Steel structure calculation method and system based on physical logic anchoring and heterogeneous decoupling

Abstract

The application provides a steel structure calculation method and system based on physical logic anchoring and heterogeneous decoupling, and relates to the technical field of intelligent steel structure calculation. The method comprises the following steps: extracting user's unstructured engineering requirements into a structured feature vector, constructing a directed acyclic computation graph along a knowledge graph; starting knowledge graph reasoning and retrieval enhancement in parallel to generate a two-way evidence, calculating a logic anchoring coefficient Lambda, which includes a number version consistency judgment, a specification timeliness dynamic weight and a strong item penalty, comparing Lambda with a dynamic reliability threshold and executing a three-branch decision to eliminate index illusion; routing the calculation instruction to a physically isolated sandbox after triple audit to determine the solution and eliminate calculation illusion; based on the logic dependence hypergraph, the dissenting node is traced back and locally recalculated, and a compliant calculation report is output. From the architecture level, the calculation illusion and index illusion in the steel structure calculation are eliminated, and the industrial numerical precision and reliable specification compliance traceability are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent computing technology for steel structures, specifically to a steel structure computing method and system based on physical logic anchoring and heterogeneous decoupling. Background Technology

[0002] Steel structure engineering calculations are a crucial branch of structural engineering, involving highly coupled design parameters, complex boundary conditions, and stringent safety and compliance requirements. While the development of deep learning and Large Language Modeling (LLM) technologies has led to initial progress in human-computer natural language interaction, the inventors have found that the index numbers of code provisions returned by general-purpose LLM models (such as Doubao) for steel structure stability calculations are frequently incorrect. For example, when asking "Which code should be used for the stability calculation of axially compressed members?", the model might return a number not found in GB 50017 or a clause that has been repealed. Furthermore, beginners lacking experience in steel structure engineering often struggle with how to ask questions, even when using LLM models—they are unclear about which parameters to provide, how to describe load conditions, and where to obtain geometric dimensions, resulting in "ineffective questioning and unusable answers." Existing RAG or prompt word engineering methods cannot fundamentally solve these problems because the root cause lies in the structural conflict between the probabilistic generation mechanism of LLM models and the deterministic requirements of engineering calculations, as well as the lack of proactive guidance in the design process.

[0003] Existing technical solutions suffer from domain adaptability deficiencies, which are mainly manifested in the following three aspects: 1. Misalignment in the mapping between engineering features and fuzzy semantic input In steel structure construction sites, initial calculation requirements are often presented in discretized, unstructured natural language descriptions. General-purpose large-scale models lack the ability to recognize the physical entity and mechanical properties of steel structures, making it difficult to accurately map such vague construction intentions into a structured set of calculation parameters that meet mechanical boundary conditions and specification requirements. This semantic alignment defect between construction intentions and standard calculation models directly leads to inaccurate starting points for calculations, failing to guarantee the correctness of subsequent calculation chains.

[0004] 2. Mechanism conflict between symbolic logic reasoning and probabilistic generative models Steel structure calculations must strictly adhere to the arithmetic logic and topological structure of national standards such as the "Steel Structure Design Standard" (GB 50017). However, general-purpose large language models, based on autoregressive generation mechanisms using probability sampling, are prone to computational illusions when dealing with engineering formulas involving strong causality and strong logic. This is particularly true when performing multi-step cascaded comprehensive verification of tensile, shear, and compressive strength; small probabilistic deviations in individual steps can be amplified at each step, ultimately causing the entire calculation path to deviate significantly from physical laws, failing to meet the stringent requirements for numerical determinism in steel structure engineering.

[0005] 3. Retrieval bottleneck between dynamic standardization systems and static knowledge distribution The regulatory framework for steel structures is complex, with intricate logical relationships among its provisions, including deep nesting, mutual exclusion, and version updates. Existing general-purpose Retrieval Augmentation (RAG) technologies, when applied to texts in this specialized field, often only achieve superficial text similarity matching due to a lack of understanding of the semantic relationship graph of steel structures. This prevents artificial intelligence (AI) from accurately interpreting the unique mapping between regulatory provisions and specific computational scenarios. It also makes it difficult to trace compliance evidence in real-time and accurately during calculations, easily leading to the illusion of fabricated or misused regulatory provisions. Consequently, the calculation results fail engineering compliance reviews due to a lack of traceable regulatory basis.

[0006] Existing similar technical solutions and their shortcomings: Patent document CN121188144B discloses a method, apparatus, and device for reducing computational illusions using a large language model in the bridge industry. This solution trains a specialized large language model based on existing bridge data and builds a model group with a general large language model. It reduces illusions through accuracy comparison after extracting unstructured document data and by selecting different models for computation based on computational complexity and verifying the results. However, this solution relies on bridge industry data and databases, making it incompatible with the steel structure field's regulatory system. Its comparison and verification mechanism is a "post-hoc" verification after the results are generated, failing to establish physical isolation or strong pre-constraints between the large language model and the numerical computation layer, thus failing to eliminate computational illusions at the architectural level. Furthermore, it does not disclose a logical dependency hypergraph local incremental recalculation mechanism applicable to collective calculations during the design phase.

[0007] Patent document CN120338721A discloses a full-process project management system based on a large language model. This solution covers the entire field of architectural engineering, but it has the following shortcomings when facing scenarios requiring precise calculations at the specification level, such as detailed design of steel structures: it lacks specific adaptation for steel structure-specific calculation scenarios; it lacks a mandatory binding mechanism between calculation results and the original text of current design specifications; it does not include technical means to prevent calculation illusions and index illusions; and it lacks accurate review and correction support when users raise objections to the calculation results.

[0008] Patent document CN121542439A discloses a hydraulic design calculation system based on a large model. This solution proposes integrating specialized calculation modules with the large model through Retrieval-Augmented Generation (RAG) technology. However, it suffers from the following drawbacks: its constructed standard knowledge base is specifically designed for the hydraulic field and is incompatible with the steel structure standard system; functionally, it relies primarily on a single RAG search, lacking the structured reasoning capabilities of a knowledge graph and struggling to handle deep logical connections between standard clauses; it places excessive demands on users' one-time input parameters and lacks a precise mechanism for tracing the calculation results back to the original standard text.

[0009] Furthermore, while the general RAG architecture proposed by MetaAI and other institutions is widely used to improve the accuracy of fact-answering in large language models, its direct application in engineering computing still has the following inherent drawbacks: ① It only performs single-path retrieval enhancement, lacks structured knowledge reasoning and cross-validation capabilities, and cannot discover and correct deep contradictions between retrieval results and underlying engineering specification logic; ② The lack of a rule engine to reasonably constrain the values ​​of retrieved parameters can easily lead to the incorrect introduction of normative clauses applicable to other scenarios into the current calculation, resulting in compliance risks; ③ It is impossible to achieve triple verification of “calculation results – corresponding standard clauses – complete calculation process” at the system level, which makes it difficult to meet the strict requirements of traceability and auditability in the engineering field.

[0010] In summary, the field of intelligent computing for steel structures urgently needs a new generation of technical solutions that can fundamentally solve the three core problems of semantic mapping distortion, computational illusion, and indexing illusion, and have efficient error correction capabilities. Summary of the Invention

[0011] This invention addresses the technical problems in existing steel structure calculation techniques, such as semantic mapping distortion, computational illusion, and indexing illusion, which arise from the direct application of general-purpose large models. It proposes a steel structure calculation method and system based on physical logic anchoring and heterogeneous decoupling. Through a heterogeneous decoupling physical isolation architecture and a dual-path heterogeneous source verification mechanism, this invention eliminates semantic mapping distortion, computational illusion, and indexing illusion in intelligent steel structure calculation at the architectural level, achieving industrial-grade numerical accuracy and reliable standard compliance traceability.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A steel structure calculation method based on physical logic anchoring and heterogeneous decoupling, the method comprising: S1. The structured feature vectors extracted from the user's unstructured engineering requirements are used to perform physical topological chain reasoning along the preset knowledge graph to construct a directed acyclic computation graph; S2. Two evidence collection paths are initiated in parallel. Path A extracts nodes representing the legal reasoning chain and their canonical identifiers from the preset knowledge graph along the computation graph. Path B calls the retrieval enhancement generation engine to retrieve the original canonical text from the vectorized canonical library and parse its standard index number. The logical anchoring coefficient Λ is calculated, and this coefficient is specifically defined as: , in, For numbering and version consistency judgment operators; To standardize unique identifiers; For standard index number; To standardize the dynamic weighting of timeliness; These are penalties set according to the level of enforceability of the provisions. For semantic similarity function, For logical description text, To standardize semantic text; Λ is compared with a dynamic reliability threshold. Compare and execute the three-branch decision: if Λ≥ Then proceed to subsequent processing if 0 < Λ < If Λ≤0, then the error output is overwritten with the canonical information of path B and the computation graph is reconstructed and re-verified until convergence. S3. For the computation instructions to be issued, execute the instruction whitelist filtering, topological logic isomorphism audit, and coefficient domain boundary hard constraint injection triple audit in sequence; the audited instructions are routed to the physically isolated operator execution environment through a one-way encrypted pipeline. Under the condition of depriving the large language model of the right to intervene in numerical operations within the process-level sandbox, the deterministic numerical solution is executed through the dynamic constraint verification engine. S4. Anchor user objections to the calculation results to the target node in the pre-maintained logical dependency directed hypergraph, and trace back along the hyperedge to extract the dependency subgraph; inject modified parameters only into the dependency subgraph and perform local incremental recalculation; force the recalculation results to S2 and S3 for compliance verification, and output a compliance calculation report after the verification is passed.

[0013] On the other hand, the present invention also provides a steel structure calculation system based on physical logic anchoring and heterogeneous decoupling, for implementing the above-mentioned method, the system comprising: The semantic awareness module extracts structured feature vectors from users' unstructured engineering needs, performs physical topological chain reasoning along a preset knowledge graph, and constructs a directed acyclic computation graph. The logic scheduling module initiates two evidence collection paths in parallel. Path A extracts nodes representing the legal reasoning chain and their canonical identifiers from the knowledge graph along the computation graph. Path B calls the retrieval enhancement generation engine to retrieve the original canonical text from the vectorized canonical library and parse its standard index number. The module also calculates the logical anchoring coefficient Λ, which is specifically defined as follows: , in, For numbering and version consistency judgment operators; To standardize unique identifiers; For standard index number; To standardize the dynamic weighting of timeliness; These are penalties set according to the level of enforceability of the provisions. For semantic similarity function, For logical description text, To standardize semantic text; Λ is compared with a dynamic reliability threshold. Compare and execute the three-branch decision: if Λ≥ Then proceed to subsequent processing if 0 < Λ < If Λ≤0, the error output is forcibly overwritten with the standard information of path B, and the computation graph is reconstructed and re-verified until convergence. The computation instructions to be issued are sequentially subjected to instruction whitelist filtering, topological logical isomorphism auditing, and coefficient domain boundary hard constraint injection triple auditing. The instructions that pass the audit are routed to the operator execution module through a one-way encrypted pipeline, and the computation results returned by the module are received. In addition, the user's objection to the computation result is anchored to the target node in the pre-maintained logical dependency directed hypergraph. The dependency subgraph is extracted by tracing back along the hyperedge. The modified parameters are injected only into the dependency subgraph and a local incremental recalculation is triggered. The recalculation result is forcibly routed back for compliance verification. The operator execution module, under the condition of depriving the large language model of its right to intervene in numerical operations within the process-level sandbox, performs deterministic numerical solutions on the received audited instructions through the dynamic constraint verification engine, and sends the calculation results back to the logic scheduling module. The logic scheduling module is also used to output a compliance calculation report after the verification is passed.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. Eliminate computational illusion at the architectural level to achieve industrial-grade numerical accuracy. This invention pioneers a heterogeneous decoupling architecture that physically isolates the semantic orchestration of large models from the numerical computation execution of the physical operator library. Compared to the computational illusion problem commonly found in the cascade verification of steel structures using logical reasoning models disclosed in the industry, this invention fundamentally eliminates the direct intervention of LLM in numerical computation at the architectural level through a third-order physical isolation mechanism, ensuring that the calculation results stably meet industrial-grade accuracy requirements.

[0015] 2. Eliminate index illusions at the logical level to ensure accurate compliance documentation. Based on the PAFO-HEA (Phantom Apocalypse Optimization-Optimization-Algorithm) dual-path heterogeneous verification, the system establishes a strong logical anchor between the calculation steps and current standards. Compared with traditional single RAG retrieval or pure knowledge graph reasoning, this invention significantly reduces the risk of mismatch and fabrication in standard citations through dual-path cross-validation and dynamic threshold determination of the logical anchoring coefficient Λ, achieving automatic identification and reliable interception of mandatory national provisions.

[0016] 3. Resolves parameter transmission gaps, significantly improving human-computer interaction efficiency. Based on a multi-stage logic state machine and feature slot detection mechanism, the system can automatically identify logical omissions in construction intentions and perform heuristic completion. Compared with traditional manual calculations or direct interaction using general large language models, the system achieves significantly improved efficiency in the entire process. Furthermore, thanks to a local backtracking correction algorithm, the system only needs to perform incremental recalculation for affected nodes during the scheme correction phase, avoiding full-scale redundant calculations as in traditional methods and greatly improving design iteration efficiency.

[0017] 4. Accumulate the experience of industry experts and improve the standardization level of detailed design. This invention transforms complex engineering calculation specifications into a reasonable knowledge ontology, enabling junior technicians to produce standardized calculation reports at the level of senior engineers under AI guidance.

[0018] 5. The combined effects of the technologies produce a synergistic qualitative change, resulting in unexpected technological effects. This invention is not a simple patchwork of existing technologies such as RAG, knowledge graphs, large models, and state machines. Through the fourth-order cooperative coupling of the fourth-order illusion resolution algorithm (PAFO-HEA), the system produces unexpected technical effects that cannot be achieved by any single technology or its conventional combination: The synergistic elimination effect of illusions: Comparative tests based on multiple sets of actual working conditions of independent steel column foundations showed that indexing illusions occurred frequently when RAG was used alone, logical errors were obvious when knowledge graph reasoning was used alone, and computational illusions still occurred when heterogeneous decoupling architecture was used alone. However, when all three worked synergistically through the fourth-order illusion elimination algorithm (PAFO-HEA), both indexing and computational illusions were suppressed to extremely low levels, greatly reducing the logical error rate. This result indicates that the second order of PAFO-HEA (dual-path heterogeneous verification and logical anchoring) enables RAG and knowledge graph to mutually correct each other.

[0019] The determinism determined by the architecture surpasses the upper limit of prompt engineering: Public data in the industry shows that the calculation deviation rate of general large models in steel structure cascade verification is generally between 2% and 10%. This invention completely eliminates the possibility of LLM directly interfering with the numerical values ​​from the architecture through the mandatory physical isolation of the third-order (deterministic numerical calculation based on physical isolation). This is a qualitative change at the architectural level, and its effect is better than prompt word optimization.

[0020] Nonlinear efficiency improvement surpasses conventional DAG optimization: In the typical scenario of "recalculating after modifying the base height," general systems need to recalculate a large number of redundant steps. While conventional DAG optimization offers some improvement, its scope remains relatively broad. This invention, however, precisely limits the recalculation scope to an extremely small core subset through fourth-order dependency graph backtracking and incremental correction. This efficiency improvement stems entirely from the fourth-order illusion resolution algorithm (PAFO-HEA)'s hypergraph modeling and precise backtracking of computational dependencies.

[0021] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0022] Figure 1 This is the overall architecture diagram of the steel structure calculation system based on physical logic anchoring and heterogeneous decoupling provided by the present invention; Figure 2 This is a flowchart of the steel structure calculation method based on physical logic anchoring and heterogeneous decoupling provided by the present invention; Figure 3 This is the multi-stage parameter-guided state machine diagram provided by the present invention; Figure 4 This is the flowchart of the second-order dual-path heterogeneous source verification and logic anchoring provided by the present invention; Figure 5 This invention provides a computational logic dependency hypergraph and a backtracking correction graph. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings, so as to more clearly understand the purpose, features and advantages of this invention. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of this invention, but are only for illustrating the essential spirit of the technical solutions of this invention. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0024] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0025] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0026] The singular forms “a” and “the” used in this specification and the appended claims include plural references unless otherwise expressly stated herein. It should be noted that the term “or” is generally used to mean “and / or” unless otherwise expressly stated herein.

[0027] In the following description, in order to clearly demonstrate the structure and working method of the present invention, a number of directional terms will be used. However, terms such as "front", "back", "left", "right", "outside", "inside", "outward", "inward", "up", and "down" should be understood as convenient terms and not as limiting terms.

[0028] The implementation details of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following content is only for the convenience of understanding the implementation details and is not necessary for implementing this solution.

[0029] Figure 1 This diagram illustrates the architecture of a steel structure calculation system based on physical logic anchoring and heterogeneous decoupling, as provided by this invention. The system's execution entity is a full-process interactive system for steel structure engineering calculations based on a large language model. It adopts a heterogeneous decoupling architecture and specifically includes the following core modules: 1. Semantic Awareness Module This module, serving as the system's input front-end, is responsible for receiving and processing unstructured engineering requirements input by users in natural language. Internally, it relies on a semantic central knowledge graph, which constructs a four-dimensional entity association network of "scene-specification-parameter-formula" to understand the physical ontology information contained in the user's intent, such as structural topology, physical conditions, and mechanical calculations, and extracts it as structured feature vectors. Simultaneously, this module connects to a condition semantic path library, which uses a directed graph structure to predefine the standard logical topology for steel structure calculations, forcibly constraining and guiding ambiguous construction intentions onto physical logical paths that conform to mechanical axioms. In the fourth-order illusion resolution algorithm (PAFO-HEA), this module corresponds to the first-order (physical topology anchoring) front-end semantic processing stage, fundamentally solving the mapping distortion problem in the process of mapping unstructured requirements to the physical logical starting point.

[0030] 2. Logical Scheduling Module This module is the core control center of the system, responsible for driving and monitoring the execution of the physical logic-anchored fourth-order illusion resolution algorithm (PAFO-HEA) throughout the entire process. Internally, it integrates two core resources: a standardized vectorized source tracing unit and a semantic central knowledge graph, and externally connects to the working condition semantic path library and the operator execution module. The specific scheduling and control responsibilities of this module in the four phases of PAFO-HEA are as follows: In the first stage (physical topology anchoring): the scheduling condition semantic path library and semantic central knowledge graph force the structured feature vectors output by the semantic perception module to be constrained onto a standardized directed acyclic computation graph, thus completing the mapping from fuzzy construction intentions to deterministic physical topology paths.

[0031] In the second stage (dual-path heterogeneous verification and logical anchoring): the standardized vectorized tracing unit and the semantic central knowledge graph are scheduled, and deterministic reasoning of the knowledge graph (path A) and RAG semantic retrieval (path B) are launched in parallel. Through cross-validation of the results of the two paths and calculation of logical anchoring coefficients, the "index illusion" is eliminated and compliance auditing is completed. Among them, the standardized vectorized tracing unit realizes the bijective association between the standard provisions and the calculation scenario based on high-dimensional vector space mapping, providing a comparison basis for retrieval; the semantic central knowledge graph uses its logical reasoning capabilities to perform knowledge-driven cross-validation of the vector retrieval results.

[0032] In the third stage (deterministic numerical computation based on physical isolation): the control front-end gateway sequentially performs instruction whitelist filtering, topological logic isomorphism auditing, and coefficient domain boundary hard constraint injection triple auditing on the computation instructions to be issued; after the audit is passed, the structured parameter package is scheduled to the operator execution module through a one-way logic injection protocol, and the latter completes the deterministic solution in the physically isolated environment.

[0033] In the fourth stage (dependency graph backtracking and incremental correction): the system receives objection instructions from users regarding specific calculation results, executes a reverse lookup mechanism through the logical dependency directed hypergraph maintained by the system, locates the smallest subset of operators that have a hyperedge dependency relationship with the objection node, triggers local incremental recalculation, and forces the recalculation results back to the second and third stages to complete the complete verification and auditing process.

[0034] 3. Operator Execution Module This module is the physically isolated execution area in the system responsible for deterministic numerical computation. In the fourth-order illusion resolution algorithm (PAFO-HEA), it corresponds to the third-order (physically isolated deterministic numerical computation) isolated execution stage. Its core is a physical operator execution container, which achieves process isolation at the operating system level. Internally, it encapsulates a deterministic engineering function library that strictly adheres to the "Steel Structure Design Standard." It communicates with upper-level modules only through a one-way encrypted channel, receiving structured parameter packets that have passed triple auditing. If the input parameter dimensions do not match, a physical dimension conflict circuit is triggered, refusing execution. The container's built-in dynamic constraint verification engine establishes multi-parameter collaborative constraint equations in real time during computation, performing dynamic boundary checks on intermediate results and the final output. Through process isolation, one-way communication, dimension verification, and dynamic constraint verification—a four-fold mechanism—this module completely deprives large language models of the authority to interfere with numerical computation at the architectural level (computational sovereignty stripping), fundamentally eliminating the "computational illusion" and ensuring that numerical results achieve industrial-grade deterministic accuracy.

[0035] During system operation, the semantic perception module transforms the user's unstructured engineering requirements into structured feature vectors. The logic scheduling module invokes the fourth-order illusion resolution algorithm (PAFO-HEA) to drive the semantic central knowledge graph and the standardized vectorized tracing unit to perform dual-path heterogeneous verification and complete triple auditing of the computation instructions to be executed. The audited deterministic computation instructions are sent to the operator execution module via a one-way encrypted pipeline. The operator execution module drives the dynamic constraint verification engine within a physically isolated container to complete the symbolic numerical solution and sends the result back to the logic scheduling module. In the entire computation process, the highly coupled and logically-driven control and scheduling tasks are all handled by the logic scheduling module. The operator execution module is only responsible for deterministic execution under controlled conditions, while the generative large language model serves only as an auxiliary tool for semantic understanding and interactive guidance, achieving true heterogeneous decoupling.

[0036] Based on the above system embodiments, this invention proposes a fourth-order illusion resolution algorithm (PAFO-HEA) based on physical logic anchoring. Figure 2 A flowchart of a steel structure calculation method based on physical logic anchoring and heterogeneous decoupling is shown. The method specifically includes the following steps: First-order: Physical topology anchoring; This phase aims to address the semantic distortion that arises when mapping unstructured engineering requirements to structured, standardized computational models, ensuring the physical accuracy of the computational logic's starting point. Specifically, it includes the following steps: S101, Semantic Feature Extraction and Intent Tensorization The system receives a natural language request from a user regarding a steel structure calculation scenario; it then performs feature extraction on the natural language request using a pre-built large language model specific to the steel structure engineering field; based on the feature extraction, it maps the natural language request into a structured feature vector V. i The structured feature vector V i It includes structural topology dimension features, physical condition dimension features, and mechanical verification dimension features parsed from the natural language request.

[0037] In this embodiment, the extraction of the structured feature vector is achieved using a sequence labeling scheme based on the pre-trained language model BERT-base-Chinese and Conditional Random Field (CRF). The output label set is defined as a steel structure feature slot system (e.g., B-LOAD represents load type, I-GEO_VAL represents geometric dimension value, etc.). The model is trained from 100,000 manually labeled construction requirement corpora. During inference, the labeling results are parsed into feature vectors containing category, value, and unit, which constitutes the structured feature vector, thereby accurately mapping unstructured natural language requirements to the starting point of structured computation.

[0038] S102. Active detection and completion based on feature slots The structured feature vector V i The system matches the data with a steel structure scenario rule base to automatically scan and identify missing key mechanical constraint feature slots in the current calculation scenario. These key mechanical constraint feature slots include, but are not limited to: load parameters (such as the type and magnitude of basic forces like axial force, bending moment, and horizontal force), geometric parameters (such as foundation bottom dimensions, foundation height, and component cross-sectional dimensions), material parameters (such as concrete strength grade, steel reinforcement grade, and steel yield strength), and boundary conditions such as seismic intensity, site category, and environmental category. In response to the detection of missing key mechanical constraint feature slots, the system triggers completion of the structured feature vector V at the front-end human-computer interaction interface based on a heuristic search algorithm. i In the process, up to the structured feature vector active parameter detection protocol; through the active parameter detection protocol, the user is guided to sequentially complete the parameter input for the load and boundary stage, the geometry and section stage, and the material and performance stage, and the feature data V input at each stage is then processed. i The included feature dimensions satisfy the requirements of mechanical logical integrity. The state transition relationships of the multi-stage sequential guidance process are described in [reference needed]. Figure 3 .

[0039] S103, Logical path integrity verification The aligned and completed structured feature vector V i The input is fed into a steel structure knowledge graph; reasoning is performed along the preset "scenario-verification-constraint-formula" physical topology chain in the steel structure knowledge graph in order to attempt to construct and output a directed acyclic computation graph with strong logical consistency. The directed acyclic computation graph The system defines all the mechanical boundary conditions that must be satisfied in the current computing scenario and the corresponding sequence of computing operators. If a complete computing path from the input to the final value cannot be constructed during the reasoning process, the system determines that the logical path integrity check has failed, terminates the subsequent operation, and refuses to activate the next-order processing flow.

[0040] The ontology of the knowledge graph defines six core entities and their relationships (e.g., "load" - [trigger verification] -> "verification item"). The physical topology chain reasoning adopts a reasoning engine based on forward chain rules. Starting from the "component" and "load" nodes anchored by the intent tensor, the verification chain is triggered and expanded through preset rules. Finally, a directed acyclic computation graph is extracted and formed by graph traversal algorithm. If the key parameter of any triggered verification item has a completely missing value path in the graph, the path integrity verification is judged to have failed.

[0041] Second-order: Dual-path heterogeneous source verification and logical anchoring based on reliability envelope; This phase aims to eliminate knowledge illusions at the index level and ensure the physical authenticity and legal accuracy of the normative provisions referenced in the computation through cross-modal alignment and mutual verification of deterministic graph reasoning paths and probabilistic semantic retrieval paths. Figure 4 As shown, the method includes the following steps: S201, Dual-path parallel evidence collection; To achieve the aforementioned cross-modal alignment and mutual verification, the system initiates two evidence collection paths in parallel: Path A is deterministic graph reasoning, which extracts the statutory reasoning chain and canonical identifiers along the first-order physical topology chain; Path B is probabilistic semantic retrieval, which uses the RAG engine to complete nearest neighbor matching and canonical number parsing in the vector space.

[0042] Directed acyclic computation graph based on the first-order output Two independent evidence extraction processes are executed in parallel: Path A - Physical Topology Anchoring Evidence: From the steel structure knowledge graph, along the aforementioned directed acyclic computation graph Extracting the set of deterministic logical path nodes This set of nodes This represents a legally established chain of reasoning that conforms to the axioms of mechanics; wherein, the set of nodes... Each node in the system is pre-loaded with a legally defined, canonical, unique identifier. The canonical unique identifier Used to uniquely identify the specific clause number of the steel structure design code on which it is based.

[0043] Path B-Semantic Space Mapping Forensics: The RAG engine is invoked to perform nearest neighbor search in the vectorized canonical library to obtain a set of canonical texts relevant to the current computing scenario. The set of standard original texts Parsed into corresponding semantic graph nodes And parse and extract each of the semantic graph nodes. The corresponding standard index number .

[0044] S202, Calculate the logic anchoring coefficient Λ of the physical drive; The logical anchoring coefficient Λ is constructed as the verification criterion, and its calculation formula is as follows: (1) Numbering and version consistency judgment operator If the specification number of the knowledge graph extraction is required to be completely consistent with the number and version number of the RAG retrieval and parsing, then... If the large model incorrectly references an obsolete version (such as the "Steel Structure Design Standard" GB 50017-2003 instead of the current GB50017-2017), or if the numbering is inconsistent, then... ,lead to This triggers subsequent forced coverage and computation graph reconstruction. This operator embodies the compliance principle of "one-vote veto" in engineering.

[0045] (2) Standardize the dynamic weight of timeliness The empowerment will be implemented in a tiered manner based on the legal validity of the regulations. This applies to currently valid regulations (latest version). For standards that are in transition (new standards have been issued but old standards have not been completely repealed); ( The specific value is dynamically calculated by the system's built-in standardized timeliness management table based on the proportion of the remaining transition period, for example... For abolished regulations, A value of 0 also triggers a forced override. This weight ensures that the system can automatically identify and reject invalid specifications.

[0046] (3) Strong penalty items A non-linear penalty mechanism is established based on the level of enforceability of the clauses to be verified. When mandatory clauses ("forced clauses") are involved, the system will deduct penalties even if the semantic similarity is high. (0.3~0.5), ensuring that minor shifts in the semantics of strong clauses are intercepted; for recommended or general clauses, deduct smaller values. Scenarios with full semantic alignment or non-canonical constraints are not deducted. This mechanism implements the "strong clause zero tolerance" principle in engineering.

[0047] (4) Semantic similarity function The value ranges from 0 to 1, and measures the set of logical nodes in path A. The original text of path B The degree of alignment in the vector space. Specifically, this is achieved by: […]. Concatenate them into a logical description text. The two texts are integrated into a standardized semantic text. Then, a pre-trained Chinese text embedding model is used to map them into vectors of the same dimension, and the cosine similarity is calculated.

[0048] S203, Dynamic threshold determination and logical backtracking correction; The logical anchoring coefficient Λ calculated in step S202 is compared with the dynamic reliability threshold. The comparison is performed, and the following judgments and corrections are made: The system presets a dynamic reliability threshold. This threshold is used to determine the reliability of the dual-path heterogeneous source verification results. It is determined using the following engineering method: (a) Constructing a test suite for obfuscating specification clauses Representative clauses were extracted from core standards such as the Steel Structure Design Standard GB 50017 and the Seismic Design Code for Buildings GB 50011. Typical noise was artificially introduced (such as tampering with parameter values, replacing component names, and fabricating standard numbers) to form paired samples of semantically correct and semantically incorrect samples.

[0049] (b) Semantic similarity distribution test The cosine similarity between the outputs of path A (knowledge graph reasoning) and path B (RAG retrieval) is calculated using a pre-trained text embedding model. The similarity distribution interval between correct and incorrect samples is statistically analyzed.

[0050] (c) Critical threshold optimization With "zero omissions in mandatory provisions" as a hard constraint, in The optimal decision boundary is searched within the range. Tests show that when the threshold is lower than... At certain times, some implicit semantic errors (such as confusing "seismic resistance level" with "seismic structural measures") cannot be intercepted; when the threshold is higher than 0.9, minor literal adjustments to the specifications that are semantically correct are easily misjudged as errors. Therefore, the preferred interval is determined to be... .

[0051] In a specific embodiment of the present invention, taking This serves as the threshold for persistent dynamic reliability. Users can also manually adjust it within the preferred range based on the project's stringency.

[0052] Based on the logical anchoring coefficient and The system performs three types of actions based on the comparison. Release: If If the current logical path is determined to have physical authenticity and the index is correct, then it is allowed to proceed to the third-order numerical calculation. Manual adjudication: If If a hidden compliance risk is identified (such as a slight deviation from a mandatory clause, or insufficient semantic matching when referencing a specification), the system will intercept the current process and generate a report, requiring manual confirmation from engineers. Forced correction: If ,(Right now or If the result after semantic deduction is negative, then an index-level fatal error is determined to exist (incorrect numbering, referencing obsolete specifications, etc.). The system will automatically extract the legally valid standard index number from path B. Compared with the original text of the standard This involves forcibly overwriting the error numbers and corresponding descriptions in the output layer of the large language model, reconstructing the directed acyclic computation graph, and re-executing dual-path forensics until... .

[0053] Third-order: Deterministic numerical computation based on physical isolation; This phase aims to physically isolate the semantic orchestration function of the large language model from rigorous numerical symbolic computation, and to strongly constrain the data interaction between the two through a "computational power gate" with multi-level supervision mechanisms, thereby fundamentally eliminating the computational illusion at the architectural level. The specific implementation steps of this method are as follows: S301: Execute multidimensional instruction auditing and physical logic alignment; Any arithmetic logic or script generated by a large language model and intended to be executed at the numerical computation layer must first undergo triple audit verification by the front-end gateway to ensure the correctness and compliance of its physical logic.

[0054] a) Whitelist filtering of engineering operator instruction sets The gateway has a pre-defined whitelist of engineering operator instructions. The gateway performs static code feature scanning on the output instructions of the large language model, allowing passage only when an instruction precisely matches a predefined engineering operator in the whitelist (e.g., strength verification of axially loaded members, bending resistance verification of flexural members, shear resistance verification of shear members, section stress verification, stability coefficient calculation, weld strength verification, bolt connection bearing capacity verification, etc.). Any unrestricted instructions, such as low-level system calls or non-predefined mathematical functions, will be instantly flagged as errors and intercepted, thus establishing an unbypassable execution barrier at the instruction level.

[0055] b) Topological logical isomorphism audit An abstract syntax tree mapping table for "standard formulas," generated from a knowledge graph, with one tree corresponding to each standard formula. For example: The formula for axial compression stability verification is as follows: Its syntax tree root node is "≤", the left subtree is the axial force N, and the right subtree is the product node. The child nodes of the product node are the stability coefficient φ, the cross-sectional area A, and the material strength f, respectively.

[0056] The gateway parses the numerical expression to be executed into a syntax tree, and then compares it with the standard syntax tree in the corresponding computing scenario. The rules are: the root node, child nodes, and internal operator types of the two are consistent, and the number of leaf nodes must be equal (for example, when there are constant terms in the expression). This is to prevent the large language model from tampering with the operators or the order of operations in the formula during the generation process.

[0057] c) Hard constraint injection at the coefficient domain boundary The legally defined value ranges of relevant parameters are dynamically extracted from the original specifications of the current computing scenario and the company's historical engineering experience documents. For example: For steel strength parameters (such as tensile and compressive strength of steel designers) Shear strength design value Design value of end face bearing strength Based on the appendix to the "Standard for Design of Steel Structures" GB 50017, the minimum and maximum allowable values ​​are extracted.

[0058] For connection calculation parameters (such as fillet weld leg size) High-strength bolt preload and anti-slip coefficient The legal construction range for the weld leg size of commonly used fillet welds in steel structures is [4, 12] mm; the legal range for the anti-slip coefficient of 8.8 grade friction type high-strength bolts is [0.35, 0.50].

[0059] The system treats these intervals as runtime environment variables and forcibly injects them into the downstream isolated computing containers, thereby fundamentally blocking the possibility of data drift during the execution phase.

[0060] S302: Physically isolated execution based on sandbox communication protocol; The computational instructions audited by S301 and the normalized parameter data packet packaged for Phase 1 survival are pushed through a one-way encrypted pipe to a completely physically isolated process-level sandbox for execution. This sandbox encapsulates a deterministic engineering function library that conforms to the "Steel Structure Design Manual," while large language modules run in a separate process space, with no direct memory sharing between the two.

[0061] (1) One-way communication mechanism Taking "steel column base plate thickness verification" as an example: The first-stage generated data package includes normalized parameters such as column axial force N=1200kN, column cross-section HM300×200, base plate dimensions 500×400×24mm, anchor bolt arrangement 4M24, steel Q355, and concrete C30. After S301 auditing is passed, the entry operator is determined to be base_plate_check. The system packages the operator name with the data package and pushes it to the sandbox through a one-way encrypted pipeline. After receiving it, the sandbox calls base_plate_check (the custom steel column base plate thickness comprehensive verification function of this invention, which internally calls bending, shear, and compression verification sub-functions in sequence) to perform the verification. Throughout the process, the sandbox does not return any intermediate calculation steps to the large language model; the large language model only receives the final conclusion (such as "24mm base plate meets the requirements") and the standard clause number on which it is based. This mechanism ensures that the large language model cannot interfere with the numerical calculation.

[0062] (2) State machine branch decision In the same example, the ingress operator `base_plate_check()` within the sandbox has a built-in state machine branch decision logic: First, it performs a base plate stiffness analysis, calculating the intermediate state variable "base plate constraint type" (cantilever plate, one-sided supported plate, two-sided supported plate, three-sided supported plate, four-sided supported plate) based on the column cross-section, base plate dimensions, and anchor bolt positions. The system automatically triggers state machine transitions based on this variable according to preset hard-coded physical logic. If it is a "cantilever plate" or "one-sided supported plate", then jump to the bending check function bending_check() (the bending check function defined in this invention). If it is a "two-sided support plate" or a "three-sided support plate", then jump to the bending verification and continue to execute the shear verification shear_check() (the shear verification function defined in this invention); If it is a "four-sided support plate", then only the pressure verification function bearing_check() (the pressure verification function defined in this invention) is executed.

[0063] After each operator completes its execution, it may output a new intermediate state, further influencing subsequent branches. The entire decision-making process is completed within a sandbox based entirely on physical feedback, and large language models cannot intervene. This mechanism solves the "selectivity illusion" problem that may occur when large language models face multi-branch decisions.

[0064] The entire computation and branch decision-making process is completed within the sandbox. The large language model ultimately only receives the final numerical calculation results for user-oriented natural language copywriting assembly.

[0065] Fourth stage: computational backtracking and incremental correction based on logical dependency hypergraph.

[0066] This stage constructs a logically dependent directed hypergraph and uses the reverse lookup mechanism of this hypergraph to achieve accurate source tracing, local backtracking, and incremental correction of the computation process. Figure 5 This invention provides a computational logic dependency hypergraph and backtracking correction graph, using column base plate verification as an example. When a user raises an objection to a calculation result (such as base plate thickness) and modifies the parameters, the system uses the result node as a starting point, searches backward along the dependency relationship for all affected nodes, forms an influence range subgraph, and performs local incremental recalculation and compliance verification only on this subgraph, thereby avoiding full recalculation and achieving efficient correction. The specific implementation steps of this method are as follows: S401: Construct a directed hypergraph of logical dependencies; During the computation, a directed hypergraph H is maintained synchronously to record the dependencies between computation steps.

[0067] The nodes of the directed hypergraph H include: input parameters, intermediate calculation results, physical operators, and specification clauses. Hyperedges are used to describe the relationship between an operator and multiple input nodes (for example, if the calculation of the base plate thickness depends on both the base plate bending moment and the steel strength, a hyperedge connects the operator to these two parameters). This hypergraph is constructed in real time during the calculation process. Through the hypergraph structure, the system achieves explicit and structured modeling of the logical dependencies of the entire calculation process, laying a data foundation for subsequent accurate backtracking.

[0068] S402: Semantic anchoring of objection points and extraction of dependency subgraphs; When a user raises an objection or requests a modification to a specific result item in the calculation report, such as the base plate thickness, the system performs the following steps: a) Semantic anchoring of objection points The system first anchors the user's objection semantics to a specific node in the directed hypergraph H through Phase 1, marking it as the target node. .

[0069] b) Sensitivity analysis based on reverse lookup mechanism by Starting from the superedge, trace back in reverse to find all direct or indirect influences. The preceding nodes, i.e., all those that are The tracing process continues until it reaches an input parameter node that can no longer be traced, depending on the parameters or intermediate results. All tagged nodes are associated with... Together they form a dependency subgraph . The scope of calculations that need to be re-evaluated has been precisely defined.

[0070] S403: Local parameter compensation and incremental update execution.

[0071] When a user modifies an input parameter based on feedback, the system performs an incremental update, avoiding a full recalculation. For example, in the calculation of the thickness of the steel column base plate, the user originally input the steel strength as Q235, and the calculated base plate thickness was 24mm. The user thought this was too large and changed the steel strength to Q355, hoping to obtain a thinner base plate.

[0072] a) Parameter update The system uses the user-modified parameter value (changing the steel strength from Q235 to Q355) as the forced bias vector to update the corresponding input parameter node in the directed hypergraph H. All other nodes in the directed hypergraph H that are not located within the dependent subgraph, along with their state data, remain persistent and are not reset by this update operation.

[0073] b) Incremental recalculation within a limited scope The system only works on dependency subgraphs The entire set of nodes (i.e., nodes affected by steel strength: base plate bending moment, base plate thickness, final conclusion, etc.) is recalculated and its state is updated. The calculation process proceeds step by step along the defined directed topological order until the updated values ​​are obtained (e.g., the base plate thickness becomes 20mm, which meets the requirements).

[0074] c) Compliance verification of the updated results Dependency Subgraph After all nodes have been recalculated, the updated calculation results and their associated calculation paths will be forcibly rerouted to the second-order dual-path heterogeneous verification and logical anchoring module for compliance verification (i.e., determined by the logical anchoring coefficient Λ), and then passed through the third-order multi-dimensional instruction audit gateway for compliance review of topology logic and coefficient domain boundaries. Only when the updated results have completely passed the above two-order verification will the system recognize them as valid results and refresh the calculation report on the user side.

[0075] After completing the partial incremental recalculation and compliance verification of S403, the system automatically triggers the standardized compliance report generation process: all finally confirmed calculation results are bidirectionally anchored to the corresponding original text of the specification clauses, the complete calculation path sequence is encapsulated, and a tamper-proof digital fingerprint is generated using a hash digest algorithm. The output is a standardized compliance calculation report that can be downloaded, archived, and reviewed for engineering projects. This report meets the industry compliance physical and logical gateway requirements for steel structure engineering, ensuring traceability, auditability, and tamper-proof calculation results.

[0076] Through the above mechanism, the scope of recalculation in typical scenarios is precisely reduced from dozens of steps in the entire process to only a few steps that are affected. Under the premise of ensuring strict compliance of the results, an order-of-magnitude improvement in recalculation efficiency is achieved, effectively solving the technical problems of state discontinuity and full recalculation.

[0077] To verify the practical technical effect of the physical logic anchored fourth-order illusion resolution algorithm (PAFO-HEA) and its system proposed in this invention, comparative tests were conducted based on typical steel structure calculation conditions. The results fully demonstrate the advanced nature and unexpected technical effects of this invention.

[0078] Regarding the core issue of illusion elimination, this invention achieves a dual breakthrough in eliminating computational and indexing illusions at both the architectural and logical levels. Existing logical reasoning models in the industry commonly exhibit significant computational illusions in steel structure cascade verification, with indexing illusions (fictitious or misused clauses) being particularly prominent. This invention, through the physical isolation of a heterogeneous decoupling architecture and a dual-path heterogeneous verification mechanism, controls the computational deviation rate of cascade verification to an extremely low level and significantly suppresses the frequency of indexing illusions, achieving automatic identification and complete interception of mandatory national clauses.

[0079] In terms of the synergistic qualitative leap in comprehensive technical effects, the fourth-order illusion resolution algorithm (PAFO-HEA) of this invention produces unexpected technical effects that cannot be achieved by any single technology or its conventional combination. When RAG is used alone, the indexing illusion problem is quite prominent; when knowledge graph reasoning is used alone, the logical error rate is high; and when the heterogeneous decoupled architecture is used alone, computational illusions still occur. However, when all three work collaboratively through the fourth-order mechanism of the fourth-order illusion resolution algorithm (PAFO-HEA), both indexing and computational illusions are suppressed to extremely low levels, and the logical error rate is greatly reduced. This result confirms that the second order of PAFO-HEA (dual-path heterogeneous verification and logical anchoring) enables RAG and knowledge graph to mutually correct each other.

[0080] In terms of computational efficiency, this invention significantly reduces the overall process time compared to traditional methods through a multi-stage logic state machine and feature slot detection mechanism. Particularly noteworthy is the non-linear efficiency improvement brought about by the fourth-order dependency graph backtracking and incremental correction mechanism: In typical scenarios where recalculation is performed after modifying the base height, traditional solutions require recalculating numerous redundant steps. While conventional DAG optimization offers some improvement, its scope remains relatively broad. This invention, however, precisely limits the recalculation scope to the smallest core subset directly related to the modified parameters, resulting in a significant improvement in secondary recalculation efficiency compared to traditional full recalculation methods.

[0081] In summary, this invention transforms fuzzy requirements into a structured computational graph through first-order physical topology anchoring; eliminates specification index illusions through second-order dual-path heterogeneous verification and logical anchoring; eliminates computational illusions through third-order physical isolation execution; and achieves efficient solution correction through fourth-order dependency graph backtracking and incremental recalculation. This invention simultaneously solves the three major problems of semantic mapping distortion, index illusions, and computational illusions in intelligent computing for steel structures at the architectural level, achieving traceable verification of specification clauses and deterministic execution of numerical calculations.

[0082] It should be noted that this system can run on one or more servers that include processors and memory. The memory stores computer programs that implement the functions of the modules described above; when these programs are executed by the processor, they implement the aforementioned method steps.

[0083] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the invention by those skilled in the art without departing from the spirit and essence of the invention. Such modifications or substitutions should all fall within the scope of the invention, or any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the invention should be covered within the protection scope of the invention. Therefore, the protection scope of the invention should be determined by the scope of the claims.

Claims

1. A steel structure calculation method based on physical logic anchoring and heterogeneous decoupling, characterized in that, The method includes: S1. The structured feature vectors extracted from the user's unstructured engineering requirements are used to perform physical topological chain reasoning along the preset knowledge graph to construct a directed acyclic computation graph; S2. Two evidence collection paths are initiated in parallel. Path A extracts nodes representing the legal reasoning chain and their canonical identifiers from the preset knowledge graph along the computation graph. Path B calls the retrieval enhancement generation engine to retrieve the original canonical text from the vectorized canonical library and parse its standard index number. The logical anchoring coefficient Λ is calculated, and this coefficient is specifically defined as: ； in, For the consistency determination operator of number and version, To standardize unique identifiers, For standard index number; To standardize the dynamic weighting of timeliness; These are penalties set according to the level of enforceability of the provisions. For semantic similarity function, For logical description text, To standardize semantic text; Λ is compared with a dynamic reliability threshold. Compare and execute the three-branch decision: if Λ≥ Then proceed to subsequent processing if 0 < Λ < If Λ≤0, then the error output will be overwritten with the canonical information of path B and the computation graph will be reconstructed and re-verified until convergence. S3. For the computation instructions to be issued, execute the instruction whitelist filtering, topological logic isomorphism audit, and coefficient domain boundary hard constraint injection triple audit in sequence; the audited instructions are routed to the physically isolated operator execution environment through a one-way encrypted pipeline. Under the condition of depriving the large language model of the right to intervene in numerical operations within the process-level sandbox, the deterministic numerical solution is executed through the dynamic constraint verification engine. S4. Anchor user objections to the calculation results to the target node in the pre-maintained logical dependency directed hypergraph, and trace back along the hyperedge to extract the dependency subgraph; inject modified parameters only into the dependency subgraph and perform local incremental recalculation; force the recalculation results to S2 and S3 for compliance verification, and output a compliance calculation report after the verification is passed.

2. The method according to claim 1, characterized in that, Step S1 includes: The structural topology dimension features, physical working condition dimension features, and mechanical verification dimension features parsed from the user's unstructured engineering requirements are mapped to the structured feature vector; The structured feature vector is matched with the steel structure scene rule base to identify missing key mechanical constraint feature slots, and an active parameter detection protocol is triggered to guide the user to complete the parameters in the order of load and boundary stage, geometry and section stage, and material and performance stage. Reasoning is performed along the physical topology chain of the preset knowledge graph. If a complete computation path from the input to the final value cannot be constructed, the logical path integrity check is determined to have failed and subsequent operations are terminated.

3. The method according to claim 1, characterized in that, The parallel initiation of two evidence collection paths in step S2 specifically includes: Path A involves extracting a set of deterministic logical path nodes from the knowledge graph along the directed acyclic computation graph. The set of nodes Each node in the code is pre-loaded with a unique identifier that points to the specific clause number of the steel structure design code it is based on. Path A extracts nodes representing legal reasoning chains and their canonical identifiers from the knowledge graph along the computation graph; Path B invokes the retrieval enhancement generation engine to perform nearest neighbor search in the vectorized canonical library, obtaining a set of canonical texts relevant to the current computing scenario. The set of standard original texts Parsed into corresponding semantic graph nodes And parse and extract the standard index number corresponding to each semantic graph node. .

4. The method according to claim 1, characterized in that, In step S2, the formula for calculating the logical anchoring coefficient Λ is as follows: (a) Numbering and version consistency judgment operator Defined as: ； (b) Standardize the dynamic weight of timeliness The release status value is determined according to the specification: ； (c) Penalties Based on the level of enforceability of the provisions: ； in, The value ranges from 0.3 to 0.

5. Take 0.1; (d) Semantic similarity function The value range is 0 to 1.

5. The method according to claim 1, characterized in that, The triple audit described in step S3 specifically includes: Instruction whitelist filtering: Static code feature scanning is performed on the calculation instructions to be issued by using a pre-defined whitelist of engineering operator instructions. Instructions are allowed to pass only if they exactly match the predefined engineering operators in the whitelist. Unrestricted instructions are blocked immediately. Topological logic isomorphism audit: The numerical operation expression to be executed is parsed into an abstract syntax tree, and the structure of the abstract syntax tree is compared with the canonical arithmetic logic topology defined by the corresponding calculation formula in the knowledge graph to verify whether it strictly adheres to the physical formula structure specified in the specification. Hard constraint injection at coefficient domain boundaries: By retrieving and enhancing the generation engine, the statutory value range of parameters related to the current calculation is dynamically extracted from the original specification and the statutory value range is forcibly injected into the downstream isolated computing container as a runtime environment variable.

6. The method according to claim 1, characterized in that, In step S3, the physically isolated operator execution environment is a process-level sandbox running independently at the operating system level. It encapsulates a deterministic engineering function library that conforms to steel structure design standards. There is no direct memory sharing between it and the process space running the large language model; they communicate only through a one-way encrypted channel. When the operator execution environment receives data packets, it first performs physical dimension verification. If the dimensions of the input parameters do not match the operator requirements, a physical dimension conflict circuit breaker is triggered, and execution is refused. Within the process-level sandbox, the engineering decision logic is hard-coded in the isolated environment. Intermediate state variables are parsed based on the preset hard-coded physical logic, and state machine transitions are automatically triggered to complete operator switching. The large language model only receives the final calculation result path without branch decision information.

7. The method according to claim 1, characterized in that, The logical dependency directed hypergraph described in step S4 is dynamically constructed during system computation: whenever a physical operator is scheduled to be executed, the system creates the corresponding operator node and establishes directed hyperedges from all input parameter nodes and preceding intermediate result nodes that the operator depends on to the operator node. The step of anchoring user objections to the calculation results to the target node and extracting the dependency subgraph includes: anchoring the semantics of user objections to the target node in the directed hypergraph through natural language understanding. ;by Starting from the hyperedge, trace back in reverse and use the reverse lookup mechanism algorithm to mark layer by layer. The process continues until all dependent nodes are reached, up to the input parameter node that can no longer be traced. Connect all marked dependency nodes with They are jointly identified as dependency subgraphs .

8. The method according to claim 1, characterized in that, The local incremental recalculation and compliance verification mentioned in step S4 specifically include: The user-modified parameter values ​​are used as a forced bias vector, which is then injected into and used to update the directed hypergraph. The values ​​of the corresponding input parameter nodes are not located in Other nodes and their state data within the system remain persistent and unchanged; Only for The entire set of nodes is recalculated and its state updated progressively along the defined directed topological order until... Get the updated value; The updated calculation results and their associated calculation paths are forcibly routed to the dual-path heterogeneous source verification and logical anchoring verification corresponding to S2, and the multi-dimensional instruction audit and compliance review corresponding to S3. The updated results are only refreshed to the user side after all the above verifications have been passed.

9. A steel structure calculation system based on physical logic anchoring and heterogeneous decoupling, characterized in that, The system for implementing the method as described in any one of claims 1 to 8 comprises: The semantic awareness module extracts structured feature vectors from users' unstructured engineering needs, performs physical topological chain reasoning along a preset knowledge graph, and constructs a directed acyclic computation graph. The logic scheduling module initiates two evidence collection paths in parallel. Path A extracts nodes representing the legal reasoning chain and their canonical identifiers from the knowledge graph along the computation graph. Path B calls the retrieval enhancement generation engine to retrieve the original canonical text from the vectorized canonical library and parse its standard index number. The module also calculates the logical anchoring coefficient Λ, which is specifically defined as follows: ,in, For numbering and version consistency judgment operators; To standardize unique identifiers; For standard index number; To standardize the dynamic weighting of timeliness; These are penalties set according to the level of enforceability of the provisions. For semantic similarity function, For logical description text, To standardize semantic text; Λ is compared with a dynamic reliability threshold. Compare and execute the three-branch decision: if Λ≥ Then proceed to subsequent processing if 0 < Λ < If Λ≤0, the error output is forcibly overwritten with the standard information of path B, and the computation graph is reconstructed and re-verified until convergence. The computation instructions to be issued are sequentially subjected to instruction whitelist filtering, topological logical isomorphism auditing, and coefficient domain boundary hard constraint injection triple auditing. The instructions that pass the audit are routed to the operator execution module through a one-way encrypted pipeline, and the computation results returned by the module are received. In addition, the user's objection to the computation result is anchored to the target node in the pre-maintained logical dependency directed hypergraph. The dependency subgraph is extracted by tracing back along the hyperedge. The modified parameters are injected only into the dependency subgraph and a local incremental recalculation is triggered. The recalculation result is forcibly routed back for compliance verification. The operator execution module, under the condition of depriving the large language model of its right to intervene in numerical operations within the process-level sandbox, performs deterministic numerical solutions on the received audited instructions through the dynamic constraint verification engine, and sends the calculation results back to the logic scheduling module. The logic scheduling module is also used to output a compliance calculation report after the verification is passed.

Images