Ship sound stealth expert model construction method and device based on knowledge graph

Abstract

The invention discloses a ship sound stealth expert model construction method based on a knowledge graph, and the method comprises the following steps: data extraction: collecting and preprocessing data from various determined heterogeneous sources, converting a processing format, cleaning initial data, and preparing data for subsequent knowledge extraction; knowledge graph construction: taking the preprocessed data as input, and coordinating and creating a sound stealth knowledge graph; and retrieval enhancement generation: providing a user-friendly environment for the expert user based on LLM and GraphRAG to interact with the expert model. The invention discloses a ship sound stealth expert model construction device based on a knowledge graph. The method realizes the vertical field question-answering effect of the ship acoustic stealth specialty, provides professional question-answering service and friendly interaction service for the user, and can be widely applied to the technical field of top acoustic design.

Description

Technical Field

[0001] This invention relates to top-level acoustic design technology, and in particular to a method and apparatus for constructing a knowledge graph-based expert model for ship acoustic stealth. Background Technology

[0002] The field of acoustic stealth design encompasses a vast knowledge system, involving multiple disciplines such as acoustics, materials science, structural mechanics, and systems engineering, exhibiting characteristics of multi-layered complexity and dispersion. This characteristic makes efficient utilization and sharing of knowledge difficult, severely impacting design efficiency and effectiveness. While industrial sectors have accumulated a wealth of technical documents and expert experience, this knowledge and data are often scattered across different systems and departments, urgently requiring integration and refinement. Therefore, based on accumulated technical materials and expert knowledge, as well as external literature, patents, and technical documents, it is crucial to extract valuable knowledge and construct an intelligent acoustic stealth knowledge management and design support system. This not only integrates dispersed knowledge and data but also provides scientific and intelligent design suggestions and decision support through the analysis and refinement of vast amounts of technical materials and expert experience, achieving systematic knowledge management and efficient sharing—a pressing technical problem for the industry. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the above-mentioned background technology and provide a method and apparatus for constructing a ship acoustic stealth expert model based on knowledge graph. This allows the invention to explore a question-and-answer system that combines a large language model with a professional knowledge base, and to explore the deep integration of a large language model and a knowledge graph. This enables the invention to achieve vertical domain question-and-answer effects in the field of ship acoustic stealth and to provide users with professional question-and-answer services and user-friendly interactive services.

[0004] This invention provides a method for constructing a ship acoustic stealth expert model based on a knowledge graph, comprising the following steps: Data extraction: collecting and preprocessing data from various heterogeneous sources, converting the processing format, cleaning the initial data, and preparing data for subsequent knowledge extraction; Knowledge graph construction: using the preprocessed data as input, coordinating and creating an acoustic stealth knowledge graph; Retrieval enhancement generation: providing a user-friendly environment for expert users to interact with the expert model based on LLM and GraphRAG.

[0005] In the above technical solution, the specific steps of the data extraction step are as follows: Formatting: The data is formatted to ensure compatibility with GraphRAG; Ordered storage: All converted TXT files are systematically stored in a designated folder prepared for GraphRAG input; Preserving entity relationships: The GraphRAG splitting method ensures that the relationships between entities within the text remain unchanged, and maintains the relationships between entities within the text throughout the splitting process through GraphRAG's linking or indexing system, allowing expert models to accurately interpret and connect the information between various parts of the document.

[0006] In the above technical solution, the formatting sub-step of the data extraction step includes the following specific processes: Unified conversion to PDF format: All received documents are initially converted to PDF format to ensure consistency and universal recognizability in subsequent processing; Converting PDF to Markdown format using marker-pdf: Each PDF document is converted to Markdown format using the marker-pdf tool, becoming an intermediate format; Optimizing the processing of large documents: Each large PDF document is carefully divided into smaller, more manageable parts; Converting Markdown to TXT format using markdown-txt: After splitting the documents, a Python script based on markdown-txt is used to convert the Markdown files to TXT format.

[0007] In the above technical solution, the knowledge graph construction steps are as follows: designing the evolution of the acoustic cloaking ontology, applying large language model technology for entity recognition and linking, extracting semantic relationships between recognized entities, and integrating knowledge from different sources into a unified graph structure.

[0008] In the above technical solution, the knowledge graph construction steps sequentially include six stages: text segmentation, knowledge graph generation, knowledge graph enhancement, community summarization, document processing, and knowledge graph visualization, as follows: Text segmentation: Long texts are split into smaller parts for subsequent information extraction and processing, ensuring that each segment has complete semantics and adapts to the processing capabilities of natural language processing models; Knowledge graph generation: Structured information, including entities, relationships, and their attributes, is extracted from the text; Knowledge graph enhancement: Community detection identifies closely connected groups in the knowledge graph using Leiden or Louvain algorithms to form communities; Graph embedding methods are used to map knowledge graph nodes to a vector space to support similarity retrieval and downstream tasks; Community summarization… Abstract: First, key entities in the community are identified by computing high-influence nodes; second, topic modeling methods are used to analyze the topics covered by the community; third, LLM is used to generate natural language summaries to facilitate user understanding and retrieval of community information; fourth, the vectorization of community summaries generates high-dimensional vectors through deep learning models for similar community retrieval, automatic clustering, personalized recommendations, and classification tasks, or clustering methods generate community-level vectors to support similar community search; document processing: knowledge graph results are stored in CSV or database format for subsequent retrieval and analysis; document links are used to establish mapping relationships between entities in the knowledge graph and original documents to ensure traceability; document vector representations are calculated based on Sentence-BERT or OpenAI Embeddings and indexed using a vector database; knowledge graph visualization: dimensionality reduction methods use UMAP or t-SNE techniques to embed and reduce the high-dimensional graph to 2D or 3D form for visualization analysis; interactive knowledge graph visualization supports users to dynamically explore entity relationships by implementing interactive knowledge graph displays.

[0009] In the above technical solution, the text segmentation step in the knowledge graph construction process includes the following segmentation methods: a token-based segmentation method to adapt to the input limitations of large language models; a semantic-based segmentation method to ensure semantic integrity; and a hierarchical segmentation strategy, which first divides the text into chapters and paragraphs, and then further subdivides it into sentences or clauses. The specific process is as follows: fixed window: segmenting the text according to a fixed length; semantic segmentation: dividing the content into blocks based on semantic similarity; recursive segmentation: iteratively segmenting the text until the conditions are met; document structure-based segmentation: segmenting by headings and paragraph structure elements; and LLM-based segmentation: using large models to intelligently segment the boundaries.

[0010] In the above technical solution, the knowledge graph generation step of the knowledge graph construction process includes the following parts: First, by using named entity recognition technology, each entity is identified from the text, and entity extraction is performed using a deep learning model or an LLM-based prompt word method; then, the relationship between entities is identified by relation extraction technology; during the processing of multi-source text, entity merging and entity normalization are performed to solve the problem of information duplication caused by different representations of the same entity.

[0011] In the above technical solution, the content of the retrieval enhancement generation step is as follows: allowing users to use natural language and prompt engineering to retrieve corresponding knowledge from the knowledge graph expert database, and receive design suggestions generated by LLM and provide visualization capabilities to assist professionals in acoustic stealth-related designs.

[0012] In the above technical solution, the specific process of the retrieval enhancement generation step is as follows: Contextualizing retrieval knowledge: The raw data retrieved from the knowledge graph is formatted and presented to the LLM in a way that makes its meaning and relevance clear. This involves providing the target entity and its direct neighbors, as well as the edges connecting them, effectively providing the LLM with a local subgraph relevant to the user query; Hint engineering guides the LLM to generate the required type of output: Hints should include domain-specific terminology related to acoustic cloaking, explicitly specifying any constraints or requirements of the generated design recommendations or solutions; Hints can also provide examples of successful design strategies or highlight common flaws that need to be avoided; Utilizing the LLM's generative capabilities to synthesize design recommendations: Based on the enhanced context and well-designed hints, the LLM generates design recommendations, proposing alternative solutions for specific acoustic cloaking challenges, or highlighting potential problems that need to be considered. The level of detail and complexity of the generated output is customized according to the user's needs and the specific context of the query.

[0013] The present invention also provides a knowledge graph-based expert model construction device for ship acoustic stealth, which has a computer program that can execute a knowledge graph-based expert model construction method for ship acoustic stealth.

[0014] The present invention provides a method and apparatus for constructing a ship acoustic stealth expert model based on knowledge graphs, which has the following beneficial effects: This invention studies a question-answering system paradigm that combines a large language model with a knowledge base. It explores the deep integration of a ship acoustic stealth knowledge graph with a large language model, achieving bidirectional interaction between the language model and the knowledge graph. This enables professional, vertically specific question-answering effects and provides a user-friendly interactive service. The professional question-answering module provides expert answers and can interpret structured data, reducing the difficulty of user comprehension. Furthermore, it can further enhance the knowledge graph with expert-verified knowledge. This new question-answering system paradigm provides users with more accurate and professional answers while maintaining a user-friendly interactive experience. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of the knowledge graph-based expert model construction method for ship acoustic stealth according to the present invention; Figure 2 This is a flowchart illustrating the data extraction steps in the knowledge graph-based expert model construction method for ship acoustic stealth of the present invention. Figure 3 This is a schematic diagram of the hierarchical text segmentation strategy in the knowledge graph construction step of the knowledge graph-based expert model construction method for ship acoustic stealth in this invention. Figure 4 This is a flowchart illustrating the retrieval enhancement generation step in the knowledge graph-based expert model construction method for ship acoustic stealth of the present invention. Figure 5 This is a schematic diagram of the architecture of the knowledge graph-based expert model construction device for ship acoustic stealth according to the present invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments, but these embodiments should not be construed as limiting the present invention.

[0017] The technical problem addressed by this invention is that while the industrial sector has accumulated a wealth of research results, technical documents, and expert experience in acoustic stealth design, this knowledge and data are often scattered across different systems and departments, lacking a unified expression and efficient sharing mechanism, which severely restricts collaborative efficiency. To address this situation, this invention aims to explore a question-and-answer system combining a large language model with a professional knowledge base, and to explore the deep integration of large language models and knowledge graphs to achieve vertical domain question-and-answer effects in the field of ship acoustic stealth, providing users with professional question-and-answer services and user-friendly interactive services.

[0018] Example 1 GraphRAG can construct expert knowledge bases through graph structures, and also utilize graph structures to assist in retrieval enhancement. Furthermore, GraphRAG can effectively process complex text data and provide high-quality answers. This invention uses GraphRAG to achieve intelligent reasoning related to acoustic cloaking knowledge, assisting in the generation and optimization of acoustic cloaking solutions, and improving the efficiency and effectiveness of acoustic cloaking design. The GraphRAG engine, as the core of the intelligent system, is responsible for indexing the acoustic cloaking knowledge graph for efficient retrieval. Upon receiving a user query, the engine converts the natural language query into a graph-based query that can be executed against the knowledge graph. Then, it uses a graph traversal algorithm to retrieve the most relevant subgraphs. Finally, it obtains the retrieved knowledge graph information and uses it to enhance the prompts provided to the large language model, thereby triggering the generation of design suggestions or solutions. To implement the above process, it can be divided into the following key steps, as detailed in the following section. Figure 1 : Data Extraction: Responsible for collecting and preprocessing data from various previously identified heterogeneous sources, including internal technical documents in various formats, captured expert knowledge, external academic papers, patent specifications, and technical documents from external entities; this part will handle format conversion, initial data cleaning, and data preparation for subsequent knowledge extraction. Knowledge graph construction: Taking preprocessed data as input and coordinating the creation of the acoustic cloaking knowledge graph, this step will include several sub-processes: design and evolution of the acoustic cloaking ontology, application of large language model technology for entity recognition and linking, extraction of semantic relationships between recognized entities, and fusion of knowledge from different sources into a unified graph structure. This part will also include a mechanism for expert review and verification of the knowledge graph content. LLM and GraphRAG-based retrieval enhancement generation: This provides expert users with a user-friendly environment to interact with expert models. The interface allows users to formulate complex queries using natural language and prompt engineering, such as thought chain reasoning templates or rule templates, to retrieve relevant knowledge from the knowledge graph expert database in a meaningful and more professional way. Users can also receive design suggestions generated by LLM and provide visualization capabilities to assist professionals in acoustic stealth-related designs.

[0019] Example 2 This embodiment is basically the same as Embodiment 1, except that: The present invention relates to a knowledge graph-based expert model construction method for ship acoustic stealth, the detailed process of which is as follows: Step 1: Detailed Data Extraction Process Predictive models like GraphRAG typically require input in a specific format. To minimize data loss during format conversion and ensure compatibility with GraphRAG, the data needs to be formatted. See below for the specific process. Figure 2 As shown: (1) Convert to PDF format: In order to standardize the document environment, all received documents are initially converted to PDF format. This step ensures consistency and universal recognizability in subsequent processing.

[0020] (2) Use marker-pdf to convert PDF to Markdown format: Use the marker-pdf tool to convert each PDF document to Markdown format. Markdown was chosen because it is simple, easy to operate, and can preserve the basic structure and content of the original document, making it an ideal intermediate format.

[0021] (3) Optimization for handling large documents: Considering the limitations of marker-pdf when handling large files, especially the issues of speed reduction and memory overflow, a document splitting strategy is implemented. Each large PDF document is carefully divided into smaller, more manageable parts. This approach not only improves processing speed but also alleviates potential memory problems.

[0022] (4) Convert Markdown to TXT format using markdown-txt: After splitting the document, use a markdown-txt-based Python script to convert the Markdown file to TXT format. This conversion step ensures that the document meets the input specifications required by GraphRAG.

[0023] (4) Ordered storage for GraphRAG: All converted TXT files are systematically stored in designated folders prepared for GraphRAG input. This organization ensures that the documents are easily accessible, in the correct order, and ready for analysis.

[0024] (5) Preservation of entity relationships: Since GraphRAG can process multiple documents simultaneously, the splitting method ensures that the relationships between entities within the text remain unchanged. These relationships are maintained throughout the splitting process through a clear linking or indexing system, allowing the model to accurately interpret and connect information between different parts of the document.

[0025] Step Two: Detailed Process of Knowledge Graph Construction The construction of a knowledge graph based on GraphRAG consists of six core components, including text segmentation, knowledge graph generation, knowledge graph enhancement, community summarization, document processing, and knowledge graph visualization.

[0026] (1) Text segmentation The first step in knowledge graph construction is to segment long texts to facilitate subsequent information extraction and processing. The core objective of text segmentation is to ensure that each segment possesses complete semantics and can adapt to the processing capabilities of natural language processing models. Currently, text segmentation primarily employs token-based segmentation methods to accommodate the input constraints of large language models such as GPT-4. Additionally, semantic-based segmentation methods, such as dependency parsing and semantic role labeling (SRL), can be used to ensure semantic integrity. Furthermore, to preserve the structured information of the text, a hierarchical segmentation strategy can be adopted, first dividing it into chapters and paragraphs, and then further subdividing it into sentences or clauses. (See [link to text segmentation strategy] for details.) Figure 3 As shown: 1) Fixed window: Split text into fixed lengths (e.g., every 200 characters).

[0027] 2) Semantic segmentation: Divide content into blocks based on semantic similarity.

[0028] 3) Recursive block division: Iteratively divide the text until a condition is met (such as minimum block size).

[0029] 4) Document structure-based segmentation: segmented by structural elements such as headings and paragraphs.

[0030] 5) LLM-based segmentation: Use a large model to intelligently segment block boundaries (such as identifying key paragraphs).

[0031] (2) Knowledge graph generation The core of knowledge graph construction lies in extracting structured information from text, including entities, relationships, and their attributes. This process mainly involves the following key technical aspects: First, Named Entity Recognition (NER) technology is used to identify entities such as names, locations, organizations, and events from text. Entity extraction can be performed using deep learning models (such as BERT-NER) or based on LLM prompts. For example, a prompt-based method can guide the model to output entity information in the following format: ("entity"<|>"Elon Musk"<|>"person"<|>"Tesla CEO") Then, relation extraction (RE) techniques are used to identify the relationships between entities. Current mainstream methods include rule-based dependency parsing, the BERT relation classifier based on deep learning, and the LLM-based Prompt relation extraction method. In multi-source text processing, entity resolution and entity normalization are also required to solve the problem of information duplication caused by different representations of the same entity. For example, "LLM" and "large model" can be normalized to the same concept. Entity resolution can be performed using rule-based matching methods (such as string similarity calculation and regular expression matching), machine learning methods (such as TF-IDF or word vector similarity calculation), and deep learning models (such as Siamese Network) to calculate representation similarity, thereby improving the accuracy of resolution.

[0032] (3) Knowledge Graph Enhancement Knowledge graph augmentation aims to optimize the knowledge graph structure, enhance retrieval capabilities, and uncover hidden information. Community detection is a key component, using algorithms like Leiden or Louvain to identify closely connected groups within the knowledge graph to form communities. Furthermore, to enhance the representational power of the knowledge graph, graph embedding methods such as Node2Vec or GraphSAGE can be used to map knowledge graph nodes to a vector space, supporting similarity retrieval and downstream tasks.

[0033] (4) Community Summary Community summaries are structured analyses of communities identified in knowledge graphs to extract their core information. First, key entities within a community can be identified by computing high-influence nodes (e.g., using the PageRank algorithm). Simultaneously, topic modeling methods, such as LDA or BERTopic, can be used to analyze the topics covered by the community. To facilitate user understanding and retrieval of community information, LLM can be used to generate natural language summaries, for example: Community X mainly involves "artificial intelligence," with core members including "Sam Altman" and "Elon Musk," and is associated with "OpenAI" and "DeepMind." The vectorized representation of community summaries is also crucial, and high-dimensional vectors can be generated using Sentence-BERT or other deep learning models. These vectors can be used for similar community retrieval, automatic clustering, personalized recommendations, and classification tasks. For example, in recommendation systems, personalized content recommendations can be provided based on matching community vectors with user historical behavior; in classification tasks, they can be used for community affiliation prediction, thereby enhancing automated knowledge organization capabilities. Alternatively, clustering methods can be used to generate community-level vectors to support similar community search.

[0034] (5) Document processing After the knowledge graph is generated, the original documents need further processing to ensure their relevance to the knowledge graph and support efficient retrieval. Document processing includes several key parts: the knowledge graph results should be stored in CSV or database (such as Neo4j, GraphDB) format for subsequent retrieval and analysis. Document linking is used to establish mapping relationships between entities in the knowledge graph and the original documents, ensuring traceability. To improve retrieval performance, document vector representations based on Sentence-BERT or OpenAI Embeddings can be used, and indexed using vector databases such as Pinecone, Weaviate, and FAISS.

[0035] (6) Knowledge graph visualization Knowledge graph visualization aims to enable users to intuitively understand the structure and relationships of knowledge graphs. Currently, commonly used visualization methods include: Dimensionality Reduction: using UMAP or t-SNE techniques to embed high-dimensional graphs into 2D or 3D forms for visualization and analysis; Interactive Knowledge Graph Visualization: using tools (such as Neo4j Bloom, D3.js + Cytoscape.js, Graphistry) to achieve interactive knowledge graph displays, supporting users to dynamically explore entity relationships.

[0036] Step 3: Detailed Process of LLM-Based Enhancement Generation The LLM-based augmented generation module is responsible for acquiring knowledge retrieved from the knowledge graph and using this knowledge to generate expert-level acoustic cloaking design suggestions or solutions. The LLM-based augmented generation process is described below. Figure 4 As shown.

[0037] (1) Contextualization of knowledge retrieval. The raw data retrieved from the knowledge graph may consist of a set of nodes and edges, which need to be formatted and presented to the LLM in a way that makes their meaning and relevance clear. This may involve providing not only the target entity, but also its direct neighbors and the edges connecting them, thereby effectively providing the LLM with a local subgraph relevant to the user query. This contextualized structure, whether as a textual description of the subgraph or a more structured format, will affect the efficiency with which the LLM utilizes it.

[0038] (2) Cueing plays a crucial role in guiding the LLM to generate the desired type of output. Well-designed cueing should include domain-specific terminology related to acoustic stealth and explicitly specify any constraints or requirements for the generated design recommendations or solutions. For example, a cue might instruct the LLM to “recommend three acoustic materials suitable for coating hulls to reduce sonar detection, taking into account factors such as effectiveness across different frequency ranges and durability in marine environments.” Cueing can also provide examples of successful design strategies or highlight common pitfalls that need to be avoided. The careful design of these cueings is essential for obtaining relevant and accurate expert advice from the LLM.

[0039] (3) Leveraging the generative capabilities of LLM to synthesize design recommendations. Based on enhanced context and well-designed prompts, LLM will generate design recommendations, propose alternative solutions for specific acoustic cloaking challenges, or highlight potential issues that need to be considered. The level of detail and complexity of the generated output can be customized according to the user's needs and the specific context of the query. The goal is to move beyond simply summarizing retrieved information and leverage its understanding of the domain to provide actionable and insightful guidance, thereby assisting acoustic engineers in the design process.

[0040] Example 3 See Figure 5 The present invention relates to a knowledge graph-based expert model construction device for ship acoustic stealth, comprising the following parts: Data Extraction Module: Responsible for collecting and preprocessing data from various previously identified heterogeneous sources, including internal technical documents in various formats, captured expert knowledge, external academic papers, patent specifications, and technical documents from external entities; this module will handle format conversion, initial data cleaning, and data preparation for subsequent knowledge extraction. Knowledge Graph Construction Module: Taking preprocessed data as input and coordinating the creation of the acoustic cloaking knowledge graph, this module will include several sub-processes: design and evolution of the acoustic cloaking ontology, application of large language model technology for entity recognition and linking, extraction of semantic relationships between identified entities, and fusion of knowledge from different sources into a unified graph structure. This module will also include a mechanism for expert review and verification of the knowledge graph content. The retrieval enhancement generation module based on LLM and GraphRAG provides expert users with a user-friendly environment to interact with expert models. This interface allows users to formulate complex queries using natural language and prompt engineering, such as thought chain reasoning templates or rule templates, to retrieve corresponding knowledge from the knowledge graph expert database in a meaningful and more professional way. It also allows users to receive design suggestions generated by LLM and provides visualization capabilities to assist professionals in acoustic stealth-related designs.

[0041] Key technical points of this invention: (1) Integrate the research and design results of multiple disciplines such as materials science, structural mechanics, and vibration acoustics in the ship acoustic stealth industry, establish a dynamic knowledge extraction framework covering materials, structure, equipment, system and overall acoustic design, form a multi-level and multi-dimensional acoustic stealth knowledge graph, and realize the closed-loop update and version iteration of the knowledge graph in the field of ship acoustic stealth through dynamic feedback of expert correction behavior in engineering verification scenarios.

[0042] (2) A question-answering system paradigm of large language model + professional knowledge base is proposed, which improves the professional capabilities of large model while retaining its ability to answer general questions. In the case of insufficient hardware resources, a smaller large model is selected to deploy a question-answering system in a professional field, so as to achieve the same or even better results as the larger large model in the professional field.

[0043] (3) Explore the deep integration of two knowledge paradigms: large language models and knowledge graphs. Achieving bidirectional linking between large language models and knowledge graphs can convert easily readable natural language into structured data, which can then be matched with structured data in the knowledge graph to enhance the professionalism of the answers.

[0044] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0045] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A method for constructing a ship acoustic stealth expert model based on knowledge graphs, characterized in that: Includes the following steps: Data extraction: Collect and preprocess data from various identified heterogeneous sources, transform the processing format, clean the initial data, and prepare data for subsequent knowledge extraction; Knowledge graph construction: Using preprocessed data as input, coordinate and create an acoustic cloaking knowledge graph; Retrieval Enhancement Generation: Based on LLM and GraphRAG, it provides a user-friendly environment for expert users to interact with expert models.

2. The method for constructing a ship acoustic stealth expert model based on knowledge graphs according to claim 1, characterized in that: The specific steps of the data extraction process are as follows: Formatting: Format the data to ensure compatibility with GraphRAG; Ordered storage: All converted TXT files are systematically stored in a designated folder prepared for GraphRAG input; Preserving Entity Relationships: The GraphRAG splitting method ensures that the relationships between entities within the text remain unchanged. Through GraphRAG's linking or indexing system, the relationships between entities within the text are maintained throughout the splitting process, allowing expert models to accurately interpret and connect information between different parts of the document.

3. The method for constructing a ship acoustic stealth expert model based on knowledge graphs according to claim 2, characterized in that: The formatting sub-step of the data extraction step includes the following specific processes: Unified conversion to PDF format: All received documents are initially converted to PDF format to ensure consistency and universal recognizability in subsequent processing; Use marker-pdf to convert PDF to Markdown format: Use the marker-pdf tool to convert each PDF document to Markdown format, creating an intermediate format; Optimized handling of large documents: Each large PDF document is carefully broken down into smaller, more manageable parts; Convert Markdown to TXT format using markdown-txt: After splitting the document, use a Python script based on markdown-txt to convert the Markdown file to TXT format.

4. The method for constructing a ship acoustic stealth expert model based on knowledge graphs according to claim 3, characterized in that: The knowledge graph construction steps are as follows: design the evolution of the acoustic cloaking ontology, apply large language model technology for entity recognition and linking, extract the semantic relationships between the recognized entities, and integrate knowledge from different sources into a unified graph structure.

5. The method for constructing a ship acoustic stealth expert model based on knowledge graphs according to claim 4, characterized in that: The knowledge graph construction process comprises six steps: text segmentation, knowledge graph generation, knowledge graph enhancement, community summarization, document processing, and knowledge graph visualization, as detailed below: Text chunking: Splitting long texts to facilitate subsequent information extraction and processing, ensuring that each chunk has complete semantics and adapting to the processing capabilities of natural language processing models; Knowledge graph generation: Extracting structured information from text, including entities, relationships, and their attributes; Knowledge Graph Augmentation: Community Detection uses Leiden or Louvain algorithms to identify closely connected groups in a knowledge graph to form communities; Using graph embedding methods, knowledge graph nodes are mapped to vector spaces to support similarity retrieval and downstream tasks; Community Summary: First, key entities in the community are identified by computing high-influence nodes; second, topic modeling methods are used to analyze the topics covered by the community; third, LLM is used to generate natural language summaries to facilitate user understanding and retrieval of community information. Fourth, the vectorization of community summaries generates high-dimensional vectors through deep learning models, which are used for similar community retrieval, automatic clustering, personalized recommendation and classification tasks, or clustering methods generate community-level vectors to support similar community search. Document processing: Knowledge graph results are stored in CSV or database format for subsequent retrieval and analysis; document links are used to establish mapping relationships between entities in the knowledge graph and the original documents to ensure traceability; document vector representations are calculated based on Sentence-BERT or OpenAI Embeddings and indexed using a vector database; Knowledge graph visualization: Dimensionality reduction methods use UMAP or t-SNE techniques to embed high-dimensional graphs into 2D or 3D forms for visualization analysis; interactive knowledge graph visualization enables users to dynamically explore entity relationships by displaying interactive knowledge graphs.

6. The method for constructing a ship acoustic stealth expert model based on knowledge graphs according to claim 5, characterized in that: In the text segmentation step of the knowledge graph construction process, the segmentation method includes the following parts: The token-based segmentation method adapts to the input constraints of large language models; Semantic-based segmentation methods ensure semantic integrity; The hierarchical segmentation strategy first divides the content into chapters and paragraphs, and then further subdivides it into sentences or clauses. The specific process is as follows: Fixed window: Splits text into segments of a fixed length; Semantic segmentation: dividing content into blocks based on semantic similarity; Recursive block segmentation: Iteratively segment the text until a condition is met; Document structure-based segmentation: divided by headings and paragraph structure elements; LLM-based partitioning: Intelligent partitioning of boundaries using a large model.

7. The method for constructing a ship acoustic stealth expert model based on knowledge graphs according to claim 6, characterized in that: The knowledge graph generation step in the knowledge graph construction process includes the following components: First, named entity recognition technology is used to identify various entities from the text, and entity extraction is performed using deep learning models or LLM-based prompt words. Then, relationship extraction techniques are used to identify the relationships between entities; In the process of processing multi-source text, entity merging and entity normalization are performed to solve the problem of information duplication caused by different representations of the same entity.

8. The method for constructing a ship acoustic stealth expert model based on knowledge graphs according to claim 7, characterized in that: The enhanced retrieval generation steps are as follows: allowing users to retrieve corresponding knowledge from the knowledge graph expert database using natural language and prompting engineering, and receiving design suggestions generated by LLM and providing visualization capabilities to assist professionals in acoustic stealth-related designs.

9. The method for constructing a ship acoustic stealth expert model based on knowledge graphs according to claim 8, characterized in that: The specific process of the retrieval enhancement generation step is as follows: Knowledge Contextualization: The raw data retrieved from the knowledge graph is formatted and presented to the LLM in a way that makes its meaning and relevance clear. This involves providing the target entities and their direct neighbors as well as the edges connecting them, effectively providing the LLM with local subgraphs that are relevant to the user query. The prompts guide the LLM to generate the required type of output: the prompts should include domain-specific terminology related to acoustic stealth, explicitly specify any constraints or requirements of the generated design recommendations or solutions; the prompts can also provide examples of successful design strategies or highlight common defects that need to be avoided; Leveraging the generative capabilities of LLM to synthesize design recommendations: Based on enhanced context and well-designed hints, LLM generates design recommendations that propose alternative solutions for specific acoustic cloaking challenges or highlight potential issues that need to be considered. The level of detail and complexity of the generated output is customized according to the user's needs and the specific context of the query.

10. A knowledge graph-based expert model construction device for ship acoustic stealth, comprising a computer program, characterized in that: The computer program is capable of executing the knowledge graph-based expert model construction method for ship acoustic stealth as described in any one of claims 1 to 9.

Images