Artificial intelligence-based electrical equipment key parameter and product matching method and related equipment

Abstract

The application discloses an electrical equipment key parameter and product matching method based on artificial intelligence and related equipment, and the method first performs OCR identification on common text by using a PP-Structure v3, and introduces a GLM-4-9B-VL model to specially analyze complex tables in the document; then a Qwen3-Embedding-4B embedding model is used to vectorize the processed text; subsequently, an optimized RAG framework is constructed by using a keyword and vector database comprehensive retrieval mode; finally, a Qwen3-8B large language model is used to perform intelligent matching and recommendation in a product library based on the retrieved key parameter information. The method solves the problems of complex text structure of technical protocols and general technical documents, easy missing reading and misreading of key parameters, and insufficient complex table analysis capability of an OCR technology in a traditional RAG framework in an electrical product engineering design process.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of electrical equipment manufacturing and artificial intelligence, and relates to a method and related equipment for matching key parameters of electrical equipment with products. Specifically, it is a method and application for identifying key parameters of electrical equipment and matching products based on artificial intelligence. Background Technology

[0002] In the process of engineering design of electrical equipment products, designers need to extract the key performance indicators and parameter requirements of the product based on the general technical specifications and special technical agreement documents provided by the product demander, and match the product specifications that best meet the requirements among the existing product models of the enterprise, and then carry out targeted engineering improvement design. This process has the following problems: (1) Parameter extraction is time-consuming. Usually, there are a large number of parameters proposed for the product, and they are presented in the form of tables in the agreement. The relevant tables usually contain multiple rows and columns, and the data nesting relationship is relatively complex. Designers have to spend a lot of time reading and analyzing all the values ​​one by one and recording them manually; (2) Parameter extraction has high requirements. Usually, the parameters proposed for the product involve multiple professional fields such as electrical, mechanical, materials, environment, and industrial design, which requires high business capabilities from designers. It is impossible for a single person to fully understand so many professional terms and connotations; (3) Parameter extraction is prone to errors. Technical agreements and specification documents usually contain a lot of text, tables and picture information. When designers follow up on multiple projects at the same time, due to differences in personal energy and ability, parameter omissions and misreadings are easy to occur, leading to subsequent design errors, waste of raw materials and cost losses.

[0003] With the rapid development of artificial intelligence (AI) technology in recent years, AI can assist technicians in identifying key parameters and matching products. However, existing AI technologies generally employ the Retrieval Augmentation (RAG) framework, but the Optical Character Recognition (OCR) models they rely on often suffer from character recognition errors and chaotic table structure parsing when processing complex tables. This leads to the inability to accurately extract key parameters from the tables in subsequent retrieval stages, severely impacting the accuracy of business decisions. Furthermore, standard RAG frameworks struggle with inaccurate retrieval recall, contextual information overload, or missing information when dealing with large amounts of specialized technical documents, affecting the efficiency and accuracy of subsequent parameter extraction and product matching. Moreover, general-purpose large language models lack expertise in the electrical equipment field, and when handling highly specialized domain knowledge, they are prone to generating "illusions" or content that does not conform to industry standards, resulting in poor reliability for direct application. Blindly adopting ultra-large models or full-scale fine-tuning to improve accuracy leads to enormous computational resource consumption and slow response speeds, making it difficult to simultaneously meet the requirements of high performance, low cost, and high efficiency in engineering practice. Therefore, how to improve the accuracy and response speed of parameter extraction while keeping the model parameter scale relatively small, and achieve an effective balance between performance and resource consumption, has become a pressing problem for the industry to solve, given the aforementioned application scenarios and existing problems. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method and related equipment for matching key parameters of electrical equipment with products. Addressing the characteristic of technical documents containing both plain text and complex tables with significant structural differences, it innovatively employs a parsing scheme that integrates "PP-Structure v3 with GLM-4-9B-VL," achieving lossless extraction of text information and reconstruction of the logical structure of complex tables. To address the issue of insufficient retrieval accuracy in specialized fields, a hybrid enhanced retrieval strategy is adopted to ensure the comprehensiveness and high relevance of contextual information. Finally, precise information extraction and matching are performed using a large language model with a moderate number of parameters, Qwen3-8B, achieving engineering-level accuracy and efficiency with limited computing power.

[0005] This invention is achieved through the following technical solution: A method for matching key parameters of electrical equipment with products based on artificial intelligence, comprising: Obtain multi-source technical documents; For multi-source technical documents, the PP-Structure v3 model and the GLM-4-9B-VL model are used to extract ordinary text information and nested data of tables, respectively. The Qwen3-Embedding-4B embedding model is used to convert plain text information and nested data of tables into vectors and store them in a vector database to build a knowledge base. Based on a pre-defined keyword database and structured question templates in the field of electrical equipment, a hybrid retrieval strategy is used to recall highly relevant text fragments from the knowledge base. The text fragments and structured question templates are input into the Qwen3-8B large language model to extract key parameters; The key parameters are matched with the product specifications in the product database using a decision tree algorithm, and the matching results are sorted into a candidate product list.

[0006] Preferably, it also includes standardizing the acquired multi-source technical documents, specifically: The documents are cleaned up, watermarks and headers / footers are removed, and the original layout structure is preserved to obtain multi-source technical documents in a unified format; the multi-source technical documents include product technical agreements and industry-standard technical specifications.

[0007] Preferably, key parameters are extracted, specifically: The text fragments and structured question templates are input into the Qwen3-8B large language model to perform accurate identification of key parameters, semantic normalization and numerical extraction. Based on the extraction of key parameters, a structured list containing the key parameters and their values ​​is generated, and a corresponding traceability report is generated. The traceability report indicates the location and original content of each key parameter in the original multi-source technical document.

[0008] Preferably, the extraction of plain text information and nested data from tables is as follows: Based on the PP-Structure v3 model, the layout analysis and optical character recognition of the multi-source technical documents are performed to extract ordinary text content. The complex table includes tables with multi-level headers, merged rows and columns, or nested data structures within cells; the GLM-4-9B-VL model parses the logical nested data of the complex table by understanding the visual features of the table and the semantics of the text.

[0009] Preferably, a knowledge base is constructed, specifically as follows: The Qwen3-Embedding-4B embedding model converts ordinary text information and nested data in tables into high-dimensional vectors, stores the vectors in a vector database, and builds an efficient index, thereby constructing a knowledge base for semantic retrieval.

[0010] Preferably, a hybrid retrieval strategy is used to retrieve highly relevant text fragments from the knowledge base, specifically as follows: Keyword matching is performed based on the keyword database, and vector retrieval is performed based on semantic similarity to obtain preliminary retrieved information fragments. A cross-encoder model is used to re-score the relevance between the preliminary retrieved information fragments and the query, and the most relevant Top-K text fragments are selected based on the scores. The hybrid retrieval strategy includes keyword matching and semantic vector retrieval; The keyword library covers professional fields such as electrical, mechanical and materials; the structured question template includes question templates with specific task instructions, output format requirements and context examples.

[0011] Preferably, key parameters are matched with product specifications in the product database using a decision tree algorithm, specifically: The key parameters are compared with the product specifications in the product database, and the parameter compliance is identified and matched with predefined business rules using a decision tree algorithm. Output a list of candidate products sorted by matching degree, and automatically associate and recommend the corresponding engineering drawings for the candidate product with the highest matching degree; The business rules are defined with parameter weights, priorities, mandatory matching conditions, and flexible tolerance ranges. The database includes multi-dimensional specifications of the company's products and their associated digital asset indexes.

[0012] An artificial intelligence-based system for matching key parameters of electrical equipment with products includes: The information access and preprocessing module acquires and preprocesses technical documents from multiple sources. The information extraction module extracts ordinary text information and nested data from tables from multiple source technical documents using the PP-Structure v3 model and the GLM-4-9B-VL model, respectively. The vector knowledge base construction module uses the Qwen3-Embedding-4B embedding model to convert ordinary text information and nested data of tables into vectors and store them in a vector database to build a knowledge base. The hybrid retrieval and reordering module, based on a pre-set keyword library and structured question templates in the field of electrical equipment, uses a hybrid retrieval strategy to recall highly relevant text fragments from the knowledge base. The key parameter extraction module inputs the text fragments and structured question templates into the Qwen3-8B large language model to extract key parameters; The parameter and product matching module matches key parameters with product specifications in the product database using a decision tree algorithm, and sorts the matching results into a candidate product list.

[0013] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a method and related equipment for matching key parameters of electrical equipment with products. Breaking through the limitations of traditional single OCR technology, this invention achieves differentiated and accurate processing of ordinary text and "complex tables" in technical documents through a parsing architecture that integrates PP-Structure v3 with GLM-4-9B-VL. Simultaneously, a hybrid enhanced retrieval strategy is adopted, balancing recall and precision while ensuring comprehensive and highly relevant context, laying a solid foundation for the accurate generation of large language models. Finally, through fine-tuning of key parameters in the RAG framework and the selection of the powerful Qwen3-8B open-source model with a moderate number of parameters, the invention significantly reduces computational resource consumption and response latency while maintaining extremely high task accuracy. This makes the solution highly practical for engineering applications and cost-effective. This invention realizes the entire process from document parsing and parameter extraction to product identification and drawing screening, forming a complete automated and intelligent business closed loop. It transforms the human role into that of "inspector" and "decision-maker," achieving a highly efficient human-machine collaborative working mode. At the same time, it frees designers from repetitive document interpretation work, allowing them to focus on creative and strategic work, thereby improving the overall efficiency of engineering design. The time spent on the initial product design stage of engineering design can be shortened from several hours to minutes, with an overall efficiency improvement of more than 70%.

[0016] Furthermore, it innovatively integrates the GLM-4-9B-VL model with traditional OCR technology. By introducing the VLM model specifically for parsing complex tables, it effectively solves the industry pain point of insufficient ability of traditional OCR technology to parse complex tables, ensuring the completeness and accuracy of key parameter extraction.

[0017] Furthermore, targeted adjustments were made to the Retrieval Enhancement Generation (RAG) architecture. For the extraction of some important parameters, a pre-defined keyword library was first used to filter and select text blocks highly relevant to the target parameters. Then, a semantic similarity was calculated and ranked using an embedding model. This compensates for the low recall rate of traditional single-vector retrieval.

[0018] Furthermore, it exhibits excellent technical versatility and adaptability. Based on a general large language model, this method effectively solves the problems of key parameter identification and product matching in the engineering design process of electrical equipment without the need for expensive industry-specific fine-tuning. It also possesses good scalability and can be adapted to different sub-sectors of electrical equipment manufacturing. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the 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.

[0020] Figure 1 Flowchart of the method for identifying key parameters of electrical equipment and matching products provided in the embodiments of the present invention; Figure 2 A schematic diagram of the RAG architecture provided in an embodiment of the present invention; Figure 3 The system architecture diagram provided for embodiments of the present invention; Figure 4 The question template containing contextual information is set for Example 1. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] This invention provides an artificial intelligence-based method for identifying key parameters and matching products in electrical equipment. Addressing the challenge of technical documents containing both plain text and complex tables with significant structural differences, it innovatively employs a parsing scheme that integrates PP-Structurev3 with GLM-4-9B-VL, achieving lossless extraction of text information and reconstruction of the logical structure of complex tables. To address the issue of insufficient retrieval accuracy in specialized fields, a hybrid enhanced retrieval strategy combining keyword matching and semantic retrieval is adopted, ensuring the comprehensiveness and high relevance of contextual information. Finally, precise information extraction and matching are performed using the large language model Qwen3-8B with a moderate number of parameters, achieving engineering-level accuracy and efficiency with limited computing power.

[0023] To achieve the above objectives, the present invention adopts the following technical solution: S1: Collection and Standardization of Multi-Source Technical Documents. The system collects product-specific technical agreements, industry-standard technical specifications, and other documents provided by product requesters. This step involves not only document aggregation but also format unification and standardization preprocessing. For example, documents in different formats (such as PDF and DOCX) are converted into a unified format suitable for subsequent parsing modules, laying the foundation for information extraction.

[0024] S2: Plain Text Information Extraction Based on PP-Structure v3. For plain text in the document, the lightweight and efficient PP-Structure v3 model is used to perform optical character recognition (OCR) and layout analysis. This step can quickly and accurately identify and extract text content, achieving efficient digitization of the document's basic information.

[0025] S3: Introduces the GLM-4-9B-VL visual language model for accurate parsing of complex tables. Addressing the challenges of handling tables with complex structures (such as multiple row / column merging and data nesting) in S2 using traditional OCR models, the GLM-4-9B-VL visual language model is specifically introduced. This model can simultaneously understand image visual features and text semantics, thereby accurately parsing the logical structure of complex tables, row / column merging relationships, and nested data within cells, ensuring the integrity and accuracy of key parameter sources.

[0026] S4: Build a vector knowledge base using Qwen3-Embedding-4B. Using the high-performance Qwen3-Embedding-4B embedding model, all text and tabular information extracted in steps S2 and S3 is converted into high-dimensional vector representations. These vectors are then stored in a vector database (such as Milvus or Chroma) and an efficient index is built, thus constructing an enterprise-specific knowledge base capable of rapid semantic retrieval.

[0027] S5: Establish a keyword library and structured question templates for parameter extraction. First, create a keyword library covering multiple professional fields such as electrical engineering, mechanical engineering, and materials science. Second, design question templates that include specific task instructions, output format requirements, and contextual examples. For example, the template would explicitly require the model to extract parameters such as "rated voltage," "insulation level," and "protection class," along with their corresponding values, and output them in a specified format.

[0028] S6: Relevant Information Retrieval Based on Hybrid Retrieval Strategy. This involves employing a hybrid retrieval strategy to retrieve the most relevant information fragments from a vector database. Keyword matching utilizes the keyword library pre-set in S5 for initial filtering, quickly identifying text blocks containing the target parameters. Semantic vector retrieval converts the user query into a vector and performs a similarity search in the vector database to retrieve semantically related document fragments.

[0029] Furthermore, after the initial retrieval, a re-ranking step is introduced, using a cross-encoder to score the relevance of the retrieval results, selecting the top-K most relevant segments, effectively removing noise, and providing high-quality context for subsequent large language models.

[0030] S7: The Qwen3-8B large language model is used to accurately identify and extract key parameters. Relevant document fragments retrieved in S6 are combined with the question template set in S5 and input into the Qwen3-8B large language model. Based on the provided context, the model performs final identification of key parameters, semantic understanding normalization, and numerical extraction. Through refined engineering optimization of the model's prompt words, the "illusion" phenomenon of the model is effectively suppressed, ensuring that its output is strictly based on technical document evidence.

[0031] S8: Generate verifiable, structured extraction results and a traceability report. The key parameters and values ​​extracted in S7 are returned to the user in a clear, structured format. Simultaneously, the system generates a traceability report, highlighting and listing the specific location (e.g., page number, coordinates) and original content of each identified parameter in the original document from S1. This greatly facilitates rapid verification and manual confirmation by designers, forming a reliable closed loop of human-machine collaboration.

[0032] S9: Establish a structured enterprise product database. Construct a database containing all existing product models and their complete specifications. This database should include not only basic electrical parameters (such as rated current and voltage) but also multi-dimensional indicators such as mechanical dimensions, material properties, and environmental adaptability. It should also be linked to digital assets such as product CAD drawings and BOM lists through indexing.

[0033] S10: Key Parameter Matching Based on Decision Tree Algorithm. The key product indicator requirements finalized in S8 are categorized and matched with the product database established in S9 according to different weights. This step uses a decision tree model, which encapsulates the business rules of domain experts and defines the weights and priorities of parameters, their mandatory nature (e.g., "rated voltage" must be a perfect match) and flexible tolerance range (e.g., "dimensional accuracy" allows for ±5% error), thereby achieving intelligent and rule-based parameter compliance judgment.

[0034] S11: Complete product identification and drawing recommendation. Based on the classification and identification results of the decision tree in S10, the system outputs a list of candidate products sorted by similarity. For the product with the highest similarity, the system automatically filters and associates it with its corresponding engineering drawings. Designers can directly make targeted design improvements based on these results, thereby reducing the time for initial product design scheme selection and screening from several hours to minutes, significantly improving engineering design efficiency.

[0035] Furthermore, this invention has refined the core parameters in the Retrieval Enhancement Generation (RAG) framework, including: (1) Optimize chunk size: Select the optimal chunk size through experimental testing to balance the contextual integrity of text chunks with retrieval accuracy.

[0036] (2) Design targeted prompts: clearly guide the large language model to focus on the extraction of key parameters of electrical equipment and product matching tasks, and reduce the generation of irrelevant information.

[0037] (3) Adjust top k (the top K results returned by the retrieval): While ensuring the comprehensiveness of relevant information, avoid redundant information from interfering with the large language model.

[0038] An artificial intelligence-based system for matching key parameters of electrical equipment with products, such as... Figure 3 As shown, it includes: The information access and preprocessing module acquires technical documents from multiple sources. The information extraction module includes the PP-Structure v3 parsing submodule and the GLM-4-9B-VL parsing submodule. For multi-source technical documents, it extracts ordinary text information and nested data of tables through the PP-Structure v3 model and the GLM-4-9B-VL model, respectively. The knowledge base construction module (vector knowledge base construction and management module) uses the Qwen3-Embedding-4B embedding model to convert ordinary text information and nested data of tables into vectors and store them in a vector database to build a knowledge base. The hybrid retrieval and reordering module, based on a pre-set keyword library and structured question templates in the field of electrical equipment, uses a hybrid retrieval strategy to recall highly relevant text fragments from the knowledge base. The key parameter and intelligent matching extraction module includes a key parameter extraction module and a parameter and product matching module; the key parameter extraction module inputs the text fragment and structured question template into the Qwen3-8B large language model to extract key parameters; The parameter and product matching module matches key parameters with product specifications in the product database using a decision tree algorithm, and sorts the matching results into a candidate product list.

[0039] The interpretable PP-Structure v3 model is a lightweight, all-in-one open-source intelligent document analysis model. Belonging to the field of document image understanding, it integrates core tasks such as layout analysis, table recognition, and optical character recognition. It can quickly and accurately segment different regions (such as text, titles, tables, and images) within a document and identify all printed text content within those regions. In this invention, its primary responsibility is to efficiently process relatively simple plain text paragraphs and basic tables in documents, achieving initial digitization of information and efficiently extracting text from most of the document's content, providing foundational data for subsequent in-depth processing.

[0040] The GLM-4-9B-VL model is a multimodal visual-language model capable of simultaneously understanding and processing visual information from images and natural language text. This model possesses powerful visual reasoning and fine-grained image understanding capabilities. For document images, it can not only read the text but also understand the overall semantics of the image, the spatial relationships and logical connections between elements. It excels at parsing objects with complex structures and irregular formats, such as merged cells and complex tables with nested relationships. Specifically designed to overcome the challenge of parsing complex tables that PP-Structure v3 struggles with, it accurately reconstructs the row and column logical structure of tables by understanding their visual layout and contextual semantics, extracting machine-readable and unambiguous table data, and ensuring the integrity of key parameter sources.

[0041] The Qwen3-8B large-scale language model is a pure text language model, a generative pre-trained model based on the Transformer architecture, adept at handling natural language understanding, reasoning, question answering, and text generation tasks. Given specific context and instructions, it can accurately extract key information, summarize content, compare differences, and output results in a structured format. Its "moderate number of parameters" achieves a good balance between performance and computational efficiency. It receives highly relevant document fragments provided by a hybrid retrieval strategy and, combined with pre-set structured question templates, performs final key parameter identification, semantic normalization, and numerical extraction tasks. The extraction results are accurate and structured, and effectively suppress the generation of content not based on context. The Qwen3-Embedding-4B embedding model is a large-scale text embedding model belonging to the category of representation learning models. It is specifically designed to convert text data into a numerical form (i.e., vectors or embeddings) that can be efficiently processed by computers. It converts all text and tabular data extracted from documents into vector representations and stores them in a vector database, thus building the "memory" center of the entire system. Based on the vectors generated by this model, subsequent semantic retrieval steps can quickly and accurately find the most semantically relevant segments from massive amounts of knowledge, providing high-quality contextual input for subsequent large-scale language model analysis. Its "4B" parameter count is among the most powerful in current embedding models, aiming to maintain acceptable inference efficiency while ensuring high-precision semantic understanding.

[0042] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0043] Example 1 This invention discloses an artificial intelligence-based method and application for electrical equipment parameter identification and product matching, belonging to the interdisciplinary field of electrical equipment manufacturing and artificial intelligence. Addressing the problems of complex text structures in technical protocols and general technical documents during electrical product engineering design, the ease with which key parameters are missed or misread, and the insufficient ability of OCR technology in the traditional RAG framework to parse complex tables, this invention proposes a technical solution that integrates the Visual Language Model (VLM) and an optimized RAG architecture. The method first utilizes PP-Structure v3 for OCR recognition of ordinary text, while simultaneously introducing the GLM-4-9B-VL model to specifically parse complex tables in the document. Next, the Qwen3-Embedding-4B embedding model is used to vectorize the processed text. Then, an optimized RAG architecture is constructed using a comprehensive retrieval method combining keywords and a vector database. Finally, the Qwen3-8B large language model is used to intelligently match and recommend products within a product database based on the retrieved key parameter information. By coordinating the optimization of core parameters such as chunksize, prompt, and top k, this invention achieves a recall and accuracy rate of 95% for key parameters in the technical protocol, significantly reducing design error rate and raw material waste, and effectively improving engineering design efficiency and quality.

[0044] The following description, using a 110kV porcelain column circuit breaker as an example, illustrates the specific implementation of the present invention in conjunction with the accompanying drawings.

[0045] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0047] The following are specific embodiments. It should be noted that these embodiments are preferred examples of the present invention and are intended for those skilled in the art to understand the present invention, but the present invention is not limited to these embodiments.

[0048] S1: Collection and standardized organization of multi-source technical documents, such as... Figure 1 As shown; (1) Document Collection: The system automatically collects technical documents provided by product requesters through enterprise document management systems or API interfaces, including product-specific technical protocols (such as DOCX and PDF formats) and industry-standard technical specifications (such as IEC and GB standards) mastered by the enterprise. Document sources can include email, cloud storage, and local databases; (2) Standardized preprocessing: The document is initially cleaned to remove irrelevant information such as watermarks, headers and footers, while retaining the original layout structure (such as chapter titles, table and image positions). After preprocessing, the document is stored in a distributed file system, and metadata (document ID, version number, page index, etc.) is generated for each document to facilitate subsequent traceability.

[0049] S2: Extraction of plain text and simple table information based on PP-Structure v3.

[0050] (1) Model configuration: The PP-Structure v3 model is used for layout analysis and optical character recognition (OCR). This model is lightweight and efficient, with approximately 10M parameters, and can be deployed in a CPU or GPU environment.

[0051] Furthermore, its pre-trained layout analysis model is used to identify text, headings, tables, and image regions in the document, and then optical character recognition (OCR) is used to extract the text content.

[0052] (2) Processing flow, such as Figure 2 As shown: After inputting a standardized document, PP-Structure v3 first outputs the document's layout structure, and then performs OCR on the identified text regions. The extracted text data is temporarily stored in an in-memory database for use in subsequent steps.

[0053] S3: Introducing the GLM-4-9B-VL visual language model for accurate parsing of complex tables.

[0054] (1) Model Deployment: For complex tables that are difficult to handle in S2 (such as tables with cross-row and cross-column merging and data nesting), deploy the GLM-4-9B-VL model. Use its visual encoder to extract visual features of the table image and use the language decoder to understand the semantic context.

[0055] (2) Parsing process: The complex table area is converted into a high-resolution image (300 DPI) and input into the GLM-4-9B-VL model. The model parses the logical structure of the table through joint vision-language reasoning and outputs structured HTML table code to ensure consistency with the original table. The parsing result is merged with the output information of S2 to form a complete document, which is then digitally represented.

[0056] The GLM-4-9B-VL model was used to test all the tabular data in the technical protocol, and the results showed a recognition accuracy of up to 98%, a significant improvement compared to traditional OCR solutions (table parsing accuracy of approximately 75%). In the text recognition stage, it maintained a stable recognition accuracy for numbers, technical terms, and special symbols in the tables, with virtually no issues such as character omissions or misrecognition. Regarding functional compliance, the model strictly adhered to the core task of "table parsing," without triggering unexpected text translation functions, thus avoiding additional information interference. In terms of table structure recognition, even with complex tables containing multiple rows and columns and nested data in the technical protocol, it could accurately restore the row and column correspondences and identify the boundaries of merged cells.

[0057] S4: Use Qwen3-Embedding-4B to build a vector knowledge base.

[0058] (1) Embedding model settings: The Qwen3-Embedding-4B model is used to generate the text and table content output by S2 and S3 as vectors. The model input is the text blocks extracted by S2 and S3, and the output is a 1024-dimensional floating-point vector.

[0059] (2) Vector Database Construction: Vectors are stored using the Milvus vector database. First, the extracted text is segmented into chunks. Experiments were conducted to optimize the chunk size to 256 characters and the overlap area to 32 characters, balancing contextual integrity and retrieval accuracy. Each text chunk is appended with metadata (such as document ID, page number, and coordinates). Then, the vectors and metadata are imported into Milvus in batches, and an IVF_FLAT index is created to achieve efficient similarity search. The knowledge base is regularly updated to cover the latest technical documents from the enterprise.

[0060] S5: Set parameters to extract a keyword library and structured question templates.

[0061] (1) Keyword library design: Based on knowledge in the field of electrical equipment, a multi-level keyword library is constructed, covering categories such as electrical (e.g., "rated voltage", "insulation level", "short circuit capacity"), mechanical (e.g., "dimensional", "weight", "installation method"), material (e.g., "conductor material", "insulation type"), and environment (e.g., "operating temperature", "protection level"). The keyword library is stored in a specified format.

[0062] (2) Question Template Optimization: Design a structured question template. The template includes task instructions and context examples. The template is iteratively tested through prompt word engineering to ensure that the model focuses on key information.

[0063] For the key parameter extraction function of 110kV porcelain column circuit breaker, nine parameters are mainly extracted: rated voltage, rated current, AC component RMS value, frequency, arc extinguishing medium, ambient altitude, maximum ambient temperature, minimum ambient temperature and pollution level. Table 1 Parameter Information Extraction Prompt

[0064] The set question template that includes contextual information, such as Figure 4 As shown: S6: Retrieval of relevant information based on a hybrid retrieval strategy.

[0065] (1) A strategy combining keyword matching and semantic vector retrieval is adopted. Using the Elasticsearch engine, Boolean queries are performed on document blocks based on the S5 keyword library to quickly filter text blocks containing the target parameter. The user query is converted into a vector using Qwen3-Embedding-4B, and cosine similarity search is performed in Milvus to return the Top-K relevant segments (K is set to 5 in the experiment).

[0066] (2) Re-ranking mechanism: A cross-encoder (such as MiniLM-L6-v2) is used to score the relevance of the search results. The model calculates the semantic relevance score between the query and each segment, retains the top-3 highest-scoring segments, and effectively removes noise. After re-ranking, the segments are arranged in descending order of relevance score, which serves as the context for subsequent large language models.

[0067] Chunk size tests were conducted with lengths of 512, 384, and 256. A relatively standard chunk size was chosen, primarily because extracting key parameters does not require lengthy text inferences. Therefore, 512 was selected as the maximum test length, and the minimum chunk size is generally around 256. The retrieval recall results for the three selected test lengths are shown in Table 2. Table 2. Recall test results for different chunk sizes

[0068] With good prompt keywords, a smaller chunk size of 256 resulted in the best document fragment recall of 77.42%. However, simply adjusting the chunk size is insufficient to achieve a higher recall rate. Therefore, keywords were added to assist in improving recall. Keyword searches were added for keywords with similar meanings such as "rated voltage," "rated current," "AC component RMS value," and "frequency," resulting in a 100% improvement in retrieval performance.

[0069] S7: The Qwen3-8B large language model is used to achieve accurate identification and extraction of key parameters.

[0070] (1) Model Deployment and Inference: The Qwen3-8B model is deployed on a GPU server, and the vLLM inference engine is used to optimize throughput. The input includes the context retrieved by S6 and the question template by S5, and the model is called via API.

[0071] (2) Optimization of prompt words: Suppress model illusions through system prompts and few-sample learning.

[0072] By employing the Qwen3 model with 8 parameters and combining it with cue word engineering, the model inference cost is reduced by approximately 90% compared to the general large model with 175 parameters, while ensuring task accuracy, thus achieving a balance between high performance and low cost.

[0073] S8: Generate verifiable structured extraction results and traceability reports.

[0074] Structured output: Parameter extraction results are returned in JSON format, including parameter name, value, unit, and confidence score. A visualization table (rendered via HTML) is also generated for users to view online.

[0075] Source tracing report generation: The system locates the position of each parameter in the original document based on metadata; it uses the OpenCV library to draw bounding boxes on the document image, highlights the areas where parameters are located, and generates a PDF report listing the parameters, original citations, page numbers, and coordinates. Users can interactively verify this through a web interface; clicking on a parameter will take them to its original location in the document.

[0076] S9: Establish a structured enterprise product database.

[0077] (1) Database design: Product data is stored using a relational database (PostgreSQL). The table structure includes fields such as product model, electrical parameters (rated current, voltage, etc.), mechanical parameters (dimensions, weight), and material properties (insulation materials, conductor types). At the same time, foreign key associations are established with external digital assets (such as the URL of CAD drawings and the ID of the BOM list).

[0078] (2) Data integration: Product data is regularly synchronized from the enterprise PLM system through ETL tools (Apache Airflow) to ensure that the database is updated in real time.

[0079] S10: Automated comparison of product parameters based on decision tree rules.

[0080] (1) Decision tree model construction: The decision tree is implemented using the Scikit-learn library, and the rules are defined by domain experts. Specifically, it is divided into mandatory parameters and flexible parameters.

[0081] (2) Priority rule: Parameters are weighted according to their importance, and the matching degree is calculated by the formula: matching degree = Σ(parameter weight × conformity).

[0082] S11: Complete product matching and drawing recommendation.

[0083] (1) Matching algorithm: The system executes the decision tree rules of S10, outputs a list of candidate products, and sorts them in descending order of matching degree. For products with a matching degree >95%, their CAD drawings are automatically associated (the drawing path is retrieved through the database index).

[0084] (2) Result delivery: Results are returned via REST API or Web interface, including product model, matching degree, parameter comparison table and drawing link. Designers can download drawings with one click and start the improvement design directly, reducing the matching time from hours to minutes.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the accompanying drawings and the above description. However, any modifications, alterations, or variations made by those skilled in the art without departing from the scope of the present invention, utilizing the disclosed technical content, are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A method for matching key parameters of electrical equipment with products based on artificial intelligence, characterized in that, include: Obtain multi-source technical documents; For multi-source technical documents, the PP-Structure v3 model and the GLM-4-9B-VL model are used to extract ordinary text information and nested data of tables, respectively. The Qwen3-Embedding-4B embedding model is used to convert plain text information and nested data of tables into vectors and store them in a vector database to build a knowledge base. Based on a pre-defined keyword database and structured question templates in the field of electrical equipment, a hybrid retrieval strategy is used to recall highly relevant text fragments from the knowledge base. The text fragments and structured question templates are input into the Qwen3-8B large language model to extract key parameters; The key parameters are matched with the product specifications in the product database using a decision tree algorithm, and the matching results are sorted into a candidate product list.

2. The method for matching key parameters of electrical equipment with products based on artificial intelligence according to claim 1, characterized in that, The acquired multi-source technical documents are standardized, specifically as follows: The documents are cleaned up, watermarks and headers / footers are removed, and the original layout structure is preserved to obtain multi-source technical documents in a unified format; the multi-source technical documents include product technical agreements and industry-standard technical specifications.

3. The method for matching key parameters of electrical equipment with products based on artificial intelligence according to claim 1, characterized in that, Extract key parameters, specifically: The text fragments and structured question templates are input into the Qwen3-8B large language model to perform accurate identification of key parameters, semantic normalization and numerical extraction. Based on the extraction of key parameters, a structured list containing the key parameters and their values ​​is generated, and a corresponding traceability report is generated. The traceability report indicates the location and original content of each key parameter in the original multi-source technical document.

4. The method for matching key parameters of electrical equipment with products based on artificial intelligence according to claim 1, characterized in that, Extract plain text information and nested data from complex tables, specifically: Based on the PP-Structure v3 model, the layout analysis and optical character recognition of the multi-source technical documents are performed to extract ordinary text content. The complex table includes tables with multi-level headers, merged rows and columns, or nested data structures within cells; the GLM-4-9B-VL model parses the nested data of the complex table by understanding the visual features of the table and the semantics of the text.

5. The method for matching key parameters of electrical equipment with products based on artificial intelligence according to claim 1, characterized in that, Building a knowledge base, specifically: The Qwen3-Embedding-4B embedding model converts ordinary text information and nested data in tables into high-dimensional vectors, stores the vectors in a vector database, and builds an efficient index, thereby constructing a knowledge base for semantic retrieval.

6. The method for matching key parameters of electrical equipment with products based on artificial intelligence according to claim 1, characterized in that, A hybrid retrieval strategy is employed to retrieve highly relevant text fragments from the knowledge base, specifically: Keyword matching is performed based on the keyword database, and vector retrieval is performed based on semantic similarity to obtain preliminary retrieved information fragments. A cross-encoder model is used to re-score the relevance between the preliminary retrieved information fragments and the query, and the most relevant Top-K text fragments are selected based on the scores. The hybrid retrieval strategy includes keyword matching and semantic vector retrieval; The keyword library covers professional fields such as electrical, mechanical and materials; the structured question template includes question templates with specific task instructions, output format requirements and context examples.

7. The method for matching key parameters of electrical equipment with products based on artificial intelligence according to claim 1, characterized in that, The key parameters are matched with the product specifications in the product database using a decision tree algorithm, specifically: The key parameters are compared with the product specifications in the product database, and the parameter compliance is identified and matched with predefined business rules using a decision tree algorithm. Output a list of candidate products sorted by matching degree, and automatically associate and recommend the corresponding engineering drawings for the candidate product with the highest matching degree; The business rules are defined with parameter weights, priorities, mandatory matching conditions, and flexible tolerance ranges. The database includes multi-dimensional specifications of the company's products and their associated digital asset indexes.

8. A system for matching key parameters of electrical equipment with products based on artificial intelligence, characterized in that, include: The information access and preprocessing module acquires and preprocesses technical documents from multiple sources. The information extraction module extracts ordinary text information and nested data from tables from multi-source technical documents using the PP-Structure v3 model and the GLM-4-9B-VL model, respectively. The vector knowledge base construction module uses the Qwen3-Embedding-4B embedding model to convert ordinary text information and nested data of tables into vectors and store them in a vector database to build a knowledge base. The hybrid retrieval and reordering module, based on a pre-set keyword library and structured question templates in the field of electrical equipment, uses a hybrid retrieval strategy to recall highly relevant text fragments from the knowledge base. The key parameter extraction module inputs the text fragments and structured question templates into the Qwen3-8B large language model to extract key parameters; The parameter and product matching module matches key parameters with product specifications in the product database using a decision tree algorithm, and sorts the matching results into a candidate product list.

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 method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images