A method and system for intelligent optimization of dynamic spectrum sharing and network performance

Abstract

The application provides a dynamic spectrum sharing and network performance intelligent optimization method and system, and relates to the technical field of wireless communication.The method comprises the following steps: constructing a dynamic spectrum sensing graph by fusing user association and interference weight through real-time collection of multi-dimensional information of users; extracting a global feature vector and a user node feature vector by using a graph neural network; generating a multi-dimensional utility weight of each user service by fusing channel quality, service priority and historical fairness based on the model; constructing a multi-objective optimization function with the weighted spectrum efficiency as the target and integrating high-priority service guarantee and edge user performance constraint by using the weight; finally, solving by using an adaptive particle swarm optimization algorithm, and outputting an optimal spectrum allocation strategy, so as to realize the collaborative optimization of efficient spectrum utilization and differentiated service quality guarantee of the application.The application can significantly improve the spectrum utilization rate and intelligently guarantee the service quality of differentiated services in the network.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically, to a method and system for dynamic spectrum sharing and intelligent network performance optimization. Background Technology

[0002] Dynamic Spectrum Sharing (DSS) is crucial for efficient spectrum utilization and smooth network evolution in 5G and future 6G networks. Its core principle is to allow networks of different standards (such as 4G LTE and 5G NR) or networks from different operators to share the same spectrum resources in the time, frequency, or spatial domains. Existing dynamic spectrum sharing schemes typically use a single, network-level performance metric as their core optimization objective, such as maximizing system spectrum efficiency or improving user speeds at the cell edge.

[0003] In related technologies, while dynamic spectrum sharing schemes can effectively improve the physical layer utilization of spectrum resources, when the algorithm aims to maximize the spectrum efficiency of cell edge users, it will continuously tilt high-quality spectrum resources towards edge users with poor channel conditions. Therefore, although the spectrum utilization index meets the requirements under single-objective optimization, it will structurally crowd out user resources located in the central area or carrying high-priority services, preventing central users from obtaining timely and certain resources, resulting in low actual network utility. Summary of the Invention

[0004] The problem addressed by this invention is how to ensure the quality of service for differentiated services in a network while achieving efficient spectrum utilization.

[0005] To address the aforementioned problems, this invention provides a method and system for dynamic spectrum sharing and intelligent network performance optimization.

[0006] In a first aspect, the present invention provides a dynamic spectrum sharing and intelligent network performance optimization method, comprising:

[0007] Obtain real-time information of all user equipment within the area to be optimized. This real-time information includes location identification information, channel status information, and the type and quality of service carried.

[0008] Based on the location identification information, channel state information, service type, and quality of service level of each user equipment, a dynamic spectrum sensing map of the area to be optimized is constructed.

[0009] Based on the dynamic spectrum sensing map, feature extraction is performed to obtain the global network state feature vector of the region to be optimized and the node feature vector of each user equipment.

[0010] Based on the global network state feature vector and the node feature vector of the user equipment, a corresponding multi-dimensional utility weight value is assigned to the service flow of each user equipment, and a multi-objective optimization function is constructed based on the multi-dimensional utility weight value;

[0011] The multi-objective optimization function is solved by particle swarm optimization algorithm to obtain the spectrum resource allocation strategy for the region to be optimized.

[0012] Optionally, constructing a dynamic spectrum sensing map of the area to be optimized based on the location identification information, channel state information, service type, and quality of service level of each user equipment includes:

[0013] The correlation degree between any two user devices is determined based on the location identification information of each user device.

[0014] Using the user equipment as nodes and the correlation degree between any two user equipments as edges, construct the initial graph structure of the region to be optimized.

[0015] The initial interference weight corresponding to the edge is determined based on the channel state information, and the initial interference weight is corrected according to the service type to obtain the final edge weight corresponding to the edge.

[0016] Assign a node priority weight to each node based on the service quality level;

[0017] The initial graph structure is updated based on the node priority weights and the final edge weights to obtain the dynamic spectrum sensing graph.

[0018] Optionally, the step of extracting features based on the dynamic spectrum sensing map to obtain the global network state feature vector of the region to be optimized and the node feature vector of each user equipment includes:

[0019] The dynamic spectrum sensing map is input into a preset graph neural network model;

[0020] Multi-round message passing is performed through the graph convolutional layer in the graph neural network model; wherein, in each round of message passing, the current feature information of the neighboring nodes of each node is aggregated according to the final edge weight of the edge in the dynamic spectrum perception graph to obtain the node feature vector of the node.

[0021] The node feature vectors of all the nodes are taken as input and global aggregation is performed through the graph pooling layer in the graph neural network model to generate the global network state feature vector.

[0022] Optionally, the step of assigning a corresponding multi-dimensional utility weight value to the service flow of each user equipment based on the global network state feature vector and the node feature vector of the user equipment includes:

[0023] The node feature vector of each user device is concatenated with the global network state feature vector to obtain the enhanced feature vector of each user device;

[0024] The enhanced feature vector is input into a multilayer perceptron, and the multilayer perceptron outputs a multidimensional weight vector. The multidimensional weight vector includes the channel quality weight component, the service priority weight component, and the historical resource acquisition fairness weight component of the user equipment.

[0025] The channel quality weight component, the service priority weight component, and the historical resource acquisition fairness weight component are normalized to obtain the multidimensional utility weight value corresponding to the service flow of the user equipment.

[0026] Optionally, constructing a multi-objective optimization function based on the multi-dimensional utility weight values ​​includes:

[0027] Based on the multidimensional utility weight values ​​and the channel state information, a spectral efficiency objective function is constructed.

[0028] Obtain the total available spectrum resources of the region to be optimized, and determine the total spectrum resource constraints based on the total available spectrum resources.

[0029] Based on the service quality level, determine the resource guarantee constraints;

[0030] Based on the location identification information, determine the performance guarantee constraints for the edge user group;

[0031] Based on the spectral efficiency objective function, the multi-objective optimization function is constructed by integrating the total spectral resource constraints, the resource guarantee constraints, and the edge user group performance guarantee constraints.

[0032] Optionally, the step of solving the multi-objective optimization function using a particle swarm optimization algorithm to obtain the spectrum resource allocation strategy for the region to be optimized includes:

[0033] Based on the structure of the spectrum resource allocation strategy, a particle swarm is initialized, wherein the position vector of each particle in the particle swarm corresponds to a candidate strategy, and the velocity vector of each particle in the particle swarm corresponds to the adjustment amount of the candidate strategy;

[0034] Based on the multi-objective optimization function, determine the fitness value of the candidate strategy corresponding to each particle;

[0035] Based on the fitness value, determine the individual position vector of each particle and the global position vector of the particle swarm;

[0036] Based on the individual position vector and the global position vector, the velocity vector and position vector of all the particles are iteratively updated, and the fitness value of the updated particles is obtained. The candidate strategy corresponding to the particle whose fitness value satisfies the convergence condition is taken as the spectrum resource allocation strategy.

[0037] Optionally, constructing a multi-objective optimization function based on the multi-dimensional utility weight values ​​includes:

[0038] Based on the multidimensional utility weight values ​​and the channel state information, a first spectral efficiency objective function is constructed.

[0039] Based on the node feature vector and the historical resources, fairness weight components are obtained, and a second fairness objective function is constructed based on the fairness weight components.

[0040] The global network state feature vector is input into the target weight prediction model to obtain the target fusion weight vector;

[0041] Based on the target fusion weight vector, the first spectral efficiency objective function and the second fairness objective function are weighted and fused to obtain the multi-objective optimization function.

[0042] Optionally, the step of inputting the global network state feature vector into the target weight prediction model to obtain the target fusion weight vector includes:

[0043] The global network state feature vector is input into the input layer of the target weight prediction model;

[0044] The high-level decision features are extracted by performing a nonlinear transformation on the input global network state feature vector through at least one hidden layer in the target weight prediction model.

[0045] The target weight prediction model outputs a model to map and normalize the high-level decision features, thereby generating the target fusion weight vector.

[0046] In this context, each weight component in the target fusion weight vector corresponds to the relative importance of different optimization objectives in the multi-objective optimization function.

[0047] Optionally, the dynamic spectrum sharing and intelligent network performance optimization method further includes:

[0048] Execute the spectrum resource allocation strategy and obtain network performance feedback data after the execution of the spectrum resource allocation strategy;

[0049] A reward signal is determined based on the difference between the network performance feedback data and the preset expected performance target;

[0050] Based on the reward signal, the model parameters of the graph neural network model and the multilayer perceptron are updated using the policy gradient method.

[0051] Secondly, the dynamic spectrum sharing and intelligent network performance optimization system of the present invention includes:

[0052] The data acquisition unit is used to acquire real-time information of all user equipment in the area to be optimized. The real-time information includes location identification information, channel status information, service type and quality of service level.

[0053] The perception map construction unit is used to construct a dynamic spectrum perception map of the area to be optimized based on the location identification information, channel state information, service type and quality of service level of each user equipment.

[0054] The feature extraction unit is used to extract features based on the dynamic spectrum sensing map to obtain the global network state feature vector of the region to be optimized and the node feature vector of each user device.

[0055] A multi-objective optimization strategy unit is used to assign corresponding multi-dimensional utility weight values ​​to the service flow of each user equipment according to the global network state feature vector and the node feature vector of the user equipment, and to construct a multi-objective optimization function based on the multi-dimensional utility weight values;

[0056] The strategy determination unit is used to solve the multi-objective optimization function using the particle swarm optimization algorithm to obtain the spectrum resource allocation strategy for the region to be optimized, and to execute the spectrum resource allocation strategy.

[0057] The dynamic spectrum sharing and intelligent network performance optimization method and system of this invention introduces multi-dimensional utility weight values, comprehensively considering multiple factors such as user location, channel state, service type, and service level, to construct a multi-objective optimization function. This improves spectrum utilization efficiency while effectively ensuring the resource needs of high-priority services and central users, achieving a balance between efficiency and fairness, and between global and local spectrum resources. Traditional dynamic spectrum sharing schemes typically use a single physical layer indicator, such as spectrum efficiency or edge user rate, as the optimization objective, easily leading to resource allocation bias towards a certain type of user, failing to consider both overall network performance and differentiated service needs. By acquiring the service type and service level of user equipment in real time and incorporating it into the dynamic spectrum perception map and feature extraction process, the resource allocation strategy can differentiate service priorities. High-real-time and high-reliability services, such as industrial control and telemedicine, can obtain more stable and timely resource guarantees, while ordinary services can receive optimized scheduling when resources are abundant. This dynamic weight allocation mechanism based on service needs significantly enhances the network's adaptability to diverse service scenarios. By constructing a dynamic spectrum-aware map and extracting global and node feature vectors, this invention can comprehensively capture the spatiotemporal changes in network status and the relationships between users, providing richer input information for resource allocation. Combining particle swarm optimization (PSO) with efficient multi-objective function solutions, the system can quickly generate near-optimal spectrum allocation strategies in complex dynamic environments, improving the network's adaptability to scenarios such as load fluctuations and changes in service distribution. By balancing the resource needs of users at different locations (center and edge) and services of different priorities, it avoids the structural problem of excessively favoring one type of user, leading to a severe performance degradation for other users. Central users and high-value services obtain the necessary resources, while edge users are supported within a reasonable range, thereby maximizing overall network utility at the system level and improving the overall experience for all types of users.

[0058] Furthermore, this invention is applicable to scenarios where multiple standards such as 4G / 5G / 6G share spectrum. It can dynamically adjust the sharing strategy based on real-time service and network status, which not only improves the overall utilization rate of spectrum resources, but also reduces the risk of service interruption caused by spectrum reorganization during network upgrades. It provides operators with a flexible and efficient resource management method for long-term network evolution and multi-generational coexistence.

[0059] In summary, this invention first constructs a perception graph of service-resource association, then extracts decision features, and based on this, establishes a multi-objective optimization model with the core objective of ensuring the experience of high-value services. Finally, it solves for the execution strategy through intelligent algorithms. This invention enables dynamic spectrum sharing to adaptively allocate resources based on real-time service combinations and network conditions, while maintaining the total network spectrum resources constant. This ensures that high-priority services receive deterministic quality of service guarantees while maximizing the overall spectrum utilization efficiency. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the dynamic spectrum sharing and intelligent network performance optimization method in an embodiment of the present invention.

[0061] Figure 2 This is a schematic diagram of the dynamic spectrum sharing and intelligent network performance optimization system in an embodiment of the present invention. Detailed Implementation

[0062] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0063] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0064] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0065] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0066] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0067] Combination Figure 1As shown in the figure, an embodiment of the present invention provides a method for dynamic spectrum sharing and intelligent network performance optimization, comprising:

[0068] Obtain real-time information of all user equipment in the area to be optimized. The real-time information includes location identification information, channel status information, service type and quality of service level.

[0069] Specifically, on the network side, such as centralized unit (CU) or wireless access network intelligent controller, real-time information from the entire network is collected and integrated through periodic signaling interactions with base stations and user equipment (UEs), such as every 1 millisecond or every 10 milliseconds. Location identification information is determined by measuring the uplink reference signal angle of arrival (ARFA) and time difference of arrival (TDOA) of UEs and combining this with the base station topology map, or by utilizing auxiliary information reported by the terminal's built-in Global Navigation Satellite System (GNSS) module. This determines the user's geographical or logical location relative to the serving base station and neighboring cells, categorized into three levels: central area, ordinary area, and edge area. Channel state information is obtained through downlink channel quality indicators (MSI), precoding matrix indicators (PMI), and rank indicators measured and reported by UEs, or uplink channel sounding reference signals measured by the base station. The service type and quality of service (QoS) level are identified by parsing the service flow descriptors issued by the core network or by monitoring the QoS flow identifiers during the radio resource control (RRC) connection establishment / reconfiguration process. Service types can be classified as enhanced mobile broadband, ultra-reliable low-latency communication (ULLS), and massive machine-type communication (MMT). The QoS level corresponds to various parameters of the 5G QoS identifier defined in the 5G standard, such as latency budget, packet error rate, and priority.

[0070] A dynamic spectrum sensing map of the area to be optimized is constructed based on the location identification information, channel state information, service type, and quality of service level of each user equipment.

[0071] Specifically, the network in the area to be optimized is modeled as an attribute heterogeneous graph. Each active user equipment (UE) is abstracted as a node in the graph. The node attribute vector is composed of the user's location identifier (e.g., area category code), channel state information (e.g., signal-to-interference-plus-noise ratio), service type (e.g., one-hot encoding), and service quality level (e.g., latency budget and priority values). Connections are established between user nodes and between user nodes and virtual cell nodes. The weights of the edges between users are determined by measuring or estimating the interference matrix based on their geographical proximity and channel interference coupling; the edges between users and cells represent service relationships. This graph structure is dynamically updated in each optimization cycle, thus forming a dynamic spectrum-aware graph that can simultaneously characterize the network's physical state, user distribution, and service demand features.

[0072] Feature extraction is performed based on the dynamic spectrum sensing map to obtain the global network state feature vector of the region to be optimized and the node feature vector of each user device.

[0073] Specifically, the constructed dynamic spectrum-aware graph is input into a well-designed graph neural network, such as an attention-based graph convolutional network, for feature learning and aggregation. This network, through multi-layer message passing and node attribute aggregation, can capture complex dependencies between nodes and the network topology. Ultimately, the output of the graph neural network comprises two parts: first, a global network state feature vector formed by global pooling of all node information in the entire graph, such as average pooling or attention pooling. This vector comprehensively represents the overall macro-state of the network, including overall load, interference level, and service mixing ratio; second, an updated node feature vector for each user node. This vector not only contains its original attributes but also incorporates contextual information from its neighboring nodes and the global network state, thus providing a more comprehensive characterization of the user's state in the current network environment.

[0074] Based on the global network state feature vector and the node feature vector of the user equipment, a corresponding multidimensional utility weight value is assigned to the service flow of each user equipment, and a multi-objective optimization function is constructed based on the multidimensional utility weight value.

[0075] Specifically, a weighting neural network or a differentiable weighting function is designed, taking the global network state feature vector extracted in the previous step and the node feature vector of each user as input. This function outputs a multi-dimensional utility weight value for each user's service flow. This weight is a vector, with each dimension corresponding to the preference intensity of different optimization objectives such as spectral efficiency, latency, reliability, and fairness.

[0076] The core logic of the weighting mechanism is as follows: for users carrying ultra-reliable low-latency communication services and located at the edge, the weights of latency and reliability dimensions will be adaptively increased; for enhanced mobile broadband users located in central areas, the weight of spectral efficiency dimension may be higher while ensuring their minimum data rate. Subsequently, based on these dynamically allocated multi-dimensional weights, multiple traditional independent optimization objectives, such as total system throughput, worst-case user rate, and service latency satisfaction rate, are weighted and fused, or constrained under a unified utility function framework, thereby constructing a comprehensive multi-objective optimization function whose mathematical expression aims to maximize the weighted comprehensive utility of all users.

[0077] The multi-objective optimization function is solved by particle swarm optimization algorithm to obtain the spectrum resource allocation strategy for the region to be optimized.

[0078] Specifically, the spectrum resource allocation problem, such as allocating resource blocks and power to different users or different standards on a shared spectrum, is encoded as particle position vectors in a particle swarm optimization algorithm. Each particle represents a candidate spectrum allocation scheme. The algorithm evaluates the merits of each particle using a fitness function, which is the multi-objective optimization function constructed in the previous step. The particle swarm, i.e., a set of randomly initialized allocation schemes, iteratively searches the solution space. Each particle updates its velocity, i.e., the adjustment direction and magnitude of the allocation scheme, based on its own historical best position and the historical best position of the entire swarm. Through hundreds to thousands of iterations, the algorithm can efficiently approximate the Pareto optimal solution set or a high-quality compromise solution of this complex, non-convex multi-objective optimization problem. Finally, the particle with the highest fitness is selected from the converged particles, and its position vector is decoded into a specific, executable spectrum resource allocation strategy, which is then distributed to each base station for execution.

[0079] In a preferred embodiment of the present invention, before the dynamic spectrum sharing and intelligent network performance optimization method runs online, a prior offline model training phase is included to initialize the parameters of the graph neural network model, multilayer perceptron, and target weight prediction model. Specifically, an offline training dataset is constructed in a network simulation platform that highly simulates the live network environment, such as one based on OPNET, NS-3, or a specially built 5G system-level simulator, or by collecting anonymized historical operating data from the live network. The data acquisition process simulates or reproduces various typical network scenarios, including different user distribution densities, mobility patterns, service mix ratios (eMBB, URLLC, mMTC), load levels, and channel conditions. For each sampling time (or simulation time slot), the following data are recorded to form a training sample: a) Full state information of the network, including the location of all user equipment, channel state information, service type, and quality of service level. This information will be used to construct a dynamic spectrum sensing map later; b) Obtain the near-optimal spectrum resource allocation strategy at that time by running a traditional, computationally complex optimization algorithm, such as a mathematical programming solver based on the Lagrange multiplier method or a known high-performance heuristic algorithm, as an expert label; c) Calculate the ideal multidimensional utility weight value that each user equipment should be assigned under this allocation strategy. This can be obtained through reverse derivation or sensitivity analysis of the optimization target to the weight, as well as the ideal target fusion weight vector corresponding to the entire network state.

[0080] Specifically, the same model structure defined in constructing the dynamic spectrum sensing map and generating multidimensional utility weights and target fusion weights is adopted, including graph neural networks, multilayer perceptrons, and target weight prediction models. The training objective is divided into two parts: supervised learning tasks (for feature extraction and weight generation): using the collected full-state information of the network (corresponding to a) as input, and a near-optimal spectrum resource allocation strategy (corresponding to b) or derived ideal multidimensional utility weights (corresponding to c) as supervision labels. By designing loss functions, for example, mean squared error loss for weight prediction; and cross-entropy loss or difference loss based on system utility for the final resource allocation result, the model learns the intermediate representation (feature vector) and weight mapping from the original network state to efficient resource allocation decisions. The simulation environment is used as the training environment, and the model combination to be trained is regarded as an agent. The agent's state is the dynamic spectrum sensing map, and its action is the generated multidimensional utility weights, which are then used to obtain the allocation strategy through a fixed, lightweight solver such as a greedy algorithm. After the environment executes this strategy, a reward is given according to a preset reward function, consistent with the reward signal definition in the online phase. Large-scale offline training is performed using algorithms such as deep deterministic policy gradient, enabling the model to learn to maximize long-term cumulative rewards, which directly optimizes the final network performance metrics.

[0081] Specifically, the training process can adopt a phased strategy: Phase 1, basic feature learning: First, supervised learning is used to pre-train the graph neural network with a large number of labeled samples, aiming to enable its output node feature vectors and global feature vectors to effectively distinguish different network states and user roles. Phase 2, decision policy learning: Then, with the graph neural network parameters fixed or a small learning rate used, the subsequent multilayer perceptron and target weight prediction model are trained, enabling them to accurately predict weights close to expert labels based on feature vectors. Phase 3, end-to-end fine-tuning and reinforcement learning: Finally, all models are de-fixed, and with a small learning rate, combined with supervised loss and reinforcement learning reward signals obtained from simulation environment interactions, the entire pipeline is fine-tuned end-to-end to further improve the synergy between modules and the generalization ability of the overall policy. Training uses stochastic gradient descent or its variants, such as the Adam optimizer, and performance is monitored on independent validation sets to prevent overfitting. After offline training, the trained model combination is evaluated using an unseen test dataset containing diverse and complex scenarios. The evaluation metrics not only include the fitting accuracy of the model's predicted weights and labels, but more importantly, the input of the model-generated weights into the particle swarm optimization algorithm to solve for the resource allocation strategy. The resulting core network performance metrics, such as system spectral efficiency, user fairness index, and high-priority service guarantee rate, are then evaluated in a simulation environment. When these metrics reach or exceed preset performance thresholds, for example, reaching over 95% of the performance of traditional optimization algorithms, the model is considered successfully trained. Subsequently, the trained model parameters (weights and biases) are solidified and deployed to the corresponding computing unit on the network side, such as Near-RT RIC, as the initial model for the online inference phase. During online runtime, this model can be directly loaded for forward inference, or an online fine-tuning mechanism based on real network feedback can be enabled in subsequent steps.

[0082] This embodiment of the dynamic spectrum sharing and intelligent network performance optimization method introduces multi-dimensional utility weight values, integrating multiple factors such as user location, channel state, service type, and service level to construct a multi-objective optimization function. This improves spectrum utilization efficiency while effectively ensuring the resource needs of high-priority services and central users, achieving a balance between efficiency and fairness, and between global and local spectrum resources. Traditional dynamic spectrum sharing schemes typically use a single physical layer indicator, such as spectrum efficiency or edge user rate, as the optimization objective, easily leading to resource allocation bias towards a certain type of user, failing to consider both overall network performance and differentiated service needs. By acquiring the service type and service level of user equipment in real time and incorporating it into the dynamic spectrum perception map and feature extraction process, the resource allocation strategy can differentiate service priorities. High-real-time, high-reliability services, such as industrial control and telemedicine, can receive more stable and timely resource guarantees, while ordinary services can receive optimized scheduling when resources are abundant. This dynamic weight allocation mechanism based on service needs significantly enhances the network's adaptability to diverse service scenarios. By constructing a dynamic spectrum-aware map and extracting global and node feature vectors, this embodiment can comprehensively capture the spatiotemporal changes in network status and the relationships between users, providing richer input information for resource allocation. Combining particle swarm optimization (PSO) with efficient multi-objective function solving, the system can quickly generate near-optimal spectrum allocation strategies in complex dynamic environments, improving the network's adaptability in scenarios such as load fluctuations and changes in service distribution. By balancing the resource needs of users at different locations (center and edge) and services with different priorities, it avoids the structural problem of excessively favoring one type of user, leading to a severe performance degradation for other users. Central users and high-value services obtain the necessary resources, while edge users are supported within a reasonable range, thereby maximizing overall network utility at the system level and improving the overall experience for all types of users.

[0083] Furthermore, this embodiment is applicable to scenarios where multiple standards such as 4G / 5G / 6G share spectrum. It can dynamically adjust the sharing strategy based on real-time service and network status, which not only improves the overall utilization rate of spectrum resources, but also reduces the risk of service interruption caused by spectrum reorganization during network upgrades. It provides operators with a flexible and efficient resource management method for long-term network evolution and multi-generational coexistence.

[0084] In summary, this embodiment first constructs a perception graph of service-resource association, then extracts decision features, and based on this, establishes a multi-objective optimization model with the core objective of ensuring the experience of high-value services. Finally, it solves the execution strategy through intelligent algorithms. This embodiment enables dynamic spectrum sharing to adaptively allocate resources based on real-time service combinations and network conditions, while ensuring that high-priority services receive deterministic quality of service guarantees, and maximizing the overall spectrum utilization efficiency, all while keeping the total network spectrum resources constant.

[0085] Optionally, constructing a dynamic spectrum sensing map of the area to be optimized based on the location identification information, channel state information, service type, and quality of service level of each user equipment includes:

[0086] The correlation degree between any two user devices is determined based on the location identification information of each user device.

[0087] Using the user equipment as nodes and the correlation degree between any two user equipments as edges, construct the initial graph structure of the region to be optimized.

[0088] The initial interference weight corresponding to the edge is determined based on the channel state information, and the initial interference weight is corrected according to the service type to obtain the final edge weight corresponding to the edge.

[0089] Assign a node priority weight to each node based on the service quality level;

[0090] The initial graph structure is updated based on the node priority weights and the final edge weights to obtain the dynamic spectrum sensing graph.

[0091] Specifically, based on the normalized relative position coordinates (X, Y) of each user equipment (UE) in its serving base station coordinate system and its geographical region label (e.g., cell center, middle, or edge), the correlation degree between any two UEs is calculated. First, the Euclidean distance between them is calculated, and a distance correlation threshold is set, such as 1 / 3 of the cell radius. If the distance is less than the threshold, the spatial correlation degree is 1; otherwise, it is 0. Simultaneously, if both belong to the cell edge region, a fixed regional correlation factor (e.g., 0.5) is added; otherwise, it is 0. Finally, the comprehensive correlation degree is calculated by summing the spatial correlation degree and the regional correlation factor, resulting in a scalar value between [0, 1.5], used to quantify the potential correlation strength between users due to geographical proximity or being in the same interference-sensitive area. All active UEs within the area to be optimized are abstracted as a set of nodes in an undirected weighted graph, with each node representing a UE. For any two different nodes, an edge is created, forming the initial topology of a complete graph. Each edge is assigned an initial edge weight, which is the correlation degree value calculated above. This complete graph structure ensures that no potential user interactions are overlooked in the initial stage, providing a complete topological foundation for subsequent fine-grained weight adjustments based on channel and service information.

[0092] Meanwhile, the determination of the initial interference weights relies on the signal-to-interference-plus-noise ratio (SINR) index in the channel state information. For each edge connecting two users, the channel quality indicator values ​​from each user to their respective serving base stations are obtained, and the potential interference coefficient is calculated. This coefficient is inversely proportional to the user's channel conditions; that is, the worse the channel conditions, the greater the potential interference weight. Then, the interference coefficients of all edges are normalized by max-min to obtain the initial interference weights. Next, adjustments are made according to the service type, with a preset service type interference sensitivity mapping table. For example, the sensitivity coefficient for ultra-reliable low-latency communication services is 1.5, for enhanced mobile broadband services it is 1.0, and for massive machine-type communication services it is 0.8. If the user service types at both ends of an edge are different, the arithmetic mean of their sensitivity coefficients is taken as the service correction factor for that edge. Finally, the final edge weight is calculated by multiplying the initial interference weight by the service correction factor, integrating three dimensions: spatial correlation, actual channel interference level, and service anti-interference requirements. The service quality level is jointly determined by the priority level (range 1 to 255, with smaller values ​​indicating higher priority) and the latency budget in the 5G QoS Identifier. Each node is assigned a priority weight, calculated by mapping the 5QI priority value to a positive coefficient (e.g., converting the priority value to a number between 0 and 1 using a linear transformation), and then multiplying it by a reference proportion of the latency budget (e.g., system reference latency divided by the user's latency budget). The more urgent the latency budget, the greater the weight. Finally, the priority weights of all nodes are normalized to a sum of 1, representing the relative priority intensity of different users in resource competition. The initial graph structure is then updated to generate a dynamic spectrum awareness graph. The node and edge sets remain unchanged, but the edge weights are updated from the initial correlation to final edge weights incorporating multi-dimensional information. Simultaneously, node attributes, i.e., node priority weights, are added to each node. The resulting graph structure is the dynamic spectrum awareness graph for this period. This graph not only includes the network topology and reflects the strength of interference and correlation between users through edge weights, but also labels the intrinsic business importance of each user through node attributes, serving as direct input for subsequent feature extraction and optimization decisions.

[0093] In a preferred embodiment of the present invention, the feasibility and adaptability of the scheme are systematically improved by introducing an offline training phase for the model and a dynamic spectrum sensing mechanism: offline training utilizes simulation or historical data, and through supervised and reinforcement learning, enables the model to pre-master the ability to generate efficient weight strategies from network states, providing a stable and high-performance initial base for online deployment; at the same time, the explicit hybrid update strategy, namely periodic updates as the main method and instant updates triggered by key events, ensures that the dynamic spectrum sensing map can be refreshed periodically with controllable overhead, and can respond quickly to changes in network topology, services and channels, thereby ensuring the timeliness, accuracy and operational efficiency of the system's decision-making in real complex network environments.

[0094] In this optional embodiment, firstly, by fusing heterogeneous information from multiple sources such as location, channel, service, and quality of service, the physical network is transformed into a weighted attribute graph. This provides a unified, efficient, and comprehensive data structure foundation for subsequent intelligent decision-making, overcoming the drawbacks of information fragmentation and difficulty in joint optimization in traditional solutions. Secondly, a dual criterion based on spatial distance and regional labels is introduced to calculate user correlation. Furthermore, the edge weights are refined by combining channel quality and service type interference sensitivity, enabling the constructed graph to dynamically and accurately reflect the potential interference coupling relationships between users and the tolerance for differences between services, providing a precise basis for interference coordination and resource avoidance. Thirdly, dynamic priority weights are assigned to each user node according to the priority and latency budget in the 5G service quality identifier, directly embedding differentiated service quality requirements into the network topology model. This allows the resource allocation algorithm to naturally perceive the urgency and importance of services from the data structure level. Finally, this dynamically updated spectrum-aware graph, as a holistic and computable state representation, successfully transforms the network resource allocation problem into a graph-based optimization problem, laying a solid foundation for subsequent optimization solutions based on graph feature extraction and swarm intelligence. This systematically supports the core objective of ensuring the quality of multi-service operations while improving spectrum efficiency.

[0095] Optionally, the step of extracting features based on the dynamic spectrum sensing map to obtain the global network state feature vector of the region to be optimized and the node feature vector of each user equipment includes:

[0096] The dynamic spectrum sensing map is input into a preset graph neural network model;

[0097] Multi-round message passing is performed through the graph convolutional layer in the graph neural network model; wherein, in each round of message passing, the current feature information of the neighboring nodes of each node is aggregated according to the final edge weight of the edge in the dynamic spectrum perception graph to obtain the node feature vector of the node.

[0098] The node feature vectors of all the nodes are taken as input and global aggregation is performed through the graph pooling layer in the graph neural network model to generate the global network state feature vector.

[0099] Specifically, the preset graph neural network model is a pre-built and offline-trained executable model framework. The input data structure of this model is defined to strictly correspond to the dynamic spectrum sensing graph, requiring an input graph containing N nodes. Each node It has an initial node feature vector, which is composed of node priority weights. It is composed of basic attributes such as normalized user channel state information values ​​and one-hot encoded service types. Each connection node and edge It has a scalar weight value This refers to the final edge weights. The network side formats and encapsulates the dynamic spectrum sensing graph completed in the current optimization cycle according to the above data structure requirements to form the standard input tensor of the model, and loads it into the input layer of the graph neural network model to start the forward propagation calculation process.

[0100] The graph neural network model contains L sequentially connected graph convolutional layers (e.g., graph attention network layers or weighted graph convolutional layers). For the Lth... Layers (from 1 to The message passing and node feature update operations performed are as follows: For any node in the graph... First, start from all its neighboring nodes. Collect them at the first Feature vectors output by the layer Then, based on the connecting edges final edge weight Calculate the feature pairs of neighboring nodes The contribution weights are determined. In one specific implementation, a weighted summation method is used for aggregation, that is, for each neighbor feature... Multiply by edge weight Summing is then performed to obtain aggregated information. ,in Represents a node The set of neighbors. Then, this aggregated information is combined with the node's... Features of the next higher level The data is then concatenated and processed using a learnable weight matrix. Perform a linear transformation and apply a nonlinear activation function. For example, ReLU, which ultimately generates nodes. In the New feature vectors of the layer The update process can be formally represented as:

[0101] ;

[0102] After such iterative message passing through layer L, each node The final output feature vector It is the node feature vector after being fully enhanced with graph structure information, which integrates its own attributes, multi-hop neighbor states and local network topology information.

[0103] Following the L-layer graph convolutional layer, a graph pooling layer is connected. This layer receives the final feature vectors from all N nodes. As input, the goal of the global aggregation operation is to compress the scattered information of all nodes into a fixed-length vector representing the state of the entire graph. A specific implementation uses a dual-path pooling strategy combining global average pooling and global max pooling: first, the arithmetic mean of the feature vectors of all nodes is calculated along the feature dimension to obtain the average feature vector. Simultaneously, the maximum value of the feature vectors of all nodes is obtained by taking the maximum value of each feature vector along the feature dimension. Then, these two vectors are concatenated (CONCAT) to form the final global network state feature vector. Average pooling captures the overall central tendency of the network, while max pooling highlights the characteristics of extreme or critical nodes in the network. The combination of the two enables… It can comprehensively characterize the overall network load level, performance bottleneck areas, and service distribution overview, providing macroscopic state information. This global vector will be output together with the feature vector of each node for use by the subsequent weight allocation module.

[0104] In this optional embodiment, firstly, the graph convolutional layer, through multiple rounds of message passing based on the final edge weights, ensures that the node feature vector of each user device not only includes its own attributes but also deeply integrates the interference relationships and state information of its multi-hop neighbors, thereby achieving accurate modeling and representation of the local network interference environment and the coupling relationships between users. Secondly, the global network state feature vector generated by aggregating all node features through the global pooling layer can comprehensively and compactly capture the overall macro-situation of the network, such as load distribution, service mixing, and overall channel quality trends, providing key context for system-level optimization. Finally, this feature extraction process transforms the original, high-dimensional, sparse dynamic spectrum sensing graph into a set of low-dimensional, dense, and semantically rich node and global feature vectors, greatly reducing the complexity of subsequent optimization problems. It provides high-quality, computable core input for multi-objective resource allocation based on intelligent algorithms and is a key bridge for the entire method to achieve intelligent decision-making and performance leap.

[0105] Optionally, the step of assigning a corresponding multi-dimensional utility weight value to the service flow of each user equipment based on the global network state feature vector and the node feature vector of the user equipment includes:

[0106] The node feature vector of each user device is concatenated with the global network state feature vector to obtain the enhanced feature vector of each user device;

[0107] The enhanced feature vector is input into a multilayer perceptron, and the multilayer perceptron outputs a multidimensional weight vector, which includes the channel quality weight component, the service priority weight component, and the historical resource acquisition fairness weight component of the user equipment.

[0108] The channel quality weight component, the service priority weight component, and the historical resource acquisition fairness weight component are normalized to obtain the multidimensional utility weight value corresponding to the service flow of the user equipment.

[0109] Specifically, assuming that after extraction by a graph neural network, the user equipment The node feature vector is Its dimensions are The global network state feature vector is Its dimensions are For each user device in the network Perform feature splicing operation: and By directly concatenating the first and last elements along the vector dimension, an enhanced feature vector is generated. This operation is mathematically represented as:

[0110] ;

[0111] in, This represents vector concatenation. The result is... It is a dimension The vector is used to enhance the feature vector. In this way, each user device's local characteristics, such as its own channel, priority, and local interference environment, are fused with the network's global context information, such as network-wide load and service distribution. This allows subsequent weight allocation decisions to be based on a complete view of both local individual states and the global system state. Simultaneously, a multilayer perceptron model is pre-defined, whose structure includes an input layer, at least one fully connected hidden layer, and an output layer. As the input to the input layer. The number of neurons in the input layer is equal to... The hidden layers use fully connected layers and non-linear activation functions, such as ReLU, to transform and abstract the features. The output layer is a fully connected layer with 3 neurons that does not use any activation functions; its output is a three-dimensional raw weight vector. These three scalar outputs correspond to the original weight components of the three dimensions, specifically:

[0112] Channel quality weight components This component is designed to reflect the user equipment. Based on its current and surrounding channel conditions, it has been encoded in the node features The degree of resource allocation that should be obtained by the user is such that the worse the channel conditions and the greater the interference, the higher the original value of this component tends to be output.

[0113] Business priority weight components This component is designed to reflect the user equipment. The service quality level (priority, latency sensitivity) of the services it carries. The higher the service priority and the more urgent the latency budget, the higher the raw value of this component tends to be output.

[0114] Historical resource access fairness weighting This component is designed to learn implicitly or in conjunction with user devices. In recent times, historical information such as the average resource share obtained in the previous few scheduling cycles can be used as part of the input features or through the model's implicit memory to compensate users who have long suffered from insufficient resource acquisition. Users with fewer historical resource acquisitions tend to have a higher original value for this component. The parameters of the multilayer perceptron are learned through offline training to optimize the overall performance of subsequent multi-objective resource allocation.

[0115] For each user device The three-dimensional original weight vector output by the multilayer perceptron is then processed. Normalization is performed to ensure the comparability of weights among different users and to facilitate subsequent optimization function calculations. The Softmax function is used to normalize these three components.

[0116] The calculation process is as follows: First, apply the exponential function to each original component to obtain:

[0117] .

[0118] Then, calculate the sum of these three exponential values:

[0119] ;

[0120] Finally, divide each index value by the sum to obtain the normalized final weight components:

[0121] , , ;

[0122] Therefore, user equipment The multidimensional utility weight values ​​are expressed as follows:

[0123] .

[0124] The weight value satisfies And each component is in Within the interval, the relative importance of the user's service flow in the three core optimization dimensions of compensating for channel disadvantages, ensuring service priority, and maintaining long-term fairness is clearly represented after integrating global and local information.

[0125] In this optional embodiment, firstly, by concatenating the node feature vector representing local characteristics with the global network state feature vector reflecting the global situation, the generated enhanced feature vector enables each user's weight allocation decision to be based on complete information that simultaneously considers their own conditions and the overall system status, overcoming the limitations of isolated decision-making. Secondly, utilizing the nonlinear fitting capability of the multilayer perceptron, original weight components containing three dimensions—channel quality, service priority, and historical fairness—are dynamically generated based on the enhanced features, enabling adaptive learning of the optimal trade-off strategy between different optimization objectives in complex network environments. Furthermore, by performing Softmax normalization on the three original weight components, standardized, additive, and comparable multidimensional utility weight values ​​for each user are finally generated. This not only clarifies the specific importance ratio of each optimization dimension to each user, ensuring the relative fairness of weights among users, but also provides directly usable and numerically stable key input parameters for subsequently constructing a unified multi-objective optimization function, thereby effectively transforming high-dimensional and complex network state information into precise guiding signals driving resource optimization allocation.

[0126] Optionally, constructing a multi-objective optimization function based on the multi-dimensional utility weight values ​​includes:

[0127] Based on the multidimensional utility weight values ​​and the channel state information, a spectral efficiency objective function is constructed.

[0128] Obtain the total available spectrum resources of the region to be optimized, and determine the total spectrum resource constraints based on the total available spectrum resources.

[0129] Based on the service quality level, determine the resource guarantee constraints;

[0130] Based on the location identification information, determine the performance guarantee constraints for the edge user group;

[0131] Based on the spectral efficiency objective function, the multi-objective optimization function is constructed by integrating the total spectral resource constraints, the resource guarantee constraints, and the edge user group performance guarantee constraints.

[0132] Specifically, when constructing the spectral efficiency objective function, it is first based on the multidimensional utility weight values ​​of each user equipment, including the normalized channel quality weight. Business priority weight And fairness weight And channel state information, such as CQI or SINR values, to calculate user In resource block Spectral efficiency This value is typically obtained using the Shannon formula or a predefined mapping table. Subsequently, a comprehensive utility weighting factor is defined. ,in and This is an adjustable scaling factor used to balance the relative importance of channel compensation and service priority; the fairness weight component is usually reflected in the constraints. Ultimately, the spectral efficiency objective function is to maximize the sum of the weighted spectral efficiencies of all users across all resource blocks: ,in The binary decision variable represents the resource block. Assign to user , Total number of users This represents the total number of resource blocks.

[0133] Obtain the total available spectrum resources for the region to be optimized, including the total bandwidth. (Hz) and total transmit power of the base station (W), in a discrete model based on resource block allocation, is equivalent to the total number of resource blocks. The total spectrum resource constraints include two aspects: First, the resource block exclusivity constraint, meaning that each resource block can be allocated to at most one user at any given time to avoid co-channel interference. This can be expressed as follows: for any resource block... ,have Second, total power constraint, assuming power is allocated to users. The resource block set is Each resource block The allocated power is Then the constraint is Power allocation can be a fixed, equal power allocation, i.e. It can also be used as an optimization variable, in which case the problem expands to joint resource block and power allocation optimization.

[0134] Based on service quality levels, such as 5QI priority values ​​and latency budgets, resource guarantee constraints are determined to ensure that high-priority or low-latency services receive minimum performance guarantees. For each user, if their 5QI priority is higher than a preset threshold or their latency budget is lower than a specific threshold, a minimum rate requirement is set. This value can be obtained by querying the service quality level and rate mapping table. The constraint is expressed as: for all users requiring service quality assurance, the following conditions must be met. In addition, scheduling delay constraints for latency-sensitive services can be introduced, such as limiting their maximum waiting time, which can be reflected in the optimization model through linear inequalities.

[0135] Based on location identification information, such as cell edge zone labels, users are divided into edge user groups, and performance guarantee constraints for these groups are determined to improve network fairness and user experience at the edge. These constraints typically set a minimum total performance threshold for each edge user group; for example, requiring the sum of the weighted spectral efficiencies of edge user groups to be no less than a certain percentage of the total weighted spectral efficiency of the entire network. ,like ,Right now ,in This represents the edge user group. Alternatively, the total number of resource blocks allocated to the edge user group can be constrained to be no less than a certain minimum proportion of the total number of resource blocks.

[0136] Based on the aforementioned spectral efficiency objective function, a multi-objective optimization function is constructed by integrating the constraints on total spectrum resources, resource guarantee, and edge user group performance guarantee. This optimization problem is formally a constrained maximization problem: maximizing the weighted sum of spectral efficiencies while satisfying all constraints. Since multiple constraints embody different objectives (such as efficiency, fairness, and guarantee), this problem can be viewed as a single-objective optimization with implicit multi-objectives, or it can be explicitly integrated through weighted summation. The final mathematical model is typically a mixed-integer linear programming or nonlinear programming problem, which can be solved using optimization solvers such as CPLEX and Gurobi, or heuristic algorithms such as genetic algorithms and greedy algorithms, outputting the optimal resource block allocation scheme. and power distribution scheme .

[0137] In a preferred embodiment of the present invention, for services with high quality of service (QoS) levels, such as services with extremely low latency or high reliability corresponding to the 5QI identifier, hard resource guarantee constraints are set. First, the set of all user equipment (UEs) carrying such services is identified. For each UE i in the set, based on the theoretical requirements of its service type, such as the guaranteed bit rate (GBR) and current channel conditions, the minimum number of resource blocks required in each scheduling cycle is calculated. This calculation can be obtained by dividing the GBR by the user's average spectral efficiency under the current channel conditions. The corresponding constraint is: the total number of resource blocks allocated to this user must not be less than this minimum requirement. This constraint ensures that the basic QoS of high-priority services is not sacrificed. Based on the location identification information of the UEs, such as the cell edge area label defined above, all UEs located at the cell edge are selected to form an edge user group E. The edge user group performance guarantee constraint aims to prevent the optimization process from excessively sacrificing the edge user experience. A minimum average performance threshold (in bps / Hz) acceptable to the edge users is set. The constraint is stated as: the average spectral efficiency of all edge user group members is not lower than this threshold. This constraint-driven optimization scheme preserves the necessary performance baseline for edge users by mathematically integrating the objectives and constraints defined in the preceding steps to construct a complete constraint optimization problem: ;

[0138] Constraint: C1 (Resource Exclusivity): Each resource block can be allocated to at most one user device at any given time.

[0139] C2 (Total Power Limit): If power is available, the total transmit power of all user equipment must not exceed the system's maximum permissible power (Pmax).

[0140] C3 (High QoS Guarantee): The total number of resource blocks allocated to each user equipment with a high QoS level. Not less than the minimum number of resource blocks required according to its theoretical calculation. , , .

[0141] C4 (Edge User Guarantee): The average spectral efficiency of the members in the set E consisting of all edge users is not lower than a preset threshold. , ,in, This refers to the number of edge users.

[0142] C5 (Variable Range): All decision variables, such as resource block allocation indicators and power allocation values, must satisfy their domain range. This optimization function model closely integrates improving spectrum efficiency (objective) with ensuring differentiated services and user fairness (constraints), forming a typical multi-objective optimization problem driven by maximizing weighted spectrum efficiency and constrained by multiple services and fairness, providing a precise mathematical description for subsequent solutions.

[0143] In this optional embodiment, firstly, multi-dimensional utility weights reflecting user channel status, service priority, and historical fairness are directly integrated into the spectrum efficiency objective function. This transforms the maximization of total system throughput from a blind process to one with clear and differentiated value orientation, achieving a fundamental shift in the optimization objective from single efficiency maximization to intelligent weighted value maximization. Secondly, by integrating the total spectrum resource constraints, resource guarantee constraints for high-quality-of-service (QoS) level services, and location-based edge user group performance guarantee constraints, the two core requirements of ensuring differentiated service quality and maintaining basic fairness among users are successfully transformed from abstract concepts into concrete and computable mathematical boundary conditions. This forces the algorithm to seek the optimal spectrum efficiency solution while meeting these hard QoS baselines. Finally, this optimization function, as a unified mathematical model, cleverly integrates the three often conflicting objectives of spectrum efficiency, service quality assurance, and user fairness within a clearly constrained framework for joint optimization. This provides a precise and complete mathematical description for subsequent efficient and automated solutions using intelligent optimization algorithms such as particle swarm optimization.

[0144] Optionally, the step of solving the multi-objective optimization function using a particle swarm optimization algorithm to obtain the spectrum resource allocation strategy for the region to be optimized includes:

[0145] Based on the structure of the spectrum resource allocation strategy, a particle swarm is initialized, wherein the position vector of each particle in the particle swarm corresponds to a candidate strategy, and the velocity vector of each particle in the particle swarm corresponds to the adjustment amount of the candidate strategy;

[0146] Based on the multi-objective optimization function, determine the fitness value of the candidate strategy corresponding to each particle;

[0147] Based on the fitness value, determine the individual position vector of each particle and the global position vector of the particle swarm;

[0148] Based on the individual position vector and the global position vector, the velocity vector and position vector of all the particles are iteratively updated, and the fitness value of the updated particles is obtained. The candidate strategy corresponding to the particle whose fitness value satisfies the convergence condition is taken as the spectrum resource allocation strategy.

[0149] Specifically, each possible allocation scheme, i.e., a binary decision matrix, is encoded as a continuous vector, called the particle's position vector. The length of this vector is determined by the product of the number of users and the number of resource blocks in the system, and each component corresponds to the potential association strength between a user and a resource block, with a value between 0 and 1. During initialization, a population of multiple particles (e.g., 50) is randomly generated. The initial position of each particle consists of random numbers uniformly distributed in the interval 0 to 1, and a velocity vector is randomly initialized for each particle to guide the direction and magnitude of position updates in subsequent iterations. To evaluate the merits of each candidate allocation scheme, a fitness function is defined. First, the continuous position vectors of the particles are converted into actual binary allocation decisions using decoding rules. A simple decoding method is to set a threshold (e.g., 0.5). If a component in the position vector exceeds this threshold, the corresponding user is determined to occupy the resource block; otherwise, they are not. Subsequently, based on the decoded allocation matrix, the objective function value of the system is calculated, i.e., the weighted spectral efficiency of all users. Simultaneously, the allocation scheme is checked to ensure it meets all preset constraints, including resource exclusivity, total power limits, minimum resource guarantees for high-priority services, and minimum average performance guarantees for edge users. For schemes that violate these constraints, a penalty function method is used: a large penalty term is deducted from the objective function value based on the severity of the violation. The final net value is the fitness value of the particle; a higher fitness indicates better overall performance.

[0150] Simultaneously, the algorithm simulates swarm intelligence in its iterative search. Each particle records its highest-fit position found historically, known as its individual historical optimum. The entire population shares and tracks the highest-fit position among all particles, known as the global historical optimum. In each iteration, each particle updates its velocity and position based on three factors: its current velocity (inertia), the attraction of its own historical optimum, and the attraction of the global historical optimum. By adjusting the weights of these factors, the algorithm achieves a balance between global exploration and local fine-grained search. After position updates, it ensures that the components remain within the effective range. This evaluation and update process is repeated until a preset convergence condition is met. The convergence condition is typically set to reaching the maximum number of iterations or the global optimal fitness value not significantly improving over multiple generations. When the iteration stops, the final global historical optimum position vector is decoded, and the resulting binary allocation matrix is ​​the optimal (or near-optimal) spectrum resource allocation strategy solved by the algorithm, which can be directly sent to the base station for execution.

[0151] Meanwhile, to improve solution efficiency and quality, this embodiment also introduces two enhancement strategies: adaptive parameter adjustment: the weight controlling the algorithm's search inertia decreases linearly with iterations, allowing the algorithm to focus on extensive exploration in the early stages and refined development in the later stages. Feasibility repair: before evaluating the fitness of particles, the decoded allocation schemes are repaired to ensure they meet key constraints. Repair mainly targets two types of conflicts: first, conflicts where the same resource block is allocated to multiple users, using a greedy rule to retain the user with the greatest utility for that resource block; second, conflicts where high-priority users do not meet resource requirements, compensated by adjusting resource blocks with the least impact on their utility from users who have already met the conditions. Repair ensures that all evaluated schemes are feasible, thereby guiding the algorithm to search efficiently within the feasible solution space.

[0152] In this optional embodiment, firstly, by innovatively encoding the discrete resource block allocation matrix into continuous particle position vectors and designing corresponding decoding rules and penalty function constraint processing mechanisms, the original high-dimensional, discrete, combinatorial optimization NP-hard problem is successfully transformed into a continuous space optimization problem suitable for efficient search by swarm intelligence algorithms, breaking through the real-time bottleneck of traditional exhaustive search or mathematical programming methods. Secondly, the particle swarm algorithm, by simulating social behavior, utilizes individual historical optimality and swarm global optimality information to guide the search direction, enabling it to quickly converge to a high-quality solution in a huge solution space. Its parallel search characteristics are particularly suitable for solving the optimization problem in the large-scale user and resource block scenario described in this paper, obtaining a near-optimal spectrum allocation strategy within an acceptable computation time. Finally, this solution process integrates the intelligent weights and multi-dimensional constraints generated in the previous steps into the fitness function, making the final spectrum resource allocation strategy not only a mathematically optimized solution but also a comprehensive optimal decision that simultaneously considers system spectrum efficiency, differentiated service quality assurance, and fairness among users. This transforms all the previous intelligent perception, feature extraction, and multi-objective modeling results into executable network instructions, completing a closed loop from network state perception to resource optimization configuration.

[0153] Optionally, constructing a multi-objective optimization function based on the multi-dimensional utility weight values ​​includes:

[0154] Based on the multidimensional utility weight values ​​and the channel state information, a first spectral efficiency objective function is constructed.

[0155] Based on the node feature vector and the historical resources, fairness weight components are obtained, and a second fairness objective function is constructed based on the fairness weight components.

[0156] The global network state feature vector is input into the target weight prediction model to obtain the target fusion weight vector;

[0157] Based on the target fusion weight vector, the first spectral efficiency objective function and the second fairness objective function are weighted and fused to obtain the multi-objective optimization function.

[0158] Specifically, firstly, the resource allocation scheme is defined by a set of decision variables to identify whether each resource block is allocated to a specific user equipment. For each possible allocation, the system calculates the achievable spectral efficiency value for that user on that resource block based on real-time channel quality, such as signal-to-interference-plus-noise ratio (SINR). Simultaneously, the system calculates a comprehensive weighting factor for each user, which integrates the user's real-time channel quality and service priority. Finally, the spectral efficiency objective function is defined as the sum of the spectral efficiency values ​​of all users across all resource blocks, multiplied by their corresponding comprehensive weighting factors. The optimization objective is to find the resource allocation scheme that maximizes this sum. The fairness objective function aims to promote fairness in resource allocation among users, paying particular attention to users with historically lower resource acquisition rates or more urgent service needs. In one embodiment, the construction is divided into two steps: First, a fairness weight component, or resource hunger, is calculated for each user. This value integrates the user's service priority and their average resource acquisition amount in recent scheduling history. The lower a user's historical acquisition amount and the higher their service priority, the higher their hunger value, indicating a more urgent current need for resources. Secondly, based on this hunger level, a fairness objective function can be constructed. One approach is to maximize the sum of the weighted product of the actual resource utility allocated to all users and their hunger level. This function tends to direct resources to users with higher hunger levels, thereby achieving a utility-based fairness improvement at the system level.

[0159] In this embodiment, a pre-trained lightweight neural network model, such as a multilayer perceptron, is used to dynamically determine the relative importance of the two objectives. The model's input is a macroscopic feature vector describing the overall state of the current network, including network load, interference level, and the proportion of users of different service types. The model's output is a two-dimensional fusion weight vector, with two weight components corresponding to the spectral efficiency objective and the fairness objective, respectively. A Softmax function ensures that their sum is 1. The model's goal during training is to learn which weight combination, under different network conditions, maximizes a higher-level system performance metric, such as the overall service request fulfillment rate, through the final resource allocation strategy. During online operation, the system inputs real-time global network state features into the prediction model to obtain dynamically adapted fusion weights for the current scenario. Subsequently, a linear weighted sum method is used to combine the spectral efficiency objective function and the fairness objective function into a scalar comprehensive objective function. For example, when the network load is light and high-throughput services dominate, the model may assign a higher weight to the spectral efficiency objective, guiding optimization to focus on increasing total capacity. Conversely, when network edge users are densely concentrated and historically unevenly distributed, the model will increase the weight of the fairness objective, guiding optimization to focus on balancing user experience. This final comprehensive objective function serves as the driving objective for subsequent intelligent optimization algorithms, such as particle swarm optimization.

[0160] In this optional embodiment, firstly, by constructing a first spectral efficiency objective function oriented towards weighted throughput and a second fairness objective function based on node characteristics and historical resources, the two core but often conflicting dimensions of efficiency and fairness are decoupled into clear and quantifiable mathematical expressions, providing an independent foundation for in-depth optimization of each dimension. Secondly, a target weight prediction model is designed. Based on the global network state feature vector reflecting the macro-level network situation, it outputs a fusion weight vector between spectral efficiency and fairness objectives in real time and automatically. This replaces the rigid mode of manually preset fixed weights in traditional methods, enabling the system to intelligently perceive whether the current network should focus on improving overall capacity, such as when the load is light, or on adjusting fairness, such as when edge users are clustered or resource allocation is severely uneven. This achieves a dynamic optimal match between the optimization strategy and the actual network scenario. Ultimately, by linearly weighting and fusing these two objective functions, a unified multi-objective optimization function is formed that can fully utilize channel conditions to pursue high spectral efficiency while proactively and flexibly ensuring the fairness of resource allocation among users. This function serves as a decision engine, driving subsequent optimization algorithms to find Pareto optimal solutions that can adaptively evolve with the network state, thereby maximizing long-term comprehensive utility at the system level.

[0161] Optionally, the step of inputting the global network state feature vector into the target weight prediction model to obtain the target fusion weight vector includes:

[0162] The global network state feature vector is input into the input layer of the target weight prediction model;

[0163] The high-level decision features are extracted by performing a nonlinear transformation on the input global network state feature vector through at least one hidden layer in the target weight prediction model.

[0164] The target weight prediction model outputs a model to map and normalize the high-level decision features, thereby generating the target fusion weight vector.

[0165] In this context, each weight component in the target fusion weight vector corresponds to the relative importance of different optimization objectives in the multi-objective optimization function.

[0166] Specifically, the target weight prediction model is a pre-trained multilayer perceptron. Its core function is to dynamically determine, based on the real-time global state of the network, whether to prioritize improving system spectral efficiency or ensuring fairness among users in the current optimization decision, and output quantified weights to guide multi-objective fusion. The input layer of this model is responsible for receiving a global feature vector representing the current macroscopic state of the network. This vector includes key indicators such as average network load, overall interference level, proportion of cell edge users, and proportion of high-priority service users. In online applications, the system collects and calculates these feature values ​​in real time and inputs them into the model in a predetermined order. The model contains at least one hidden layer, for example, composed of two fully connected layers, each with hundreds of neurons. Each hidden layer performs complex weighted combinations on its inputs and transforms them using nonlinear activation functions such as ReLU. This series of nonlinear processing enables the model to automatically learn and extract deeper and more abstract feature patterns from the original network state indicators, thereby understanding the optimal trade-off between efficiency and fairness under different network conditions, such as high load and high interference, or low load and multiple center users. The model's output layer is a special fully connected layer with the same number of neurons as the number of optimization objectives to be fused; in this embodiment, there are two, corresponding to the spectral efficiency objective and the fairness objective, respectively. This layer first performs a linear combination of the high-level abstract features output from the last hidden layer to obtain an initial two-dimensional score vector. Then, the softmax function is used to normalize this score, ensuring that both output components are positive and sum to 1. The final output is a two-dimensional dynamic fusion weight vector.

[0167] The first component of this vector represents the weight of the spectral efficiency objective, and the second component represents the weight of the fairness objective. Their values ​​directly reflect the model's assessment of the current network situation. For example, when the network load is extremely heavy and the user experience at the edge is generally poor, the model may output a higher fairness weight (e.g., 0.7) and a lower efficiency weight (e.g., 0.3), guiding resource allocation towards fairness. Conversely, when the network load is light and high-throughput services dominate, the model may output a higher efficiency weight (e.g., 0.8), guiding optimization to focus on improving overall capacity. These dynamically generated weights will be directly used to construct the final weighted composite objective function, achieving adaptive and intelligent adjustment to the optimization direction.

[0168] In this optional embodiment, firstly, the model receives a global network state feature vector containing macroscopic information such as network load, interference distribution, and service composition through the input layer, laying a comprehensive data foundation for weight decision-making. Secondly, leveraging the nonlinear transformation capability of at least one hidden layer, the model can deeply mine and abstract the complex relationships between these original features, extracting high-level decision features for judging whether the current network should prioritize efficiency or fairness, achieving an intelligent upgrade from state description to decision logic. Furthermore, the output layer transforms the high-level decision features into a set of non-negative target fusion weight vectors with a sum of 1 through linear mapping and Softmax normalization. This vector precisely represents the relative importance of optimization objectives such as spectral efficiency and fairness in the current network environment in a quantitative manner. Finally, this model, as the core of adaptive decision-making, replaces manually preset fixed weights, enabling the multi-objective optimization function to flexibly adjust its optimization orientation according to the real-time network situation. This ensures that the resource allocation strategy that best suits the current global interests is always generated in a complex dynamic network environment, achieving precise matching and dynamic coordination between the optimization focus and the actual network needs.

[0169] Optionally, the dynamic spectrum sharing and intelligent network performance optimization method further includes:

[0170] Execute the spectrum resource allocation strategy and obtain network performance feedback data after the execution of the spectrum resource allocation strategy;

[0171] A reward signal is determined based on the difference between the network performance feedback data and the preset expected performance target;

[0172] Based on the reward signal, the model parameters of the graph neural network model and the multilayer perceptron are updated using the policy gradient method.

[0173] Specifically, in a preferred embodiment of the present invention, a performance monitoring module is deployed on the network side. After the spectrum resource allocation strategy is executed and a complete scheduling cycle has elapsed, this module collects measurement reports from base stations and user equipment (UEs) to obtain key network performance feedback data. This data specifically includes: the average throughput actually achieved by each UE during the cycle, the packet latency and packet loss rate of the services carried by each UE, the service quality requirements of high-priority services such as URLLC being met, the overall spectrum efficiency of the system (i.e., the ratio of total throughput to total bandwidth), and the average throughput of edge user groups. These observations together constitute the original set for evaluating the effectiveness of the current resource allocation strategy.

[0174] To quantify the effectiveness of the strategy, the system predefines a reward function designed to guide the learning process towards preset system performance targets. Specifically, a set of expected target values ​​is first set, such as the expected system spectral efficiency, the expected minimum throughput for edge users, and the expected satisfaction rate of high-priority services. The reward function calculates the weighted difference between each actual feedback data point and its corresponding expected target, transforming it into a scalar reward signal. A negative penalty is introduced if the average throughput for edge users fails to meet the target. The larger the final calculated reward value, the better the overall performance of the current allocation strategy and the closer it is to the desired ideal state.

[0175] In this embodiment, the entire process—from constructing a dynamic spectrum-aware map and extracting features to generating multi-dimensional utility weights that drive resource allocation—is considered as a policy defined by neural network parameters. This policy takes the original network state as input and outputs the utility weights that ultimately guide resource allocation. To optimize this policy, this embodiment employs a policy gradient method for online learning and updating. The specific steps are as follows: In each decision cycle, a resource allocation policy is generated and executed based on the real-time network state using the current policy parameters; after policy execution, feedback data is collected through a performance monitoring module, and a reward signal is calculated; based on the obtained reward signal, the gradient estimate of the policy parameters is calculated. In practical implementation, this step typically utilizes the automatic differentiation function of deep learning frameworks, treating the reward signal as a loss and calculating the gradient through backpropagation; the gradient ascent algorithm, such as the Adam optimizer, is used to update the policy parameters. The parameters are adjusted along the direction that yields higher rewards, and the learning rate controls the update step size.

[0176] In this embodiment, through the above-mentioned continuous execution-evaluation-update cycle, the parameters of the strategy are continuously optimized, thereby enabling the system to learn to automatically generate multi-dimensional utility weights that can better balance spectrum efficiency, user fairness and service quality assurance, and ultimately achieve self-optimization and continuous improvement of network resource allocation strategy.

[0177] In this optional embodiment, firstly, by acquiring multi-dimensional network performance feedback data including throughput, latency, packet loss rate, and specific service quality satisfaction after executing the strategy, the system can accurately and quantitatively evaluate the actual effect of each resource allocation decision, providing real-world experience samples for model optimization. Secondly, based on these feedback data and the difference between preset expected targets, such as target spectral efficiency and edge user rate thresholds, a composite reward signal is constructed. This signal integrates high-level service objectives and network KPIs into a differentiable scalar guidance signal, transforming the complex multi-objective optimization problem into a unified objective driving model parameter learning. Finally, using the policy gradient method, guided by this reward signal, the parameters of the graph neural network model and the multilayer perceptron are simultaneously updated online and progressively through backpropagation. This allows the entire weight generation strategy to continuously learn from the successes and failures of historical decisions, dynamically adjusting its feature extraction and weight mapping logic. As a result, the system can not only achieve multi-objective trade-offs in a single decision but also continuously adapt to changes in the network environment and business models over time, achieving autonomous evolution and continuous improvement of long-term performance.

[0178] Combination Figure 2 As shown, the dynamic spectrum sharing and intelligent network performance optimization system of this invention includes:

[0179] The data acquisition unit is used to acquire real-time information of all user equipment in the area to be optimized. The real-time information includes location identification information, channel status information, service type and quality of service level.

[0180] The perception map construction unit is used to construct a dynamic spectrum perception map of the area to be optimized based on the location identification information, channel state information, service type and quality of service level of each user equipment.

[0181] The feature extraction unit is used to extract features based on the dynamic spectrum sensing map to obtain the global network state feature vector of the region to be optimized and the node feature vector of each user equipment.

[0182] A multi-objective optimization strategy unit is used to assign corresponding multi-dimensional utility weight values ​​to the service flow of each user equipment according to the global network state feature vector and the node feature vector of the user equipment, and to construct a multi-objective optimization function based on the multi-dimensional utility weight values;

[0183] The strategy determination unit is used to solve the multi-objective optimization function using the particle swarm optimization algorithm to obtain the spectrum resource allocation strategy for the region to be optimized, and to execute the spectrum resource allocation strategy.

[0184] The dynamic spectrum sharing and intelligent network performance optimization system of the present invention has the same advantages over the prior art as the aforementioned dynamic spectrum sharing and intelligent network performance optimization method, and will not be repeated here.

[0185] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for dynamic spectrum sharing and intelligent network performance optimization, characterized in that, include: Obtain real-time information of all user equipment in the area to be optimized, including location identification information, channel status information, service type and quality of service level; Based on the location identification information, channel state information, service type, and quality of service level of each user equipment, a dynamic spectrum sensing map of the area to be optimized is constructed; specifically, this includes determining the correlation degree between any two user equipments based on the location identification information of each user equipment. Using the user equipment as nodes and the correlation degree between any two user equipments as edges, construct the initial graph structure of the region to be optimized. The initial interference weight corresponding to the edge is determined based on the channel state information, and the initial interference weight is corrected according to the service type to obtain the final edge weight corresponding to the edge. Assign a node priority weight to each node based on the service quality level; The initial graph structure is updated based on the node priority weights and the final edge weights to obtain the dynamic spectrum sensing graph; Feature extraction is performed based on the dynamic spectrum sensing map to obtain the global network state feature vector of the region to be optimized and the node feature vector of each user device; specifically, this includes inputting the dynamic spectrum sensing map into a preset graph neural network model; Multi-round message passing is performed through the graph convolutional layer in the graph neural network model; wherein, in each round of message passing, the current feature information of the neighboring nodes of each node is aggregated according to the final edge weight of the edge in the dynamic spectrum perception graph to obtain the node feature vector of the node. The node feature vectors of all the nodes are taken as input and global aggregation is performed through the graph pooling layer in the graph neural network model to generate the global network state feature vector. Based on the global network state feature vector and the node feature vector of the user equipment, a corresponding multidimensional utility weight value is assigned to the service flow of each user equipment, and a multi-objective optimization function is constructed based on the multidimensional utility weight value; wherein, the step of assigning a corresponding multidimensional utility weight value to the service flow of each user equipment based on the global network state feature vector and the node feature vector of the user equipment includes: concatenating the node feature vector of each user equipment with the global network state feature vector to obtain the enhanced feature vector of each user equipment; The enhanced feature vector is input into a multilayer perceptron, and the multilayer perceptron outputs a multidimensional weight vector. The multidimensional weight vector includes the channel quality weight component, the service priority weight component, and the historical resource acquisition fairness weight component of the user equipment. The channel quality weight component, the service priority weight component, and the historical resource acquisition fairness weight component are normalized to obtain the multidimensional utility weight value corresponding to the service flow of the user equipment. The step of constructing a multi-objective optimization function based on the multi-dimensional utility weight values ​​includes: constructing a spectral efficiency objective function based on the multi-dimensional utility weight values ​​and the channel state information; Obtain the total available spectrum resources of the region to be optimized, and determine the total spectrum resource constraints based on the total available spectrum resources. Based on the service quality level, determine the resource guarantee constraints; Based on the location identification information, determine the performance guarantee constraints for the edge user group; Based on the spectral efficiency objective function, the multi-objective optimization function is constructed by integrating the total spectral resource constraint, the resource guarantee constraint, and the edge user group performance guarantee constraint. The multi-objective optimization function is solved by particle swarm optimization algorithm to obtain the spectrum resource allocation strategy for the region to be optimized.

2. The dynamic spectrum sharing and intelligent network performance optimization method according to claim 1, characterized in that, The step of solving the multi-objective optimization function using a particle swarm optimization algorithm to obtain the spectrum resource allocation strategy for the region to be optimized includes: Based on the structure of the spectrum resource allocation strategy, a particle swarm is initialized, wherein the position vector of each particle in the particle swarm corresponds to a candidate strategy, and the velocity vector of each particle in the particle swarm corresponds to the adjustment amount of the candidate strategy; Based on the multi-objective optimization function, determine the fitness value of the candidate strategy corresponding to each particle; Based on the fitness value, determine the individual position vector of each particle and the global position vector of the particle swarm; Based on the individual position vector and the global position vector, the velocity vector and position vector of all the particles are iteratively updated, and the fitness value of the updated particles is obtained. The candidate strategy corresponding to the particle whose fitness value satisfies the convergence condition is taken as the spectrum resource allocation strategy.

3. The dynamic spectrum sharing and intelligent network performance optimization method according to claim 1, characterized in that, The construction of a multi-objective optimization function based on the multi-dimensional utility weight values ​​includes: Based on the multidimensional utility weight values ​​and the channel state information, a first spectral efficiency objective function is constructed. Based on the node feature vector and the historical resources, fairness weight components are obtained, and a second fairness objective function is constructed based on the fairness weight components. The global network state feature vector is input into the target weight prediction model to obtain the target fusion weight vector; Based on the target fusion weight vector, the first spectral efficiency objective function and the second fairness objective function are weighted and fused to obtain the multi-objective optimization function.

4. The dynamic spectrum sharing and intelligent network performance optimization method according to claim 3, characterized in that, The step of inputting the global network state feature vector into the target weight prediction model to obtain the target fusion weight vector includes: The global network state feature vector is input into the input layer of the target weight prediction model; The high-level decision features are extracted by performing a nonlinear transformation on the input global network state feature vector through at least one hidden layer in the target weight prediction model. The target weight prediction model outputs a model to map and normalize the high-level decision features, thereby generating the target fusion weight vector. In this context, each weight component in the target fusion weight vector corresponds to the relative importance of different optimization objectives in the multi-objective optimization function.

5. The method for dynamic spectrum sharing and intelligent network performance optimization according to claim 1, characterized in that, Also includes: Execute the spectrum resource allocation strategy and obtain network performance feedback data after the execution of the spectrum resource allocation strategy; A reward signal is determined based on the difference between the network performance feedback data and the preset expected performance target; Based on the reward signal, the model parameters of the graph neural network model and the multilayer perceptron are updated using the policy gradient method.

6. A dynamic spectrum sharing and intelligent network performance optimization system, characterized in that, include: The data acquisition unit is used to acquire real-time information of all user equipment in the area to be optimized. The real-time information includes location identification information, channel status information, service type and quality of service level. A perception graph construction unit is used to construct a dynamic spectrum perception graph of the region to be optimized based on the location identification information, channel state information, service type, and quality of service level of each user equipment. Specifically, this includes: determining the correlation degree between any two user equipments based on the location identification information of each user equipment; constructing an initial graph structure of the region to be optimized, using the user equipment as nodes and the correlation degree between any two user equipments as edges; determining the initial interference weights corresponding to the edges based on the channel state information, and correcting the initial interference weights according to the service type to obtain the final edge weights; setting node priority weights for each node according to the quality of service level; and updating the initial graph structure according to the node priority weights and the final edge weights to obtain the dynamic spectrum perception graph. The feature extraction unit is used to extract features based on the dynamic spectrum sensing map to obtain the global network state feature vector of the region to be optimized and the node feature vector of each user device. Specifically, it includes: inputting the dynamic spectrum sensing map into a preset graph neural network model; performing multiple rounds of message passing through the graph convolutional layer in the graph neural network model; wherein, in each round of message passing, based on the final edge weights of the edges in the dynamic spectrum sensing map, aggregating the current feature information of the adjacent nodes of each node to obtain the node feature vector of that node; and using the node feature vectors of all nodes as input, performing a global aggregation operation through the graph pooling layer in the graph neural network model to generate the global network state feature vector. A multi-objective optimization strategy unit is used to allocate corresponding multi-dimensional utility weight values ​​to the service flows of each user equipment based on the global network state feature vector and the node feature vector of the user equipment, and to construct a multi-objective optimization function based on the multi-dimensional utility weight values. The allocation of corresponding multi-dimensional utility weight values ​​to the service flows of each user equipment based on the global network state feature vector and the node feature vector of the user equipment includes: concatenating the node feature vector of each user equipment with the global network state feature vector to obtain an enhanced feature vector for each user equipment; inputting the enhanced feature vector into a multilayer perceptron, and outputting a multi-dimensional weight vector through the multilayer perceptron, the multi-dimensional weight vector containing channel quality weight components, service priority weight components, and historical resource acquisition fairness weight components of the user equipment; and normalizing the channel quality weight components, the service priority weight components, and the historical resource acquisition fairness weight components to obtain the multi-dimensional utility weight values ​​corresponding to the service flows of the user equipment. The step of constructing a multi-objective optimization function based on the multi-dimensional utility weight values ​​includes: constructing a spectral efficiency objective function based on the multi-dimensional utility weight values ​​and the channel state information; obtaining the total available spectrum resources of the region to be optimized, and determining a total spectrum resource constraint based on the total available spectrum resources; determining a resource guarantee constraint based on the quality of service level; determining an edge user group performance guarantee constraint based on the location identification information; and constructing the multi-objective optimization function by integrating the total spectrum resource constraint, the resource guarantee constraint, and the edge user group performance guarantee constraint based on the spectral efficiency objective function. The strategy determination unit is used to solve the multi-objective optimization function using the particle swarm optimization algorithm to obtain the spectrum resource allocation strategy for the region to be optimized, and to execute the spectrum resource allocation strategy.

Images