For each data point p i , obtain the local feature of p i through the multi-layer perceptron;

[0018] Perform MLP transformation on all data points p i within the neighborhood N(i) of each data point p j and obtain the maximum eigenvalue;

[0019] Concatenate the local feature of p i and the maximum eigenvalue within the neighborhood N(i) to obtain the concatenated feature;

[0020] Input the obtained concatenated feature into the multi-layer perceptron and the max pooling layer in sequence, and output the radar point cloud feature.

[0021] Further, obtain the RIS power delay distribution data graph, and use the CNN convolutional layer and position encoding to extract the RIS power delay distribution feature, specifically including the following steps:

[0022] Obtain the RIS power delay distribution data graph P r ;

[0023] Input the power delay distribution data graph P r into the two-dimensional CNN convolutional layer and flatten the convolutional result;

[0024] Concatenate the flattened convolutional result with the power delay distribution data graph Pr The position encoding features are spliced and combined to obtain the RIS power delay distribution features:

[0025]

[0026] Among them, represents the RIS power delay distribution feature; Flatten represents the flattening operation; Conv2D represents the two-dimensional convolutional layer; E pos represents the position encoding operation.

[0027] Furthermore, the position encoding operation is implemented by the following formula:

[0028]

[0029]

[0030] Among them, m ∈ {0, 1, 2,...., M - 1} represents the spatial position index value, where M is the maximum spatial position index, and i ∈ {0, 1, …, ⌊D c / 2⌋ - 1} represents the feature dimension index, and D c represents the number of convolutional output channels.

[0031] Furthermore, a cross-modal fusion module is used to perform feature fusion on the extracted radar point cloud features and RIS power delay distribution features, specifically including:

[0032] Construct a target correlation feature matrix using the attention mechanism:

[0033]

[0034] Among them, represents the radar point cloud feature at time t, represents the RIS power delay distribution feature at time t, represents the query space matrix obtained after linear transformation of the radar feature; represents the key space matrix obtained after linear transformation of the RIS power delay distribution feature; d k represents the dimension compression factor of the attention mechanism, and D r represents the radar point cloud feature dimension, represents the RIS power delay distribution feature dimension;

[0035] Based on the target correlation feature matrix, the transformed feature of the RIS power delay distribution feature converted to the radar point cloud dimension is obtained, and it is fused with the radar point cloud feature to obtain the fused feature :

[0036]

[0037] Among them, W v represents the value space matrix obtained after the linear change of the RIS power delay distribution characteristics, and LayerNorm represents the layer normalization function.

[0038] Furthermore, use the Transformer encoder to encode the obtained fusion features to obtain encoded features containing spatio-temporal dependencies, specifically including the following expressions:

[0039] H enc = TransformerEncoder(H fuse )

[0040] Among them, TransformerEncoder represents the Transformer encoder, and H enc represents the obtained fusion encoded features.

[0041] Furthermore, calculate the dynamic weight of each target, and use the dynamic weight to perform weighted aggregation on the encoded features to obtain weighted aggregation features, specifically including the following steps:

[0042] Calculate the weight of each target:

[0043]

[0044] Among them, represents the data value of the i-th target point in the fusion encoded feature at time t, represents the data value of the i-th target point in the RIS power delay distribution data sequence at time t, tanh is the hyperbolic tangent function, v is a learnable parameter, and W w is a learnable matrix;

[0045] Perform weighted summation of the target weight and the corresponding data points in the fusion encoded feature sequence to obtain weighted aggregation features:

[0046]

[0047] Among them, F H represents the weighted aggregation feature, and H i represents the i-th corresponding data point in the fusion encoded feature sequence.

[0048] Furthermore, use the Transformer decoder to decode the weighted aggregation feature and obtain the beam prediction result, specifically including:

[0049] Use the Transformer decoder to decode the weighted aggregation feature;

[0050] The output result of the decoder is passed through a linear layer and an activation function to obtain the amplitude and phase prediction results of the RIS:

[0051]

[0052] Among them, W out is the output of the Transformer decoder;

[0053] The amplitude and phase can be output separately:

[0054]

[0055] Among them, A pred represents the amplitude, represents the phase output.

[0056] The present invention also discloses a beam prediction system based on radar environment perception and RIS power distribution pattern, including a computer program, and when the computer program runs, it can implement the beam prediction method described in any one of the present invention.

[0057] It can be seen from the above technical solutions that compared with the prior art, the present invention discloses a beam prediction method based on radar environment perception and RIS power distribution pattern, and has the following beneficial effects:

[0058] 1. Improve the beam prediction accuracy

[0059] By combining radar data with the PDP optimization of the RIS, the present invention can accurately predict the beam direction in a dynamic environment, significantly improving the accuracy of beam prediction. Compared with the prior art (such as traditional CSI-based methods), the present invention improves the prediction accuracy, especially in complex mobile scenarios and high-dynamic environments.

[0060] 2. Effectively suppress multipath interference and beam splitting

[0061] The present invention utilizes the PDP optimization of the RIS. By dynamically adjusting the phase, amplitude, and polarization of the reflection units, the energy of the main path signal is enhanced, and the interference caused by multipath propagation is reduced, especially the beam splitting problem. Experiments show that the sidelobe interference is reduced by 15 dB, the signal attenuation is reduced, and the stability and communication quality of the system are effectively improved.

[0062] 3. Improve the calculation efficiency and reduce the overhead

[0063] Through deep learning model compression (such as knowledge distillation technology) and optimization, the amount of training data and the inference time are reduced, and the computational complexity is reduced by 40%. Compared with the prior art, especially traditional CSI-based deep learning models, the present invention can significantly improve the calculation efficiency without sacrificing accuracy and is applicable to high-speed mobile scenarios.

[0064] 4. Enhance dynamic environment adaptability

[0065] The system of the present invention can perceive environmental changes in real time (such as target movement, obstacle position change), and adaptively adjust the beam prediction strategy through reinforcement learning (RL) to ensure continuous high-quality signal transmission in different dynamic environments. Compared with traditional static models, the present invention has stronger robustness and adaptability, and is particularly suitable for application scenarios such as fast-changing 6G vehicle-to-everything and industrial Internet of Things.

[0066] 5. Reduce costs and energy consumption

[0067] The present invention reduces the hardware requirements and consumption of computing resources by optimizing the system architecture and deep learning model, thereby reducing the energy consumption and cost of the overall system. This advantage makes the technology have strong economy and scalability during large-scale deployment, especially having obvious advantages in high-density network environments.

[0068] 6. Optimize beam control strategy

[0069] Combined with the reinforcement learning algorithm, the system of the present invention can optimize the beam control strategy according to real-time feedback, and has significant advantages in ensuring system stability and efficiency. Compared with traditional fixed or manually adjusted beam control methods, the present invention can adapt to environmental changes and improves the long-term stable operation ability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0071] Figure 1 It is a schematic diagram of the overall process of the beam prediction method based on radar environment perception and RIS power distribution pattern provided by the embodiment of the present invention.

[0072] Figure 2 It is a schematic diagram of the specific steps for obtaining radar point cloud features by using a multi-layer perceptron and a max pooling layer provided by the embodiment of the present invention.

[0073] Figure 3 It is a schematic diagram of feature fusion by using a cross-modal fusion module provided by the embodiment of the present invention.

[0074] Figure 4 It is a comparison diagram of the robustness effects of the method of the present invention and the existing method in different noise environments provided by the embodiment of the present invention.

[0075] Figure 5 This is a comparison chart of the inference prediction speed effect between the method of the present invention and the existing method provided by the embodiment of the present invention.

[0076] Figure 6 This is a comparison chart of the prediction accuracy effect between the method of the present invention and the existing method provided by the embodiment of the present invention. Detailed implementation manners

[0077] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0078] Refer to Figures 1 - 3 , the embodiment of the present invention discloses a beam prediction method based on radar environmental perception and RIS power distribution mode, including the following steps:

[0079] Obtain the radar point cloud data at each time step in the radar time series data, and use a multi-layer perceptron and a max pooling layer to extract features from the radar point cloud data at each time step to obtain radar point cloud features;

[0080] Obtain the RIS power delay distribution data map, and use a CNN convolutional layer and position encoding to extract RIS power delay distribution features;

[0081] Use a cross-modal fusion module to fuse the extracted radar point cloud features and RIS power delay distribution features to obtain fused features;

[0082] Use a Transformer encoder to encode the obtained fused features to obtain encoded features containing spatio-temporal dependencies;

[0083] Calculate the dynamic weight of each target, and use the dynamic weight to perform weighted aggregation on the encoded features to obtain weighted aggregation features;

[0084] Use a Transformer decoder to decode the weighted aggregation features and obtain the beam prediction result.

[0085] Further, obtaining the radar point cloud data at each time step in the radar time series data includes the following radar point cloud data sets:

[0086] P t ={p i |p i =(x i ,yi , z i , v y,i , σ i ), ∈ R 6}, where the radar point cloud data set P t each data point p i includes the position (x i , y i , z i ), velocity v y,i and radar cross section area σ i , R 6 represents 6 dimensions of the radar point cloud data.

[0087] Further, a multi-layer perceptron and a max pooling layer are used to extract features from the radar point cloud data at each time step, specifically including:

[0088] For each data point p i , the local feature of p i is obtained through a multi-layer perceptron;

[0089] All data points p i within the neighborhood N(i) of each data point p j are subjected to MLP transformation, and the maximum eigenvalue is obtained;

[0090] The local feature of p i and the maximum eigenvalue within the neighborhood N(i) are concatenated to obtain a concatenated feature;

[0091] The obtained concatenated feature is sequentially input into a multi-layer perceptron and a max pooling layer, and the radar point cloud feature is output, where N is the number of targets, D r is the feature dimension, and is represented by the formula: .

[0092] In this step, the neighborhood N(i) is determined by using the fixed variable radius method, and all data points within a certain radius r of the current data point are used as the neighborhood range of this data point.

[0093] Further, an RIS power delay distribution data map is obtained, and an RIS power delay distribution feature is extracted by using a CNN convolutional layer and position encoding, specifically including the following steps:

[0094] Obtain the RIS power delay distribution data map P r ;

[0095] Input the power delay distribution data map P r into a two-dimensional CNN convolutional layer and flatten the convolutional result;

[0096] Concatenate the flattened convolution result with the power delay profile data graph P r 's position encoding features to obtain the RIS power delay profile features:

[0097] ;

[0098] Among them, represents the RIS power delay profile features; Flatten represents the flattening operation; Conv2D represents the two-dimensional convolutional layer; E pos represents the position encoding operation.

[0099] Specifically, in the embodiments of the present invention, the following method is adopted for position encoding:

[0100] ;

[0101] ;

[0102] Among them, m ∈ {0, 1, 2,...., M - 1} represents the spatial position index value, where M is the maximum spatial position index, and i ∈ {0, 1,…, ⌊D c / 2⌋ - 1} represents the feature dimension index, and D c represents the number of convolutional output channels.

[0103] Furthermore, use the cross-modal fusion module to perform feature fusion on the extracted radar point cloud features and RIS power delay profile features, specifically including:

[0104] Use the attention mechanism to construct the target correlation feature matrix:

[0105] ;

[0106] Among them, represents the radar point cloud features at time t, represents the RIS power delay profile features at time t, represents the query space matrix obtained after the linear transformation of the radar features; represents the key space matrix obtained after the linear transformation of the RIS power delay profile features; d k represents the dimension compression factor of the attention mechanism, and D r represents the dimension of the radar point cloud features, represents the dimension of the RIS power delay profile features; in this step, the purpose of the correlation feature matrix is to establish the data correlation between the radar point cloud features and the RIS power delay profile, so that data of two different modalities can be fused.

[0107] Then, based on the target correlation feature matrix obtained in the previous step, the conversion features of the RIS power delay distribution features transformed into the radar point cloud dimension are obtained and fused with the radar point cloud features to obtain the fused features. :

[0108]

[0109] Among them, W v represents the value space matrix obtained after the linear transformation of the RIS power delay distribution features, and LayerNorm represents the layer normalization function.

[0110] Furthermore, the obtained fused features are encoded using a Transformer encoder to obtain the encoded features containing spatio-temporal dependencies, which specifically include the following expressions:

[0111] H enc = TransformerEncoder(H fuse )

[0112] Among them, TransformerEncoder represents the Transformer encoder, and H enc represents the obtained fused encoded features.

[0113] Furthermore, calculate the dynamic weight of each target, and use the dynamic weight to perform weighted aggregation on the encoded features to obtain the weighted aggregation features, which specifically include the following steps:

[0114] Calculate the weight of each target:

[0115] ;

[0116] Among them, represents the data value of the i-th target point in the fused encoded features at time t, represents the data value of the i-th target point in the RIS power delay distribution data sequence at time t, tanh is the hyperbolic tangent function, v is a learnable parameter, and W w is a learnable matrix;

[0117] Weight the target weight with the corresponding data points in the fused encoded feature sequence and perform weighted summation to obtain the weighted aggregation features:

[0118] ;

[0119] Among them, F H represents the weighted aggregation features, and H i represents the i-th corresponding data point in the fused encoded feature sequence.

[0120] Further, a Transformer decoder is used to decode the weighted aggregated features and obtain the beam prediction result, which specifically includes:

[0121] Using the Transformer decoder to decode the weighted aggregated features;

[0122] The output result of the decoder is passed through a linear layer and an activation function to obtain the amplitude and phase prediction results of the RIS: ; where W out is the output of the Transformer decoder;

[0123] The amplitude and phase can be output separately:

[0124]

[0125] where A pred represents the amplitude, represents the phase output.

[0126] The overall loss function of the model in the present invention is as follows:

[0127] ;

[0128] where λ1, λ2, and λ3 represent weight coefficients. The first term of the loss function represents the beam reconstruction loss, the second term represents the power distribution similarity loss, and the third term represents the final loss of the key target. represents the beam prediction value, Y represents the beam true value, pr and respectively represent the true and predicted power distribution similarities, KL represents the KL divergence, p i and respectively represent the true and predicted target positions.

[0129] Further, physical constraint optimization is added after the prediction output layer of the present invention to further optimize the obtained prediction result, which specifically includes:

[0130] Amplitude constraint: A opt = min(max(A pred , 0), A max );

[0131] Phase constraint: .

[0132] During the physical constraint process, iteration is performed by the projected gradient descent method to satisfy the final constraint conditions:

[0133] ;

[0134] .

[0135] To verify the effectiveness of the present invention, the method of the present invention is compared with existing similar methods, with reference to Figure 4 , Figure 5 , Figure 6 . The beam prediction method combining radar environmental perception and RIS power distribution pattern disclosed by the present invention is superior to existing similar methods in terms of robustness, inference speed, and inference accuracy.

[0136] The present invention also discloses a beam prediction system based on radar environmental perception and RIS power distribution pattern, including a computer program, which can implement the beam prediction method described in any one of the present invention when running.

[0137] When the present invention runs the above script containing the above method on a device terminal (such as a computer terminal), it specifically may include the following steps.

[0138] 1. Initialize the environment

[0139] Global variable declaration: At the beginning of the script, the global variable SELECTED_DATA_PATH is declared to store the path of the selected data set. The code uses global SELECTED_DATA_PATH for global variable declaration.

[0140] Clean the workspace: The clear; clc; command clears all variables in the workspace and empties the command window to ensure that the script runs in a clean environment.

[0141] Add subdirectories to the path: Use addpath(genpath(pwd)) to add the current directory and all its subdirectories to the MATLAB search path to ensure that subsequent functions can be called smoothly.

[0142] 2. Create a directory structure

[0143] Check and create directories: The code checks and creates the following directories through mkdir:

[0144] oconfigs

[0145] oresults

[0146] oresults / models

[0147] oresults / figures If these directories do not exist, the script will create them for storing configuration files, experimental results, trained models, and graphic files.

[0148] 3. Scan available data sets

[0149] Scanning the dataset directory: The code scans the current working directory and its subdirectories for available datasets through scan_data_files(pwd). data_paths = scan_data_files(pwd); will return the paths and the number of files of all available datasets.

[0150] Selecting a dataset: If multiple datasets are scanned, the script will ask the user to enter the number of the selected dataset and provide an option to select all data. If the user does not enter a selection, the first dataset will be selected by default. The code uses input to get the user input and selects the corresponding dataset path.

[0151] 4. Creating a label file

[0152] Label file check: The script checks whether the dataset selected by the user contains a label file (test_labels.mat). If the file does not exist, the system will prompt the user whether to create a label file.

[0153] Generating a label file: The user can choose three label generation methods: exhaustive search, hierarchical search, or the optimal method. The code will call create_labels_file(selected_path, methods{method_choice}); to create the label file.

[0154] 5. Conducting a method comparison experiment

[0155] Experiment comparison selection: At this step, the system will ask the user whether to conduct a method comparison experiment. If the user selects "Yes", the script will call the compare_methods function to compare different beam prediction methods and save the results to the results / figures directory.

[0156] 6. Selecting the running mode

[0157] Mode selection: The user selects the running mode at this step, providing the following four options:

[0158] ①Full experimental process

[0159] ②Beam prediction performance comparison

[0160] ③Training a deep learning predictor

[0161] ④Visualizing the dataset

[0162] The user enters the mode number through input, and the script executes the corresponding task according to the selected mode.

[0163] 7. Execute corresponding tasks

[0164] Full experimental process: If the user selects Mode 1, the script will call the run_all_experiments() function to execute the full experimental process including data loading, preprocessing, training, evaluation, and result saving.

[0165] Beam prediction performance comparison: If the user selects Mode 2, the script will execute compare_methods('results / figures') to compare different beam prediction methods and save the results in the results / figures directory.

[0166] Train the deep learning predictor: If the user selects Mode 3, the script will call train_dl_predictor() to train the deep learning prediction model and optimize beam prediction.

[0167] Visualize the dataset: If the user selects Mode 4, the script will call visualize_dataset(data_config.radar_train_dir) to visualize the radar data in the dataset and generate power maps, obstacle maps, and PDP maps.

[0168] 8. End execution

[0169] Execution completed: After all tasks are executed, the script displays a prompt "Execution completed!" indicating that the script has ended successfully.

[0170] The various embodiments in this specification are described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and reference can be made to the description of the method part for the relevant parts.

[0171] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A beam prediction method based on radar environment perception and RIS power distribution pattern, characterized in that, Including the following steps: Obtain the radar point cloud data at each time step in the radar time series data, and use a multi-layer perceptron and a max pooling layer to extract features from the radar point cloud data at each time step to obtain radar point cloud features; Obtain the RIS power delay distribution data map, and use a CNN convolutional layer and position encoding to extract RIS power delay distribution features; Use a cross-modal fusion module to fuse the extracted radar point cloud features and RIS power delay distribution features to obtain fused features; Use a Transformer encoder to encode the obtained fused features to obtain encoded features containing spatio-temporal dependencies; Calculate the dynamic weight of each target, and use the dynamic weight to perform weighted aggregation on the encoded features to obtain weighted aggregation features; Use a Transformer decoder to decode the weighted aggregation features and obtain the beam prediction result.

2. The beam prediction method based on radar environmental perception and RIS power distribution pattern according to claim 1, wherein Obtain the radar point cloud data at each time step in the radar time series data, including the following radar point cloud data sets: P t ={p i |p i =(x i ,y i ,z i ,v y,i ,σ i ) ∈R 6}, where the radar point cloud data set P t each data point p i in it includes position (x i , y i , z i ), velocity v y,i and radar cross section area σ i , R 6 represents six dimensions of the radar point cloud data.

3. The beam prediction method based on radar environment perception and RIS power distribution pattern according to claim 2, wherein Use a multi-layer perceptron and a max pooling layer to extract features from the radar point cloud data at each time step, specifically including: For each data point p i , the local feature of p i is obtained through a multi-layer perceptron; For each data point p i for all data points p j within the neighborhood N(i) of, perform the MLP transformation and obtain the maximum eigenvalue; Concatenate the local features of p i and the maximum eigenvalue within the neighborhood N(i) to obtain the concatenated features; Input the obtained concatenated features into a multi-layer perceptron and a max pooling layer in sequence, and output radar point cloud features.

4. The beam prediction method based on radar environmental perception and RIS power distribution pattern according to claim 1, characterized in that Obtain the RIS power delay distribution data map, and use a CNN convolutional layer and position encoding to extract RIS power delay distribution features, specifically including the following steps: Obtain the RIS power delay profile data graph P r ; Input the power delay profile data graph P r into a two-dimensional CNN convolutional layer and flatten the convolutional result; Concatenate the flattened convolutional result with the position encoding features of the power delay profile data graph P r to obtain the RIS power delay profile features: ； Among them, represents the RIS power delay distribution characteristic; Flatten represents the flattening operation; Conv2D represents the two-dimensional convolutional layer; E pos represents the position encoding operation.

5. The beam prediction method based on radar environment perception and RIS power distribution pattern according to claim 4, wherein, The position encoding operation is implemented by the following formula: ； ； Among them, \(m\in\{0,1,2,\ldots,M - 1\}\) represents the spatial position index value, where \(M\) is the maximum spatial position index, and \(i\in\{0,1,\ldots,\lfloor D c / 2\rfloor - 1\}\) represents the feature dimension index, and \(D c represents the number of convolutional output channels.

6. The beam prediction method based on radar environment perception and RIS power distribution pattern according to claim 1, wherein Use a cross-modal fusion module to fuse the extracted radar point cloud features and RIS power delay distribution features, specifically including: Use an attention mechanism to construct a target association feature matrix: ； Among them, represents the radar point cloud feature at time t, represents the RIS power delay distribution feature at time t, represents the query space matrix obtained after linear transformation of the radar feature; represents the key space matrix obtained after linear transformation of the RIS power delay distribution feature; d k represents the dimension compression factor of the attention mechanism, D r represents the dimension of the radar point cloud feature, represents the dimension of the RIS power delay distribution feature; The conversion features obtained by converting the RIS power delay distribution features based on the target correlation feature matrix to the radar point cloud dimension are fused with the radar point cloud features to obtain the fused features : ； Among them, W v represents the value space matrix obtained after the linear change of the RIS power delay distribution characteristics, and LayerNorm represents the layer normalization function.

7. The beam prediction method based on radar environment perception and RIS power distribution pattern according to claim 6, using a Transformer encoder to encode the obtained fused features to obtain encoded features containing spatio-temporal dependencies, specifically including the following expressions: H enc = TransformerEncoder(H fuse ) Among them, TransformerEncoder represents the Transformer encoder, and H enc represents the obtained fused encoded features.

8. The beam prediction method based on radar environment perception and RIS power distribution pattern according to claim 7, wherein Calculate the dynamic weight of each target, and use the dynamic weight to perform weighted aggregation on the encoded features to obtain weighted aggregation features, specifically including the following steps: Calculate the weight of each target: ； Among them, represents the data value of the i-th target point in the fused encoded feature at time t, represents the data value of the i-th target point in the RIS power delay distribution data sequence at time t, tanh is the hyperbolic tangent function, v is a learnable parameter, and W w is a learnable matrix; Sum the target weight with the corresponding data points in the fused encoded feature sequence to obtain a weighted aggregated feature: ； Among them, F H represents the weighted aggregation feature, and H i represents the i-th data point corresponding in the fused coding feature sequence.

9. The beam prediction method based on radar environment perception and RIS power distribution pattern according to claim 8, wherein, using a Transformer decoder to decode the weighted aggregation features and obtain the beam prediction result, specifically including: Use a Transformer decoder to decode the weighted aggregation features; Pass the output result of the decoder through a linear layer and an activation function to obtain the amplitude and phase prediction results of the RIS: ； Among them, W out is the output of the Transformer decoder; The amplitude and phase can be output separately: ； Among them, A pred represents the amplitude, represents the phase output.

10. A beam prediction system based on radar environmental perception and RIS power distribution pattern, characterized in that, Including a computer program, which can implement the beam prediction method according to any one of claims 1-9 when running.

Images