A method, system, equipment, and medium for managing auxiliary design documents for high-end equipment.

Abstract

This invention provides a method, system, device, and medium for managing auxiliary design documents for high-end equipment, belonging to the field of high-end equipment manufacturing technology. The method first processes drawings to extract technical symbols and terminology; performs semantic understanding on the documents, generates tags, and disambiguates them; divides document domains using an ontology database; constructs a three-level tree-structured knowledge base to achieve cross-document entity links; receives user search terms; performs parallel full-text and semantic searches; weights and fuses the results; and outputs the final result. Through hybrid retrieval and knowledge base construction, the method achieves rapid retrieval of design documents, model matching, and shortens the R&D cycle, improving mechanical design efficiency, ensuring the structured knowledge base and high relevance of search results, and meeting the precise needs of high-end equipment manufacturing for design document management.

Description

Technical Field

[0001] This invention belongs to the field of high-end equipment manufacturing technology, specifically relating to a method, system, equipment, and medium for managing auxiliary design documents for high-end equipment. Background Technology

[0002] The high-end equipment manufacturing industry encompasses aerospace, rail transportation, precision machine tools, and energy equipment, characterized by technology intensity, high integration, and strong customization. With the development of intelligent manufacturing and the Industrial Internet, enterprises are increasingly demanding knowledge management and intelligent assisted design. Leading companies in the industry are driving a transformation from experience-driven to data and knowledge-driven approaches, requiring the construction of knowledge management systems capable of understanding professional semantics, enabling cross-document association, and supporting intelligent retrieval.

[0003] In related technologies, high-end equipment design documents are mostly in PDF and Word formats. When extracting keywords, it is generally impossible to understand the meaning of the keywords, the corresponding outline level, and the entity attributes. This can lead to ambiguity in key entities. For example, damping may refer to the damping coefficient in a control theory-related design document, but to the damping orifice in a hydraulic system document, resulting in label confusion and the appearance of incorrect documents during subsequent searches.

[0004] Furthermore, the same keyword in different documents is unrelated, requiring designers to separately open design manuals, strength calculation sheets, and procurement agreements to obtain the corresponding keyword, resulting in low reuse efficiency. Full-text search can only match documents containing the keyword and cannot identify semantically similar content, leading to incomplete search results and affecting user experience. Summary of the Invention

[0005] This invention provides a method for managing auxiliary design documents for high-end equipment. The method combines keyword precision and semantic relevance to improve retrieval accuracy; it also shortens the time for design document retrieval and model matching, thus helping to shorten the R&D cycle.

[0006] The methods include: Step S101: Process the drawings in the auxiliary design of high-end equipment and extract the technical symbols and professional terms from the drawings; Step S102: Perform semantic understanding on the documents in the auxiliary design of high-end equipment, extract the document's theme, key entities, outline and main content, generate tags, and achieve key entity disambiguation through LoRa fine-tuning; Step S103: Based on the generated tags, and combined with the ontology library in the mechanical design field, the document content corresponding to the tags is divided into domains to obtain multiple different domains; Step S104: Based on technical symbols, professional terms, tags and domains, construct a three-level tree-structured knowledge base, use a graph database to store the relationships between nodes at each level in the knowledge base, and implement cross-document entity links; Step S105: Receive search query terms submitted by the user for the auxiliary design of high-end equipment; Step S106: Based on the three-level tree-structured knowledge base, for the received search query terms, perform full-text search and large-model semantic search in parallel to obtain full-text search results and semantic search results respectively; Step S107: Perform weighted fusion of the obtained full-text search results and semantic search results to obtain the final fusion search results; Step S108: Output the final fusion search results to the user to complete the search response for high-end equipment auxiliary design related content.

[0007] Preferably, S102 specifically includes the following steps: Use a document parsing library to process the document, remove format tags, headers, footers and non-body content, and obtain the text content sequence; The transformed text content sequence is input into the large language model. The input sequence is encoded through the multi-layer Transformer network structure inside the large language model to generate a deep representation vector containing contextual semantic information. Based on the generated deep representation vectors, entity recognition and classification heads are configured in the model output layer; the recognition head traverses the text sequence, identifies and extracts entities belonging to a preset category set, which includes material type, process parameters, and component name; Configure text summarization and topic retrieval flags in the output layer of the large language model; By using a low-rank adaptive method based on LoRa, a large language model is adjusted using a dataset containing disambiguation data of entities in the mechanical field. By adjusting the parameters of the large language model, the model can map key entities with the same glyphs but different meanings to labels according to the context.

[0008] Preferably, S103 specifically includes the following steps: The tags generated in S102 are compared with the ontology library in the field of mechanical design to identify the professional terms and concepts in the tags; The model clustering algorithm is used to group the tags, and the tags with similar semantics are grouped into the same group; Based on the clustering results and the hierarchical relationships in the domain ontology library, the domain corresponding to each group is determined. Perform semantic association analysis on the tags within each domain to construct a semantic network of the tags within the domain; The constructed semantic network is mapped to a graph database to store the association between domains and tags.

[0009] Preferably, S104 specifically includes the following steps: Based on the multiple domains obtained from the domain partitioning in step S103, corresponding domain nodes are created in the graph database, and a unique identifier and attribute are assigned to each domain node. For each domain, based on the topic tags corresponding to the documents contained therein, create topic nodes in the graph database and establish parent-child relationship edges between these topic nodes and their respective domain nodes; The technical symbols and professional terms extracted in step S101 and the key entities extracted in step S102 are created as entity nodes in the graph database, and parent-child relationship edges are established between the entity nodes and their respective topic nodes. Store the document itself as a document node in the graph database, establish containment edges between the document node and the entity node describing the content, and associate the document node with its respective topic node and domain node. By traversing the entity nodes in the graph database, different entity nodes referring to the same physical object are identified, and cross-document link edges are established for different entity nodes to achieve cross-document entity links.

[0010] Preferably, S1045 specifically includes the following steps: Filter the set of nodes with the same entity type attribute from the graph database; For each selected entity node, extract its stored attribute information and associated semantic fragment content; The attribute information of each entity node is concatenated with the semantic fragment content to generate the node's feature vector; Calculate the similarity of feature vectors of different nodes under the same entity type, and find node pairs whose similarity exceeds a preset threshold; For node pairs with similarity exceeding a threshold, establish cross-document link edges and unify the identifier of one of the nodes.

[0011] Preferably, S106 specifically includes the following steps: Map user search queries to the hierarchical structure of a multi-level tree-like knowledge base, and extract domain, topic, or entity keywords associated with the query terms; The full-text search function of the graph database is invoked to match the extracted keywords in node attributes and edge relationships to obtain nodes and associated paths that meet the conditions. Input the user's search query into the large language model to generate semantic vectors for the query; The semantic retrieval function of the large language model is invoked to compare the semantic vector of the query with the semantic vector of the knowledge base node to obtain semantically similar nodes and related paths; The nodes and paths obtained from full-text search and semantic search are organized separately and output as full-text search results and semantic search results, respectively.

[0012] Preferably, S107 specifically includes the following steps: Receive a list of search results returned by the full-text search engine and semantic search module. Each result contains a document identifier and its original relevance score. Based on the position of each result in the original search list, a ranking-based score is calculated for both the full-text search results and the semantic search results. Normalize the ranking-based scores; Read the preset full-text search weight coefficients and semantic search weight coefficients, and calculate the weighted average of the normalized ranking score and the corresponding weight coefficients to obtain a fusion score for each search result; Based on the calculated fusion score, all search results from different channels are sorted in descending order to generate the final fusion search result list.

[0013] This application also provides a high-end equipment auxiliary design document management system, the system comprising: The drawing information extraction module is used to process drawings in the auxiliary design of high-end equipment and extract technical symbols and professional terms from the drawings. The semantic understanding module is used to perform semantic understanding on documents in the auxiliary design of high-end equipment, extract the document's theme, key entities, outline and main content, generate tags, and achieve key entity disambiguation through LoRa fine-tuning. The domain partitioning module, based on the generated tags and combined with the ontology library of the mechanical design domain, partitions the document content corresponding to the tags into multiple different domains; The knowledge base construction module builds a three-level tree-structured knowledge base based on technical symbols, professional terms, tags, and domains. It uses a graph database to store the relationships between nodes at each level of the knowledge base and enables cross-document entity links. The search term receiving module is used to receive search query terms submitted by users for the auxiliary design of high-end equipment; The hybrid retrieval module, based on a three-level tree-structured knowledge base, performs full-text retrieval and large-model semantic retrieval in parallel for the received search query terms, and obtains full-text retrieval results and semantic retrieval results respectively. The weighted fusion module is used to weight and fuse the obtained full-text search results and semantic search results to obtain the final fused search results. The search results output module is used to output the final fused search results to the user, completing the search response for content related to high-end equipment auxiliary design.

[0014] According to another embodiment of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the high-end equipment auxiliary design document management method.

[0015] According to another embodiment of this application, a storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the high-end equipment auxiliary design document management method.

[0016] As can be seen from the above technical solutions, the present invention has the following advantages: This invention provides a high-end equipment auxiliary design document management method that improves the efficiency of drawing information conversion by extracting technical symbols and professional terms from drawings. It achieves deep semantic understanding of documents based on a large language model, ensuring the completeness of information extraction such as themes and entities; and resolves key entity ambiguities through LoRa fine-tuning. Domain division is performed using a mechanical design domain ontology, allowing documents to be categorized by professional domain, improving retrieval efficiency. A tree-structured knowledge base provides clear knowledge hierarchy, and a graph database storage enables node association and cross-document entity links, improving reusability. Parallel execution of full-text search and semantic search covers keyword matching and semantic similarity matching, improving retrieval recall. Weighted fusion optimizes weights and supports dynamic weight adjustment when a single search type fails, improving the relevance of the fused results. It directly outputs search results with document paths and domains, supporting targeted responses for scenarios such as cooperation agreement parsing and model matching, reducing user operation steps and improving ease of use. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 Flowchart for the management of auxiliary design documents for high-end equipment; Figure 2 Schematic diagram of a document management system for assisting in the design of high-end equipment; Figure 3 This is a schematic diagram of an electronic device. Detailed Implementation

[0019] This invention provides a high-end equipment auxiliary design document management method that parses unstructured data such as unified regulations, calculation sheets, product manuals, and model numbers involved in mechanical design. It employs CRNN+Attention dual-engine OCR technology, specifically optimizing technical symbols and professional terms in CAD drawings to identify corresponding entity content. For the document portion, it extracts the theme, key entities, main content, and outline of unstructured documents, generating corresponding content tags. Based on the tag data, a large model is used to divide the data into domains, with different domains containing related tags and documents on different themes, collectively constructing a tree-like knowledge base. Optionally, the retrieval part uses Elasticsearch full-text search and DeepSeek-R1 large-model semantic retrieval, employing the ReciprocalRankFusion (RRF) algorithm for hybrid ranking, enabling rapid searching of relevant design content. This improves the speed of parsing cooperation agreements, retrieving design documents, and matching similar models, shortening the mechanical design R&D cycle.

[0020] The following describes in detail the high-end equipment auxiliary design document management method involved in this application. Specific details such as particular system structures and technologies are presented for illustrative purposes rather than limiting, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details.

[0021] It should be understood that, when used in this specification, terms include indicating the presence of a described feature, integral, step, operation, element, and / or component, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms include, encompass, have, and variations thereof mean including but not limited to, unless otherwise specifically emphasized.

[0022] The statements such as "one embodiment" or "some embodiments" described in this application mean that one or more embodiments of this application include the specific features, structures, or characteristics described in that embodiment. Therefore, the statements such as "in one embodiment," "in some embodiments," "in other embodiments," and "in still other embodiments" in this application do not necessarily refer to the same embodiment, but rather mean one or more, but not all, embodiments, unless otherwise specifically emphasized.

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 The diagram shows a flowchart of a method for managing auxiliary design documents for high-end equipment in a specific embodiment. The method includes: Step S101: Process the drawings in the auxiliary design of high-end equipment and extract the technical symbols and professional terms from the drawings.

[0025] In some embodiments, an AutoCAD format parsing library is used to read CAD drawing files for high-end equipment auxiliary design, obtaining information such as the layer structure, lines, annotations, and text boxes of the drawings. Based on OCR technology, features are extracted from common technical symbols in the drawings, and an attention mechanism is used to focus on densely annotated areas. Technical terminology text in the drawings is identified, and the extracted technical symbols are classified according to symbol type and parameter value format. Technical terms are organized and stored according to terminology name and attribute format, with each record associated with the storage path and unique ID of the corresponding drawing.

[0026] Step S102: Perform semantic understanding on the documents in the auxiliary design of high-end equipment, extract the document's theme, key entities, outline and main content, generate tags, and achieve key entity disambiguation through LoRa fine-tuning.

[0027] In some embodiments, Apache POI is used to parse Word documents and Poppler to parse PDF documents, extracting text content and paragraph levels. The text is then segmented into paragraphs, and the data is input into the model to extract the topic, key entities, design basis, parameter calculations, verification data, and content summaries. A dataset of entity ambiguities in the mechanical design domain is constructed. The LoRa fine-tuning method is used to freeze the main parameters of the model, updating only the low-rank matrix of the attention layer. A learning rate of 5e-5 and 3 training epochs are set to disambiguate the extracted key entities. Structured labels are generated according to domain category, topic, disambiguated entity, and outline format, and these labels are associated with document IDs and storage paths.

[0028] This embodiment, based on LoRa fine-tuning, adapts to the mechanical design field with a small number of trainable parameters, ensuring the unique identifier of the same entity, converting unstructured documents into structured tags containing unique entities, and establishing the association between documents and tags.

[0029] Step S103: Tags. Combining the ontology library in the mechanical design field, the document content corresponding to the tags is divided into domains to obtain multiple different domains.

[0030] In some embodiments, using a mechanical design domain ontology as the standard, automated domain classification of tags is achieved by matching latent domain keywords with core domain keywords. For special tags that fail to match, manual review and correction are introduced to ensure that the domain division conforms to actual design scenarios. Finally, the relationship between domains, tags, and documents is established. The top-level domain nodes of the three-level knowledge base provide direct data, reducing the complexity of knowledge base construction.

[0031] Step S104: Based on technical symbols, professional terms, tags and domains, construct a three-level tree-structured knowledge base, use a graph database to store the relationships between nodes at each level in the knowledge base, and implement cross-document entity links.

[0032] In some embodiments, a node structure of domains, topics, entities / technical symbols is constructed in the graph database, and directional relationships between nodes are established through association fields to form a tree-like knowledge network; cross-document entity links are matched by the same entity ID to break the isolation of knowledge in a single document; index creation optimizes retrieval efficiency and ensures that the knowledge base can be accessed efficiently.

[0033] This three-level tree structure clearly defines the knowledge hierarchy, facilitating user understanding and retrieval. The graph database's relational storage supports cross-document knowledge association, allowing users to retrieve all related document information when searching for a specific entity; index creation reduces retrieval response time from seconds to milliseconds, improving the efficiency of subsequent retrieval processes.

[0034] Step S105: Receive search query terms submitted by the user for high-end equipment auxiliary design; the search query terms include query content corresponding to cooperation agreement parsing, design document retrieval, and matching requirements of similar models.

[0035] In some embodiments, user query terms are received via a web search interface. After selecting a scenario, the user enters query terms. The input query terms are preprocessed, retaining technical terms and recording the scenario identifier corresponding to the query terms. The query ID, query terms, scenario identifier, and reception time are stored in a MySQL search request table. The search request table records query information, enabling full-process traceability.

[0036] Step S106: Based on the three-level tree-structured knowledge base, perform full-text search and large-model semantic search in parallel for the received search query terms, and obtain full-text search results and semantic search results respectively.

[0037] In some embodiments, full-text search is based on keyword matching to quickly locate knowledge nodes containing query terms; semantic search is based on vector similarity to capture the semantic association between query terms and knowledge, covering scenarios where keywords do not match but are semantically similar. Celery task scheduling enables parallel execution of dual searches, avoiding the cumulative time consumption of serial searches and improving search speed; finally, two types of search results are output respectively, with semantic search addressing the needs of semantically similar queries, and the two types of searches complementing each other to improve result coverage.

[0038] Step S107: Weight and fuse the obtained full-text search results and semantic search results to obtain the final fused search results.

[0039] In some embodiments, weight configuration is verified based on actual retrieval scenarios to ensure the rationality of the fusion logic. Normalization eliminates differences between different scoring scales, making the weighted calculation objective; group calculation enables information integration of multiple retrieval results for a single document; anomaly detection handles special scenarios; low-relevance results are removed to improve quality, thereby enhancing the usability of the final results and user browsing efficiency.

[0040] Step S108: Output the final fusion search results to the user to complete the search response for high-end equipment auxiliary design related content.

[0041] In some embodiments, the final result set is converted into a user-readable interface display format, supplementing it with preview, download, and other functions to improve ease of use. Related document recommendations, based on cross-document links in the knowledge base, extend the scope of knowledge available to users and reduce redundant searches. User behavior logging and empty result prompts enable continuous system optimization, improving search accuracy and user satisfaction.

[0042] In one embodiment of the present invention, based on step S102, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. Step S102 specifically includes the following steps: Step S1021: Use a document parsing library to process the document, remove format tags, headers, footers and other non-body content, and obtain a text content sequence.

[0043] In some embodiments, for a document, the parser analyzes the file's object tree, distinguishes between the text stream, path, and metadata, and extracts the text stream sequentially. Optionally, the parser can also extract text content based on parsing XML information.

[0044] Step S1022: Input the transformed text content sequence into the large language model. The input sequence is encoded using the multi-layer Transformer network structure within the large language model to generate a deep representation vector containing contextual semantic information.

[0045] The large language model can use DeepSeek for document semantic understanding, extracting document topics and key entities. A Sentence-BERT encoder is used to generate semantic vectors for queries and documents, supporting high-precision semantic matching. ChatGLM3-6B, combined with an ontology library based on mechanical design, assists in determining the association rules between tags and domains. ERNIE 4.0, combined with multimodal capabilities, assists in recognizing technical symbols in CAD drawings, supplementing the semantic representation of symbols recognized by OCR.

[0046] In some embodiments, after the text is segmented into tokens, each token is converted into a vector. The Transformer encoder consists of multiple layers, each performing self-attention and feedforward neural network operations. The self-attention mechanism dynamically aggregates global information by calculating the association weights between each token in a sequence and all other tokens. This allows the model to fully consider the influence of all words in its context when encoding a word, thereby generating a deeply understood, context-sensitive representation vector.

[0047] Step S1023: Based on the generated deep representation vector, configure the entity recognition and classification head in the model output layer. The recognition head traverses the text sequence, identifies and extracts entities belonging to a preset category set, which includes material type, process parameters, and component name.

[0048] In some embodiments, a linear classification layer is superimposed on the deep representation vector of each token output by the large language model. The classification layer determines the entity type of each token. Optionally, transition rules between entity labels are learned to globally optimize the entire output label sequence, ensuring the rationality and consistency of the results. This enables the accurate location and classification of key technical information items from massive amounts of text, achieving rapid information location and computer-processable processing.

[0049] Step S1024: Configure text summarization and topic retrieval identifiers in the output layer of the large language model; The topic retrieval identifier is based on deep representation vectors, which generate topic tags that describe the core ideas of the document and extract the main chapter titles that constitute the logical structure of the document to form a document outline.

[0050] In some embodiments, topic extraction can be performed by calculating the similarity between sentence vectors and the overall document vector, using these as topics. Outline extraction can identify each sentence. The model accomplishes this by learning typical sentence structure, position, and formatting features of headings.

[0051] Step S1025: Using a LoRa-based low-rank adaptive method, the large language model is tuned using a dataset labeled with entity disambiguation data from the mechanical domain. By adjusting the parameters of the large language model, it is enabled to map key entities with the same glyphs but different meanings to unique, unambiguous identifiers based on context.

[0052] In some embodiments, LoRA (Low-Rank Adaptation of Large Language Models) freezes the original weights of the pre-trained large model without adjusting the full set of parameters. During forward propagation, the original output is augmented with the output of the low-rank adapter. In this way, the model trains a large number of parameters, aligning and transforming the knowledge of those parameters with the mechanical design domain.

[0053] For example, the model learns that a nut in a specific context refers to a standard part model with a specific specification, rather than an object from everyday life. In this way, by analyzing the context, the model correctly identifies the nut in the sentence "Please select GB / T 6170 M12 nut" as a standard part model, and distinguishes it from the nut in the sentence "This project is the most critical nut".

[0054] It also teaches the model that H7 refers to a tolerance zone designation in a specific context, rather than a simple alphanumeric combination. By analyzing the context, the model correctly identifies H7 in the sentence "axis diameter φ50H7" as a tolerance zone designation and distinguishes it from the H in the sentence "room height 2.7H".

[0055] This embodiment enables general data pre-trained models to acquire domain-specific knowledge and terminology, thereby improving the accuracy of entity disambiguation and domain semantic understanding.

[0056] In one embodiment of the present invention, based on step S103, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S103 specifically includes the following steps: Step S1031: Compare the tags generated in S102 with the ontology library in the field of mechanical design to identify the technical terms and concepts in the tags.

[0057] In some embodiments, the tag text is converted into a vector representation, and then cosine similarity is calculated between it and the term vectors in the domain ontology library. This can accurately identify the specialized information in the tags, providing a precise semantic basis for subsequent domain partitioning.

[0058] Step S1032: Use a model clustering algorithm to group the labels, and group semantically similar labels into the same group.

[0059] In some embodiments, semantic similarity between tags is calculated, and then the tags are grouped into different groups based on the similarity. Clustering algorithms achieve optimal grouping results by continuously optimizing intra-cluster similarity and inter-cluster dissimilarity. This can automatically group semantically related tags together, improving the efficiency and accuracy of domain partitioning.

[0060] Step S1033: Based on the clustering results and the hierarchical relationships in the domain ontology library, determine the domain corresponding to each group.

[0061] In some embodiments, clustered label groups are matched with nodes in the ontology library, and path analysis and semantic similarity calculation are used to determine the domain corresponding to the label groups. This leverages the domain ontology library to ensure the rationality of domain partitioning and its compliance with domain specifications.

[0062] Step S1034: Perform semantic association analysis on the tags in each domain to construct a semantic network of tags within the domain.

[0063] In some embodiments, by calculating the co-occurrence frequency and semantic relevance between tags and constructing weighted edges between tags, the potential semantic connections between tags can be revealed.

[0064] Step S1035: Map the constructed semantic network to the graph database and store the association between domains and labels.

[0065] In some embodiments, domains and labels are used as nodes in a graph database, semantic relationships are stored as edges, and an inverted index is created for fast querying.

[0066] In one embodiment of the present invention, based on step S104, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S104 specifically includes the following steps: Step S1041: Based on the multiple domains obtained from the domain partitioning in step S103, create corresponding domain nodes in the graph database and assign a unique identifier and attributes to each domain node.

[0067] In some embodiments, the system reads the output of S103 and creates a node in the graph database for each partitioned domain. The node is configured as a data container and can have attributes attached. These attributes give the domain node semantic information, facilitating querying.

[0068] Step S1042: For each domain, create a topic node in the graph database based on the topic tags corresponding to the documents contained therein, and establish parent-child relationship edges between the topic node and the domain node to which it belongs.

[0069] In some embodiments, after the domain nodes are created, a knowledge subtree is built under each domain. All documents belonging to that domain are traversed, and their topic tags are collected. For each unique topic, a topic node is created, and a directed edge is created from that topic node to its parent domain node. In this way, by traversing and querying the graph database, all topics under a given domain can be found, and the corresponding domains for those topics can be determined.

[0070] Step S1043: Create entity nodes in the graph database using the technical symbols and professional terms extracted in S101 and the key entities extracted in S102, and establish parent-child relationship edges between the entity nodes and their respective topic nodes.

[0071] In some embodiments, technical symbols, technical terms, and key entities extracted from drawings and documents are created as entity nodes.

[0072] Optionally, the technical symbol can be H7 / g6, the technical term can be quenching treatment, and the key entity can be hydraulic cylinder or 45# steel.

[0073] Based on the semantic context of the entity node in the document, the entity node is associated with the corresponding topic node. For example, the entity node "gear module" will establish a relationship edge with the topic node "gear design".

[0074] Step S1044: Store the document itself as a document node in the graph database, establish containment edges between the document node and the entity node describing its content, and associate the document node with its respective topic node and domain node.

[0075] In some embodiments, the metadata of the original design document is created as document nodes. Relationship edges are established. Specifically, this may involve connecting document nodes to their respective domain nodes and topic nodes. Document nodes are also connected to all entity nodes they contain. In this way, the document is fully integrated into the knowledge graph. Through document nodes, the domains, topics, and entities of all nodes are associated with specific knowledge sources. When a user retrieves any entity or topic, they can locate all the original documents it contains, ensuring knowledge traceability.

[0076] Step S1045: By traversing the entity nodes in the graph database, identify different entity nodes that refer to the same physical object, and establish cross-document link edges for different entity nodes, thereby realizing cross-document entity links.

[0077] In some embodiments, all entity nodes are traversed to find entities that refer to the same object but may appear in different documents under different names or in different forms. This constructs an interconnected knowledge network. By querying any entity, users can discover all related information scattered across different documents, improving knowledge discovery capabilities and overall utilization. For example, querying a model number can simultaneously find all relevant parameters in related design manuals, calculation sheets, and cooperation agreements.

[0078] Step S1045 of the present invention further includes the following steps: S10451: Filter the set of nodes with the same entity type attribute from the graph database. For example, filter all nodes whose entity type is hydraulic cylinder or material type.

[0079] S10452: For each selected entity node, extract its stored attribute information and associated semantic fragment content.

[0080] In some embodiments, the values ​​of attribute fields such as model and material are read from each entity node, as well as the semantic fragment content associated with the node through edges.

[0081] S10453: Concatenate the attribute information of each entity node with the semantic fragment content to generate the feature vector of that node.

[0082] In some embodiments, the attribute information of each node is concatenated with semantic fragment content to form a complete description string, which is then input into a pre-trained model to generate a fixed-dimensional feature vector. This transforms textual node information into a computer-processable vector form, capturing the semantic features of the nodes through vector representation.

[0083] S10454: Calculate the similarity of feature vectors of different nodes under the same entity type, and find node pairs whose similarity exceeds a preset threshold.

[0084] In some embodiments, vector similarity is used to quantify the semantic similarity of nodes. The higher the similarity, the more likely the two nodes represent the same physical concept. This quantitative indicator objectively judges entity identity, improving the accuracy of links.

[0085] S10455: For node pairs with similarity exceeding the threshold, establish cross-document link edges and unify the identifier of one of the nodes.

[0086] In some embodiments, different nodes of the same entity are connected by graph edges, and a unified identifier is used to eliminate entity ambiguity, ensuring that the same entity is unique in the knowledge base. This enables cross-document entity association, returns the same entity only once during retrieval, avoids duplicate results, eliminates inconsistencies in entity descriptions, and improves the consistency of the knowledge base.

[0087] In one embodiment of the present invention, based on step S106, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. Step S106 specifically includes the following steps: S1061: Map user search queries to the hierarchical structure of a multi-level tree-like knowledge base, and extract domain, topic, or entity keywords associated with the query terms.

[0088] In some embodiments, the semantics of user query terms are analyzed and mapped to domains, topics, and entities in the knowledge base. This transforms the user's natural language requirements into specific hierarchical locations within the knowledge base, clarifying the target scope of the retrieval.

[0089] S1062: Call the full-text search function of the graph database, match the extracted keywords in node attributes and edge relationships, and obtain nodes and associated paths that meet the conditions.

[0090] In some embodiments, the full-text search capabilities of graph databases are utilized to locate nodes and relational paths containing query keywords, supporting full-text search requirements based on keyword matching. Relying on the relational data of the graph structure, the search results inherently possess contextual relationships. S1063: Input the user's search query terms into the large language model to generate semantic vectors for the query terms.

[0091] S1064: Call the semantic retrieval function of the large language model, compare the query semantic vector with the semantic vector of the knowledge base node, and obtain semantically similar nodes and related paths.

[0092] In some embodiments, the semantic vectors of all nodes in the knowledge base are pre-imported into the vector database; using the query vector obtained in S1063, the nodes with the highest similarity are retrieved in the vector database, and the paths of these nodes in the graph database are associated to capture semantic-level associations and improve the retrieval recall rate.

[0093] S1065: Organize the nodes and paths obtained from full-text search and semantic search respectively, and output them as full-text search results and semantic search results.

[0094] In some embodiments, the full-text search results of S1062 are organized into a list of node IDs, associated paths, and keyword matching scores; the semantic search results of S1064 are organized into a list of node IDs, associated paths, and semantic similarity scores. Separating the results of the two search methods provides input for the weighted fusion step. The full-text search results reflect the keyword matching degree, while the semantic search results reflect the semantic relevance, thus improving the accuracy of the fusion.

[0095] In one embodiment of the present invention, based on step S107, the following is a possible embodiment and its specific implementation will be described in a non-limiting manner. S107 specifically includes the following steps: Step S1071: Receive a list of search results returned by the full-text search engine and the semantic search module. Each result contains a document identifier and its original relevance score.

[0096] In some embodiments, the outputs of two retrieval channels are received in parallel: a list of full-text search results from Elasticsearch, where each document has a relevance score calculated using the BM25 algorithm; and a list of semantic search results from a vector database, where each document has a semantic similarity score calculated using cosine similarity, optionally ranging from -1 to 1, or from 0 to 1, with scores closer to 1 indicating greater similarity. Paired document identifiers are then created, and the original scores are temporarily stored.

[0097] Step S1072: Calculate a ranking-based score for each result based on its position in the original search list for both the full-text search result and the semantic search result.

[0098] In some embodiments, since the original scoring systems for full-text search and semantic search are completely different, direct weighted summation cannot be implemented. This embodiment uses the ranking position of the results as the metric. For each document in the results list, a ranking score is calculated based on its ranking position. Optionally, the reciprocal of the ranking is calculated, i.e., ranking score = 1 / ranking position. For example, the document ranked 1st has a score of 1 / 1 = 1.0, the document ranked 2nd has a score of 1 / 2 = 0.5, and so on.

[0099] The key prerequisite for effective integration is to unify the correlation measures from different sources and with different dimensions into comparable scores based on ranking position.

[0100] Step S1073: Normalize the ranking-based scores to eliminate the order-of-magnitude differences caused by the different original scoring systems.

[0101] Step S1074: Read the preset full-text search weight coefficient and semantic search weight coefficient, and perform a weighted calculation on the normalized ranking score and the corresponding weight coefficient to obtain a fusion score for each search result.

[0102] In some embodiments, two pre-determined weighting coefficients, optimized through extensive experiments, are read from a configuration file or database: a full-text search weight α and a semantic search weight β, for example, α=0.6, β=0.4, and α+β=1. Then, for each document that appears in both search result lists, its final fusion score is calculated as: Fusion Score = α * Full-text Search Normalized Score + β * Semantic Search Normalized Score. For documents that appear in only one list, only the weighted value of the channel for which a score is given is calculated.

[0103] Configurable weights control the influence of the two retrieval strategies on the final results. Adjustments can be made based on the specific application scenario; for example, increasing the weight of full-text retrieval when high-precision matching is needed, and increasing the weight of semantic retrieval when semantic expansion is required, thereby achieving the optimal overall retrieval effect.

[0104] Step S1075: Based on the calculated fusion score, sort all the search results from different channels in descending order to generate the final fusion search result list.

[0105] In some embodiments, after calculating the fusion score of all candidate documents, the documents and their fusion scores are treated as a whole set and sorted in descending order of fusion score using an efficient ranking algorithm. After sorting, an ordered list of search results is generated. This list combines the accuracy of keyword matching with the intelligence of semantic matching; the documents at the top are considered the best answers that simultaneously meet both search criteria, improving user search efficiency and experience.

[0106] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0107] The following are embodiments of the high-end equipment auxiliary design document management system provided in this disclosure. This system and the high-end equipment auxiliary design document management method of the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the high-end equipment auxiliary design document management system, please refer to the embodiments of the high-end equipment auxiliary design document management method.

[0108] like Figure 2 As shown, the system includes: The drawing information extraction module 201 is used to process drawings in the auxiliary design of high-end equipment and extract technical symbols and professional terms from the drawings; Semantic understanding module 202 is used to perform semantic understanding on documents in high-end equipment auxiliary design, extract the document's theme, key entities, outline and main content, generate tags, and achieve key entity disambiguation through LoRa fine-tuning; The domain partitioning module 203, based on the generated tags and combined with the ontology library of the mechanical design domain, partitions the document content corresponding to the tags into multiple different domains; The knowledge base construction module 204 constructs a three-level tree-structured knowledge base based on technical symbols, professional terms, tags, and domains. It uses a graph database to store the relationships between nodes at each level of the knowledge base and enables cross-document entity links. The term receiving module 205 is used to receive the term search submitted by the user for the auxiliary design of high-end equipment. The hybrid retrieval module 206, based on a three-level tree-structured knowledge base, performs full-text retrieval and large-model semantic retrieval in parallel for the received retrieval query terms, and obtains full-text retrieval results and semantic retrieval results respectively. The weighted fusion module 207 is used to perform weighted fusion on the obtained full-text search results and semantic search results to obtain the final fusion search results. The search result output module 208 is used to output the final fused search results to the user, completing the search response for content related to high-end equipment auxiliary design.

[0109] like Figure 3 As shown, this application also provides an electronic device, including a display module 103, a memory 102, a processor 101, a communication module 104, and a computer program stored in the memory and executable on the processor 101. When the processor 101 executes the program, it implements the steps of the high-end equipment auxiliary design document management method.

[0110] In embodiments of the present invention, electronic devices include, but are not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments described and / or claimed herein.

[0111] In this embodiment, processor 101 may be implemented using at least one of an application-specific integrated circuit, a programmable logic device, a field-programmable gate array, a processor, a controller, a microcontroller, a microprocessor, or an electronic unit designed to perform the functions described herein. In some cases, such an implementation may be implemented within a controller. For software implementation, implementations such as processes or functions may be implemented with separate software modules that allow the performance of at least one function or operation. Software code may be implemented by a software application (or program) written in any suitable programming language, and the software code may be stored in memory and executed by the controller.

[0112] The display module 103 is used to display information input by the user or information provided to the user. The display module 103 may include a display panel, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like.

[0113] The memory 102 can be used to store software programs and various data. The memory 102 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0114] The communication module 104 transmits radio signals to and / or receives radio signals from at least one of a base station, an external terminal, and a server. Such radio signals may include voice call signals, video call signals, or various types of data sent and / or received according to text and / or multimedia messages.

[0115] The present invention also provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the high-end equipment auxiliary design document management method.

[0116] The storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0117] The storage medium stores a program product capable of implementing the methods described above in this specification. In some possible implementations, various aspects of this disclosure can also be implemented as a program product comprising program code that, when run on a terminal device, causes the terminal device to perform the steps described in the exemplary methods section of this specification according to various exemplary embodiments of this disclosure.

[0118] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for managing auxiliary design documents for high-end equipment, characterized in that, The methods include: Step S101: Process the drawings in the auxiliary design of high-end equipment and extract the technical symbols and professional terms from the drawings; Step S102: Perform semantic understanding on the documents in the auxiliary design of high-end equipment, extract the document's theme, key entities, outline and main content, generate tags, and achieve key entity disambiguation through LoRa fine-tuning; Step S103: Based on the generated tags, and combined with the ontology library in the mechanical design field, the document content corresponding to the tags is divided into domains to obtain multiple different domains; Step S104: Based on technical symbols, professional terms, tags and domains, construct a three-level tree-structured knowledge base, use a graph database to store the relationships between nodes at each level in the knowledge base, and implement cross-document entity links; Step S105: Receive search query terms submitted by the user for the auxiliary design of high-end equipment; Step S106: Based on the three-level tree-structured knowledge base, for the received search query terms, perform full-text search and large-model semantic search in parallel to obtain full-text search results and semantic search results respectively; Step S107: Perform weighted fusion of the obtained full-text search results and semantic search results to obtain the final fusion search results; Step S108: Output the final fusion search results to the user to complete the search response for high-end equipment auxiliary design related content.

2. The method for managing high-end equipment auxiliary design documents according to claim 1, characterized in that, S102 specifically includes the following steps: Use a document parsing library to process the document, remove format tags, headers, footers and non-body content, and obtain the text content sequence; The transformed text content sequence is input into the large language model. The input sequence is encoded through the multi-layer Transformer network structure inside the large language model to generate a deep representation vector containing contextual semantic information. Based on the generated deep representation vectors, entity recognition and classification heads are configured in the model output layer; The recognition head traverses the text sequence, identifies and extracts entities belonging to a preset category set, which includes material type, process parameters, and component name; Configure text summarization and topic retrieval flags in the output layer of the large language model; By using a low-rank adaptive method based on LoRa, a large language model is adjusted using a dataset containing disambiguation data of entities in the mechanical field. By adjusting the parameters of the large language model, the model can map key entities with the same glyphs but different meanings to labels according to the context.

3. The method for managing high-end equipment auxiliary design documents according to claim 1, characterized in that, S103 specifically includes the following steps: The tags generated in S102 are compared with the ontology library in the field of mechanical design to identify the professional terms and concepts in the tags; The model clustering algorithm is used to group the tags, and the tags with similar semantics are grouped into the same group; Based on the clustering results and the hierarchical relationships in the domain ontology library, the domain corresponding to each group is determined. Perform semantic association analysis on the tags within each domain to construct a semantic network of the tags within the domain; The constructed semantic network is mapped to a graph database to store the association between domains and tags.

4. The method for managing high-end equipment auxiliary design documents according to claim 1, characterized in that, S104 specifically includes the following steps: Based on the multiple domains obtained from the domain partitioning in step S103, corresponding domain nodes are created in the graph database, and a unique identifier and attribute are assigned to each domain node. For each domain, based on the topic tags corresponding to the documents contained therein, create topic nodes in the graph database and establish parent-child relationship edges between these topic nodes and their respective domain nodes; The technical symbols and professional terms extracted in step S101 and the key entities extracted in step S102 are created as entity nodes in the graph database, and parent-child relationship edges are established between the entity nodes and their respective topic nodes. Store the document itself as a document node in the graph database, establish containment edges between the document node and the entity node describing the content, and associate the document node with its respective topic node and domain node. By traversing the entity nodes in the graph database, different entity nodes referring to the same physical object are identified, and cross-document link edges are established for different entity nodes to achieve cross-document entity links.

5. The method for managing auxiliary design documents for high-end equipment according to claim 4, characterized in that, S1045 specifically includes the following steps: Filter the set of nodes with the same entity type attribute from the graph database; For each selected entity node, extract its stored attribute information and associated semantic fragment content; The attribute information of each entity node is concatenated with the semantic fragment content to generate the node's feature vector; Calculate the similarity of feature vectors of different nodes under the same entity type, and find node pairs whose similarity exceeds a preset threshold; For node pairs with similarity exceeding a threshold, establish cross-document link edges and unify the identifier of one of the nodes.

6. The method for managing auxiliary design documents for high-end equipment according to claim 1, characterized in that, S106 specifically includes the following steps: Map user search queries to the hierarchical structure of a multi-level tree-like knowledge base, and extract domain, topic, or entity keywords associated with the query terms; The full-text search function of the graph database is invoked to match the extracted keywords in node attributes and edge relationships to obtain nodes and associated paths that meet the conditions. Input the user's search query into the large language model to generate semantic vectors for the query; The semantic retrieval function of the large language model is invoked to compare the semantic vector of the query with the semantic vector of the knowledge base node to obtain semantically similar nodes and related paths; The nodes and paths obtained from full-text search and semantic search are organized separately and output as full-text search results and semantic search results, respectively.

7. The method for managing high-end equipment auxiliary design documents according to claim 1, characterized in that, S107 specifically includes the following steps: Receive a list of search results returned by the full-text search engine and semantic search module. Each result contains a document identifier and its original relevance score. Based on the position of each result in the original search list, a ranking-based score is calculated for both the full-text search results and the semantic search results. Normalize the ranking-based scores; Read the preset full-text search weight coefficients and semantic search weight coefficients, and calculate the weighted average of the normalized ranking score and the corresponding weight coefficients to obtain a fusion score for each search result; Based on the calculated fusion score, all search results from different channels are sorted in descending order to generate the final fusion search result list.

8. A high-end equipment auxiliary design document management system, characterized in that, The system is used to implement the high-end equipment auxiliary design document management method as described in any one of claims 1 to 7; The system includes: The drawing information extraction module is used to process drawings in the auxiliary design of high-end equipment and extract technical symbols and professional terms from the drawings. The semantic understanding module is used to perform semantic understanding on documents in the auxiliary design of high-end equipment, extract the document's theme, key entities, outline and main content, generate tags, and achieve key entity disambiguation through LoRa fine-tuning. The domain partitioning module, based on the generated tags and combined with the ontology library of the mechanical design domain, partitions the document content corresponding to the tags into multiple different domains; The knowledge base construction module builds a three-level tree-structured knowledge base based on technical symbols, professional terms, tags, and domains. It uses a graph database to store the relationships between nodes at each level of the knowledge base and enables cross-document entity links. The search term receiving module is used to receive search query terms submitted by users for the auxiliary design of high-end equipment; The hybrid retrieval module, based on a three-level tree-structured knowledge base, performs full-text retrieval and large-model semantic retrieval in parallel for the received search query terms, and obtains full-text retrieval results and semantic retrieval results respectively. The weighted fusion module is used to weight and fuse the obtained full-text search results and semantic search results to obtain the final fused search results. The search results output module is used to output the final fused search results to the user, completing the search response for content related to high-end equipment auxiliary design.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the high-end equipment auxiliary design document management method as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the high-end equipment auxiliary design document management method as described in any one of claims 1 to 7.

Images