A risk-adaptive intelligent compiling method for construction full content type

Abstract

The present application belongs to the technical field of building construction informatization, especially a risk adaptive intelligent preparation method for construction full content type, comprising the following steps: S1: classifying the construction scheme content into text type, table type and picture type according to the two-dimensional classification of 'type-universality', and executing special preprocessing. The risk adaptive intelligent preparation method for construction full content type covers 'text + table + picture' through setting, automatically calculates the characteristic table parameters, realizes 80% efficiency improvement, provides drawing guidance for the characteristic pictures, reduces the manual supplementary workload by 90%, solves the limitation of the prior art that only handles text, achieves full content coverage and reduces manual supplementary effect.

Description

Technical Field

[0001] This invention relates to the field of information technology in building construction, and in particular to a risk-adaptive intelligent compilation method for all types of construction. Background Technology

[0002] The construction technical plan is a core document guiding the safe and efficient implementation of a project. Its content includes text descriptions, parameter tables, and professional drawings, and it must strictly adapt to the project's risk level. In the existing technology, patent application number 202510839037.1 published on the China Patent Network proposes an intelligent compilation scheme of "RAG + large model fine-tuning of text paragraphs," but it still suffers from the following unresolved technical bottlenecks:

[0003] Limited content processing scope: It only focuses on the intelligent generation of "text paragraphs" and does not cover the "parameter tables" (such as material requirement plans and equipment configuration tables) and "professional pictures" (such as construction site layout diagrams and process flow diagrams) that account for 30%-50% of the construction plan. A lot of manual supplementation is required, and the efficiency improvement is limited.

[0004] RAG retrieval lacks risk adaptation: It adopts a fixed cosine similarity retrieval logic and does not adjust the retrieval priority according to the risk level of the construction project (such as extremely high-risk deep foundation pits and first-level high-risk high-formwork). High-risk projects are prone to retrieving non-critical knowledge, and the accuracy rate of standard citation is only about 80%, which poses a safety hazard.

[0005] The large model has strong generalization ability for fine-tuning: the fine-tuning data cleaning only includes general steps such as "noise removal and segmentation", and has not been optimized for the characteristics of construction documents such as "dense terminology and many repeated parameter annotations"; the reward model only generalizes and compares "superior / inferior answers", and has not strengthened core construction requirements such as "accurate reference to standard clauses and completeness of risk parameters";

[0006] Lack of end-to-end support: There is no real-time intelligent Q&A function, and compilers need to manually consult the specifications when encountering technical problems (such as determining parameters for special working conditions); no large model selection and adaptation solution is provided, and construction companies (especially small and medium-sized enterprises) find it difficult to implement due to the high cost of deploying a single model;

[0007] Low knowledge update efficiency: No incremental update mechanism was designed. After a new standard is released or a good historical solution is added, the entire index library needs to be reconstructed, resulting in a long adaptation cycle and failing to meet the high-frequency update requirements of construction standards.

[0008] Existing technologies cannot meet the requirements for compiling construction plans that are "full-content, high-risk, low-cost, and quick-adaptable," and there is an urgent need for a risk-adaptive intelligent compilation method that is applicable to all types of construction content. Summary of the Invention

[0009] To address the technical problem that existing construction scheme compilation methods cannot meet the requirements of "full content, high risk, low cost, and rapid adaptation" in construction scheme compilation, this invention proposes a risk-adaptive intelligent compilation method for all types of construction content.

[0010] This invention proposes a risk-adaptive intelligent compilation method for all types of construction projects, comprising the following steps:

[0011] S1: Classify the construction plan content into text, table, and image categories according to the "type-generality" two-dimensional classification, and perform special preprocessing;

[0012] S2: Based on the project risk level classification, construct a dynamically weighted RAG index library, perform similarity search with risk factors, and output the top k search results;

[0013] S3: Perform in-depth customized fine-tuning of the basic large model in the construction field, including unsupervised fine-tuning specific to construction and supervised fine-tuning based on core construction requirements;

[0014] S4: Combine the preprocessed content vector, RAG search result vector, and risk level vector to form a context, generate text, table, and image content according to type, and perform compliance verification;

[0015] S5: Provides full-process multi-round intelligent question answering support, while also enabling incremental updates to the knowledge base.

[0016] Preferably, in step S1, the specific preprocessing includes: text preprocessing using a construction-specific word segmenter, with the word segmentation formula being...

[0017] Tokenizer const This is a construction-specific word segmenter based on BERT pre-training, where I is the input text and t is the input text. n For tokens containing construction terminology, i n For corpus indexing;

[0018] In table preprocessing, characteristic table parameters are calculated using formulas. Calculate, V struc For the structural volume, ρ mat For the material density, δ loss This is the construction loss coefficient;

[0019] The image preprocessing process includes drawing guidelines for outputting characteristic images, including requirements for size, scale, and key annotations.

[0020] Preferably, in step S2, the project risk level is divided into four levels, and each of the four levels corresponds to a knowledge graph weight W. K and the weights W of four unstructured document libraries. D ;

[0021] Similarity retrieval introduces risk factor γ risk The similarity score is set at 1.2 for top-tier, 1.0 for first-tier, 0.9 for second-tier, and 0.8 for third-tier. The corrected similarity formula is as follows: ;

[0022] Search results by Sim dyn Sort in descending order, taking the top 10 for Special Grade, the top 8 for Grade 1, and the top 6 for Grade 2 and below.

[0023] Preferably, in step S3, the loss function for construction-specific unsupervised fine-tuning is: λ=0.3 is the penalty coefficient for construction terminology, Dict const A dictionary of construction terminology;

[0024] The loss function of the reward model with supervised fine-tuning for core construction requirements is: K=5, y w The winning answer, which includes normative references, parameters, and risk measures, is y. l The answer that is inferior is the better one.

[0025] Preferably, in step S4, the context construction formula is: V pre For preprocessed content vectors, V RAG V is the vector of RAG search results. risk This represents a risk level vector.

[0026] After the content is generated, it must be verified that "the integrity of the standard references is 100%, the logic of the table parameters is correct, and the image format is compliant". Only if the verification pass rate is ≥98% can it be output.

[0027] Preferably, in step S5, the context weights of multi-round intelligent question answering are calculated according to... Calculate α=0.7, T=8 as the maximum number of dialogue rounds, and t as the current round;

[0028] The formula for incremental update of the knowledge base is: Vectorization and index addition are performed only on new data, and the consistency check conflict rate is ≤2%.

[0029] Preferably, the construction loss coefficient δ loss In this context, the steel reinforcement is 3%, the concrete is 2%, and the formwork is 5%; the resolution requirement for the characteristic image drawing guide is ≥300dpi.

[0030] Preferably, the risk level classification of the project is based on the "Regulations on the Classification and Control of Safety Risks in Construction":

[0031] The highest level is defined as a deep foundation pit ≥20m or a main tower height ≥200m; the first level is defined as a deep foundation pit 10-20m or a main tower height 100-200m; the second level is defined as a deep foundation pit 5-10m; and the third level is defined as a deep foundation pit <5m.

[0032] Preferably, the base model for the construction-specific unsupervised fine-tuning is Qwen2.5-7B, with training parameters of batch size 16, learning rate 2e-5, and number of iterations 50 rounds.

[0033] The labeled data for supervised fine-tuning was completed by 2 senior engineers and 3 editors with ≥8 years of experience. The training iterations were 30 rounds with a learning rate of 1e-5.

[0034] Preferably, the basic large model is selected from at least one of Qwen2.5, ByteDance, and Baidu Wenxin Yiyan;

[0035] The incremental update trigger conditions for the knowledge base are: within 7 days after the release of a new specification, within 5 days after the acceptance of a newly added excellent historical solution, and when a total of 30 expert experience summaries are accumulated.

[0036] The beneficial effects of this invention are as follows:

[0037] 1. By setting up coverage of "text + table + image", the parameters of the feature table are automatically calculated, improving efficiency by 80%. The feature image provides drawing guidance, reducing the workload of manual supplementation by 90%. This solves the limitation of existing technologies that only process text, achieving full content coverage and reducing the need for manual supplementation.

[0038] 2. By setting up risk-adaptive retrieval, the accuracy of the standard is improved. When used, the accuracy rate of standard reference for extremely risky projects reaches 98%, which is higher than the existing technology's 80%. This avoids the safety hazards caused by inaccurate retrieval due to the existing technology's 80% accuracy rate, and meets the requirements for construction risk management.

[0039] 3. By setting up real-time Q&A and low-cost implementation, the multi-round Q&A response time is less than 1 second, supporting small and medium-sized enterprises to choose low-cost models, solving the problems of lack of existing technical support and high costs. Through rapid knowledge updates and adaptation to standard iterations, the incremental update cycle of the knowledge base is shortened to 1 day, meeting the high-frequency update requirements of construction standards and improving the timeliness of the solution. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of a risk-adaptive intelligent compilation method for all types of construction content proposed in this invention;

[0041] Figure 2 This is a flowchart illustrating the steps of a risk-adaptive intelligent compilation method for all types of construction content proposed in this invention.

[0042] Figure 3 This is a diagram illustrating the dynamic weighted RAG retrieval principle of a risk-adaptive intelligent compilation method for all types of construction content proposed in this invention.

[0043] Figure 4 This invention presents a customized fine-tuning flowchart for a risk-adaptive intelligent compilation method for all types of construction projects, as proposed in this invention.

[0044] Figure 5 This is a flowchart of the full-content-type differentiation processing of a risk-adaptive intelligent compilation method for all construction content types proposed in this invention;

[0045] Figure 6 This invention proposes a risk-adaptive intelligent compilation method for all types of construction projects. Figure 1 Enlarged view of the structure at point A1;

[0046] Figure 7 This invention proposes a risk-adaptive intelligent compilation method for all types of construction projects. Figure 1 Enlarged view of the structure at point A2;

[0047] Figure 8 This invention proposes a risk-adaptive intelligent compilation method for all types of construction projects. Figure 2 Enlarged view of the structure at point B1;

[0048] Figure 9 This invention proposes a risk-adaptive intelligent compilation method for all types of construction projects. Figure 2 Enlarged view of the structure at point B2. Detailed Implementation

[0049] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0050] Reference Figures 1-9 A risk-adaptive intelligent compilation method for all types of construction projects includes the following steps:

[0051] S1: Classify the construction plan content into three categories based on "type-generality": text (general / specific), table (general / specific), and image (common / specific), and perform specific preprocessing.

[0052] In step S1, the specific preprocessing includes: text preprocessing using a construction-specific word segmenter, with the word segmentation formula being...

[0053] Tokenizerconst This is a construction-specific word segmenter based on BERT pre-training, where I is the input text and t is the input text. n For tokens containing construction terminology, i n For corpus indexing;

[0054] In table preprocessing, characteristic table parameters are calculated using formulas. Calculate, V struc For the structural volume, ρ mat For the material density, δ loss This is the construction loss coefficient;

[0055] The image preprocessing process includes drawing guidelines for outputting characteristic images, including requirements for size, scale, and key annotations.

[0056] Furthermore, the construction loss coefficient δ loss In the process, the steel reinforcement is 3%, the concrete is 2%, and the formwork is 5%; the resolution requirement for the characteristic image drawing guide is ≥300dpi, and the format is AutoCAD .dwg or PNG / JPG.

[0057] S2: Based on the project risk level classification, construct a dynamically weighted RAG index library, perform similarity search with risk factors, and output the top k search results;

[0058] In step S2, the project risk level is divided into four levels, and each of the four levels corresponds to a knowledge graph weight W. K And the weights W of four unstructured document libraries D .

[0059] The project risk level is divided into four levels: Extra High, Level 1, Level 2, and Level 3, corresponding to the knowledge graph weight W. K The weights W of the unstructured document library are 0.8, 0.6, 0.4, and 0.3 respectively. D The values ​​are 0.2, 0.4, 0.6, and 0.7 respectively.

[0060] Similarity retrieval introduces risk factor γ risk The similarity score is set at 1.2 for top-tier, 1.0 for first-tier, 0.9 for second-tier, and 0.8 for third-tier. The corrected similarity formula is as follows: ;

[0061] Search results by Sim dyn Sort in descending order, taking the top 10 for Special Grade, the top 8 for Grade 1, and the top 6 for Grade 2 and below.

[0062] Furthermore, the risk level classification of the project is based on the "Regulations on the Classification and Control of Safety Risks in Construction": Special level is deep foundation pit ≥20m or main tower height ≥200m, Level 1 is deep foundation pit 10-20m or main tower height 100-200m, Level 2 is deep foundation pit 5-10m, and Level 3 is deep foundation pit <5m.

[0063] S3: Perform in-depth customized fine-tuning of the basic large model in the construction field, including unsupervised fine-tuning specific to construction and supervised fine-tuning based on core construction requirements;

[0064] In step S3, the loss function for construction-specific unsupervised fine-tuning is: λ=0.3 is the penalty coefficient for construction terminology, Dict const A dictionary of construction terminology;

[0065] The loss function of the reward model with supervised fine-tuning for core construction requirements is: K=5, y w The winning answer, which includes normative references, parameters, and risk measures, is y. l The answer that is inferior is the better one.

[0066] Furthermore, the base model for the construction-specific unsupervised fine-tuning is Qwen2.5-7B, with training parameters of batch size 16, learning rate 2e-5, and 50 iterations. The labeled data for supervised fine-tuning was completed by 2 senior engineers and 3 editors with ≥8 years of experience, with 30 training iterations and a learning rate of 1e-5.

[0067] S4: Combine the preprocessed content vector, RAG search result vector, and risk level vector to form a context, generate text, table, and image content according to type, and perform compliance verification;

[0068] In step S4, the context construction formula is as follows: V pre For preprocessed content vectors, V RAG V is the vector of RAG search results. risk This represents a risk level vector.

[0069] After the content is generated, it must be verified that "the integrity of the standard references is 100%, the logic of the table parameters is correct, and the image format is compliant". Only if the verification pass rate is ≥98% can it be output.

[0070] S5: Provides full-process multi-round intelligent question answering support, while also enabling incremental updates to the knowledge base.

[0071] In step S5, the context weights of the multi-round intelligent question answering are determined according to... Calculate α=0.7, T=8 as the maximum number of dialogue rounds, and t as the current round;

[0072] The formula for incremental update of the knowledge base is: Vectorization and index addition are performed only on new data, and the consistency check conflict rate is ≤2%.

[0073] The basic model is selected from at least one of Qwen2.5, ByteDance, and Baidu Wenxin Yiyan; the incremental update trigger conditions for the knowledge base are: within 7 days after the release of the new specification, within 5 days after the acceptance of the newly added excellent historical solutions, and when 30 expert experience summaries are accumulated.

[0074] By setting up coverage of "text + table + image", the parameters of the feature table are automatically calculated, improving efficiency by 80%. The feature image provides drawing guidance, reducing the workload of manual supplementation by 90%. This overcomes the limitation of existing technologies that only process text, achieving full content coverage and reducing the need for manual supplementation.

[0075] By setting up risk-adaptive retrieval, the accuracy of the standards is improved. When used, the accuracy rate of standard citation for projects with extremely high risks reaches 98%, compared to 80% with existing technologies. This avoids safety hazards caused by inaccurate retrieval due to the 80% accuracy rate of existing technologies, and meets the requirements for construction risk management.

[0076] By setting up real-time Q&A and low-cost deployment, the system supports small and medium-sized enterprises to choose low-cost models (such as ByteDance's Doubao, which reduces deployment costs by 60%) through multi-round Q&A response time of less than 1 second. This solves the problems of lack of existing technical support and high costs. Furthermore, through rapid knowledge updates and adaptation to standard iterations, the incremental update cycle of the knowledge base is shortened to 1 day (compared to 7 days for existing technologies), meeting the high-frequency update requirements of construction standards and improving the timeliness of the solution.

[0077] Example 1

[0078] A risk-adaptive intelligent compilation method for all types of construction projects includes the following steps:

[0079] S1: Comprehensive construction content classification and specialized preprocessing. It receives user-input compilation instructions, categorizes construction plan content into a two-dimensional "type-generality" classification, and performs specialized preprocessing, covering all content types including text, tables, and images.

[0080] S11: Definition of two-dimensional classification:

[0081] Text-based documents are divided into general texts (such as quality assurance systems and acceptance standard clauses) and specific texts (such as key and difficult points in construction and process adaptation instructions).

[0082] Tables: These are divided into general tables (such as efficiency analysis tables and risk factor lists, with fixed formats) and characteristic tables (such as material requirement plans and equipment configuration tables, where parameters need to be calculated on a project-by-project basis).

[0083] Image type: divided into common images (such as schematic diagrams of protective piles and standard drawings of rain shelters, which can be reused) and characteristic images (such as schematic diagrams of engineering structures and construction site layout diagrams, which require professional drawing);

[0084] S12: Specialized Pre-processing Operations:

[0085] Text preprocessing: A word segmentation tool specifically designed for the construction field was used (based on BERT pre-training, with 100,000 construction texts as training data). The word segmentation formula is as follows:

[0086]

[0087] Tokenizer const This is a construction-specific word segmenter based on BERT pre-training, where I is the input text and t is the input text. n For tokens containing construction terms (such as "hydraulic climbing formwork" and "curing temperature difference"), i n For corpus indexing.

[0088] Table preprocessing: General tables use templates from a library with fixed formats; characteristic tables extract parameters and calculate factors (e.g., material requirement = structural volume × material density × loss coefficient), with the following formula: V struc For the structural volume (user input), ρ mat For material density (retrieved from knowledge base), δ loss This is the construction loss factor (e.g., 3% for steel reinforcement and 2% for concrete).

[0089] Image preprocessing: Common images are retrieved from the knowledge base (format PNG / JPG, resolution ≥300dpi); characteristic images generate "drawing guidelines" (including size, scale, key annotation requirements, such as "the layout diagram of the main tower sprinkler system needs to indicate the position of the windproof buckle").

[0090] S2: Risk-Adaptive Dynamic Weighted RAG Index Construction and Retrieval. Based on project risk levels, a dynamic weighted RAG retrieval system is constructed to improve the accuracy of high-risk project retrieval.

[0091] S21: Project Risk Level Classification and Weight Mapping

[0092] According to the "Regulations on the Classification and Control of Safety Risks in Construction Projects", project risks are classified into four levels: Special Level (e.g., deep foundation pit ≥ 20m, main tower height ≥ 200m), Level 1 (deep foundation pit 10-20m, main tower height 100-200m), Level 2 (deep foundation pit 5-10m), and Level 3 (deep foundation pit < 5m). The weighting mapping rules are as follows:

[0093] Special Grade 0.8 0.2 Regulatory Clauses > Historical Cases Level 1 0.6 0.4 Regulatory clauses = historical cases Level 2 0.4 0.6 Historical Cases > Regulatory Clauses Level 3 0.3 0.7 Primarily historical cases

[0094] S22: Dynamic Weighted RAG Retrieval Process:

[0095] Vector embedding: A pre-trained Sentence-BERT model (768 vector dimensions) from the construction domain was used. The embedding formula is as follows:

[0096] E const This is a construction-specific embedded model, where T is the token sequence after word segmentation.

[0097] Index Construction and Storage: Historical solutions are divided into blocks according to "chapter-risk level" (e.g., "deep foundation pit support - extremely high risk"), and stored in the Chroma vector database after vectorization. The index construction formula is as follows:

[0098] V c,risk Risk is an index vector with risk labels. tag For risk level labels (such as "Extraordinary").

[0099] Dynamic weighted retrieval: combining "content vector + risk level" retrieval, similarity calculation incorporates the risk factor γ. risk (For Grade A, use 1.2; for Grade B, use 1.0; for Grade C, use 0.9; for Grade D, use 0.8). The corrected formula is:

[0100] Press Sim dyn Sort in descending order and retrieve the first k results (k=10 for top-tier, k=8 for first-tier, and k=6 for second-tier and below). The formula is:

[0101] .

[0102] S3: Deeply customized large-scale model fine-tuning in the construction field. Addressing the characteristics of strong dependence on construction specifications and strict parameter requirements, a "two-stage + two-constraint" fine-tuning strategy is designed.

[0103] S31: Phase 1: Unsupervised fine-tuning for construction (adapting to language habits)

[0104] Data screening: Select 5,000 excellent construction plans from the past 5 years, 300 current national / industry standards (such as GB50204 and JGJ276), and 200 experience summaries of senior engineers (with ≥10 years of experience).

[0105] Specialized data cleaning:

[0106] a1: Remove noise specific to construction documents (such as repeated "approval comments" and "headers and footers");

[0107] a2: Extract core process parameters (such as "curing time ≥ 14 days" and "wind resistance level ≥ 6") to form a parameter dictionary;

[0108] a3: Standardized construction terminology (e.g., unifying "climbing formwork" as "hydraulic climbing formwork" and "concrete" as "concrete");

[0109] Fine-tuning training: Using Qwen2.5-7B as the base model, the training task was "continuing construction text writing". A construction term penalty term was added to the loss function (to avoid generating non-standard terms). The formula is as follows: ;

[0110] Where λ=0.3 is the penalty coefficient, Dictconst is the dictionary of construction terms, and I(·) is the indicator function (1 for non-terms, 0 otherwise); training parameters: batch size 16, learning rate 2e-5, number of iterations 50 rounds.

[0111] S32: Second phase: Supervision and fine-tuning of core construction requirements (strengthening professionalism);

[0112] Data construction: 10,000 "Construction Questions-Answers" samples were annotated by 2 construction experts (senior engineers) + 3 senior compilers (≥8 years of experience). Annotation standards:

[0113] The answer must include "the standard clause number (such as Clause 7.4.1 of GB50204-2015) + technical parameters + risk control measures";

[0114] Prioritize high-risk scenario samples (such as "dewatering parameters for extremely high-risk deep foundation pits");

[0115] Reward Model (RM) Training: The evaluation metrics are "canonical reference integrity (weight 0.4), parameter accuracy (0.3), and risk measure integrity (0.3)". The loss function formula is as follows: Where C(K,2) is the number of combinations of choosing 2 from K=5 answers, and y w For the winning answer (meeting 3 criteria), y l For the inferior answer (missing ≥1 indicator), r θ (·) represents the reward model score, and σ is the sigmoid function;

[0116] Fine-tuning and optimization: Based on the InstructGPT framework, the unsupervised model is fine-tuned using labeled samples, with 30 training iterations and a learning rate of 1e-5.

[0117] S4: Full content type scheme generation and differentiation processing, integrating RAG search results and fine-tuning the model to generate a full content scheme of "text-table-image" and perform differentiated adaptation:

[0118] S41: Context Construction: Concatenate "preprocessed content vector + RAG search result vector + risk tag", the formula is: V pre For preprocessed content vectors, V RAG V is the vector of RAG search results. risk This is a risk level vector (e.g., the highest risk level is [1,0,0,0]).

[0119] S42: Generating content by category:

[0120] Text generation: Call the construction-specific prompt word template (example: "Role: Senior Bridge Engineer; Goal: Generate text for the concrete curing of the main tower with special risk; Constraints: including curing time, temperature difference control, wind protection measures, and one standard reference for each point"), input the fine-tuning model, and generate parameters: temperature coefficient 0.2 (to ensure stability), maximum length 1000 characters, and repetition penalty 1.3;

[0121] Table generation: General table filling template; Featured tables based on S12 P cal Automatically calculates parameters and generates Markdown format (supports Excel export);

[0122] Image generation: Common images can be directly inserted; for characteristic images, output "drawing guide + blank template" (such as AutoCAD format .dwg) to prompt the user to draw and then import.

[0123] S43: Content Validation: Automatically validate “completeness of standard references (100% required), logical consistency of table parameters (e.g., material requirement > 0), and compliance of image formats (resolution ≥ 300dpi)”. Output is only allowed if the validation pass rate is ≥ 98%.

[0124] S5: End-to-end multi-round intelligent question answering and incremental knowledge base updates, providing real-time technical support and efficient knowledge updates, solving the problems of lack of existing technical support and slow updates.

[0125] S51: Multi-round intelligent question answering (adapted to construction technical issues):

[0126] Problem Analysis: Using a construction intent recognition model (trained with 20,000 construction problems), we analyze "core entities (such as climbing formwork systems), problem types (such as safety thresholds), and risk levels (such as extreme risk)".

[0127] Context management: Redis is used to cache the last 8 rounds of dialogue, and the context weight is calculated using an exponential decay formula: , where α=0.7 (recent dialogues have higher weight), T=8 (maximum round), and t is the current round.

[0128] Answer generation: Based on the dynamic weight retrieval results of S2, an answer containing "conclusion + standard reference + case reference" is generated (Example: "The climbing formwork operation should be stopped when the wind level is level 6 or above (referencing Article 5.3.4 of JGJ276-2012), and referring to a special-risk main tower scheme in 2023 (No. QT-2023-05), the operation should be stopped when the instantaneous wind speed is ≥10.8m / s").

[0129] S52: Incremental update of knowledge base:

[0130] Update triggers: within 7 days of the release of a new standard, within 5 days of the acceptance of a newly added excellent historical solution, or when a total of 30 expert experience summaries are accumulated;

[0131] Incremental indexing: Only vectorization and index addition are performed on new data, without requiring a full database reconstruction. The formula is: Among them, Index old For the existing index, V new For the new data vector, Risk new New data risk labels;

[0132] Consistency check: Check for conflicts between new data and existing knowledge (such as changes in standard parameters). The principle of "new standard first" is adopted, and the update can be completed only if the conflict rate is ≤2%.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A risk-adaptive intelligent compilation method for all types of construction projects, characterized in that, Includes the following steps: S1: Classify the construction plan content into text, table, and image categories according to the "type-generality" two-dimensional classification, and perform special preprocessing; S2: Based on the project risk level classification, construct a dynamically weighted RAG index library, perform similarity search with risk factors, and output the top k search results; S3: Perform in-depth customized fine-tuning of the basic large model in the construction field, including unsupervised fine-tuning specific to construction and supervised fine-tuning based on core construction requirements; S4: Combine the preprocessed content vector, RAG search result vector, and risk level vector to form a context, generate text, table, and image content according to type, and perform compliance verification; S5: Provides full-process multi-round intelligent question answering support, while also enabling incremental updates of the knowledge base; In step S1, the specific preprocessing includes: text preprocessing using a construction-specific word segmenter, with the word segmentation formula being... Tokenizer const This is a construction-specific word segmenter based on BERT pre-training, where I is the input text and t is the input text. n For tokens containing construction terminology, i n For corpus indexing; In table preprocessing, characteristic table parameters are calculated using formulas. Calculate, V struc For the structural volume, ρ mat For the material density, δ loss This is the construction loss coefficient; Drawing guidelines for outputting characteristic images in image preprocessing, including requirements for size, scale, and key annotations; In step S2, the project risk level is divided into four levels, and each of the four levels corresponds to a knowledge graph weight W. K and the weights W of four unstructured document libraries. D ; Similarity retrieval introduces risk factor γ risk The similarity score is set at 1.2 for top-tier, 1.0 for first-tier, 0.9 for second-tier, and 0.8 for third-tier. The corrected similarity formula is as follows: ; Search results by Sim dyn Sort in descending order; In step S3, the loss function for construction-specific unsupervised fine-tuning is: λ=0.3 is the penalty coefficient for construction terminology, Dict const A dictionary of construction terminology; The loss function of the reward model with supervised fine-tuning for core construction requirements is: K=5, y w The winning answer, which includes normative references, parameters, and risk measures, is y. l The inferior answer; In step S4, the context construction formula is as follows: V pre For preprocessed content vectors, V RAG V is the vector of RAG search results. risk This represents a risk level vector. After content is generated, it must be verified that "the integrity of the standard references is 100%, the logic of the table parameters is correct, and the image format is compliant". Only if the verification pass rate is ≥98% can it be output. In step S5, the context weights of the multi-round intelligent question answering are determined according to... Calculate α=0.7, T=8 as the maximum number of dialogue rounds, and t as the current round; The formula for incremental update of the knowledge base is: Vectorization and index addition are performed only on new data, and the consistency check conflict rate is ≤2%.

2. The risk-adaptive intelligent compilation method for all types of construction work according to claim 1, characterized in that: The construction loss coefficient δ loss In this context, the steel reinforcement is 3%, the concrete is 2%, and the formwork is 5%; the resolution requirement for the characteristic image drawing guide is ≥300dpi.

3. The risk-adaptive intelligent compilation method for all types of construction work as described in claim 1, characterized in that: The risk level classification of the project is based on the "Regulations on the Classification and Control of Safety Risks in Construction". The highest level is defined as a deep foundation pit ≥20m or a main tower height ≥200m; the first level is defined as a deep foundation pit 10-20m or a main tower height 100-200m; the second level is defined as a deep foundation pit 5-10m; and the third level is defined as a deep foundation pit <5m.

4. The risk-adaptive intelligent compilation method for all types of construction work as described in claim 1, characterized in that: The base model for the construction-specific unsupervised fine-tuning is Qwen2.5-7B, with training parameters of batch size 16, learning rate 2e-5, and number of iterations 50 rounds. The labeled data for supervised fine-tuning was completed by 2 senior engineers and 3 editors with ≥8 years of experience. The training iterations were 30 rounds with a learning rate of 1e-5.

5. The risk-adaptive intelligent compilation method for all types of construction work according to claim 1, characterized in that: The incremental update trigger conditions for the knowledge base are: within 7 days after the release of a new specification, within 5 days after the acceptance of a newly added excellent historical solution, and when a total of 30 expert experience summaries are accumulated.

Images