Hybrid RIS adaptive vortex beam generation system and method based on large model optimization

Abstract

The invention discloses a hybrid RIS self-adaptive vortex beam generation system based on large model optimization, and the system comprises a model pre-training module which carries out the pre-training of a DeepSeek-R1 model through selected data by employing a few-sample learning and prompt engineering technology, obtains an edge communication model, and carries out the dynamic optimization of the edge communication model through a cGAN network; the channel prediction module extracts channel characteristics according to the CSI attention module, and performs prediction in combination with an edge communication model to obtain mixed RIS auxiliary channel state information; and the adaptive vortex beam generation module calculates the total power consumption and the maximum reachable sum rate of the system according to the prediction result, calculates a base station beam forming matrix and a mixed RIS vortex beam phase shift matrix by adopting a beam forming network in combination with a safe energy efficiency maximization target, and generates an adaptive vortex beam. Through large model optimization and multi-task learning, adaptability and communication performance of the system in a complex channel environment are improved, and low-power-consumption and high-precision beam forming and dynamic optimization are realized.

Description

Technical Field

[0001] The technical field related to the present invention is the field of wireless communication technology, specifically referring to a hybrid RIS adaptive vortex beam generation system and method optimized based on a large model. Background Art

[0002] In the field of wireless communication technology, with the continuous development of 6G communication technology and wireless sensor networks, the application of RIS (Reconfigurable Intelligent Surface) has gradually become an important means to improve the efficiency of communication systems.

[0003] In the prior art, the RIS architecture is mainly divided into two types: passive and active. Passive RIS is limited by fixed reflection characteristics and cannot dynamically adapt to the rapidly changing wireless environment, resulting in the "multiplicative fading" effect and limited long-distance transmission performance. For example, in Chinese Patent Application No. CN118413259A, although a STAR-RIS assisted NOMA communication method is proposed, this scheme focuses on transmission rate optimization and ignores the joint optimization of system energy efficiency and multipath channels, making it difficult to meet the dual requirements of low energy consumption and high coverage in 6G networks; while active RIS can actively adjust signals, but has low energy efficiency and a sharp increase in power consumption when applied on a large scale. In addition, traditional reflective metasurfaces can only generate plane beams and cannot generate complex waveforms such as vortex beams, which limits the improvement of spatial degrees of freedom and spectral efficiency. In terms of system optimization, existing solutions mostly adopt a single-task optimization paradigm and lack the systematicness of multi-task joint optimization. For example, in Chinese Patent Application No. CN117527608A, this scheme adopts a single-task optimization paradigm, fails to integrate tasks such as channel prediction, beamforming, and energy efficiency optimization, and lacks a multi-user joint antenna selection strategy. Traditional solutions rely on static models and independent optimization modules, resulting in low flexibility and resource utilization rate of the system in a dynamic environment and being unable to achieve global optimality through multi-task learning. Regarding the problem of computational complexity, in high-dimensional dynamic channel prediction, the existing computational resource consumption is relatively high, making it difficult to meet the high-efficiency requirements of communication systems. For example, in Chinese Patent Application No. CN119070937A, although this scheme reduces part of the computational complexity through an extrapolation method, it still relies on the initial calculation of ray tracing and does not solve the pilot overhead problem in large-scale antenna systems.

[0004] Therefore, inventing a hybrid RIS adaptive vortex beam generation system optimized based on a large model can solve problems such as the passivity and energy efficiency contradiction of the traditional RIS architecture, insufficient waveform generation and spatial multiplexing capabilities, lack of systematicness in multi-task joint optimization, and excessive pilot overhead and calculation complexity, achieve low-power basic beamforming, high-precision dynamic optimization, and more efficient spatial multiplexing, reduce part of the computational complexity, and improve the adaptability and accuracy of the large model in practical applications. Summary of the Invention

[0005] The objective of the present invention is to provide a hybrid RIS adaptive vortex beam generation system and method optimized based on a large model. The present invention can solve problems such as the passivity and energy efficiency contradiction of traditional RIS architectures, insufficient waveform generation and spatial multiplexing capabilities, lack of systematic multi-task joint optimization, and excessive pilot overhead and calculation complexity. It improves the adaptability and communication performance of the system in complex channel environments and realizes low-power, high-precision beamforming and dynamic optimization.

[0006] To achieve this objective, a hybrid RIS adaptive vortex beam generation system optimized based on a large model designed by the present invention includes:

[0007] The model pre-training module is used to select communication channel data samples based on hybrid RIS from historical downlink CSI (Channel State Information) data through data selection technology, select vortex beam communication data samples from the millimeter-wave communication channel measurement dataset, and use the communication channel data samples based on hybrid RIS and the vortex beam communication data samples to pre-train the DeepSeek-R1 model using few-shot learning and prompting engineering techniques to obtain the DeepSeek-R1 edge communication model. Generate a channel matrix and beam configuration through a cGAN (Conditional Generative Adversarial Network) network, and dynamically optimize the DeepSeek-R1 edge communication model using the channel matrix and beam configuration;

[0008] The channel prediction module is used to extract and preprocess channel features from communication channel data samples based on hybrid RIS according to the set CSI attention module to obtain processed channel feature data, and predict the processed channel feature data based on the dynamically optimized DeepSeek-R1 edge communication model to obtain hybrid RIS-assisted channel state information;

[0009] The adaptive vortex beam generation module is used to calculate the total power consumption and maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system according to the hybrid RIS-assisted channel state information; according to the hybrid RIS-assisted channel state information, the set objective of maximizing the security energy efficiency of the hybrid RIS-assisted wireless transmission system, the total power consumption and maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system, and combine with the beamforming network to calculate the base station beamforming matrix and hybrid RIS vortex beam phase shift matrix that maximize the energy efficiency, thereby generating an adaptive vortex beam.

[0010] Preferably, the method for obtaining the DeepSeek-R1 edge communication model based on dynamic optimization is as follows:

[0011] Optimize the prediction accuracy of the DeepSeek-R1 edge communication model twice. At the same time, perform multi-task joint optimization on the hybrid RIS adaptive vortex beam generation system based on the joint loss function. In the multi-task joint optimization hybrid RIS adaptive vortex beam generation system, generate the channel matrix and beam configuration according to the generator of the cGAN network in the DeepSeek-R1 edge communication model after two optimizations and the set parameters. Compare the difference between the channel matrix generated by the generator and the channel state information collected in real time through the discriminator in the cGAN network of the DeepSeek-R1 edge communication model after the second optimization. At the same time, compare the difference between the generated beam configuration and the beam configuration collected in real time. When the difference between the generated channel matrix and the channel state information collected in real time is less than the set channel state threshold, and the difference between the generated beam configuration and the beam configuration collected in real time is less than the set beam configuration threshold, obtain the DeepSeek-R1 edge communication model after dynamic optimization. Otherwise, after updating the parameters of the cGAN network generator and discriminator, regenerate the channel matrix and beam configuration according to the generator of the cGAN network and the set parameters.

[0012] Preferably, the specific method for optimizing the prediction accuracy of the DeepSeek-R1 edge communication model twice is as follows:

[0013] Obtain the channel state information loss function by calculating the error between the hybrid RIS-assisted channel state information and the channel state information of the hybrid RIS collected in real time, and optimize the prediction accuracy of the DeepSeek-R1 edge communication model according to minimizing the channel state information loss function to obtain the DeepSeek-R1 edge communication model after the first optimization;

[0014] Calculate the loss function of the beamforming network model based on the set five-dimensional constraints of hybrid RIS-assisted communication, the total power consumption of the hybrid RIS-assisted wireless transmission system, and the maximum achievable sum rate;

[0015] During the generation process of the adaptive vortex beam, feedback the beamforming matrix and the hybrid RIS phase shift matrix that maximize the energy efficiency to the cGAN network in the DeepSeek-R1 edge communication model after the first optimization, and perform the second optimization on the prediction accuracy of the DeepSeek-R1 edge communication model after the first optimization through the discriminator in the cGAN network;

[0016] The specific process of multi-task joint optimization:

[0017] The process of obtaining the hybrid RIS-assisted channel state information in the channel prediction module and the process of obtaining the base station beamforming matrix and the hybrid RIS vortex beam phase shift matrix that maximize the energy efficiency in the adaptive vortex beam generation module are trained in an alternating training manner until the channel state information loss function and the loss function of the beamforming network model reach a set equilibrium state; on the premise that the channel state information loss function and the loss function of the beamforming network model reach the set equilibrium state, the joint loss function is calculated by combining the channel state information loss function and the loss function of the beamforming network model, and multi-task joint optimization is achieved by minimizing the joint loss function.

[0018] Advantages of the present invention:

[0019] The present invention proposes a hybrid RIS adaptive vortex beam generation system based on large model optimization, which adopts a step-by-step optimization idea, decomposes the beamforming and phase shift matrix optimization processes into multiple steps, and each step is independently optimized. The step-by-step processing method avoids the high computational complexity brought by simultaneously processing all complex problems, makes the optimization problem of each step simpler, can be quickly solved, and meets the real-time requirements of the communication system; through the multi-task learning framework, multiple tasks such as channel prediction, beamforming, and energy efficiency management are comprehensively considered. In different steps, the channel prediction model, beamforming network, and energy efficiency management strategy are respectively optimized to ensure that the generated control strategy can not only ensure the efficient and stable operation of the communication system, but also take into account the energy efficiency and other requirements of the system, and realize the comprehensive optimized operation of the communication system; by using intelligent algorithms to solve the optimization function, a better solution can be found in a short time. Through the combination of step-by-step optimization and intelligent algorithms, the generated control strategy is more accurate and reliable, can effectively cope with complex scenarios and dynamic changes in the communication system, and reduce the risk of communication interruption and performance degradation. The present invention solves the problems of single-objective optimization, excessively high computational complexity, and lack of systematic integration in the prior art by introducing large model optimization and multi-task learning, can efficiently and accurately generate hybrid RIS adaptive vortex beam generation strategies, meet the requirements of the communication system in complex environments, and contribute to improving the performance and reliability of the communication system. Description of the Drawings

[0020] Figure 1 It is a schematic structural diagram of the present invention;

[0021] Figure 2 It is a schematic diagram of the training process of the DeepSeek-R1 edge communication model of the present invention;

[0022] Figure 3 It is a communication system framework;

[0023] Figure 4 It is a flowchart of the cGAN network;

[0024] Figure 5 It is a schematic diagram of the fusion training scheme. Specific implementation manners

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Therefore, the detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0026] The following further elaborates on the present invention with reference to the accompanying drawings and specific embodiments:

[0027] Embodiment 1

[0028] A hybrid RIS adaptive vortex beam generation system optimized based on a large model, as Figure 1 shown, includes:

[0029] The model pre-training module is used to select communication channel data samples based on hybrid RIS from historical downlink CSI data through data selection technology, select vortex beam communication data samples from the millimeter-wave communication channel measurement dataset, and use the communication channel data samples based on hybrid RIS and the vortex beam communication data samples to pre-train the DeepSeek-R1 model using few-shot learning and prompt engineering techniques to obtain the DeepSeek-R1 edge communication model. Generate a channel matrix and beam configuration through the cGAN network, and dynamically optimize the DeepSeek-R1 edge communication model using the channel matrix and beam configuration;

[0030] The channel prediction module is used to extract and preprocess the channel features of the communication channel data samples based on hybrid RIS according to the set CSI attention module to obtain the processed channel feature data, and predict the processed channel feature data based on the dynamically optimized DeepSeek-R1 edge communication model to obtain the hybrid RIS-assisted channel state information;

[0031] The adaptive vortex beam generation module is used to calculate the total power consumption and the maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system according to the hybrid RIS-assisted channel state information; according to the hybrid RIS-assisted channel state information, the set goal of maximizing the secure energy efficiency of the hybrid RIS-assisted wireless transmission system, the total power consumption and the maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system, combined with the beamforming network, calculate the base station beamforming matrix and the hybrid RIS vortex beam phase shift matrix that maximize the energy efficiency, and generate adaptive vortex beams through the base station beamforming matrix and the hybrid RIS vortex beam phase shift matrix that maximize the energy efficiency.

[0032] In the above technical solution, the historical uplink and downlink CSI dataset can provide a channel dataset for the training of the DeepSeek-R1 model of the channel prediction module, helping the model predict different channel information, and at the same time calculating and generating the optimal adaptive vortex beam under different channel states. The millimeter-wave communication channel measurement dataset comes from the millimeter-wave band contained in the vortex beam, mainly the 5G band, and also includes the 6G band (terahertz band). However, since 6G is still in development and the dataset of the terahertz band is small, the communication channel dataset of the millimeter-wave band is selected.

[0033] In the above technical solution, the parameters of various communication performances selected from the historical uplink and downlink CSI data based on the hybrid RIS communication channel data samples and the vortex beam communication data samples from the millimeter-wave communication channel measurement dataset are: data with a signal-to-noise ratio between 10 dB and 30 dB to ensure that the communication system can perform effective channel prediction under different signal qualities; data with a channel coherence time greater than 50 ms to cover channel environments with relatively high and low time-varying characteristics; data with a multipath delay spread greater than 10 nanoseconds to capture the multipath propagation effect in complex environments; data with a channel capacity greater than 5 bps / Hz to cover high-bandwidth and high-efficiency wireless communication scenarios.

[0034] In the above technical solution, the few-shot learning technology enables the model to quickly adapt to new scenarios with a very small amount of training data, enhancing the generalization ability; the prompt engineering technology optimizes the input prompt, reduces redundant information, and improves the model's recognition and processing efficiency of key information.

[0035] In the above technical solution, the performance of the hybrid RIS adaptive vortex beam generation system optimized based on a large model is improved through multi-task joint optimization. First, the error between the hybrid RIS-assisted channel state information and the real-time collected channel state information is calculated to obtain the channel state information loss function, and based on this, the DeepSeek-R1 edge communication model is preliminarily optimized. Then, according to the five-dimensional constraint conditions of hybrid RIS-assisted communication, the total system power consumption, and the maximum achievable sum rate, the loss function of the beamforming network model is calculated. Next, the base station beamforming matrix and the hybrid RIS phase shift matrix that maximize the energy efficiency are fed back to the cGAN network in the preliminarily optimized model, and the discriminator is used for the second optimization. At the same time, in an alternating training manner, the channel prediction module and the adaptive vortex beam generation module are jointly trained until the channel state information loss function and the loss function of the beamforming network model reach a set balance state. At this balance state, the joint loss function is calculated by combining the two loss functions, and multi-task joint optimization is achieved by minimizing this joint loss function to improve the adaptability and communication performance of the system in a complex channel environment, achieving low-power, high-precision beamforming and dynamic optimization.

[0036] In the above technical solution, the maximum achievable sum rate generally refers to the highest data transmission rate that can be theoretically achieved in a specific communication system or network.

[0037] In the above technical solution, the method for obtaining the DeepSeek-R1 edge communication model after dynamic optimization is as follows:

[0038] The prediction accuracy of the DeepSeek-R1 edge communication model is optimized twice, and at the same time, multi-task joint optimization is performed on the hybrid RIS adaptive vortex beam generation system based on the joint loss function. In the hybrid RIS adaptive vortex beam generation system with multi-task joint optimization, according to the generator and the set parameters in the cGAN network of the DeepSeek-R1 edge communication model after two optimizations, the channel matrix and beam configuration are generated. The discriminator in the cGAN network of the DeepSeek-R1 edge communication model after the second optimization compares the differences between the channel matrix generated by the generator and the real-time collected channel state information, and at the same time compares the differences between the generated beam configuration and the real-time collected beam configuration. When the difference between the generated channel matrix and the real-time collected channel state information is less than the set channel state threshold, and the difference between the generated beam configuration and the real-time collected beam configuration is less than the set beam configuration threshold, the DeepSeek-R1 edge communication model after dynamic optimization is obtained. Otherwise, after updating the parameters of the cGAN network generator and discriminator, the channel matrix and beam configuration are regenerated according to the generator of the cGAN network and the set parameters.

[0039] In the above technical solution, the set parameters are a dynamic beam optimization environment parameter set, specifically referring to: real-time channel state information, user location, beamforming requirements, and interference information.

[0040] In the above technical solution, the generator and discriminator of the cGAN network are described in detail to generate and evaluate the process of the channel matrix and beam configuration, ensuring that the difference between the generated channel matrix and beam configuration and the real-time collected channel state information and beam configuration is within an acceptable range, to determine whether the generated channel matrix and beam configuration meet the requirements, and if not, regenerate them, thereby ensuring the accuracy and adaptability of the model.

[0041] In the above technical solution, the specific ways to optimize the prediction accuracy of the DeepSeek-R1 edge communication model twice are as follows:

[0042] The channel state information loss function is obtained by calculating the error between the hybrid RIS-assisted channel state information and the real-time collected hybrid RIS-based channel state information, and the prediction accuracy of the DeepSeek-R1 edge communication model is optimized according to minimizing the channel state information loss function to obtain the DeepSeek-R1 edge communication model after the first optimization;

[0043] The loss function of the beamforming network model is calculated based on the set five-dimensional constraints of hybrid RIS-assisted communication, the total power consumption of the hybrid RIS-assisted wireless transmission system, and the maximum achievable sum rate;

[0044] During the adaptive vortex beam generation process, the beamforming matrix that maximizes the energy efficiency and the hybrid RIS phase shift matrix are fed back to the cGAN network in the DeepSeek-R1 edge communication model after the first optimization, and the prediction accuracy of the DeepSeek-R1 edge communication model after the first optimization is optimized for the second time through the discriminator in the cGAN network;

[0045] The specific process of multi-task joint optimization:

[0046] The process of obtaining the hybrid RIS-assisted channel state information in the channel prediction module and the process of obtaining the base station beamforming matrix and the hybrid RIS vortex beam phase shift matrix that maximize the energy efficiency in the adaptive vortex beam generation module are trained in an alternating training manner until the channel state information loss function and the loss function of the beamforming network model reach a set balanced state; on the premise that the channel state information loss function and the loss function of the beamforming network model reach a set balanced state, the joint loss function is calculated by combining the channel state information loss function and the loss function of the beamforming network model, and multi-task joint optimization is achieved by minimizing the joint loss function.

[0047] In the above technical solution, the multi-tasks in the joint optimization of multi-tasks refer to simultaneously processing and optimizing multiple interrelated tasks during the optimization of a communication system. The tasks specifically include:

[0048] A channel prediction task, which predicts the future channel state based on historical channel state information (CSI) to cope with the dynamically changing wireless environment; a beamforming task, which calculates the optimal base station beamforming matrix according to the predicted channel state information to improve the quality and efficiency of signal transmission; an energy efficiency management task, which optimizes the energy efficiency of the system and reduces power consumption on the premise of ensuring communication quality; a waveform generation task, which is used to generate complex waveforms such as vortex beams to increase the spatial degrees of freedom and spectral efficiency and achieve more efficient spatial multiplexing.

[0049] In the above technical solution, each task in the joint optimization of multi-tasks is often processed independently in the traditional communication system optimization. However, in the present invention, through a multi-task learning framework, they are integrated into a unified optimization process. Through a joint loss function, the optimization objectives of these tasks are comprehensively considered, enabling the system to simultaneously optimize multiple aspects such as channel prediction, beamforming, and energy efficiency management in a dynamic environment, thereby improving the overall performance.

[0050] In the above technical solution, the achievement of relatively high performance means that when the loss function drops to a stable lower value, and after several consecutive iterations, the change rate of the loss value is less than 0.1% or 0.01%, it can be considered that the loss value has tended to be stable; when it reaches a certain specific lower value, such as the loss function < 0.01, it can be regarded that the loss value has reached a lower level, and the specific threshold can be set according to the actual situation.

[0051] In the above technical solution, the discriminator in the cGAN network is used to perform a second optimization on the prediction accuracy of the DeepSeek-R1 edge communication model after the first optimization. Specifically, the discriminator in the cGAN network is used to further optimize the generator in the cGAN network.

[0052] In the above technical solution, as Figure 5 shown, before the alternating training, phased training needs to be carried out. Phased training means independently training the channel prediction module and the adaptive vortex beam generation module respectively to ensure that each sub-network module can achieve relatively high performance in its respective task.

[0053] In the above technical solution, the process of obtaining the hybrid RIS-assisted channel state information in the channel prediction module and the process of obtaining the base station beamforming matrix and the hybrid RIS vortex beam phase shift matrix for maximizing the energy efficiency in the adaptive vortex beam generation module are trained through an alternating training method until the loss functions of the channel state information and the beamforming network model both converge to a certain accuracy, and then enter the joint training stage, as Figure 5 shown, until the equilibrium state. In the joint training stage, the errors of the loss functions of the channel state information and the beamforming network model no longer fluctuate significantly, and the system begins to optimize the collaborative performance of the two tasks through the joint loss function. Then, combining the loss functions of the channel state information and the beamforming network model, the joint loss function is calculated.

[0054] In the above technical solution, a certain accuracy means that in the alternating training stage, the errors of the channel prediction and the adaptive vortex beam generation module reach a predetermined threshold (1e-3), the decline rate of the errors tends to be gentle, but the equilibrium state has not been completely reached, and there is still room for further optimization.

[0055] In the above technical solution, the optimized equilibrium state means that the loss functions of the channel state information and the beamforming network model both converge, and neither of them changes significantly, indicating that the two tasks have reached the state of collaborative optimization.

[0056] In the above technical solution, the process of obtaining the hybrid RIS-assisted channel state information in the channel prediction module and the process of obtaining the base station beamforming matrix and the hybrid RIS vortex beam phase shift matrix for maximizing the energy efficiency in the adaptive vortex beam generation module are trained through an alternating training method, and a joint loss function is introduced for joint optimization to ensure that the maximum energy efficiency of the system is effectively guaranteed during the simultaneous optimization of channel prediction and adaptive vortex beam generation.

[0057] In the above technical solution, joint optimization and weighted loss functions are adopted to ensure that multiple objectives, including channel prediction, beamforming, and energy efficiency management, are simultaneously optimized during the fine-tuning process; the joint optimization objective function can dynamically adjust the weights according to specific application scenarios, effectively avoiding overfitting or local optimization situations, achieving multi-objective optimization, ensuring the balance among channel state prediction, beamforming, phase shift matrix optimization, and energy efficiency management, and ensuring that the system can obtain the best performance in different wireless environments.

[0058] In the above technical solution, the large model is fine-tuned based on real-time feedback data to ensure that each adjustment can maximize the performance of the model; real-time channel feedback and beam optimization feedback are key components of the fine-tuning. By collecting information such as user data, channel status, and system performance, the fine-tuning process can adjust parameters according to this real-time data, thereby improving the adaptability and response ability of the system. Especially in high-interference and multipath propagation environments, the fine-tuning strategy can quickly respond to these environmental changes. By using an adaptive learning rate, the fine-tuning process can ensure that the model maintains a balance between stability and learning speed and ensure accurate signal processing and optimization in complex environments.

[0059] In the above technical solution, the prediction accuracy of the DeepSeek-R1 edge communication model is improved through two optimizations to ensure its adaptability and communication performance in complex channel environments, significantly improving the prediction accuracy and adaptability of the model, ensuring the stable operation of the system under different channel conditions, and at the same time optimizing beamforming to improve communication quality and energy efficiency.

[0060] In the above technical solution, before the second optimization of the DeepSeek-R1 edge communication model, a cGAN network is introduced into the DeepSeek-R1 edge communication model to generate a channel matrix and beam configuration according to conditions such as real-time channel state information, user location, beamforming requirements, and interference information; the channel matrix and beam configuration generated by the generator will be evaluated by the discriminator to evaluate the difference between the channel matrix and beam configuration and the real-time collected channel and beam configuration. Through this evaluation and feedback, the cGAN network can continuously optimize the generation results and improve the accuracy and adaptability of the model.

[0061] In the above technical solution, the processed channel feature data is input into the optimized DeepSeek-R1 edge communication model for predicting channel state information. The backbone network of the DeepSeek-R1 edge communication model consists of stacked Transformer decoder layers, and each layer includes a multi-head self-attention mechanism, a feed-forward network, an addition layer, and layer normalization. The structure of the Transformer model is flexible, and the number of stacked layers and the dimensions of each layer can be adjusted according to requirements. During the training process of the DeepSeek-R1 edge communication model, the multi-head attention and feed-forward layers of the LLM (Large Language Model) are frozen to retain its general knowledge, and only the addition layer, layer normalization, and position embedding layer are fine-tuned to adapt to the channel prediction task.

[0062] In the above technical solution, the DeepSeek-R1 edge communication model is pre-trained and optimized based on the LLM, and the LLM is a general large language model.

[0063] In the above technical solution, the adaptability and communication performance of the system in a complex channel environment are improved through multi-task joint optimization. This can not only comprehensively optimize tasks such as channel prediction, beamforming, and energy efficiency management, but also enhance the system performance, ensure stable operation, strengthen adaptability, and effectively cope with complex scenarios and dynamic changes.

[0064] In the above technical solution, the specific method for extracting and preprocessing channel features from communication channel data samples according to the set CSI attention module is as follows:

[0065] A training environment is established through the downlink channel model, and channel feature data is obtained through channel estimation in the downlink channel model, reference signals in the downlink channel model, and the CSI feedback mechanism in the downlink channel model. Among them, the downlink channel model between the base station and the user terminal is expressed as:

[0066]

[0067] Among them, N is the number of clusters, M n is the number of paths in the nth cluster, β n,m is the complex path gain of the mth path in the nth cluster, τ n,m is the delay of the mth path, ν n,m is the Doppler frequency offset of the mth path in the nth cluster, v is the speed of the mobile user, f is the signal frequency, φ n,m is the angle between the path and the moving direction of the user, c is the speed of light, Φ n,m is the random phase of the mth path in the nth cluster, a(θ n,m ,φ n,m ) is the steering vector of the path, which is used to describe the propagation direction of the signal in space, and j is the imaginary unit;

[0068] Design a CSI attention module based on a convolutional neural network and an SE module to extract channel features from communication channel data samples and perform preprocessing of channel features. Channel features are extracted from communication channel data samples through a convolutional neural network, and its specific formula is:

[0069]

[0070] Among them, X fm is the initially extracted feature, is the input CSI data, K is the number of antennas, N is the time step, P is the feature dimension. After the convolution operation, the feature dimension P of the original Xi becomes P′, Conv is the two-dimensional convolution operation, ReLU is the activation function. The convolution layer extracts the channel features within each patch through the sliding window method and integrates the channel features between different patches;

[0071] The SE module is used to further process the characteristics of the input CSI data, aiming to assign weights to different patches. The SE module consists of two parts: the squeezing part and the excitation part:

[0072] The squeezing part applies global average pooling to generate a global description of each channel feature map;

[0073] The excitation part establishes the relationship between channel features through two fully connected layers. First, the dimension is compressed through the ReLU activation function, and then the dimension is restored through another fully connected layer. The sigmoid activation function is used to generate the attention weight tensor, and its specific formula is:

[0074]

[0075] where, X SE represents the weight information of each feature;

[0076] Then, the features are scaled using the attention weights generated by sigmoid, and its specific formula is:

[0077] X sca = X fm ⊙ X SE

[0078] where, X sca represents the scaled feature combining the feature weights and the convolutional features. ⊙ represents element-wise multiplication. Finally, the attention weight tensor is generated. By cascading the CSI attention module multiple times, the effectiveness of feature extraction is further enhanced, and the processed channel feature data is obtained.

[0079] In the above technical solution, a patch refers to dividing the input data into small blocks or local regions for extracting local features.

[0080] In the above technical solution, the downlink channel model is used to accurately simulate the actual physical effects encountered during the signal transmission process from the base station to the user terminal, such as the Doppler effect, time delay, path gain, and multipath propagation, etc. These physical phenomena are described by the parameters in the model (such as complex path gain, Doppler frequency offset, delay, phase, etc.), providing a real and complex training environment for channel prediction.

[0081] In the above technical solution, the channel characteristics provided by the downlink channel model are the basis for feature extraction in the subsequent CNN convolutional layer. The convolutional layer extracts multi-dimensional features from the CSI data output from the model, such as time-domain and frequency-domain features, etc. By extracting these physical characteristics, CNN can generate richer feature representations and enhance the prediction ability of the model.

[0082] In the above technical solution, a downlink channel model is constructed to establish a training environment to simulate the actual wireless channel situation; in the channel prediction training module, the downlink channel model is used for channel estimation, and through the reference signal and CSI feedback mechanism in the downlink channel, channel feature data is extracted and obtained, and these channel feature data are used to train the model to predict future channel state information.

[0083] In the above technical solution, X fm is a feature tensor obtained through convolutional layer operations, which represents the preliminary features extracted from the input channel state information (CSI). Under the action of the convolutional layer, the local features of the signal are captured and integrated, including the features in the time domain and frequency domain; X SE is the CSI data feature processed by the SE module, which is the attention weight tensor output by the SE module and represents the weight information of each feature. After the SE module globally describes the features, it assigns the weight of each feature, thereby enhancing the useful features and suppressing the unimportant or redundant features; X sca is the CSI feature X extracted through convolutional operations fm and the attention weight generated by the SE module SE is a feature tensor obtained by element-wise multiplication, which represents the scaled features. It is the final output that fuses feature extraction and attention weights and is used to scale the weights of the initially extracted features, amplify the important features, and suppress the unimportant features.

[0084] In the above technical solution, by designing a CSI attention module based on a convolutional neural network and an SE module, channel features can be more effectively extracted from communication channel data samples and preprocessed; by cascading the CSI attention module multiple times, the effectiveness of feature extraction can be further enhanced, enabling the model to better adapt to different channel conditions.

[0085] In the above technical solution, the specific method for calculating the total power consumption and the maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system is as follows:

[0086] Assume that the first T units of the hybrid RIS are active RIS units, and the remaining N - T units are passive RIS units, and the maximum amplification gain is α. Then the hybrid RIS phase shift matrix can be expressed as:

[0087] R = Ψ + Φ

[0088] Ψ = C T Α

[0089] Φ = I N Θ

[0090]

[0091] Among them, R is the hybrid RIS phase shift matrix, Ψ is the vortex beam reflection coefficient matrix of the active RIS unit, and Φ is the general beamforming reflection coefficient matrix of the passive RIS unit; C T is the N×N identity matrix, I N is the N×N identity matrix; represents the diagonal matrix of the vortex beam reflection coefficients of the active RIS unit, where is the reflection coefficient of the nth active RIS reflection unit, including the amplitude adjustment A n and the spiral phase adjustment of the vortex beam is the topological charge number of the vortex beam, θ n is the azimuth angle of the nth active RIS reflection unit relative to the center; represents the diagonal matrix of the general beamforming reflection coefficients of the passive RIS unit, where is the coefficient of the nth reflection element, and for i ∈ {1, 2, …, T} there is when i ∈ {T + 1, T + 2, …, N}, there is

[0092] such as Figure 3 As shown, through the uplink, the base station can obtain the channel information from the user to the RIS and from the user to the base station. In the uplink transmission, the signal sent by the base station passes through the direct channel and the reflection channel G, and arrives at the radio frequency module of the kth sensor end. The noise is amplified at the hybrid RIS. Then, the received signal of the radio frequency module of the kth sensor end can be expressed as:

[0093]

[0094] Among them, y k is the received signal of the radio frequency module of the kth sensor end, is the channel between the base station and the hybrid RIS, is the channel between the hybrid RIS and the radio frequency module of the kth sensor end, is the channel between the base station and the radio frequency module of the kth sensor end. R is the vortex beam reflection coefficient matrix of the hybrid RIS, is the additive white Gaussian noise at the radio frequency module of the kth sensor end, is the noise introduced by the active RIS unit of the hybrid RIS, and x represents the signal vector transmitted by the base station;

[0095] The signal-to-interference-plus-noise ratio of the kth downlink sensor end radio frequency module can be expressed as:

[0096]

[0097] where \(r\) k is the effective RIS noise power of the \(k\)-th user, and \(\gamma\) k is the signal-to-interference-plus-noise ratio of the \(k\)-th downlink sensor-end radio frequency module;

[0098] The achievable sum rate of all the sensor-end radio frequency modules is expressed as:

[0099]

[0100] where \(B\) is the channel bandwidth and \(V\) is the achievable sum rate of all the sensor-end radio frequency modules;

[0101] The calculation of the total power consumption of the hybrid RIS-assisted wireless transmission system includes the power consumption of the hybrid RIS and the total power consumption of the system. The power consumption of the hybrid RIS can be expressed as:

[0102]

[0103] \(y\) r \(=\varPsi Gx+\varPsi n\) c

[0104] where \(P\) RIS is the power consumption of the hybrid RIS, and \(y\) r is the reflected signal of the hybrid RIS;

[0105] The total power consumption of the system can be expressed as:

[0106]

[0107] where \(a,b\in(0,1]\) are the amplifier efficiencies of the base station and the active RIS, \((\|\mathbf{W}\) c \| 2 +\|\mathbf{W}\) r \| 2 ) is the transmit power of the base station, \(P\) c is the hardware power consumed by the sensor-end radio frequency module and the base station, \(P\) r is the circuit power consumed by the hybrid RIS, \(\mathbf{W}\) c and \(\mathbf{W}\) r are the jointly optimized transmit beamforming matrices of the base station, \(\mathbf{W}\) c is the beamforming matrix transmitted by the base station, and \(\mathbf{W}\) r represents the auxiliary beamforming matrix of the reflected signal of the hybrid RIS;

[0108] In the above technical solution, calculating the total power consumption and the maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system can comprehensively evaluate and optimize the energy consumption and communication performance of the system. The calculation of the total power consumption covers the base station transmission power, the transmission energy consumption of the hybrid RIS, the static power consumption, etc., which helps to maximize the energy efficiency in system design and optimization, reduce the operating cost and energy consumption. The calculation of the maximum achievable sum rate reflects the highest communication capacity of the system under ideal conditions, providing a key basis for optimizing parameters such as beamforming and phase shift matrix, thereby improving the communication quality and spectral efficiency. By reasonably controlling the total power consumption and pursuing the maximum achievable sum rate, the energy consumption and performance of the system can be effectively balanced to meet the requirements of different application scenarios. In addition, the hybrid RIS scheme is superior to the traditional passive or active RIS scheme in terms of energy efficiency and communication performance, and can effectively overcome the signal fading problem and improve the system security rate.

[0109] In the above technical solution, the set goal of maximizing the secure energy efficiency refers to constructing a problem model for maximizing the secure energy efficiency of the hybrid RIS-assisted wireless transmission system, including the following steps:

[0110] The construction of the problem of maximizing the secure energy efficiency of the hybrid RIS-assisted wireless transmission system includes jointly optimizing the transmit beamforming matrix W of the base station under the constraints of the beam pattern gain, the base station transmission power, and the power and amplitude of the hybrid RIS c and W r and the phase shift matrix R of the RIS to maximize the system energy efficiency. Among them, the beamforming matrix of the base station

[0111] Establish the problem model for maximizing the secure energy efficiency of the hybrid RIS-assisted wireless transmission system as P1:

[0112]

[0113]

[0114] where η is the system energy efficiency, C1 in the five-dimensional constraint condition of the hybrid RIS-assisted communication represents that the maximum transmission power of the base station is P max ; C2 restricts the minimum threshold Γ of the sensing beam pattern gain, where ρ(θ0) represents the beam pattern gain and a(θ0) represents the steering vector at direction θ0; C3 is the total output power limit of the hybrid RIS; C4 is the amplitude constraint of the active RIS unit in the hybrid RIS, and C5 is the amplitude constraint of the passive RIS unit in the hybrid RIS.

[0115] In the above technical solution, by setting the goal of maximizing the secure energy efficiency, the security and energy efficiency of the system are comprehensively considered, enabling the communication system to efficiently utilize energy while ensuring the security of information transmission, thereby improving the overall performance of the system; by maximizing the secure energy efficiency, power can be reasonably allocated to different users or communication links under limited power resources, improving the resource utilization rate of the system. In an environment with eavesdroppers or untrusted relay nodes, maximizing the secure energy efficiency can effectively protect user information and reduce the risk of information leakage. Moreover, the goal of maximizing the secure energy efficiency is applicable to various communication scenarios, such as two-way relay networks, hybrid RIS-assisted communication systems, etc., and can meet the security and energy-saving requirements in different scenarios; by optimizing the energy efficiency, the system's dependence on energy can be reduced, and the reliability of the system in an energy-constrained environment can be improved.

[0116] In the above technical solution, the specific method for dynamic optimization according to the cGAN network is as follows:

[0117] The loss function of the cGAN network is used to optimize the parameters of the generator and the discriminator. The expression of the loss function of the cGAN network is:

[0118]

[0119] Among them, D(H, R) is the output of the discriminator, indicating the probability that the input channel matrix H and the beamforming matrix R are real data. H gen is the channel matrix generated by the generator, R gen is the beam configuration generated by the generator, H real is the real channel matrix, and R real is the real beam configuration;

[0120] Maximizing the discriminator loss function D θ is used to optimize the discriminator parameter θ, and minimizing the in the generator loss function is used to optimize the generator parameter The total objective function of the cGAN network for optimizing the generator parameters and the discriminator parameters can be expressed as:

[0121]

[0122] Among them, represents the generator loss function is minimized, represents the discriminator loss function D θ is maximized. When the value of minimizing the generator loss function meets the set channel state threshold, and the value of maximizing the discriminator loss function D θ meets the set beam configuration threshold, the optimization of the model is completed.

[0123] In the above technical solution, maximizing the loss of the discriminator enables the discriminator to better distinguish real samples and generated samples. The goal of the discriminator is to correctly classify, that is, to identify the false data generated by the generator and try to distinguish the real data as much as possible. By maximizing the loss of the discriminator, the discriminator will improve its ability to identify the difference between the generated data and the real data as much as possible during training, which is equivalent to encouraging the discriminator to make accurate judgments on the false data generated by the generator. The discriminator continuously strengthens its classification ability through this maximization process and can better identify forged data.

[0124] In the above technical solution, minimizing the loss of the generator enables the generator to generate samples as similar as possible to the real data, making it impossible for the discriminator to distinguish between the generated data and the real data. By minimizing the loss of the generator, the generator tries to generate samples closer and closer to the real data, thereby improving the quality of the generated samples. The generator improves the quality of the generated samples by minimizing its loss function. As the data generated by the generator gets closer and closer to the real data, it becomes more and more difficult for the discriminator to distinguish between true and false data. The decrease in the loss of the generator represents a gradual improvement in the quality of the generated data.

[0125] In the above technical solution, by defining the loss function of the cGAN network, it details how to optimize the parameters of the generator and the discriminator to improve the accuracy of the generated channel matrix and beam configuration. It also determines whether the optimization of the generator and the discriminator meets the requirements by setting thresholds to ensure the robustness and stability of the model under different conditions.

[0126] In the above technical solution, the specific method for calculating the joint loss function is:

[0127] The specific method for obtaining the channel state information loss function by calculating the error between the predicted hybrid RIS-assisted channel state information and the real-time collected channel state information is:

[0128] The channel feature data output by the pre-trained DeepSeek-R1 edge communication model is X LLM , X LLM is a matrix belonging to space, where T represents the time step and d represents the feature dimension;

[0129] The dimension conversion is performed through two fully connected layers in the pre-trained DeepSeek-R1 edge communication model, and the conversion formula is:

[0130]

[0131] where L is the predicted channel length, K is the number of antennas, and FC is the fully connected layer;

[0132] Rearrange the above output to obtain the rearranged data feature X re , X re is a matrix belonging to space. Separate the real and imaginary parts of the data and perform a denormalization operation on the result to generate the final channel prediction result The specific calculation formula is as follows:

[0133]

[0134] where X de is the denormalized data, is the predicted channel matrix; "," represents the index for separating different dimensions, and ":" represents selecting all slices of that dimension;

[0135] Calculate the prediction result and the error between the actual channel H f . The error feedback can be measured by the following loss function:

[0136]

[0137] where, ‖·‖2 represents the L2 norm, λ is the regularization coefficient, is the regularization term of the model parameter θ;

[0138] Based on the set five-dimensional constraints of hybrid RIS-assisted communication, the total system power consumption, and the maximum achievable sum rate, the specific calculation formula for the loss function of the beamforming network model is:

[0139] The loss function of the beamforming network is expressed as:

[0140]

[0141] where, represents the loss function of the beamforming network model, T is the number of training samples in a mini-batch in mini-batch processing, η is the system energy efficiency, P max is the maximum transmit power, ρ(θ0) is the beam pattern gain, Γ is the gain threshold, P RIS is the hybrid RIS power, P0 is the power limit value, max(0,Γ - ρ(θ0)), max(0,P RIS - P0), and represent the penalty terms for the five-dimensional constraints of hybrid RIS-assisted communication, and λ1, λ2, λ3, λ4, λ5 are the hyperparameter penalty coefficients corresponding to the respective constraints. max(0,x) means taking x when x > 0 and taking 0 when x ≤ 0;

[0142] According to the channel state information loss function and the loss function of the beamforming network model The specific formula for calculating the combined loss function is:

[0143]

[0144] wherein represents the combined loss function, λ1 and λ2 represent the adjustment parameters of the loss function, and the GradNorm (Gradient Normalization) dynamic weight adjustment algorithm is used to automatically adjust the weights λ1 and λ2 of the loss function.

[0145] In the above technical solution, the dimension conversion refers to converting the original matrix of T (time step) × d (feature dimension) into a matrix of 2K (number of antennas) × L (predicted channel length).

[0146] In the above technical solution, by combining the channel state information loss function and the loss function of the beamforming network model, a method for comprehensively evaluating the model performance is provided, ensuring that multiple objectives are optimized simultaneously during the training process; also, by using the GradNorm dynamic weight adjustment algorithm, the weights of the loss function are automatically adjusted, enabling the model to better balance the optimization between different tasks.

[0147] In the above technical solution, the actual application scenarios of the hybrid RIS adaptive vortex beam generation system are applicable to various scenarios in 5G and 6G communication systems, such as wireless coverage in high-density urban environments, device communication in industrial Internet of Things, etc.; in the 5G communication system, the present invention can significantly improve the spectral efficiency and communication quality; in the 6G communication system, the present invention can achieve low-power, high-precision beamforming and dynamic optimization, meeting the requirements of high-frequency, low-latency, and high-data-rate.

[0148] In the above technical solution, the hybrid RIS adaptive vortex beam generation system has good scalability and can adapt to different communication environments and channel conditions. By adjusting the system parameters and optimization objectives, it can flexibly respond to different application scenarios. For example, in a high-interference environment, the system robustness can be improved by increasing the channel state threshold and beam configuration threshold.

[0149] In the above technical solution, the hybrid RIS adaptive vortex beam generation system takes various measures in terms of communication security, such as encrypting the channel state information and beam configuration data to prevent signal interference and data leakage. In terms of reliability, the system can quickly respond to channel changes through real-time feedback and dynamic optimization, ensuring the stable operation of the system.

[0150] Embodiment 2

[0151] An optimization method based on large models, such as Figure 2 shown, through data selection technology, preprocess the historical uplink and downlink CSI data and millimeter-wave communication channel measurement data sets. Select communication channel data samples based on hybrid RIS from the historical uplink and downlink CSI data, and select vortex beam communication data samples from the millimeter-wave communication channel measurement data set. Pre-train the DeepSeek-R1 model according to the communication channel data samples based on hybrid RIS and the vortex beam communication data samples to obtain the DeepSeek-R1 edge communication model. Predict the hybrid RIS-assisted channel state information according to the DeepSeek-R1 edge communication model. Combine the predicted hybrid RIS-assisted channel state information with the adaptive vortex beam generation module to obtain the optimal vortex beam. Then, jointly optimize and train the above modules to obtain the joint optimization model (DeepSeek-R1 edge communication model).

[0152] A hybrid RIS adaptive vortex beam generation method, which includes the following steps:

[0153] Select communication channel data samples based on hybrid RIS from the historical uplink and downlink CSI data through data selection technology, and select vortex beam communication data samples from the millimeter-wave communication channel measurement data set. Use the communication channel data samples based on hybrid RIS and the vortex beam communication data samples to pre-train the DeepSeek-R1 model by using few-shot learning and prompting engineering techniques to obtain the DeepSeek-R1 edge communication model. Generate a channel matrix and beam configuration through the cGAN network, and dynamically optimize the DeepSeek-R1 edge communication model by using the channel matrix and beam configuration;

[0154] Extract and preprocess the channel features of the communication channel data samples based on hybrid RIS according to the set CSI attention module to obtain the processed channel feature data. Predict the processed channel feature data based on the dynamically optimized DeepSeek-R1 edge communication model to obtain the hybrid RIS-assisted channel state information;

[0155] Calculate the total power consumption and maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system according to the hybrid RIS-assisted channel state information; according to the hybrid RIS-assisted channel state information, the set goal of maximizing the security energy efficiency of the hybrid RIS-assisted wireless transmission system, the total power consumption and maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system, combine the beamforming network to calculate the base station beamforming matrix and hybrid RIS vortex beam phase shift matrix for maximizing the energy efficiency, thereby generating an adaptive vortex beam.

[0156] Embodiment 3

[0157] A computer program product includes a computer program which, when executed by a processor, implements the steps of the method described in Embodiment 2.

[0158] The content not described in detail in this specification belongs to the prior art well-known to those skilled in the art.

[0159] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a system for implementing the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0161] These computer program instructions can also be stored in a computer-readable memory capable of guiding a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction system that implements the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0162] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than limit the scope of its protection. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that after reading the present invention, various changes, modifications or equivalent substitutions can still be made to the specific implementation manners of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A hybrid RIS adaptive vortex beam generation system based on large model optimization, characterized in that: It includes: The model pre-training module is used to select communication channel data samples based on hybrid RIS from historical downlink CSI data through data selection technology, select vortex beam communication data samples from millimeter wave communication channel measurement data sets, use the communication channel data samples based on hybrid RIS and vortex beam communication data samples, use few-sample learning and prompt engineering technology to pre-train the DeepSeek-R1 model, obtain the DeepSeek-R1 edge communication model, generate the channel matrix and beam configuration through the cGAN network, and use the channel matrix and beam configuration to dynamically optimize the DeepSeek-R1 edge communication model; The channel prediction module is used to extract and preprocess the channel characteristics of the communication channel data samples based on the hybrid RIS according to the set CSI attention module to obtain the processed channel characteristic data, and predict the processed channel characteristic data based on the dynamically optimized DeepSeek-R1 edge communication model to obtain the hybrid RIS auxiliary channel state information; The adaptive vortex beam generation module is used to calculate the total power consumption and the maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system according to the hybrid RIS-assisted channel state information; According to the hybrid RIS-assisted channel state information, the set hybrid RIS-assisted wireless transmission system security energy efficiency maximization target, the total power consumption and the maximum achievable rate of the hybrid RIS-assisted wireless transmission system, the base station beamforming matrix and the hybrid RIS vortex beam phase shift matrix with maximized energy efficiency are calculated in combination with the beamforming network, thereby generating an adaptive vortex beam.

2. The hybrid RIS adaptive vortex beam generation system based on large model optimization according to claim 1, characterized in that: The acquisition method based on the dynamically optimized DeepSeek-R1 edge communication model is: The prediction accuracy of the DeepSeek-R1 edge communication model is optimized twice, and the hybrid RIS adaptive vortex beam generation system is multi-task jointly optimized based on the joint loss function. In the hybrid RIS adaptive vortex beam generation system with multi-task joint optimization, the channel matrix and beam configuration are generated according to the generator and the set parameters of the cGAN network in the DeepSeek-R1 edge communication model after the two optimizations. The discriminator in the cGAN network in the DeepSeek-R1 edge communication model after the second optimization compares the difference between the channel matrix generated by the generator and the channel state information collected in real time, and compares the difference between the generated beam configuration and the beam configuration collected in real time. When the difference between the generated channel matrix and the channel state information collected in real time is less than the set channel state threshold, and the difference between the generated beam configuration and the beam configuration collected in real time is less than the set beam configuration threshold, the dynamically optimized DeepSeek-R1 edge communication model is obtained. Otherwise, after updating the parameters of the cGAN network generator and the discriminator, the channel matrix and beam configuration are regenerated according to the generator and the set parameters of the cGAN network.

3. The hybrid RIS adaptive vortex beam generation system based on large model optimization according to claim 2, characterized in that: The specific method of twice optimizing the prediction accuracy of the DeepSeek-R1 edge communication model is as follows: The channel state information loss function is obtained by calculating the error between the hybrid RIS-assisted channel state information and the real-time collected channel state information based on the hybrid RIS, and the prediction accuracy of the DeepSeek-R1 edge communication model is optimized according to minimizing the channel state information loss function to obtain the DeepSeek-R1 edge communication model after the first optimization. Based on the set five-dimensional constraints of hybrid RIS-assisted communication, the total power consumption and maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system are calculated to obtain the loss function of the beamforming network model. It is used to feed back the beamforming matrix and the hybrid RIS phase shift matrix that maximize the energy efficiency to the cGAN network in the DeepSeek-R1 edge communication model after the first optimization in the process of adaptive vortex beam generation, and to perform a second optimization on the prediction accuracy of the DeepSeek-R1 edge communication model after the first optimization through the discriminator in the cGAN network; The specific process of multi-task joint optimization: The process of obtaining the hybrid RIS-assisted channel state information in the channel prediction module and the process of obtaining the base station beamforming matrix and the hybrid RIS vortex beam phase shift matrix for maximizing energy efficiency in the adaptive vortex beam generation module are trained in an alternating training manner until the channel state information loss function and the loss function of the beamforming network model reach a set equilibrium state; on the premise that the channel state information loss function and the loss function of the beamforming network model reach a set equilibrium state, the channel state information loss function and the loss function of the beamforming network model are combined to calculate the joint loss function, and multi-task joint optimization is achieved by minimizing the joint loss function.

4. The hybrid RIS adaptive vortex beam generation system based on large model optimization according to claim 1, characterized in that: The specific method of extracting and preprocessing the channel characteristics of the communication channel data samples according to the set CSI attention module to obtain the processed channel characteristic data is as follows: A training environment is established through a downlink channel model, and channel characteristic data is obtained through channel estimation in the downlink channel model, reference signals in the downlink channel model, and CSI feedback mechanisms in the downlink channel model. The downlink channel model between the base station and the user end is expressed as: Where N is the number of clusters, M n is the number of paths in the nth cluster, β n,m is the complex path gain of the mth path in the nth cluster, τ n,m is the delay of the mth path, ν n,m is the Doppler frequency shift of the mth path in the nth cluster, v is the speed of the mobile user, f is the signal frequency, φ n,m is the angle between the path and the user's moving direction, c is the speed of light, Φ n,m is the random phase of the mth path in the nth cluster, a(θ n,m ,φ n,m ) is the pointing vector of the path, j is the imaginary unit; A CSI attention module based on convolutional neural network and SE module is designed to extract channel features from communication channel data samples and preprocess channel features. Channel features are extracted from communication channel data samples through convolutional neural network. The specific formula is as follows: Among them, X fm For the initially extracted features, is the input CSI data, K is the number of antennas, N is the time step, P is the feature dimension. After the convolution operation, the feature dimension P of the original Xi becomes P′. Conv is a two-dimensional convolution operation, ReLU is the activation function, and the convolution layer extracts the channel features within each patch through a sliding window method and integrates the channel features between different patches. The SE module is used to further process the input CSI data features in order to assign weights to different patches. The SE module consists of two parts: the squeeze part and the excitation part: Global average pooling is applied to the squeezed part to generate a global description of each channel feature map; The excitation part establishes the relationship between channel features through two fully connected layers. First, the dimension is compressed through the ReLU activation function, and then the dimension is restored through another fully connected layer. The sigmoid activation function is used to generate the attention weight tensor. The specific formula is: Among them, X SE Represents the weight information of each feature; Then use the attention weight generated by sigmoid to scale the features. The specific formula is: X sca =X fm ⊙X SE Among them, X sca It is represented as a scaled feature that combines feature weights and convolutional features. ⊙ represents element-level multiplication, which ultimately generates an attention weight tensor. By connecting the CSI attention modules in series multiple times, the processed channel feature data is obtained.

5. The hybrid RIS adaptive vortex beam generation system based on large model optimization according to claim 4, characterized in that: The specific method for calculating the total power consumption and the maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system is: Assuming that the first T units of the hybrid RIS are active RIS units, the remaining NT units are passive RIS units, and the maximum amplification gain is α, the hybrid RIS phase shift matrix can be expressed as: R=Ψ+Φ Ψ=C T A Φ=I N I Where R is the hybrid RIS phase shift matrix, Ψ is the vortex beam reflection coefficient matrix of the active RIS unit, and Φ is the general beamforming reflection coefficient matrix of the passive RIS unit; C T is the N×N unit matrix, I N is the N×N unit matrix; represents the diagonal matrix of the vortex beam reflection coefficients of the active RIS unit, where is the reflection coefficient of the nth active RIS reflection unit, including the amplitude adjustment A n Helical phase modulation of vortex beams l is the topological charge of the vortex beam, θ n is the azimuth angle of the nth active RIS reflector unit relative to the center; represents the general beamforming reflection coefficient diagonal matrix of the passive RIS unit, where is the coefficient of the nth reflective element, and for i∈{1,2,…,T} When i∈{T+1,T+2,…,N}, we have In uplink transmission, the base station sends signals via direct channels and reflection channel G, Arriving at the k-th sensor end RF module, the received signal of the k-th sensor end RF module can be expressed as: Among them, y k is the received signal of the RF module at the kth sensor end, is the channel between the base station and the hybrid RIS, is the channel between the hybrid RIS and the k-th sensor-side RF module, is the channel between the base station and the kth sensor end RF module, R is the vortex beam reflection coefficient matrix of the hybrid RIS, is the additive white Gaussian noise at the RF module of the kth sensor, is the noise introduced by the active RIS unit of the hybrid RIS, and x represents the signal vector transmitted by the base station; The signal-to-interference-to-noise ratio of the RF module at the kth downlink sensor can be expressed as: Among them, r k is the effective RIS noise power of the kth user, γ k is the signal-to-interference-noise ratio of the RF module at the kth downlink sensor end; The reachable sum rate of all sensor-side RF modules is expressed as: Where B is the channel bandwidth, V is the reachable rate of all sensor-side RF modules; The calculation of the total power consumption of the hybrid RIS-assisted wireless transmission system includes the power consumption of the hybrid RIS and the total power consumption of the system. The power consumption of the hybrid RIS can be expressed as: the r =ΨGx+Ψn c Among them, P RIS is the power consumption of hybrid RIS, y r is the reflection signal of the hybrid RIS; The total power consumption of the system can be expressed as: Where a,b∈(0,1] is the amplifier efficiency of the base station and active RIS, (‖W c ‖ 2 +‖W r ‖ 2 ) is the transmission power of the base station, P c is the hardware power consumed by the sensor-side RF module and the base station, P r is the circuit power consumed by the hybrid RIS, W c and W r To jointly optimize the base station’s transmit beamforming matrix, W c is the beamforming matrix transmitted by the base station, W r Table 1 is the auxiliary beamforming matrix of the hybrid RIS reflected signal.

6. The hybrid RIS adaptive vortex beam generation system based on large model optimization according to claim 5, characterized in that: The set security energy efficiency maximization goal refers to constructing a security energy efficiency maximization problem model of a hybrid RIS-assisted wireless transmission system, including the following steps: The construction of the safety energy efficiency maximization problem of the hybrid RIS-assisted wireless transmission system includes jointly optimizing the base station's transmit beamforming matrix W under the constraints of beam pattern gain, base station transmit power and hybrid RIS power and amplitude. c and W r And the phase shift matrix R of RIS, which maximizes the system energy efficiency, where the beamforming matrix of the base station is The model of the security energy efficiency maximization problem of the hybrid RIS-assisted wireless transmission system is established as P1: Among them, η is the system energy efficiency, and C1 in the five-dimensional constraint condition of hybrid RIS-assisted communication means that the maximum transmission power of the base station is P max ; C2 limits the minimum threshold Γ of the perceived beam pattern gain, where ρ(θ0) represents the beam pattern gain and a(θ0) represents the steering vector at the direction θ0; C3 is the total output power limit of the hybrid RIS; C4 is the amplitude constraint of the active RIS unit in the hybrid RIS, and C5 is the amplitude constraint of the passive RIS unit in the hybrid RIS.

7. The hybrid RIS adaptive vortex beam generation system based on large model optimization according to claim 2, characterized in that: The specific method of dynamic optimization based on the cGAN network is: The loss function of the cGAN network is used to optimize the parameters of the generator and the discriminator. The loss function expression of the cGAN network is: Among them, D(H,R) is the output of the discriminator, which indicates the probability of whether the input channel matrix H and beamforming matrix R are real data. gen is the channel matrix generated by the generator, R gen is the beam configuration generated by the generator, H real is the real channel matrix, R real is the actual beam configuration; Maximize D in the loss function θ Used to optimize the discriminator and minimize the loss function For optimizing the generator, the overall objective function of the cGAN network can be expressed as: in, Represents the generator loss function minimize, Denotes the discriminator loss function D θ Maximize, when the generator loss function The minimized value satisfies the set channel state threshold, and the discriminator loss function D θ When the maximized value meets the set beam configuration threshold, the model optimization is completed.

8. The hybrid RIS adaptive vortex beam generation system based on large model optimization according to claim 3, characterized in that: The specific method for calculating the joint loss function is: The specific method of obtaining the channel state information loss function by calculating the error between the predicted hybrid RIS-assisted channel state information and the real-time collected channel state information is: The channel feature data output by the pre-trained DeepSeek-R1 edge communication model is X LLM , X LLM Is a The matrix of space, where T represents the time step and d represents the feature dimension; The dimension conversion is performed through two fully connected layers in the pre-trained DeepSeek-R1 edge communication model. The conversion formula is: Among them, L is the predicted channel length, K is the number of antennas, and FC is the fully connected layer; The above output Rearrange to get the rearranged data feature X re , X re Is a The matrix of the space separates the real and imaginary parts of the data and denormalizes the results to generate the final channel prediction results. The specific calculation formula is: Among them, X de is the denormalized data, is the predicted channel matrix, "," represents the index separating different dimensions, and ":" represents selecting all slices of the dimension; Calculate the prediction results With the actual channel H f The error between , the error feedback can be measured by the following loss function: Among them, ‖·‖2 represents the L2 norm, λ is the regularization coefficient, is the regularization term of the model parameter θ; Based on the set five-dimensional constraints of hybrid RIS-assisted communication, the total system power consumption and the maximum achievable sum rate, the specific calculation formula of the loss function of the beamforming network model is: The loss function of the beamforming network is It is expressed as: in, It is expressed as the loss function of the beamforming network model, T is the number of training samples in a micro-batch in micro-batch processing, η is the system energy efficiency, P max is the maximum transmit power, ρ(θ0) is the beam pattern gain, Γ is the gain threshold, P RIS is the hybrid RIS power, P0 is the power limit value, max(0,Γ-ρ(θ0)), max(0,P RIS -P0), and represents the penalty term of the five-dimensional constraint condition of hybrid RIS-assisted communication, λ1, λ2, λ3, λ4, and λ5 are the hyperparameter penalty coefficients of the corresponding constraints, and max(0,x) means taking x when x>0 and taking 0 when x≤0; According to the channel state information loss function And the loss function of the beamforming network model The specific formula for calculating the joint loss function is: in, Represents the joint loss function, λ1 and λ2 represent the adjustment parameters of the loss function, and the GradNorm dynamic weight adjustment algorithm is used to automatically adjust the weights λ1 and λ2 of the loss function.

9. A hybrid RIS adaptive vortex beam generation method based on large model optimization, characterized in that: It includes the following steps: The communication channel data samples based on hybrid RIS are selected from the historical downlink CSI data through data selection technology, and the vortex beam communication data samples are selected from the millimeter wave communication channel measurement data set. The communication channel data samples based on hybrid RIS and the vortex beam communication data samples are used to pre-train the DeepSeek-R1 model using few-sample learning and prompt engineering technology to obtain the DeepSeek-R1 edge communication model. The channel matrix and beam configuration are generated through the cGAN network, and the DeepSeek-R1 edge communication model is dynamically optimized using the channel matrix and beam configuration. According to the set CSI attention module, the channel features of the communication channel data samples based on hybrid RIS are extracted and preprocessed to obtain the processed channel feature data. The processed channel feature data are predicted based on the dynamically optimized DeepSeek-R1 edge communication model to obtain the hybrid RIS auxiliary channel state information. Calculate the total power consumption and maximum achievable sum rate of the hybrid RIS-assisted wireless transmission system according to the hybrid RIS-assisted channel state information; According to the hybrid RIS-assisted channel state information, the set hybrid RIS-assisted wireless transmission system security energy efficiency maximization target, the total power consumption and the maximum achievable rate of the hybrid RIS-assisted wireless transmission system, the base station beamforming matrix and the hybrid RIS vortex beam phase shift matrix with maximized energy efficiency are calculated in combination with the beamforming network, thereby generating an adaptive vortex beam.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in claim 9 are implemented.

Images