Knowledge graph driven interpretable satellite beam allocation method and device

Abstract

The invention discloses a knowledge graph driven interpretable satellite beam allocation method and equipment, and belongs to the technical field of satellite communication and 5G NR collaborative network resource scheduling. Scene symbolization is performed on user states, features are coded into scene feature character strings, and the scene feature character strings are coded into scene feature character strings; scene features, a beam allocation strategy and award performance thereof are written into a knowledge graph in a graph structure form, and multi-service and multi-scene-oriented adaptive satellite beam scheduling is realized by combining similar scene retrieval and model hot start, so that the interpretability of a strategy decision process is realized, and the self-adaptive satellite beam scheduling efficiency is improved. While the differentiated service quality requirements of multiple types of services of voice, video and games are met, the system throughput and the resource utilization rate are improved, the service fairness among users is improved, and the service time delay and the training convergence time are reduced, so that the overall transmission burden of a satellite and a ground network is reduced, and the service quality is improved. And finally, the comprehensive service quality of the user in the space-ground integrated convergence network environment is improved.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication and 5G NR collaborative network resource scheduling technology, specifically involving a knowledge graph-driven interpretable satellite beam allocation method and device. Background Technology

[0002] With the rapid development of low-Earth orbit satellite constellations and integrated space-ground networks, integrated space-ground communication is gradually becoming a key means to improve global coverage and service continuity. The simultaneous access of numerous satellites, ground base stations, and a massive number of heterogeneous terminals results in highly dynamic network topologies and strong fluctuations in service loads both spatially and temporally. Traditional beam allocation and resource scheduling strategies, relying on static planning or empirical rules, struggle to adapt to complex scenario changes in a timely manner, easily leading to overloaded beams and idle beams, resulting in low resource utilization, link congestion, and degraded user service quality. Existing satellite resource scheduling schemes typically focus on single or a few scenarios, optimizing throughput and coverage metrics through fixed weights or heuristic algorithms. However, in scenarios involving multiple services (such as voice, video, and gaming services with multiple quality-of-service (QoS) levels) and multi-cell collaboration, it is difficult to simultaneously consider multiple dimensions of metrics, including throughput, latency, and fairness among users. Furthermore, the lack of a systematic characterization of service models and scenario structures leads to a reliance on human experience in strategy design, resulting in insufficient generalization capabilities.

[0003] In recent years, deep reinforcement learning has been introduced into satellite beam allocation and wireless resource management. It can learn beam-user group matching strategies through interaction with the environment, thereby improving system throughput and automation to a certain extent. However, existing reinforcement learning solutions generally suffer from the following problems: First, most methods are trained on a fixed or idealized traffic scenario, lacking explicit modeling and differentiation of the "scenario." When traffic patterns or business composition change, training from scratch or long retraining is required, resulting in slow and unstable convergence. Second, traditional reinforcement learning models do not explicitly record "which strategy works best in which scenario." The large amount of interactive experience generated during training is difficult to structure and reuse across scenarios, leading to knowledge transfer issues and high computational and time costs. Most existing solutions rely only on numerical state vectors, lacking abstract descriptions of high-level semantic features such as business distribution, priority levels, latency sensitivity, and resource requirements. Operations personnel find it difficult to understand why a specific beam allocation strategy is chosen at a certain moment, especially unable to explain "which business model the current scenario belongs to" or "how the strategy has performed historically in similar scenarios" from the perspective of business model or scenario structure. This results in a "black box" characteristic for the strategy decision-making process, insufficient interpretability, and is also detrimental to subsequent network operation and maintenance optimization and audit tracing. Summary of the Invention

[0004] The purpose of this invention is to provide a knowledge graph-driven, interpretable satellite beam allocation method and device to solve the technical problems of poor experience reusability, interpretability and auditability of existing methods.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A knowledge graph-driven, interpretable satellite beam assignment method includes the following steps: Training task nodes, scene feature nodes, and model nodes are created sequentially in the graph database, and corresponding relationship edges are established to construct a knowledge graph. The training results of historical satellite beam allocation strategies are written into the knowledge graph to form a knowledge base. The training results of the satellite beam allocation strategies include training tasks, scene features, and model information. The similarity query is performed between the scene features of the current scene and the scene features already existing in the knowledge base. If the similarity exceeds a preset threshold, the corresponding scene in the knowledge base is directly loaded as the similar scene of the current scene. If the similarity is lower than the preset threshold, the strategy is retrained for the current scene. After the training is completed, the new training task, scene features and model information are added to the knowledge graph. Based on the training task nodes, scene feature nodes, and model nodes corresponding to the current scenario, generate explanatory information output for operation and maintenance personnel.

[0006] Furthermore, the training task node includes task identifier, traffic scenario, timestamp, final model path, final / best average reward, convergence round, number of training rounds, and reward type; The scene feature nodes include the total number of scenes, diversity, average priority level, average latency sensitivity level, average resource requirement level, business model distribution, and dominant business distribution. The model node includes the model file path, size, and creation time.

[0007] Furthermore, the scene features are obtained through symbolic satellite communication scene representation, and the steps include: Obtain current environmental status data; Based on the current environmental status data, the business activity and data proportion are classified according to different service quality categories of business types; Context features are constructed based on user status and classification results.

[0008] Furthermore, the complete environmental status data includes a user status matrix and an active user indicator vector. The user status matrix includes user identifier, access base station or satellite identifier, service type marker, user geographical location, geometric relationship with base station and satellite, amount of data to be transmitted in the queue, amount of historically transmitted data, current reachable link capacity, current available link capacity, and the most recent transmission result.

[0009] Furthermore, based on the complete environmental status data, the business activity and data share are classified according to different service quality categories and business types, including: For each business type, collect samples of business activity and business data percentages; Maintain a set of multi-quantile statistical structures for the business activity and business data proportion of each type of business; The quantile values ​​are updated based on the relative size of the new sample in the current business distribution; Based on the sample values ​​of each business and the updated quantile range, business activity and data share are divided into multiple levels.

[0010] Furthermore, the generation of the satellite beam allocation strategy includes: Output the satellite beam-user group allocation matrix based on the current environmental conditions; Users are divided into satellite access users, base station access users and unserved users according to the allocation matrix, and an actual allocation matrix is ​​constructed at the resource block dimension. Complete resource mapping and update user queue status while satisfying power constraints; A near-end policy optimization algorithm is used to jointly train the policy network and the value network to generate a beam allocation policy. The near-end policy optimization algorithm aims to maximize the cumulative discount reward during training and updating, and constructs an instant reward function.

[0011] Furthermore, the aforementioned near-end policy optimization algorithm is used to jointly train the policy network and the value network to generate a beam assignment policy, including: The policy network and value network share the front-end feature extraction layer. Through trajectory acquisition, advantage estimation, pruning constraint update and mean square error optimization, a stable and efficient allocation strategy is gradually obtained.

[0012] Secondly, the present invention provides a knowledge graph-driven interpretable satellite beam assignment system, comprising a construction module, a matching module, and an output module, wherein: The construction module is used to sequentially create training task nodes, scene feature nodes, and model nodes in the graph database, establish corresponding relationship edges, construct a knowledge graph, and write the training results of historical satellite beam allocation strategies into the knowledge graph to form a knowledge base. The training results of the satellite beam allocation strategies include training tasks, scene features, and model information. Matching module: Used to perform similarity queries between the scene features of the current scene and the scene features already existing in the knowledge base. If the similarity exceeds a preset threshold, the corresponding scene in the knowledge base is directly loaded as the similar scene of the current scene; if the similarity is lower than the preset threshold, the strategy is retrained for the current scene. After the training is completed, the new training task, scene features and model information are added to the knowledge graph. Output module: Used to generate explanatory information output for operation and maintenance personnel based on the training task nodes, scene feature nodes and model nodes corresponding to the current scene.

[0013] Thirdly, a terminal device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0014] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: This invention presents a knowledge graph-driven, interpretable satellite beam allocation method. By storing scene features, beam allocation strategies, and their reward performance in a graph structure within a knowledge graph, and through similar scene retrieval and model hot-start, a reinforcement learning algorithm is guided to generate beam allocation strategies oriented towards user groups. At each strategy selection, explanatory information based on scene features and historical model performance is output, achieving traceable and understandable satellite beam allocation decisions. The proposed allocation scheme can meet the differentiated needs of various services such as voice, video, and games while improving system throughput and resource utilization, enhancing service fairness among users, reducing service latency and training convergence time, thereby alleviating the overall transmission burden of satellite and terrestrial networks, improving the comprehensive service quality for users in a converged networking environment, and ensuring good interpretability and auditability of the satellite beam allocation process through explicit output of explanatory information.

[0016] The symbolization of scene features has the following advantages: Dimensionality reduction and comparability: Compressing complex continuous states into discrete features with fixed structures creates manageable and comparable "scene nodes," facilitating cross-time / cross-task alignment and comparison, and enabling graph storage and retrieval migration. Scene feature strings can serve as unique keys for scene nodes, supporting similar scene retrieval; upon a match, historical models can be directly loaded for a warm start, reducing training from scratch and random exploration, resulting in stronger interpretability and auditability. Symbolization provides "readable semantic labels" that can output explanatory information during decision-making, making it easier for operations and maintenance personnel to understand and track information. Attached Figure Description

[0017] Figure 1 This is a flowchart of a knowledge graph-driven interpretable satellite beam allocation method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the system architecture of a knowledge graph-driven interpretable satellite beam allocation method according to an embodiment of the present invention; Figure 3 This is a flowchart of the satellite beam allocation method based on knowledge graphs according to the present invention. Figure 4 This is a simulation experiment diagram of the satellite beam allocation method based on knowledge graphs according to the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0020] It should be noted that the terms "first" and "second" in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] The present invention will now be described in further detail with reference to the accompanying drawings: like Figure 1 As shown, a knowledge graph-driven interpretable satellite beam assignment method includes the following steps: Step 1: In the graph database, training task nodes, scene feature nodes, and model nodes are created sequentially, and corresponding relationship edges are established to construct a knowledge graph. The training results of historical satellite beam allocation strategies are written into the knowledge graph to form a knowledge base. The training results of the satellite beam allocation strategies include training tasks, scene features, and model information. The knowledge graph is defined as follows: It is organized around three types of information: "training task—scenario features—model". Training task nodes record task identifier, traffic scenario, timestamp, final model path, final / best average reward, convergence rounds, training rounds, and reward type. Scenario feature nodes use scenario feature strings as unique identifiers and record the statistical results of symbolic scenario trajectories during training, including the total number of scenarios, diversity, average priority level, average latency sensitivity level, average resource requirement level, business model distribution, and dominant business distribution. Model nodes record the model file path, size, and creation time. These nodes are connected through the relationships "training task → scenario features" and "training task → model", forming a searchable knowledge base. Symbolic scene information is collected and accumulated during training. After training, the information from the entire training process is statistically analyzed to generate scene features, including statistical information on pattern distribution, priority / latency / resource requirements, etc. In the graph database, training task nodes, scene feature nodes, and model nodes are created sequentially, and corresponding relationship edges are established. The training results of each satellite beam allocation strategy are written into the knowledge graph in the form of structured knowledge to form a traceable and reusable experience base; the training results of the satellite beam allocation strategy include at least the training task, scene features, and model information.

[0022] Scene feature nodes are obtained through symbolic satellite communication scenes, and the steps include: Obtain current environmental status data; Based on the current environmental status data, the business activity and data proportion are classified according to different service quality categories of business types; Context features are constructed based on user status and classification results.

[0023] Specifically, in each scheduling time slot, the current communication environment state is collected to obtain an environment state set. User state matrix Record user-related information, including: user identifier, access base station or satellite identifier, service type marker, user geographical location (latitude, longitude, altitude), geometric relationship with the base station and satellite (distance, elevation angle), amount of data to be transmitted in the queue, amount of historically transmitted data, current reachable link capacity, and the most recent transmission result. Active user indicator vector. Used to mark whether each user has a business request in the current time slot. A value of 1 indicates that the user is an active user, and a value of 0 indicates that the user has no business request at present. Based on the relative position and obstruction between the user and the satellite and ground base station, and combined with the wireless channel and modulation and coding model, the link loss, interference and noise factors are estimated to obtain the link capacity available to each user in the current time slot. The results are written into the corresponding field in the user state matrix to form a complete current environment state for subsequent decision-making. For different service quality categories of business types (such as voice, video, and games), we continuously track and classify business activity and data share. Based on the aforementioned business statistics, the symbolization generator constructs scene features σ_t. To facilitate subsequent knowledge graph storage and retrieval, multiple discretized fields are concatenated in a fixed order to form a symbolic representation of the scene, mainly including: Business distribution mode field: for example, "video business-dominated", "game business-dominated", "balanced across multiple businesses"; Dominant Business Field: Marks the currently dominant business type; Priority level field: Based on the QCI priority and activity ratio of different businesses, the overall priority of the scenario is discretized into multiple level intervals; Latency sensitivity field: Based on the ratio of average queuing latency to the service level protocol threshold, the scenario latency sensitivity is discretized into multiple levels; Resource demand level field: Based on the ratio of average queued data volume to link capacity, the resource stress level of the scenario is classified.

[0024] Based on complete environmental status data, business activity and data share are categorized according to different service quality categories and business types, including: For each business type, two types of sample data are continuously collected: business activity sample, which represents the proportion of users currently having this business queue to all active users; and business data proportion sample, which represents the proportion of the total number of bits in this business queue to the total number of bits in all business queues. Maintain a set of multi-quantile statistical structures for the business activity and business data proportion of each type of business; For each type of business, a set of multi-quantile statistical structures is maintained to characterize the distribution across different level intervals for "activity level" and "data share". This structure uses a small number of marker values ​​and location counters to approximate the distribution of business volume in low, medium, and high intervals online. Whenever a new sample (including business activity samples and business data proportions) is received, the label values ​​are automatically adjusted according to the relative size of the sample in the current business distribution, and the quantile values ​​are updated to achieve real-time tracking of the changing trends of business activity and data proportions. Based on the sample values ​​of each business and the updated quantile ranges mentioned above, business activity and data share are divided into multiple levels, such as: low, lower, higher, and high.

[0025] Ultimately, scene features are encoded into a structured string, serving as a unique key for scene features in the knowledge graph. This symbolization process compresses complex continuous states into searchable and comparable discrete features, providing a foundation for knowledge graph-based retrieval and transfer, while also providing structured scene semantic tags for subsequent interpretation modules. This allows the system to provide a readable description of the "current scene type and dominant business" along with the output strategy.

[0026] The process of obtaining the satellite beam allocation strategy includes: Based on the current environmental status Output a satellite beam-user group allocation matrix The rows and columns of the matrix correspond one-to-one with the number of satellites in the system and the number of available beams for each satellite. Each element represents "the user group identifier served by a specific beam of a satellite in the current time slot." The allocation matrix is ​​parsed according to the order of satellite index and beam index. When a user group is repeatedly allocated, the beam allocated first is retained. According to the allocation matrix All users are divided into three categories: satellite access users served by a satellite beam; base station access users served by a ground base station; and users who are not currently served in their time slot. Based on this, an actual allocation matrix is ​​constructed at the resource block level: for satellite access users, their corresponding resource block positions are marked as satellite occupied; for base station access users, their corresponding resource block positions are marked as base station occupied; and unserved users are marked as idle at their respective positions. During resource mapping, both the maximum power constraint of a single beam and the total power constraint of the entire system are considered simultaneously, and resource allocation is completed under the premise of meeting power limits. Subsequently, the link quality indicators of each user are calculated based on the pre-set channel model and interference model, and mapped to the actual achievable throughput. This updates the queue length and service queue status of each user, and updates the environmental state for the next time step.

[0027] Satellite beam allocation decision-making employs a near-end policy optimization algorithm to jointly train the policy network and the value network. Both share a front-end feature extraction layer, which typically consists of multiple fully connected layers. The back-ends output the beam allocation policy distribution and state value estimates, respectively, including: In a given scenario, the current strategy interacts with the environment to collect trajectory data including state, action, reward, and next state. The advantage estimation algorithm is used to evaluate the merits of the decision at each time step to obtain the preference direction required when updating the policy; In each iteration, the policy parameters are updated using an objective function with pruning constraints to limit the difference between the old and new policies and ensure the stability of the training process. The value network is updated based on the mean square error index, enabling it to more accurately estimate state values ​​and improve overall learning efficiency. Through the above training process, a satellite beam allocation strategy is gradually obtained.

[0028] During model training, the optimization objective is to maximize the cumulative reward of the discount. In each time slot, an instant reward is constructed based on current throughput performance, fairness of resource allocation among users, and multi-dimensional indicators of user service satisfaction. The optimization objective function is:

[0029] In the formula, π represents the policy, which is the rule or distribution of actions taken by the agent under different states. The goal is to find the optimal policy.

[0030] This represents the expectation under policy π, which is the average return of all possible trajectories (states, actions, and reward sequences) under that policy.

[0031] This represents the cumulative sum from time step t=0 to t=T (i.e., the reward for the entire round).

[0032] Discount factor, 0 < γ ≤ 1: When γ is close to 1, more emphasis is placed on future rewards; When γ is small (e.g., 0.9, 0.8), more emphasis is placed on immediate rewards.

[0033] The immediate reward received at time step t.

[0034] Step 2: Perform a similarity query between the scene features of the current scene and the scene features already existing in the knowledge base. If the similarity exceeds a preset threshold, directly load the corresponding scene in the knowledge base as a similar scene for the current scene; if the similarity is lower than the preset threshold, retrain the strategy for the current scene. After training is completed, the new training task, scene features and model information are added to the knowledge graph. like Figure 3As shown, when faced with new business scenarios, existing models are quickly retrieved from the knowledge graph by querying the similarity of scenario features. First, a symbolization generator is invoked to generate a feature representation of the current scenario; The query module uses the scene features as query conditions, prioritizes matching the associated scene features and model nodes in the knowledge graph, calculates the similarity between historical scenes and the current scene, and uses cosine similarity as the similarity measure. like Figure 2 As shown, when the similarity exceeds the preset threshold, it is considered to have hit a similar scene, skip the retraining process, directly load the corresponding model weights from the knowledge graph for inference evaluation, and record this model reuse event in the knowledge graph; When a suitable model is not found or the similarity is insufficient, the system enters the training branch to retrain the policy for the current scene. After training is completed, the new complete training results (new training tasks, scene features, and model information) are written back into the knowledge graph to enrich the system's knowledge accumulation.

[0035] Step 3: Based on the training task nodes, scene feature nodes, and model nodes corresponding to the current scene, generate explanatory information output for operation and maintenance personnel; Based on the matching results of the current scenario, the module generates explanatory information for operations and maintenance personnel from the feature nodes, the matched historical training task nodes, and their associated model nodes. This explanatory information includes at least: the current scenario's business distribution pattern, dominant business type, scenario priority level, latency sensitivity level, resource requirements level, and the training task identifier, training time, and average reward level of the selected model in similar scenarios. The explanatory module combines this information into structured or natural language form and outputs it through the operations and maintenance interface or logs, ensuring the interpretability and auditability of the satellite beam allocation decision-making process.

[0036] The invention will be further illustrated below with simulation experiments, and the results are as follows: Figure 4 As shown.

[0037] The hardware platform of this invention is: 13th Gen Intel Core i7-13700KF, 64GB RAM, NVIDIA GeForce RTX 3090; the software platform is: Ubuntu-22.04.1, Python 3.10.16.

[0038] The simulation experiment in this case study was run in a virtual environment, modeling the complete air-space-ground communication process. Figure 4 The training rewards for using a knowledge graph-based satellite beam assignment method are shown. Figure 4As can be seen, the reward curve of the satellite beam allocation method based on knowledge graphs is consistently higher than that of the satellite beam allocation method without knowledge graphs. This indicates that knowledge graphs play a certain role in improving the training process. In a new training session, the knowledge graph obtains the initial strategy from similar scenarios, so its training process is faster than starting training from scratch, achieving higher rewards in a shorter time. This improves system throughput and resource utilization, enhances service fairness among users, reduces service latency, alleviates the overall transmission burden of satellite and terrestrial networks, and improves the overall service quality for users in a converged networking environment.

[0039] This invention proposes a knowledge graph-driven, interpretable satellite beam allocation method. On one hand, by symbolizing user location, link quality, and the underlying state of service queues, scenario features capable of characterizing service distribution, priority, and resource requirements are constructed, abstracting complex operational scenarios into manageable and comparable "scenario nodes." On the other hand, during reinforcement learning training, scenario features, beam allocation strategies, and their reward performance are continuously written into the knowledge graph in a graph structure. Through similar scenario retrieval across training tasks and model hot-start, historical experience is structurally accumulated and transferred across scenarios, enabling the rapid provision of high-quality beam allocation strategies when facing new service models. This invention provides a novel beam allocation mechanism integrating "scenario knowledge management + reinforcement learning decision-making" for integrated space-ground networks. Technically, it achieves adaptive and interpretable satellite beam scheduling under multi-service and multi-scenario conditions, helping to improve system throughput and resource utilization, enhance service fairness among users, shorten training convergence time, and ultimately improve the overall service quality for users in a converged network environment.

[0040] In another embodiment of the present invention, a knowledge graph-driven interpretable satellite beam assignment system is provided, comprising a construction module, a matching module, and an output module, wherein: The construction module is used to sequentially create training task nodes, scene feature nodes, and model nodes in the graph database, establish corresponding relationship edges, construct a knowledge graph, and write the training results of historical satellite beam allocation strategies into the knowledge graph to form a knowledge base. The training results of the satellite beam allocation strategies include training tasks, scene features, and model information. Matching module: Used to perform similarity queries between the scene features of the current scene and the scene features already existing in the knowledge base. If the similarity exceeds a preset threshold, the corresponding scene in the knowledge base is directly loaded as the similar scene of the current scene; if the similarity is lower than the preset threshold, the strategy is retrained for the current scene. After the training is completed, the new training task, scene features and model information are added to the knowledge graph. Output module: Used to generate explanatory information output for operation and maintenance personnel based on the training task nodes, scene feature nodes and model nodes corresponding to the current scene.

[0041] Figure 2 The overall system architecture of this invention is demonstrated: the system operates on a satellite-terrestrial network, and a state acquisition and symbolization module, a reinforcement learning control module, and a knowledge graph construction and maintenance module are set up in the ground control center. The state acquisition module obtains information such as user state and computing resources from the network side and generates state vectors and scene signatures; the reinforcement learning control module combines reward design to output control decisions such as beam allocation, and passes training indicators, model information, etc. to the knowledge graph module for retrieval, matching similar scenes from the knowledge graph database to achieve model hot start, forming a closed loop of "perception-decision-update-storage".

[0042] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can be implemented in one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs) containing computer-usable program code. The form of a computer program product implemented on ROM (optical memory, etc.).

[0043] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0044] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0045] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or substitutions to the specific implementation of the invention, but these changes, modifications or substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A knowledge graph-driven, interpretable satellite beam assignment method, characterized in that, Includes the following steps: Training task nodes, scene feature nodes, and model nodes are created sequentially in the graph database, and corresponding relationship edges are established to construct a knowledge graph. The training results of historical satellite beam allocation strategies are written into the knowledge graph to form a knowledge base. The training results of the satellite beam allocation strategies include training tasks, scene features, and model information. The similarity query is performed between the scene features of the current scene and the scene features already existing in the knowledge base. If the similarity exceeds a preset threshold, the corresponding scene in the knowledge base is directly loaded as the similar scene of the current scene. If the similarity is lower than the preset threshold, the strategy will be retrained for the current scene. After the training is completed, the new training task, scene features and model information will be added to the knowledge graph. Based on the training task nodes, scene feature nodes, and model nodes corresponding to the current scenario, generate explanatory information output for operation and maintenance personnel.

2. The knowledge graph-driven interpretable satellite beam assignment method according to claim 1, characterized in that, The training task nodes include task identifier, traffic scenario, timestamp, final model path, final / best average reward, convergence round, number of training rounds, and reward type; The scene feature nodes include the total number of scenes, diversity, average priority level, average latency sensitivity level, average resource requirement level, business model distribution, and dominant business distribution. The model node includes the model file path, size, and creation time.

3. The knowledge graph-driven interpretable satellite beam assignment method according to claim 1, characterized in that, The scene features are obtained by symbolizing satellite communication scenes, and the steps include: Obtain current environmental status data; Based on the current environmental status data, the business activity and data proportion are classified according to different service quality categories of business types; Context features are constructed based on user status and classification results.

4. The knowledge graph-driven interpretable satellite beam assignment method according to claim 3, characterized in that, The complete environmental status data includes a user status matrix and an active user indicator vector. The user status matrix includes user identifier, access base station or satellite identifier, service type marker, user geographical location, geometric relationship with base station and satellite, amount of data to be transmitted in the queue, amount of historically transmitted data, current reachable link capacity, current available link capacity, and the most recent transmission result.

5. The knowledge graph-driven interpretable satellite beam assignment method according to claim 3, characterized in that, Based on the complete environmental status data, business activity and data share are categorized according to different service quality categories and business types, including: For each business type, collect samples of business activity and business data percentages; Maintain a set of multi-quantile statistical structures for the business activity and business data proportion of each type of business; The quantile values ​​are updated based on the relative size of the new sample in the current business distribution; Based on the sample values ​​of each business and the updated quantile range, business activity and data share are divided into multiple levels.

6. The knowledge graph-driven interpretable satellite beam assignment method according to claim 1, characterized in that, The generation of the satellite beam allocation strategy includes: Output the satellite beam-user group allocation matrix based on the current environmental conditions; Users are divided into satellite access users, base station access users and unserved users according to the allocation matrix, and an actual allocation matrix is ​​constructed at the resource block dimension. Complete resource mapping and update user queue status while satisfying power constraints; A near-end policy optimization algorithm is used to jointly train the policy network and the value network to generate a beam allocation policy. The near-end policy optimization algorithm aims to maximize the cumulative discount reward during training and updating, and constructs an instant reward function.

7. The knowledge graph-driven interpretable satellite beam assignment method according to claim 6, characterized in that, The near-end policy optimization algorithm is used to jointly train the policy network and the value network to generate a beam assignment policy, including: The policy network and value network share the front-end feature extraction layer. Through trajectory acquisition, advantage estimation, pruning constraint update and mean square error optimization, a stable and efficient allocation strategy is gradually obtained.

8. A knowledge graph-driven interpretable satellite beam assignment system, characterized in that, It includes a building module, a matching module, and an output module, where: The construction module is used to sequentially create training task nodes, scene feature nodes, and model nodes in the graph database, establish corresponding relationship edges, construct a knowledge graph, and write the training results of historical satellite beam allocation strategies into the knowledge graph to form a knowledge base. The training results of the satellite beam allocation strategies include training tasks, scene features, and model information. Matching module: Used to perform similarity queries between the scene features of the current scene and the scene features already existing in the knowledge base. If the similarity exceeds a preset threshold, the corresponding scene in the knowledge base is directly loaded as the similar scene of the current scene; if the similarity is lower than the preset threshold, the strategy is retrained for the current scene. After the training is completed, the new training task, scene features and model information are added to the knowledge graph. Output module: Used to generate explanatory information output for operation and maintenance personnel based on the training task nodes, scene feature nodes and model nodes corresponding to the current scene.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements as claimed in claim 1.

7. The steps of any of the methods described.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements as described in claim 1.

7. The steps of any of the methods described.

Images