A Data Retrieval Enhancement Generation Method and Apparatus for Satellite Communications Based on Large Models

Abstract

This invention relates to the fields of artificial intelligence and natural language processing, and particularly to a method and apparatus for enhancing data retrieval and generation in the satellite communication field based on a large model. The process includes: data cleaning and metadata annotation, two-layer text clustering construction, two-layer knowledge graph construction, standardized user semantic input, knowledge graph retrieval, and answer optimization. The proposed two-layer semantic clustering structure and two-layer knowledge graph construction mechanism can effectively represent multi-granularity semantic units in the satellite communication field, supporting cross-document and cross-standard knowledge organization and retrieval. The introduction of a large language model to complete and standardize entities and relationships results in a higher degree of automation in the knowledge graph construction process and better relationship extraction quality. This invention enables intelligent parsing, accurate retrieval, and high-quality answer generation for complex natural language queries in the satellite communication field, possessing significant semantic understanding, reasoning, and closed-loop optimization capabilities.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, and in particular to a data retrieval enhancement generation method and apparatus based on a large model in the field of satellite communications. Background Technology

[0002] In the field of modern satellite communications, accurate data retrieval and processing are fundamental to supporting key tasks such as system design, protocol development, fault diagnosis, and technological evolution. With the increasing complexity of satellite communication systems and the surge in related literature, standards, and test data, accurately extracting the required information from massive, multi-source, and heterogeneous satellite communication data has become a critical problem that engineers urgently need to solve. Currently, technical documents in the satellite communication field typically include 3GPP-NTN standard documents, satellite communication network protocols, satellite communication system specifications, system architecture diagrams, and experimental data reports. These documents are complex, diverse in expression, and possess high semantic depth, making it difficult for traditional information retrieval methods to meet the needs of in-depth, semantic queries.

[0003] Traditional keyword matching methods fail to understand the semantics of user queries, performing only simple matching based on keywords, resulting in low accuracy of search results. This is especially true when dealing with specialized satellite communication terminology and complex sentences, where search results are often unsatisfactory. Most existing information retrieval systems lack deep semantic understanding of natural language, failing to effectively interpret the true intent of user queries. Their ability to handle synonyms, polysemous words, and contextual relationships is weak, leading to poor relevance of search results. For example, "satellite network architecture" and "satellite communication network structure" are closely related in the field of satellite communication, but traditional search methods may fail to connect the two, thus missing important relevant information.

[0004] With the continuous growth of satellite communication data, traditional information retrieval methods have proven inefficient in processing large-scale datasets and unable to quickly respond to user queries. Existing technologies experience a significant decrease in response speed when processing millions or even hundreds of millions of data points, failing to meet the demands of real-time satellite monitoring and rapid decision-making. Current retrieval methods generally lack recommendation functionality, failing to provide customized search results based on users' historical queries and preferences, resulting in a poor user experience. Satellite communication information retrieval needs results tailored to the specific needs of engineers or researchers. For example, engineers might be interested in fault diagnosis methods for a particular satellite system, while researchers might be more concerned with the latest research advancements in satellite communications. Current technologies cannot effectively recommend information based on users' historical behavior and preferences, forcing users to spend considerable time filtering information.

[0005] Satellite communication-related information is distributed across multiple databases, version repositories, and document systems, resulting in information fragmentation, heterogeneous formats, and inconsistent standards, creating severe information silos. For example, technical documents from different manufacturers and research institutions often employ their own custom structures and terminology standards, making it difficult for existing technologies to uniformly parse and integrate them, thus affecting the retrieval system's ability to fuse cross-source data. Meanwhile, while Large Language Models (LLMs) possess some language understanding and generation capabilities, they are limited by token length settings and cannot process complete satellite communication system design documents or multi-chapter specifications in one go. Furthermore, their responses, lacking structured knowledge support, are prone to generating "illusions"—generating logically correct but technically inaccurate content, or even fabricating non-existent data or processes, impacting the reliability of engineering applications.

[0006] In Large Language Models (LLM), token segmentation is a key step in splitting text into processable units (Tokens). Different segmentation methods have their own characteristics. The following is an analysis of the currently popular token segmentation methods and their advantages and disadvantages: (1) Lexicon-based tokenization: This method matches words in the text with a predefined dictionary and segments them according to rules (such as forward maximum matching and reverse maximum matching). This method does not require training and directly segments words quickly based on the dictionary, making it suitable for lightweight tasks; the segmentation rules are clear, the results are easy to understand, and it is easy to adjust manually; it supports multiple languages, and only needs to update the dictionary to adapt to new words. However, it cannot solve the problem of "one word with multiple meanings" or out-of-vocabulary (OOV) words; it does not consider the contextual semantics, and the segmentation results may not conform to the actual meaning; it requires manual updating of the dictionary, making it difficult to process new words in the domain in real time. (2) Statistical tokenization: This method trains the model through a corpus and learns the probability or sequence pattern of word occurrence. It can automatically identify out-of-vocabulary words through large-scale data training; it uses contextual probability to reduce segmentation errors; it adapts to the dynamic changes of new words and reduces manual intervention. However, its training requires a large amount of labeled data and the model inference speed is slow; models such as HMM are limited by the Markov assumption and it is difficult to capture the semantic association in long texts; the model parameters are complex and the word segmentation decision process is not intuitive. (3) Subword Tokenization, which splits words into subword units to balance the size of the vocabulary and semantic integrity. New words can be processed by combining subwords; the vocabulary of subword tokenization is smaller, reducing the storage and computational overhead of the model; the subword unit has both word semantics and character flexibility. However, the complex word segmentation of this method may destroy the semantic integrity; the subword models trained on corpora from different domains are very different and need to be adjusted accordingly. (4) Byte-level Tokenization, which segments text by bytes or characters (e.g., each Chinese character is a token, and English is segmented by letters). Any text can be segmented into byte units without maintaining a vocabulary; it is applicable to all languages ​​and does not need to be optimized for specific languages. However, its segmentation granularity is too fine and lacks word-level semantic information; the number of tokens surges after long text segmentation, increasing the computational load of the model. (5) Dynamic word segmentation based on Transformer, combined with pre-trained language models to dynamically adjust the word segmentation strategy and optimize token segmentation according to context. Some models can dynamically adjust the word segmentation granularity; the word segmenter and the pre-trained model are optimized together to improve the overall performance. However, this method has poor versatility, and the word segmenter of a specific model is difficult to transfer directly to other tasks or domains; the word segmentation logic is bound to the model parameters, making debugging and interpretation difficult.

[0007] Therefore, how to provide a method and apparatus for enhancing the generation of satellite communication data based on a large language model, which can improve the semantic understanding and expression of knowledge in the satellite communication field, and realize unified modeling, semantic reasoning and efficient question answering of large-scale, heterogeneous satellite communication data, is a key problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to provide a data retrieval enhancement generation method for the satellite communication field based on a large model. By constructing a two-layer semantic clustering structure and a two-layer knowledge graph, the generation capability of the large model and the reasoning capability of the graph are integrated to achieve intelligent parsing, accurate retrieval, and high-quality answer generation for complex natural language queries in the satellite communication field. It has significant semantic understanding capability, reasoning capability, and closed-loop optimization capability.

[0009] The technical solution adopted in this invention is as follows:

[0010] An enhanced data retrieval and generation method for satellite communications based on a large model includes the following steps:

[0011] S1: Perform structured cleaning on multi-source heterogeneous data in the field of satellite communication, forming multiple text blocks and annotating metadata; among them, multi-source heterogeneous data in the field of satellite communication includes the collection of 3GPP protocol documents, satellite network architecture white papers and inter-satellite link technology reports;

[0012] S2: Perform document-level clustering on the text blocks to generate document topic clusters, and perform fine-grained secondary clustering within the document topic clusters to form fine-grained semantic clusters, resulting in a two-layer clustering result;

[0013] S3: Based on the results of two-layer clustering and the cleaned block data, a two-layer knowledge graph is constructed. Atomic nodes correspond to text blocks, and super nodes correspond to fine-grained semantic clusters. Entity relationships are extracted using an offline large model to supplement semantic associations, forming a two-layer knowledge graph that supports domain knowledge reasoning.

[0014] S4: Receives natural language questions input by the user, performs word segmentation, stop word removal, entity recognition, and semantic standardization processing to generate a standardized semantic representation;

[0015] S5: Based on the entity information of the user's question, perform multi-dimensional retrieval in the two-layer knowledge graph, extract relevant entity nodes and their associated paths, construct a knowledge subgraph and extract relevant atomic node text;

[0016] S6: Combining the knowledge subgraph and related atomic node text, an initial answer is generated using a large language model, and language quality checks, professional terminology consistency verification, and version control are performed to form the final optimized answer.

[0017] Furthermore, S1 specifically includes the following sub-steps:

[0018] S11: Collect diverse documents from authoritative data sources in the field of satellite communications, including 3GPP protocol documents, satellite network architecture white papers, and inter-satellite link technology reports. The collection process adopts distributed crawling and API call methods. The diverse documents collected are uniformly converted into a unified plain text format, including removing format noise and preliminary identification and chapter segmentation of chart information.

[0019] S12: For texts exceeding the set length, a segmentation method based on semantic boundary detection is adopted to split the document into multiple semantically independent text blocks, each of which contains a complete semantic unit;

[0020] S13: Automatically extract and label metadata based on text content, including the document ID, chapter title, key term tags, and publication time, to form a metadata set.

[0021] Furthermore, S2 specifically includes the following sub-steps:

[0022] S21: Divide each text block High-dimensional vector representations are generated using a pre-trained large language model. ;

[0023] S22: Use hierarchical clustering algorithm to analyze all text vectors Perform document-level clustering to divide the document into several document topic clusters:

[0024]

[0025] Intra-cluster text similarity satisfies:

[0026]

[0027] In the formula, The set similarity threshold is 1≤j≤m, where m is the number of document topic clusters;

[0028] S23: In each document subject cluster Internally, the K-Means algorithm is used to subdivide text vectors into semantically consistent fine-grained semantic clusters:

[0029]

[0030] Fine-grained cluster intra-vector similarity satisfies:

[0031]

[0032] In the formula, 1≤ ≤ , This represents the number of fine-grained semantic clusters.

[0033] S24: The silhouette coefficient is used to evaluate the clustering results. The calculation formula is as follows:

[0034]

[0035] In the formula, 𝑎 represents the average distance within a cluster, and 𝑏 represents the minimum average distance between clusters;

[0036] S25: According to Adjust the clustering parameters and return to S23 for iteration until the similarity requirement is met.

[0037] Furthermore, S3 specifically includes the following sub-steps:

[0038] S31: Divide the text into chunks Each text block in Transformed into atomic nodes in the spectrum and attach metadata information The smallest semantic unit of the graph includes the document identifier, the location of the chunk, the corresponding topic cluster number, the timestamp, and the version number.

[0039] S32: Fine-grained semantic clusters processed by the two-layer clustering algorithm Merging sets of atomic nodes to construct supernodes Each supernode corresponds to a semantic topic or concept class;

[0040] S33: Extract domain entities based on atomic node text. It automatically generates natural language semantic relationships between entities using a local open-source large model and constructs connection edges between entities;

[0041] S34: Add node and edge types to the graph structure to construct a complete two-layer knowledge graph. ,in A set of nodes, including a set of atomic nodes. Supernode set and entity node set ; It is a set of edges, including edges representing semantic relationships between entities. Structural edges of atomic nodes and their supernodes And cross-domain semantic join edges used to represent cross-document references, number jumps, and chronological order. .

[0042] Furthermore, S4 specifically includes the following sub-steps:

[0043] S41: Receive natural language questions input by the user, and perform word segmentation on the natural language question text using a method that combines a domain dictionary and a statistical model, so that the segmented word sequence satisfies the requirement of maximizing the overall segmentation probability;

[0044] S42: Remove meaningless words from the word segmentation results and dynamically adjust them in conjunction with the domain stop word list;

[0045] S43: Combine local open-source large model with rule template matching to extract key named entities from the query, including technical terms, communication protocol names and system component names;

[0046] S44: Map the extracted entity information to a standard domain vocabulary or ontology to generate a normalized semantic representation.

[0047] Furthermore, S5 specifically includes the following sub-steps:

[0048] S51: Based on the named entities identified in the user question, search for the corresponding entity nodes in the two-layer knowledge graph to obtain a set of atomic nodes that are directly related to the semantics of the user question; where the entity types include protocol names, technical components, and parameter items;

[0049] S52: Statistically determine the affiliation of the hit entity nodes in the fine-grained semantic clusters, identify the semantic clusters in which the entities are most concentrated, and represent the main semantic scope of the current problem;

[0050] S53: Within the identified semantic clusters, collect other entity nodes that share edges with the hit entity, and combine them with the relational paths between entities in the graph, including inheritance, application, and interoperability, to expand and obtain semantically related atomic nodes.

[0051] S54: Aggregate related atomic nodes and upstream and downstream relationship paths to construct a compact and semantically related knowledge subgraph;

[0052] S55: Perform unified extraction, deduplication, and semantic integration of the original text fragments corresponding to the atomic nodes in the knowledge subgraph.

[0053] Furthermore, S6 specifically includes the following sub-steps:

[0054] S61: Input the domain knowledge subgraph structure constructed in S5 and the extracted relevant atomic node text into the large language model to generate a preliminary natural language answer based on knowledge support;

[0055] S62: If the user query involves the evolution of communication protocol versions or standard replacement, extract the corresponding data based on the timestamp or version tag to make the answer content consistent with the target version; generate the final optimized answer, and the user can view the answer content, source node information, version tag and semantic path hints.

[0056] A data retrieval enhancement generation device based on a large model in the field of satellite communications includes a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they implement the steps of the above-mentioned method.

[0057] The advantages of this invention compared to the prior art are as follows:

[0058] The two-layer semantic clustering structure and two-layer knowledge graph construction mechanism proposed in this invention can effectively represent multi-granularity semantic units in the field of satellite communication and support cross-document and cross-standard knowledge organization and retrieval.

[0059] This invention introduces a large language model to complete and standardize entities and relations, making the knowledge graph construction process more automated and the relation extraction quality better.

[0060] This invention is based on an entity-driven graph subgraph retrieval method, which can more accurately locate the high-density semantic clusters involved in the user's query intent and improve the relevance and coverage of information extraction.

[0061] The large-scale model generated by this invention integrates graph paths and original corpora, ensuring the traceability, professionalism, and logic of the output content. It is suitable for complex knowledge service tasks such as multi-document collaborative reading, protocol standard comparison, and key technology Q&A in satellite communication scenarios, significantly reducing the workload of manual organization and Q&A, and has broad engineering application prospects and promotional value. Attached Figure Description

[0062] Figure 1 This is a flowchart of a data retrieval enhancement generation method for the satellite communication field based on a large model, as proposed in this invention.

[0063] Figure 2 This invention presents an information processing flow for a data retrieval enhancement generation method based on a large model in the field of satellite communications. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments.

[0065] Example 1: Reference Figure 1-2 As shown, this invention provides a data retrieval enhancement generation method based on a large model in the field of satellite communications, comprising the following steps:

[0066] S1: Perform structured cleaning on multi-source heterogeneous data in the field of satellite communication to form multiple text blocks and annotate metadata; among them, multi-source heterogeneous data in the field of satellite communication includes the collection of 3GPP protocol documents, satellite network architecture white papers and inter-satellite link technology reports.

[0067] In this embodiment, S1 specifically includes:

[0068] S11: Collect diverse documents from authoritative data sources in the field of satellite communications, including 3GPP protocol documents, satellite network architecture white papers, and inter-satellite link technology reports. The collection process adopts distributed crawling and API call methods to ensure the timeliness and integrity of the data. The diverse documents collected are uniformly converted into a unified plain text format, mainly including removing format noise, preliminary identification of chart information, and chapter segmentation.

[0069] S12: For texts exceeding the set length, a segmentation method based on semantic boundary detection is adopted to split the document into multiple semantically independent text blocks. Each text block contains a complete semantic unit, such as a technical definition, protocol description, or design scheme, which facilitates subsequent clustering analysis and knowledge extraction.

[0070] S13: Automatically extract and label metadata based on text content, such as the document ID, chapter title, key term tags, publication time, etc., to form a metadata set.

[0071] S2: Perform document-level clustering on the text blocks to generate document topic clusters, and perform fine-grained secondary clustering within the document topic clusters to form fine-grained semantic clusters, resulting in a two-layer clustering result.

[0072] In this embodiment, S2 specifically includes:

[0073] S21: Divide each text block High-dimensional vector representations are generated using pre-trained large language models (such as BERT and SimCSE). ;

[0074] S22: Use hierarchical clustering algorithm to analyze all text vectors Perform document-level clustering to divide the document into several document topic clusters:

[0075]

[0076] Intra-cluster text similarity satisfies:

[0077]

[0078] In the formula, The set similarity threshold is 1≤j≤m, where m is the number of document topic clusters;

[0079] S23: In each document subject cluster Internally, the K-Means algorithm is further used to subdivide the text vectors into more semantically consistent fine-grained semantic clusters:

[0080]

[0081] Fine-grained clusters exhibit higher vector similarity, satisfying the following:

[0082]

[0083] In the formula, 1≤ ≤ , This represents the number of fine-grained semantic clusters.

[0084] S24: The clustering results are evaluated using the silhouette coefficient, calculated using the following formula:

[0085]

[0086] In the formula, 𝑎 represents the average distance within a cluster, and 𝑏 represents the minimum average distance between clusters;

[0087] S25: Adjust the clustering parameters based on the K value, return to S23 for iteration, until the similarity requirement is met.

[0088] S3: Based on the results of two-layer clustering and the cleaned block data, a two-layer knowledge graph is constructed. Atomic nodes correspond to text blocks, and super nodes correspond to fine-grained semantic clusters. Entity relationships are extracted using an offline large model to supplement semantic associations, forming a two-layer knowledge graph that supports domain knowledge reasoning.

[0089] In this embodiment, S3 specifically includes:

[0090] S31: Divide the text into chunks Each text block in Transformed into atomic nodes in the spectrum and attach metadata information The smallest semantic unit of the graph includes, but is not limited to, document identifiers, chunk locations, corresponding topic cluster labels, timestamps, and version numbers.

[0091] S32: Fine-grained semantic clusters processed by the two-layer clustering algorithm Merging sets of atomic nodes to construct supernodes To achieve higher levels of semantic abstraction in knowledge representation, each supernode corresponds to a semantic topic or concept class;

[0092] S33: Extract domain entities based on atomic node text. It automatically generates natural language semantic relationships between entities using a local open-source large model and constructs connection edges between entities;

[0093] S34: Add the following node and edge types to the graph structure to construct a complete two-layer knowledge graph. ,in A set of nodes, including a set of atomic nodes. Supernode set and entity node set ; It is a set of edges, including edges representing semantic relationships between entities. Structural edges of atomic nodes and their supernodes and edges used to represent cross-domain semantic connections such as cross-document references, number jumps, and chronological order. The overall structure of the graph is shown below:

[0094]

[0095] It also adds cross-document navigation edges (such as terminology references and numbering intersections) and time-series connection edges (sequential events and version iterations) to achieve semantic navigation;

[0096] By introducing path rules based on logical reasoning, interpretable and composable relationship links are constructed in a two-layer knowledge graph, supporting complex knowledge reasoning and implicit semantic discovery. It is particularly suitable for scenarios such as concept combination, function inheritance, and performance derivation in the field of satellite communication.

[0097] S4: Receives natural language questions input by the user, performs word segmentation, stop word removal, entity recognition, and semantic standardization processing to generate a standardized semantic representation.

[0098] In this embodiment, S4 specifically includes:

[0099] S41: Receive the natural language query input by the user, perform word segmentation on the query text, and achieve high-precision word segmentation using a method combining a domain dictionary and a statistical model; assuming the input query is a sentence... The word segmentation process is to find the set of segmentation points. ,in This makes the segmented word sequence This satisfies the requirement of maximizing the overall segmentation probability;

[0100] S42: Perform stop word removal on the word segmentation results to remove meaningless words, and dynamically adjust in combination with the domain stop word list to improve the accuracy of subsequent entity recognition;

[0101] S43: Combine local open-source large model with rule template matching to extract key named entities in the query, including technical terms, communication protocol names, system component names, etc., to provide basic semantic anchors for subsequent graph retrieval;

[0102] S44: Map the extracted entity information to a standard domain vocabulary or ontology to achieve semantic normalization, which facilitates subsequent matching and reasoning with nodes in the knowledge graph.

[0103] S5: Based on the entity information of the user's question, perform multi-dimensional retrieval in the two-layer knowledge graph, extract relevant entity nodes and their associated paths, construct a knowledge subgraph and extract relevant atomic node text.

[0104] In this embodiment, S5 specifically includes:

[0105] S51: Based on the named entities identified in the user question, search for the corresponding entity nodes in the two-layer knowledge graph to obtain a set of atomic nodes that are directly related to the semantics of the user question; where the entity types include protocol names, technical components, parameter items, etc.

[0106] S52: Statistically analyze the affiliation of the hit entity nodes in the second-level graph (fine-grained semantic clusters), identify the semantic clusters with the most concentrated distribution of entities, and consider them to represent the main semantic scope of the current problem;

[0107] S53: Within the identified semantic clusters, collect other entity nodes that share edges with the hit entity, and combine them with the relational paths between entities in the graph (such as "inherited from", "applied to", "interoperable", etc.) to expand and obtain semantically related atomic nodes;

[0108] S54: Aggregate related atomic nodes and upstream and downstream relationship paths to construct a compact and semantically related knowledge subgraph for subsequent knowledge-driven answer generation;

[0109] S55: Perform unified extraction, deduplication, and semantic integration of the original text fragments corresponding to the atomic nodes in the knowledge subgraph.

[0110] S6: Combining the knowledge subgraph and related atomic node text, an initial answer is generated using a large language model, and language quality checks, professional terminology consistency verification, and version control are performed to form the final optimized answer.

[0111] In this embodiment, S6 specifically includes:

[0112] S61: Input the domain knowledge subgraph structure constructed in S5 and the extracted relevant atomic node text into the large language model to generate a preliminary natural language answer based on knowledge support;

[0113] S62: If the user query involves the evolution of communication protocol versions or standard replacement, extract the corresponding data based on the timestamp or version tag to make the answer content consistent with the target version; generate the final optimized answer, and the user can view the answer content, source node information, version tag and semantic path hints.

[0114] This invention achieves high-precision question-and-answer generation in satellite communication scenarios by constructing a domain knowledge graph and combining semantic retrieval and large model generation capabilities. This effectively improves the efficiency of teaching, operation and maintenance support, and engineering knowledge acquisition, and helps to upgrade the intelligent system of scientific research and engineering training.

[0115] Example 2: The present invention also includes a data retrieval enhancement and generation device based on a large model in the field of satellite communication, including a memory and one or more processors. The memory stores executable code. When the one or more processors execute the executable code, they are used to implement all the steps of the data retrieval enhancement and generation method based on a large model in the field of satellite communication described in Example 1. These include modules such as multi-source heterogeneous data cleaning, two-layer semantic clustering construction, two-layer knowledge graph construction, natural language query parsing, knowledge graph-driven retrieval, answer generation optimization, and user feedback closed loop. The device has high scalability, high accuracy, and good semantic reasoning capabilities.

[0116] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0117] Example 3: This example provides a computer-readable storage medium storing a program. When executed by a processor, the program implements the data retrieval enhancement generation method for satellite communication based on a large model described in Example 1. The storage medium can be in the form of a disk, optical disk, flash memory, EEPROM, SSD, etc., and is suitable for scenarios such as servers, teaching terminals, and portable smart devices.

[0118] The program modules include: query parsing module, knowledge graph matching module, hybrid retrieval module, evidence reordering module, large model generation module, answer optimization and tracing module, etc., which can be deployed locally or remotely called cloud large model service for collaborative computing via API.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data retrieval enhancement generation method based on a large model in the field of satellite communications, characterized in that, Includes the following steps: S1: Perform structured cleaning on multi-source heterogeneous data in the field of satellite communication, forming multiple text blocks and annotating metadata; among them, multi-source heterogeneous data in the field of satellite communication includes the collection of 3GPP protocol documents, satellite network architecture white papers and inter-satellite link technology reports; S2: Perform document-level clustering on the text blocks to generate document topic clusters, and perform fine-grained secondary clustering within the document topic clusters to form fine-grained semantic clusters, resulting in a two-layer clustering result; S3: Based on the results of two-layer clustering and the cleaned block data, a two-layer knowledge graph is constructed. Atomic nodes correspond to text blocks, and super nodes correspond to fine-grained semantic clusters. Entity relationships are extracted using an offline large model to supplement semantic associations, forming a two-layer knowledge graph that supports domain knowledge reasoning. S4: Receives natural language questions input by the user, performs word segmentation, stop word removal, entity recognition, and semantic standardization processing to generate a standardized semantic representation; S5: Based on the entity information of the user's question, perform multi-dimensional retrieval in the two-layer knowledge graph, extract relevant entity nodes and their associated paths, construct a knowledge subgraph and extract relevant atomic node text; S6: Combining the knowledge subgraph and related atomic node text, the initial answer is generated using a large language model, and language quality checks, professional terminology consistency verification, and version control are performed to form the final optimized answer; Specifically, S2 includes the following sub-steps: S21: Divide each text block High-dimensional vector representations are generated using a pre-trained large language model. ; S22: Use hierarchical clustering algorithm to analyze all text vectors Perform document-level clustering to divide the document into several document topic clusters: Intra-cluster text similarity satisfies: In the formula, The set similarity threshold is 1≤j≤m, where m is the number of document topic clusters; S23: In each document subject cluster Internally, the K-Means algorithm is used to subdivide text vectors into semantically consistent fine-grained semantic clusters: Fine-grained cluster intra-vector similarity satisfies: In the formula, 1≤ ≤ , This represents the number of fine-grained semantic clusters. S24: The silhouette coefficient is used to evaluate the clustering results. The calculation formula is as follows: In the formula, 𝑎 represents the average distance within a cluster, and 𝑏 represents the minimum average distance between clusters; S25: According to Adjust the clustering parameters and return to S23 for iteration until the similarity requirement is met; Specifically, S3 includes the following sub-steps: S31: Divide the text into chunks Each text block in Transformed into atomic nodes in the spectrum and attach metadata information The smallest semantic unit of the graph includes the document identifier, the location of the chunk, the corresponding topic cluster number, the timestamp, and the version number. S32: Fine-grained semantic clusters processed by the two-layer clustering algorithm Merging sets of atomic nodes to construct supernodes Each supernode corresponds to a semantic topic or concept class; S33: Extract domain entities based on atomic node text. It automatically generates natural language semantic relationships between entities using a local open-source large model and constructs connection edges between entities; S34: Add node and edge types to the graph structure to construct a complete two-layer knowledge graph. ,in A set of nodes, including a set of atomic nodes. Supernode set and entity node set ; It is a set of edges, including edges representing semantic relationships between entities. The structural edges of atomic nodes and their supernodes And cross-domain semantic join edges used to represent cross-document references, number jumps, and chronological order. .

2. The data retrieval enhancement generation method for satellite communication based on a large model according to claim 1, characterized in that, S1 specifically includes the following sub-steps: S11: Collect diverse documents from authoritative data sources in the field of satellite communications, including 3GPP protocol documents, satellite network architecture white papers, and inter-satellite link technology reports. The collection process adopts distributed crawling and API call methods. The system performs a unified format conversion on the diverse documents collected, transforming documents of various formats into a unified plain text format, including removing format noise and performing preliminary identification and chapter segmentation of chart information; S12: For texts exceeding the set length, a segmentation method based on semantic boundary detection is adopted to split the document into multiple semantically independent text blocks, each of which contains a complete semantic unit; S13: Automatically extract and label metadata based on text content, including the document ID, chapter title, key term tags, and publication time, to form a metadata set.

3. The data retrieval enhancement generation method for satellite communication based on a large model according to claim 1, characterized in that, S4 specifically includes the following sub-steps: S41: Receive natural language questions input by the user, and perform word segmentation on the natural language question text using a method that combines a domain dictionary and a statistical model, so that the segmented word sequence satisfies the requirement of maximizing the overall segmentation probability; S42: Remove meaningless words from the word segmentation results and dynamically adjust them in conjunction with the domain stop word list; S43: Combine local open-source large model with rule template matching to extract key named entities from the query, including technical terms, communication protocol names and system component names; S44: Map the extracted entity information to a standard domain vocabulary or ontology to generate a normalized semantic representation.

4. The data retrieval enhancement generation method for satellite communication based on a large model according to claim 1, characterized in that, S5 specifically includes the following sub-steps: S51: Based on the named entities identified in the user question, search for the corresponding entity nodes in the two-layer knowledge graph to obtain a set of atomic nodes that are directly related to the semantics of the user question; where the entity types include protocol names, technical components, and parameter items; S52: Statistically determine the affiliation of the hit entity nodes in the fine-grained semantic clusters, identify the semantic clusters in which the entities are most concentrated, and represent the main semantic scope of the current problem; S53: Within the identified semantic clusters, collect other entity nodes that share edges with the hit entity, and combine them with the relational paths between entities in the graph, including inheritance, application, and interoperability, to expand and obtain semantically related atomic nodes. S54: Aggregate related atomic nodes and upstream and downstream relationship paths to construct a compact and semantically related knowledge subgraph; S55: Perform unified extraction, deduplication, and semantic integration of the original text fragments corresponding to the atomic nodes in the knowledge subgraph.

5. The data retrieval enhancement generation method for satellite communication based on a large model according to claim 1, characterized in that, S6 specifically includes the following sub-steps: S61: Input the domain knowledge subgraph structure constructed in S5 and the extracted relevant atomic node text into the large language model to generate a preliminary natural language answer based on knowledge support; S62: If the user query involves the evolution of communication protocol versions or standard changes, extract the corresponding data based on the timestamp or version tag to ensure that the answer content is consistent with the target version; The final optimized answer is generated, and users can view the answer content, source node information, version tags, and semantic path hints.

6. A data retrieval enhancement and generation device based on a large model in the field of satellite communication, characterized in that, The method includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the steps of the method according to any one of claims 1-5.

Images