Intelligent reflecting surface assisted non-orthogonal multiple access communication resource optimization system and method based on physical information graph neural network

Abstract

The application discloses an intelligent reflecting surface assisted non-orthogonal multiple access communication resource optimization system and method based on a physical information graph neural network, and relates to the technical fields of wireless communication and artificial intelligence. After collecting channel state information, the method constructs a two-part graph with the intelligent reflecting surface reflecting unit and the user as nodes by a physical information graph neural network module, takes the cascade channel gain from the base station to the reflecting unit and from the reflecting unit to the user as an edge attribute, and calculates the attention weight accordingly. After updating the node representation through message passing, the output is sent to a hierarchical multi-agent reinforcement learning module. The module outputs the intelligent reflecting surface phase control action in the slow time scale, outputs the power allocation and user scheduling action in the fast time scale, and performs communication transmission. The application also realizes cross-scene deployment through federated transfer learning. The application reduces the resource optimization complexity, improves the spectrum efficiency, energy efficiency and cross-scene generalization ability.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication and artificial intelligence, specifically to an intelligent reflector-assisted nonorthogonal multiple access communication resource optimization system and method based on physical information graph neural networks, applicable to intelligent wireless resource management scenarios in sixth-generation mobile communication systems. Background Technology

[0002] As sixth-generation mobile communication systems evolve towards ultra-large-scale connectivity, high spectral efficiency, and high energy efficiency, traditional communication systems relying on fixed wireless propagation environments can no longer meet the communication demands of massive terminal access and dynamic network environments. Intelligent Reflector System (RIS) reconstructs the wireless propagation environment through a large number of tunable passive reflector elements, while Non-Orthogonal Multiple Access (NOMA) improves spectral efficiency and user connectivity through power domain multiplexing. Combining RIS with NOMA allows for the manipulation of the equivalent channel gain of different users using RIS to enhance channel differences between users. This benefits user sequencing, power allocation, and serial interference cancellation in NOMA, improving system throughput and service fairness, thus becoming an important research direction in the field of wireless communication.

[0003] Existing resource optimization methods for RIS-assisted NOMA communication typically employ traditional mathematical optimization techniques such as alternating optimization, semidefinite relaxation, and continuous convex approximation, decomposing the original non-convex problem into several subproblems and solving them iteratively. While these methods are effective in small-scale scenarios, in multi-cell, large-scale RIS, and massive user scenarios, the high dimensionality of the RIS phase shift variable, coupled with NOMA power allocation, user scheduling, and quality of service constraints, causes the optimization problem to exhibit mixed-integer, non-convex, and high-dimensional coupling characteristics, drastically increasing the solution complexity. Furthermore, traditional methods require numerous iterations to converge, making it difficult to meet the millisecond-level real-time scheduling requirements of dynamic networks; when channel states change, remodeling and solving are necessary, resulting in high online computational overhead and difficulties in engineering deployment.

[0004] To reduce the complexity of online optimization, existing technologies attempt to employ deep reinforcement learning methods, treating the RIS auxiliary channel as a black box mapping relationship and learning the relationship between phase shift configuration and system performance through a large number of samples. However, this black-box learning does not utilize the physical structure of wireless propagation. As the number of RIS reflector units increases, the action space expands dramatically, resulting in low sample efficiency, slow convergence, and unstable training. Some solutions use graph neural networks to process network topology relationships to improve the ability to model large-scale scenes, but they usually rely on data-driven methods to learn edge weights or attention weights, failing to directly embed physical priors such as electromagnetic propagation path contribution and RIS cascade channel amplitude into the network structure. Therefore, it remains difficult to achieve a good balance between training efficiency, generalization, and physical interpretability.

[0005] Furthermore, existing models are typically trained in specific environments, and their performance degrades significantly when deployed to new cells, new frequency bands, or different RIS configurations. This necessitates the re-collection of large amounts of target scenario data and retraining or large-scale fine-tuning, resulting in high deployment costs. Simultaneously, existing methods primarily focus on spectral efficiency, neglecting the comprehensive energy efficiency under constraints such as RIS control power consumption, AI inference computing power consumption, and quality of service. This makes it difficult to achieve coordinated optimization between spectral efficiency, energy efficiency, and quality of service. Summary of the Invention

[0006] To address the challenges in coordinating intelligent reflector phase shift configuration, non-orthogonal multiple access power allocation, user scheduling, and energy efficiency optimization in existing intelligent reflector-assisted non-orthogonal multiple access communication systems, this invention aims to provide a resource optimization method and system for intelligent reflector-assisted non-orthogonal multiple access communication based on a physical information graph neural network. This solution addresses the technical problems of high complexity in joint resource optimization, low efficiency and poor interpretability of black-box reinforcement learning samples, insufficient generalization ability due to the graph neural network's failure to fully utilize electromagnetic propagation physical priors, and lack of comprehensive energy efficiency considerations and difficulties in cross-scenario deployment.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A smart reflector-assisted nonorthogonal multiple access communication resource optimization system based on a physical information graph neural network includes: a base station module, a smart reflector module, a smart reflector control module, a user terminal module, a channel state information acquisition module, an equivalent channel calculation module, a nonorthogonal multiple access user ranking and scheduling module, a multi-objective optimization modeling module, a physical information graph neural network module, a hierarchical multi-agent reinforcement learning module, a federated transfer learning module, a policy output module, a memory, and a processor. The smart reflector module includes multiple tunable reflector units. The smart reflector control module is connected to the smart reflector module, and the channel state information acquisition module is connected to the base station module. The equivalent channel calculation module is connected to the channel state information acquisition module; the non-orthogonal multiple access user ranking and scheduling module is connected to the equivalent channel calculation module; the multi-objective optimization modeling module is connected to the non-orthogonal multiple access user ranking and scheduling module; the physical information graph neural network module is connected to the multi-objective optimization modeling module; the hierarchical multi-agent reinforcement learning module is connected to the physical information graph neural network module and the multi-objective optimization modeling module; the federated transfer learning module is connected to the physical information graph neural network module and the hierarchical multi-agent reinforcement learning module; the policy output module is connected to the hierarchical multi-agent reinforcement learning module; and it is also connected to the base station module and the intelligent reflector control module, respectively.

[0009] This invention also provides a method for optimizing nonorthogonal multiple access communication resources based on a physical information graph neural network, which includes the following steps:

[0010] S1: The channel state information acquisition module acquires the channel state information between the base station and the user, the base station and the smart reflector, and the smart reflector and the user, and sends the acquired channel state information to the equivalent channel calculation module;

[0011] S2: The equivalent channel calculation module calculates the user's equivalent channel gain based on the direct link and the intelligent reflector-assisted reflection link, and sends the calculation results to the non-orthogonal multiple access user sorting and scheduling module.

[0012] S3: The non-orthogonal multiple access user ordering and scheduling module determines the user cluster partitioning, user ordering and scheduling results based on the user's equivalent channel gain, service quality requirements and service status, and sends the determined user cluster partitioning, user ordering and scheduling results to the multi-objective optimization modeling module;

[0013] S4: The multi-objective optimization modeling module constructs a resource optimization model with spectral efficiency and effective energy efficiency as joint objectives, and sets service quality constraints, transmit power constraints, phase state constraints and decision delay constraints. The constructed model information is then sent to the physical information graph neural network module and the hierarchical multi-agent reinforcement learning module.

[0014] S5: The physical information graph neural network module constructs a bipartite graph with intelligent reflective surface reflective units as the first type of nodes and users as the second type of nodes. The channel gain from the base station to the reflective unit and the channel gain from the reflective unit to the user are used as edge attributes. Attention weights are calculated based on the edge attributes. The hidden representation of the nodes is updated through message passing. The intelligent reflective surface phase adjustment vector and state value estimate are output to the hierarchical multi-agent reinforcement learning module.

[0015] S6: The hierarchical multi-agent reinforcement learning module outputs intelligent reflector phase control actions on a slow time scale and outputs non-orthogonal multiple access power allocation and user scheduling actions on a fast time scale, and sends the phase control actions, power allocation actions and user scheduling actions to the policy output module.

[0016] S7: The strategy output module sends the intelligent reflector phase shift configuration command to the intelligent reflector control module, and sends the user power allocation command and user scheduling command to the base station module;

[0017] S8: The base station module performs non-orthogonal multiple access superposition transmission according to the received instructions, the intelligent reflector control module adjusts the phase of each reflector unit according to the received instructions, and the user terminal module receives the signal and performs serial interference cancellation.

[0018] Furthermore, in step S5, the attention weights are calculated based on the edge attributes as follows: Let the channel amplitude from the b-th base station to the l-th reflecting unit in the m-th smart reflector be... The channel amplitude from the l-th reflecting unit in the m-th smart reflector to the k-th user is: Then edge attributes Attention weight ,in This is a learnable temperature parameter.

[0019] Furthermore, the effective energy efficiency in step S4 is expressed as: , where R sum P is the total system speed. total This is the total power consumption of the system, and P tx For base station transmission power, For power amplifier efficiency, P circuit For the power consumption of base station and user-side circuits, P RIS To control the power consumption of the intelligent reflective surface reflective unit, P AI Power consumption for AI inference calculations.

[0020] Furthermore, in step S6, the reward function of the hierarchical multi-agent reinforcement learning module is set as a combination of the system's joint optimization objective and the service quality penalty term: The joint optimization objective is to optimize both spectral efficiency and energy efficiency. The service quality breach penalty coefficient, R is the minimum rate requirement for user k. k This represents the actual achievable rate for user k.

[0021] Furthermore, in step S6, the hierarchical multi-agent reinforcement learning module is trained using a centralized training and decentralized execution method, and a physical consistency regularization term is introduced during the training phase. The physical consistency regularization term is used to penalize the deviation between the composite channel gain predicted by the model and the composite channel gain calculated based on the physical channel model.

[0022] Furthermore, this method also includes a federated transfer learning step: training a basic resource optimization model in the source scenario using a digital twin environment; fine-tuning the model in the target scenario using a small amount of measured channel data, freezing the physical attention layer in the physical information graph neural network during fine-tuning, and only updating the final readout layer or part of the policy output layer; when there are multiple deployment sites, each site uploads model parameters or gradients without sharing the original user data and original channel data, and the central server performs federated aggregation to obtain the updated global model.

[0023] Furthermore, in step S6, the update period for the slow time scale is 50 milliseconds, and the update period for the fast time scale is 1 millisecond.

[0024] Furthermore, in step S6, the intelligent reflector phase control action output by the hierarchical multi-agent reinforcement learning module is a phase increment. It is not the absolute phase value.

[0025] Furthermore, in the bipartite graph constructed in step S5, the edges between the intelligent reflective surface reflection unit nodes and user nodes only exist between node pairs with physical propagation paths, and the attention weights are adaptively adjusted using the edge attributes as physical priors and learnable temperature parameters during calculation.

[0026] Compared with traditional solutions, the present invention has the following advantages:

[0027] (1) This invention unifies the modeling of intelligent reflector phase configuration, non-orthogonal multiple access power allocation, user scheduling and energy efficiency optimization, and uses a hierarchical multi-agent reinforcement learning approach for solving. Compared with traditional iterative optimization methods such as alternating optimization, semidefinite relaxation and continuous convex approximation, this application does not require repeatedly solving high-dimensional non-convex optimization problems in each decision cycle, which can reduce online computational complexity and improve resource allocation efficiency.

[0028] (2) By jointly optimizing the phase of the intelligent reflector, the power allocation of non-orthogonal multiple access and user scheduling, the present invention enables the direct link and the intelligent reflector-assisted reflection link to work together, thereby enhancing the desired signal, suppressing interference and improving the user's equivalent channel conditions, which is beneficial to improving the overall transmission rate and spectrum efficiency of the system.

[0029] (3) In the optimization process, this invention not only considers the system transmission rate, but also comprehensively considers the base station transmission power consumption, the intelligent reflector control power consumption, and the artificial intelligence inference calculation power consumption to construct an effective energy efficiency index. Therefore, this method can improve communication performance while controlling the total power consumption of the system, and achieve coordinated optimization between spectrum efficiency and energy efficiency.

[0030] (4) This invention embeds the cascaded channel structure between the base station and the smart reflector, and between the smart reflector and the user, into a graph neural network, enabling the model to directly utilize prior information about wireless propagation. Compared to completely black-box deep reinforcement learning methods, this application can reduce the training sample requirements, accelerate model convergence, and reduce training difficulty.

[0031] (5) The physical information graph neural network in this invention uses intelligent reflective surface reflective units and users as graph nodes, and electromagnetic propagation path contribution as edge attributes and attention weights, so that the network structure corresponds to the actual wireless propagation process. Therefore, the phase configuration and power allocation results output by the model have stronger physical interpretability.

[0032] (6) This invention introduces user service quality constraints into the resource optimization objective and sets a service quality breach penalty term in the reward function, so that the model can improve the overall system performance while taking into account the minimum rate requirements of weak users and marginal users, thereby improving the user service quality satisfaction rate and system fairness.

[0033] (7) This invention adopts a federated transfer learning approach, training the base model in the source scenario and fine-tuning it using a small amount of data in the target scenario. Since the physical propagation structure has a certain degree of universality, this method can reduce the retraining cost in new cell, new frequency band or new smart reflector deployment scenarios and improve the model's cross-scenario generalization ability.

[0034] (8) This invention adopts a centralized training and distributed execution structure. During the training phase, global information can be used to learn collaborative optimization strategies, while during the execution phase, each base station agent makes rapid decisions based solely on its local state. After training is completed, the intelligent reflector phase, user power allocation, and user scheduling results can be output through forward inference of the neural network, thus reducing real-time control latency and facilitating deployment in practical wireless communication systems. Attached Figure Description

[0035] This manual includes the following figures, which illustrate the following:

[0036] Figure 1 A schematic diagram of the structure of a smart reflector-assisted nonorthogonal multiple access communication system;

[0037] Figure 2 This is a flowchart of the resource optimization method for this application;

[0038] Figure 3 This is a schematic diagram of the neural network structure of the physical information graph;

[0039] Figure 4 Diagram of a hierarchical multi-agent reinforcement learning decision-making framework;

[0040] Figure 5 Flowchart for deploying federal migration learning. Detailed Implementation

[0041] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, in order to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention, and to facilitate its implementation.

[0042] like Figure 1As shown, the intelligent reflector-assisted non-orthogonal multiple access communication resource optimization system based on a physical information graph neural network provided by this invention includes a base station module, an intelligent reflector module, an intelligent reflector control module, a user terminal module, a channel state information acquisition module, an equivalent channel calculation module, a non-orthogonal multiple access user ranking and scheduling module, a multi-objective optimization modeling module, a physical information graph neural network module, a hierarchical multi-agent reinforcement learning module, a federated transfer learning module, a policy output module, a memory, and a processor. The base station module and the user terminal module are connected via a wireless communication link, and the base station module and the intelligent reflector module are coupled via a wireless propagation path. The intelligent reflector module and the user terminal module are also coupled via a wireless propagation path. The intelligent reflector control module is connected to the intelligent reflector module via a wired or wireless control link and sends phase control commands to the intelligent reflector module. The channel state information acquisition module is connected to the base station module and acquires channel state information between the base station and the user, between the base station and the intelligent reflector, and between the intelligent reflector and the user. The equivalent channel calculation module is connected to the channel state information acquisition module; the non-orthogonal multiple access (NMO) user ranking and scheduling module is connected to the equivalent channel calculation module; the multi-objective optimization modeling module is connected to the NMO user ranking and scheduling module; the physical information graph neural network module is connected to the multi-objective optimization modeling module; the hierarchical multi-agent reinforcement learning module is connected to both the physical information graph neural network module and the multi-objective optimization modeling module; the federated transfer learning module is connected to both the physical information graph neural network module and the hierarchical multi-agent reinforcement learning module; and the policy output module is connected to the hierarchical multi-agent reinforcement learning module and, respectively, to the base station module and the intelligent reflector control module. A computer program is stored in the memory, and when the processor executes the computer program, it implements the resource optimization method described in this invention.

[0043] like Figure 2 The diagram shown is an overall flowchart of the method of the present invention, which includes the following steps.

[0044] S1: The channel state information acquisition module acquires channel state information between the base station and the user, the base station and the smart reflector, and the smart reflector and the user, and sends the acquired channel state information to the equivalent channel calculation module. Specifically, the channel state information acquisition module acquires channel state information between each base station, smart reflector, and user, including the direct channel from the base station to the user, the channel from the base station to the smart reflector, the channel from the smart reflector to the user, user quality of service requirements, the maximum transmit power of the base station, the number and phase constraints of the smart reflector reflector units, the current user queue status or service requirements, the current phase status of the smart reflector, and system power consumption parameters. The system power consumption parameters include base station transmit power consumption, power amplifier efficiency, base station circuit power consumption, smart reflector unit control power consumption, and artificial intelligence inference calculation power consumption. In one implementation, the direct channel from the base station to the user and the channel from the base station to the smart reflector are estimated using uplink sounding reference signals, utilizing the channel reciprocity of a time-division duplex system. The channel from the smart reflector to the user is obtained through user feedback, i.e., the user measures the downlink reference signal and feeds back the channel state information. User quality of service requirements include the minimum rate requirement, maximum latency requirement, and minimum reliability requirement for each user.

[0045] S2: The equivalent channel calculation module calculates the user's equivalent channel gain based on the direct link and the intelligent reflector-assisted reflection link, and sends the calculation result to the non-orthogonal multiple access user ordering and scheduling module. Specifically, the equivalent channel calculation module calculates the equivalent channel from the base station to the user based on the direct channel from the base station to the user, the channel from the base station to the intelligent reflector, the channel from the intelligent reflector to the user, and the phase shift matrix of the intelligent reflector. The equivalent channel consists of two parts: the first part is the direct link from the base station to the user, and the second part is the auxiliary link after reflection by one or more intelligent reflectors. When establishing the network model, the base station is denoted as the b-th base station, and N is configured. t The transmitting antenna; the smart reflector is denoted as the m-th smart reflector, and each smart reflector includes L m There are tunable reflector elements; the user terminal is denoted as the k-th user, and is a single-antenna terminal. The phase shift matrix of each smart reflector is expressed as... ,in This represents the phase of the l-th reflecting element in the m-th smart reflector, satisfying the unit modulus constraint. In one embodiment, the phase is a continuous phase, i.e. ∈[0,2π); In another implementation, the phase is selected from a finite set of phases, for example , where Q represents the number of selectable phase states. In one implementation, the equivalent channel from the b-th base station to the k-th user is represented as . ,in This indicates the direct channel from the base station to the user. This represents the channel from the base station to the m-th smart reflector. This represents the channel from the m-th smart reflector to the k-th user. Let represent the phase shift matrix of the m-th intelligent reflective surface.

[0046] S3: The non-orthogonal multiple access (NMO) user sequencing and scheduling module determines the user cluster partitioning, user sequencing, and scheduling results based on the user's equivalent channel gain, quality of service requirements, and service status, and sends the determined user cluster partitioning, user sequencing, and scheduling results to the multi-objective optimization modeling module. Specifically, the NMO user sequencing and scheduling module divides users into one or more NMO user clusters based on the equivalent channel gain of each user. Within each NMO user cluster, the user decoding order is determined based on the user's equivalent channel gain, quality of service requirements, or service priority. The base station transmits power domain superimposed signals to multiple users on the same time-frequency resource block. For multiple users within the same user cluster, the base station allocates different power ratios to each user and transmits the signals using a non-orthogonal multiple access superimposed coding method. The user side uses a serial interference cancellation method for signal detection. In one embodiment, the base station transmits the signal as follows: C b This represents the user cluster served by the b-th base station. This represents the transmit power allocated to the i-th user. Represents the beamforming vector. This represents the data symbol sent to the i-th user. Considering non-ideal serial interference cancellation, a residual interference factor β∈[0,0.1] is set to characterize the residual interference caused by the decoded signal not being completely eliminated. When β=0, it indicates ideal serial interference cancellation, and when β>0, it indicates the existence of residual interference.

[0047] S4: The multi-objective optimization modeling module constructs a resource optimization model with spectral efficiency and effective energy efficiency as joint objectives, and sets service quality constraints, transmit power constraints, phase state constraints, and decision delay constraints. The constructed model information is then sent to the physical information graph neural network module and the hierarchical multi-agent reinforcement learning module. Specifically, the multi-objective optimization modeling module constructs a resource optimization objective function based on the user reachable rate, total system power consumption, and user service quality constraints. The user reachable rate is calculated using the Shannon formula. In one implementation, the signal-to-interference-plus-noise ratio (SNR) for the k-th user after serial interference cancellation is: Where σ² represents noise power, and the achievable rate for the user is R. k =log2(1+γ k) The system spectral efficiency is obtained by summing the achievable rates of all users, and the total system rate is... The system's effective energy efficiency is expressed as... , where P totalThis represents the total system power consumption. Total system power consumption includes... , where P tx P represents the base station transmit power, η represents the power amplifier efficiency, and P represents the base station transmit power. circuit P represents the power consumption of the base station and user-side circuits. RIS P represents the power consumption of the intelligent reflective surface reflective unit control. AI This represents the computational power consumed when the artificial intelligence model performs reasoning and decision-making. The multi-objective optimization modeling module constructs a joint optimization objective. Where α∈[0,1] represents the trade-off coefficient between spectral efficiency and energy efficiency, R ref and E ref These represent the reference spectral efficiency and reference energy efficiency used for normalization, respectively. Constraints include minimum rate or minimum quality of service constraints for each user, maximum transmit power constraints for each base station, phase constraints for each reflector element of the smart reflector, decoding order constraints for non-orthogonal multiple access users, decision delay constraints, and user scheduling constraints.

[0048] S5: The physical information graph neural network module constructs a bipartite graph with intelligent reflector units as the first type of nodes and users as the second type of nodes. The channel gain from the base station to the reflector unit and the channel gain from the reflector unit to the user are used as edge attributes. Attention weights are calculated based on these edge attributes. Node hidden representations are updated through message passing, and the intelligent reflector phase adjustment vector and state value estimate are output to the hierarchical multi-agent reinforcement learning module. For example... Figure 3 The diagram shows the structure of a Physical Information Graph (PIP) neural network. The PIP neural network module constructs a bipartite graph consisting of intelligent reflector units and user nodes for each intelligent reflector and its affected users. Intelligent reflector units are classified as the first type of node, and users as the second type. The characteristics of intelligent reflector unit nodes include current phase state, phase increment, and reflector unit location or number. The characteristics of user nodes include direct channel access, equivalent channel gain, quality of service requirements, and user service status. Edges are established between intelligent reflector unit nodes and user nodes. Edge attributes are jointly determined by the channel gain from the base station to the intelligent reflector unit and the channel gain from the intelligent reflector unit to the user. Edges only exist between node pairs with a physical propagation path. In one implementation, a physical propagation path is considered to exist when both the channel gain from the base station to the reflector unit and the channel gain from the reflector unit to the user are greater than a preset threshold. The physical edge attributes between the l-th reflector unit and the k-th user are represented as follows: ,in This represents the channel amplitude from the b-th base station to the l-th reflecting element in the m-th smart reflector. This represents the channel amplitude from the l-th reflecting unit in the m-th smart reflector to the k-th user. Further, attention weights are calculated based on the physical edge attributes. Where γ is a learnable temperature parameter. The attention weight is not entirely obtained through data training, but is determined by the contribution of the electromagnetic propagation path and adjusted through the learnable temperature parameter. The attention weight is calculated using the edge attribute as a physical prior and adaptively adjusted through the learnable temperature parameter. The physical information graph neural network module updates the hidden representations of the intelligent reflector unit nodes and user nodes through message passing. In this way, the bilinear cascaded structure in the intelligent reflector auxiliary channel is explicitly embedded in the graph neural network, so that the neural network no longer completely relies on training samples to learn the propagation law. The physical information graph neural network module sends the output intelligent reflector phase adjustment vector and state value estimate to the hierarchical multi-agent reinforcement learning module.

[0049] S6: The hierarchical multi-agent reinforcement learning module outputs intelligent reflector phase control actions on a slow time scale and non-orthogonal multiple access power allocation and user scheduling actions on a fast time scale, and sends the phase control actions, power allocation actions, and user scheduling actions to the policy output module. Figure 5 The diagram shown illustrates a hierarchical multi-agent reinforcement learning decision-making framework. Specifically, the hierarchical multi-agent reinforcement learning module first decomposes the resource optimization problem into a hierarchical Markov decision process, breaking it down into slow-timescale decision-making and fast-timescale decision-making. Slow-timescale decision-making is used for intelligent reflector phase shift configuration. Since user location and large-scale propagation environment changes relatively slowly, the intelligent reflector phase is updated according to a longer period, such as every T... s Updated every millisecond; fast timescale decision-making is used for non-orthogonal multiple access power allocation and user scheduling. Due to the rapid changes in small-scale fading, service queues, and user instantaneous states, power allocation is updated at shorter intervals, such as every terahertz (T). f Updated every millisecond, where T s >T f In one implementation, T s Set to 50 milliseconds, T f Set to 1 millisecond; in high-speed movement scenarios, T s It can be adjusted to 20 milliseconds, T in low-speed scenarios sThe time interval can be adjusted to 100 milliseconds. This hierarchical processing avoids the high-dimensional action space problem caused by a single agent simultaneously controlling all intelligent reflector phases and all user power. The hierarchical multi-agent reinforcement learning module divides each base station and its associated intelligent reflectors into one agent. Each agent is responsible for non-orthogonal multiple access power allocation, user scheduling, and intelligent reflector phase control associated with that base station within its cell. Each agent's state space includes the direct link channel state information of users in the cell, the intelligent reflector auxiliary link channel state information of users in the cell, the equivalent channel gain of each user, the service quality satisfaction of each user, the current intelligent reflector phase state, the current non-orthogonal multiple access user cluster division, and neighboring cell interference information. Each agent's action space includes the non-orthogonal multiple access user power allocation ratio, user scheduling actions, and intelligent reflector phase adjustment amounts. The intelligent reflector phase action is a phase increment. Instead of directly outputting the absolute phase, using phase increment output ensures a smoother phase shift adjustment process and reduces the control overhead caused by frequent large-amplitude phase switching. The reward function of the hierarchical multi-agent reinforcement learning module is set as a combination of the system's joint optimization objective and the service quality penalty term. Where λ represents the service quality breach penalty coefficient, R represents the minimum rate requirement of the k-th user. k This represents the actual reachable rate for the k-th user.

[0050] The hierarchical multi-agent reinforcement learning module trains the intelligent decision-making model using a centralized training and distributed execution approach. For example... Figure 4As shown, during the training phase, the states and actions of each agent are input into a centralized evaluation network. This network evaluates the impact of joint actions on the global objective function based on the global network state, actions of each base station, actions of each intelligent reflector, and performance feedback from each user. In one implementation, a multi-agent deep deterministic policy gradient algorithm is used for training. Each agent includes a policy network and an evaluation network. The policy network takes its local state as input and outputs actions, while the evaluation network takes the global state and the actions of all agents as input and outputs a Q-value. The training process includes the following sub-steps: First, initialize the parameters of the physical information graph neural network, policy network, and evaluation network; then, reset the environment and obtain the initial state at the beginning of each training round; next, each agent selects an action based on the current state, choosing the smart reflector phase action on a slow time scale and the non-orthogonal multiple access power allocation action on a fast time scale; execute the action and obtain user rate, energy efficiency, service quality satisfaction, and reward value; store the state, action, reward, and next state in the experience replay cache; randomly sample a batch of training samples from the experience replay cache; update the evaluation network to approximate the action value under the current policy; update the policy network to output resource control actions that can improve the reward value; add a physical consistency regularization term to ensure that the network's internal representation is consistent with the smart reflector-assisted equivalent channel structure; and softly update the target network parameters. During the execution phase, each agent outputs resource control actions for its own cell based only on its local observation state, including non-orthogonal multiple access power allocation, user scheduling, and smart reflector phase adjustment.

[0051] S7, the strategy output module sends the intelligent reflector phase shift configuration command to the intelligent reflector control module, and sends the user power allocation command and user scheduling command to the base station module.

[0052] S8: The base station module performs non-orthogonal multiple access (NOA) superposition transmission according to the received instructions. The intelligent reflector control module adjusts the phase of each reflector unit according to the received instructions. The user terminal module receives the signal and performs serial interference cancellation. Specifically, the base station module performs NOA superposition transmission according to the received instructions, superimposing the data symbols of multiple users according to the power allocation ratio, then precoding them through beamforming vectors, and finally transmitting them through the radio frequency link. The intelligent reflector control module adjusts the phase state of each reflector unit according to the received instructions through the controller. In one embodiment, the intelligent reflector controller is implemented using a field-programmable gate array (FPGA), which can control the phase of each reflector unit with microsecond-level precision. The user terminal module performs serial interference cancellation according to the decoding order. Specifically, the user first decodes the user data with the strongest signal, subtracts this data from the received signal, then decodes its own data, and so on until it completes its own data reception.

[0053] In addition to the steps described above, this invention also includes a federated transfer learning step, whereby the federated transfer learning module adapts the trained resource optimization model to different deployment scenarios. For example... Figure 5 The diagram shows the deployment flowchart for federated transfer learning. In the source scenario, a basic resource optimization model is trained using a digital twin environment or simulation environment. The digital twin environment generates training samples based on the cell layout, building distribution, user location, smart reflector location, and channel model. In the target scenario, the model is fine-tuned using a small amount of measured channel data. During fine-tuning, the physical attention layer in the physical information graph neural network (i.e., the edge attribute and attention weight calculation layer) is frozen, and only the final readout layer or part of the policy output layer is updated, thus adapting to the target scenario while retaining knowledge of the electromagnetic propagation structure. When multiple deployment sites exist, each site uploads model parameters or model gradients without sharing the original user data and original channel data, and the central server performs federated aggregation.

[0054] After training or fine-tuning, the real-time acquired channel state information, user quality of service requirements, and current smart reflector state are input into the intelligent decision-making model. The policy output module outputs the following control results: phase configuration of each smart reflector reflector unit. User ordering in each non-orthogonal multiple access user cluster, and power allocation ratio for each user. User scheduling results, and, if necessary, output base station beamforming vectors or beam selection results.

[0055] The role and effect of the embodiments

[0056] This invention embeds the physical structure of the intelligent reflector-assisted channel into a graph neural network, enabling the model to directly utilize known propagation laws and reducing reliance on a large number of training samples. By employing a physical attention mechanism instead of purely data-driven attention learning, the network prioritizes reflection units and paths that contribute significantly to the user's equivalent channel, resulting in better physical interpretability and higher training efficiency. Through hierarchical timescale decomposition, RIS phase configuration and NOMA power allocation are decided on slow and fast timescales respectively, reducing the action space dimension of a single decision. A multi-agent collaborative mechanism is used, leveraging global information for collaborative optimization during training and independent decision-making using local information during execution, balancing global performance and distributed deployment efficiency. A joint optimization objective of spectral efficiency and effective energy efficiency is constructed to strike a balance between communication performance, energy consumption, and computational overhead. Service quality constraints and reward / penalty mechanisms ensure user service needs are met. Federated transfer learning enhances cross-scenario deployment capabilities. Therefore, this invention collaboratively addresses the problems of high complexity, low sample efficiency, weak generalization ability, insufficient energy efficiency considerations, and difficulties in real-time deployment in existing technologies from multiple levels, including physical modeling, network structure, reinforcement learning decision-making, energy efficiency optimization, and cross-scenario transfer learning.

[0057] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution; or the direct application of the inventive concept and technical solution to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A smart reflector-assisted nonorthogonal multiple access communication resource optimization system based on a physical information graph neural network, characterized in that, include: The system comprises a base station module, an intelligent reflector module, an intelligent reflector control module, a user terminal module, a channel state information acquisition module, an equivalent channel calculation module, a non-orthogonal multiple access user ranking and scheduling module, a multi-objective optimization modeling module, a physical information graph neural network module, a hierarchical multi-agent reinforcement learning module, a federated transfer learning module, a policy output module, a memory, and a processor. The intelligent reflector module includes multiple tunable reflector units. The intelligent reflector control module is connected to the intelligent reflector module. The channel state information acquisition module is connected to the base station module. The equivalent channel calculation module is connected to the channel state information acquisition module. The non-orthogonal multiple access user ranking and scheduling module... The multi-access user ranking and scheduling module is connected to the equivalent channel calculation module; the multi-objective optimization modeling module is connected to the non-orthogonal multi-access user ranking and scheduling module; the physical information graph neural network module is connected to the multi-objective optimization modeling module; the hierarchical multi-agent reinforcement learning module is connected to both the physical information graph neural network module and the multi-objective optimization modeling module; the federated transfer learning module is connected to both the physical information graph neural network module and the hierarchical multi-agent reinforcement learning module; the policy output module is connected to the hierarchical multi-agent reinforcement learning module; and is also connected to the base station module and the intelligent reflector control module, respectively.

2. The optimization method for the intelligent reflector-assisted nonorthogonal multiple access communication resource optimization system based on physical information graph neural network according to claim 1, characterized in that, Includes the following steps: S1: The channel state information acquisition module acquires the channel state information between the base station and the user, the base station and the smart reflector, and the smart reflector and the user, and sends the acquired channel state information to the equivalent channel calculation module; S2: The equivalent channel calculation module calculates the user's equivalent channel gain based on the direct link and the intelligent reflector-assisted reflection link, and sends the calculation results to the non-orthogonal multiple access user sorting and scheduling module. S3: The non-orthogonal multiple access user ordering and scheduling module determines the user cluster partitioning, user ordering and scheduling results based on the user's equivalent channel gain, service quality requirements and service status, and sends the determined user cluster partitioning, user ordering and scheduling results to the multi-objective optimization modeling module; S4: The multi-objective optimization modeling module constructs a resource optimization model with spectral efficiency and effective energy efficiency as joint objectives, and sets service quality constraints, transmit power constraints, phase state constraints and decision delay constraints. The constructed model information is then sent to the physical information graph neural network module and the hierarchical multi-agent reinforcement learning module. S5: The physical information graph neural network module constructs a bipartite graph with intelligent reflective surface reflective units as the first type of nodes and users as the second type of nodes. The channel gain from the base station to the reflective unit and the channel gain from the reflective unit to the user are used as edge attributes. Attention weights are calculated based on the edge attributes. The hidden representation of the nodes is updated through message passing. The intelligent reflective surface phase adjustment vector and state value estimate are output to the hierarchical multi-agent reinforcement learning module. S6: The hierarchical multi-agent reinforcement learning module outputs intelligent reflector phase control actions on a slow time scale and outputs non-orthogonal multiple access power allocation and user scheduling actions on a fast time scale, and sends the phase control actions, power allocation actions and user scheduling actions to the policy output module. S7: The strategy output module sends the intelligent reflector phase shift configuration command to the intelligent reflector control module, and sends the user power allocation command and user scheduling command to the base station module; S8: The base station module performs non-orthogonal multiple access superposition transmission according to the received instructions, the intelligent reflector control module adjusts the phase of each reflector unit according to the received instructions, and the user terminal module receives the signal and performs serial interference cancellation.

3. The intelligent reflector-assisted nonorthogonal multiple access communication resource optimization method based on physical information graph neural network according to claim 2, characterized in that, In step S5, the attention weight is calculated based on the edge attributes as follows: Let the channel amplitude from the b-th base station to the l-th reflecting unit in the m-th smart reflector be... The channel amplitude from the l-th reflecting unit in the m-th smart reflector to the k-th user is: Then edge attributes Attention weight ,in This is a learnable temperature parameter.

4. The intelligent reflector-assisted nonorthogonal multiple access communication resource optimization method based on physical information graph neural network according to claim 2, characterized in that, The effective energy efficiency in step S4 is expressed as: , where R sum P is the total system speed. total This is the total power consumption of the system, and P tx For base station transmission power, For power amplifier efficiency, P circuit For the power consumption of base station and user-side circuits, P RIS To control the power consumption of the intelligent reflective surface reflective unit, P AI Power consumption for AI inference calculations.

5. The intelligent reflector-assisted nonorthogonal multiple access communication resource optimization method based on physical information graph neural network according to claim 2, characterized in that, In step S6, the reward function of the hierarchical multi-agent reinforcement learning module is set as a combination of the system joint optimization objective and the service quality penalty term: The joint optimization objective is to optimize both spectral efficiency and energy efficiency. The service quality breach penalty coefficient, R is the minimum rate requirement for user k. k This represents the actual achievable rate for user k.

6. The intelligent reflector-assisted nonorthogonal multiple access communication resource optimization method based on physical information graph neural network according to claim 2, characterized in that, In step S6, the hierarchical multi-agent reinforcement learning module is trained using a centralized training and decentralized execution method. A physical consistency regularization term is introduced during the training phase. The physical consistency regularization term is used to penalize the deviation between the composite channel gain predicted by the model and the composite channel gain calculated based on the physical channel model.

7. The intelligent reflector-assisted nonorthogonal multiple access communication resource optimization method based on physical information graph neural network according to claim 2, characterized in that, It also includes a federated transfer learning step: training a basic resource optimization model in the source scenario using a digital twin environment; fine-tuning the model in the target scenario using a small amount of measured channel data, freezing the physical attention layer in the physical information graph neural network during fine-tuning, and only updating the final readout layer or part of the policy output layer; when there are multiple deployment sites, each site uploads model parameters or gradients without sharing the original user data and original channel data, and the central server performs federated aggregation to obtain the updated global model.

8. The intelligent reflector-assisted nonorthogonal multiple access communication resource optimization method based on physical information graph neural network according to claim 2, characterized in that, In step S6, the update period for the slow time scale is 50 milliseconds, and the update period for the fast time scale is 1 millisecond.

9. The intelligent reflector-assisted nonorthogonal multiple access communication resource optimization method based on physical information graph neural network according to claim 2, characterized in that, In step S6, the intelligent reflector phase control action output by the hierarchical multi-agent reinforcement learning module is a phase increment. It is not the absolute phase value.

10. The intelligent reflector-assisted nonorthogonal multiple access communication resource optimization method based on physical information graph neural network according to claim 2, characterized in that, The bipartite graph constructed in step S5 has edges between intelligent reflective surface reflective unit nodes and user nodes that exist only between node pairs with physical propagation paths, and the attention weights are calculated using the edge attributes as physical priors and adaptively adjusted through learnable temperature parameters.

Images