A method and system for trusted intelligent question answering of multi-source heterogeneous knowledge in an industrial field
Patent Information
- Application Number
- CN202610383619.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-17
- Publication Date
- 2026-10-16
AI Technical Summary
[0005]针对现有技术的以上缺陷或改进需求，本发明提供了一种工业领域多源异构知识的可信智能问答方法及系统，其目的在于构建一个面向工业设备领域异构文档的、全流程可信增强的知识图谱检索与问答（KG-RAG）框架，由此解决现有技术中通用图检索增强生成框架在应对工业领域异构文档时，因数据专业性强、多源矛盾冗余而导致的知识构建可信度低、问答结果幻觉严重的技术问题
1. 本发明的核心在于构建了贯穿数据预处理、知识提取、图谱融合及问答生成全过程的统一可信机制，并通过统一的信任权重体系实现信任度从数据源头到最终答案的逐级传递与量化管理。基于此，本发明不仅能够提高工业领域知识构建与知识问答的整体可靠性，而且能够在问答生成过程中有效抑制专业领域中的幻觉问题，显著提升最终输出答案的准确性、可信性和稳定性。
Smart Images
Figure CN122925536A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of intelligent management and knowledge engineering of industrial equipment, and more specifically, relates to a reliable intelligent question-answering method and system for multi-source heterogeneous knowledge in the industrial field. Background Technology
[0002] In complex industrial sectors such as high-end equipment manufacturing, energy and chemical engineering, and rail transportation, high-quality operation, maintenance, testing, and fault diagnosis of core equipment and systems are crucial for ensuring production safety and operational efficiency. Currently, knowledge-intensive tasks in these sectors (such as test report preparation, root cause analysis, and process plan review) heavily rely on domain experts manually reviewing and analyzing massive amounts of technical documents. These documents typically include international / national standards, manufacturer technical manuals, historical test reports, maintenance records, and real-time monitoring data, exhibiting characteristics of being multi-sourced, heterogeneous, and highly specialized. Traditional manual methods are not only inefficient but also limited by personal experience, making it difficult to systematically and accurately integrate and utilize the deep-seated knowledge, complex logic, and professional criteria scattered across various documents, easily leading to decision-making delays or biased conclusions.
[0003] In recent years, large language models and their retrieval-enhanced generation techniques have provided new pathways for the automated solution of the aforementioned problems. Graph retrieval-enhanced generation techniques support deep reasoning by constructing knowledge graphs rich in entity relationships, demonstrating potential in handling complex question-answering tasks in specialized domains. However, industrial document data presents core challenges such as numerous principle and formula descriptions, dense technical terminology, information scattered across heterogeneous sources of structured / semi-structured / unstructured data, and potential version conflicts or descriptive inconsistencies between different sources. When existing general-purpose frameworks such as GraphRAG are directly applied to such scenarios, they exhibit significant shortcomings in the accuracy of entity and relation extraction, the reliability of multi-source information fusion, and the ability to identify and resolve contradictory and redundant data. This results in the generated results still facing serious "illusion" problems, failing to meet the stringent requirements of industrial scenarios for knowledge reliability and decision auditability.
[0004] Therefore, there is an urgent need to design a reliable intelligent question-answering method and system for multi-source heterogeneous knowledge in the industrial field. Summary of the Invention
[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a reliable intelligent question-answering method and system for multi-source heterogeneous knowledge in the industrial field. Its purpose is to construct a knowledge graph retrieval and question-answering (KG-RAG) framework with full-process reliability enhancement for heterogeneous documents in the industrial equipment field. This solves the technical problems of low reliability of knowledge construction and serious illusion of question-answering results caused by the high specialization of data and the redundancy of multiple sources when dealing with heterogeneous documents in the industrial field in general graph retrieval enhancement generation frameworks.
[0006] To achieve the above objectives, according to one aspect of the present invention, a reliable intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field is provided, comprising the following steps: S1: Domain-specific data preprocessing: Perform domain-specific data preprocessing on heterogeneous documents in the industrial domain to generate standardized data blocks carrying metadata; S2: Trust-aware knowledge extraction and graph construction: Input the standardized data block generated in step S1 into the trust-aware extraction agent, extract entities and relationships through the agent, calculate the initial trust weight, and construct a heterogeneous knowledge subgraph with trust labels; S3: Trusted Knowledge Fusion and Enhanced Graph Generation: Entity alignment and relation alignment are performed on the multiple heterogeneous knowledge subgraphs generated in step S2. Cross-graph trust verification and conflict adjudication are then used to generate a global trusted enhanced knowledge graph. S4: Trust-guided retrieval and question-answer generation: Based on the global trust-enhanced knowledge graph generated in step S3, in response to user queries, it outputs highly trustworthy answers through trust-weighted retrieval, evidence ranking, and answer generation.
[0007] Preferably, step S1 includes: S11: Cleaning, unit normalization, and condition correction of structured data; S12: Parse, discover patterns, and transform unified representations of semi-structured data; S13: Parse unstructured documents, reconstruct text, and perform legally-aware semantic segmentation; S14: Add metadata, including source, authority level, version, and condition context, to each processed data block in steps S11 to S13 to form the standardized data block, which serves as the basis for subsequent credibility calculation.
[0008] Preferably, step S2 includes: S21: Construct a heterogeneous graph representation framework based on the standardized data blocks generated in step S1; S22: Based on the heterogeneous graph representation framework, combined with the data type features and semantic features in the standardized data block, the agent's execution instructions are dynamically generated through trust perception, and the trustworthiness of the relationship is evaluated based on the generated content; S23: Based on the instructions, perform entity extraction and relation extraction on the standardized data block, calculate entity credibility and relation credibility, and fill and update the heterogeneous knowledge subgraph according to the entity credibility and relation credibility until the construction of the heterogeneous knowledge subgraph is completed.
[0009] Preferably, step S3 specifically includes: S31: Perform entity and relationship alignment on the multiple heterogeneous knowledge subgraphs generated in step S2, and construct entity-node mapping; S32: Based on the entity-node mapping, detect conflicting descriptions of the same entity or relationship between different heterogeneous knowledge subgraphs, and perform cross-graph trust verification; S33: Based on the cross-graph trust verification results, combined with the source authority level, revision history and conditional context, adjudicate the conflict description to filter out noisy knowledge and inaccurate node relationships; S34: Based on the conflict resolution results, update the trust weights of nodes and edges through iterative trust propagation to generate a globally trustworthy enhanced knowledge graph.
[0010] Preferably, step S4 specifically includes: S41: In response to a user query, extract the query entity from the user query and form the query input for subsequent trusted guided retrieval and question-answering reasoning; S42: Based on the query entity, retrieve the evidence subgraph according to the trust weight in the global trust-enhanced knowledge graph; when the query entity cannot be extracted, perform intensive retrieval to supplement and recall candidate evidence; S43: Perform credible reordering and evidence voting on the evidence subgraph and candidate evidence retrieved by the retrieval to generate candidate answers; S44: Perform clause-level credibility verification on the candidate answers; when the verification result is lower than a preset threshold, output a low confidence prompt and source evidence.
[0011] Preferably, in step S22, the agent's execution instructions are dynamically generated through trusted perception, and the relationship trustworthiness is evaluated. The calculation formula is as follows:
[0012] in, Based on the confidence level of the basic relationship, The credibility weights of the entities at both ends of the relationship. The basic confidence weighting coefficient This refers to the credibility of the relationship.
[0013] Preferably, in step S34, the confidence weight The calculation formula is as follows:
[0014] in, For the edge In the The credibility weight of the next iteration for arrive The set of all paths For a single path, For path edges The current weight, The path decay factor, Yes Path decay.
[0015] Preferably, in step S41, after extracting the query entity from the user query and forming the query input, trusted guided retrieval and question-answering reasoning are performed based on the global trusted enhanced knowledge graph. The expression for the trusted guided retrieval, reasoning, and question-answering process is as follows:
[0016] In this system, the user input question is q, the trusted enhanced knowledge graph is G, the text block index is T, and the question-answering generation model is M. This represents the final answer output by the system. This represents the set of evidence that is relevant to the problem and has been filtered through trust constraints.
[0017] According to another aspect of the present invention, a reliable intelligent question-answering system for multi-source heterogeneous knowledge in the industrial field is provided, employing the above-described method, including: Domain-specific data preprocessing modules are used to generate standardized data blocks; The Trust Awareness Knowledge Extraction and Graph Construction Module is used to construct heterogeneous knowledge subgraphs with initial trust weights. The trusted knowledge fusion and enhanced graph generation module is used to generate a global trusted enhanced knowledge graph; A trust-guided retrieval and question-answering generation module is used to generate highly credible answers; The modules mentioned above are connected by a unified trust weight system to achieve a reliable knowledge flow throughout the entire process from the data source to the final answer.
[0018] In summary, compared with the prior art, the reliable intelligent question-answering method and system for multi-source heterogeneous knowledge in the industrial field provided by the present invention have the following beneficial effects: 1. The core of this invention lies in constructing a unified and reliable mechanism that runs through the entire process of data preprocessing, knowledge extraction, graph fusion, and question-answer generation. It achieves the step-by-step transfer and quantitative management of trust from the data source to the final answer through a unified trust weight system. Based on this, this invention not only improves the overall reliability of knowledge construction and question answering in the industrial field, but also effectively suppresses illusion problems in professional fields during question-answer generation, significantly improving the accuracy, reliability, and stability of the final output answer.
[0019] 2. This invention addresses the characteristics of complex structures, diverse expressions, dense technical terminology, and the inclusion of formulas, drawings, and principle explanations in multi-source heterogeneous documents in the industrial field. Through domain-specific data preprocessing and an agent-assisted, reliable knowledge extraction mechanism, it achieves more accurate extraction of entity, relationship, and attribute information from complex industrial documents. Compared to traditional graph retrieval enhancement generation methods, this invention better adapts to the processing needs of industrial professional documents, improving the accuracy and robustness of knowledge extraction while ensuring the efficiency of knowledge graph construction. This lays a higher-quality data foundation for subsequent graph construction and question-answering reasoning.
[0020] 3. This invention further utilizes authoritative cross-validation, conflict resolution, and trust adjudication mechanisms to effectively integrate redundant, contradictory, and inconsistent information from multi-source heterogeneous data, constructing a higher-quality, clearer-structured, and more credible knowledge graph. Simultaneously, by introducing a credible question-answering mechanism based on evidence weighting and evidence voting, the retrieval results are evaluated, ranked, and synthesized across multiple dimensions to improve the consistency, interpretability, and auditability of evidence during the answer generation process. Therefore, this invention is particularly suitable for highly credible knowledge-based question-answering tasks in highly specialized scenarios such as industrial equipment testing, operation and maintenance, and technical document management. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the overall framework of the present invention; Figure 2 This is a flowchart illustrating the data preprocessing specific to this invention. Figure 3 This is a flowchart of the knowledge graph generation process assisted by the trusted mechanism of this invention; Figure 4 This is a schematic diagram of the entity-node-edge-weight mapping of the present invention; Figure 5 This is a flowchart of the knowledge graph-enhanced retrieval and question-answering scheme of the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0023] This invention is constructed using intelligent test report compilation and expert knowledge Q&A as the application scenario. In its daily work, the testing center needs to process massive amounts of transformer technical documents from multiple sources, including but not limited to: international / national standards (such as IEC, GB), technical manuals and type test reports provided by equipment manufacturers, annual on-site inspection and diagnostic reports, maintenance records, and test data from different laboratories. These documents vary in format (PDF, Word, Excel, database entries), and the information may have version iterations or contradictory descriptions (such as differences in records of the same equipment parameters at different times), and contain a large number of professional formulas, schematic diagrams, and technical criteria. In the traditional model, compiling an in-depth test report or answering a complex technical inquiry requires manual review, comparison, and synthesis of the above documents, which is time-consuming, labor-intensive, and prone to oversights due to limitations in personal experience. Directly applying existing large language models or general RAG systems often results in "illusions" due to inaccurate understanding of technical terminology, confusion of similar concepts, or adoption of incorrect data, failing to meet the extremely high reliability requirements of the industrial field for conclusions.
[0024] Please see Figure 1 This invention provides a reliable intelligent question-answering method and system framework for multi-source heterogeneous knowledge in the industrial field. The system architecture is designed based on multi-level, modular, and standardized principles, aiming to improve the system's scalability, robustness, and reusability.
[0025] This invention provides a trusted intelligent question-answering system for multi-source heterogeneous knowledge in the industrial field, comprising: Domain-specific data preprocessing modules are used to generate standardized data blocks; The Trust Awareness Knowledge Extraction and Graph Construction Module is used to construct heterogeneous knowledge subgraphs with initial trust weights. The trusted knowledge fusion and enhanced graph generation module is used to generate a global trusted enhanced knowledge graph; A trust-guided retrieval and question-answering generation module is used to generate highly credible answers; The modules mentioned above are connected by a unified trust weight system to achieve a reliable knowledge flow throughout the entire process from the data source to the final answer.
[0026] Specifically, firstly, the system collects multi-source heterogeneous data through a data acquisition module, including sensor data from industrial sites, production batch and time information collected by RFID, structured database records, and unstructured documents (such as process documents, quality reports, and maintenance logs). This data undergoes domain-specific preprocessing, including cleaning, formatting, standardization, and semantic alignment, providing high-quality input for subsequent knowledge extraction. Domain-specific preprocessing is then applied to the multi-source heterogeneous documents in the industrial domain to form standardized data blocks carrying trusted metadata. Entities and relationships are extracted based on these standardized data blocks, and a heterogeneous knowledge subgraph with initial trust weights is constructed. Multiple heterogeneous knowledge subgraphs are aligned, verified, adjudicated, and their weights are propagated and updated to generate a globally trusted enhanced knowledge graph. Based on this globally trusted enhanced knowledge graph, trusted guided retrieval, evidence fusion, and answer generation are performed, and corresponding source evidence is output. These stages are interconnected, forming a complete trusted closed loop from source data processing to final question-and-answer output.
[0027] The core of this system is a knowledge extraction and graph construction module based on trust perception. By introducing agent collaboration and a workflow driven by a large language model (LLM), it automatically extracts entities and relationships from structured, semi-structured, and unstructured documents. All extraction results are aggregated, aligned, and conflict-resolved through an agent layer, and finally, the LLM performs comprehensive reasoning and answer generation, forming consistent and trustworthy knowledge fragments. These fragments are continuously fed into the knowledge graph, dynamically updating its structure.
[0028] The entire system achieves clear decoupling and flexible orchestration of data flow and task flow through modular design. Intelligent agents serve as reusable functional units, and instructions act as configurable control logic. Combined with a trustworthy weight evaluation mechanism, these elements jointly ensure the interpretability of the knowledge extraction process and the credibility of the results. This architecture not only supports the automated construction of domain knowledge graphs from massive amounts of multimodal data but also lays a solid, dynamic, and evolvable knowledge foundation for subsequent upper-level applications such as semantic retrieval, intelligent reasoning, and decision support.
[0029] This embodiment also provides a reliable intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field, specifically including: S1: Domain-specific data preprocessing: Perform domain-specific data preprocessing on heterogeneous documents in the industrial domain to generate standardized data blocks carrying metadata; Step S1 specifically includes: S11: Clean, normalize, and conditionally correct the structured data; Specifically, preprocessing is performed on structured data with fixed patterns and clearly defined fields, such as test records and fault monitoring logs, to transform the original tables into high-quality, mappable intermediate states, and to annotate key metadata for subsequent steps.
[0030] Data cleaning and standardization are performed on missing and outlier values. To address the dynamic nature of engineering standards, unit normalization (e.g., unifying voltage units to "kilovolts") and status correction (e.g., correcting temperature reference values according to standards) are implemented to ensure data consistency.
[0031] Schema parsing and metadata appending parse the fixed schema (table structure, column definitions, primary and foreign key constraints) of the source data. The core task is to append key trust metadata to each row of records, including: data source (specific database table), authority level (such as national standard, manufacturer standard), version / timestamp, and applicable condition context. This step outputs a well-organized, structured set of data records with complete metadata, whose explicit schema provides the foundation for direct mapping to downstream modules.
[0032] S12: Parse, discover patterns, and transform unified representations of semi-structured data; Specifically, it parses self-descriptive data without a fixed pattern, such as XML and JSON format configuration files and tagged log files. Data parsing and pattern discovery utilize appropriate parsers (such as XML DOM parsers and JSON parsers) to extract data elements and their hierarchical / tag relationships. By analyzing tags, keys, and nesting structures, the inherent data patterns are dynamically discovered.
[0033] Data transformation and metadata appending convert the parsed data into a unified intermediate representation (such as a set of key-value pairs or a simple table). Simultaneously, trust metadata is appended to each data item, including: source file, path / tag information, authority level, generation timestamp, and conditional context. This step outputs a semi-structured data set with a clear schema and rich metadata.
[0034] S13: Parse unstructured documents, reconstruct text, and perform legally-aware semantic segmentation; Specifically, it processes texts such as technical manuals and experimental reports, focusing on the construction and preservation of deep semantic context.
[0035] By using optical character recognition (OCR) and document parsing technologies, the original document is converted into plain text, while preserving the logical structure of chapters, sections, lists, etc. as much as possible.
[0036] Combining industry regulations, standard clauses, and professional semantic features, a regulation-aware dynamic semantic segmentation strategy is adopted to segment the text, so that each text block can maintain the integrity of clauses, statements, and test item descriptions as much as possible; and by setting context overlap between adjacent text blocks, key information is avoided from being cut off at the block boundaries, resulting in standardized text data blocks suitable for subsequent knowledge extraction.
[0037] S14: Add metadata, including source, authority level, version, and condition context, to each processed data block from S11 to S13 to form the standardized data block, which serves as the basis for subsequent credibility calculations. Specifically, for structured, semi-structured, and text data blocks processed through S11 to S13, metadata such as source document or source table, clause or clause identifier, authority level, version number or timestamp, and conditional context are uniformly appended. Furthermore, preparatory alignment information for subsequent knowledge graph entity mapping and text indexing is constructed, forming a standardized set of data blocks carrying trusted metadata. This provides direct support for subsequently constructing "node-text block" aligned connections and implementing hybrid retrieval.
[0038] The input corpus encompasses multi-source, heterogeneous data from the industrial equipment sector, including national standards, industry standards, manufacturer manuals, inspection and testing reports, and operational logs. To overcome the challenge of handling the inherent differences in such data with a single workflow, a divide-and-conquer strategy is adopted in the preprocessing stage, designing separate pipelines for structured and unstructured data, while uniformly adhering to regulatory awareness principles to ensure that the output contains source traceability and conditional constraint information. This invention employs a graph structure model as a unified knowledge representation framework; please refer to [link / reference]. Figure 2 .
[0039] Through the above process, this invention transforms raw, heterogeneous industrial equipment data into two types of standardized data blocks (structured data blocks and indexed text blocks) rich in domain context and rigorously labeled with trust metadata such as source, authority, version, and conditions. These outputs form a high-quality foundation for the downstream trust-aware knowledge extraction and graph construction modules to perform efficient and accurate entity relationship extraction and initial trust weight calculation. They also directly support the auditability of the final knowledge graph and the efficiency of hybrid retrieval.
[0040] S2: Trust-aware knowledge extraction and graph construction: Input the standardized data block generated in step S1 into the trust-aware extraction agent, extract entities and relationships through the agent, assign initial trust weights, and construct a heterogeneous knowledge subgraph with trust labels; Step S2 includes: S21: Construct a heterogeneous graph representation framework based on the standardized data blocks generated in step S1; Specifically, when constructing a heterogeneous graph of domain specialization under the trusted weight dimension, a knowledge graph is first used to represent the input, and the constructed data heterogeneous graph... Represented as:
[0041] in, Represents the set of all entity nodes. Represents the set of all relation edges. Indicates the element type; Element type The expression is as follows:
[0042] in, As the head node, For the relationship, The tail node, For relation weights, , These are heterogeneous types for nodes and edges, respectively.
[0043] In a heterogeneous graph, for each edge, a meta-type attribute can be obtained, including the head node. ,relation Tail node The metatype. Unlike knowledge graphs of regular language text, it adds a trusted weight to each entity node. This is used to implement subsequent trust mechanisms to adapt to industrial equipment data. Trust weights are obtained synchronously with entity nodes during the entity relationship extraction phase; this process is described in detail below.
[0044] S22: Based on the heterogeneous graph representation framework, combined with the data type features and semantic features in the standardized data block, the agent's execution instructions are dynamically generated through trust perception, and the trustworthiness of the relationship is evaluated based on the generated content; Specifically, regarding the introduction of the necessary agents for improvement in the knowledge graph stage, please refer to [link / reference]. Figure 3 This is used to implement a trusted mechanism without introducing multi-task agents in other stages. Since the traditional GraphRAG can work without agents, from a practical point of view, it avoids introducing unnecessary agents to reduce unnecessary consumption and latency.
[0045] This paper introduces an extraction agent to construct the heterogeneous graph, thereby optimizing entity association information and semantic dependencies in the knowledge graph. The process of generating heterogeneous graphs based on the extraction agent can be defined as follows: , where the proxy function Combined with observation This includes structured data such as explicit graph connections and unstructured data such as text information. A heterogeneous graph generation action is then generated. intelligent agent Input and output are:
[0046] in, , is a high-confidence entity set and relation triple, is a trusted weight set, is an agent, which corresponds to trainable parameters , is a text block, is the current knowledge graph subgraph, is a subgraph trusted weight set, is an entity extraction model, is a relation extraction model.
[0047]
[0048] Herein, W is a weight set, is an entity set weight, is a relation triple weight.
[0049] In this process, entity extraction is first performed based on the agent:
[0050]
[0051] Herein, is the extracted basic entity set, is the extracted basic weight, is a text block, is an entity extraction model, is an instruction vector, is the i-th entity, is the type corresponding to the entity, is the NER model confidence, is the entity credibility, is the set of other entities in the same sentence, and are entity credibility balance coefficients.
[0052] Indirectly, better instructions are required, so an instruction-driven credible perception module is designed, and the instruction for the extraction agent to perform credible perception is dynamically generated, so that high-confidence entities and relations obtain higher weights in instruction generation. Specifically, the text and the subgraph are concatenated in their textual descriptions, and an instruction vector is generated through a lightweight Attention layer. Finally, from the basic entity set , the part that satisfies is selected to form a high-confidence entity set , is a confidence threshold.
[0053] In the relation extraction stage:
[0054]
[0055]
[0056] in, For relational triples, , For the entities at both ends of the relationship, For the relationship between the two entities, For relation extraction models, Based on the confidence level of the basic relationship, As a confidence threshold, The credibility weights of the entities at both ends of the relationship. The basic confidence weighting coefficient This refers to the credibility of the relationship.
[0057] S23: Based on the instructions, perform entity extraction and relation extraction on the standardized data block, calculate entity credibility and relation credibility, and fill and update the heterogeneous knowledge subgraph according to the entity credibility and relation credibility until the construction of the heterogeneous knowledge subgraph is completed.
[0058] S23 specifically includes: The heterogeneous subgraph is updated. The method for updating the heterogeneous subgraph during the generation process is as follows:
[0059] in, For subgraph, The updated subgraph, For a set of entities with high confidence, For relational triples.
[0060] Accordingly, after the subgraph is updated, the existing relationships are searched and validated again. The weight update method during this process is as follows:
[0061] in, This represents the entity after iterative updates. This represents the entity after iterative updates. This represents an entity in an iterative update process. This indicates the relationship after iterative updates. This indicates the relationship before the iteration update. This represents the relationship in the iterative update process. Update parameters for entity weights. Update the parameters for relation weights.
[0062] This problem model significantly improves the quality of the knowledge graph while maintaining computational efficiency through a reliable weighting mechanism.
[0063] S3: Trusted knowledge fusion and enhanced graph generation: Trusted knowledge fusion is performed on multiple heterogeneous knowledge subgraphs generated in step S2. Through cross-graph trust verification and conflict adjudication, a global trusted enhanced knowledge graph is generated. Step S3 specifically includes: S31: Perform entity and relationship alignment on the multiple heterogeneous knowledge subgraphs generated in step S2, and construct entity-node mapping; Specifically, after the heterogeneous graph is generated, heterogeneous graph fusion is carried out on multiple heterogeneous knowledge subgraphs to obtain the final global knowledge graph.
[0064] Please see Figure 4 First, the target entities and sets of relationships at this stage:
[0065] in, This is the merged global node set. The number of subgraphs. For data source identification, For node ID connector, These are nodes in the original subgraph. For the set of nodes in the subgraph, For entity e, the hub node For a unique set of entities, A collection containing entity nodes. For the link threshold, This is the merged global edge set. For the node pairs in the original subgraph, Let Gi be the original set of edges.
[0066] This method distinguishes between nodes and entities because, before heterogeneous graph fusion, the same entity may contain multiple nodes, thus requiring entity alignment and relationship alignment. The entity-node mapping is constructed as follows:
[0067] in, A collection containing entity nodes. A function for normalizing node text. For nodes in the subgraph, The set of nodes for all subgraphs. For nodes.
[0068] The above formula establishes a mapping from entities (text content) to nodes (nodes in the graph).
[0069] S32: Based on the entity-node mapping, detect conflicting descriptions of the same entity or relationship between different heterogeneous knowledge subgraphs, and perform cross-graph trust verification; Specifically, traditional heterogeneous graphs often directly fuse entities based on strategies such as instruction alignment to obtain the final graph. However, faulty or erroneous knowledge in heterogeneous graphs can introduce noise errors, and this noise is often found in structured data. Therefore, this invention does not perform direct fusion. Instead, it uses discrete heterogeneous knowledge subgraphs to perform cross-validation on descriptions of the same entity or relationship between different subgraphs. This detects noise, inaccurate entities, and inconsistent relationships introduced by faulty data. During cross-validation, the existing trusted weights of each node and edge are used as the validation criteria to judge the consistency between different entities and relationships, thereby identifying potential conflicting descriptions.
[0070] S33: Based on the cross-graph trust verification results, combined with the source authority level, revision history and conditional context, adjudicate the conflict description to filter out noisy knowledge and inaccurate node relationships; Specifically, upon detecting contradictory descriptions of the same entity or relationship across different subgraphs, a trusted adjudication is performed on the conflicting knowledge by combining the source authority level, revision history, and conditional context information from the metadata. This trusted adjudication is used to filter, retain, or eliminate multiple candidate nodes or relationships to remove noise from the original subgraphs, identify and eliminate inaccurate node relationships, thereby improving the accuracy and consistency of the fused knowledge graph.
[0071] S34: Based on the conflict resolution results, update the trust weights of nodes and edges through iterative trust propagation to generate a globally trustworthy enhanced knowledge graph.
[0072] Specifically, after completing entity alignment, relation alignment, cross-graph trust verification, and conflict resolution, the trust weights of the nodes and edges involved in the fusion process are iteratively updated as follows:
[0073] in, For the edge In the The credibility weight of the next iteration for arrive The set of all paths For a single path, For path edges The current weight, The path decay factor, Yes Path decay.
[0074] The introduction of an iterative mechanism in the trusted weight update aims to address the semantic consistency problem in graph fusion. Cross-heterogeneous graph fusion may introduce noise (such as homonymous entities), which is corrected through neighbor weight feedback during the iterative process.
[0075] Through the above steps, the present invention can achieve reliable alignment, conflict resolution, and weight propagation and update among multiple heterogeneous knowledge subgraphs, thereby reducing the impact of noisy knowledge on the fusion results and improving the accuracy, consistency, and reliability of the global knowledge graph.
[0076] S4: Trust-guided retrieval and question-answering generation: Based on the global trust-enhanced knowledge graph generated in step S3, in response to user queries, a highly trustworthy answer is generated through a trust weight-guided retrieval and evidence voting mechanism.
[0077] Step S4 includes: S41: In response to a user query, extract the query entity from the user query and form the query input for subsequent trusted guided retrieval and question-answering reasoning; Specifically, the system receives user-input question text, performs word segmentation, semantic parsing, and domain terminology recognition on the question text, and extracts query entities related to industrial equipment, test items, parameter indicators, fault types, operating status, and standard clauses. The extracted results are then normalized and mapped using a domain dictionary, entity sets in a knowledge graph, and contextual semantics to form a query entity set for knowledge graph retrieval. For ambiguous query entities, disambiguation processing can be performed using the question context and prior credibility information of the entity to improve the accuracy of subsequent retrieval. The expression for the trust-guided retrieval reasoning question-answering process studied in this invention is as follows:
[0078] in, For the trust-guided approach, the user input question is q, the trust-enhanced knowledge graph is G, the text block index is T, and the question-answering generation model is M. This represents the final answer output by the system. This represents the set of evidence that is relevant to the problem and has been filtered through trust constraints.
[0079] S42: Based on the query entity, retrieve the evidence subgraph according to the trust weight in the global trust-enhanced knowledge graph; when the query entity cannot be extracted, perform intensive retrieval to supplement and recall candidate evidence; Specifically, based on the query entity extracted in step S41, matching is performed in the global trusted enhanced knowledge graph to retrieve nodes, edges, and their neighborhood relationships associated with the query entity, resulting in a candidate evidence subgraph. During the retrieval process, node trust weights, edge trust weights, entity semantic relevance, and source metadata are used as retrieval criteria to filter and rank the candidate evidence. For user questions where query entities cannot be effectively extracted or where direct matching results are lacking in the knowledge graph, intensive retrieval is further performed on the standardized text data blocks generated in step S1 to supplement and recall relevant evidence, thereby forming a candidate evidence set combining graph retrieval and text retrieval.
[0080] S43: Perform credible reordering and evidence voting on the evidence subgraph and candidate evidence retrieved by the retrieval to generate candidate answers; Specifically, after generating trustworthy weight elements during the graph generation stage, an optimized question-answering retrieval strategy was further designed based on these elements. This optimization ensures that the retrieval results significantly increase the proportion of highly trustworthy content while maintaining recall. Firstly, based on a natural language question... and the knowledge graph constructed above Give a guide After that, it was possible to obtain the map Relevant entities, relationships, and logical paths are extracted. Next, the guiding text is fed into a graph neural network augmented LLM to generate the final answer. The process is modeled as follows:
[0081] Where Q represents the natural language problem. For the constructed knowledge graph, The provided guide is able to extract from the graph Extract relevant entities, relationships, and logical paths. The guiding text is provided to the LLM to generate the final answer. S44: Perform clause-level credibility verification on the candidate answers; when the verification result is lower than a preset threshold, output a low confidence prompt and source evidence.
[0082] To address the problems of high evidence noise, weak relation constraints, and insufficient answer interpretability in multi-source heterogeneous knowledge environments, which are inherent in traditional retrieval enhancement generation methods, and the knowledge graph in step S3, a customized retrieval and question-answering strategy incorporating trusted assistance was developed. Please refer to the details of the question-answering retrieval strategy. Figure 5 Please refer to Table 1 for the pseudocode, as follows:
[0083] in, For text block indexing (the mapping from entity to text block, where M is the question-answering generation model (such as LLM)). To support the list of evidence text blocks. E_q is the set of entities extracted from the question, S is the knowledge subgraph after weight filtering, and τ is the weight threshold. The initial candidate text blocks are represented by `ans`, which is the answer result consisting of three parts: `text`, `source`, and `confidence`, representing the answer text, source tag, and confidence score, respectively. `evidence_text` is the concatenated evidence text, and `augmented_input` is the augmented text input to the model.
[0084] In summary, this invention constructs a trusted intelligent question-answering method and system for multi-source heterogeneous knowledge in the industrial field by introducing a unified trusted mechanism in each stage of data preprocessing, knowledge extraction, knowledge fusion, and question-answer generation. This method can effectively process industrial documents from different sources, with varying structures, and containing redundant or even conflicting content, forming a trusted enhanced knowledge graph carrying source, authority level, version information, and conditional context. Based on this graph, trusted retrieval, evidence aggregation, and answer generation for user questions are achieved.
[0085] Compared to existing technologies, this invention not only improves the accuracy and consistency of entity relationship extraction, knowledge fusion, and graph construction in complex documents in the industrial field, but also enhances the accuracy, reliability, interpretability, and auditability of question-and-answer results through a trusted weight-guided retrieval, ranking, and clause-level verification mechanism, effectively suppressing illusion problems in professional fields. Therefore, this invention is particularly suitable for high-reliability knowledge base construction and intelligent question-and-answer tasks in scenarios such as industrial equipment design, testing, and maintenance, and has significant practical value.
[0086] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A reliable intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field, characterized in that: Includes the following steps: S1: Domain-specific data preprocessing: Perform domain-specific data preprocessing on heterogeneous documents in the industrial domain to generate standardized data blocks carrying metadata; S2: Trust-aware knowledge extraction and graph construction: Input the standardized data block generated in step S1 into the trust-aware extraction agent, extract entities and relationships through the agent, assign initial trust weights, and construct a heterogeneous knowledge subgraph with trust labels; S3: Trusted knowledge fusion and enhanced graph generation: Trusted knowledge fusion is performed on multiple heterogeneous knowledge subgraphs generated in step S2. Through cross-graph trust verification and conflict adjudication, a global trusted enhanced knowledge graph is generated. S4: Trust-guided retrieval and question-answering generation: Based on the global trust-enhanced knowledge graph generated in step S3, in response to user queries, a highly trustworthy answer is generated through a trust weight-guided retrieval and evidence voting mechanism.
2. The trusted intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field as described in claim 1, characterized in that: Step S1 includes: S11: Cleaning, unit normalization, and condition correction of structured data; S12: Parse, discover patterns, and transform unified representations of semi-structured data; S13: Parse unstructured documents, reconstruct text, and perform legally-aware semantic segmentation; S14: Attach metadata, including source, authority level, version, and condition context, to each processed data block in steps S11 to S13 to form the standardized data block.
3. The trusted intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field as described in claim 1, characterized in that: Step S2 includes: S21: Construct a heterogeneous graph representation framework based on the standardized data blocks generated in step S1; S22: Based on the heterogeneous graph representation framework, combined with the data type features and semantic features in the standardized data block, the agent's execution instructions are dynamically generated through trust perception, and the trustworthiness of the relationship is evaluated based on the generated content. S23: Based on the instructions, perform entity extraction and relation extraction on the standardized data block, calculate entity credibility and relation credibility, and fill and update the heterogeneous knowledge subgraph according to the entity credibility and relation credibility until the construction of the heterogeneous knowledge subgraph is completed.
4. The trusted intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field as described in claim 3, characterized in that: Step S3 includes: S31: Perform entity alignment and relationship alignment on the multiple heterogeneous knowledge subgraphs generated in step S2 to construct entity-node mapping; S32: Based on the entity-node mapping, detect conflicting descriptions of the same entity or relationship between different heterogeneous knowledge subgraphs, and perform cross-graph trust verification; S33: Based on the cross-graph trust verification results, combined with the source authority level, revision history and conditional context, adjudicate the conflict description to filter out noisy knowledge and inaccurate node relationships; S34: Based on the conflict resolution results, update the trust weights of nodes and edges through iterative trust propagation to generate a globally trustworthy enhanced knowledge graph.
5. A reliable intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field as described in claim 1, characterized in that: Step S4 includes: S41: In response to a user query, extract the query entity from the user query and form the query input for subsequent trusted guided retrieval and question-answering reasoning; S42: Based on the query entity, retrieve the evidence subgraph according to the trust weight in the global trust-enhanced knowledge graph; when the query entity cannot be extracted, perform intensive retrieval to supplement and recall candidate evidence: S43: Perform credible reordering and evidence voting on the evidence subgraph and candidate evidence retrieved by the retrieval to generate candidate answers; S44: Perform clause-level credibility verification on the candidate answers; when the verification result is lower than a preset threshold, output a low confidence prompt and source evidence.
6. The trusted intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field as described in claim 3, characterized in that: In step S22, the agent's execution instructions are dynamically generated through trusted perception, and the trustworthiness of the relationships is evaluated based on the generated content. The trustworthiness of candidate relationships is calculated, and the relationship trustworthiness is determined. The calculation formula is as follows: in, Based on the confidence level of the basic relationship, The credibility weights of the entities at both ends of the relationship. The basic confidence weighting coefficient This refers to the credibility of the relationship.
7. The trusted intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field as described in claim 4, characterized in that: In step S34, the credibility weight The calculation formula is as follows: in, For the edge In the The credibility weight of the next iteration for arrive The set of all paths For a single path, For path edges The current weight, The path decay factor, Yes Path decay.
8. A reliable intelligent question-answering method for multi-source heterogeneous knowledge in the industrial field as described in claim 5, characterized in that: In step S41, after extracting the query entity from the user query and forming the query input, trust-guided retrieval and question-answering reasoning are performed based on the global trust-enhanced knowledge graph. The expression for the trust-guided retrieval, reasoning, and question-answering process is as follows: in, For the trust-guided approach, the user input question is q, the trust-enhanced knowledge graph is G, the text block index is T, and the question-answering generation model is M. This represents the final answer output by the system. This represents the set of evidence that is relevant to the problem and has been filtered through trust constraints.
9. A trusted intelligent question-answering system for multi-source heterogeneous knowledge in the industrial field, employing the method described in any one of claims 1-8, characterized in that: include: Domain-specific data preprocessing modules are used to generate standardized data blocks; The Trust Awareness Knowledge Extraction and Graph Construction Module is used to construct heterogeneous knowledge subgraphs with initial trust weights. The trusted knowledge fusion and enhanced graph generation module is used to generate a global trusted enhanced knowledge graph; A trust-guided retrieval and question-answering generation module is used to generate highly credible answers; The modules mentioned above are connected by a unified trust weight system to achieve a reliable knowledge flow throughout the entire process from the data source to the final answer.