Method and system for solving large model calculation illusion in complex engineering field

Abstract

The present application relates to the field of artificial intelligence and engineering calculation technology, and specifically provides a method and system for solving the calculation illusion of large models in complex engineering fields, aiming to solve the problems that large models cannot correctly identify the intention of engineers, the reasoning conclusion is prone to error and the conclusion lacks credibility in the prior art. In the present application, the input engineering calculation requirement text is analyzed for intention, and a structured task description is generated; the task description is retrieved in the formula knowledge base, and the calculation formula is selected based on the set threshold; the computer algebra system is called to perform deterministic symbolic calculation to solve; the calculation result is triple-verified, and the verification record, the calculation formula used in the calculation process, the parameters and the calculation result are integrated into an evidence chain report. The present application can accurately identify the intention of engineers, select the correct engineering formula, complete the specific calculation process by the computer algebra system to ensure the reliability of the calculation result, and provide a traceable evidence chain.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and engineering computing technology, specifically relating to a method and system for solving computational illusions in large models in complex engineering fields. Background Technology

[0002] With the development of artificial intelligence and natural language processing technologies, intelligent agents based on large language models (LLM) are gradually being applied in the engineering field. They can help engineers quickly obtain information through natural language interaction and improve the efficiency of design, calculation and decision-making.

[0003] However, since large language models are essentially based on statistical probability rather than strict factual logic, coupled with the limitations of training data and insufficient reasoning ability, for problems requiring long-chain logical reasoning (such as mathematical problems and complex logic problems), large models may make mistakes in the intermediate steps or lose key information, ultimately arriving at a seemingly reasonable but completely wrong conclusion, which is the "big model illusion". Therefore, the following problems will exist: (1) Large models cannot correctly identify the engineer's intention and thus choose the wrong engineering formula; (2) For relatively complex engineering formulas, large models are prone to producing seemingly reasonable but completely wrong conclusions, which is fatal to the engineering field; (3) The conclusions produced by large models are unpredictable, cannot provide effective evidence to support the conclusions, and lack credibility. Summary of the Invention

[0004] To address the problems in the prior art, this invention provides a method and system for solving computational illusions in large models in complex engineering fields. It can accurately identify computational needs, select the correct engineering formulas, ensure the reliability of computational results, and provide a traceable chain of evidence.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] In a first aspect, the present invention provides a method for resolving computational illusions in large models within complex engineering domains, comprising:

[0007] The input engineering calculation requirement text is parsed to generate a structured task description; the task description includes engineering domain tags, calculation objectives, a list of known parameters, and parameter units;

[0008] According to the task description, a search is performed in the constructed formula knowledge base to match applicable candidate formulas, and the candidate formulas are sorted by confidence level to obtain a list of candidate formulas.

[0009] Determine whether the highest confidence level is less than a set threshold; if it is less, a calculation formula is manually selected from the candidate formula list; otherwise, the candidate formula corresponding to the highest confidence level is selected as the calculation formula.

[0010] Based on the calculation formula and the parameters in the known parameter list, the calculation result is obtained by calling a computer algebra system to perform deterministic symbolic calculation;

[0011] The calculation results are subjected to triple verification, including dimensional consistency verification, numerical range rationality verification, and boundary condition verification. If the triple verification fails, the abnormal items are marked and the calculation formula is reselected for iterative calculation and verification until the triple verification passes and a verification record is generated.

[0012] The calculation results that have passed triple verification, the calculation formula used, the standard number referenced by the calculation formula, the parameters used in the calculation formula, and the verification record are integrated into an evidence chain report and returned.

[0013] A further solution: The formula knowledge base stores a set of formulas that have been manually reviewed; all formulas in the formula knowledge base are organized hierarchically according to engineering fields, and each formula record includes the formula expression, applicable working conditions, engineering experience value range, parameter definition domain, dimensional information, and referenced standard number.

[0014] A further approach: The step of matching applicable candidate formulas in the formula knowledge base to obtain a list of candidate formulas sorted by confidence level includes:

[0015] The task description is parsed into search criteria;

[0016] Based on the search criteria, a two-layer recall method combining hard constraint filtering and semantic similarity retrieval is used to retrieve an applicable set of candidate formulas from the formula knowledge base.

[0017] For each candidate formula in the candidate formula set, the matching degree between the candidate formula and the task description is calculated from multiple dimensions. The dimensions include engineering domain matching degree, target physical quantity matching degree, coverage of the formula input parameters by the known parameter list, whether the parameters in the known parameter list fall into the domain of the formula parameters, dimensional consistency, and semantic similarity.

[0018] The matching scores calculated from each dimension are weighted, summed, and normalized to obtain the confidence score of each candidate formula.

[0019] Each candidate formula is sorted according to the confidence level to obtain the candidate formula list.

[0020] Based on the above scheme, by comprehensively evaluating the candidate formulas from multiple dimensions, the suitability of the candidate formulas to the task description can be quantitatively characterized from different levels, avoiding the problem of formula misselection caused by relying on only a single matching condition, and improving the accuracy and reliability of formula matching.

[0021] A further solution: The content of the dimension consistency verification is as follows: Based on the dimension information corresponding to each parameter in the calculation formula, the physical dimensions of the parameters in the known parameter list and the calculation result are derived, and it is determined whether the dimension of the derived output result is consistent with the target dimension corresponding to the calculation target.

[0022] The content of the data range rationality verification is as follows: compare the calculation result with the engineering experience value range of the calculation formula, and compare whether the calculation result is within the reasonable value range of the corresponding engineering parameter;

[0023] The boundary condition verification involves verifying the parameters in the known parameter list and the calculation results based on the parameter domain and applicable working conditions corresponding to the calculation formula. This includes determining whether the parameters in the known parameter list satisfy the domain constraints of the calculation formula, whether the engineering working conditions satisfy the applicable working conditions of the calculation formula, and whether the calculation results exceed the engineering experience value range of the calculation formula.

[0024] Based on the above scheme, the three verifications in this invention are progressive and complementary. The dimension consistency verification focuses on finding incorrect formula selection; the data range rationality verification focuses on finding incorrect parameter or unit input; and the boundary condition verification focuses on finding cases where the value exceeds the defined domain or the premise for the applicable formula is not valid. The three verifications work together to ensure that only calculation results that simultaneously meet the requirements of correct type, reasonable value, and valid premise can be released, effectively preventing erroneous results from flowing into subsequent engineering decision-making stages.

[0025] A further proposed solution is to dynamically adjust the threshold value based on the safety level of the engineering field; the threshold value ranges from 0.7 to 0.95, and the value of the threshold value is proportional to the safety level of the engineering field.

[0026] A further option: The computer algebra system includes a symbolic computation subsystem and third-party professional engineering analysis software;

[0027] The complexity of the computational task corresponding to the aforementioned formula is determined, and a hierarchical solution strategy is adopted based on the task complexity:

[0028] For tasks determined to be simple calculations, the symbolic computation subsystem is invoked to perform the solution and obtain the calculation result;

[0029] For tasks deemed to involve complex calculations, third-party professional engineering analysis software is invoked to perform the solution and obtain the calculation results.

[0030] Based on the above scheme, regardless of whether the symbolic computation subsystem or third-party professional engineering analysis software is called, the present invention performs deterministic calculations based on established formulas and engineering models, rather than using large language models for numerical reasoning. This fundamentally eliminates the computational illusion caused by statistical reasoning while taking into account both computational efficiency and the ability to solve complex engineering problems.

[0031] A further solution includes: when determining whether all triple checks have passed, the following steps are also taken: if any one of the triple checks fails, an anomaly is marked, the parameters are manually corrected, the calculation is performed again based on the corrected parameters to obtain the calculation result, and the triple checks are performed again until all triple checks have passed.

[0032] A further approach involves using a domain-fine-tuned large language model combined with an engineering terminology dictionary to preprocess the input engineering calculation requirement text, extract keywords, perform engineering semantic disambiguation, and map structured slots to generate a structured task description.

[0033] A further solution: The input preprocessing involves: segmenting the engineering calculation requirement text into sentences and words, and performing text normalization to obtain a normalized sequence of text to be parsed;

[0034] The keyword extraction content is as follows: using a dual-channel approach that combines engineering terminology dictionary matching with domain fine-tuning of large language model named entity recognition, multiple named entities are identified from the normalized text sequence to be parsed, including engineering discipline category words, calculation type words, target physical quantity words, known parameter words, and the numerical values ​​and units of parameter words;

[0035] The engineering semantic disambiguation process involves: using a domain-fine-tuned large language model to perform semantic determination on ambiguous named entities, thereby identifying the unique engineering semantics of each named entity;

[0036] The structured slot mapping consists of: pre-defining a structured slot template containing engineering domain labels, calculation targets, a list of known parameters, and parameter units; filling the structured slot template with disambiguated named entities according to semantic type; after filling, performing missing detection on the required slots; if a calculation target or key known parameter is detected to be missing, the missing information is manually supplemented; if all required slots are fully filled, a structured task description is generated.

[0037] Based on the above scheme, the engineering terminology dictionary matching channel can ensure the accuracy of the recognition of the included professional terms, and the domain fine-tuning large language model named entity recognition channel can supplement the recognition of colloquial or non-standard expressions not included in the dictionary. The two channels work in parallel to improve the recall rate of recognition.

[0038] In a second aspect, the present invention provides a system for resolving computational illusions in large models within complex engineering domains, for performing a method for resolving computational illusions in large models within complex engineering domains as described in any of the first aspects; the system includes:

[0039] Interactive module, used for human interaction;

[0040] The intent parsing module is used to receive the engineering calculation requirement text input through the interaction module and perform intent parsing to generate a structured task description; wherein, the task description includes engineering domain tags, calculation objectives, a list of known parameters, and parameter units;

[0041] The formula matching module is configured with a pre-built formula knowledge base; the formula matching module is used to search in the formula knowledge base according to the task description, match applicable candidate formulas and sort them by confidence level, and return a list of candidate formulas; and select the calculation formula from the list of candidate formulas based on a set threshold.

[0042] The symbolic computation engine module is used to perform deterministic symbolic computation by calling a computer algebra system to obtain the computation result based on the computation formula and the parameters in the known parameter list.

[0043] The result verification module is used to perform triple verification on the calculation results, including dimension consistency verification, numerical range rationality verification, and boundary condition verification. If the triple verification fails, the abnormal item is marked and the calculation formula is reselected for iterative calculation and verification until the triple verification passes and a verification record is generated.

[0044] The evidence tracing module is used to integrate the calculation results that have passed triple verification, the calculation formula used, the standard number referenced by the calculation formula, the parameters used in the calculation formula, and the verification record into an evidence chain report and return it through the interaction module.

[0045] The beneficial effects of this invention are as follows:

[0046] (1) The present invention constructs an engineering calculation decision mechanism consisting of the intent parsing module, the formula knowledge base module and the formula matching module, which can convert the input engineering calculation requirement text into a structured task description and perform matching and confidence evaluation based on a professional formula knowledge base; when the matching result has insufficient confidence, it actively initiates interactive confirmation for manual (engineer) selection, thereby avoiding formula selection errors caused by the large model intent misunderstanding, and improving the accuracy of engineering calculation task identification and formula calling.

[0047] (2) The present invention uses the symbolic computation engine module to replace the large model in performing formula derivation and numerical calculation, limiting the large model to the task understanding and process arrangement stage, while entrusting the specific calculation process to the computer algebra system. This fundamentally avoids the model illusion, intermediate step errors and key information omissions that occur in the large model during complex engineering calculations, multi-step logical reasoning and long chain calculations, and significantly improves the accuracy and certainty of the calculation results.

[0048] (3) The present invention sets up a result verification mechanism. After the calculation is completed, the dimension consistency verification, the numerical range rationality verification and the boundary condition verification are performed in sequence to perform multiple verifications on the calculation results. This can detect problems such as formula application errors, abnormal parameter inputs and dimension mismatches, and calculation results deviating from the engineering experience range in a timely manner before the results are output. This effectively reduces the risk of erroneous results entering the engineering design and decision-making process and improves the reliability and security of the system operation.

[0049] (4) The present invention uses the evidence tracing module to associate and manage the calculation results, the calculation formulas used, the source of the parameters used, the verification records, and the standard number referenced by the calculation formulas, and generates an auditable evidence chain report, so that engineers can clearly trace the basis for the formation of the calculation results and the verification process, effectively solving the problems of lack of basis, difficulty in verification and lack of traceability of the output results of existing large models, and improving the credibility and auditability of engineering calculation results.

[0050] (5) This invention combines the semantic understanding capability of large models with the regular and deterministic engineering computing capability through the collaborative processing flow of “natural language understanding - formula intelligent matching - deterministic symbol calculation - result verification - evidence tracing”. This not only improves the human-computer interaction efficiency of engineering computing, but also effectively suppresses the reasoning illusion problem of large models in complex engineering scenarios, thereby meeting the application requirements of engineering fields for computational accuracy, reliability and traceability. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of a method for solving computational illusions in large models in complex engineering fields according to the present invention;

[0053] Figure 2 This is a flowchart of obtaining the candidate formula list in this invention;

[0054] Figure 3 This is a schematic diagram of a system for solving computational illusions in large models in complex engineering fields according to the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the protection scope of the present invention.

[0056] In this invention, "human" refers to the operator who provides the system with information such as engineering calculation requirements, parameter information, and missing information. In the embodiments, the human is an engineering technician with the corresponding professional background, such as an electrical engineer, structural engineer, mechanical engineer, thermal engineer, or other professionals engaged in engineering design, analysis, calculation, and verification. However, it should be understood that the human is not limited to the aforementioned professional engineers, but can also be a technician, designer, maintenance personnel, or other user with the corresponding operating permissions.

[0057] Example 1:

[0058] like Figure 1 As shown, this embodiment provides a method for solving computational illusions in large models in complex engineering fields, including steps S100 to S600, as detailed below:

[0059] Step S100: Perform intent parsing on the engineering calculation requirement text input by the engineer to generate a structured task description; the task description includes engineering domain tags, calculation objectives, a list of known parameters, and parameter units;

[0060] Specifically, in step S100, a domain-fine-tuned large language model combined with an engineering terminology dictionary is used to process the engineering calculation requirement text input by the engineer in natural language as follows:

[0061] (1) Input preprocessing: The engineering calculation requirement text is processed by sentence segmentation and word segmentation, and text normalization operation is performed. The text normalization operation includes unifying full-width and half-width characters, standardizing the writing of numbers and units of measurement, and cleaning special symbols and redundant whitespace, thereby eliminating the interference of text noise on subsequent recognition and obtaining a normalized text sequence to be parsed.

[0062] (2) Keyword extraction: Each entry in the engineering terminology dictionary includes the term name, synonyms and aliases, the engineering field to which it belongs, the corresponding physical quantity type and standard unit of measurement; a dual-channel approach combining engineering terminology dictionary matching and domain fine-tuning large language model named entity recognition is adopted to identify named entities such as engineering discipline category words, calculation type words, target physical quantity nouns, known parameter nouns and the numerical values ​​and units of parameter nouns from the standardized text sequence to be parsed; wherein, the engineering terminology dictionary matching channel ensures the accuracy of the recognition of the included professional terms, and the domain fine-tuning large language model named entity recognition channel supplements the recognition of colloquial or non-standard expressions not included in the dictionary, and the recognition results of the two channels are combined to improve the recognition recall rate.

[0063] (3) Engineering semantic disambiguation: In response to the common phenomena of polysemy, abbreviation ambiguity and unit ambiguity in engineering context, the semantic determination of ambiguous named entities is performed by combining the context and the identified engineering domain labels and calling the domain fine-tuning large language model. When the same term has different meanings in different engineering domains, its unique semantics are determined based on the domain indicator words in the context. When the same abbreviation corresponds to multiple terms, the correct term is selected based on the engineering domain to which it belongs. When there are multiple interpretations of the parameter unit, the correct unit is selected based on the physical quantity type corresponding to the parameter, thereby eliminating ambiguity and determining the unique engineering semantics of each named entity.

[0064] (4) Structured slot mapping: A structured slot template containing engineering domain labels, calculation targets, a list of known parameters, and parameter units is predefined; the disambiguated named entities are filled into the corresponding slots in the structured slot template according to their semantic types; engineering discipline category words are mapped to engineering domain labels, calculation type words and target physical quantity names are mapped to calculation targets, and known parameter names and their values ​​are mapped to a list of known parameters. The units of each parameter are normalized and converted before being filled into the parameter unit slots; after filling, the required slots are checked for missing information. If the calculation target or key known parameters are missing, the engineer fills in the missing information. If all required slots are completely filled, the structured task description is generated.

[0065] Step S200: According to the task description, search the constructed formula knowledge base, match applicable candidate formulas, sort them by confidence level, and obtain a candidate formula list.

[0066] Preferably, the formula knowledge base stores a set of formulas reviewed by professionals, including engineering formulas from various engineering fields. All formulas in the knowledge base are organized hierarchically by engineering field, and each formula record includes the formula expression, applicable working conditions, engineering experience value range, parameter domain, dimensional information, and referenced standard number. The formula knowledge base provides a source of retrieval data for matching applicable candidate formulas.

[0067] Based on the above solutions, such as Figure 2 As shown, step S200, which involves matching applicable candidate formulas in the formula knowledge base to obtain a list of candidate formulas sorted by confidence level, includes:

[0068] S201. Parse the task description into search conditions;

[0069] S202. Based on the search criteria, a two-layer recall method combining hard constraint filtering and semantic similarity retrieval is used to retrieve an applicable set of candidate formulas from the formula knowledge base.

[0070] The hard constraint filtering filters the formula knowledge base based on the engineering domain tags in the task description, the target physical quantity corresponding to the calculation target, and the list of known parameters, and removes formulas that do not match the engineering domain, whose target physical quantity is inconsistent with the formula output quantity, or whose known parameters cannot cover the input parameters required by the formula.

[0071] The semantic similarity retrieval encodes the task description and the text of each formula record in the formula knowledge base, such as applicable working conditions, parameter definitions, and dimensional information, into semantic vectors. By calculating the vector similarity, formulas with similar semantics but different expressions are recalled as a supplement to the hard constraint filtering results.

[0072] After merging and deduplicating the hard constraint filtering and the semantic similarity retrieval results, a candidate formula set is formed;

[0073] S203. For each candidate formula in the candidate formula set, calculate the matching degree between the candidate formula and the task description from multiple dimensions;

[0074] The dimensions include engineering domain matching degree, target physical quantity matching degree, coverage of the known parameter list to the formula input parameters, whether the parameters in the known parameter list fall into the domain of the formula parameters, dimensional consistency, and semantic similarity.

[0075] S204. The matching degrees calculated for each dimension are weighted, summed, and normalized to obtain the confidence level of each candidate formula.

[0076] S205. Sort each candidate formula according to the confidence level from high to low or from low to high to obtain the candidate formula list.

[0077] In one embodiment, step S204 calculates the confidence level of each candidate formula using a weighted scoring method. The various dimensions—engineering domain matching degree, target physical quantity matching degree, coverage of the formula input parameters by the known parameter list, whether the parameters in the known parameter list fall within the formula parameter domain, dimensional consistency, and semantic similarity—each correspond to a preset weight. , , , , and ,in: .

[0078] The preset weights can be configured or dynamically adjusted according to computational needs, engineering field characteristics, historical computational samples, and the size of the formula knowledge base.

[0079] The original scoring formula for each candidate formula is as follows:

[0080] ;

[0081] Wherein, S represents the original score value; The score representing the degree of matching in the engineering field; The score value representing the matching degree of the target physical quantity; The score representing parameter coverage; The score value represents the matching degree of the domain of the formula parameters; The score value represents the consistency of dimensions; A score representing semantic similarity;

[0082] As a multi-dimensional scoring scheme, it can be:

[0083] Engineering field match: Completely identical. Consistent with the superior domain, Inconsistent .

[0084] Target physical quantity matching degree: completely consistent. ; Belongs to the same type of physical quantity, Inconsistent .

[0085] Parameter coverage (the degree to which the known parameter list covers the formula input parameters):

[0086] .

[0087] Formula parameter domain matching degree (whether the parameters in the known parameter list fall within the formula parameter domain): All fall within the domain. Partially fell in, None of them fall into the category. .

[0088] Dimensional consistency: Completely consistent. Inconsistent .

[0089] Semantic similarity: can be calculated by fine-tuning a large language model in a specific domain, and .

[0090] After obtaining the raw scores of all candidate formulas, the Min-Max normalization method is used to normalize the raw scores of all candidate formulas, resulting in the normalized confidence score of each candidate formula:

[0091] ;

[0092] Among them, the This represents the confidence level of the j-th candidate formula; the This represents the original score of the j-th candidate formula; Represents the minimum original score in the candidate formula set; This represents the maximum original score in the set of candidate formulas.

[0093] Step S300: Determine whether the highest confidence level is less than a set threshold; wherein, the set threshold is dynamically adjusted according to the security level of the engineering field; the value range of the threshold is 0.7~0.95, and the value of the threshold is proportional to the security level of the engineering field, that is, the higher the security level of the engineering field, the larger the value of the threshold.

[0094] If the highest confidence level in the candidate formula list is less than the threshold, the sorted candidate formula list, along with the confidence level of each candidate formula and the reference standard number (source basis), will be presented to the engineer, who will then select the calculation formula from the candidate formula list.

[0095] If the highest confidence level in the candidate formula list is greater than or equal to the threshold, then the candidate formula corresponding to the highest confidence level is selected as the calculation formula.

[0096] Step S400: Based on the calculation formula and the parameters in the known parameter list, perform deterministic symbolic calculation by calling a computer algebra system to obtain the calculation result;

[0097] It should be noted that deterministic symbolic computation refers to the symbolic derivation and numerical solution process performed by a Computer Algebra System (CAS) according to predefined mathematical rules and operational logic. For the same input formula and parameters, the CAS always produces the same computational result, unaffected by random sampling, probabilistic inference, or model generation processes. Therefore, it features unique results, reproducible processes, and verifiable computational accuracy.

[0098] Specifically, the large language model is mainly used to complete cognitive decision-making tasks such as engineering calculation requirement analysis, parameter extraction, formula retrieval, and formula matching, rather than directly undertaking the final mathematical calculations.

[0099] Compared with direct use of large language models for numerical reasoning, deterministic symbolic computation avoids the instability of results caused by model illusion, numerical drift, and the accumulation of rounding errors, thereby improving the reliability of engineering calculation results and the security of engineering applications.

[0100] Furthermore, the computer algebra system includes a symbolic computation subsystem and third-party professional engineering analysis software;

[0101] The symbolic computation subsystem is a self-developed system customized for common engineering computation scenarios. It does not rely on external services and features fast response speed, low operating overhead, controllable result format, and is unaffected by external network conditions and the availability of third-party services, thereby ensuring the efficient and stable execution of a large number of simple computation tasks.

[0102] The third-party professional engineering analysis software includes, but is not limited to, marine structure analysis software SACS, floating body hydrodynamic analysis software MOSES, and building structure finite element analysis software ETABS. The third-party professional engineering analysis software has mature, complete and engineering practice-verified solution capabilities in complex engineering such as structural analysis, hydrodynamic analysis, and finite element analysis, which can ensure the accuracy and reliability of the calculation results of complex engineering.

[0103] In this embodiment, the complexity of the computational task corresponding to the calculation formula is determined first. The determination criteria include the types of operations involved in the calculation, the expression structure, the solution method, and whether it relies on a professional engineering analysis model. When the computational task only involves operations such as the four arithmetic operations, exponentiation and root extraction, evaluation of simple algebraic expressions, and univariate linear equations, which can be directly solved by general symbolic operations, it is determined to be a simple computation. When the computational task involves operations such as structural strength and stability analysis, fatigue analysis, dynamic response analysis, hydrodynamic analysis of floating bodies, and finite element analysis, which require solving based on a professional engineering analysis model, it is determined to be a complex computation.

[0104] Then, a hierarchical solution strategy is adopted based on the task complexity:

[0105] For tasks determined to be simple calculations, the symbolic computation subsystem is invoked to perform the solution and obtain the calculation result;

[0106] For tasks deemed complex, a predefined standardized interface is used to call third-party professional engineering analysis software to perform the solution and obtain the calculation results. Specifically, the calculation formulas and parameters in the known parameter list are converted into corresponding calculation models and input data according to the input specifications of the called software, submitted to the third-party professional engineering analysis software for solution, and the calculation results are read and parsed after the solution is completed.

[0107] Step S500: Perform the following triple verification on the calculation results obtained in step S400:

[0108] The first step is to verify the consistency of dimensions: based on the dimension information corresponding to each parameter in the calculation formula, the physical dimensions of the parameters in the known parameter list and the calculation results are derived, and it is determined whether the dimensions of the derived output results are consistent with the target dimensions corresponding to the calculation target.

[0109] The second step is to verify the reasonableness of the data range: compare the calculation results with the engineering experience value range of the calculation formula, and compare whether the calculation results are within the reasonable value range of the corresponding engineering parameters.

[0110] The third step is boundary condition verification: Based on the parameter definition domain and applicable working conditions corresponding to the calculation formula, the parameters in the known parameter list and the calculation results are verified to determine whether the parameters in the known parameter list meet the domain constraints of the calculation formula, whether the engineering working conditions meet the applicable working conditions of the calculation formula, and whether the calculation results exceed the engineering experience value range of the calculation formula.

[0111] A further step is to determine whether all three checks mentioned above have passed.

[0112] If all three checks pass, a check record is generated; then step S600 is executed.

[0113] If any one of the triple checks fails, an exception is marked, and the process returns to step S300, where the engineer reselects a calculation formula from the candidate formula list. Then, step S400 is executed, and based on the reselected calculation formula and the parameters in the known parameter list described in step S100, a deterministic symbolic calculation is performed by calling a computer algebra system to obtain the calculation result. Then, step S500 is executed, and the calculation result obtained in step S400 is subjected to triple check, and it is determined whether the triple check passes.

[0114] Until all three checks pass;

[0115] It should be noted that in order for all triple checks to pass, the cycle of steps S300-S400-S500 may be executed more than once.

[0116] In another alternative case, if any one of the triple checks fails, an anomaly is marked, the engineer corrects the parameters, and the calculation result is obtained by re-executing step S400 based on the corrected parameters, and then the triple check is performed in step S500 until all three checks pass.

[0117] The following example, using the calculation of the maximum bending stress of a simply supported beam with a rectangular cross-section subjected to a concentrated force at mid-span, illustrates the above triple verification:

[0118] Selected calculation formula: ; wherein, the The maximum bending stress is given by M, where M is the mid-span bending moment, W is the section modulus of bending, F is the mid-span concentrated force, L is the beam span, b is the width of the rectangular section, and h is the height of the rectangular section. The parameters provided by the engineer are: mid-span concentrated force F = 10000 N, beam span L = 2 m, rectangular section width b = 0.05 m, and rectangular section height h = 0.1 m. The maximum bending stress is obtained by using a computer algebra system. That is, 60MPa.

[0119] Perform triple verification on the calculation results:

[0120] The first step is dimensional consistency verification: deriving the physical dimensions of each term in the calculation formula. (Molecule) The dimensions are: force × length = The dimension of bending moment is denominator; The dimensions are: That is, the dimension of the section bending modulus; dividing the two yields The value is consistent with the standard dimension of stress, Pascal (Pa). This re-verification passed, indicating that the selected calculation formula is correct at the dimensional level and no dimensional inconsistency occurred due to formula mismatch.

[0121] If the derived output dimensions do not match the expected stress dimensions, it is determined that the formula selection is incorrect and marked as abnormal.

[0122] The second step is to verify the reasonableness of the data range: the calculated result of 60 MPa is compared with the range of engineering experience values ​​recorded in the calculation formula. For commonly used structural steel beams, the reasonable range of their working stress is preset to be approximately 0–235 MPa (with the yield strength of Q235 steel of approximately 235 MPa as the upper limit reference). In this example, the calculated result of 60 MPa falls within this range and has a reasonable safety margin, so this second verification passes.

[0123] If an engineer mistakenly inputs the mid-span concentrated force F as 10,000,000 N, the calculated result will reach approximately 60,000 MPa, far exceeding the above-mentioned engineering experience range. In this case, the numerical range will be judged as abnormal, and the engineer will be prompted to check the input parameters, thereby intercepting absurd results caused by incorrect parameter or unit input.

[0124] The third step is boundary condition verification: Boundary conditions are verified for the input parameters and the applicable premises of the calculation formula. Firstly, the parameter domain is verified: the rectangular section width b = 0.05 m, the rectangular section height h = 0.1 m, and the beam span L = 2 m are all positive values, and the denominators are... First, the value is not zero, satisfying the domain requirement of the formula. Second, the applicable premise of the calculation formula is verified, confirming that the load type in this example is a concentrated force at mid-span and the support method is simply supported, consistent with the assumptions of the selected calculation formula. Third, the physical limit is verified, the calculated result of 60 MPa is less than the material yield strength of 235 MPa, the beam is within the elastic range, and the applicable premise of the selected calculation formula is valid. All the above boundary conditions are met, and the re-verification passes.

[0125] If the rectangular cross-section dimension is zero, resulting in a zero denominator, or if the calculated stress exceeds the yield strength, causing the elasticity assumption to be invalid, then the calculation formula is deemed to be outside its applicable boundary and marked as abnormal.

[0126] In this example, the calculated result of 60 MPa passes the dimension consistency check, data range rationality check, and boundary condition check in sequence, and a check record is generated.

[0127] Step S600: The calculation results that have passed the triple verification, the calculation formula used, the specification number referenced by the calculation formula, the parameters used in the calculation formula, and the verification record are integrated into an evidence chain report and returned to the engineer.

[0128] Example 2:

[0129] like Figure 3 As shown, this embodiment provides a system for resolving computational illusions in large models within complex engineering domains, used to execute a method for resolving computational illusions in large models within complex engineering domains as described in Embodiment 1; the system includes:

[0130] The interaction module is used to interact with engineers;

[0131] The intent parsing module is used to receive engineering calculation requirement text input by engineers through the interaction module, parse the intent, generate and output a structured task description; wherein, the task description includes engineering domain tags, calculation objectives, a list of known parameters, and parameter units;

[0132] The formula matching module is configured with a pre-built formula knowledge base. This module receives the task description, searches the knowledge base, matches applicable candidate formulas, sorts them by confidence level, and returns a list of candidate formulas. The module also determines whether the highest confidence level is greater than a set threshold. If it is, the candidate formula corresponding to the highest confidence level is selected as the calculation formula. Otherwise, the module initiates an interactive confirmation with the engineer, who then selects the calculation formula from the candidate formula list and outputs the calculation formula.

[0133] The symbolic computation engine module is used to receive the computation formula and the parameters in the known parameter list, and to obtain the computation result by calling the computer algebra system to perform deterministic symbolic computation.

[0134] The result verification module is used to receive the calculation result and perform triple verification, including dimension consistency verification, numerical range rationality verification, and boundary condition verification; if all three verifications pass, a verification record is generated; and the calculation result and the verification record are output.

[0135] If any one of the triple checks fails, the exception is marked and fed back to the formula matching module. The matching module then initiates another interactive confirmation with the engineer through the interaction module, allowing the engineer to reselect the calculation formula from the candidate formula list and output it. The symbolic calculation engine module then performs the calculation again based on the reselected formula and outputs the result. The result verification module then performs the triple check on the calculation result again until all three checks pass.

[0136] The evidence tracing module is used to receive the calculation results that have passed triple verification, the calculation formula used, the standard number referenced by the calculation formula, the parameters used in the calculation formula, and the verification records output by the result verification module, and integrate them into an evidence chain report, which is then returned to the engineer through the interaction module.

[0137] In this embodiment, the intent parsing module has a built-in engineering terminology dictionary organized hierarchically by engineering field.

[0138] The intent parsing module interprets the engineering calculation requirement text and, during the structured slot mapping process, if it detects that the calculation target or key known parameters are missing, it initiates a supplementary query to the engineer through the interaction module to obtain the missing information.

[0139] For another scheme in the result verification module, if any one of the triple verifications fails, an anomaly is marked. The result verification module prompts the engineer to check the input parameters through the interaction module to correct them. The symbolic computation engine module calls the computer algebra system again to perform deterministic symbolic computation based on the corrected parameters to obtain the calculation result. The calculation result is then transmitted to the result verification module, which performs the triple verification on the calculation result until all three verifications pass.

[0140] This invention is not limited to the above-described optional embodiments. Anyone can derive other various forms of products under the guidance of this invention. However, regardless of any changes made in their shape or structure, any technical solution that falls within the scope of the claims of this invention shall be protected by this invention.

Claims

1. A method for solving computational illusions in large models in complex engineering fields, characterized in that, include: The input engineering calculation requirement text is parsed to generate a structured task description; the task description includes engineering domain tags, calculation objectives, a list of known parameters, and parameter units; According to the task description, a search is performed in the constructed formula knowledge base to match applicable candidate formulas, and the candidate formulas are sorted by confidence level to obtain a list of candidate formulas. Determine whether the highest confidence level is less than the set threshold; If it is less than, then a calculation formula will be manually selected from the list of candidate formulas; Otherwise, the candidate formula corresponding to the highest confidence level is selected as the calculation formula; Based on the calculation formula and the parameters in the known parameter list, the calculation result is obtained by calling a computer algebra system to perform deterministic symbolic calculation; The calculation results are subject to triple verification, including dimensional consistency verification, numerical range rationality verification, and boundary condition verification. If all three checks fail, an anomaly is marked and a new calculation formula is selected for iterative calculation and verification until all three checks pass and a verification record is generated. The calculation results that have passed triple verification, the calculation formula used, the standard number referenced by the calculation formula, the parameters used in the calculation formula, and the verification record are integrated into an evidence chain report and returned.

2. The method for solving computational illusions in large models in complex engineering fields according to claim 1, characterized in that, The formula knowledge base stores a set of formulas that have been manually reviewed; all formulas in the formula knowledge base are organized hierarchically by engineering field, and each formula record includes the formula expression, applicable working conditions, engineering experience value range, parameter definition domain, dimensional information, and referenced standard number.

3. The method for solving computational illusions in large models in complex engineering fields according to claim 2, characterized in that, The step of matching applicable candidate formulas in the formula knowledge base to obtain a list of candidate formulas sorted by confidence level includes: The task description is parsed into search criteria; Based on the search criteria, a two-layer recall method combining hard constraint filtering and semantic similarity retrieval is used to retrieve an applicable set of candidate formulas from the formula knowledge base. For each candidate formula in the candidate formula set, the matching degree between the candidate formula and the task description is calculated from multiple dimensions. The dimensions include engineering domain matching degree, target physical quantity matching degree, coverage of the formula input parameters by the known parameter list, whether the parameters in the known parameter list fall into the domain of the formula parameters, dimensional consistency, and semantic similarity. The matching scores calculated from each dimension are weighted, summed, and normalized to obtain the confidence score of each candidate formula. Each candidate formula is sorted according to the confidence level to obtain the candidate formula list.

4. The method for solving computational illusions in large models in complex engineering fields according to claim 2, characterized in that, The content of the dimension consistency verification is as follows: based on the dimension information corresponding to each parameter in the calculation formula, the physical dimensions of the parameters in the known parameter list and the calculation result are derived, and it is determined whether the dimension of the derived output result is consistent with the target dimension corresponding to the calculation target. The content of the data range rationality verification is as follows: compare the calculation result with the engineering experience value range of the calculation formula, and compare whether the calculation result is within the reasonable value range of the corresponding engineering parameter; The boundary condition verification involves verifying the parameters in the known parameter list and the calculation results based on the parameter domain and applicable working conditions corresponding to the calculation formula. This includes determining whether the parameters in the known parameter list satisfy the domain constraints of the calculation formula, whether the engineering working conditions satisfy the applicable working conditions of the calculation formula, and whether the calculation results exceed the engineering experience value range of the calculation formula.

5. The method for solving computational illusions in large models in complex engineering fields according to claim 2, characterized in that, The set threshold is dynamically adjusted according to the safety level of the engineering field; the value of the threshold ranges from 0.7 to 0.95, and the value of the threshold is proportional to the safety level of the engineering field.

6. The method for solving computational illusions in large models in complex engineering fields according to claim 1, characterized in that, The computer algebra system includes a symbolic computation subsystem and third-party professional engineering analysis software; The complexity of the computational task corresponding to the aforementioned formula is determined, and a hierarchical solution strategy is adopted based on the task complexity: For tasks determined to be simple calculations, the symbolic computation subsystem is invoked to perform the solution and obtain the calculation results; For tasks deemed to involve complex calculations, third-party professional engineering analysis software is invoked to perform the solution and obtain the calculation results.

7. The method for solving computational illusions in large models in complex engineering fields according to claim 1, characterized in that, The process of determining whether all triple checks have passed also includes: if any one of the triple checks fails, an anomaly is marked, the parameters are manually corrected, the calculation is performed again based on the corrected parameters to obtain the calculation result, and the triple checks are performed again until all triple checks have passed.

8. The method for solving computational illusions in large models in complex engineering fields according to claim 1, characterized in that, A domain-fine-tuned large language model combined with an engineering terminology dictionary is used to preprocess the input engineering calculation requirement text, extract keywords, perform engineering semantic disambiguation, and structured slot mapping to generate a structured task description.

9. A method for solving computational illusions in large models in complex engineering fields according to claim 8, characterized in that, The input preprocessing involves: segmenting the engineering calculation requirement text into sentences and words, and performing text normalization to obtain a normalized sequence of text to be parsed; The keyword extraction content is as follows: using a dual-channel approach that combines engineering terminology dictionary matching with domain fine-tuning of large language model named entity recognition, multiple named entities are identified from the normalized text sequence to be parsed, including engineering discipline category words, calculation type words, target physical quantity words, known parameter words, and the numerical values ​​and units of parameter words; The content of the engineering semantic disambiguation is as follows: semantic determination of ambiguous named entities is performed through a domain-fine-tuned large language model to determine the unique engineering semantics of each named entity; The structured slot mapping consists of: pre-defining a structured slot template containing engineering domain labels, calculation targets, a list of known parameters, and parameter units; filling the structured slot template with disambiguated named entities according to semantic type; after filling, performing missing detection on the required slots; if a calculation target or key known parameter is detected to be missing, the missing information is manually supplemented; if all required slots are fully filled, a structured task description is generated.

10. A system for solving computational illusions in large models in complex engineering fields, characterized in that, A system for performing a method for resolving computational illusions in large models in complex engineering fields as described in any one of claims 1-9; the system comprising: Interactive module, used for human interaction; The intent parsing module is used to receive the engineering calculation requirement text input through the interaction module and perform intent parsing to generate a structured task description; wherein, the task description includes engineering domain tags, calculation objectives, a list of known parameters, and parameter units; The formula matching module is configured with a pre-built formula knowledge base; the formula matching module is used to search in the formula knowledge base according to the task description, match applicable candidate formulas and sort them by confidence level, and return a list of candidate formulas; and select the calculation formula from the list of candidate formulas based on a set threshold. The symbolic computation engine module is used to perform deterministic symbolic computation by calling a computer algebra system to obtain the computation result based on the computation formula and the parameters in the known parameter list. The result verification module is used to perform triple verification on the calculation results, including dimension consistency verification, numerical range rationality verification, and boundary condition verification. If the triple verification fails, the abnormal item is marked and the calculation formula is reselected for iterative calculation and verification until the triple verification passes and a verification record is generated. The evidence tracing module is used to integrate the calculation results that have passed triple verification, the calculation formula used, the standard number referenced by the calculation formula, the parameters used in the calculation formula, and the verification record into an evidence chain report and return it through the interaction module.

Images