A standard intelligent retrieval and compliance analysis system and method for the field of nuclear fusion

Abstract

The application discloses a standard intelligent retrieval and compliance analysis system and method for the field of nuclear fusion, which comprises a standard knowledge graph construction and management module, a field semantic analysis engine, a multi-strategy hybrid retrieval module, a compliance analysis and evaluation module and a user interaction and feedback optimization module. The field semantic analysis engine and the multi-strategy hybrid retrieval module are connected through a structured query object for data transmission. The standard knowledge graph construction and management module defines five entity types and five relationship types and executes a cascading update mechanism. The field semantic analysis engine adopts a double-channel parallel analysis architecture and realizes semantic weight factor calculation through synonym and hyponym extension. The multi-strategy hybrid retrieval module executes a three-stage retrieval pipeline. The compliance analysis and evaluation module performs multi-dimensional compliance scoring, multi-standard conflict detection and compliance link tracing. The application can realize collaborative processing of knowledge graph structure retrieval, semantic text retrieval and numerical parameter matching, improve retrieval efficiency, enhance semantic understanding and improve compliance guarantee capability, and realize precise adaptive optimization of the system.

Description

Technical Field

[0001] This invention relates to the field of standardized information processing and artificial intelligence, specifically to a standard intelligent retrieval and compliance analysis system and its working method based on knowledge graph and semantic retrieval technologies for controlled nuclear fusion engineering. Background Technology

[0002] Nuclear fusion engineering is a complex engineering field that involves multiple disciplines and technologies. Taking China's magnetic confinement fusion research as an example, the large-scale fusion engineering projects currently underway in China include the Comprehensive Research Facility for Key Fusion Reactor Systems (CRAFT), the Compact Fusion Experimental Device (BEST), and the planned China Fusion Engineering Demonstration Reactor (CFEDR). The design, manufacturing, testing, and operation of these projects involve the application and compliance of numerous technical standards.

[0003] In terms of standard sources, the standard system involved in nuclear fusion engineering exhibits a significant multi-source heterogeneous characteristic: it includes national mandatory standards (GB) and national recommended standards (GB / T), as well as energy industry standards (NB), nuclear industry standards (EJ), nuclear fusion-specific standards, group standards, local standards, and enterprise standards from various units; at the same time, it also needs to refer to international standards and specifications such as the SDC-IC structural design guidelines issued by the ITER organization, the RCC-MRx design and construction specifications for French nuclear facilities, and the ASME standards of the American Society of Mechanical Engineers. The internal standard list of the Institute of Plasma Physics, Hefei Institutes of Physical Science, Chinese Academy of Sciences alone exceeds 300 items, covering more than ten engineering subsystems such as superconducting magnets, vacuum chambers, first walls, divertors, remote maintenance, power supplies and magnet power supplies, auxiliary heating (ECRH, LHCD, ICRH, NBI), diagnostics, integrated control, cryogenics, water cooling, and assembly.

[0004] These standards exhibit complex semantic relationships. Regarding citations, the welding procedure qualification standard NB / T 47014-2011 references the tensile testing standard GB / T 228.1-2021 and the impact testing standard GB / T 229-2020 in its normative references, while the non-destructive testing standard NB / T 47013 series for pressure equipment is referenced by the pressure vessel manufacturing standard GB / T 150.4-2011. In terms of substitution, standards are updated as technology advances and management needs dictate; when a new standard replaces an older one, all upstream standards referencing the older version also need to be updated accordingly. Regarding cross-application, the same material (e.g., 316LN stainless steel) may need to meet both Chinese national standards and RCC-MRx or ASME specifications when used in the manufacture of fusion devices. Currently, engineers primarily rely on personal experience and manual review when searching for standards and assessing compliance, resulting in low search efficiency, high risk of omissions, and difficulty in identifying cross-disciplinary connections.

[0005] Regarding existing standard retrieval tools, the most commonly used public platforms by engineers fall into two main categories. One is the National Standards Full-Text Public Access System, which provides online reading of the full text of national standards and linear search functions based on standard numbers and title keywords, but lacks semantic understanding capabilities for specific engineering fields such as nuclear fusion. The other is industry standard query platforms, represented by the Gongbiao.com website, which supports browsing by industry category and keyword search, but also lacks intelligent semantic matching capabilities. The common limitations of these platforms are: they only provide independent retrieval of single standards, failing to demonstrate the citation network and cross-application relationships between standards; they lack the function of matching precision ranges based on engineering parameters; and they do not distinguish the different levels of constraint between mandatory and recommended standards in engineering compliance.

[0006] At the system architecture level, existing standard retrieval systems generally adopt a flat architecture of "single retrieval engine + results list". This architecture treats retrieval as an atomic operation—receiving keyword input, processing it through a single matching algorithm, and directly outputting a results list, lacking the ability to perform layered processing and multi-dimensional collaborative optimization of the retrieval process. Specifically, this architecture has three structural defects: First, it lacks a multi-level retrieval collaboration mechanism, failing to effectively integrate the structured relationship retrieval of knowledge graphs with text semantic retrieval, resulting in retrieval results that either rely too heavily on text matching and ignore the structural connections between standards, or are based solely on category browsing and lack semantic understanding depth; Second, it lacks a compliance assessment function for retrieval results, as the system only returns matching results and cannot automatically assess the applicability of standards from multiple dimensions; Third, it lacks a feedback-driven adaptive optimization architecture, as the system's retrieval model parameters remain unchanged once set, and user feedback cannot be fed back to the various functional modules of the system for differentiated adjustments and continuous optimization.

[0007] At the algorithm level, existing standard retrieval systems generally adopt traditional retrieval methods based on keyword matching, which have two fundamental limitations: First, they cannot handle semantic equivalence issues. The field of nuclear fusion uses a large number of mixed Chinese and English terms, and the same concept has multiple equivalent expressions (such as "first wall" and "plasma-facing material" being semantically equivalent). Traditional methods treat these expressions as completely different words. Second, they cannot handle the relationship between hierarchical concepts. For example, when a user enters "auxiliary heating", it cannot automatically match standard text that only contains "ECRH". Summary of the Invention

[0008] The technical solution of this invention is used to solve how to achieve hierarchical fusion processing of knowledge graph structure retrieval, semantic text retrieval and numerical parameter matching, and overcome the limitations of the flat architecture of the single retrieval engine in the existing system.

[0009] The present invention solves the above-mentioned technical problems through the following technical means:

[0010] This invention provides a standard intelligent retrieval and compliance analysis system for the field of nuclear fusion, comprising: The standard knowledge graph construction and management module is used to store standard entities and their attributes, as well as semantic relationships between standards; Domain semantic parsing engine, used to parse user-input natural language queries into structured query objects; A multi-strategy hybrid retrieval module is used to receive the structured query object and execute the retrieval process to generate a candidate standard set; The compliance analysis and assessment module is used to perform compliance scoring, conflict detection, and link tracing on the candidate standard set. The user interaction and feedback optimization module is used to receive user input, display evaluation results, and collect user feedback. The structured query object includes semantic vectors, numerical constraint sets, and entity label sets. The domain semantic parsing engine and the multi-strategy hybrid retrieval module exchange data through the structured query object.

[0011] Furthermore, the standard knowledge graph construction and management module includes multiple entity types and multiple relationship types. The multiple entity types include standard entities, subsystem entities, material entities, process entities, and life cycle stage entities. The multiple relationship types include reference relationships, substitution relationships, cross-application relationships, supplementary relationships, and applicable relationships.

[0012] Furthermore, the standard knowledge graph construction and management module is also used to execute a cascading update mechanism, which includes the following steps: finding all upstream standard nodes in the graph that reference the changed standard through referencing relationships; adding a pending review status marker to each standard in the upstream standard node set and recording the triggering reason; recursively checking whether the upstream standard is referenced by a higher-level standard, and if so, continuing to propagate the pending review status marker upwards until it reaches the root node with no in-degree or a checked node; generating a cascading impact report, listing all affected standard links and their current status. The time complexity of the cascading update mechanism is O(V+E), where V is the number of standard nodes and E is the number of referencing edges.

[0013] Furthermore, the domain semantic parsing engine adopts a dual-channel parallel parsing architecture, including a semantic parsing channel and a numerical parameter extraction channel. The semantic parsing channel includes word segmentation, synonym and hyponym expansion, entity label recognition, and semantic weight factor calculation. The semantic weight factor is set as follows: 1.0 for exact match, 0.85 for synonym match, 0.6 for hyponym match, and 0.7 for hyponym match. The synonym and hyponym expansion uses a hash table index, with a query time complexity of O(1). The numerical parameter extraction channel uses regular expressions to extract temperature range, pressure parameters, and thickness tolerance parameters, and normalizes the extracted numerical parameters into an interval representation.

[0014] Furthermore, the multi-strategy hybrid retrieval module executes a three-level retrieval pipeline. The first level of the three-level retrieval pipeline is graph structure retrieval, which searches for a set of standard nodes associated with the entity tag set from the standard knowledge graph construction and management module. The second level of the three-level retrieval pipeline is semantic weighted retrieval, which uses the TFMEAN-IDF algorithm to calculate the text similarity score. The formula for calculating the text similarity score is: Text Similarity Score = Term Frequency × Inverse Document Frequency × Semantic Weight Factor. The third level of the three-level retrieval pipeline is numerical parameter filtering, which calculates the precision matching score.

[0015] Furthermore, the compliance analysis and assessment module performs multi-dimensional compliance scoring, multi-standard conflict detection, and compliance link tracing. The multi-dimensional compliance score is calculated as follows: Comprehensive Applicability Score = α × Text Similarity Score + β × Graph Structure Matching Score + γ × Precision Matching Score + δ × Mandatory Level Correction Factor, where α is 0.40, β is 0.25, γ is 0.20, and δ is 0.15. The multi-standard conflict detection checks for differences in the precision parameter ranges of standard pairs with cross-applicability relationships, categorizing them into alert, warning, and severe levels based on the parameter deviation ratio. A parameter deviation ratio within 10% is considered alert, 10% to 30% is warning, and over 30% is severe. The compliance link tracing records the semantic expansion of the query terms, standard text matching, graph path, and detailed scores for each dimension.

[0016] Furthermore, the user interaction and feedback optimization module implements a three-way feedback loop mechanism, including a first feedback loop path, a second feedback loop path, and a third feedback loop path. In the first feedback loop path, the user interaction and feedback optimization module transmits missing standard supplements and relationship corrections to the standard knowledge graph construction and management module. In the second feedback loop path, the user interaction and feedback optimization module transmits newly discovered synonyms and hyponym pairs to the domain semantic parsing engine. In the third feedback loop path, the user interaction and feedback optimization module transmits retrieval accuracy statistics and threshold adjustment suggestions to the multi-strategy hybrid retrieval module.

[0017] This invention also provides a standard intelligent retrieval and compliance analysis method based on the above system, comprising the following steps: S1. Receive natural language queries input by the user through the user interaction and feedback optimization module; S2. The natural language query is parsed in parallel through a domain semantic parsing engine to generate a structured query object; S3. Receive the structured query object through the multi-strategy hybrid retrieval module and execute a three-level retrieval pipeline to generate a candidate standard set; S4. The candidate standard set is evaluated through a multi-dimensional compliance score, multi-standard conflict detection, and compliance link tracing using the compliance analysis and evaluation module. S5. Display the scoring result set, conflict report, and reasoning chain through the user interaction and feedback optimization module; S6. User feedback is collected through the user interaction and feedback optimization module, and used as a three-way feedback loop mechanism for the standard knowledge graph construction and management module, the domain semantic parsing engine, and the multi-strategy hybrid retrieval module.

[0018] Furthermore, S2 includes parallel processing of the semantic parsing channel and the numerical parameter extraction channel. S3 includes first-level graph structure retrieval, second-level semantic weighted retrieval, and third-level numerical parameter filtering, wherein the upper limit of the candidate set size for the first-level graph structure retrieval is 50, the upper limit of the candidate set size for the second-level semantic weighted retrieval is 20, and the minimum threshold for the accuracy matching score for the third-level numerical parameter filtering is 0.1. S4 includes multi-dimensional compliance scoring, multi-standard conflict detection, and compliance link tracing. S6 includes transmitting graph expansion information to the standard knowledge graph construction and management module, transmitting dictionary expansion information to the domain semantic parsing engine, and transmitting strategy optimization information to the multi-strategy hybrid retrieval module; parameter adjustment is triggered when the weighted accuracy of a certain type of query is lower than 0.7.

[0019] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned standard intelligent retrieval and compliance analysis method. The computer-readable storage medium is a non-volatile storage medium.

[0020] Compared with the prior art, the present invention has the following advantages: (1) System architecture level: The five-module hierarchical architecture proposed in this invention overcomes the limitations of the existing system's flat architecture of "single search engine + result list" by defining clear data interfaces between modules and a unified data model of structured query objects, and realizes the organic synergy of knowledge graph structure retrieval, semantic text retrieval and numerical parameter matching.

[0021] (2) Knowledge graph level: This invention is the first to construct a standard knowledge graph for the field of nuclear fusion, defining five entity types and five relationship types. In particular, the cascading update mechanism can automatically propagate the "pending review" mark when the standard version changes, which significantly reduces the risk of referencing outdated standards in the project.

[0022] (3) Semantic retrieval level: The dual-channel parsing architecture of the present invention ensures that the extraction of semantic information and numerical parameters do not interfere with each other, and the unified data model of structured query objects provides a standardized input interface for subsequent multi-level retrieval; the TFMEAN-IDF semantic weighted retrieval algorithm effectively solves the semantic mismatch problem of professional terms in the field of nuclear fusion by introducing a domain terminology dictionary and semantic weight factors.

[0023] (4) Search architecture level: The three-level search pipeline of the present invention, "graph coarse screening → semantic fine ranking → parameter filtering", improves search efficiency while ensuring search accuracy by refining the candidate set step by step. The size of the candidate set is gradually reduced from 300+ in the whole database to less than 20 in the end, which greatly reduces the amount of semantic matching computation.

[0024] (5) Compliance analysis level: The multi-standard conflict detection function of the present invention can automatically discover situations where candidate standards make different requirements for the same technical object, and classify and mark them according to the severity; the compliance link tracing function provides a complete reasoning path for each search result, making the scoring results interpretable.

[0025] (6) Feedback optimization level: The multi-way feedback return mechanism of the present invention enables user feedback to act differently on three different levels: knowledge graph layer (graph expansion), semantic parsing layer (dictionary expansion) and retrieval strategy layer (threshold adjustment), realizing the precise adaptive optimization of the system, rather than the coarse-grained optimization of the prior art that can only adjust a single weight parameter. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall system architecture and data flow of Embodiment 1 of the present invention.

[0027] Figure 2 This is a schematic diagram of the standard knowledge graph structure and cascading update mechanism in Embodiment 1 of the present invention.

[0028] Figure 3 This is a schematic diagram of the three-level retrieval pipeline and compliance scoring process in Embodiment 1 of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] The meanings of the labels in the attached diagram are as follows: Figure 1 In this module, M1 represents the standard knowledge graph construction and management module, M2 represents the domain semantic parsing engine, M3 represents the multi-strategy hybrid retrieval module, M4 represents the compliance analysis and evaluation module, and M5 represents the user interaction and feedback optimization module. Figure 2 In the diagram, 201 represents a standard entity, 202 represents a subsystem entity, 203a represents a material entity, and 203b represents a process entity. Figure 3 It demonstrates the complete data processing flow from user input to result output.

[0031] Example 1 like Figure 1As shown, this embodiment provides a standard intelligent retrieval and compliance analysis system for the nuclear fusion field, including a standard knowledge graph construction and management module M1, a domain semantic parsing engine M2, a multi-strategy hybrid retrieval module M3, a compliance analysis and evaluation module M4, and a user interaction and feedback optimization module M5.

[0032] I. System Overall Architecture The five modules interact through well-defined data interfaces, forming a complete processing chain of "query input → semantic parsing → multi-level retrieval → compliance assessment → result display → feedback feedback". The core data flow path of the system is as follows: the user inputs a natural language query through M5 → M2 performs dual-channel parsing of the query to generate a structured query object Q (containing a semantic vector V_sem and a numerical constraint set C_num) → M3 receives the structured query object Q and generates a candidate standard set R through a three-level retrieval pipeline (graph coarse screening, semantic fine ranking, and parameter filtering) → M4 performs multi-dimensional compliance scoring, conflict detection, and link tracing on the candidate set R, outputting a result set S with scores and interpretable reasoning paths → M5 displays the result set S to the user and collects feedback. After being classified by M5, the feedback data is applied to M1 (graph expansion), M2 (lexicon expansion), and M3 (threshold adjustment) through three independent feedback paths, respectively, to achieve multi-level adaptive optimization of the system.

[0033] II. Standard Knowledge Graph Construction and Management Module (M1) like Figure 2 As shown, M1 is responsible for the construction, storage, and dynamic maintenance of the standard knowledge graph in the field of nuclear fusion. The knowledge graph defines five entity types and five relation types.

[0034] The five entity types include: (1) Standard Entity 201. Each standard entity contains attribute fields such as standard number, Chinese name, English name, publication date, implementation date, repeal date, mandatory level, scope of application text, precision parameters, and keyword list. The mandatory level is divided into four levels: strictly mandatory (corresponding to GB mandatory national standards, score 1.0), industry mandatory (corresponding to mandatory clauses in NB / EJ industry standards, score 0.8), recommended (corresponding to GB / T recommended national standards, score 0.5), and reference (corresponding to international standards such as ITER SDC-IC and RCC-MRx, score 0.3). Precision parameters are stored in the form of numerical ranges, for example, the temperature range is represented as [300, 500], and the pressure parameter is represented as [10, 15].

[0035] (2) Subsystem entity 202 represents the engineering subsystem of the nuclear fusion device. In this embodiment, fifteen subsystems are defined, including superconducting magnet, vacuum chamber, first wall, divertor, remote maintenance, power supply and magnet power supply, ECRH, LHCD, ICRH, NBI, diagnostics, integrated control, cryogenic, water cooling, and assembly.

[0036] (3) Material entities 203a, such as 316LN, 316L, CuCrZr, tungsten alloy, NbTi, Nb3Sn, low activation steel CLAM, etc.

[0037] (4) Process entity 203b, such as welding, TIG welding, electron beam welding, non-destructive testing, ultrasonic testing, radiographic testing, etc.

[0038] (5) The life cycle stage entities include eight stages: design, procurement, manufacturing, inspection, installation, commissioning, operation and decommissioning.

[0039] The five types of relationships include: (1) Citation relationship (cites) indicates that the standard cites another standard in the normative reference clause, such as NB / T47014-2011 citing GB / T 228.1-2021.

[0040] (2) Replacement relationship, which means that the new version of the standard replaces the old version of the standard, for example, GB / T 228.1-2021 replaces GB / T 228.1-2010.

[0041] (3) Cross-applies indicate that the same object is subject to multiple standard systems at the same time. For example, NB / T 47014-2011 and RCC-MRx 2025 are both applicable to the TIG welding process of 316LN material.

[0042] (4) Supplements indicate that one standard provides supplementary provisions for content not covered by another standard.

[0043] (5) Applies to: This indicates the mapping relationship between the standard and the subsystem, material, process, and life cycle stage. For example, NB / T 47014-2011 applies to the vacuum chamber subsystem, 316LN material, TIG welding process, and manufacturing stage.

[0044] The core innovation of M1 lies in its cascading update mechanism. When the state of a standard entity changes (e.g., it is replaced by a new standard or it is repealed), M1 automatically executes the following cascading processing flow: Step C1 involves finding the set of all upstream standard nodes in the graph that reference the standard through the `cites` relationship. Specifically, for each changing standard node, all reference edges pointing to that node are traversed in reverse order to obtain the set of source nodes.

[0045] Step C2: Automatically add a "Pending Review" status marker to each standard in the upstream standard node set and record the triggering reason. The pending review status marker includes information such as the marking time, the changed standard number, and the change type. The triggering reason is recorded as "The referenced GB / T 228.1-2010 has been replaced by GB / T 228.1-2021".

[0046] Step C3 recursively checks whether the upstream standard is referenced by an even higher-level standard. If so, it continues to propagate the "pending review" flag upwards until it reaches the root node with zero in-degree or a node that has already been checked. The recursion terminates when the current node has zero in-degree or when the current node already has a pending review flag.

[0047] Step C4 generates a cascading impact report, listing all affected standard links and their current status. The cascading impact report includes the total number of affected standards, a list of standards at each level, a propagation path diagram, and recommended review priorities.

[0048] This mechanism employs a reverse breadth-first traversal algorithm with a time complexity of O(V+E), where V is the number of standard nodes and E is the number of reference edges. In this embodiment with 300 standards, the cascading update mechanism can complete the update in milliseconds. This mechanism ensures that when the underlying standards change, all standards in the entire reference chain can be promptly marked and reviewed, avoiding the risk of referencing outdated standards in the project.

[0049] III. Domain Semantic Parsing Engine (M2) like Figure 3 As shown in the upper part, M2 is responsible for parsing the user's input natural language technical requirements into a structured query object Q. The core innovation of M2 lies in its dual-channel parallel parsing architecture, where two channels process the same input text simultaneously, extracting semantic information and numerical parameter information respectively, and finally merging them into a unified structured query object.

[0050] Channel 1: Semantic Parsing Channel The channel performs the following processing steps: Step S1 involves mixed Chinese and English word segmentation, using the jieba word segmentation engine and loading a specialized terminology dictionary for the nuclear fusion field. Specialized terms in the dictionary (such as "first wall," "divertor," and "electron cyclotron resonance heating") are assigned extremely high frequency weights to ensure they are not incorrectly segmented.

[0051] Step S2: Synonym and Hypothesis Expansion. The domain terminology dictionary pre-constructs approximately 30 sets of synonym mappings (e.g., "first wall" = "FW"; "plasma facing material" = "PFM"); and approximately 10 sets of hypothesis relationships (e.g., the hypothesis of the hyperthesis "auxiliary heating" is "ECRH", "LHCD", "ICRH", "NBI"). The system queries the synonym index and hypothesis index for each word in the word segmentation results, incorporating semantically equivalent and semantically related words into the expanded query set. The index uses a hash table structure, with a query time complexity of O(1).

[0052] Step S3, entity tag recognition, automatically identifies subsystem names, material grades, and process types within the expanded vocabulary set, generating an entity tag set E_meta to drive the first-level graph retrieval of M3. Entity tag recognition employs a rule-based matching method, performing precise matching through a predefined entity dictionary.

[0053] Step S4: The semantic weight factor w_s is calculated and assigned different weights according to the matching type: exact match 1.0, synonym match 0.85, hypernym match 0.6, and hyponym match 0.7, generating a weighted semantic vector V_sem. The asymmetric design of the semantic weight factor is based on the following technical principles: when a user inputs a specific term, the standard that includes its hypernym usually still applies, therefore the weight of hyponym matching (0.7) is higher than the weight of hypernym matching (0.6); while when a user inputs a hypernym, the specific standard for a particular hyponym may not be what the user needs.

[0054] Channel 2: Numerical Parameter Extraction Channel This channel employs regular expressions and pattern matching techniques to extract engineering numerical parameters in parallel from the same input text. Supported parameter types include temperature range (e.g., "300-500°C", "Operating temperature not exceeding 800°C"), pressure parameters (e.g., "25MPa", "Design pressure 10-15MPa"), and thickness / tolerance parameters (e.g., "12mm", "Wall thickness ±0.05mm"). Each extracted numerical parameter is normalized into a uniform interval representation [low, high] and labeled with its parameter type, forming a numerical constraint set C_num.

[0055] The outputs of the two channels converge at the merging layer of M2, generating a structured query object Q={V_sem, C_num, E_meta}. This structured query object Q is passed to M3 as a unified data interface, driving the subsequent three-level retrieval pipeline. The advantages of the dual-channel architecture are: semantic parsing and numerical extraction do not interfere with each other, avoiding the interference of numerical parameters on the word segmentation results in a single processing pipeline, and the processing delay of semantic expansion on numerical extraction.

[0056] IV. Multi-Strategy Hybrid Retrieval Module (M3) As shown Figure 3 in the middle part, M3 receives the structured query object Q output by M2, and gradually refines the candidate set through a three-level retrieval pipeline, and finally outputs a candidate standard set R that has been screened in multiple dimensions. The design idea of the three-level pipeline is "structure first, then semantics, and then parameters" - using the structural relationship of the knowledge graph to quickly narrow the search space, then using semantic similarity for fine sorting, and finally using numerical parameters for hard constraint filtering.

[0057] Level 1: Graph Structure Retrieval (Coarse Screening) At this level, the entity label set E_meta is extracted from the structured query object Q, and the set of standard nodes R1_direct that is directly associated with these entity labels through the applies_to relationship is found in the knowledge graph of M1. Subsequently, one-hop expansion is performed along the citation relationship (cites) and supplementary relationship (supplements), and the standards that have a direct citation or supplementary relationship with the standards in R1_direct are also included in the candidate set, obtaining the extended candidate set R1. If E_meta is empty (that is, no clear subsystem, material, or process entity is recognized in the user input), this level is skipped, and R1 is defaulted to the entire standard library. The candidate set size control strategy at this level is: when the size of R1 exceeds the preset upper limit N1 (default N1 = 50), it is sorted and truncated according to the number of matching dimensions of the applies_to relationship (the one that matches the subsystem + material + process simultaneously takes precedence over the one that only matches a single dimension).

[0058] Level 2: Semantic Weighted Retrieval (Fine Ranking) At this level, TFMEAN-IDF semantic weighted retrieval is performed on the candidate set R1 output by the first level. The weighted similarity between the text content (standard name, scope of application description, keywords, etc.) of each standard in R1 and the semantic vector V_sem in the structured query object Q is calculated. The calculation formula is score_text = TF × IDF × w_s, where TF is the term frequency, IDF is the inverse document frequency, and w_s is the semantic weight factor calculated by M2. The core difference between the TFMEAN-IDF algorithm and the traditional TF-IDF is that: the traditional TF-IDF regards "first wall" and "first wall" as literally completely different words, and the similarity is zero; while TFMEAN-IDF recognizes the semantic equivalence relationship between the two through the synonym index and assigns a matching weight of 0.85. Similarly, the matching between hypernyms and hyponyms is also assigned a non-zero weight. Sort in descending order of score_text and take the top N2 candidate standards (default N2 = 20) to obtain the fine-ranked candidate set R2.

[0059] Level 3: Numerical Parameter Filtering This level utilizes the set of numerical constraints C_num in the structured query object Q to perform parameter matching filtering on the candidate criteria in R2. For each numerical constraint [q_low, q_high] in C_num, the corresponding type precision parameters [s_low, s_high] stored in the candidate criterion entity are checked, and a precision matching score score_precision is calculated: a score of 1.0 is given when the applicable range of the criterion completely covers the range of the query parameters (i.e., s_low ≤ q_low and q_high ≤ s_high); a score of 0.0 is given when there is partial overlap, based on the overlap ratio, with a minimum of 0.3; and a score of 0 is given when there is no overlap. Criteria without precision parameter records have a default score of 0.5. If C_num is empty (the user input does not contain numerical parameters), this level is skipped, and R3 = R2. Otherwise, criteria with a score_precision lower than the minimum threshold (default 0.1) are removed, resulting in the final candidate set R3.

[0060] In the three-level pipeline, the upper limit of the candidate set size N1 and N2 for each level and the minimum threshold for the third level are all configurable parameters. The system provides default values, and users can adjust them in the advanced settings of M5. They can also be adaptively adjusted by the feedback loop path three of M5 based on the cumulative retrieval accuracy statistics.

[0061] V. Compliance Analysis and Assessment Module (M4) like Figure 3 As shown in the lower part, M4 receives the candidate standard set R3 and its intermediate scores at each level output by M3, and performs three core functions: multi-dimensional compliance scoring, multi-standard conflict detection, and compliance link tracing.

[0062] (1) Multi-dimensional compliance scoring M4 calculates the comprehensive applicability score S = α·S_text + β·S_graph + γ·S_precision + δ·S_mandatory for each candidate standard in R3. Here, S_text is the second-level TFMEAN-IDF text similarity score; S_graph is the knowledge graph structure matching score, calculated based on the number of matching dimensions of the apply_to relationship in the first level (matching subsystems + materials + processes yields the highest score); S_precision is the third-level precision range matching score; and S_mandatory is the mandatory level correction factor, assigned according to the standard's four-level mandatory classification. The default weight parameters are α=0.40, β=0.25, γ=0.20, and δ=0.15, which users can adjust in real time in the M5 interface. Specifically, when a recommended standard GB / T is referenced in contract clauses or regulatory documents, M4 supports automatically upgrading its mandatory level from "Recommended" to "Contractually Mandatory" (score increases from 0.5 to 0.9).

[0063] (2) Multi-standard conflict detection While calculating the score, M4 automatically detects whether there are standard pairs in R3 that impose different requirements on the same technical object. The detection logic is as follows: iterates through the standard pairs in R3. When two standards are associated through the cross_applies relationship (i.e., applicable to the same material or process) and their respective accuracy parameter ranges are inconsistent (e.g., the national standard requires a thickness tolerance of ±0.05mm while RCC-MRx requires ±0.03mm), it is marked as a potential conflict, and a conflict report is generated. The conflict report includes: the number and name of the conflicting standard pair, the specific parameters of the conflict, the required values ​​of each standard, and the severity rating of the conflict (divided into three levels according to the parameter deviation ratio: prompt level, warning level, and severe level; deviation within 10% is prompt level, 10%-30% is warning level, and more than 30% is severe level).

[0064] (3) Compliance Link Traceability M4 generates a complete inference path for each search result, from the user's query to the standard. The inference path records which words in the query underwent semantic expansion, which words in the standard text were matched, which relational paths in the graph were traversed, and detailed scores for each dimension, presented to the user in a visual link format. The purpose of this feature is to make the scoring results interpretable—users not only see a total score, but also understand "why this standard was retrieved" and "how each score of this standard was derived."

[0065] VI. User Interaction and Feedback Optimization Module (M5) M5 provides a web-based user interface responsible for three functions: query input, result display, and feedback collection. For query input, M5 offers a natural language input box, one-click example query entry, and an advanced filtering panel (allowing for combined filtering by subsystem, material, process, lifecycle stage, and mandatory level). For result display, M5 shows the scoring result set output by M4, including detailed scoring for each dimension, applicable scope text, accuracy parameters, a list of related standards, conflict reports, and inference chains, and supports exporting results in Excel, CSV, and JSON formats.

[0066] The core innovation of M5 lies in its multi-path feedback mechanism. For example... Figure 1 As shown, M5 provides feedback buttons for each search result (with three levels of evaluation: relevant, partially relevant, and irrelevant), as well as entry points for supplementing missing criteria and correcting terminology. The collected feedback data is processed by M5's feedback classifier and then transmitted to different target modules through three independent feedback paths: Return Path 1 (M5→M1, Graph Expansion): When a user submits new standard information through the missing standard supplementation entry, M5 pushes the basic attributes of the standard to M1. M1 automatically creates new standard entity nodes and, based on the new standard's scope of application description and keywords, automatically generates candidate relationship recommendations using existing apply_to mapping rules (such as recommending which subsystems and materials the standard is associated with). After user confirmation, these recommendations are written into the graph.

[0067] Backflow Path Two (M5→M2, Dictionary Expansion): When a user rates a standard as "relevant" but the standard is retrieved through a graph relationship rather than direct matching, M5 automatically extracts professional terms that appear in the standard text but are not included in the M2 terminology dictionary. These terms are then recommended as candidate synonyms or new terms to the system administrator for review and addition to the dictionary.

[0068] Path 3 (M5→M3, retrieval strategy optimization): M5 periodically calculates the weighted accuracy of accumulated feedback data (1 point for relevance, 0.5 points for partial relevance, and 0 points for irrelevance). When the accuracy of a certain type of query (classified by subsystem or process) falls below a preset threshold (default 0.7), M5 generates suggestions for adjusting the retrieval parameters for that type of query, including: the adjustment direction for the upper limit of the candidate set size N1 and N2 of the three-level pipeline, the adjustment direction for the minimum threshold of the third level, and fine-tuning suggestions for the four weight parameters α, β, γ, and δ of the compliance score. These adjustment suggestions are then applied to the M3 configuration after administrator confirmation.

[0069] Feedback data is persistently stored in the form of timestamped JSON records. Each record contains a query ID, query content, standard number, evaluation level, feedback type (evaluation / supplementation / correction), and user identifier.

[0070] Workflow Example The user inputs a natural language query via M5: "TIG welding process standard for 316LN material". M5 passes the query text to M2. M2's semantic parsing channel segments the query, obtaining "316LN", "material", "TIG welding", "process", and "standard". The synonym and hyponym expansion step queries the synonym index, expanding "TIG welding" to "tungsten inert gas shielded welding" and "GTAW". The entity label recognition step identifies the material entity "316LN" and the process entity "TIG welding", generating the entity label set E_meta. The semantic weight factor calculation step assigns a perfect match weight of 1.0 to "316LN", a perfect match weight of 1.0 to "TIG welding", a synonym matching weight of 0.85 to "tungsten inert gas shielded welding", and a synonym matching weight of 0.85 to "GTAW". The numerical parameter extraction channel did not identify any numerical parameters, and the numerical constraint set C_num is empty. M2 generates a structured query object Q, which contains a semantic vector V_sem, an empty set of numerical constraints C_num, and an entity label set E_meta, and passes it to M3.

[0071] M3 performs a first-level graph structure retrieval. It extracts the entity tag set E_meta from the structured query object Q, containing "316LN" and "TIG welding". In M1, it searches for standard nodes associated with "316LN" or "TIG welding" through the apply_to relationship, obtaining the directly associated standard set R1_direct, which includes standards such as NB / T 47014-2011 and EJ / T 888-2017. A one-hop expansion is performed along the reference relationships, incorporating standards such as GB / T 228.1-2021 and GB / T 229-2020 referenced by NB / T 47014-2011 into the candidate set, resulting in an expanded candidate set R1, containing 48 standards.

[0072] M3 performs a second-level semantic weighted retrieval. Weighted similarity is calculated between the text content and semantic vector V_sem of each standard in R1. For NB / T 47014-2011, the standard name is "Welding Procedure Qualification for Pressure Equipment," and the keywords include "welding," "TIG," and "316LN." When calculating score_text, "TIG" completely matches "TIG welding" in the query, with a weight of 1.0; "316LN" completely matches "316LN" in the query, with a weight of 1.0; "welding" and "TIG welding" in the query have a hyponymous relationship, with a weight of 0.6. After combining TF and IDF, the score_text of NB / T 47014-2011 is 0.92. The top 20 standards, sorted in descending order of score_text, are used to obtain the refined candidate set R2.

[0073] M3 performs the third level of numerical parameter filtering. Since the numerical constraint set C_num is empty, this level is skipped, and R3 equals R2.

[0074] M4 calculates the overall applicability score S for each candidate standard in R3. For NB / T 47014-2011, S_text is 0.92, S_graph is 1.0 (matching both material and process dimensions), S_precision is 0.5 (no precision parameter record), and S_mandatory is 0.8 (industry-mandatory level). The overall applicability score S = 0.40 × 0.92 + 0.25 × 1.0 + 0.20 × 0.5 + 0.15 × 0.8 = 0.738.

[0075] M4 performs multi-standard conflict detection. It checks if there are standard pairs in R3 associated through the `cross_applies` relationship. It finds that both NB / T 47014-2011 and RCC-MRx 2025 are applicable to TIG welding of 316LN material, but their requirements for the tensile strength of the welded joint differ. NB / T 47014-2011 requires a minimum of 490 MPa, while RCC-MRx 2025 requires a minimum of 520 MPa, a deviation of 6.1%, classifying it as a warning-level conflict. A conflict report is generated.

[0076] M4 performs compliance traceability. An inference path is generated for NB / T 47014-2011: the query term "TIG welding" is expanded with synonyms to "tungsten inert gas welding" (weight 0.85) and "GTAW" (weight 0.85), matching the standard keyword "TIG" (weight 1.0); the query term "316LN" matches "316LN" in the standard's scope of application (weight 1.0); the graph's apply_to relationship links the material entity "316LN" and the process entity "TIG welding" (S_graph=1.0); text similarity score S_text=0.92; precision matching score S_precision=0.5 (no precision parameter); mandatory level correction factor S_mandatory=0.8 (industry mandatory level); comprehensive applicability score S=0.738.

[0077] M5 displays the scoring results set, including the overall score of NB / T 47014-2011 (0.738), detailed scores for each dimension, applicable scope text, accuracy parameters, a list of related standards, a conflict report, and a visualization of the inference chain. Users rate NB / T 47014-2011 as relevant. M5 collects feedback data and transmits it to M3 via a third feedback loop for statistical analysis of retrieval accuracy and generation of threshold adjustment suggestions.

[0078] The system using this embodiment improves retrieval efficiency, gradually reducing the candidate set size from 300 records in the entire database to less than 20 records, and reducing semantic matching computation by 93%. Semantic understanding is enhanced; the TFMEAN-IDF algorithm effectively solves the semantic mismatch problem of professional terms through synonym indexing and semantic weight factors, achieving a query time complexity of O(1). Compliance assurance capabilities are improved; the cascading update mechanism automatically propagates the review-pending marker when standards change, with a processing time in milliseconds, reducing the risk of referencing expired standards; multi-standard conflict detection is graded according to the deviation ratio, automatically identifying potential conflicts. Adaptive optimization capabilities are improved; the three-way feedback feedback mechanism allows user feedback to differentiate and impact the knowledge graph layer, semantic parsing layer, and retrieval strategy layer, achieving multi-level precise optimization.

[0079] Example 2 The difference from Example 1 is that this example focuses on demonstrating the specific working process of the cascading update mechanism.

[0080] Assuming that standard GB / T 228.1-2010 is replaced by the new standard GB / T 228.1-2021, the standard knowledge graph construction and management module M1 detects that the obsolescence date field of GB / T 228.1-2010 has changed from null to "2021-04-01". The cascading update mechanism is automatically triggered.

[0081] Step C1: Locate the set of all upstream standard nodes in the graph that reference GB / T 228.1-2010 through referencing relationships. Traverse all referencing edges pointing to GB / T 228.1-2010 in reverse order to obtain the set of source nodes, which includes standards such as NB / T47014-2011 and GB / T 150.4-2011.

[0082] Step C2: Add a pending review status marker to each standard in the upstream standard node set and record the triggering reason. Add a pending review marker to NB / T 47014-2011, with the marker time "2021-04-02 10:15:30", the changed standard number "GB / T 228.1-2010", the change type "replacement", and the triggering reason "the referenced GB / T 228.1-2010 has been replaced by GB / T 228.1-2021". Add a similar pending review marker to GB / T 150.4-2011.

[0083] Step C3 recursively checks whether upstream standards are referenced by even more upstream standards. It checks if NB / T 47014-2011 is referenced by other standards and finds that GB / T 9711-2017 references NB / T 47014-2011. A pending review flag is added to GB / T 9711-2017, triggered by the reason "The referenced NB / T 47014-2011 references the superseded GB / T 228.1-2010". The recursion terminates as it finds that GB / T 9711-2017 has zero in-degree.

[0084] Step C4: Generate a cascading impact report. The cascading impact report contains a total of 3 affected standards. The list of standards at each level is as follows: Level 1 standards NB / T 47014-2011 and GB / T 150.4-2011; Level 2 standard GB / T 9711-2017. The propagation path diagram is GB / T 228.1-2010 → NB / T 47014-2011 → GB / T 9711-2017. It is recommended to review these standards in ascending order of propagation level.

[0085] The cascading update mechanism executes in milliseconds with a time complexity of O(V+E), where V represents 3 standard nodes and E represents 2 referencing edges. A cascading impact report is provided to engineers to guide them in timely updates to the standards on the referencing chain.

[0086] The technical advantage of this embodiment is that when the standard version changes, the cascading update mechanism automatically identifies and marks all affected standard reference chains, avoiding the risk of omissions in manual inspection, improving processing efficiency to the millisecond level, and ensuring the timeliness and accuracy of standard application.

[0087] Example 3 The difference from Example 1 is that this example focuses on demonstrating the specific working process of the three-level retrieval pipeline in queries containing numerical parameters.

[0088] Users can enter natural language queries through the M5 user interaction and feedback optimization module to find the standard that the tensile strength of the 316LN material TIG welded joint in the vacuum chamber is not less than 500MPa.

[0089] The semantic parsing channel of the domain semantic parsing engine M2 segments the query, obtaining "vacuum chamber", "316LN", "material", "TIG welding", "joint", "tensile strength", "not less than", "500MPa", and "standard". The synonym and hyponym expansion step queries the synonym index, expanding "TIG welding" to "tungsten inert gas shielded welding" and "GTAW". The entity tag recognition step identifies the subsystem entity "vacuum chamber", the material entity "316LN", and the process entity "TIG welding", generating the entity tag set E_meta. The semantic weight factor calculation step assigns a perfect match weight of 1.0 to "vacuum chamber", "316LN", and "TIG welding", and a synonym match weight of 0.85 to "tungsten inert gas shielded welding" and "GTAW".

[0090] The domain semantic parsing engine M2 uses regular expressions to extract numerical parameters. It identifies "not less than 500MPa," normalizes it to the interval [500, ∞], labels it as tensile strength type, and forms the numerical constraint set C_num.

[0091] The domain semantic parsing engine M2 generates a structured query object Q, which includes a semantic vector V_sem, a set of numerical constraints C_num, and a set of entity labels E_meta, and passes it to the multi-strategy hybrid retrieval module M3.

[0092] The multi-strategy hybrid retrieval module M3 performs the first-level graph structure retrieval. It extracts the entity tag set E_meta from the structured query object Q, containing "vacuum chamber," "316LN," and "TIG welding." In the standard knowledge graph construction and management module M1, it searches for standard nodes associated with "vacuum chamber," "316LN," or "TIG welding" through the apply_to relationship, obtaining the directly associated standard set R1_direct, which includes standards such as NB / T 47014-2011 and GB / T 150.4-2011. A one-hop expansion along the reference relationship yields an expanded candidate set R1, containing 52 standards. Since 52 exceeds the preset upper limit N1 of 50, it is truncated based on the number of matching dimensions in the apply_to relationship. NB / T 47014-2011 matches all three dimensions—vacuum chamber, 316LN, and TIG welding—and has the highest sorting priority. After truncation, R1 is obtained, containing 50 standards.

[0093] The multi-strategy hybrid retrieval module M3 performs the second-level semantic weighted retrieval. Weighted similarity is calculated between the text content and semantic vector V_sem of each standard in R1. For NB / T 47014-2011, the standard name is "Welding Process Qualification for Pressure Equipment," and the keywords include "welding," "TIG," "316LN," and "tensile strength." When calculating score_text, "TIG" completely matches "TIG welding" in the query, with a weight of 1.0; "316LN" completely matches "316LN" in the query, with a weight of 1.0; "tensile strength" completely matches "tensile strength" in the query, with a weight of 1.0; and "welding" and "TIG welding" in the query have a hyponymous relationship, with a weight of 0.6. After combining TF and IDF, the score_text of NB / T 47014-2011 is 0.95. The top 20 standards, sorted in descending order of score_text, are used to obtain the refined candidate set R2.

[0094] The multi-strategy hybrid retrieval module M3 performs third-level numerical parameter filtering. For each candidate standard in R2, the precision parameter of the tensile strength type is checked. For NB / T 47014-2011, the precision parameter is [490, ∞], indicating a tensile strength of not less than 490 MPa. The query parameter is [500, ∞]. The scope of the standard does not fully encompass the query parameter range because 490 < 500. The overlap interval is calculated as [500, ∞], the query interval is [500, ∞], the overlap ratio is 100%, and the score_precision is 1.0. For RCC-MRx 2025, the precision parameter is [520, ∞], the scope of the standard fully encompasses the query parameter range, and the score_precision is 1.0. Standards with a score_precision below the minimum threshold of 0.1 are removed, resulting in the final candidate set R3, containing 18 standards.

[0095] The compliance analysis and assessment module M4 calculates the overall applicability score S for each candidate standard in R3. For NB / T47014-2011, S_text is 0.95, S_graph is 1.0 (matching subsystem, material, and process dimensions simultaneously), S_precision is 1.0, and S_mandatory is 0.8 (industry-mandatory level). The overall applicability score S = 0.40 × 0.95 + 0.25 × 1.0 + 0.20 × 1.0 + 0.15 × 0.8 = 0.950. For RCC-MRx 2025, S_text is 0.88, S_graph is 1.0, S_precision is 1.0, and S_mandatory is 0.3 (reference level). The overall applicability score S = 0.40 × 0.88 + 0.25 × 1.0 + 0.20 × 1.0 + 0.15 × 0.3 = 0.847.

[0096] The compliance analysis and assessment module M4 performs multi-standard conflict detection. It checks for standard pairs linked through the `cross_applies` relationship in R3. It finds that NB / T 47014-2011 and RCC-MRx 2025 both apply to the vacuum chamber subsystem, 316LN material, and TIG welding process, but their requirements for the tensile strength of the welded joints differ. NB / T 47014-2011 requires a minimum of 490 MPa, while RCC-MRx 2025 requires a minimum of 520 MPa, a deviation of 6.1%, classifying it as a warning-level conflict. A conflict report is generated.

[0097] The User Interaction and Feedback Optimization Module M5 displays a set of scoring results, sorted in descending order of overall applicability score. NB / T47014-2011 ranks first with a score of 0.950, and RCC-MRx 2025 ranks second with a score of 0.847. The conflict report indicates a warning-level conflict between the two.

[0098] The technical advantage of this embodiment lies in the fact that the three-level retrieval pipeline, through progressive refinement via graph coarse screening, semantic ranking, and parameter filtering, reduces the candidate set size from 52 to 50, then to 20, and finally to 18 in queries containing numerical parameters, improving retrieval efficiency by 65%. Numerical parameter filtering ensures that the retrieval results meet engineering accuracy requirements, and the accuracy matching score accurately reflects the degree of overlap between the standard's applicable scope and the query requirements. Multi-standard conflict detection automatically identifies parameter differences between different standard systems for the same technical object, classifying them according to the deviation ratio, making it easy for users to quickly identify the severity of conflicts.

[0099] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention.

Claims

1. A standard intelligent retrieval and compliance analysis system for the field of nuclear fusion, characterized in that, include: The standard knowledge graph construction and management module is used to store standard entities and their attributes, as well as semantic relationships between standards; Domain semantic parsing engine, used to parse user-input natural language queries into structured query objects; A multi-strategy hybrid retrieval module is used to receive the structured query object and execute the retrieval process to generate a candidate standard set; The compliance analysis and assessment module is used to perform compliance scoring, conflict detection, and link tracing on the candidate standard set. The user interaction and feedback optimization module is used to receive user input, display evaluation results, and collect user feedback. The structured query object includes semantic vectors, numerical constraint sets, and entity label sets. The domain semantic parsing engine and the multi-strategy hybrid retrieval module exchange data through the structured query object.

2. The standard intelligent retrieval and compliance analysis system according to claim 1, characterized in that, The standard knowledge graph construction and management module includes multiple entity types and multiple relationship types.

3. The standard intelligent retrieval and compliance analysis system according to claim 2, characterized in that, The various entity types include standard entities, subsystem entities, material entities, process entities, and lifecycle stage entities.

4. The standard intelligent retrieval and compliance analysis system according to claim 2, characterized in that, The various relationship types include reference relationships, substitution relationships, cross-application relationships, supplementary relationships, and applicable relationships.

5. The standard intelligent retrieval and compliance analysis system according to claim 1, characterized in that, The standard knowledge graph construction and management module is also used to execute a cascading update mechanism.

6. The standard intelligent retrieval and compliance analysis system according to claim 5, characterized in that, The cascading update mechanism includes the following steps: Find the set of all upstream standard nodes in the graph that reference the changed standard through referencing relationships; Add a pending review status marker to each standard in the upstream standard node set and record the triggering reason; Recursively check whether the upstream standard is referenced by a higher-level standard. If so, continue to propagate the pending status flag upwards until it reaches the root node with no in-degree or the already checked node. Generate a cascading impact report, listing all affected standard links and their current status.

7. The standard intelligent retrieval and compliance analysis system according to claim 5, characterized in that, The time complexity of the cascading update mechanism is O(V+E), where V is the number of standard nodes and E is the number of reference edges.

8. The standard intelligent retrieval and compliance analysis system according to claim 1, characterized in that, The domain semantic parsing engine adopts a dual-channel parallel parsing architecture.

9. The standard intelligent retrieval and compliance analysis system according to claim 8, characterized in that, The dual-channel parallel parsing architecture includes a semantic parsing channel and a numerical parameter extraction channel. The semantic parsing channel includes word segmentation, synonym and hyponym expansion, entity label recognition, and semantic weight factor calculation.

10. The standard intelligent retrieval and compliance analysis system according to claim 9, characterized in that, The semantic weight factor is set to the following values: 1.0 for exact match, 0.85 for synonym match, 0.6 for hypernym match, and 0.7 for hyponym match.

11. The standard intelligent retrieval and compliance analysis system according to claim 9, characterized in that, The synonym and hyponym expansion uses a hash table index, with a query time complexity of O(1).

12. The standard intelligent retrieval and compliance analysis system according to claim 8, characterized in that, The numerical parameter extraction channel uses regular expressions to extract temperature range, pressure parameters, and thickness tolerance parameters, and normalizes the extracted numerical parameters into an interval representation.

13. The standard intelligent retrieval and compliance analysis system according to claim 1, characterized in that, The multi-strategy hybrid retrieval module executes a three-level retrieval pipeline.

14. The standard intelligent retrieval and compliance analysis system according to claim 13, characterized in that, The first stage of the three-level retrieval pipeline is graph structure retrieval, which searches for the set of standard nodes associated with the entity tag set from the standard knowledge graph construction and management module.

15. The standard intelligent retrieval and compliance analysis system according to claim 13, characterized in that, The second level of the three-level retrieval pipeline is semantic weighted retrieval, which uses the TFMEAN-IDF algorithm to calculate text similarity scores.

16. The standard intelligent retrieval and compliance analysis system according to claim 13, characterized in that, The third stage of the three-level retrieval pipeline is numerical parameter filtering, which calculates the accuracy matching score.

17. The standard intelligent retrieval and compliance analysis system according to claim 1, characterized in that, The compliance analysis and assessment module performs multi-dimensional compliance scoring, multi-standard conflict detection, and compliance link tracing.

18. The standard intelligent retrieval and compliance analysis system according to claim 17, characterized in that, The calculation formula for the multi-dimensional compliance score is: Comprehensive Applicability Score = α × Text Similarity Score + β × Graph Structure Matching Score + γ × Precision Matching Score + δ × Mandatory Level Correction Factor.

19. The standard intelligent retrieval and compliance analysis system according to claim 17, characterized in that, The multi-standard conflict detection examines the difference in accuracy parameter ranges between standard pairs that have an overlapping applicability relationship.

20. The standard intelligent retrieval and compliance analysis system according to claim 17, characterized in that, The compliance link tracing records the semantic expansion of the query terms, standard text matching, graph path, and detailed scores for each dimension.

21. The standard intelligent retrieval and compliance analysis system according to claim 1, characterized in that, The user interaction and feedback optimization module implements a three-way feedback feedback mechanism.

22. The standard intelligent retrieval and compliance analysis system according to claim 21, characterized in that, The three-way feedback loop mechanism includes a first feedback loop path, in which the user interaction and feedback optimization module transmits missing standard supplements and relationship corrections to the standard knowledge graph construction and management module.

23. The standard intelligent retrieval and compliance analysis system according to claim 21, characterized in that, The three-way feedback feedback mechanism includes a second feedback feedback path, in which the user interaction and feedback optimization module transmits newly discovered synonyms and hyponym pairs to the domain semantic parsing engine.

24. The standard intelligent retrieval and compliance analysis system according to claim 21, characterized in that, The three-way feedback feedback mechanism includes a third feedback feedback path, in which the user interaction and feedback optimization module transmits retrieval accuracy statistics and threshold adjustment suggestions to the multi-strategy hybrid retrieval module.

25. A standard intelligent retrieval and compliance analysis method based on the system described in claim 1, characterized in that, Includes the following steps: S1. Receive natural language queries input by the user through the user interaction and feedback optimization module; S2. The natural language query is parsed in parallel through a domain semantic parsing engine to generate a structured query object; S3. Receive the structured query object through the multi-strategy hybrid retrieval module and execute a three-level retrieval pipeline to generate a candidate standard set; S4. The candidate standard set is subjected to multi-dimensional compliance scoring, multi-standard conflict detection, and compliance link tracing through the compliance analysis and evaluation module; S5. Display the scoring result set, conflict report, and reasoning chain through the user interaction and feedback optimization module; S6. User feedback is collected through the user interaction and feedback optimization module, and used as a three-way feedback loop mechanism for the standard knowledge graph construction and management module, the domain semantic parsing engine, and the multi-strategy hybrid retrieval module.

26. The standard intelligent retrieval and compliance analysis method according to claim 25, characterized in that, The S2 includes parallel processing of the semantic parsing channel and the numerical parameter extraction channel.

27. The standard intelligent retrieval and compliance analysis method according to claim 25, characterized in that, The S3 includes a first-level graph structure retrieval, a second-level semantic weighted retrieval, and a third-level numerical parameter filtering.

28. The standard intelligent retrieval and compliance analysis method according to claim 25, characterized in that, The S4 includes multi-dimensional compliance scoring, multi-standard conflict detection, and compliance link tracing.

29. The standard intelligent retrieval and compliance analysis method according to claim 25, characterized in that, S6 includes transmitting graph expansion information to the standard knowledge graph construction and management module, transmitting dictionary expansion information to the domain semantic parsing engine, and transmitting strategy optimization information to the multi-strategy hybrid retrieval module.

Images