Multi-source information fusion retrieval method and system combined with information credibility portrait

Abstract

The invention discloses a multi-source information fusion retrieval method and system combined with an information credibility portrait, and belongs to the technical field of intelligent information retrieval and auxiliary decision making. The method comprises the steps that the credibility portrait is constructed according to a predefined credibility index, and multi-source knowledge classification is completed; obtaining a candidate knowledge point set corresponding to query, and dividing the candidate knowledge point set into contrast knowledge point groups for conflict detection; according to a conflict detection result, performing complementary enhancement on the contrast knowledge point group without conflicts according to multi-source knowledge classification to form a fused knowledge point set; according to multi-source knowledge classification, credibility dynamic weighted fusion and human-in-loop conflict recheck and correction are carried out on the contrast knowledge point groups containing conflicts, and after conflict resolution is completed, a fused knowledge point set is obtained; and using the fused knowledge point set as a basis to feed back a retrieval result corresponding to the query. According to the method, the credibility of retrieval and decision results can be improved by utilizing the multi-source information through credibility enhancement and conflict resolution of the multi-source information.

Description

Technical Field

[0001] This invention belongs to the field of intelligent information retrieval and decision support technology, specifically relating to a multi-source information fusion retrieval method and system that combines information credibility profiles. Background Technology

[0002] In the operation and management of various complex systems and business processes, operators often need to process large amounts of real-time operational data, alarm information, and numerous professional technical documents (such as operating procedures, emergency plans, technical standards, system manuals, operation reports, and troubleshooting manuals) simultaneously under high pressure and time constraints. When real-time data streams, alarm events, and document clauses overlap, the scale and complexity of the information can easily exceed an individual's instantaneous processing capacity, leading to information overload. On the one hand, the document structure is complex and the terminology is dense, resulting in high costs for understanding and retrieval; on the other hand, even experienced personnel find it difficult to reliably extract key constraints and effective conclusions in a short period of time, and the decision-making process still relies heavily on personal memory and experience, thus increasing the risk of omissions and misjudgments in sudden or high-pressure scenarios.

[0003] To alleviate the aforementioned problems, existing technologies have developed knowledge graphs, question-answering systems, and intelligent retrieval tools for specific domains. These tools assist users in locating document content through keyword matching, semantic retrieval, and template-based question answering, and to some extent support simple cross-document correlation analysis. In recent years, some systems have also integrated information from multiple sources, such as technical standards, maintenance records, and log information, attempting to unify retrieval and ranking. However, these systems still have significant shortcomings in the fusion and verification of multi-source information: First, results from different sources are often presented side-by-side in long lists, lacking a quality characterization of source differences. Second, when multi-source information is contradictory, incomplete, or inconsistent in version, there is often a lack of effective conflict detection, complementarity, and automated conflict resolution mechanisms, leading to "finding more" actually exacerbating information overload. More importantly, existing multi-source retrieval and fusion methods generally treat different information sources as homogeneous data, primarily ranking or splicing based on relevance scores. This makes it difficult to explicitly characterize the differences in authority, timeliness, logical consistency, and traceability of each candidate piece of information, thus hindering the provision of interpretable and reusable automatic adjudication criteria when conflicts arise. Summary of the Invention

[0004] The purpose of this invention is to address the problems in the prior art by providing a multi-source information fusion retrieval method and system that combines information credibility profiles. By introducing information credibility profiles, information sources are classified and graded. On this basis, consistency verification and conflict detection are performed on multi-source information to enhance the credibility of multi-source information and resolve conflicts, thereby improving the credibility of retrieval and decision-making results by utilizing multi-source information.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a multi-source information fusion retrieval method combining information credibility profiling is provided, including: Build a credibility profile based on predefined credibility indicators and complete multi-source knowledge classification; Obtain the set of candidate knowledge points corresponding to the query, and divide the set of candidate knowledge points into groups of corresponding knowledge points for conflict detection; Based on the conflict detection results, the non-conflicting control knowledge point group is enhanced by multi-source knowledge classification to form a fused knowledge point set; the conflicting control knowledge point group is dynamically weighted and fused according to multi-source knowledge classification, and human-in-the-loop conflict review and correction are performed. After the conflict resolution is completed, the fused knowledge point set is obtained. The fusion of knowledge points serves as the basis for the query to return the corresponding search results.

[0006] As a preferred embodiment, the predefined credibility indicators include: Authority score W 1. Defined as the basic reliability level of information in terms of professional credibility and standardization, the assignment method is to be offline calibrated and normalized according to the knowledge source category and authority level; Timeliness weight W 2. Defined as the degree of matching between information and the current task time window and the activity of updates. The assignment method is to parse the release time, revision time or version number of the knowledge source, calculate it according to the time decay function and normalize it after combining the update frequency correction. Logical consistency score W 3, defined as the degree of consistency and conflict intensity when cross-referenced with other sources, is assigned by calculating the consistency score and normalizing it based on cross-source comparison results and inference chain verification statistical support and conflict signals; Traceability score W 4. Defined as the degree of location and auditability of information basis and formation process, the value is assigned by calculating and normalizing the completeness of source reference information and process metadata.

[0007] As a preferred embodiment, the steps of performing credibility-weighted fusion of conflicting knowledge point groups according to multi-source knowledge classification, and human-in-the-loop conflict review and correction, to obtain a fused knowledge point set after conflict resolution, include: Based on the established multi-source knowledge classification and metadata, a multi-index credibility profile is constructed for each knowledge point from the corresponding dimensions of credibility indicators. A dynamic weighted fusion strategy is designed to score and compare different candidate conclusions, and a quantifiable and traceable adjudication result is given when conflicts occur. The following weighted credibility function is used to fuse and assign values ​​to each knowledge point: In the formula, the parameters α , β , c , d The weights are generated through weight selection and adaptive fine-tuning. The process of weight selection and adaptive fine-tuning is as follows: the user query is processed to identify the task type and extract scene elements, and a set of basic weights is given based on the pre-configured "task-weight mapping table". α 0, β 0, c 0, d 0), fine-tuning the basic weights based on the conflict level and metadata completeness of the current candidate knowledge point group, and further adjusting the weights. α , β , c , d Normalization is performed to ensure weight comparability and guarantee a+b+c+d =1; After conflict resolution, for situations where the two or more parties in the conflict control group have similar overall credibility and it is difficult to form a stable ranking, or where the conflict involves key security boundaries, core parameter thresholds, or significantly affects operational steps, and even if there are differences in credibility, a decision cannot be made directly based on rules alone, a human-in-the-loop conflict review and correction mechanism is introduced. The scoring results and auditable evidence are output to domain experts or key users, forming a closed-loop adjudication process of "data-model-human". When manual review is triggered, the conflict comparison knowledge point group is presented to the experts, including conflict summary, comparison view and source and credibility score of each candidate item. The experts can choose to adopt a conclusion, give a conditional merging conclusion, request supplementary evidence or re-search, or return to regenerate. The ruling result is fixed as the conclusion and basis annotation of the current conflict group.

[0008] As a preferred option, for authoritative scores W 1. Set basic scores based on the knowledge source category, assigning higher basic weights to textbooks and authoritative publications (0.8-1.0); assigning medium-high basic weights to standards and regulations (0.7-0.9); and assigning lower basic weights to internet search results and content generated by large models (0.3-0.6). Timeliness weight WThe calculation of W2 is based on the latest update timestamp and update frequency of the knowledge point. Specifically, it analyzes the publication time, revision time, or version number of the knowledge source and calculates the time difference Δt from the current time. A preset decay function maps Δt to a time freshness score between [0,1]: content updated within the last year is assigned a weight close to 1, gradually decreasing within 1-3 years, and documents older than 3 years or without update records are reduced to a preset lower limit. For content with multiple version iterations or frequent maintenance records, a certain bonus can be given to the basic freshness score, resulting in a comprehensive W2. The calculation expression is as follows: Δt The time difference between the knowledge point and the current time; f min : Lower limit of basic score for timeliness; n ver The number of times a knowledge point was updated during the statistical period; or max The maximum number of version updates for all knowledge points within the statistical period; or ver Version iteration bonus coefficient;

[0009] Version iteration normalization:

[0010] Then, timeliness weight W 2. Calculate using the following formula:

[0011] Logical consistency score W 3. Calculated through cross-source comparison results; for each knowledge point, the number of supporting matches and the number of conflicting matches are calculated; where the number of supporting matches is the number of times it is consistent with or cited by knowledge points from other sources, and the number of conflicting matches is the number of times it is explicitly contradictory to or marked as suspicious by knowledge points from other sources; consistency and contradiction between two knowledge points are judged through a large language model, and normalized by subtracting a set proportion of conflict penalty coefficient from "number of supporting matches / total number of comparisons", the calculation expression is as follows: N sup Number of supporting matches for knowledge points; N conf Number of conflict matching for knowledge points; N tot Total number of comparisons, satisfying N tot=N sup +N conf ; l Conflict penalty coefficient, satisfying 0 ≤ l ≤1; Then, logical consistency is determined. W 3. Calculate using the following formula:

[0012] Traceability score W 4. Calculations are performed based on the completeness and traceability of the knowledge point's metadata; a set of key metadata fields are assigned to each knowledge record during the calculation. W At time 4, scores are awarded based on the completeness and clarity of key metadata fields. The more complete the key metadata fields and the clearer the path, the higher the score; if only a vague source can be traced, then... W 4. Reduce to the preset lower limit, the calculation expression is as follows: N filled The number of metadata fields that have been filled in for the knowledge points; N total The total number of metadata fields expected to be filled in; b human : Whether the indicator variable has been manually labeled. If manually labeled, it is 1; otherwise, it is 0. or human Additional points awarded manually; Then traceability score W 4. Calculate using the following formula: .

[0013] As a preferred approach, multi-source knowledge is classified according to a predefined credibility index in the following manner: Category A: Authoritative books and textbooks; Category B: Standards, Procedures, and Technical Documents; Category C: Online search engine API results; Category D: Content generated from large models; Among them, categories A and B constitute a reliable offline knowledge base, category C provides online supplementary information that meets timeliness requirements, and category D provides structured generative reasoning solutions.

[0014] As a preferred approach, for Category A and Category B knowledge, hierarchical parsing is performed according to "Document ID-Chapter-Paragraph-Sentence", retaining page numbers and paragraph positions, and extracting formulas and FF charts in the form of placeholders and explanatory text, uniformly forming structured knowledge points with fields of "knowledge point number, text content, type label, document ID, chapter and page number", and the offline knowledge base is composed of all knowledge points; After completing the construction of the offline knowledge base, each standard knowledge point is encoded into a vector and a vector index is established. When a user initiates a query or the system generates a sub-question, the query is first preprocessed and then encoded into a query vector. An approximate nearest neighbor search is performed on the vector index to return several candidate knowledge points of type A and type B that are closest to the query semantic distance. For Category C knowledge, relevant online documents are retrieved in real time when a user submits a query or a sub-question is generated internally by calling the search engine API; semantic vector matching is performed on the query and candidate webpage titles and summaries, and filtering is performed in combination with site whitelists, source domains, and publication time rules to retain knowledge points that are relevant to the query and have reliable sources, and metadata is recorded; For Category D knowledge, when a user initiates a query, the original query text is input into a pre-deployed large language model to generate candidate answers or explanations. During the generation process, the generation behavior is constrained, and the model is required to explicitly provide preconditions or scope of application. The generated results, along with the call time, model version, and main parameters, are recorded as knowledge points with independent source identification.

[0015] As a preferred embodiment, the step of complementarizing and enhancing the non-conflicting contrastive knowledge point groups according to multi-source knowledge classification to form a fused knowledge point set includes: For each B-type knowledge point, perform structured analysis, retrieve the most semantically relevant principle and background evidence fragments from the A-type knowledge points, input the "original knowledge point + evidence fragment + output template constraint" into the large language model, generate structured enhanced knowledge points within the evidence scope, and explicitly associate each enhanced explanation with the corresponding evidence fragment knowledge point. Link the original text of the B-type knowledge point with the above structured explanation field to form a traceable enhanced B-type knowledge point B' record. For Category C knowledge, the search engine API is used to obtain relevant webpage titles, summaries, and text fragments, and paragraphs containing key information are extracted from whitelisted sites or authoritative domains. A large language model is invoked to extract structured knowledge points from each online fragment according to a pre-defined unified field template, and the terminology and unit units are standardized. The structured results are then compared with relevant knowledge points in A+B' using vector recall and rule verification. If the core conclusions and key constraints are consistent, the corresponding online entries are added as the latest supplements or supporting evidence to the corresponding topic. If an online entry provides time-sensitive information missing in A+B', only the missing fields are completed, and the original citation is retained. If there is a significant contradiction with Category A+B' knowledge points in terms of key parameters, applicable conditions, or conclusion polarity, it is marked as a conflict candidate and removed or downgraded. The retained online knowledge points are semantically clustered to remove duplicates, forming a unified, timestamped, and traceable C' set. For knowledge of type D, the large language model outputs candidate answers in structured entries and breaks down the candidate answers into atomic statements. For each statement, the most relevant evidence fragments are recalled in A+B'+C'. Combining rule verification and semantic support discrimination of the large language model, the statements are labeled as supportive, contradictory, or uncertain. Contradictory items are directly deleted or downgraded, and uncertain items trigger secondary generation. The set of retained statements constitutes D', and each statement is accompanied by supporting evidence citations and verification conclusions for tracing and review.

[0016] As a preferred approach, the step of using a fusion knowledge point set as the basis for querying the corresponding search results includes the following steps: Users input queries via natural language, which can take the form of single-sentence questions and answers, complex questions with multiple constraints, or handling requests that include business status descriptions. The input is standardized to generate a structured query description. At the same time, the context is obtained from the business side and attached to the current session as structured fields, forming a unified driving object for the query description and the business context. Based on a unified driving object, the system retrieves baseline knowledge points from offline knowledge base retrieval pathways for types A and B, the latest dynamic information fragments from online knowledge pathways for type C, and candidate explanations or suggestions generated by the large model pathway for type D. It then performs semantic alignment and topic merging of multi-source candidate knowledge points, followed by conflict detection driven by rules and the large model, outputting topic-related knowledge point groups, consistency or conflict markers, and conflict location information. Consistent groups are enhanced through complementary fusion to form a structured enhanced knowledge set. Conflict groups are resolved through dynamic weighted reliability calculations and triggered human-in-the-loop verification and closed-loop correction, forming a unified reliable knowledge package for this query. The query description, business context, and unified trusted knowledge package are combined to construct a constrained prompt input, which drives the large language model to generate task-oriented answers or auxiliary decision-making suggestions, ensuring that the generated results are consistent with the upstream fusion and adjudication conclusions and match the actual business context. In the results presentation stage, in addition to outputting natural language answers, traceable and auditable information support is also provided simultaneously, enabling users to understand the basis and applicable boundaries of the answers while obtaining them.

[0017] As a preferred embodiment, the steps of obtaining the candidate knowledge point set corresponding to the query and dividing the candidate knowledge point set into a matching knowledge point group for conflict detection include: For each candidate knowledge point, extract the metadata representation of "subject object + parameter + applicable conditions" from the text, and encode the metadata text into a semantic vector; perform Top-K nearest neighbor matching and clustering in the vector space, and group knowledge points that are semantically similar and discuss the same object, the same parameter or the same condition into the same comparison knowledge point group. Entries below the similarity threshold are considered not to belong to the same topic comparison range and are not included in the conflict determination of the comparison knowledge point group. For each knowledge point group, pairwise comparisons are performed on the knowledge points, and a computable hard rule test is conducted. Then, for samples that cannot be covered by the hard rule test or are ambiguous, a large model is invoked for semantic discrimination, specifically including: Unify units and dimensions, align parameters and conditions with the same name; check whether intervals intersect or contain each other for numerical claims, and check whether they are mutually exclusive for enumerated or state-type claims; check whether polarity words and mandatory words are opposite in normative expressions. Two knowledge points to be compared within the same knowledge point group are used as evidence pairs. Fixed prompt words are used to ask the large model to output three categories: consistent, conflicting, or uncertain, and at the same time return the conflict trigger fragment. When the large model judges it as conflicting and the confidence level is higher than the threshold, the control pair is marked as conflicting. If the output is uncertain, it is retained as uncertain and an explanation is provided. The system will output the results of the knowledge point group division, the consistency label within the group, the conflict type, and the location of the conflict point.

[0018] Secondly, a multi-source information fusion retrieval system combining information credibility profiling is provided, including: The credibility profile building and multi-source knowledge classification module is used to build credibility profiles according to predefined credibility indicators and complete multi-source knowledge classification. The conflict detection module is used to obtain the set of candidate knowledge points corresponding to the query, and divide the set of candidate knowledge points into a group of reference knowledge points for conflict detection. The module for obtaining the fusion knowledge point set is used to perform complementary enhancement on the non-conflicting control knowledge point group according to the multi-source knowledge classification based on the conflict detection results, forming a fusion knowledge point set; and to perform dynamic weighted fusion of credibility and human-in-the-loop conflict review and correction on the control knowledge point group containing conflicts according to the multi-source knowledge classification, and obtain the fusion knowledge point set after completing the conflict resolution. The search results feedback module is used to provide feedback on the search results based on the integrated knowledge point set.

[0019] Compared with the prior art, the present invention has at least the following beneficial effects: This invention constructs a credibility profile based on predefined credibility indicators and performs multi-source knowledge classification accordingly, achieving classified management of information from multiple sources. For user queries, semantic retrieval is used to obtain candidate fragments from various information sources. Semantically similar information is clustered into candidate knowledge point sets, and consistency verification and conflict detection are performed based on their credibility profiles. According to the conflict detection results, non-conflicting reference knowledge point groups are enhanced through complementary multi-source knowledge classification to form a fused knowledge point set. Conflicting reference knowledge point groups undergo dynamic credibility weighted fusion and human-in-the-loop conflict review and correction through multi-source knowledge classification, resulting in a fused knowledge point set after conflict resolution. Through the above processing, a unified fused knowledge point set is obtained. Based on this, the intelligent information retrieval system generates answers and decision-making support information for the current task, explicitly providing the main information sources and their credibility clues, enabling users to understand the basis and scope of application while obtaining the answer. This invention combines a multi-source information fusion retrieval method with information credibility profiling to improve the credibility of retrieval and decision-making results by utilizing information from multiple sources. It overcomes the shortcomings of existing technologies in multi-source information credibility control, result fusion and verification. By verifying the consistency of multi-source heterogeneous information, supplementing complementarity and resolving conflicts, it provides high-quality and traceable fusion results for upper-level retrieval and decision-making systems, thereby improving the credibility and interpretability of the output from a mechanism perspective. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 The flowchart of the multi-source information fusion retrieval method combining information credibility profiling in this embodiment of the invention is shown in the figure. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, those skilled in the art can obtain other embodiments without creative effort.

[0023] Please see Figure 1 The multi-source information fusion retrieval method of this invention, which combines information credibility profiling, mainly includes the following steps: S1. Construct a credibility profile according to predefined credibility indicators and complete multi-source knowledge classification; S2. Obtain the set of candidate knowledge points corresponding to the query, and divide the set of candidate knowledge points into a group of corresponding knowledge points for conflict detection. S3. Based on the conflict detection results, the non-conflicting control knowledge point group is enhanced by multi-source knowledge classification to form a fused knowledge point set; the conflicting control knowledge point group is dynamically weighted and fused according to multi-source knowledge classification, and human-in-the-loop conflict review and correction are performed. After the conflict resolution is completed, the fused knowledge point set is obtained. S4. Use the integrated knowledge point set as the basis to provide feedback on the corresponding search results.

[0024] In one possible implementation, the predefined credibility index in step S1 includes: Authority score W 1. Defined as the basic reliability level of information in terms of professional credibility and standardization, the assignment method is to be offline calibrated and normalized according to the knowledge source category and authority level; Timeliness weight W 2. Defined as the degree of matching between information and the current task time window and the activity of updates. The assignment method is to parse the release time, revision time or version number of the knowledge source, calculate it according to the time decay function and normalize it after combining the update frequency correction. Logical consistency score W 3, defined as the degree of consistency and conflict intensity when cross-referenced with other sources, is assigned by calculating the consistency score and normalizing it based on cross-source comparison results and inference chain verification statistical support and conflict signals; Traceability score W 4. Defined as the degree of location and auditability of information basis and formation process, the value is assigned by calculating and normalizing the completeness of source reference information and process metadata.

[0025] To facilitate standardized processing, all credibility indicators are normalized to the [0,1] interval. The specific calculation methods for the four indicators are as follows: For authoritative scores W1. Set basic scores based on the knowledge source category, assigning higher basic weights to textbooks and authoritative publications (0.8-1.0); assigning medium-high basic weights to standards and regulations (0.7-0.9); and assigning lower basic weights to internet search results and content generated by large models (0.3-0.6). Timeliness weight W The calculation of W2 is based on the latest update timestamp and update frequency of the knowledge point. Specifically, it analyzes the publication time, revision time, or version number of the knowledge source and calculates the time difference Δt (in years or months) from the current time. A preset decay function (piecewise or exponential decay function) maps Δt to a time freshness score between [0,1]. Content updated within the last year is assigned a weight close to 1, gradually decreasing within 1-3 years, and documents older than 3 years or without update records are reduced to a preset lower limit. For content with multiple version iterations or frequent maintenance records, a certain bonus can be given to the basic freshness score, resulting in a comprehensive W2. The calculation expression is as follows: Δt The time difference between the knowledge point and the current time (unit: years); f min The lower limit of the basic score for timeliness is set at 0.2. n ver The number of times a knowledge point was updated during the statistical period; or max The maximum number of version updates for all knowledge points within the statistical period; or ver Version iteration bonus coefficient, ranging from 0.05 to 0.1;

[0026] Version iteration normalization:

[0027] Then, timeliness weight W 2. Calculate using the following formula:

[0028] Logical consistency score W3. Calculated through cross-source comparison results; for each knowledge point, the number of supporting matches and the number of conflicting matches are calculated; where the number of supporting matches is the number of times it is consistent with or cited by knowledge points from other sources, and the number of conflicting matches is the number of times it is explicitly contradictory to or marked as suspicious by knowledge points from other sources; consistency and contradiction between two knowledge points are judged through a large language model, and normalized by subtracting a set proportion of conflict penalty coefficient from "number of supporting matches / total number of comparisons", the calculation expression is as follows: N sup Number of supporting matches for knowledge points; N conf Number of conflict matching for knowledge points; N tot Total number of comparisons, satisfying N tot =N sup +N conf ; l Conflict penalty coefficient, satisfying 0 ≤ l ≤1; Then, logical consistency is determined. W 3. Calculate using the following formula:

[0029] If a piece of knowledge appears only in a few sources and is repeatedly identified as conflicting, its W 3. Reduce; conversely, multi-source consistent support for knowledge W 3 is close to 1.

[0030] Traceability score W 4. Calculations are performed based on the completeness and traceability of the knowledge point's metadata; a set of key metadata fields are assigned to each knowledge record, such as document ID, book title / standard number, chapter / clause number, page number, publication date, URL, or database primary key. (This is part of the calculation.) W At 4 o'clock, scores are awarded based on the completeness and clarity of key metadata fields. The more complete the key metadata fields and the clearer the path, the higher the score; if only a vague source can be traced (such as "web search results" without a specific origin), then... W 4. Reduce to a preset lower limit. In the specific implementation, the base score is calculated based on "number of filled metadata fields / total number of expected fields", and additional points are given to items that have been manually annotated or recorded through a strict process, ultimately forming a normalized score. W 4. The calculation expression is as follows: N filledThe number of metadata fields that have been filled in for the knowledge points; N total The total number of metadata fields expected to be filled in; b human : Whether the indicator variable has been manually labeled. If manually labeled, it is 1; otherwise, it is 0. or human Additional points awarded manually; Then traceability score W 4. Calculate using the following formula:

[0031] Based on the typical performance of information sources in the four credibility-related quality attributes of authority, timeliness, logical consistency, and traceability, this invention performs hierarchical modeling of multi-source information, classifying knowledge sources into the following four categories: Category A consists of authoritative books and textbooks, including classic textbooks and authoritative monographs in the field. These books have a complete theoretical system and a clear structure, and are highly authoritative and logically consistent, but have limited coverage of the latest engineering practices. Category B consists of standards, procedures, and technical documents, including national / industry standards, enterprise procedures, technical specifications, and operation manuals. These documents are highly relevant to engineering, can directly guide practice, and are sensitive to version and timeliness. Category C consists of online search engine API results, including web page content and online reports obtained through general search engines or enterprise internal document search interfaces. These results have broad coverage and high timeliness, but their quality varies and there is a lot of noise. Category D consists of content generated by a large model. These are answers, explanations, and suggestions generated by a large language model after receiving a query. They are natural in expression and highly structured, but there is a risk of false generation and inference bias.

[0032] Among them, types A and B constitute a reliable offline knowledge base, type C provides online supplementary information that meets timeliness requirements, and type D provides structured generative reasoning solutions. These four types of information sources together serve as the basis for subsequent intelligent fusion and conflict resolution.

[0033] For Category A and Category B knowledge, this invention adopts a unified process to establish an offline professional knowledge base: using tools such as PDFplumber, the original PDF file is parsed hierarchically according to "document ID-chapter-paragraph-sentence", retaining page numbers and paragraph positions, and extracting formulas and FF charts in the form of placeholders and explanatory text, uniformly forming structured knowledge points with fields of "knowledge point number, text content, type label, document ID, chapter and page number", and the offline knowledge base is composed of all knowledge points; After completing the construction of the offline knowledge base, this invention uses bge-large-zh to encode each standard knowledge point into a 1024-dimensional vector and builds a vector index based on FAISS. When a user initiates a query or the system generates a sub-question internally, the query is first preprocessed (noise symbols are removed, terminology is standardized, and key parameters and entities are retained), and then encoded into a 1024-dimensional query vector. An approximate nearest neighbor search is performed on the vector index to return the top 5 to 10 candidate knowledge points of type A and type B that are closest to the query semantic distance. For Category C knowledge, relevant online documents are retrieved in real time when a user submits a query or a sub-question is generated internally by calling search engine APIs (such as Google Search API or enterprise internal document search interface); semantic vector matching is performed on the query and candidate webpage titles and summaries using bge-large-zh or lightweight coding models, and low-quality results are filtered in combination with site whitelists, source domains, publication time and other rules, and finally about 2 to 4 Category C knowledge points that are highly relevant to the query and have reliable sources are retained, and their URLs, crawling time and summaries and other metadata are recorded; For Category D knowledge, when a user initiates a query, the original query text (with optional additional task scenarios, user roles, and key constraints) is input into a pre-deployed large language model, which generates 1 to 3 candidate answers or explanations. During the generation process, the generation behavior is constrained by setting temperature, maximum length, and fixed output templates, and the model is required to explicitly provide preconditions or applicable scope as much as possible. The generated results, along with the call time, model version, and main parameters, are recorded as knowledge points with independent source identification.

[0034] In one possible implementation, after obtaining the set of candidate knowledge points corresponding to the same query in step S2, consistency checks and conflict detection are performed on the candidate content, and the verification label (consistent / conflicting / uncertain) and conflict point location are output for each knowledge point. This embodiment of the invention proposes a vector-based semantic alignment and topic merging method, including extracting metadata representations of "subject object + parameter + applicable conditions" (e.g., device / component name, parameter name, operating conditions) from the text of each candidate knowledge point, and encoding the metadata text into a semantic vector using bge-large-zh; performing Top-K nearest neighbor matching and clustering in the vector space, grouping knowledge points that are semantically similar and discuss the same object, parameter, or condition into the same comparison knowledge point group; entries with a similarity threshold below a certain threshold (e.g., 0.75~0.85) are considered not to belong to the comparison scope and are not included in the conflict determination of the comparison knowledge point group. Meanwhile, this invention proposes a conflict detection method based on rules and a large language model, which includes pairwise comparison of knowledge points within each knowledge point group, performing a computable hard rule test, and then calling a large model for semantic discrimination for samples that cannot be covered by the hard rule test or that are ambiguous. Specifically, it includes: Rule consistency check: Unify units and dimensions (e.g., m) km, ℃ K), align parameters and conditions with the same name; for numerical claims, check whether the intervals intersect or contain each other (if the intervals do not intersect, it is considered a conflict); for enumeration or state claims, check whether they are mutually exclusive; for normative expressions, check whether polarity words and mandatory words are opposite (e.g., "should / should not", "allow / prohibit", "must / strictly prohibit" are opposite and are considered a conflict). Large model conflict detection: Within the same set of knowledge points to be compared, two knowledge points are used as evidence pairs. The large model is asked to output three categories using fixed prompts: consistent, conflicting, or uncertain, and to return conflict trigger fragments (e.g., contradictory parameter values, opposite constraints, mutually exclusive conclusions). When the large model determines that the pair is conflicting and the confidence level is higher than a threshold (e.g., 0.7), the pair is marked as conflicting. If the output is uncertain, it is retained as uncertain and accompanied by an explanation (missing conditions, inconsistent definitions, insufficient context, etc.). Output: The results of knowledge point grouping, consistency labels within the group (consistent / conflicting / uncertain), conflict types (numerical / conditional / polarity / definitional differences), and conflict point locations (corresponding original text fragments and metadata references) serve as input for subsequent processing.

[0035] For the control knowledge point groups in S2 that have been determined to have no logical conflicts and are internally related, since each control group contains knowledge points from different sources, based on the aforementioned multi-source knowledge classification, this invention proposes an automated complementary information supplementation method based on a large language model and a vectorization model. This method can gradually enhance the internal consistency and source traceability of knowledge points, providing a high-quality fusion knowledge point set for the subsequent final answer generation.

[0036] In one possible implementation, step S3, which involves complementary enhancement of the non-conflicting contrastive knowledge point group according to multi-source knowledge classification to form a fused knowledge point set, includes: A Method for Enhancing the Interpretation of Type B Regulations Clauses Based on Type A Theoretical Knowledge In this step, the invention uses a large language model to perform interpretive enhancement on the B-type procedure clauses, generating an enhanced version B'. Specifically, each B-type knowledge point is structurally parsed, extracting key terms, parameter names, thresholds / units, and applicable conditions. Based on this, the bge-large-zh model is used to recall the most semantically relevant principle-based and background evidence fragments (Top-K) from the A-type knowledge points. The "original knowledge point + evidence fragment + output template constraints" are input into the large language model, requiring the model to generate structured enhanced knowledge points only within the scope of the evidence, including but not limited to "background explanation, principle basis, boundary conditions, and precautions." Each enhanced explanation is explicitly associated with the corresponding evidence fragment knowledge point (e.g., citing evidence number / location). The original text of the B-type knowledge point is linked with the above-mentioned structured interpretation fields to form a traceable B' record.

[0037] A Complementary Enhancement and Filtering Correction Method for Class C Online Results Based on A+B' Knowledge Base Using the generated A+B' knowledge points as the semantic foundation, the online search results for category C are subjected to "structured extraction - consistency verification - complementary completion" to obtain the corrected C': The search engine API is used to obtain the webpage titles, summaries, and text fragments related to the query, and paragraphs containing key information such as "parameter thresholds, applicable conditions, version changes, announcement numbers, and key points of handling" are extracted from whitelisted sites or authoritative domains; a large language model is invoked to extract each online fragment into structured knowledge points (object / component, parameter name-unit-value range, applicable prerequisites, exception conditions, key conclusions, publication time / version number, source URL and original citation location) according to a preset unified field template, and the terminology is corrected. The system is standardized with unit dimensions. The structured results are then compared with the relevant knowledge points in A+B' using vector recall and rule verification. If the core conclusions and key constraints are consistent, the corresponding online entries are attached to the corresponding topics as the latest supplements or evidence. If the online entries provide time-sensitive information missing in A+B' (such as new version thresholds, latest announcement numbers, and updated applicable boundaries), only the missing fields are filled in and the original citations are retained. If there is a significant contradiction with A+B' type knowledge points in terms of key parameters, applicable conditions, or conclusion polarity, they are marked as conflict candidates and removed or downgraded. The retained online knowledge points are semantically clustered to remove duplicates, forming a unified set C' with timestamps and traceable sources.

[0038] Constraint Verification and Correction Methods for Generated Content of Large Models For content generated by the large model in category D, the system adopts a controlled process of "generation-verification-rewriting" to obtain D': The large language model outputs candidate answers in structured entries (divided into conclusion points / steps) and breaks down the candidate answers into atomic statements; for each statement, the most relevant evidence fragments are recalled in A+B'+C', and the statements are marked as supporting, contradictory, or uncertain by combining rule verification (scope / unit / preconditions / prohibited items) and semantic support discrimination of the large language model. Contradictory items are directly deleted or downgraded, and uncertain items trigger secondary generation (backfilling with available evidence and constraints, requiring the statement to be rewritten only within the scope of evidence). The set of retained statements constitutes D', and each statement is accompanied by supporting evidence citations and verification conclusions for subsequent tracing and review.

[0039] After step S2, the system has completed consistency judgment and conflict detection for candidate knowledge points from different knowledge sources, and aggregated items judged to have explicit or implicit contradictions into "conflict comparison knowledge point groups" (e.g., giving incompatible intervals for the same parameter, giving opposite conclusions for the same operation, or having mutually exclusive applicable conditions). For this type of conflict group, the system needs to further complete "conflict resolution": under the constraint of credibility index, determine the main conclusion that should be accepted (or give a conditional conclusion), and retain the unaccepted items with a mark and audit record, thereby achieving controllable conflict resolution.

[0040] Credibility-based dynamic weighted fusion method and automatic knowledge point conflict resolution In one possible implementation, step S3 introduces an automated credibility assessment mechanism based on knowledge source credibility profiles: based on the source categories and metadata determined in step S1, a multi-indicator credibility profile is constructed for each knowledge point from the dimensions of authority, timeliness, logical consistency and traceability, and on this basis, a dynamic weighted fusion strategy is designed to automatically score and compare different candidate conclusions, and to give a quantifiable and traceable ruling result when conflicts occur.

[0041] Step S3 describes the dynamic weighted fusion of the conflicting knowledge point groups according to multi-source knowledge classification, followed by human-in-the-loop conflict review and correction. After conflict resolution, a fused knowledge point set is obtained, including: The following weighted credibility function is used to merge and assign values ​​to each knowledge point: In the formula, the parameters α , β , c , dThe weights are generated through weight selection and adaptive fine-tuning. The process of weight selection and adaptive fine-tuning is as follows: the user query is subjected to task type identification and scenario element extraction (such as: knowledge Q&A / fault diagnosis / debugging analysis / compliance verification, whether it is a security sensitive issue, whether it has strong timeliness, whether it has strong traceability requirements, etc.), and a set of basic weights is given according to the pre-configured "task-weight mapping table". α 0, β 0, c 0, d 0), for example, compliance checks improve α and d Real-time processing improves β Analytical problems with dense conflicts or long reasoning chains improve c The basic weights are fine-tuned based on the degree of conflict and the completeness of metadata for the current candidate knowledge point groups (if the conflict is severe, the weights are increased). c If the source field is missing, adjust upwards. d (constraint weights), and on α , β , c , d Normalization is performed to ensure weight comparability and guarantee a+b+c+d =1; Through the above-mentioned automatic conflict resolution and dynamic weighted fusion method based on credibility profile, this invention can not only enhance the complementarity of multi-source knowledge, but also quantitatively compare and automatically screen the conclusions given by different sources, and give highly credible and consistent conclusions in most common scenarios.

[0042] Human-in-the-loop conflict review and closed-loop correction in high-uncertainty scenarios Even after conflict resolution, there may still be two types of "high uncertainty" conflict control groups: First, the two or more parties involved in the conflict have similar overall credibility, making it difficult to form a stable ranking; Second, the conflict involves key security boundaries, core parameter thresholds, or significantly affects operational procedures, and even if there are differences in credibility, a decision cannot be made directly based solely on rules.

[0043] In response to the above situation, the present invention introduces a human-in-the-loop conflict review and correction mechanism, which outputs the scoring results and auditable evidence to domain experts or key users, forming a closed-loop adjudication process of "data-model-human". When manual review is triggered, the conflict comparison knowledge point group is presented to the experts, including a conflict summary (conflict points and difference fields), a comparison view (showing the aligned logical chain and key sentence evidence side by side), and the source and credibility score of each candidate item. The experts can perform operations such as adopting a conclusion, giving a conditional merging conclusion (supplementing applicable premises / thresholds), requesting supplementary evidence or re-searching, and returning for regeneration. They can also provide supplementary explanations. The ruling result is fixed as the conclusion and basis annotation of the current conflict group.

[0044] Through the above design, this invention organically combines automated scoring with manual review: in conventional scenarios, it relies on automatic adjudication to improve efficiency, and in high-risk or high-uncertainty scenarios, it introduces a human-in-the-loop approach to ensure controllability and traceability, thereby significantly improving the reliability and engineering usability of multi-source conflict resolution.

[0045] In one possible implementation, when using the fused knowledge point set as the basis for the retrieval results in step S4, the process of "user query - multi-source retrieval - consistency verification - (complementary fusion + conflict resolution) - interpretable output" is integrated into one, achieving reliable compression, traceable organization, and auditable presentation of multi-source information, providing reliable support for rapid decision-making in information overload scenarios, including: Users input queries via natural language, which can take the form of single-sentence questions and answers, complex questions with multiple constraints, or handling requests that include business status descriptions. The input undergoes standardized processing: noise reduction and error correction, standardization of technical terms, and extraction of key entities and constraints (such as equipment / components, parameter names and units, thresholds, time windows, status conditions, and regional ranges) to generate a structured query description. At the same time, the context (such as user roles and permissions, business modules, key operational status summaries, and major alarm overviews) is obtained from the business side and attached to the current session as structured fields, forming a unified driving object for the query description and business context. Based on a unified driving object, the system retrieves baseline knowledge points from offline knowledge base retrieval pathways for types A and B, the latest dynamic information fragments from online knowledge pathways for type C, and candidate explanations or suggestions generated by the large model pathway for type D. It then performs semantic alignment and topic merging of multi-source candidate knowledge points, followed by conflict detection driven by rules and the large model, outputting topic-related knowledge point groups, consistency or conflict markers, and conflict location information. Consistent groups are enhanced through complementary fusion to form a structured enhanced knowledge set. Conflict groups are resolved through dynamic weighted credibility calculations, triggering human-in-the-loop review and closed-loop correction when necessary, ultimately forming a unified credible knowledge package for this query (including main conclusion, applicable conditions, key constraints, evidence citations, credibility clues, and adjudication records).

[0046] The query description, business context, and unified trusted knowledge package are combined to construct a constrained prompt input, which drives the large language model to generate task-oriented answers or auxiliary decision-making suggestions. The output content may include: key conclusions, recommended steps, precautions and risk warnings, branch suggestions under different conditions, and key parameter boundaries and applicable scope, thereby ensuring that the generated results are consistent with the upstream fusion and adjudication conclusions and match the actual business context. During the results presentation phase, in addition to outputting natural language answers, traceable and auditable supporting information is provided simultaneously: each key conclusion is labeled with its main supporting knowledge points and source category (A / B / C / D), document / clause / page number or URL, publication time / version information, comprehensive credibility clues, conflict resolution methods (automatic adjudication or human-loop confirmation), and necessary uncertainty prompts. This information can be presented through tags, folded cards, footnotes, or structured evidence lists, enabling users to "quickly obtain answers" while clearly understanding the "basis and applicable boundaries of the answers," thereby significantly improving the credibility, transparency, and usability of the system output.

[0047] The invention will be further explained below through a Q&A scenario on troubleshooting ship mechanical and electrical equipment malfunctions.

[0048] This embodiment describes the application of the present invention in a knowledge-based question-and-answer scenario for fault diagnosis and maintenance of marine electromechanical equipment. This embodiment deploys the "multi-source information classification, intelligent fusion, and human-in-the-loop verification enhancement method combining information credibility profiling" into the marine electromechanical support system. This provides highly reliable knowledge-based question-and-answer services for fault diagnosis, troubleshooting, and handling to engine room crew, watch engineers, and maintenance personnel, alleviating information overload caused by the coexistence of multiple engine models, multiple versions of manuals, and historical records. The steps are as follows: S1 Credibility Profile Construction and Multi-Source Knowledge Classification S1.1 Definition of Credibility Profile In this embodiment, the "credibility profile" is used to characterize the reliability level and applicability boundaries of each knowledge point in the ship maintenance scenario. The credibility profile includes at least four indicators: authority, timeliness, logical consistency, and traceability, as well as their comprehensive scores, and is linked to the source location (e.g., specific manual version, chapter / clause, page number, or announcement number and publication date) for subsequent fusion weighting, conflict resolution, and audit traceability.

[0049] S1.2 Definition of Knowledge Source Types This embodiment categorizes knowledge sources into four types and models them accordingly: Category 1 (Theory and Mechanism): Equipment working principles, system mechanism explanations, typical fault mechanism analysis, teaching materials, etc., used to provide causal explanations and boundary conditions.

[0050] The second category (procedures and manuals): manufacturer's operation and maintenance manuals, fault code explanations, company maintenance procedures, standard operating procedures, emergency response cards, etc., are used to directly guide on-site operations.

[0051] The third category (online updates and notifications): Manufacturer technical bulletins, revision notes, service announcements, and the latest case entries in the company's internal knowledge portal, used to supplement and update information and new problem patterns.

[0052] The fourth category (large language model generation): Candidate diagnostic approaches, investigation steps, risk warnings, etc., generated by the large language model based on user questions, are used as "candidate content to be verified" and are not directly used as factual evidence.

[0053] S1.3 Establishment of Offline Professional Knowledge Base (First / Second Type of Knowledge Source) The system will store the first and second categories of data offline: including instruction manuals and maintenance manuals for main units, auxiliary units, generator sets, pumps and valves, cooling and lubrication subsystems, fault code manuals, maintenance procedures and emergency response cards, and historical maintenance records (which can be anonymized).

[0054] When loading documents offline, the system performs structured parsing and breaks them down into fine-grained knowledge points, retaining metadata such as chapter / clause numbers, page numbers, version numbers, applicable working conditions, key parameter ranges and units; it performs semantic deduplication and merging of knowledge points that are repeated or highly similar across documents, forming a unified offline knowledge point set and index.

[0055] S1.4 Semantic Retrieval and Query Methods for Offline Knowledge Bases (Type 1 / Type 2) After a user submits a question, the system performs terminology standardization and entity extraction (equipment name, model, fault code, parameter name, operating condition description, etc.) on the query. Then, it uses a semantic encoding model to vectorize the query and offline knowledge points, and retrieves the most semantically relevant set of knowledge points from the vector index. The returned results also carry the source category, clause / page number / version position, and association link (e.g., "fault code - possible cause - inspection steps - security restrictions"), providing structured input for subsequent consistency judgment and fusion.

[0056] S1.5 Online Knowledge Acquisition Methods (Category 3 / Category 4 Knowledge Sources) The third type of online acquisition: The system retrieves the latest entries by searching for "equipment / model + fault code / symptom + keywords" through the enterprise's internal document retrieval interface, the manufacturer's announcement entry, or a controlled online retrieval interface. It captures metadata such as title, abstract, publication date, and source domain / organization, and performs site whitelisting and time window filtering to obtain a small number of highly relevant online fragments.

[0057] The fourth type of generation and acquisition: The system calls the pre-deployed large language model, takes the user's original problem (with a small amount of working condition summary attached) as input to generate a small number of candidate diagnosis / treatment suggestions, and records the model version, generation parameters and timestamp as objects for subsequent consistency verification and validation.

[0058] S2 Consistency Verification and Conflict Detection Method for Multi-Source Knowledge S2.1 Vectorized Semantic Alignment and Same-Topic Merging The system converts candidate knowledge points from four sources into comparable semantic representations and aligns and merges them according to semantic similarity: items with different expressions but the same theme are aggregated into the same "comparison knowledge point group" (e.g., multi-source descriptions around the same fault code, the same component, and the same parameter threshold), providing comparison units for subsequent conflict detection.

[0059] S2.2 Conflict Detection Based on Rules and Large Language Models For each set of knowledge points to be compared, the system first uses rules to quickly screen for conflicts: aligning the numerical ranges and units of parameters with the same name, checking whether the applicable working conditions are mutually exclusive, identifying whether the polarities of conclusions such as "allowed / prohibited" and "should / should not" are opposite, and checking whether the order of handling is mutually exclusive. For cases where the rules are difficult to determine, the system calls a large language model to perform semantic discrimination, outputs "consistent / conflicting / uncertain" and locates the conflict point (whether the conflict occurs in the preconditions, parameter values ​​or operational conclusions), and finally forms a set of "conflicting knowledge point sets".

[0060] S3 Knowledge Point Complement and Fusion Method Based on Multi-Source Knowledge Classification S3.1 A Method for Enhancing the Interpretation of Second-Type Procedural Clauses Based on First-Type Theoretical Knowledge The system directly calls the large language model to perform interpretive enhancement on the second type of procedure clauses: For each procedure clause, it first retrieves evidence fragments that match its terminology, parameters and working conditions from the first type of mechanism knowledge, and then inputs "procedure clause + evidence fragment + fixed structured template" into the large language model, which generates an enhanced clause record (such as generating fields such as "principle basis, applicable boundaries, precautions, common mistakes, and association tips with related clauses") to form an enhanced version, which is used to complete "how to do it" into "why do it this way, and under what conditions do it this way".

[0061] S3.2 Online Result Complementary Enhancement and Filtering Correction Method Based on the Knowledge Base of "Theory + Enhancement Procedure" The system first "makes usable" the third type of online fragments, and then "absorbs or removes" them: through information extraction and structuring with a large language model, the key objects, parameter thresholds, applicable conditions, versions, and release dates in the online fragments are extracted as standard fields; then, they are aligned and compared with the knowledge base of "theory + enhanced procedures". If the online content provides more recent version revision points, batch difference explanations, typical case symptoms, and troubleshooting tips, and does not contradict the base, it is attached to the corresponding topic as a supplementary point; if the online content has obvious conflicts with the base in terms of key thresholds, prohibition conditions, and processing order, or if the source is unreliable, it is marked as conflicting and its adoption priority is reduced or it is removed, thus obtaining the online set after complementary enhancement and filtering correction.

[0062] S3.3 Constraint Verification and Correction Methods for Large Model Generated Content The system breaks down the candidate answers generated by the large language model into several verifiable statements (reasons, steps, thresholds, risk points, etc.) and checks them one by one with the upstream aligned knowledge base: those that can be supported by evidence and do not violate security restrictions are retained; those that contradict the evidence or violate hard constraints are deleted or require the model to be rewritten within the scope of the evidence; finally, the generated content after verification and correction is obtained, and the corresponding evidence source is attached to each retained statement for easy traceability.

[0063] S4 A Human-In-Loop Multi-Source Knowledge Conflict Resolution Method Based on Credibility Indicators S4.1 Credibility-based dynamic weighted fusion method and automatic knowledge point conflict resolution The system calculates the overall credibility of each candidate item in the conflict-of-contrast knowledge point group and performs dynamic weighted judgment. The overall credibility is obtained by linear weighting of four indicators: Overall credibility = α ·Authority+ β Timeliness+ c Logical consistency+ d • Traceability.

[0064] The weighting coefficients are derived from: the system's pre-configured task types (e.g., "fault handling Q&A" emphasizes authority and traceability), user roles and risk level configurations (e.g., security-sensitive issues increase the weight of logical consistency), and adaptive adjustments from writing back historical adjudication data (human-in-the-loop adjudication results can be used to fine-tune weights and thresholds).

[0065] In cases where there are significant differences in overall credibility and the severity of conflict is low, the system automatically selects the main conclusion and retains the unaccepted items as "backup information" in the audit record.

[0066] S4.2 Human-in-the-loop conflict review and closed-loop correction in high-uncertainty scenarios When conflicts involve safety-sensitive items, significant differences in key thresholds, or insufficient confidence in automatic adjudication, the system generates a conflict summary and comparison view (conclusions from different sources, applicable operating conditions, citation locations, credibility indicators, and scores), which is then submitted to the duty engineer or senior marine engineer for review. Manual adjudication allows for options such as adoption, merging and revision, or returning for supplementary retrieval; the adjudication results are written back to the system to update the weight settings, thresholds, and source priorities for similar conflicts, forming a closed-loop optimization.

[0067] S5 Integrated Knowledge Application System for Intelligent Retrieval and Decision Support S5.1 User Query Input and Context Modeling Users initiate troubleshooting questions in natural language (e.g., "The generator is showing a certain fault code and the current is fluctuating, how do I troubleshoot it?"). The system extracts the equipment name / model, fault code, parameter trends, and current operating conditions, and can access the business context (alarm summary, key sensor readings, and recent maintenance record summary) to form a structured query description, driving subsequent multi-source retrieval and processing links.

[0068] S5.2 Response Generation and Result Presentation Based on Fusion and Verification Results The system uses the fused trusted knowledge as a constrained context to generate the final response: it outputs recommended investigation paths, key checkpoints, security restrictions and risk warnings, along with clues about the source and credibility of the main evidence (source category, version / release date, citation location, comprehensive score, and whether it has been manually reviewed), presented in a structured manner to facilitate on-site execution and traceability.

[0069] The invention will be further illustrated below using a drone search and rescue scenario.

[0070] This embodiment describes the application of the present invention in a drone search and rescue scenario. This embodiment will be deployed on a drone search and rescue collaboration platform to provide on-site search and rescue personnel with highly reliable retrieval and auxiliary decision-making services for mission planning, risk assessment, and standardized operations. This is used to alleviate information overload caused by the overlap of weather, airspace restrictions, drone manuals, and procedural specifications. The steps are as follows: S1 Credibility Profile Construction and Multi-Source Knowledge Classification S1.1 Definition of Credibility Profile Credibility profiles are used to characterize the reliability and applicability of each piece of knowledge / constraint in search and rescue missions. They include authority, timeliness, logical consistency, traceability, and comprehensive score, and are linked to the source and validity period (e.g., the effective time window of weather warnings, the validity period of airspace restrictions, and the issuing agency).

[0071] S1.2 Definition of Knowledge Source Types Category 1 (Principles and Methods): Flight principles, meteorological fundamentals, search and rescue methods, sensor imaging mechanisms and applicable boundaries, etc.

[0072] Category 2 (Specifications and Manuals): Aircraft flight manuals, ground station operating procedures, safety regulations, emergency response cards, mission procedures and compliance requirements, etc.

[0073] The third category (online dynamic information): real-time public information or institutional interface information such as weather warnings, temporary airspace restrictions, maps / topography / road hydrology, etc.

[0074] The fourth category (large language model generation): The model generates candidate search strategies, division of labor suggestions, risk warnings, etc. based on the query, which serve as candidate content to be verified.

[0075] S1.3 Establishment of Offline Professional Knowledge Base (First / Second Type of Knowledge Source) The system stores the model manual, link configuration instructions, standard search and rescue procedures and checklists, emergency response plans, sensor usage instructions and identification precautions offline; it parses and breaks down the information into fine-grained knowledge points and retains the version number, chapter and clause, applicable operating conditions (wind speed, visibility, load, range, etc.) and key parameter range. After semantic deduplication and merging, a searchable offline index is formed.

[0076] S1.4 Semantic Retrieval and Query Methods for Offline Knowledge Bases (Type 1 / Type 2) The system extracts and standardizes elements such as aircraft model, payload, mission area, weather conditions, and range constraints from queries. It then retrieves the most relevant manual clauses, process steps, and principle boundaries through semantic retrieval and returns the source location and related links (constraints - operation steps - risk warnings).

[0077] S1.5 Online Knowledge Acquisition Methods (Category 3 / Category 4 Knowledge Sources) The third type of online acquisition involves obtaining the latest information through interfaces such as meteorological services, map geographic information, and temporary airspace restrictions, and recording the source, update time, coverage area, and validity period.

[0078] The fourth type of generation and acquisition involves calling a large language model to generate a small number of candidate suggestions based on the user's question and task summary (region, time window, device model, battery life constraints, etc.), and recording the model version, generation parameters, and timestamp.

[0079] S2 Consistency Verification and Conflict Detection Method for Multi-Source Knowledge S2.1 Vectorized Semantic Alignment and Same-Topic Merging Candidate entries from multiple sources are aligned and grouped into a control group based on semantic similarity (e.g., around themes such as "whether takeoff is possible", "maximum permissible wind speed / rainfall conditions", "return-to-base power threshold", "whether a certain area is accessible"), to unify the expression of objects, units and condition descriptions.

[0080] S2.2 Conflict Detection Based on Rules and Large Language Models First, use rules to screen for conflicts (incompatible threshold ranges, mutually exclusive validity periods, contradictory conclusions of no-fly / restriction, mutually exclusive preconditions, mutually exclusive steps, etc.). Then, for items that are difficult to determine by rules, call a large language model to identify consistency / conflict / uncertainty and locate the conflict points, forming a conflict comparison knowledge point group.

[0081] S3 Knowledge Point Complement and Fusion Method Based on Multi-Source Knowledge Classification S3.1 A Method for Enhancing the Interpretation of Second-Type Procedural Clauses Based on First-Type Principles The system directly calls the large language model, organizes "process clauses + principle evidence + fixed templates" as input, and generates interpretive enhancement content (principles, applicable boundaries, precautions, common misuse points, etc.) for the second type of standardized steps, improving the understandability and enforceability of the clauses under on-site conditions.

[0082] S3.2 Online Result Complementary Enhancement and Filtering Correction Method Based on the Knowledge Base of "Principles + Enhancement Procedures" The system structures the third type of online dynamic information into usable constraints (effective time window, regional scope, threshold, risk level, etc.) and aligns it with the knowledge base: consistent and more timely and specific online information is absorbed as supplementary constraints (such as updated short-term warnings and revisions to temporary restriction scopes); online information that conflicts with key restrictions of the base or whose sources do not meet the trust requirements is downgraded or removed, while the source and validity period are retained for auditing.

[0083] S3.3 Constraint Verification and Correction Methods for Large Model Generated Content The generated suggestions are split and verified according to "strategy / division of labor / threshold / risk point": those that meet the upstream constraints and can be supported by evidence are retained; those that violate airspace restrictions, meteorological risk constraints or aircraft type restrictions are deleted or required to be rewritten within the scope of evidence, and the final output is the verification result with evidence reference.

[0084] S4 A Human-In-Loop Multi-Source Knowledge Conflict Resolution Method Based on Credibility Indicators S4.1 Credibility-based dynamic weighted fusion method and automatic knowledge point conflict resolution The system calculates overall credibility based on four indicators: authority, timeliness, logical consistency, and traceability. It dynamically weights the data by selecting weighting coefficients according to task type, role, and risk level. For conflict control groups that can be automatically adjudicated, the system outputs the main conclusion and supporting evidence, and retains unaccepted items for auditing.

[0085] S4.2 Human-in-the-loop conflict review and closed-loop correction in high-uncertainty scenarios For conflict groups involving security-sensitive items or those where the automatic determination of confidence is insufficient, the system generates a conflict summary and comparison view, which is submitted to the on-site manager for review. The review results are written back to the system to update weights, thresholds, and source priorities, forming a closed-loop optimization.

[0086] S5 Integrated Knowledge Application System for Intelligent Retrieval and Decision Support S5.1 User Query Input and Context Modeling Search and rescue personnel initiate queries (such as "Wind speed is increasing and there is a short-term heavy rainfall warning. Should the search and rescue continue? How should the flight path and division of labor be adjusted?"). The system extracts the aircraft type, range, battery level, payload, region, time window, and meteorological / geographical constraints, and integrates real-time telemetry summaries, link quality, covered areas, and alarm summaries to form a structured query description.

[0087] S5.2 Response Generation and Result Presentation Based on Fusion and Verification Results The system generates executable suggestions within a constrained, trusted knowledge context: search strategy and priority areas, drone division of labor, key safety restrictions (return-to-home battery threshold, maximum permissible wind speed / rainfall conditions, validity period and scope of temporary restrictions), risk warnings and alternative solutions, along with clues about the source and credibility (source category, update time / validity period, citation location, comprehensive score, whether manually reviewed), presented in a structured form for rapid execution and traceability.

[0088] Another embodiment of the present invention proposes a multi-source information fusion retrieval system that combines information credibility profiling, comprising: The credibility profile building and multi-source knowledge classification module is used to build credibility profiles according to predefined credibility indicators and complete multi-source knowledge classification. The conflict detection module is used to obtain the set of candidate knowledge points corresponding to the query, and divide the set of candidate knowledge points into a group of reference knowledge points for conflict detection. The module for obtaining the fusion knowledge point set is used to perform complementary enhancement on the non-conflicting control knowledge point group according to the multi-source knowledge classification based on the conflict detection results, forming a fusion knowledge point set; and to perform dynamic weighted fusion of credibility and human-in-the-loop conflict review and correction on the control knowledge point group containing conflicts according to the multi-source knowledge classification, and obtain the fusion knowledge point set after completing the conflict resolution. The search results feedback module is used to provide feedback on the search results based on the integrated knowledge point set.

[0089] It should be noted that the information interaction and execution process between the above-mentioned module units are based on the same concept as the method embodiment. For details on their specific functions and technical effects, please refer to the method embodiment section. They will not be repeated here.

[0090] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0091] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A multi-source information fusion retrieval method combining information credibility profiling, characterized in that, include: Build a credibility profile based on predefined credibility indicators and complete multi-source knowledge classification; Obtain the set of candidate knowledge points corresponding to the query, and divide the set of candidate knowledge points into groups of corresponding knowledge points for conflict detection. Based on the conflict detection results, the non-conflicting control knowledge point groups are complemented and enhanced according to multi-source knowledge classification to form a fused knowledge point set; The set of fused knowledge points is obtained by dynamically weighting and fusing the knowledge points in the comparison group that contain conflicts according to the multi-source knowledge classification, and by reviewing and correcting the human-in-the-loop conflicts. The fusion of knowledge points serves as the basis for the query to return the corresponding search results.

2. The multi-source information fusion retrieval method combining information credibility profiling according to claim 1, characterized in that, The predefined credibility metrics include: Authority score W 1. Defined as the basic reliability level of information in terms of professional credibility and standardization, the assignment method is to be offline calibrated and normalized according to the knowledge source category and authority level; Timeliness weight W 2. Defined as the degree of matching between information and the current task time window and the activity of updates. The assignment method is to parse the release time, revision time or version number of the knowledge source, calculate it according to the time decay function and normalize it after combining the update frequency correction. Logical consistency score W 3, defined as the degree of consistency and conflict intensity when cross-referenced with other sources, is assigned by calculating the consistency score and normalizing it based on cross-source comparison results and inference chain verification statistical support and conflict signals; Traceability score W 4. Defined as the degree of location and auditability of information basis and formation process, the value is assigned by calculating and normalizing the completeness of source reference information and process metadata.

3. The multi-source information fusion retrieval method combining information credibility profiling according to claim 2, characterized in that, The steps of performing credibility-based dynamic weighted fusion of conflicting knowledge point groups according to multi-source knowledge classification, and human-in-the-loop conflict review and correction, to obtain a fused knowledge point set after conflict resolution, include: Based on the established multi-source knowledge classification and metadata, a multi-index credibility profile is constructed for each knowledge point from the corresponding dimensions of credibility indicators. A dynamic weighted fusion strategy is designed to score and compare different candidate conclusions, and a quantifiable and traceable adjudication result is given when conflicts occur. The following weighted credibility function is used to fuse and assign values ​​to each knowledge point: In the formula, the parameter α , β , γ , δ The weights are generated through weight selection and adaptive fine-tuning. The process of weight selection and adaptive fine-tuning is as follows: the user query is processed to identify the task type and extract scene elements, and a set of basic weights is given based on the pre-configured "task-weight mapping table". α 0, β 0, γ 0, δ 0), fine-tuning the basic weights based on the conflict level and metadata completeness of the current candidate knowledge point group, and further adjusting the weights. α , β , γ , δ Normalization is performed to ensure weight comparability and guarantee α+β+γ+δ =1; After conflict resolution, for situations where the two or more parties in the conflict control group have similar overall credibility and it is difficult to form a stable ranking, or where the conflict involves key security boundaries, core parameter thresholds, or significantly affects operational steps, and even if there are differences in credibility, it is not possible to make a decision directly based on rules alone, a human-in-the-loop conflict review and correction mechanism is introduced. The scoring results and auditable evidence are output to domain experts or key users to form a closed-loop adjudication process of "data-model-human". When manual review is triggered, the conflict comparison knowledge point group is presented to the experts, including conflict summary, comparison view and source and credibility score of each candidate item. The experts can choose to adopt a conclusion, give a conditional merging conclusion, request supplementary evidence or re-search, or return to regenerate. The ruling result is fixed as the conclusion and basis annotation of the current conflict group.

4. The multi-source information fusion retrieval method combining information credibility profiling according to claim 2, characterized in that: For authoritative scores W 1. Set basic scores based on the knowledge source category, assigning higher basic weights to textbooks and authoritative publications, with a higher basic weight of 0.8~1.0; and assigning medium to high basic weights to standards and regulations, with a medium to high basic weight of 0.7~0.

9. Internet search results and content generated by large models are assigned a low base weight, which is 0.3 to 0.

6. Timeliness weight W The calculation of 2 is based on the latest update timestamp and update frequency of the knowledge point. Specifically, it involves parsing the publication time, revision time or version number of the knowledge source and calculating the time difference Δt from the current time. The time freshness score between [0,1] is mapped to Δt using a preset decay function: content updated within the last year is assigned a weight close to 1, gradually decreasing within 1 to 3 years, and documents older than 3 years or with no update records are reduced to a preset lower limit; for content with multiple version iterations or frequent maintenance records, a certain bonus can be given to the basic freshness score, resulting in W2; the calculation expression is as follows: Δt The time difference between the knowledge point and the current time; f min : Lower limit of basic score for timeliness; n ver The number of times a knowledge point was updated during the statistical period; η max The maximum number of version updates for all knowledge points within the statistical period; η ver Version iteration bonus coefficient; Version iteration normalization: Therefore, timeliness weight W 2. Calculate using the following formula: Logical consistency score W 3. Calculated through cross-source comparison results; for each knowledge point, the number of supporting matches and the number of conflicting matches are calculated; where the number of supporting matches is the number of times it is consistent with or cited by knowledge points from other sources, and the number of conflicting matches is the number of times it is explicitly contradictory to or marked as suspicious by knowledge points from other sources; consistency and contradiction between two knowledge points are judged through a large language model, and normalized by subtracting a set proportion of conflict penalty coefficient from "number of supporting matches / total number of comparisons", the calculation expression is as follows: N sup Number of supporting matches for knowledge points; N conf Number of conflict matching for knowledge points; N tot Total number of comparisons, satisfying N tot =N sup +N conf ; λ Conflict penalty coefficient, satisfying 0 ≤ λ ≤1; Then, logical consistency is determined. W 3. Calculate using the following formula: Traceability score W 4. Calculations are performed based on the completeness and traceability of the knowledge point's metadata; a set of key metadata fields are assigned to each knowledge record during the calculation. W At time 4, scores are awarded based on the completeness and clarity of key metadata fields. The more complete the key metadata fields and the clearer the path, the higher the score; if only a vague source can be traced, then... W 4. Reduce to the preset lower limit, the calculation expression is as follows: N filled The number of metadata fields that have been filled in for the knowledge points; N total The total number of metadata fields expected to be filled in; b human : Whether the indicator variable has been manually labeled. If manually labeled, it is 1; otherwise, it is 0. η human Additional points awarded manually; Then traceability score W 4. Calculate using the following formula: 。 5. The multi-source information fusion retrieval method combining information credibility profiling according to claim 2, characterized in that, Based on predefined credibility metrics, multi-source knowledge is classified in the following manner: Category A: Authoritative books and textbooks; Category B: Standards, Procedures, and Technical Documents; Category C: Online search engine API results; Category D: Content generated from large models; Among them, categories A and B constitute a reliable offline knowledge base, category C provides online supplementary information that meets timeliness requirements, and category D provides structured generative reasoning solutions.

6. The multi-source information fusion retrieval method combining information credibility profiling according to claim 5, characterized in that, For Category A and Category B knowledge, hierarchical parsing is performed according to "Document ID-Chapter-Paragraph-Sentence", page numbers and paragraph positions are retained, and formulas and FF charts are extracted in the form of placeholders and explanatory texts to form a unified structured knowledge point with fields of "knowledge point number, text content, type label, document ID, chapter and page number". All knowledge points constitute the offline knowledge base. After completing the construction of the offline knowledge base, each standard knowledge point is encoded into a vector and a vector index is established. When a user initiates a query or the system generates a sub-question, the query is first preprocessed and then encoded into a query vector. An approximate nearest neighbor search is performed on the vector index to return several candidate knowledge points of type A and type B that are closest to the query semantic distance. For Category C knowledge, relevant online documents are retrieved in real time when a user submits a query or when sub-questions are generated internally by calling the search engine API. Semantic vector matching is performed on the query and candidate webpage titles and summaries, and filtering is performed in combination with site whitelists, source domains, and publication time rules to retain knowledge points that are relevant to the query and have reliable sources, and metadata is recorded. For Category D knowledge, when a user initiates a query, the original query text is input into a pre-deployed large language model to generate candidate answers or explanations. During the generation process, the generation behavior is constrained, and the model is required to explicitly provide preconditions or scope of application. The generated results, along with the call time, model version, and main parameters, are recorded as knowledge points with independent source identification.

7. The multi-source information fusion retrieval method combining information credibility profiling according to claim 5, characterized in that, The steps of complementarizing and enhancing non-conflicting contrasting knowledge point groups according to multi-source knowledge classification to form a fused knowledge point set include: For each B-type knowledge point, perform structured analysis, retrieve the most semantically relevant principle and background evidence fragments from the A-type knowledge points, input the "original knowledge point + evidence fragment + output template constraint" into the large language model, generate structured enhanced knowledge points within the evidence scope, and explicitly associate each enhanced explanation with the corresponding evidence fragment knowledge point. Link the original text of the B-type knowledge point with the above structured explanation field to form a traceable enhanced B-type knowledge point B' record. For Category C knowledge, the search engine API is used to obtain relevant webpage titles, summaries, and text fragments, and paragraphs containing key information are extracted from whitelisted sites or authoritative domains. A large language model is invoked to extract structured knowledge points from each online fragment according to a pre-defined unified field template, and the terminology and unit units are standardized. The structured results are then compared with relevant knowledge points in A+B' using vector recall and rule verification. If the core conclusions and key constraints are consistent, the corresponding online entries are added as the latest supplements or supporting evidence to the corresponding topic. If an online entry provides time-sensitive information missing in A+B', only the missing fields are completed, and the original citation is retained. If there is a significant contradiction with Category A+B' knowledge points in terms of key parameters, applicable conditions, or conclusion polarity, it is marked as a conflict candidate and removed or downgraded. The retained online knowledge points are semantically clustered to remove duplicates, forming a unified, timestamped, and traceable C' set. For knowledge of type D, the large language model outputs candidate answers in structured entries and breaks down the candidate answers into atomic statements. For each statement, the most relevant evidence fragments are recalled in A+B'+C'. Combining rule verification and semantic support discrimination of the large language model, the statements are labeled as supportive, contradictory, or uncertain. Contradictory items are directly deleted or downgraded, and uncertain items trigger secondary generation. The set of retained statements constitutes D', and each statement is accompanied by supporting evidence citations and verification conclusions for tracing and review.

8. The multi-source information fusion retrieval method combining information credibility profiling according to claim 7, characterized in that, The process of using a fusion knowledge point set as the basis for querying the corresponding search results includes the following steps: Users input queries via natural language, which can take the form of single-sentence questions and answers, complex questions with multiple constraints, or handling requests that include business status descriptions. The input is standardized to generate a structured query description. At the same time, the context is obtained from the business side and attached to the current session as structured fields, forming a unified driving object for the query description and the business context. Based on a unified driving object, the system calls the offline knowledge base retrieval pathways of categories A and B to obtain baseline knowledge points, the latest dynamic information fragments obtained from the online knowledge pathway of category C, and the candidate explanations or suggestions generated by the large model pathway for category D knowledge. The system then performs semantic alignment and topic merging of the multi-source candidate knowledge points, and performs conflict detection driven by rules and the large model. It outputs topic-related knowledge point groups, as well as consistency or conflict markers and conflict location information. The consistency groups are then enhanced through complementary fusion to form a structured enhanced knowledge set. The conflict group is resolved by dynamic weighting of credibility, and the triggering human-in-the-loop review and closed-loop correction are performed to form a unified and credible knowledge package for this query. The query description, business context, and unified trusted knowledge package are combined to construct a constrained prompt input, which drives the large language model to generate task-oriented answers or auxiliary decision-making suggestions, ensuring that the generated results are consistent with the upstream fusion and adjudication conclusions and match the actual business context. In the results presentation stage, in addition to outputting natural language answers, traceable and auditable information support is also provided simultaneously, enabling users to understand the basis and applicable boundaries of the answers while obtaining them.

9. The multi-source information fusion retrieval method combining information credibility profiling according to claim 1, characterized in that, The steps of obtaining the candidate knowledge point set corresponding to the query and dividing the candidate knowledge point set into a matching knowledge point group for conflict detection include: For each candidate knowledge point, extract the metadata representation of "subject object + parameter + applicable conditions" from the text, and encode the metadata text into a semantic vector; perform Top-K nearest neighbor matching and clustering in the vector space, and group knowledge points that are semantically similar and discuss the same object, the same parameter or the same condition into the same comparison knowledge point group. Entries below the similarity threshold are considered not to belong to the same topic comparison range and are not included in the conflict determination of the comparison knowledge point group. For each knowledge point group, pairwise comparisons are performed on the knowledge points, and a computable hard rule test is conducted. Then, for samples that cannot be covered by the hard rule test or are ambiguous, a large model is invoked for semantic discrimination, specifically including: Unify units and dimensions, align parameters and conditions with the same name; check whether intervals intersect or contain each other for numerical claims, and check whether they are mutually exclusive for enumerated or state-type claims; check whether polarity words and mandatory words are opposite in normative expressions. Two knowledge points to be compared within the same knowledge point group are used as evidence pairs. Fixed prompt words are used to ask the large model to output three categories: consistent, conflicting, or uncertain, and at the same time return the conflict trigger fragment. When the large model judges it as conflicting and the confidence level is higher than the threshold, the control pair is marked as conflicting. If the output is uncertain, it is retained as uncertain and an explanation is provided. The system will output the results of the knowledge point group division, the consistency label within the group, the conflict type, and the location of the conflict point.

10. A multi-source information fusion retrieval system that combines information credibility profiling, characterized in that, include: The credibility profile building and multi-source knowledge classification module is used to build credibility profiles according to predefined credibility indicators and complete multi-source knowledge classification. The conflict detection module is used to obtain the set of candidate knowledge points corresponding to the query, and divide the set of candidate knowledge points into a group of reference knowledge points for conflict detection. The integrated knowledge point set acquisition module is used to complement and enhance the non-conflicting control knowledge point groups according to multi-source knowledge classification based on the conflict detection results, forming an integrated knowledge point set; The set of fused knowledge points is obtained by dynamically weighting and fusing the knowledge points in the comparison group that contain conflicts according to the multi-source knowledge classification, and by reviewing and correcting the human-in-the-loop conflicts. The search results feedback module is used to provide feedback on the search results based on the integrated knowledge point set.

Images