An intelligent question-answering system and method for iterative reasoning type pressure pipeline design

Abstract

The application discloses an intelligent question and answer system and method for pressure pipeline design based on iterative reasoning, which comprises a knowledge base construction module, a retrieval and iteration module, an answer synthesis and tracing module, and an interaction and knowledge evolution module. The knowledge base construction module parses the pressure pipeline design specification document into structured clause records, and constructs a knowledge graph and a vector index containing specification number, version identification, chapter path, clause number and cross-reference relationship. The retrieval and iteration module performs intent recognition and constraint extraction on user questions, and performs multiple rounds of iterative retrieval to address evidence coverage deficiencies, specification conflicts or version timeliness deficiencies. The answer synthesis and tracing module generates answers under strong constraints of specification original text, calculates fact consistency scores, and outputs results with additional multi-dimensional tracing information after meeting the standards. The interaction and knowledge evolution module receives user feedback, dynamically adjusts knowledge confidence, and writes high-value external knowledge back to the knowledge base after passing the audit.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and natural language processing technology, and relates to an iterative reasoning-based intelligent question-answering system and method for pressure pipeline design. Background Technology

[0002] Pressure pipeline design involves numerous national and industry standards and specifications, which are complex and frequently updated. Engineers must frequently consult relevant clauses during the design process to ensure that the design meets safety and compliance requirements. Traditional retrieval methods rely on keyword matching or manual document review, making it difficult to accurately understand the multiple constraints and implicit intentions in users' natural language questions. Existing intelligent question-answering systems mostly employ end-to-end generation models, lacking strong constraints on the original standard text, which can easily lead to factual biases or citation errors. While some systems incorporate knowledge graphs, they fail to effectively handle cross-references between standards, version timeliness, and conflicts in numerical set constraints. Furthermore, when questions involve multiple operating conditions or multiple clause dependencies, a single search often fails to provide complete coverage, resulting in incomplete answers. Existing technologies also lack effective mechanisms for utilizing user feedback, failing to dynamically integrate verified external knowledge into the knowledge base. Therefore, there is an urgent need for an intelligent question-answering method that can combine structured standard knowledge, support multi-round iterative reasoning, rigorously verify factual consistency, and possess continuous evolution capabilities to meet the high-reliability technical decision support needs of the pressure pipeline design field. Summary of the Invention

[0003] To address the problems existing in the background technology, this invention proposes an iterative reasoning-based intelligent question-and-answer system and method for pressure pipeline design.

[0004] The first aspect of this application provides an iterative reasoning-based intelligent question-answering system for pressure pipeline design, comprising:

[0005] The knowledge base construction module is used to parse pressure pipeline design specification documents into structured clause records and build a knowledge graph and vector index containing specification number, version identifier, chapter path, clause number and cross-reference relationship;

[0006] The retrieval and iteration module is used to identify the intent and extract the constraints of the natural language questions raised by the user, and to perform a preliminary retrieval based on the knowledge graph and vector index. If the retrieval results have insufficient information coverage, conflicts in numerical or set constraints of normative clauses, or lack of version timeliness, the reflection and planning submodule is activated to generate supplementary query strategies and execute multiple rounds of iterative retrieval.

[0007] The answer synthesis and tracing module is used to integrate internal and external evidence obtained through retrieval, generate a preliminary answer under the strong constraints of facts and normative original texts in the knowledge graph, and calculate the factual consistency score between the answer and the source information fragment; when the score is not lower than a preset threshold, the answer with additional multi-dimensional tracing information is output.

[0008] The interaction and knowledge evolution module is used to receive user feedback on answers, dynamically adjust the confidence of relationships in the knowledge graph and the credibility of external information sources based on the feedback, and write high-value external information back to the knowledge base construction module in a structured form after the review mechanism is passed.

[0009] Optionally, the knowledge base construction module performs the following processing flow:

[0010] The input pressure piping design specification document is parsed in a structured manner, and a structured record containing clause numbers, titles, appendix identifiers, table references and footnotes is generated according to the clause granularity.

[0011] Identify cross-references within and across specifications, construct reference relationship edges, and associate source clauses with target clauses;

[0012] The terminology is standardized by combining domain dictionaries, and entities and their relationships are extracted using sequence labeling and rule-based methods to form triples and associate them with the corresponding clause records.

[0013] Semantically vectorize the clauses and key sentences, create a vector index, and associate it with the clause records using a unique identifier;

[0014] Based on the ontology definition, unified concepts, attributes, and relational constraints are defined, and the extracted results are populated into the knowledge graph.

[0015] Receive the review conclusions from the interaction and knowledge evolution modules, and write back the externally verified new knowledge to the clause records, knowledge graph and vector index with the same structure fields after the review is approved.

[0016] Optionally, the retrieval and iteration module includes an input parsing submodule, a multi-strategy evidence retrieval and ranking submodule, a problem decomposition and synthesis submodule, and a reflection and planning submodule;

[0017] The input parsing submodule is used to complete the intent parsing of user questions, extract key entities and constraints, and resolve referential issues in combination with the dialogue context.

[0018] The multi-strategy evidence recall and ranking submodule is used to perform dense semantic retrieval and sparse keyword retrieval, and to normalize and weight the scores of the two to generate a ranking of candidate evidence.

[0019] The problem decomposition and synthesis submodule is used to decompose a user problem into several sub-problems when it is detected that the user problem contains multiple working conditions, multiple constraints, or cross-term dependencies, and then perform constraint alignment and logical synthesis after each sub-problem is retrieved.

[0020] The reflection and planning submodule is used to quantify information gaps and generate supplementary query strategies when there is insufficient evidence coverage, regulatory conflicts, or version timeliness issues. These strategies include hierarchical tracing paths, horizontal correlation paths, version timeliness verification paths, internal in-depth mining paths, or external authoritative source query paths.

[0021] Optionally, the reflection and planning submodule is activated based on one of the following triggering conditions:

[0022] Calculate the information dispersion index of the retrieved normative clauses in terms of their level of validity or chapter distribution. When the index is lower than a preset threshold, it is determined that the evidence coverage is insufficient.

[0023] For two or more highly relevant clauses, a conflict probability score is calculated on the numerical range or set constraint of the same parameter. When the score exceeds a preset threshold, it is determined that there is a specification conflict.

[0024] When a user's question contains time-sensitive keywords and the internal knowledge base lacks valid version metadata, it is determined that the version timeliness is missing.

[0025] Optionally, the conflict probability score is obtained by weighted fusion of numerical constraint conflict degree, set constraint conflict degree, and logical contradiction degree; for numerical constraints Conflict level is defined as:

[0026] ;

[0027] in, Represents the degree of conflict of numerical constraints; This represents the interval formed by the lower and upper limits of the numerical constraint of a certain parameter in the first clause of the specification; This represents the interval formed by the lower and upper limits of the numerical constraints for the same parameter in the second clause of the specification; This represents the length of the intersection of the two intervals. It represents the length of the union of two intervals.

[0028] Set constraints Define the degree of conflict:

[0029] ;

[0030] in, Represents the degree of conflict of constraints in a set; This represents the set of allowed or restricted values ​​for a parameter in the first clause of the specification. This represents the set of allowed or restricted values ​​for the same parameter in Article 2 of the regulations; The number of elements in the intersection of two sets; It represents the number of elements in the union of two sets.

[0031] The overall probability of conflict is:

[0032] ;

[0033] in, Represents the overall probability of conflict; Represents a non-linear activation function; The preset weight coefficients representing the degree of conflict of numerical constraints; The preset weight coefficients representing the degree of conflict of set constraints; Preset weighting coefficients representing the degree of logical contradiction; This represents the numerical constraint conflict degree obtained from the aforementioned calculation; This represents the set constraint conflict degree obtained from the aforementioned calculation; It represents the degree of logical contradiction.

[0034] Optionally, the answer synthesis and tracing module performs the following operations when generating answers:

[0035] Evidence from core specifications and external sources is hierarchically merged and aligned.

[0036] For factual and definitional questions, answers are generated by extracting the original text of the clauses; for questions that require cross-clause synthesis, language is organized without deviating from the original meaning of the evidence.

[0037] The generated answer is aligned with each source information fragment item by item, and the factual consistency score is calculated.

[0038] The factual consistency score is obtained by weighted fusion of numerical constraint consistency, set constraint consistency, logical implication consistency and contradiction penalty term, and the weights are dynamically set according to the authority level, timeliness and relevance of the information fragment;

[0039] When the score is not lower than the preset threshold, the answer is output along with multi-dimensional traceability information, including the specification number and clause number, the name of the external source, the issuing organization, the publication date, and the reason for adoption.

[0040] Optionally, the formula for calculating the factual consistency score is:

[0041] ;

[0042] in, Representative generates answer With evidence set The score for factual consistency between them; This represents the initial answer generated by the system; This represents the set of source evidence used to generate the answer; Representative evidence set The total number of source information fragments contained therein; The summation index refers to the k-th source information segment. The weight represents the k-th source information segment; This represents the result of the consistency verification function, which is used to determine the answer. With the k-th source information fragment Are they consistent? This represents the k-th specific source information fragment in the evidence set.

[0043] Optionally, the interaction and knowledge evolution module performs the following operations:

[0044] Provide a user feedback interface to receive user ratings, corrections, or comments on answers;

[0045] The feedback is mapped as a quantized signal and used to update the confidence of relations in the knowledge graph and the credibility of external sources;

[0046] When a user gives a positive rating to an answer that contains external information that exceeds a preset threshold, or when the manual review interface marks it as an important supplement, the new knowledge structure integration process is triggered.

[0047] The structured integration process includes: extracting parameter names, parameter values, and applicable conditions from external information; mapping them to attribute fields defined in the ontology through slot filling; submitting for review in manual assistance mode; and creating or updating knowledge graph nodes and relationships after approval, and attaching source and version metadata.

[0048] A second aspect of this application provides an iterative reasoning-based intelligent question-answering method for pressure pipeline design, comprising:

[0049] The knowledge base construction module parses the pressure pipeline design specification document into structured clause records and builds a knowledge graph and vector index containing specification number, version identifier, chapter path, clause number and cross-reference relationship;

[0050] The retrieval and iteration module performs intent recognition and constraint extraction on the natural language questions raised by the user, and conducts a preliminary retrieval based on the knowledge graph and vector index. If the retrieval results have insufficient information coverage, conflicts in numerical or set constraints of normative clauses, or lack of version timeliness, the reflection and planning submodule is activated to generate supplementary query strategies and execute multiple rounds of iterative retrieval.

[0051] The internal and external evidence retrieved by the answer synthesis and tracing module is integrated to generate a preliminary answer under the strong constraints of facts and normative original texts in the knowledge graph, and the factual consistency score between the answer and the source information fragment is calculated; when the score is not lower than a preset threshold, the answer with additional multi-dimensional tracing information is output.

[0052] The interaction and knowledge evolution module receives user feedback on the answers and dynamically adjusts the confidence of relationships in the knowledge graph and the credibility of external information sources based on the feedback. After the review mechanism is approved, high-value external information is written back to the knowledge base construction module in a structured form.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] This invention achieves precise transformation from standard text to computable knowledge by constructing a structured knowledge base for pressure pipeline design specifications. The system employs a multi-strategy fusion retrieval and iterative reasoning mechanism to effectively address multiple constraints, operating conditions, and cross-clause dependencies in user questions, significantly improving the completeness of evidence coverage. By quantifying information dispersion indicators and conflict probability scores, the system can proactively identify insufficient evidence coverage, conflicts in numerical or set constraints of standard clauses, and missing version timeliness, generating targeted supplementary query strategies to avoid erroneous or biased conclusions. The answer generation process is strictly constrained by the facts in the knowledge graph and the original standard text, and introduces a fact consistency score verification mechanism to ensure a high degree of consistency between the output content and the source information. Each answer is accompanied by multi-dimensional traceability information, supporting complete tracing of standard number, clause number, external source name, issuing agency, publication date, and reasons for adoption, enhancing the auditability of engineering decisions. The interaction and knowledge evolution module transforms user feedback into structured knowledge update signals, enabling dynamic writing back of high-value external information under the protection of a manual review mechanism, giving the system continuous evolution capabilities. Overall, this invention improves the accuracy, logical rigor, compliance reliability, and timeliness of intelligent question-and-answer systems for pressure pipeline design, effectively supporting engineering technicians in making safe and efficient technical decisions. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of an iterative reasoning-based intelligent question-and-answer system for pressure pipeline design, according to one embodiment of the present invention.

[0056] Figure 2 This is a flowchart of an iterative reasoning-based intelligent question-and-answer method for pressure pipeline design, according to one embodiment of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] In one embodiment, such as Figure 1 As shown, an iterative reasoning-based intelligent question-and-answer system for pressure pipeline design is provided. This iterative reasoning-based intelligent question-and-answer system for pressure pipeline design corresponds one-to-one with the iterative reasoning-based intelligent question-and-answer system and method for pressure pipeline design described in the following embodiments. The iterative reasoning-based intelligent question-and-answer system for pressure pipeline design includes: a knowledge base construction module, a retrieval and iteration module, an answer synthesis and tracing module, and an interaction and knowledge evolution module. Detailed descriptions of each functional module are as follows:

[0059] The knowledge base construction module parses pressure pipeline design specification documents into structured clause records and constructs a knowledge graph and vector index containing specification numbers, version identifiers, chapter paths, clause numbers, and cross-reference relationships. The retrieval and iteration module identifies the intent and extracts constraints from user-submitted natural language questions, and performs a preliminary search based on the knowledge graph and vector index. If the search results have insufficient information coverage, conflicts in specification clause values ​​or set constraints, or missing version timeliness, the reflection and planning submodule is activated to generate supplementary query strategies and execute multiple rounds of iterative retrieval. The answer synthesis and tracing module integrates internal and external evidence obtained from the search, generates a preliminary answer under strong constraints of facts in the knowledge graph and the original specification text, and calculates the factual consistency score between the answer and the source information fragment. When the score is not lower than a preset threshold, an answer with additional multi-dimensional tracing information is output. The interaction and knowledge evolution module is used to receive user feedback on answers, dynamically adjust the confidence of relationships in the knowledge graph and the credibility of external information sources based on the feedback, and write high-value external information back to the knowledge base construction module in a structured form after the review mechanism is passed.

[0060] In this invention, during system operation, the knowledge base construction module first accumulates structured knowledge and forms a searchable index, then the retrieval and iteration module organizes and verifies evidence around the user's question, then the answer synthesis and tracing module generates and verifies the answer under the constraint of evidence, and finally the interaction and knowledge evolution module feeds user feedback and expert review results back to the knowledge base and retrieval strategy, thus forming a closed loop of knowledge accumulation, evidence acquisition, answer generation, and feedback evolution.

[0061] The knowledge base construction module transforms pressure pipeline design specifications into a structured, computable, and traceable knowledge system. The module first receives the original specification documents as input, including textual documents such as national or industry-issued pressure pipeline design standards, regulations, and technical guidelines. The module then performs structured parsing on the specifications, with parsing granularity down to the clause level. Each clause is extracted independently, generating a corresponding structured record. This structured record includes the clause number, clause title, chapter / section path, appendix identifier, referenced table numbers, and footnote content. All fields are stored in a standardized format to ensure consistency and retrieval in subsequent processing.

[0062] After completing the clause-level structuring, the knowledge base construction module further identifies cross-references within the specification and references across specifications. For example, if a clause mentions "see Clause 5.2.3" or "in accordance with Chapter 4 of GB / T20801.3," the knowledge base construction module automatically locates the referenced target clause through a combination of regular expression matching and semantic analysis, and establishes a reference relationship edge between the source clause and the target clause. This reference relationship edge includes the reference type, direction, and contextual fragments, forming a directed graph structure that provides topological support for subsequent reasoning.

[0063] Subsequently, the knowledge base construction module invokes a pre-built pressure piping domain dictionary to standardize the technical terms in the clause text. This domain dictionary covers core terms such as material grades, pipe fitting types, design parameters, and operating conditions, ensuring that the same concept is mapped to a unified standard name under different expressions. Based on this, the knowledge base construction module uses a sequence labeling model combined with manually defined extraction rules to identify entities and their attributes from the clause text. Entities include equipment type, material name, pressure rating, temperature range, and safety factor; attributes include numerical values, units, applicable conditions, and limitations. Entities and attributes form subject-verb-object triples, each triple being associated with a unique identifier in its source clause, achieving a strong binding between knowledge and the original text.

[0064] To further support semantic retrieval, the knowledge base construction module performs semantic vectorization on the full text of each clause and its key sentences. This vectorization employs a pre-trained language model based on domain fine-tuning, mapping the text to fixed-dimensional real-valued vectors. All vectors are organized into a vector index, supporting fast retrieval via near-nearest neighbor representation. Each item in the vector index is bidirectionally associated with the aforementioned structured records, triples, and citation relationships through a unique clause identifier, ensuring that retrieval results can be traced back to the original specification position.

[0065] Meanwhile, the knowledge base construction module integrates the extracted knowledge according to a predefined ontology model. This ontology model clearly defines the core concepts in the field of pressure pipeline design, the hierarchical relationships between concepts, the data types and value constraints of attributes, and the cardinality and logical rules of relationships. All extracted triples must pass ontology consistency verification before being written into the knowledge graph. After successful verification, the triples are populated into the corresponding nodes and edges of the knowledge graph, forming a structured semantic network.

[0066] Furthermore, the knowledge base construction module possesses dynamic evolution capabilities. When the interaction and knowledge evolution module returns new knowledge that has been reviewed and confirmed, the knowledge base construction module receives the review conclusion. If the review conclusion is passed, the new knowledge is written into the clause record database, the knowledge graph node relationship table, and the vector index according to the same field format as the aforementioned structured records. The new knowledge originates from external authoritative sources verified by user feedback, and its writing process maintains the same structure and association mechanism as the original normative knowledge, ensuring the overall consistency and maintainability of the knowledge base.

[0067] In one embodiment, the system receives the full-text PDF document of the "Code for Design of Industrial Metal Piping" GB 50316-2000 (2008 edition). The knowledge base construction module first parses it into 287 independent clauses. Each record contains fields such as "Clause Number: 6.1.2", "Title: Selection of Piping Components", "Appendix Identifier: None", "Table Reference: Table 6.1.2-1", and "Footnote: Note 1: Carbon steel shall comply with NACE MR0175 requirements in wet H2S environments". Then, it identifies that clause 7.3.4 references clause 6.1.2 and establishes a directed reference edge. Using the domain dictionary, "CS" is normalized to "carbon steel", and the triple <carbon steel, applicable environmental limitations, wet H2S> is extracted through sequence labeling and associated with clause 7.3.4. The clause text is vectorized and stored in the FAISS index with ID T-734. Ontology verification confirms that "carbon steel" belongs to the "metallic materials" subclass and "wet H2S" belongs to the "corrosive environment" instance, and the relationship conforms to the definition. Ultimately, this knowledge was fully incorporated into the knowledge graph.

[0068] Through the above processing flow, the present invention enables the system to accurately convert standardized text into structured knowledge, supports high-precision semantic retrieval and logical reasoning, avoids omissions or misjudgments caused by traditional keyword matching, and improves the accuracy and interpretability of the question-answering system.

[0069] The retrieval and iteration module is the core component of the system for achieving accurate question answering and deep reasoning. Internally, it consists of an input parsing submodule, a multi-strategy evidence retrieval and ranking submodule, a question decomposition and synthesis submodule, and a reflection and planning submodule working together. Each submodule processes user input sequentially and triggers multiple rounds of iterative retrieval when necessary to gradually approach a complete, consistent, and traceable answer.

[0070] The input parsing submodule first receives questions from users in natural language. This submodule performs semantic understanding on the question text to identify the user's query intent. This intent includes, but is not limited to, parameter queries, compliance assessments, design selection suggestions, or clause interpretations. Simultaneously, this submodule extracts key entities from the question, such as pipe materials, design pressure, operating temperature, and media type, and identifies implicit constraints, such as numerical ranges, set restrictions, or logical relationships. Furthermore, when the user's question is within a multi-turn dialogue context, this submodule performs referential resolution processing, restoring pronouns or omitted components to their corresponding specific entities or conditions, ensuring that subsequent searches are based on complete semantics.

[0071] After obtaining the parsed structured query, the multi-strategy evidence recall and ranking submodule initiates a dual-channel retrieval mechanism. The first channel is dense semantic retrieval, which uses the clause vector index generated by the aforementioned knowledge base construction module to calculate the similarity between the user query vector and all clause vectors, returning the candidate clauses with the most semantically similar results. The second channel is sparse keyword retrieval, which matches keywords in the query with terms in the clause text based on the inverted index, retaining highly accurate matching results. Each channel outputs a candidate list with its original score. Subsequently, this submodule normalizes the two types of scores to eliminate differences in units, and then performs a weighted fusion according to preset weights to generate a uniformly ranked candidate evidence list. This fusion strategy balances semantic generalization ability and keyword precision matching ability, improving the recall quality.

[0072] The problem decomposition and synthesis submodule is activated when it detects a user's problem having a complex structure. This complex structure includes situations involving multiple operating conditions, multiple interdependent design parameters, or dependence on different specification chapters or even different standard documents. For example, a user might ask, "In a carbon steel pipe with a design pressure of 4 MPa, an operating temperature of 350 degrees Celsius, and a medium of wet hydrogen sulfide, which clauses should be followed for wall thickness calculation?" This question contains three constraints and involves multiple dimensions such as material suitability, strength calculation, and corrosion protection. In this case, the submodule automatically decomposes the original problem into several sub-problems, such as "the suitability clauses for carbon steel in a wet hydrogen sulfide environment" and "the clause containing the carbon steel pipe wall thickness calculation formula at 4 MPa and 350 degrees Celsius." Each sub-problem independently enters the search process, obtaining its own evidence set. Subsequently, the submodule performs constraint alignment on the search results of each sub-problem, checks parameter consistency, and synthesizes the results according to logical rules to form a comprehensive evidence set covering all dimensions of the original problem.

[0073] The Reflection and Planning submodule is responsible for monitoring the quality of the current evidence set and initiating iterative optimization when defects are found. This submodule continuously evaluates evidence coverage, standard consistency, and timeliness reliability. The reflection mechanism is triggered when any of the following situations are detected: 1) Insufficient evidence coverage, manifested as search results concentrated in a single chapter while ignoring other related areas; 2) Standard conflicts, i.e. different highly relevant clauses give mutually exclusive numerical ranges or set restrictions on the same parameter; 3) Lack of version timeliness, i.e. the user's question clearly points to a specific year version, but the search results do not indicate valid version information or mix multiple version contents. Once triggered, this submodule quantifies the current information gap and generates targeted supplementary query strategies. The strategies include hierarchical tracing paths, i.e. tracing upwards along the standard reference relationship to the basic definition clauses; horizontal association paths, i.e. finding parallel clauses on related topics in the same chapter; version timeliness verification paths, i.e. limiting the search scope to the standard document of a specified year version; internal in-depth mining paths, i.e. performing secondary fine-grained analysis on ambiguous expressions in the preliminary results; and external authoritative source query paths, i.e. calling certified industry databases or standard publishing platforms to obtain supplementary information. Each strategy is transformed into a new structured query and re-enters the retrieval process until the evidence set meets the requirements for completeness and consistency.

[0074] In one embodiment, a user queries "GB 50316 regarding the usage restrictions of stainless steel pipes in chloride ion environments." The input parsing submodule identifies the intent as "usage restriction query," with key entities being "stainless steel" and "chloride ion environment," and the constraint being "according to GB 50316." The multi-strategy evidence recall and ranking submodule initially retrieves clause 8.2.5 and Appendix C, clause 3. The problem decomposition and synthesis submodule finds that the former specifies material selection, while the latter specifies chloride ion concentration limits, but the correspondence is not clearly defined. The reflection and planning submodule detects insufficient information coverage and generates a lateral association path strategy, supplementing the search with related terms such as "austenitic stainless steel" and "stress corrosion cracking," ultimately recalling clause 8.2.7 and its notes, clearly stating that 304 stainless steel is prohibited when the chloride ion concentration exceeds 50 mg / L. Based on this, the system synthesizes a complete answer and marks the sources of each conclusion.

[0075] Through the coordinated operation of the above four-level sub-modules, the system can handle complex, fuzzy, or multi-constraint engineering problems, avoid the one-sidedness of a single search, and significantly improve the comprehensiveness, logical rigor, and standard compliance of the answers.

[0076] The activation of the reflection and planning submodule relies on a quantitative assessment of the quality of the current search results. This assessment is based on three clearly defined triggering conditions: insufficient evidence coverage, normative conflicts, and lack of version timeliness. Each condition is determined using calculable indicators to ensure the objectivity and reproducibility of the system's decisions.

[0077] When assessing insufficient evidence coverage, the reflection and planning submodule calculates the information dispersion index of the retrieved normative clauses in terms of their legal force level or chapter distribution. The legal force level refers to the legal or administrative force level of the normative document, such as national standards, industry standards, or local regulations; the chapter distribution refers to the position of the clauses within the normative document, such as different technical chapters like general design principles, material selection, and strength calculation. This module first statistically analyzes the legal force level category and chapter path of each search result, calculating the degree of concentration of their distribution. If the vast majority of results are concentrated in a single chapter or the same legal force level, it indicates that the current search has failed to cover the multi-dimensional technical aspects of the issue. In this case, the module calculates the knowledge density entropy as the information dispersion index, using the formula: ,in, Representing knowledge density entropy, as a core indicator for measuring information dispersion, the lower the value, the more concentrated the information distribution, and the lower the dispersion, which may mean that the search results are limited to a single source or chapter, resulting in insufficient evidence coverage. This represents the set of search results, that is, the set of standard clauses recalled by the system based on the user's question; This represents the total number of terms returned by the search, i.e., the number of candidate evidences recalled in this search operation; This represents the index variable in the summation operation, referring to the m-th item in the search results; represents the normalized weight of the m-th clause in the search results; the weight reflects the relevance or importance of the clause to the user's question, and the sum of the weights of all clauses is usually normalized to 1; represents a very small positive number used to prevent logarithmic singularity, which, in mathematical calculations, if When the value is close to 0, directly calculating the logarithm can lead to unstable values ​​or infinity. Therefore, we introduce... This ensures the stability of numerical calculations. When this indicator falls below a preset threshold, the system determines that the evidence coverage is insufficient and triggers a supplementary search.

[0078] When assessing regulatory conflicts, the reflection and planning submodule focuses on whether two or more highly relevant clauses are consistent in their constraints on the same parameter. These parameters include quantifiable or enumerable technical elements such as design pressure, allowable stress, temperature limits, and material composition. The module first identifies numerical ranges or sets of constraints for the same parameter within these clauses. For numerical ranges… and Calculate its conflict degree For set constraints and Calculate its conflict degree At the same time, the degree of logical contradiction is assessed by combining natural language inference models. Overall Conflict Probability It is obtained by weighting and summing the three terms using the nonlinear activation function σ, i.e. ,in , , These are preset weighting coefficients. When When the preset threshold is exceeded, the system determines that there is a specification conflict and needs to activate the conflict resolution strategy.

[0079] When assessing version timeliness deficiencies, the reflection and planning submodule first analyzes whether user questions contain time-related keywords, such as "latest version," "after 2020," or "currently valid." If such keywords exist, it checks whether all relevant specification clauses in the internal knowledge base have attached valid version metadata, including release date, implementation date, obsolescence status, and replacement relationship. If key clauses in the search results lack version identifiers, or contain multiple versions and the currently applicable version cannot be determined, the system determines that the version timeliness is deficient. This determination does not rely on subjective judgment but is based on the completeness and consistency verification of metadata fields.

[0080] In one embodiment, a user asks, "According to the latest version of GB / T 20801, what are the restrictions on the use of austenitic stainless steel pipes in chlorine-containing media?" The input parsing submodule identifies the timeliness keyword "latest version." The initial search returns three clauses, labeled as the 2006 version, the 2020 version, and no version information. Because there is a clause without version information, and the 2006 version has been superseded by the 2020 version but is not marked as obsolete, the system determines that the version timeliness is missing. The reflection and planning submodule then generates a version timeliness verification path, limiting subsequent searches to only the full text of the 2020 version and excluding content without version identification, ultimately obtaining accurate and valid restriction clauses.

[0081] Through the three triggering mechanisms described above, this invention enables the system to proactively identify the limitations of a single search, avoiding erroneous conclusions due to incomplete information, contradictory standards, or version confusion. This mechanism significantly improves the reliability, compliance, and timeliness of question-and-answer results, giving the system the rigorous reasoning characteristics required for engineering practice.

[0082] The answer synthesis and sourcing module is responsible for transforming multi-source retrieved evidence into a final deliverable question-and-answer result, ensuring its high reliability in terms of semantics, logic, and normative basis. The module first performs hierarchical merging and alignment of evidence from core specifications and external sources. Core specifications refer to structured national or industry pressure pipeline design standards in the system's knowledge base; external sources refer to certified authoritative technical databases, standard publishing platforms, or regulatory agency announcements. Evidence is categorized into two levels based on source type: internal clauses and external information. Within each level, fine-grained alignment is performed based on parameter themes, constraint types, and logical roles to eliminate redundancy and ambiguity.

[0083] For factual and definitional questions, such as "What is the definition of allowable stress?" or "The content of Clause 4.2.1 of GB 50316", the answer synthesis and sourcing module directly extracts the original text of the clauses to generate answers. The extraction process preserves the original wording, units, and limiting conditions without rewriting or summarizing, ensuring complete consistency between the answer and the standard text. For questions requiring cross-clause synthesis, such as "What material should be selected in a steam pipeline with a design pressure of 6 MPa and an operating temperature of 400 degrees Celsius?", which involves multiple clauses such as strength calculations, temperature suitability, and material selection, the answer synthesis and sourcing module organizes the language without deviating from the original meaning of the evidence. The organization process follows a logical order: first stating the preconditions, then citing the conclusions of each clause in sequence, and finally forming a coherent technical recommendation. All expressions are strictly limited to the semantic boundaries of the source evidence.

[0084] After generating the initial answer, the answer synthesis and source tracing module performs factual consistency verification. It aligns each technical claim in the generated answer with each source information fragment item by item. For example, if the answer mentions "carbon steel is not suitable for wet hydrogen sulfide environments," the system locates clause 7.3.4 and its footnotes, and verifies whether this statement is supported by the original text. Consistency verification is achieved by calculating a factual consistency score. This score is a weighted fusion of numerical constraint consistency, set constraint consistency, logical implication consistency, and a contradiction penalty item. Numerical constraint consistency determines whether the numerical range of parameters in the answer is completely contained within the range specified in the source clause; set constraint consistency determines whether the permitted or prohibited items listed in the answer belong to the set defined by the source clause; logical implication consistency evaluates whether the answer statement is logically implied by the source text based on a natural language inference model; the contradiction penalty item sets the overall score to zero when any logical contradiction is detected. The weights of each score are not fixed but dynamically set based on the authority level, timeliness, and time decay time constant of the corresponding information fragment, as well as its relevance to the user's question. The level of authority is determined by the type of information source, such as national standards being higher than enterprise guidelines; timeliness is calculated by the time difference between the publication date and the current time, using an exponential decay function; relevance is inherited from the relevance score in the initial retrieval stage.

[0085] When the factual consistency score is not lower than a preset threshold, the system determines the answer to be credible and outputs it. Simultaneously, multi-dimensional traceability information is added. This traceability information includes the standard number and clause number, such as Clause 6.1.2 of GB 50316-2000; if an external source is cited, it includes the external source name, issuing organization, publication date, and reason for adoption. The reason for adoption explains why the external information was included in the answer, for example, "This data has been cited by the NACE MR0175 standard and used to explain material selection in a wet H2S environment." All traceability fields are mapped to specific statements in the answer, supporting sentence-by-sentence tracing.

[0086] In one embodiment, a user asks, "Is 304 stainless steel usable in a water medium with a chloride ion concentration of 80 mg / L?" The system retrieves GB / T 20801.2-2020, clause 5.3.2, which states that "the risk of stress corrosion should be assessed when the chloride ion concentration of austenitic stainless steel exceeds 50 mg / L." Simultaneously, the external source, Appendix M of ASME B31.3-2022, explicitly states that "304 stainless steel is prohibited in water environments with a chloride ion concentration greater than 50 mg / L." The answer synthesis and tracing module first merges the two pieces of evidence, identifying that their constraints are consistent. Since the question involves compliance judgment, the module's language is "304 stainless steel cannot be used in a water medium with a chloride ion concentration of 80 mg / L," and both sources are labeled. Fact consistency verification confirms that the answer value of 80 mg / L exceeds the 50 mg / L limit, and the logic complies with the prohibition, without contradiction. Finally, the answer and the complete tracing chain are output.

[0087] Through the aforementioned mechanism, this invention ensures that each output answer has clear, verifiable specifications or authoritative basis, avoiding subjective assumptions or information distortion. This design significantly enhances the auditability, traceability, and legal compliance of engineering Q&A, meeting the stringent requirements for rigorous technical decision-making in the field of pressure pipeline design.

[0088] The factual consistency score quantifies the overall factual agreement between the generated answer and all source information fragments. Its calculation is based on a weighted average mechanism to ensure that highly authoritative, timely, and relevant evidence has a greater impact on the final score. The formula for calculating this score is as follows: ,in, Representative generates answer With evidence set The score for factual consistency between them; This represents the initial answer generated by the system; This represents the set of source evidence used to generate the answer; Representative evidence set The total number of source information fragments contained therein; The summation index refers to the k-th source information segment. This represents the weight of the k-th source information segment. This weight comprehensively reflects the segment's authority, timeliness, and relevance. This represents the result of the consistency verification function. This function is used to determine the answer. With the k-th source information fragment Whether they are consistent or not, usually a value of 0 indicates inconsistency, and a value of 1 indicates consistency; This represents the k-th specific source information fragment in the evidence set.

[0089] For example, the system will first generate answer A. g The system is decomposed into several verifiable technical claim units. Each claim unit corresponds to a specific parameter, constraint, or logical judgment. Subsequently, the system traverses each source information fragment in the evidence set I. Perform a consistency verification operation. The function uses structured comparison to determine whether the answer is supported by the given fragment. If all technical claims in the answer are supported... If direct evidence is found, and there is no extraneous or contradictory content, then... Returns 1 otherwise; otherwise, returns 0. This verification process does not rely on semantic similarity, but rather on precise numerical range inclusion relationships, set subset relationships, or logical implication relationships to ensure the deterministic nature of the result. Weight It is not a fixed value, but rather dynamically calculated from three dimensions. The first dimension is the level of authority, reflecting the normative validity of the source information fragment. National standard clauses have a higher level of authority than industry guidelines, and regulatory agency announcements have a higher level of authority than ordinary technical forums. The second dimension is timeliness, calculated using the time difference between the publication date and the current time. The calculation uses an exponential decay model. ,in This represents a timeliness score based on the publication date, with a value ranging from 0 to 1; Represents the natural exponential function; This represents the difference between the publication date of the k-th segment and the current system time, i.e., the time span; This represents the preset decay time constant. This constant controls the rate of decay. The smaller the value, the faster the weight of older documents decreases, reflecting a preference for the new specifications. Each of the three dimensions is multiplied by a preset coefficient. , The weights are then added together with γ to obtain the final weight. . The comprehensive weight represents the k-th source information segment; These represent preset non-negative real coefficients, which are used to adjust the proportions of authority level, timeliness, and relevance in the total weight; The authoritative rating representing the source of the k-th segment; The timeliness score for the k-th segment is determined by the publication date. The semantic relevance score between the user's question and the k-th fragment is typically inherited from the score in the initial retrieval phase.

[0090] In one embodiment, a user asks, "Is it permissible to use carbon steel pipes designed for a temperature of 425 degrees Celsius?" The system generates the answer, "Carbon steel pipes must not be used when the design temperature exceeds 425 degrees Celsius." Evidence set I contains two fragments: This refers to Clause 6.1.3 of GB 50316-2000, published in 2000, which states that "the maximum service temperature of carbon steel is 425 degrees Celsius." This is an internal technical memo from an engineering company, issued in 2024, stating that "the recommended operating temperature for carbon steel should not exceed 400 degrees Celsius." The system monitors Ag and... The verification process revealed that the answer "Do not use above 425 degrees Celsius" was incorrect. The statement "the highest temperature is 425 degrees Celsius" is logically equivalent, therefore... .right During verification, the answer did not mention the 400-degree Celsius limit and did not adopt the suggestion, but it also did not contradict it. However, due to... This is a suggested rather than mandatory rule, and since the answer did not cite its content, the system determines that this fragment does not support the current answer's claim. Weight The price is relatively high because it originates from national standards and remains valid even after the expiration of its shelf life. Due to its low level of authority, it is considered lower. Ultimately If the value is close to 1, the output threshold is met.

[0091] Through this computational mechanism, the system can distinguish the contribution of core specifications and auxiliary information, avoiding interference from low-authority or outdated information in the evaluation of answer credibility. Simultaneously, binarization... The function ensures that consistency judgments have clear boundaries, preventing erroneous approvals caused by fuzzy matching. This design enables the system to output answers that strictly conform to the original specifications even in complex, multi-source environments, significantly improving the compliance and security of engineering decisions.

[0092] The interaction and knowledge evolution module is a key component of the system's continuous learning and knowledge updating. Its core function is to transform user feedback into actionable knowledge evolution signals and dynamically expand the system's knowledge base while ensuring the rigor of the specifications. This module first provides a user feedback interface, allowing users to rate the system's output answers, submit corrections, or add technical comments. The rating uses multi-level discrete options to reflect the user's comprehensive judgment on the accuracy, completeness, and practicality of the answer; corrections can propose revisions for specific clause citation errors, parameter value deviations, or logical contradictions; comments are used to supplement empirical explanations in engineering practice or point out marginal scenarios not covered by the specifications.

[0093] Upon receiving user feedback, this module maps it into structured quantitative signals. For ratings, the system converts them into confidence adjustment factors; for error corrections, the system extracts the entities, parameters, and relationships involved and generates knowledge correction proposals to be verified; for comments, the system identifies whether they contain formalizable technical rules. All quantitative signals are associated with the source information fragments upon which the original answer depends. Subsequently, the system automatically updates the confidence values ​​of the corresponding relationships in the knowledge graph based on the feedback strength and source reliability. If the feedback points to external source content, the global credibility score of that external source is updated simultaneously. Both confidence and credibility are normalized real numbers used to influence weight allocation in subsequent retrieval and consistency verification processes.

[0094] When certain conditions are met, the system triggers a new knowledge structure integration process. These conditions include: a user giving a positive evaluation higher than a preset threshold to an answer containing external information, and this evaluation coming from a registered engineer account with professional qualifications; or the system administrator explicitly marking the feedback as important supplementary information through a manual review interface. Once triggered, the module initiates the structure integration process. This process first precisely extracts technical elements from the external information, including parameter names, parameter values, and their applicable conditions. For example, extracting elements such as "duplex stainless steel," "upper limit of chloride ion concentration," "200 mg / L," and "applicable to seawater cooling systems" from user-submitted industry reports. Subsequently, the system calls a predefined ontology model and maps these elements to corresponding attribute fields through a slot-filling mechanism. The ontology model explicitly defines the legal attributes and data types of various entities in the pressure pipeline domain, ensuring that the mapping results conform to semantic constraints.

[0095] After mapping is completed, the system generates structured knowledge proposals in manual-assisted mode and submits them to the review interface. The review interface is open to technical experts with review privileges and allows viewing of original feedback, extraction results, ontology mapping paths, and potential conflict detection reports. Reviewers can approve, reject, or request modifications to the proposals. If the review is successful, the system writes the new knowledge into the knowledge base with a structure completely consistent with the original specification knowledge. Specific operations include: creating virtual clause entries in the clause record database, labeling their source as external validation knowledge; creating or updating corresponding nodes and relationships in the knowledge graph, setting initial confidence levels; generating a semantic vector for the knowledge in the vector index and associating it with a unique identifier; and attaching complete source metadata, including the external source name, issuing organization, publication date, user submission time, and reviewer identifier. All newly added content supports subsequent source tracing and version tracking.

[0096] In one embodiment, a user submitted feedback stating that "according to the latest version of NACE SP0472, super duplex steel UNSS32750 is still safe to use in seawater with a chloride ion concentration of up to 1000 mg / L," and attached a screenshot of the original standard. This user is a system-certified corrosion engineer. Upon receiving the feedback, the system extracts the parameters "UNSS32750," "upper limit of chloride ion concentration," "1000 mg / L," and "applicable medium: seawater." Through slot filling, this is mapped to the "chloride ion concentration resistance" attribute of the "materials" class in the system body. The proposal is pushed to the review interface, and experts approve it after verifying Section 4.2 of NACE SP0472-2023. The system then creates a new material node in the knowledge graph, establishing a relationship with concepts such as "seawater conditions" and "stress corrosion cracking," and marking the source as NACE SP0472-2023 and the user's submission record. Subsequently, when users query material selection for high-chlorine environments, this knowledge can be retrieved and referenced normally.

[0097] Through the aforementioned mechanism, the system has transformed from a static, standardized knowledge base to a dynamically evolving knowledge system. User feedback is no longer an isolated evaluation but rather an effective input driving the continuous improvement of the knowledge base. This design, while maintaining the authority of the standards, effectively incorporates proven external technological advancements, significantly enhancing the system's adaptability to emerging materials, new operating conditions, and cross-standard scenarios, thereby strengthening its long-term usability and technological foresight in engineering practice.

[0098] Specific limitations regarding the iterative reasoning-based intelligent question-and-answer system for pressure pipeline design can be found in the following section on the limitations of the iterative reasoning-based intelligent question-and-answer system and method for pressure pipeline design, and will not be repeated here. Each module in the aforementioned iterative reasoning-based intelligent question-and-answer system for pressure pipeline design can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0099] In one embodiment, such as Figure 2 As shown, an iterative reasoning-based intelligent question-answering system and method for pressure pipeline design are provided, which can be applied to... Figure 2 Taking China as an example, the following specific steps will be used:

[0100] S10: The knowledge base construction module parses the pressure pipeline design specification document into structured clause records and builds a knowledge graph and vector index containing specification number, version identifier, chapter path, clause number and cross-reference relationship.

[0101] S20: The retrieval and iteration module performs intent recognition and constraint extraction on the natural language questions raised by the user, and performs a preliminary retrieval based on the knowledge graph and vector index; if the retrieval results have insufficient information coverage, conflicts in numerical or set constraints of normative clauses, or lack of version timeliness, the reflection and planning submodule is activated to generate supplementary query strategies and execute multiple rounds of iterative retrieval.

[0102] S30: The internal and external evidence obtained by the answer synthesis and tracing module are integrated and retrieved to generate a preliminary answer under the strong constraints of facts and normative original texts in the knowledge graph, and the factual consistency score between the answer and the source information fragment is calculated; when the score is not lower than a preset threshold, the answer with additional multi-dimensional tracing information is output.

[0103] S40: The interaction and knowledge evolution module receives user feedback on the answer, dynamically adjusts the confidence of relationships in the knowledge graph and the credibility of external information sources based on the feedback, and writes high-value external information back to the knowledge base construction module in a structured form after the review mechanism is passed.

[0104] Specifically, the iterative reasoning-based intelligent question-answering method for pressure pipeline design described in this invention achieves the entire process from standard text to verifiable, traceable, and evolving engineering questions and answers through four processing stages. First, the method uses a knowledge base construction module to perform deep structuring processing on the original pressure pipeline design standard document. The standard document includes national and industry-issued standard texts, such as GB 50316 and the GB / T 20801 series. The knowledge base construction module uses a parsing engine combining rules and machine learning to segment the document by clause granularity and generate a structured record for each clause. The structured record includes the standard number, version identifier, chapter path, clause number, main text content, table references, and footnote information. Simultaneously, the system automatically identifies cross-references within clauses and across documents, such as "see Clause 7.2.1" or "in accordance with TSG D0001," and constructs directed reference edges to associate source and target clauses. Based on this, combined with a domain ontology model, entities and their attribute relationships are extracted to form triplet knowledge, which is then populated into the knowledge graph. In addition, each clause and key sentence is converted into a semantic vector, and a vector index is built. This vector index is bidirectionally linked to structured records and knowledge graph nodes through unique identifiers, providing support for subsequent semantic retrieval.

[0105] After a user poses a natural language question, the system enters the second phase, where the retrieval and iteration module performs intent understanding and evidence retrieval. The retrieval and iteration module first performs semantic parsing of the question, identifying the user's intent type, such as parameter query, compliance judgment, or design selection, and extracting key entities and constraints, such as "design pressure 4 MPa," "medium is wet hydrogen sulfide," and "material is carbon steel." Simultaneously, it combines multi-turn dialogue context to resolve referential inconsistencies, ensuring semantic integrity. Subsequently, the system performs dense semantic retrieval and sparse keyword retrieval in parallel, using vector indexes and inverted indexes respectively to obtain candidate clauses, and normalizes and weights the scores of the two types of results to generate a preliminary evidence ranking. If the preliminary retrieval results have insufficient information coverage, conflicts in numerical ranges or set constraints of regulatory clauses, or lack of effective version metadata leading to uncertainty in timeliness, the reflection and planning submodule is activated. This submodule quantifies the current information gaps and generates supplementary query strategies, including hierarchical tracing paths that trace upwards along the citation chain, association paths that extend horizontally to relevant technical chapters, timeliness verification paths that limit specific year versions, deep paths that perform fine-grained mining of ambiguous statements, or supplementary paths that call on external authoritative sources. Based on these strategies, it initiates multiple rounds of iterative retrieval until the evidence set meets the requirements for completeness and consistency.

[0106] The third stage, performed by the answer synthesis and sourcing module, generates answers and verifies credibility. This module integrates multi-source evidence from core specifications and certified external sources, stratifying and logically aligning it by technical theme. For factual or definitional questions, the original text of the specification is directly extracted as the answer; for questions requiring cross-clause synthesis, the language is organized strictly without deviating from the original meaning of the evidence to ensure technical logical coherence. After generating the preliminary answer, the system compares it item by item with each source information fragment, calculating a factual consistency score. This score is derived from a weighted fusion of numerical constraint consistency, set constraint consistency, logical implication consistency, and contradiction penalty items, and dynamically sets weights based on the authority level, timeliness, and time decay time constant of each information fragment, as well as its relevance to the question. When the factual consistency score is not lower than a preset threshold, the system outputs the final answer, along with multi-dimensional sourcing information, including specification number and clause number, external source name, issuing organization, publication date, and reasons for adoption, ensuring that each technical conclusion can be traced back to its original basis.

[0107] The fourth stage involves continuous system optimization through the interaction and knowledge evolution module. This module provides a user feedback interface to receive user ratings, corrections, or technical comments on answers. Feedback information is mapped into quantitative signals to dynamically adjust the confidence level of corresponding relationships in the knowledge graph and the credibility of external information sources. When a user gives a positive evaluation of an answer containing external information that exceeds a preset threshold, or when it is marked as an important supplement by a manual review interface, the system triggers a new knowledge structure integration process. This process extracts parameter names, parameter values, and applicable conditions from the external information, maps them to attribute fields defined in the ontology through slot filling, and after manual review and approval, writes them back to the knowledge base construction module in a structure completely consistent with the original specification knowledge, synchronously updating the clause records, knowledge graph nodes, and vector indexes, and attaching complete source and version metadata.

[0108] In a complete implementation, a user asked, "Can No. 20 steel pipes be used under operating conditions of 450 degrees Celsius and a medium containing moist H2S?" The system resolved the entities "No. 20 steel" and the constraints "450 degrees Celsius" and "moist H2S". The initial search returned GB 50316, Clause 6.1.2 regarding the operating temperature limits for carbon steel, and Clause 7.3.4 regarding environmental restrictions in moist H2S. The reflection module found that these two clauses did not clearly define the conclusions under the combined operating conditions, indicating insufficient information coverage. Therefore, a lateral correlation path was generated, supplementing the search with relevant content from NACE MR0175. The answer synthesis module confirmed that No. 20 steel must meet hardness and heat treatment requirements in moist H2S, and that 450 degrees Celsius exceeds its recommended upper limit; therefore, a negative answer was generated, with three sources noted. The user, a registered engineer, gave a high evaluation and supplemented the search with a project acceptance report as supporting evidence. The interactive module triggers the review process. After expert confirmation, the information "The maximum applicable temperature of No. 20 steel is 425 degrees Celsius (wet H2S conditions)" is written back into the knowledge base as structured knowledge. Subsequent similar questions can directly reference this newly added knowledge.

[0109] Through the above four-stage closed-loop process, the iterative reasoning-based intelligent question-and-answer method for pressure pipeline design realizes the transformation from static specifications to dynamic intelligent question-and-answer, significantly improving the accuracy, interpretability, and timeliness of pressure pipeline design consultation, and effectively supporting engineering technicians in making compliant, safe, and efficient technical decisions.

[0110] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc. 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 used as 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.

[0111] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An iterative reasoning-based intelligent question-and-answer system for pressure pipeline design, characterized in that, include: The knowledge base construction module is used to parse pressure pipeline design specification documents into structured clause records and build a knowledge graph and vector index containing specification number, version identifier, chapter path, clause number and cross-reference relationship; The retrieval and iteration module is used to identify the intent and extract the constraints of the natural language questions raised by the user, and to perform a preliminary retrieval based on the knowledge graph and vector index. If the retrieval results have insufficient information coverage, conflicts in numerical or set constraints of normative clauses, or lack of version timeliness, the reflection and planning submodule is activated to generate supplementary query strategies and execute multiple rounds of iterative retrieval. The answer synthesis and tracing module is used to integrate internal and external evidence obtained through retrieval, generate a preliminary answer under the strong constraints of facts and normative original texts in the knowledge graph, and calculate the factual consistency score between the answer and the source information fragment; when the score is not lower than a preset threshold, the answer with additional multi-dimensional tracing information is output. The interaction and knowledge evolution module is used to receive user feedback on answers, dynamically adjust the confidence of relationships in the knowledge graph and the credibility of external information sources based on the feedback, and write high-value external information back to the knowledge base construction module in a structured form after approval.

2. The iterative reasoning-based intelligent question-and-answer system for pressure pipeline design according to claim 1, characterized in that, The knowledge base construction module performs the following processing flow: The input pressure piping design specification document is parsed in a structured manner, and a structured record containing clause numbers, titles, appendix identifiers, table references and footnotes is generated according to the clause granularity. Identify cross-references within and across specifications, construct reference relationship edges, and associate source clauses with target clauses; The terminology is standardized by combining domain dictionaries, and entities and their relationships are extracted using sequence labeling and rule-based methods to form triples and associate them with the corresponding clause records. Semantically vectorize the clauses and key sentences, create a vector index, and associate it with the clause records using a unique identifier; Based on the ontology definition, unified concepts, attributes, and relational constraints are defined, and the extracted results are populated into the knowledge graph. Receive the review conclusions from the interaction and knowledge evolution modules, and write back the externally verified new knowledge to the clause records, knowledge graph and vector index with the same structure fields after the review is approved.

3. The iterative reasoning-based intelligent question-and-answer system for pressure pipeline design according to claim 1, characterized in that, The retrieval and iteration module includes an input parsing submodule, a multi-strategy evidence recall and ranking submodule, a problem decomposition and synthesis submodule, and a reflection and planning submodule. The input parsing submodule is used to complete the intent parsing of user questions, extract key entities and constraints, and resolve referential issues in combination with the dialogue context. The multi-strategy evidence recall and ranking submodule is used to perform dense semantic retrieval and sparse keyword retrieval, and to normalize and weight the scores of the two to generate a ranking of candidate evidence. The problem decomposition and synthesis submodule is used to decompose a user problem into several sub-problems when it is detected that the user problem contains multiple working conditions, multiple constraints, or cross-term dependencies, and then perform constraint alignment and logical synthesis after each sub-problem is retrieved. The reflection and planning submodule is used to quantify information gaps and generate supplementary query strategies when there is insufficient evidence coverage, regulatory conflicts, or version timeliness issues. These strategies include hierarchical tracing paths, horizontal correlation paths, version timeliness verification paths, internal in-depth mining paths, or external authoritative source query paths.

4. The iterative reasoning-based intelligent question-and-answer system for pressure pipeline design according to claim 1, characterized in that, The reflection and planning submodule is activated based on one of the following triggering conditions: Calculate the information dispersion index of the retrieved normative clauses in terms of their level of validity or chapter distribution. When the index is lower than a preset threshold, it is determined that the evidence coverage is insufficient. For two or more highly relevant clauses, a conflict probability score is calculated on the numerical range or set constraint of the same parameter. When the score exceeds a preset threshold, it is determined that there is a specification conflict. When a user's question contains time-sensitive keywords and the internal knowledge base lacks valid version metadata, it is determined that the version timeliness is missing.

5. The iterative reasoning-based intelligent question-and-answer system for pressure pipeline design according to claim 4, characterized in that, The information dispersion index is knowledge density entropy, and its calculation formula is as follows: ； in, Entropy represents the knowledge density and indicates the dispersion of information distribution. Represents the set of search results; Represents the total number of terms returned by the search; This represents the index variable in the summation operation, referring to the m-th item in the search results; represents the normalized weight of the m-th clause in the search results, where the sum of the weights of all clauses is normalized to 1; It represents a tiny positive number used to prevent logarithmic singularity.

6. The iterative reasoning-based intelligent question-and-answer system for pressure pipeline design according to claim 4, characterized in that, The conflict probability score is obtained by weighted fusion of numerical constraint conflict degree, set constraint conflict degree, and logical contradiction degree; for numerical constraints... Conflict level is defined as: ； in, Represents the degree of conflict of numerical constraints; This represents the interval formed by the lower and upper limits of the numerical constraint of a certain parameter in the first clause of the specification; This represents the interval formed by the lower and upper limits of the numerical constraints for the same parameter in the second clause of the specification; This represents the length of the intersection of the two intervals. It represents the length of the union of two intervals. Set constraints Define the degree of conflict: ； in, Represents the degree of conflict of constraints in a set; This represents the set of allowed or restricted values ​​for a parameter in the first clause of the specification. This represents the set of allowed or restricted values ​​for the same parameter in Article 2 of the regulations; The number of elements in the intersection of two sets; It represents the number of elements in the union of two sets. The overall probability of conflict is: ； in, Represents the overall probability of conflict; Represents a non-linear activation function; The preset weight coefficients representing the degree of conflict of numerical constraints; The preset weight coefficients corresponding to the degree of conflict of set constraints; Preset weighting coefficients representing the degree of logical contradiction; This represents the numerical constraint conflict degree obtained from the aforementioned calculation; This represents the set constraint conflict degree obtained from the aforementioned calculation; It represents the degree of logical contradiction.

7. The iterative reasoning-based intelligent question-and-answer system for pressure pipeline design according to claim 1, characterized in that, The answer synthesis and tracing module performs the following operations when generating answers: Evidence from core specifications and external sources is hierarchically merged and aligned. For factual and definitional questions, answers are generated by extracting the original text of the clauses. For issues requiring comprehensive consideration across multiple clauses, organize the language appropriately without deviating from the original meaning of the evidence; The generated answer is aligned with each source information fragment item by item, and the factual consistency score is calculated. The factual consistency score is obtained by weighted fusion of numerical constraint consistency, set constraint consistency, logical implication consistency and contradiction penalty term, and the weights are dynamically set according to the authority level, timeliness and relevance of the information fragment; When the score is not lower than the preset threshold, the answer is output along with multi-dimensional traceability information, including the specification number and clause number, the name of the external source, the issuing organization, the publication date, and the reason for adoption.

8. The iterative reasoning-based intelligent question-and-answer system for pressure pipeline design according to claim 7, characterized in that, The formula for calculating the factual consistency score is as follows: ； in, Representative generates answer With evidence set The score for factual consistency between them; This represents the initial answer generated by the system; This represents the set of source evidence used to generate the answer; Representative evidence set The total number of source information fragments contained therein; The summation index refers to the k-th source information segment. The weight represents the k-th source information segment; This represents the result of the consistency verification function, which is used to determine the answer. With the k-th source information fragment Are they consistent? This represents the k-th specific source information fragment in the evidence set.

9. The iterative reasoning-based intelligent question-and-answer system for pressure pipeline design according to claim 1, characterized in that, The interaction and knowledge evolution module performs the following operations: Provide a user feedback interface to receive user ratings, corrections, or comments on answers; The feedback is mapped as a quantized signal and used to update the confidence of relations in the knowledge graph and the credibility of external sources; When a user gives a positive rating to an answer that contains external information that exceeds a preset threshold, or when the manual review interface marks it as an important supplement, the new knowledge structure integration process is triggered. The structured integration process includes: extracting parameter names, parameter values, and applicable conditions from external information; mapping them to attribute fields defined in the ontology through slot filling; submitting for review in manual assistance mode; and creating or updating knowledge graph nodes and relationships after approval, and attaching source and version metadata.

10. An iterative reasoning-based intelligent question-answering method for pressure pipeline design, characterized in that, include: The knowledge base construction module parses the pressure pipeline design specification document into structured clause records and builds a knowledge graph and vector index containing specification number, version identifier, chapter path, clause number and cross-reference relationship; The retrieval and iteration module performs intent recognition and constraint extraction on the natural language questions raised by the user, and conducts a preliminary retrieval based on the knowledge graph and vector index. If the retrieval results have insufficient information coverage, conflicts in numerical or set constraints of normative clauses, or lack of version timeliness, the reflection and planning submodule is activated to generate supplementary query strategies and execute multiple rounds of iterative retrieval. The internal and external evidence retrieved by the answer synthesis and tracing module is integrated to generate a preliminary answer under the strong constraints of facts and normative original texts in the knowledge graph, and the factual consistency score between the answer and the source information fragment is calculated; when the score is not lower than a preset threshold, the answer with additional multi-dimensional tracing information is output. The interaction and knowledge evolution module receives user feedback on the answers and dynamically adjusts the confidence of relationships in the knowledge graph and the credibility of external information sources based on the feedback. After the review mechanism is approved, high-value external information is written back to the knowledge base construction module in a structured form.

Images