A large language model agent method for satellite communication operation and control

Abstract

The present invention relates to a large language model agent method for satellite communication operation and control, belonging to the field of operation and maintenance control and network management of satellite communication systems. The method comprises: constructing an agent, the agent comprising a personalized long-term memory component, a graph retrieval component, a workflow planning component, and an intent correction component; adding the constructed agent to the large language model in the form of a plug-in; after the large language model reads user input, first using the intent correction component to correct the user's incorrect intent to obtain a corrected intent input, then using the personalized long-term memory component and the graph retrieval component to obtain enhanced retrieval data, and finally transmitting the corrected intent input and enhanced retrieval data to the workflow planning component to obtain operation and control parameters for distribution to satellites, terminals, or earth stations. The present invention can solve the problem that operation and maintenance control and network management in the field of satellite communications are currently mainly manual and have a low level of intelligence.

Description

Technical Field

[0001] The present invention relates to the fields of operation, maintenance, control and network management in satellite communications, and in particular to a large language model agent method for satellite communication operation and control. Background Art

[0002] Large language models (LLMs), as foundational models such as Deepseek, Qwen, and GPT, have demonstrated a powerful ability to solve practical problems through complex step-by-step reasoning. Specifically designed LLMs have begun to be integrated into ground communication networks, thanks to in-depth research on AI agents. However, current research in satellite communications has not kept pace with advances in LLM reasoning and agent capabilities. LLM research on satellite communications tasks is rare, and there is also little research on how to integrate agents with actual satellite systems.

[0003] In order for the LLM model to generate more accurate and reliable answers when solving complex real-world problems, a specially designed agent system is usually required to access external tools such as databases, private models, and search engines. AI agent systems generally consist of three modules: planning, memory, and tool use. In these modules, the LLM model plays a role in reasoning and decision-making. At the same time, some studies have been the first to explore general-purpose task agents: Manus, AutoGPT, and AutoGen are designed to handle a wide range of general tasks by breaking down user queries into actionable components. The essence of communication is to provide services, and the LLM model agent has brought about changes in the relationship between communication services and users. Agent systems based on the LLM model need to be designed based on industry problems and have already shown great potential in applications in finance, education, medical diagnosis, and other fields.

[0004] Specifically, there is currently no discussion or definition of how to integrate LLM models with satellite communication networks, and no design of relevant agent architectures to solve practical problems in the field of satellite communication, such as satellite network planning, network management, traffic scheduling, and resource allocation. There is also no design solution for whether LLM models can be used to derive solutions to satellite communication operation and control systems through step-by-step reasoning, nor is there a design solution for generating reliable network parameters through agent calls. Summary of the Invention

[0005] To address these issues, this paper proposes a large language model agent approach for satellite communication operations and control. This paper pioneers an AI agent architecture suitable for satellite communication operations, maintenance, and network management scenarios. This architecture can serve as the operations and control center within satellite communication systems, enabling one-click, natural language-based, full-process control of satellites, terminals, and earth stations.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A large language model agent method for satellite communication operation and control includes the following steps:

[0008] Step 1: Build an agent. The agent includes a personalized long-term memory component, a graph retrieval component, a workflow planning component, and an intent correction component. The personalized long-term memory component records the preferences and historical behaviors of different types of users, and the workflow planning component stores the workflows of different tasks.

[0009] Step 2: Add the built agent to the large language model as a plug-in;

[0010] In step 3, after the large language model reads the user input, it first uses the intent correction component to correct the user's incorrect intent to obtain the corrected intent input. Then, after being processed jointly by the personalized long-term memory component and the graph retrieval component, enhanced retrieval data is obtained. Finally, the corrected intent input and enhanced retrieval data are passed to the workflow planning component to obtain the operation and control parameters sent to the satellite, terminal, or earth station.

[0011] Furthermore, the large language model is a long-context large language model using 128k Tokens.

[0012] Furthermore, the preferences of different types of users recorded in the personalized long-term memory component are pre-organized fixed data, and the historical behaviors of different types of users recorded in the personalized long-term memory component are historical behavior memories of different users.

[0013] Furthermore, the intent correction component is used to verify and correct the user's input needs. After the user's intention needs are input, the intent parameters are obtained through reasoning and parsing of the large language model, and then the intent correction component corrects the intent parameters after parsing of the large language model to make them correct values.

[0014] The intent correction component includes the intent translation module, the named entity recognition module, the template matching module, and the conflict detection module. It works as follows:

[0015] Convert the intent text into predefined standard parameters through the intent translation module;

[0016] The named entity recognition module replaces the fuzzy entities used by communication users with predefined standard entities;

[0017] The template matching module is used to align parameter templates. The Agent call parameter names are aligned with the actual delivered parameters through intent template matching.

[0018] The conflict detection module verifies resource occupancy to avoid conflicts.

[0019] Furthermore, the graph retrieval component constructs a knowledge vector library through satellite communication documents and implements offline data retrieval using the Graph-RAG retrieval method based on entity relationship information in the satellite communication system;

[0020] Said satellite communication documents include satellite system design documents, user manuals, and satellite communication teaching materials;

[0021] In the graph construction phase, a large language model is used to extract entities and relationships from structured text, and the connections between entities and relationships are displayed in the form of a graph. During retrieval, entities and relationships are modeled jointly as units.

[0022] In the inference phase, vectors are generated based on the user's query, relevant entities are retrieved, entities and corresponding relationships are indexed, and the knowledge blocks corresponding to the corresponding entities are indexed based on the relationships, thereby enhancing the performance of the large language model.

[0023] Furthermore, the workflow planning component takes a user intent as input, and decomposes the predefined user intent into satellite operation and control workflows through predefined connection relationships. For each workflow, different models are called in multiple steps to solve different problems and finally obtain output results; the models include traffic prediction model, resource allocation model, and anti-interference model.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. This invention implements an AI Agent architecture suitable for satellite communication operation and maintenance control and network management scenarios, which can solve the current problem of manual operation and maintenance in the field of satellite communication and network management, which has a low degree of intelligence.

[0026] 2. This invention introduces personalized long-term memory components, graph retrieval components, workflow planning components, and intent correction components into the AI ​​Agent architecture, significantly improving the accuracy and reliability of Agent calls.

[0027] 3. The present invention can serve as the operation and control center in a satellite communication system, and can achieve one-click natural language-based full-process management and control of satellites, terminals, and earth stations. DETAILED DESCRIPTION

[0028] The technical solution of the present invention is described in detail below with reference to specific embodiments.

[0029] A large language model agent method for satellite communication operation and control includes the following steps:

[0030] Step 1: Build an agent. The agent includes a personalized long-term memory component, a graph retrieval component, a workflow planning component, and an intent correction component. The personalized long-term memory component records the preferences and historical behaviors of different types of users, and the workflow planning component stores the workflows of different tasks.

[0031] Step 2: Add the built agent to the large language model as a plug-in;

[0032] In step 3, after the large language model reads the user input, it first uses the intent correction component to correct the user's incorrect intent to obtain the corrected intent input. Then, it uses the personalized long-term memory component and the graph retrieval component to obtain enhanced retrieval data. Finally, the corrected intent input and enhanced retrieval data are passed to the workflow planning component to obtain the operation and control parameters sent to the satellite, terminal, or earth station.

[0033] This method builds the following components for the agent of the large language model:

[0034] The personalized long-term memory component records the preferences and historical behaviors of different types of users, and provides contextual prompts to the model during each model inference.

[0035] This component records the preferences and historical behavior of each user type, providing contextual information to the LLM model during each inference. The LLM model uses a long-context LLM model with 128k tokens. The personalized long-term memory component sets different preferences for different users because satellite communications involve a wide variety of users with varying behavioral habits, such as those at sea, in vehicles, and in aircraft. Furthermore, the personalized long-term memory component stores a memory of each user's previous behavior. For example, maritime users typically request a fixed number of satellites and prefer Ka-band satellites. Because the preferences and historical behavior of different satellite users are relatively fixed, the personalized settings of this component enable the LLM model to distinguish between different types of users during inference, providing valuable insights to aid analysis and decision-making.

[0036] The graph retrieval component builds a knowledge vector library through satellite communication documents, encodes the entity relationship information in the satellite communication system based on the graph, and realizes offline data retrieval through this component.

[0037] This method uses the Graph-RAG retrieval method to construct a graph retrieval component. By integrating offline data retrieval, the graph retrieval component can limitedly expand the capabilities of satellite communication agents. Graphs can encode heterogeneous and relational information in satellite communication systems, making them well-suited for real-world applications. Furthermore, satellite, beam, and terminal names typically do not appear in the LLM pre-training phase. The introduction of this component enables the satellite communication agent to identify named entities that were not present during the LLM pre-training phase. Specifically, this method constructs a graph retrieval component using accumulated satellite communication documentation. Satellite communication documentation includes satellite system design documents, user manuals, and satellite communication textbooks. During the graph construction phase, LLM extracts entities and relationships from structured text, displaying the connections between entities and relationships in the form of a graph. This allows entities and relationships to be jointly modeled as units during retrieval. During the inference phase, vectors are generated based on the user's query, and relevant entities are retrieved. This is used to index the entities and their corresponding relationships. The knowledge blocks corresponding to the corresponding entities are then indexed based on the relationships, enhancing the performance of the LLM.

[0038] The workflow planning component takes a user intent input, decomposes it into satellite operation and control workflows through predefined connection relationships, and calls the corresponding model execution in steps to obtain the output results.

[0039] The workflow planning component, a key design element for addressing complex satellite communication user intent, forms a fixed workflow to handle complex operations and maintenance management tasks. This component calls different models through predefined connections, breaking down predefined user intent into workflows. These models are then called in multiple steps to address different issues. These models include traffic prediction, resource allocation, and anti-interference models for decision support. The workflow planning component not only provides templates for LLM tool calls but also includes constraints for system-level agent calls, thereby improving the accuracy of multi-step calls for complex satellite communication tasks.

[0040] The intent correction component verifies and corrects user demand input, including intent translation, named entity recognition, template matching, conflict detection and other steps, improving the accuracy of intent parsing.

[0041] The intent correction component verifies and corrects user input. After a user enters an intent, LLM reasoning and analysis are used to derive intent parameters. This component then corrects these LLM-analyzed intent parameters to the correct value, effectively improving intent analysis accuracy and simplifying the traditional intent management process.

[0042] The intent correction component includes steps such as intent translation, template matching, and conflict detection. The intent translation step converts the intent into parameters understandable by the satellite communication agent. For example, in the query "Query beam coverage in the Hawaii area," "Hawaii" is translated into the specific latitude and longitude [20N, -157W]. The named entity recognition step replaces the fuzzy entities used by communication users with entities recognizable to the agent, such as using "TT03" instead of "Tiantong-1 Satellite 03." Matching and correction through named entity recognition significantly improves the robustness and usability of the actual system. The template matching step aligns the parameter templates, aligning the agent call parameter names generated by the model with the actual parameters issued through intent template matching, ensuring the correct number of parameters and keywords. The conflict detection step verifies resource conflicts. For example, when applying for a satellite beam frequency resource, conflict detection checks whether the frequency is already occupied to avoid conflicts.

[0043] Among them, the fuzzy entity is obtained after the user's initial input intention is translated.

[0044] When implementing this method, the satellite communication operation and control system Agent is first initialized, the system prompt word is configured, the tool list registration is completed, the multi-scenario Agent architecture is defined, and the basic LLM model is initialized.

[0045] Then, the four aforementioned components are introduced into the satellite communication operation and control system agent. Each time a user's intent is input, it is inferred by the basic LLM model and processed by the personalized long-term memory component, graph retrieval component, multi-agent workflow planning component, and intent correction component to obtain the output result.

[0046] Based on the generated parameters in the output results, the Agent calls the tool to send the generated parameters to the corresponding satellite operation and control and network management systems to complete the operation and control of satellites, terminals and earth stations.

[0047] Here's a more concrete example:

[0048] A large language model agent method for satellite communication operation and control uses a general large language model (LLM) and an agent to achieve one-click network construction for satellite communication. The method specifically includes the following steps:

[0049] Step 1: Initialize the satellite communication operation and control system Agent, configure the system prompt word, complete the tool list registration, define the multi-scenario Agent architecture, and initialize the basic LLM model. Specifically:

[0050] Initialize the satellite communication operation and control system Agent, configure the system prompt, complete the tool list registration, define the multi-scenario Agent architecture, and initialize the basic LLM model. Among them, a typical configuration of the system prompt is "You are an expert in the field of satellite communication; after the user enters the requirements, you need to: first, decide which registered tool to call; then, generate the corresponding parameters according to the input requirements. Required parameters must be generated. If not, you need to inquire again; finally, call the corresponding tool and perform a function call." Tool list registration is to register the available tools in the field of satellite communication operation and control systems into the Agent framework. Typical tools included in the tool list include: mission planning tools, beam adjustment tools, resource query tools, resource allocation tools, resource scheduling tools, instruction injection tools, etc. Initialize the LLM model, including setting the temperature parameter to 0.3 and the core sampling probability parameter to 0.7.

[0051] Step 2: Build a personalized long-term memory component. This component records the preferences and historical behaviors of different types of users in a personalized manner, and provides contextual prompts to the model during each model inference.

[0052] This component records the preferences and historical behavior of each user type, providing contextual information to the LLM model during each inference. The LLM model uses a long-context LLM model with 128k tokens. The personalized long-term memory component sets different preferences for different users. This is because satellite communications involve a wide variety of users with varying behavioral habits, such as those at sea, in vehicles, and in aircraft. Furthermore, the personalized long-term memory component stores a memory of each user's previous behavior. For example, maritime users typically request a fixed number of satellites and prefer Ka-band satellites. Because the preferences and historical behavior of different satellite users are relatively fixed, the personalized settings of this component enable LLM inference to distinguish between different user types, providing valuable insights to aid analysis and decision-making.

[0053] Step 3: Build a graph retrieval component, construct a knowledge vector library through satellite communication documents, encode the entity relationship information in the satellite communication system based on the graph, and implement offline data retrieval through this component.

[0054] This method uses the Graph-RAG retrieval method to construct a graph retrieval component. By integrating offline data retrieval, the graph retrieval component can limitedly expand the capabilities of satellite communication agents. Graphs can encode heterogeneous and relational information in satellite communication systems, making them well-suited for real-world applications. Furthermore, satellite, beam, and terminal names typically do not appear in the LLM pre-training phase. The introduction of this component enables the satellite communication agent to identify named entities that were not present during the LLM pre-training phase. Specifically, this method constructs a graph retrieval component using accumulated satellite communication documentation. Satellite communication documentation includes satellite system design documents, user manuals, satellite communication textbooks, and user manuals. During the graph construction phase, LLM extracts entities and relationships from structured text, displaying the connections between entities and relationships in the form of a graph. This allows entities and relationships to be jointly modeled as units during retrieval. During the inference phase, vectors are generated based on the user's query, and relevant entities are retrieved. This is used to index the entities and their corresponding relationships. The knowledge blocks corresponding to the corresponding entities are then indexed based on the relationships, enhancing the performance of the LLM.

[0055] Step 4: Build a multi-agent workflow planning component. This component decomposes a user intention input into a satellite operation and control workflow through predefined connection relationships, and calls the corresponding model execution step by step to obtain the output results.

[0056] This workflow planning component serves as a key design for addressing complex satellite communication user intent and forms a fixed workflow to handle complex operations and maintenance management tasks. This component calls different models through predefined connections, breaking down predefined user intent into workflows. These models are then called in multiple steps to address different issues. These models include traffic prediction, resource allocation, and anti-interference models for decision support. The workflow planning component not only provides templates for LLM tool calls but also includes constraints for system-level agent calls, thereby improving the accuracy of multi-step calls for complex satellite communication tasks.

[0057] Step 5: Build an intent correction component, which verifies and corrects user demand input, including intent translation, named entity recognition, template matching, conflict detection, and other steps, improving the accuracy of intent parsing.

[0058] The intent correction component verifies and corrects user input. After a user enters an intent, LLM reasoning and analysis are used to derive intent parameters. This component then corrects these LLM-analyzed intent parameters to the correct value, effectively improving intent analysis accuracy and simplifying the traditional intent management process.

[0059] The intent correction component includes steps such as intent translation, template matching, and conflict detection. The intent translation step converts the intent into parameters understandable by the satellite communication agent. For example, in the query "Query beam coverage in the Hawaii area," "Hawaii" is translated into the specific latitude and longitude [20N, -157W]. The named entity recognition step replaces the fuzzy entities used by communication users with entities recognizable to the agent, such as using "TT03" instead of "Tiantong-1 Satellite 03." Matching and correction through named entity recognition significantly improves the robustness and usability of the actual system. The intent template matching step aligns the parameter templates, aligning the agent call parameter names generated by the model with the actual parameters issued through intent template matching, ensuring the correct number of parameters and keywords. The conflict detection step verifies resource conflicts. For example, when applying for a satellite beam frequency resource, conflict detection checks whether the frequency is already occupied to avoid conflicts.

[0060] Step 6: Introduce the above four components into the satellite communication operation and control system agent. Every time a user's intention is input, the basic LLM model is used to infer it. After being processed by the personalized long-term memory component, graph retrieval component, workflow planning component, and intention correction component, the output result is obtained.

[0061] Step 7: Based on the generated parameters in the output results, the Agent calls the tool to send the generated parameters to the corresponding satellite operation and control and network management systems to complete the operation and control of the satellite, terminal and earth station.

[0062] In view of the fact that network management in the current satellite communication field relies more on the demand input of operation and maintenance personnel, and the need to further combine the intelligent satellite communication with the reasoning ability of the LLM model, the present invention proposes a large language model agent method for satellite communication operation and control. In the field of satellite communication operation and maintenance management, the present invention implements a satellite communication network construction method based on the large language model, realizes a full process closed loop from operation and maintenance requirements to the generation and issuance of actual parameters for network construction, achieves the goal of fast and agile response network construction, and shortens the traditional artificial satellite communication network construction time from minutes to seconds.

[0063] This paper proposes a large language model agent-based approach for satellite communication operations and control. This approach addresses the current manual nature of operations and maintenance control and network management in satellite communications, which often lacks intelligence. This paper pioneers an AI agent architecture suitable for these scenarios, incorporating personalized long-term memory, graph retrieval, multi-agent workflow planning, and intent correction components. This significantly improves the accuracy and reliability of agent invocations. This architecture can serve as the operations and control center in satellite communication systems, enabling one-click, natural language-based, full-process control of satellites, terminals, and earth stations.

[0064] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0065] The above-described embodiments merely represent several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A large language model agent method for satellite communication operation and control, characterized in that: The following steps are involved: Step 1: Build an agent. The agent includes a personalized long-term memory component, a graph retrieval component, a workflow planning component, and an intent correction component. The personalized long-term memory component records the preferences and historical behaviors of different types of users, and the workflow planning component stores the workflows of different tasks. The graph retrieval component constructs a knowledge vector library through satellite communication documents and uses the Graph-RAG retrieval method based on the entity relationship information in the satellite communication system to achieve offline data retrieval; Said satellite communication documents include satellite system design documents, user manuals, satellite communication teaching materials, and user manuals; In the graph construction phase, a large language model is used to extract entities and relationships from structured text, and the connections between entities and relationships are displayed in the form of a graph. During retrieval, entities and relationships are modeled jointly as units. In the inference phase, vectors are generated based on the user's query, relevant entities are retrieved, entities and their corresponding relationships are indexed, and knowledge blocks corresponding to the corresponding entities are indexed based on the relationships, thus enhancing the performance of large language models. Step 2: Add the built agent to the large language model as a plug-in; In step 3, after the large language model reads the user input, it first uses the intent correction component to correct the user's incorrect intent to obtain the corrected intent input. Then, after being processed jointly by the personalized long-term memory component and the graph retrieval component, enhanced retrieval data is obtained. Finally, the corrected intent input and enhanced retrieval data are passed to the workflow planning component to obtain the operation and control parameters sent to the satellite, terminal, or earth station.

2. A large language model agent method for satellite communication operation and control according to claim 1, characterized in that: The large language model is a long-context large language model using 128k tokens.

3. A large language model agent method for satellite communication operation and control according to claim 1, characterized in that: The preferences of different types of users recorded in the personalized long-term memory component are pre-organized fixed data, and the historical behaviors of different types of users recorded in the personalized long-term memory component are the historical behavior memories of different users.

4. A large language model agent method for satellite communication operation and control according to claim 1, characterized in that: The intent correction component is used to verify and correct the user's input needs. After the user's intention needs are input, the intent parameters are obtained through inference and analysis of the large language model. The intent correction component then corrects the intent parameters after the large language model analysis to make them correct values. The intent correction component includes the intent translation module, the named entity recognition module, the template matching module, and the conflict detection module. It works as follows: Convert the intent text into predefined standard parameters through the intent translation module; The named entity recognition module replaces the fuzzy entities used by communication users with predefined standard entities; The template matching module is used to align parameter templates. The Agent call parameter names are aligned with the actual delivered parameters through intent template matching. The conflict detection module verifies resource occupancy to avoid conflicts.

5. The large language model agent method for satellite communication operation and control according to claim 1, characterized in that: The workflow planning component takes a user intent as input and decomposes the predefined user intent into satellite operation and control workflows through predefined connection relationships. For each workflow, different models are called in multiple steps to solve different problems and finally obtain the output results; the models include traffic prediction model, resource allocation model, and anti-interference model.