Large language model anti-seismic report examination method and device based on domain knowledge enhancement

Abstract

The invention relates to the technical field of large language models, in particular to a large language model anti-seismic report examination method and device based on domain knowledge enhancement, and provides a large language model anti-seismic report examination method based on domain knowledge enhancement. The method comprises the steps that firstly, a building earthquake-proof professional field large language model is trained, then data preprocessing is conducted on an earthquake-proof examination report, and finally examination is conducted on the basis of the trained building earthquake-proof professional field large language model, so that accumulated earthquake-proof special examination experience and the large language model are fused, and a professional LLM of the vertical field is built; strict examination can be carried out according to standards and files, interference of human errors and subjective factors is effectively avoided, accuracy and reliability of examination results are improved, examination efficiency is improved, a more scientific and accurate basis is provided for expert examination and internal pre-examination, the earthquake resistance and disaster prevention capacity of a construction project is further improved, and the earthquake disaster risk is reduced.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, specifically a method and device for reviewing seismic reports based on domain knowledge enhancement using a large language model. Background Technology

[0002] The seismic performance of buildings is directly related to the safety of people's lives and property and social stability. In order to improve the seismic resistance and disaster prevention capabilities of construction projects, reduce the risk of earthquake disasters, and protect the safety of people's lives and property, corresponding special review requirements for the seismic design of building projects have been issued, especially in earthquake-prone areas.

[0003] For example, a certain province stipulates that "Class A buildings and Class B buildings of medium or larger size, as well as buildings exceeding the applicable scope of the standard but not exceeding the limits, must undergo a special seismic fortification review." Currently, the seismic review involves the owner and design unit submitting review materials, such as a "Seismic Review Report," which is then reviewed by experts organized by the housing and construction authorities. These experts, based on their professional experience and expertise, manually review a large amount of material, including the submitted report, design drawings, and calculation sheets, comparing them with relevant standards and documents. The following problems have been identified:

[0004] (1) Inefficient, with a large workload for review, requiring a lot of time and manpower, especially when dealing with large and complex projects;

[0005] (2) Easy to overlook. Due to the large amount of review materials, manual review is prone to oversights, which may pose a hidden danger to the seismic safety performance of these important buildings.

[0006] (3) Limited experience and knowledge, many types of projects to be reviewed, and a lot of technical content involved, which may exceed the experience and knowledge of the reviewers, requiring them to consult a large amount of information on short notice, which affects the quality and efficiency of the review.

[0007] Recently, several open-source AI Large Language Models (LLMs) have been launched both domestically and internationally. LLMs have already seen successful applications in education, healthcare, finance, and other fields, demonstrating enormous application potential and offering new perspectives and methods for the construction industry. LLMs can integrate large amounts of domain-specific knowledge and data, relying on deep learning and natural language processing technologies to more accurately understand and analyze information in documents. Meanwhile, the "Seismic Resistance Review Report" is primarily text and tables, requiring less multimodal computation and having relatively low hardware requirements. The "Seismic Resistance Review Report" is a long text, requiring complex semantic segmentation and enhanced retrieval, which is precisely where LLM's technical advantages lie, providing an opportunity to conduct seismic resistance report reviews by setting up domain-knowledge-enhanced large language models. Summary of the Invention

[0008] The purpose of this invention is to provide a method and device for reviewing seismic reports based on a large language model with domain knowledge enhancement, addressing the aforementioned problems.

[0009] The technical solution adopted in this invention is as follows: a method for reviewing seismic resistance reports based on a large language model with domain knowledge enhancement, comprising the following steps:

[0010] S1. Establish a comprehensive language model for the field of seismic resistance in buildings, and set up a knowledge base, a review scope database, a review completeness database, and a review compliance database for the field of seismic resistance in buildings;

[0011] S2. Train the large language model in the field of seismic resistance of buildings to obtain sub-models for review tasks, including structural category classification model, design method identification model, review scope determination model, integrity determination model and compliance determination model;

[0012] S3. Perform data preprocessing on the seismic review report and construct a project profile;

[0013] S4. Review based on the trained large language model of building seismic resistance and output the review results.

[0014] Furthermore, S1 includes the following sub-steps:

[0015] S1.1. Collect national, industry, and local seismic design codes, building type review cases, and expert review opinions; use Python's natural language processing library to parse text and tables.

[0016] S1.2. Perform unstructured data processing on the parsed document;

[0017] S1.3. Establish a review scope database based on laws, regulations, and building seismic design standards;

[0018] S1.4. Establish a review integrity library based on the design conditions and calculation results that must be included under different structural types and design methods in seismic resistance of buildings;

[0019] S1.5. Construct a review compliance library, based on unstructured data after processing, and build a four-tuple knowledge graph in the form of subjects, relations, sub-subjects and objects;

[0020] S1.6. Transform the data in the knowledge base, review scope base, review integrity base, and review compliance base of the building seismic resistance professional field into high-dimensional vectors through an embedding model, and store them in a vector database for semantic retrieval.

[0021] Furthermore, in S1.2, the unstructured data processing of the parsed document includes setting metadata, experience pairs, and a terminology database;

[0022] For the code provisions of local seismic design codes in the parsed documents, they are atomized and decomposed according to "chapter-clause-sub-item", and mandatory provisions and applicable structural types are marked as metadata;

[0023] For the review cases of building types in the parsed documents, the four elements of "project parameters - violation points - corresponding clauses - rectification suggestions" are extracted in a structured manner to form experience pairs for the model to learn from;

[0024] The terminology database includes a mapping table of professional terms related to building types.

[0025] Furthermore, in S1.5, the normative clauses obtained from S1.2 after data processing and annotation are stored in a knowledge graph format of four tuples in the form of subject, relation, sub-subject, and object;

[0026] The subject can have multiple relationships, and each relationship can correspond to multiple sub-sub ...

[0027] Furthermore, in S1.3, the scope of review includes building classification, building use or building level, building size, and constraints;

[0028] The building classification includes disaster prevention and relief buildings, public buildings, and residential buildings;

[0029] The disaster prevention and relief buildings include medical buildings, disease prevention and control center buildings, disaster prevention and emergency command centers, emergency shelter buildings, fire truck garages and their duty rooms;

[0030] The public and residential buildings mentioned include educational buildings, scientific research buildings, sports buildings, cultural and entertainment buildings, commercial buildings, office buildings, convention and exhibition buildings, museum buildings, archives, information center buildings, elderly care institutions, and children's welfare homes.

[0031] Furthermore, in S1.4, the integrity review library is divided into two sub-libraries: design conditions and calculation results, and each review point R... i Mark t i and m i ;

[0032] The t i Indicates the structural category of the project, where i represents one of the following: reinforced concrete frame structure, reinforced concrete frame-shear wall structure, reinforced concrete shear wall structure, reinforced concrete tube structure, steel frame structure, steel braced structure, steel frame-braced structure, steel frame-shear wall panel structure, and steel tube structure.

[0033] The m i This indicates the project design method, where i represents one of the following: seismic design or seismic isolation and damping design.

[0034] Furthermore, in S2, an open-source large language model is selected as the base to train the large language model in the field of seismic resistance of buildings in S1. Through instruction fine-tuning training, sub-models for the review task are generated, including structural category classification model, design method identification model, review scope judgment model, integrity judgment model and compliance judgment model.

[0035] LoRA was used to inject seismic review question-and-answer pairs into the trained large language model, which was used to teach the large language model domain knowledge and reasoning patterns.

[0036] Before generating any review results, RAG must first retrieve at least one normative clause or case from the vectorized database as a basis.

[0037] Furthermore, step S3 includes the following sub-steps:

[0038] S3.1. Data preprocessing: Python's natural language processing library is used to clean the text and table data of the review report. After preprocessing, the data is uniformly converted into a format suitable for input to a large language model.

[0039] S3.2. Using the structural category classification model, retrieve project information, determine the structural category, and match the annotations t in the integrity review database. i The review points;

[0040] S3.3. Identify the model through design methods, retrieve project information, determine the design method, and match it with the m marked in the integrity review library. i The provisions;

[0041] S3.4. Level 1 review: Through the review scope determination model, retrieve project information, automatically extract project information of the report to be reviewed, associate with the review scope database, and determine whether the project falls within the review scope.

[0042] S3.5. Secondary review, based on annotation t i and the annotation m i This triggers a condition, invokes the integrity determination model, and associates the integrity data marked with t in the integrity review library. i and the annotation m i Review point R i Search for R in the report i If relevant parameters are found, the process proceeds to Level 3 review (S3.6), as shown in the following formula:

[0043] R i =R 1i +R 2i ={R1∈(t 1i ∩m 1i )+R2∈(t 2i ∩m2i )};

[0044] In the above formula, Ri refers to the review point in the review integrity library;

[0045] R 1i This refers to the review points related to the integrity of the review library and design conditions;

[0046] R 2i This refers to the review points related to the integrity of the database and the calculation results;

[0047] t i This refers to the review points in the integrity review library that are marked with structural categories;

[0048] t 1i This refers to review points in the integrity database that are relevant to the design conditions and have been marked with structural categories;

[0049] t 2i This refers to review points that are related to the integrity of the review database and the calculation results, and that are marked with structural categories;

[0050] m 1i This refers to review points in the integrity library that are related to the design conditions and have been marked with the design methodology.

[0051] m 2i This refers to review points in the integrity database that are relevant to the calculation results and have been marked with the design method.

[0052] If not retrieved, the review conclusion will output that this parameter is missing.

[0053] S3.6. Three-level review: Establish the correspondence between the parameters of the integrity assessment model and the compliance assessment model, and incorporate R... i The corresponding four-tuple relationship in the compliance judgment model knowledge base will be t i Corresponding to the main body of the quadruple, m i The remaining parameters correspond to the quadruple subclasses, and the quantized parameters correspond to the quadruple objects.

[0054] Furthermore, in S4, after reviewing based on the trained large language model of the building seismic resistance professional field, a Word version of the review conclusion is output, organized, and presented. For the compliant parts in the review report, compliance is marked and displayed in the background; for the problematic parts, the location of the problem, the type of problem, and the specific seismic code clause violated are output in the review conclusion. At the same time, if there are mandatory clauses marked, the mandatory clauses are output and displayed in bold.

[0055] This application also provides a large language model seismic report review device based on domain knowledge enhancement, which is suitable for performing the above-mentioned large language model seismic report review method based on domain knowledge enhancement.

[0056] The beneficial effects of this invention are as follows:

[0057] This paper presents a domain-knowledge-enhanced large language model-based method for reviewing seismic resistance reports. It integrates accumulated seismic resistance review experience with the large language model to build a professional LLM in the vertical domain. This method can conduct rigorous reviews against standards and documents, effectively avoiding interference from human error and subjective factors, improving the accuracy and reliability of review results, increasing review efficiency, and providing a more scientific and accurate basis for expert review and internal pre-review. This further enhances the seismic resistance and disaster prevention capabilities of construction projects and reduces the risk of earthquake disasters. Attached Figure Description

[0058] Figure 1 Flowchart of a method for reviewing seismic reports based on domain knowledge-enhanced large language models.

[0059] Figure 2 A diagram illustrating the relationships between subject, relation, sub-subject, and object.

[0060] Figure 3 This is a schematic diagram of an embodiment.

[0061] Figure 4 A schematic diagram of the integrity library structure for review.

[0062] Figure 5 A diagram illustrating the matching of the integrity database and the compliance determination model knowledge base for review.

[0063] Figure 6 This is a schematic diagram of the structure of an electronic device. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described in the accompanying drawings can generally be arranged and designed in various different configurations.

[0065] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0066] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0067] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0068] like Figure 1 As shown, the seismic report review method based on a large language model enhanced with domain knowledge includes the following steps:

[0069] S1. Establish a comprehensive language model for the field of seismic resistance in buildings, and set up a knowledge base, a review scope database, a review completeness database, and a review compliance database for the field of seismic resistance in buildings;

[0070] S2. Train the large language model in the field of seismic resistance of buildings to obtain sub-models for review tasks, including structural category classification model, design method identification model, review scope determination model, integrity determination model and compliance determination model;

[0071] S3. Perform data preprocessing on the seismic review report and construct a project profile;

[0072] S4. Review based on the trained large language model of building seismic resistance and output the review results.

[0073] The purpose of this design is to provide a domain-knowledge-enhanced large language model-based method for reviewing seismic resistance reports. This method integrates accumulated experience in seismic resistance review with the large language model to build a professional LLM for the vertical domain. It can conduct rigorous reviews against standards and documents, effectively avoiding interference from human error and subjective factors, improving the accuracy and reliability of review results, increasing review efficiency, and providing a more scientific and accurate basis for expert review and internal pre-review. This will further improve the seismic resistance and disaster prevention capabilities of construction projects and reduce the risk of earthquake disasters.

[0074] This approach addresses the shortcomings of existing seismic report review processes, such as inefficiency, heavy workload, and the need for significant time and manpower. It is particularly problematic when dealing with large and complex projects, where oversights are common. The sheer volume of review materials makes manual review prone to errors, potentially jeopardizing the seismic safety performance of important buildings. Furthermore, the limited experience and knowledge of reviewers, coupled with the diverse types of projects and the wide range of technical content involved, sometimes exceed their expertise and require extensive on-the-spot research, impacting both the quality and efficiency of the review process.

[0075] In this embodiment, step S1 includes the following sub-steps:

[0076] S1.1. Collect national, industry, and local seismic design codes, building type review cases, and expert review opinions; use Python's natural language processing library to parse text and tables.

[0077] S1.2. Perform unstructured data processing on the parsed document;

[0078] S1.3. Establish a review scope database based on laws, regulations, and building seismic design standards;

[0079] S1.4. Establish a review integrity library based on the design conditions and calculation results that must be included under different structural types and design methods in seismic resistance of buildings;

[0080] S1.5. Construct a review compliance library, based on unstructured data after processing, and build a four-tuple knowledge graph in the form of subjects, relations, sub-subjects and objects;

[0081] S1.6. Transform the data in the knowledge base, review scope base, review integrity base, and review compliance base of the building seismic resistance professional field into high-dimensional vectors through an embedding model, and store them in a vector database for semantic retrieval.

[0082] Among them, S1.1 includes local seismic design codes, review cases for various building types, and expert review opinions. Preprocessing of text and tables, such as document parsing, was performed using Python's NLTK natural language processing library and OCR+LayoutLM.

[0083] In S1.2, unstructured data processing of the parsed document includes setting metadata, experience pairs, and a terminology database;

[0084] For the code provisions of local seismic design codes in the parsed documents, they are atomized and decomposed according to "chapter-clause-sub-item", and mandatory provisions and applicable structural types are marked as metadata;

[0085] For the review cases of building types in the parsed documents, the four elements of "project parameters - violation points - corresponding clauses - rectification suggestions" are extracted in a structured manner to form experience pairs for the model to learn from;

[0086] The terminology database includes a mapping table of professional terms related to building types.

[0087] It should be noted that the aforementioned mapping table for professional terms is used to resolve issues such as synonyms, abbreviations, etc. For example, "the ratio of maximum inter-story displacement to story height" corresponds to "inter-story drift angle" and "Δu / h"; [Δu / h] corresponds to "Δu / h limit" and "inter-story drift angle limit".

[0088] Meanwhile, in this embodiment, in S1.3, the scope of review includes building classification, building use or building level, building size, and constraints;

[0089] The building classification includes disaster prevention and relief buildings, public buildings, and residential buildings;

[0090] In S1.4, the integrity review library is divided into two sub-libraries: design conditions and calculation results, and each review point R... i Mark t i and m i ;

[0091] The t i This indicates the structural category of the project; i represents a type of reinforced concrete shear wall structure or steel frame structure.

[0092] The m i This indicates the project design method, where 'i' represents one of the following: seismic design or seismic isolation and damping design.

[0093] In S1.5, the normative clauses obtained from the data processing and annotation of S1.2 are stored in a knowledge graph format of four tuples in the form of subject, relation, sub-subject and object;

[0094] like Figure 2 As shown, the subject can have multiple relationships, and each relationship can correspond to multiple sub-sub ...

[0095] In practice, such as Figure 3 and Figure 6 As shown, according to building classification (including primary and secondary categories, purpose or level, building scale, and constraints such as building area or building height), the inpatient building area of ​​a tertiary hospital is >0m². 2 Or, if the height is greater than 0m, a special review is required. The inpatient building area of ​​a secondary hospital is greater than 5000m². 2 If the height is greater than 24m, a special review is required.

[0096] like Figure 3 As shown, based on the provisions of the "Technical Specification for Concrete Structures of High-Rise Buildings", the quaternion storage is as follows:

[0097] {“Reinforced concrete frame structure”, “Limit of elastic inter-story drift angle”, “Height not greater than 150m”, “1 / 550”};

[0098] {“Reinforced concrete frame structure”, “Limit of elastic inter-story drift angle”, “Height not less than 250m”, “1 / 500”}.

[0099] In the large language model training part, S2 uses an open-source large language model as the base to train the large language model in the field of seismic engineering in S1. Through fine-tuning of instructions, sub-models for the review task are generated, including structural category classification model, design method identification model, review scope determination model, integrity determination model and compliance determination model.

[0100] LoRA was used to inject seismic review question-and-answer pairs into the trained large language model, which was used to teach the large language model domain knowledge and reasoning patterns.

[0101] Before generating any review results, RAG must first retrieve at least one normative clause or case from the vectorized database as a basis.

[0102] It should be noted that the use of sub-models for review tasks, including structural category classification model, design method identification model, review scope determination model, integrity determination model, and compliance determination model, reduces training difficulty and improves the execution accuracy of each task.

[0103] Meanwhile, LoRA (Low-Rank Adaptation) technology is used to inject thousands of high-quality seismic review question-and-answer pairs into the LLM, efficiently imparting domain knowledge and reasoning patterns to the model.

[0104] The format is:

[0105] Input: Report excerpt + review task;

[0106] Output: Judgment result + Legal clause ID + Reasoning process;

[0107] Comparison: The judgment result is compared with the preset result.

[0108] The specific example is shown below:

[0109] Input: Report excerpt: "Project overview (e.g., specific details: New office building for Chengdu XX Vocational Education School, covering an area of ​​4000 square meters, 6 floors in total, each floor 4.2m)", Review task: "Determine whether this project requires a seismic-resistant special review")

[0110] Output: Review required. According to Appendix 1 of the "Technical Guidelines for Seismic Design Review of Housing Construction Projects in Sichuan Province", the building is classified as an educational building and its purpose is an office building. The building height is 6 × 4.2m = 25.2m > 24m, so review is required.

[0111] Comparison: Judgment result: Requires review; Preset result: Requires review. Comparison: Correct.

[0112] Furthermore, Retrieval-Enhanced Generation (RAG) is employed: This forces the model to retrieve at least one regulatory provision or case from the vectorized knowledge base before generating any judgment result. This significantly improves the reliability and traceability of review conclusions and effectively suppresses model "illusions."

[0113] In actual implementation, in S3, such as Figure 4 and Figure 5As shown, it includes the following sub-steps:

[0114] S3.1. Data preprocessing: Python's natural language processing library is used to clean the text and table data of the review report. After preprocessing, the data is uniformly converted into a format suitable for input to a large language model.

[0115] S3.2. Using the structural category classification model, retrieve project information, determine the structural category, and match the annotations t in the integrity review database. i The review points;

[0116] S3.3. Identify the model through design methods, retrieve project information, determine the design method, and match it with the m marked in the integrity review library. i The provisions;

[0117] S3.4. Level 1 review: Through the review scope determination model, retrieve project information, automatically extract project information of the report to be reviewed, associate with the review scope database, and determine whether the project falls within the review scope.

[0118] For example, in the "Project Overview" section of the "Special Review Report on Seismic Resistance of the Inpatient Building Project of XX Hospital", the project description is as follows: "This project is the expansion project of the outpatient department of XX tertiary hospital in XX city. The project construction site is located on XX Street, XX city, Sichuan Province. The project is 42.4m high and adopts a reinforced concrete frame structure."

[0119] Extracting the information as {"tertiary hospital", "outpatient", "height 42.4m"}, and associating it with S1.3, we get: the building classification is "medical building", the use is "tertiary hospital outpatient", and the constraint is "building area 5514m²". 2 "If all conditions are met, it is determined that the case falls within the scope of review and therefore requires review."

[0120] S3.5. Secondary review, based on annotation t i and the annotation m i This triggers a condition, invokes the integrity determination model, and associates the integrity data marked with t in the integrity review library. i and the annotation m i Review point R i Search for R in the report i If relevant parameters are found, the process proceeds to Level 3 review (S3.6), as shown in the following formula:

[0121] R i =R 1i +R 2i ={R1∈(t 1i ∩m 1i )+R2∈(t 2i ∩m 2i )};

[0122] If not retrieved, the review conclusion will output that this parameter is missing.

[0123] In the above formula, Ri refers to the review point in the review integrity library;

[0124] R 1i This refers to the review points related to the integrity of the review library and design conditions;

[0125] R 2i This refers to the review points related to the integrity of the database and the calculation results;

[0126] t i This refers to the review points in the integrity review library that are marked with structural categories;

[0127] t 1i This refers to review points in the integrity database that are relevant to the design conditions and have been marked with structural categories;

[0128] t 2i This refers to review points that are related to the integrity of the review database and the calculation results, and that are marked with structural categories;

[0129] m 1i This refers to review points in the integrity library that are related to the design conditions and have been marked with the design methodology.

[0130] m 2i This refers to review points that are related to the integrity of the library and the calculation results and are marked with the design method.

[0131] Meanwhile, models such as structural category judgment models and design method identification models output t. i =Reinforced concrete frame structure, m i =Seismic design, then secondary review, R 1i = (Structural design service life, seismic fortification level, safety level), R 2i = (Period ratio, inter-story drift angle).

[0132] According to R i The search retrieved relevant texts such as "structural design service life 50 years", "seismic fortification level II", "period ratio 0.89", and "inter-story drift angle 1 / 550", which are consistent with R. 1i and R 2i contrast.

[0133] Output: "Security level" is missing.

[0134] S3.6. Three-level review: Establish the correspondence between the parameters of the integrity assessment model and the compliance assessment model, and incorporate R... i The corresponding four-tuple relationship in the compliance judgment model knowledge base will be t i Corresponding to the main body of the quadruple, m iThe remaining parameters correspond to the quadruple subclasses, and the quantized parameters correspond to the quadruple objects.

[0135] For example, "period ratio 0.89" corresponds to:

[0136] R i =Period ratio correspondence;

[0137] t i =The main structure corresponding to the reinforced concrete frame structure;

[0138] m i =Seismic design + Class A height corresponding sub-structure;

[0139] A quantization parameter of 0.89 corresponds to a quadruple object;

[0140] The processed quadruple is: {"reinforced concrete frame", "Class A height", "cycle ratio", "0.89"}. Compared with the clause quadruple {"Class A height high-rise reinforced concrete frame", "cycle ratio limit", "0.9"} in the "integrity library", it is determined to be compliant.

[0141] Finally, in S4, after reviewing the large language model of the building seismic resistance professional field based on the training, the review conclusion is output in Word format, and then organized and presented. For the parts of the review report that are deemed compliant, compliance is marked and displayed in the background; for the parts with problems, the review conclusion outputs the location of the problem, the type of problem, and the specific seismic code clauses violated. At the same time, if there are mandatory clauses marked, the mandatory clauses are output and displayed in bold.

[0142] In the example, the user uploads a "Seismic Resistance Special Review Report". The background data is preprocessed and a trained professional domain language model is used to review the processed "Seismic Resistance Special Review Report". Finally, after the large language model completes the review, it outputs the review results.

[0143] Taking the review of the "Special Review Report on Seismic Resistance of the Inpatient Building Project of XX Tertiary Hospital in XX City, Sichuan Province" as an example, the implementation process of this invention is explained as follows:

[0144] The system deployment includes:

[0145] Knowledge base configuration: Includes current standards and specifications, including more than 80 mandatory provisions, more than 200 general provisions, and 30 review cases of similar projects;

[0146] Model training: Based on the LLaMA3-70B model, the pedigree model was fine-tuned using 300 question-answer pairs and review cases. The LoRA rank was set to 32, the training epochs were 8, and the validation set accuracy was 94.3%.

[0147] Hardware environment: 8×A100GPU server.

[0148] Review and implement the steps, including:

[0149] 1. Report Input: Users upload a PDF review report (more than 40 pages in total, including more than 10 tables);

[0150] 2. Pretreatment;

[0151] OCR+LayoutLM is used to process text and tabular data and extract relevant data.

[0152] 3. Structural Category Classification:

[0153] Call the structural category classification model, extract the information "reinforced concrete frame", and output ti=reinforced concrete frame structure.

[0154] 4. Design Method Identification:

[0155] The design method is invoked to identify the model and extract information. The extracted information is {"Tertiary Hospital", 8-degree zone}, which belongs to "two zones and eight types of buildings". The output is mi = seismic isolation design.

[0156] 5. First-level review:

[0157] In the "Project Overview" section of the report, the extracted information is {"Tertiary Hospital", "Outpatient Department", "Height 42.4m"}. By associating with S1.1.3 and matching the knowledge base rules, the statement "Tertiary Hospital Outpatient Department → falls within the scope of review" is determined to require review.

[0158] 6. Secondary review:

[0159] According to the label t i and m i Related R i Retrieve these R from the report i item;

[0160] R 1i : Structural design service life, durability service life, seismic fortification category, safety level, foundation design level;

[0161] R 2i Period ratio, inter-story drift angle, period before and after seismic isolation under seismic fortification action;

[0162] The search revealed that the foundation design grade was missing, so the output "Foundation design grade missing" was sent to the "Result Output Model".

[0163] 7. Three-level review:

[0164] During the secondary review, the text "cycle ratio 0.91" was retrieved, and the corresponding quadruple was {"Grade A high-rise reinforced concrete frame", "cycle ratio", "0.91"}. Compared with the knowledge stored in the knowledge base, which is {"Grade A high-rise reinforced concrete frame", "cycle ratio limit", "0.9"}, it was determined to be non-compliant.

[0165] 8. Output Results:

[0166] Generate a review report. Report content:

[0167] ① Lack of foundation design grade;

[0168] ② The project's "cycle ratio" is in violation of regulations. According to the "Technical Specification for Concrete Structures of High-Rise Buildings", the torsion ratio of reinforced concrete buildings of Class A height shall not exceed 0.9.

[0169] like Figure 6 The diagram shown is a structural schematic of an electronic device according to an embodiment of this application. The electronic device includes a memory and a processor. The memory stores computer-readable instructions, and the processor executes the computer-readable instructions stored in the memory to implement the domain knowledge-enhanced large language model seismic report review method described in any of the above embodiments.

[0170] In one embodiment of this application, the electronic device further includes a bus and a computer program stored in the memory and executable on the processor, such as a domain knowledge-enhanced large language model seismic report review program.

[0171] The figure illustrates an electronic device with memory and a processor. Those skilled in the art will understand that the structure shown does not constitute a limitation on the electronic device and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0172] It should be noted that the memory in the electronic device stores multiple computer-readable instructions to implement a large language model seismic report review method based on domain knowledge enhancement, and the processor can execute the multiple instructions to implement this method.

[0173] Specifically, the processor's implementation method for the above instructions can be found in the description of the relevant steps in the corresponding embodiment of the figure, and will not be repeated here.

[0174] Those skilled in the art will understand that the schematic diagram is merely an example of an electronic device and does not constitute a limitation on the electronic device. The electronic device may be a bus-type structure or a star-type structure. The electronic device may also include more or fewer other hardware or software than shown in the diagram, or different component arrangements. For example, the electronic device may also include input / output devices, network access devices, etc.

[0175] It should be noted that electronic devices are merely examples. Other existing or future electronic products that are suitable for this application should also be included within the scope of protection of this application and are incorporated herein by reference.

[0176] The memory includes at least one type of readable storage medium, which can be non-volatile or volatile. The readable storage medium includes flash memory, portable hard drives, multimedia cards, card-type memory (e.g., SD or DX memory), magnetic storage, magnetic disks, optical disks, etc. In some embodiments, the memory can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory can be an external storage device of the electronic device, such as a plug-in portable hard drive, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. The memory can be used not only to store application software and various types of data installed on the electronic device, such as the code of a domain knowledge-enhanced large language model earthquake resistance report review program, but also to temporarily store data that has been output or will be output.

[0177] In some embodiments, a processor can be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions. This includes combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor is the control unit of an electronic device, connecting various components of the device through various interfaces and lines. It performs various functions and processes data by running or executing programs or modules stored in the memory (e.g., executing a customer push program based on intellectual property change data) and by accessing data stored in the memory.

[0178] The processor executes the operating system of the electronic device and various installed applications. The processor executes the applications to implement the steps in the above-described embodiments of the domain knowledge-enhanced large language model seismic report review method, such as the steps shown in the figure.

[0179] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete this application. The one or more modules / units may be a series of computer-readable instruction segments capable of performing a specific function, which describe the execution process of the computer program in an electronic device. For example, the computer program may be divided into a receiving module, a preprocessing module, a projection module, and a determining module.

[0180] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute portions of the domain knowledge-enhanced large language model seismic report review method described in the various embodiments of this application.

[0181] When modules / units integrated into an electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware devices. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above.

[0182] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for reviewing seismic resistance reports based on a large language model with enhanced domain knowledge, characterized in that: Includes the following steps: S1. Establish a comprehensive language model for the field of seismic resistance in buildings, and set up a knowledge base, a review scope database, a review completeness database, and a review compliance database for the field of seismic resistance in buildings; S2. Train the large language model in the field of seismic resistance of buildings to obtain sub-models for review tasks, including structural category classification model, design method identification model, review scope determination model, integrity determination model and compliance determination model; S3. Perform data preprocessing on the seismic review report and construct a project profile; S4. Review based on the trained large language model of building seismic resistance and output the review results.

2. The method for reviewing seismic resistance reports based on a large language model with domain knowledge enhancement according to claim 1, characterized in that, S1 includes the following sub-steps: S1.

1. Collect national, industry, and local seismic design codes, building type review cases, and expert review opinions, and use Python's natural language processing library to parse the text and tables. S1.

2. Perform unstructured data processing on the parsed document; S1.

3. Establish a review scope database based on laws, regulations, and building seismic design standards; S1.

4. Establish a review integrity library based on the design conditions and calculation results that must be included under different structural types and design methods in seismic resistance of buildings; S1.

5. Construct a review compliance library, based on unstructured data after processing, and build a four-tuple knowledge graph in the form of subjects, relations, sub-subjects and objects; S1.

6. Transform the data in the knowledge base, review scope base, review integrity base, and review compliance base of the building seismic resistance professional field into high-dimensional vectors through an embedding model, and store them in a vector database for semantic retrieval.

3. The method for reviewing seismic resistance reports based on a large language model with domain knowledge enhancement according to claim 2, characterized in that, In step S1.2, the unstructured data processing of the parsed document includes setting metadata, experience pairs, and a terminology database; For the code provisions of local seismic design codes in the parsed documents, they are atomized and decomposed according to "chapter-clause-sub-item", and mandatory provisions and applicable structural types are marked as metadata; For the review cases of building types in the parsed documents, the four elements of "project parameters - violation points - corresponding clauses - rectification suggestions" are extracted in a structured manner to form experience pairs for the model to learn from; The terminology database includes a mapping table of professional terms related to building types.

4. The method for reviewing seismic resistance reports based on a large language model with domain knowledge enhancement according to claim 3, characterized in that, In S1.5, the normative clauses obtained from S1.2 after data processing and annotation are stored in a knowledge graph format of four tuples in the form of subject, relation, sub-subject and object; The subject can have multiple relationships, and each relationship can correspond to multiple sub-sub ...

5. The method for reviewing seismic resistance reports based on a large language model with domain knowledge enhancement according to claim 2, characterized in that, In S1.3, the scope of review includes building classification, building use or building level, building size, and constraints; The building classification includes disaster prevention and relief buildings, public buildings, and residential buildings; The disaster prevention and relief buildings include medical buildings, disease prevention and control center buildings, disaster prevention and emergency command centers, emergency shelter buildings, fire truck garages and their duty rooms; The public and residential buildings mentioned include educational buildings, scientific research buildings, sports buildings, cultural and entertainment buildings, commercial buildings, office buildings, convention and exhibition buildings, museum buildings, archives, information center buildings, elderly care institutions, and children's welfare homes.

6. The method for reviewing seismic resistance reports based on a large language model with domain knowledge enhancement according to claim 2, characterized in that, In S1.4, the integrity review library is divided into two sub-libraries: design conditions and calculation results, and each review point R... i Mark t i and m i ; The t i Indicates the structural category of the project, where i represents one of the following: reinforced concrete frame structure, reinforced concrete frame-shear wall structure, reinforced concrete shear wall structure, reinforced concrete tube structure, steel frame structure, steel braced structure, steel frame-braced structure, steel frame-shear wall panel structure, and steel tube structure. The m i This indicates the project design method, where i represents one of the following: seismic design or seismic isolation and damping design.

7. The method for reviewing seismic resistance reports based on a large language model with domain knowledge enhancement according to claim 1, characterized in that, In S2, an open-source large language model is selected as the base to train the large language model in the field of seismic resistance of buildings in S1. Through fine-tuning training with instructions, sub-models for the review task are generated, including structural category classification model, design method identification model, review scope determination model, integrity determination model and compliance determination model. LoRA was used to inject seismic review question-and-answer pairs into the trained large language model, which was used to teach the large language model domain knowledge and reasoning patterns. Before generating any review results, RAG must first retrieve at least one normative clause or case from the vectorized database as a basis.

8. The method for reviewing seismic resistance reports based on a large language model with domain knowledge enhancement according to claim 2, characterized in that, S3 includes the following sub-steps: S3.

1. Data preprocessing: Python's natural language processing library is used to clean the text and table data of the review report. After preprocessing, the data is uniformly converted into a format suitable for input to a large language model. S3.

2. Using the structural category classification model, retrieve project information, determine the structural category, and match the annotations t in the integrity review database. i The review points; S3.

3. Identify the model through design methods, retrieve project information, determine the design method, and match it with the m marked in the integrity review library. i The provisions; S3.

4. Level 1 review: Through the review scope determination model, retrieve project information, automatically extract project information of the report to be reviewed, associate with the review scope database, and determine whether the project falls within the review scope. S3.

5. Secondary review, based on annotation t i and the annotation m i This triggers a condition, invokes the integrity determination model, and associates the integrity data marked with t in the integrity review library. i and the annotation m i Review point R i Search for R in the report i If relevant parameters are found, the process proceeds to Level 3 review (S3.6), as shown in the following formula: R i =R 1i +R 2i ={R1∈(t 1i ∩ m 1i ) + R2∈(t 2i ∩ m 2i )}; In the above formula, Ri refers to the review point in the review integrity database; R 1i This refers to the review points related to the integrity of the review library and design conditions; R 2i This refers to the review points related to the integrity of the database and the calculation results; t i This refers to the review points in the integrity review library that are marked with structural categories; t 1i This refers to review points in the integrity database that are relevant to the design conditions and have been marked with structural categories; t 2i This refers to review points that are related to the integrity of the review database and the calculation results, and that are marked with structural categories; m 1i This refers to review points in the integrity library that are related to the design conditions and have been marked with the design methodology. m 2i This refers to review points in the integrity database that are relevant to the calculation results and have been marked with the design method. If not retrieved, the review conclusion will output that this parameter is missing. S3.

6. Three-level review: Establish the correspondence between the parameters of the integrity assessment model and the compliance assessment model, and incorporate R... i The corresponding four-tuple relationship in the compliance judgment model knowledge base will be t i Corresponding to the main body of the quadruple, m i The remaining parameters correspond to the quadruple subclasses, and the quantized parameters correspond to the quadruple objects.

9. The method for reviewing seismic reports based on a large language model with domain knowledge enhancement according to claim 2, characterized in that, In S4, after reviewing based on the trained large language model of building seismic resistance, a Word version of the review conclusion is output, organized, and presented. For the compliant parts in the review report, compliance is marked and displayed in the background. For the problematic parts, the review conclusion outputs the location of the problem, the type of problem, and the specific seismic code clauses violated. At the same time, if there are mandatory clauses marked, the mandatory clauses are output and displayed in bold.

10. A large language model-based seismic report review device with domain knowledge enhancement, characterized in that: The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the domain knowledge-enhanced large language model seismic report review method as described in any one of claims 1 to 9.

Images