A multi-dimensional quality-improving construction scheme document generation method

Abstract

The present disclosure relates to a multi-dimensional quality improvement construction scheme document generation method. The method comprises the following steps: obtaining construction demand data; establishing a construction scheme document automatic generation model; training the construction scheme document automatic generation model; inputting the construction demand data into the trained construction scheme document automatic generation model, and outputting a construction scheme document by the construction scheme document automatic generation model. The construction scheme document automatic generation model comprises at least a proximal policy optimization model, and a reward function is set in the proximal policy optimization model. According to the technical scheme, the construction scheme document can be automatically generated by using the construction scheme document automatic generation model. The quality of the construction scheme document is improved in multiple dimensions.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of building technology, in particular to a construction scheme document generation method for improving quality in multiple dimensions. BACKGROUND

[0002] In the construction process, the preparation of the construction special scheme is a crucial link. In the traditional preparation of the construction special scheme, although some artificial intelligence technologies are used, such as using a general large language model (LLM) to automatically generate documents. However, due to the large vocabulary and huge quantity of construction documents in the construction field, the quality of the generated construction documents is low, and contradictions often occur in the documents. For example, the building area determined in the "project profile" chapter changes in the "material plan" chapter, which is contradictory. Grammatical errors and ambiguities often occur in the documents, seriously affecting the quality of the documents.

[0003] With the rapid development of the construction industry and the increasing complexity of engineering projects, higher requirements are placed on the quality of the construction special scheme. Therefore, there is an urgent need to develop a high-quality construction special scheme preparation method. SUMMARY

[0004] To overcome the problems in the related art, the present disclosure provides a construction scheme document generation method for improving quality in multiple dimensions to solve the problem of low efficiency and poor quality of manually sorting construction scheme documents in the related art.

[0005] According to a first aspect of an embodiment of the present disclosure, a construction scheme document generation method for improving quality in multiple dimensions is provided, comprising:

[0006] obtaining construction requirement data;

[0007] establishing a construction scheme document automatic generation model;

[0008] training the construction scheme document automatic generation model;

[0009] inputting the construction requirement data into the trained construction scheme document automatic generation model, and the construction scheme document automatic generation model outputs a construction scheme document;

[0010] The construction scheme document automatic generation model at least includes a proximal policy optimization model;

[0011] In the proximal policy optimization model, a reward function is set, and the reward function calculation formula is as follows:

[0012] R_total =α×R_logic +β×R_consistency +γ×R_compliance +δ×R_quality;

[0013] R_logic is a multi-dimensional logical association reward, where the multi-dimensional factors include schedule matching degree, material usage deviation degree, and cost.

[0014] R_consistency is the cross-paragraph consistency reward;

[0015] R_compliance is the specification compliance reward;

[0016] R_quality is a reward for text quality and diversity.

[0017] α is the multidimensional logical association reward weight;

[0018] β is the cross-segment consistency reward weight within the scheme;

[0019] γ is the compliance reward weight;

[0020] δ represents the reward weight for text quality and diversity.

[0021] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects:

[0022] Compared with existing technologies, the technical solution of this application utilizes an automatic construction plan document generation model to automatically generate construction plan documents. This significantly improves the generation efficiency and quality of construction plan documents. Notably, by employing a near-end strategy optimization model, which includes a reward function, this application's reward function comprises multi-dimensional logical association rewards, cross-paragraph consistency rewards, standard compliance rewards, and text quality and diversity rewards, each with corresponding weights. Multi-dimensional logical association rewards tightly bind information from different dimensions within the document. Cross-paragraph consistency rewards reduce inconsistencies within the document, ensuring a unified stance, information, and viewpoint throughout, resulting in a more coherent reading experience. Standard compliance rewards ensure the document adheres more closely to pre-defined rules or standards. Text quality and diversity rewards improve basic text fluency and accuracy, reducing grammatical errors and ambiguities; they also prevent monotonous repetition by providing richer examples and argumentation methods, avoiding the repetition of the same expressions or cases, thus enhancing document readability. This contributes to improving the quality of construction documents.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0025] Figure 1 This is a flowchart illustrating a method for generating and compiling construction plan documents with multi-dimensional quality improvement, according to an exemplary embodiment. Detailed Implementation

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0027] It should be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0028] With the rapid development of the construction industry and the increasing complexity of engineering projects, higher demands are being placed on the quality and efficiency of construction-specific plans. Therefore, there is an urgent need to develop an efficient, accurate, and up-to-date method for preparing construction-specific plans. The rapid development of artificial intelligence technology offers new approaches and possibilities for solving this problem.

[0029] Based on this, this application proposes a method for generating and compiling construction plan documents with multi-dimensional quality improvement, see appendix. Figure 1 The method includes the following steps:

[0030] In step S102, construction requirement data is obtained.

[0031] In this embodiment, the construction requirements data can be engineering parameters, including building area and construction period. For example, one set of engineering parameters includes: a frame structure residential building with a building area of ​​20,000 square meters and a construction period of 300 days.

[0032] In step S104, an automatic model for generating construction plan documents is established.

[0033] In step S106, the automatic generation model for the construction plan document is trained.

[0034] In step S108, the construction requirement data is input into the trained construction plan document automatic generation model, and the construction plan document automatic generation model outputs the construction plan document.

[0035] In this embodiment, the construction plan document automatically generates a model, which can automatically fill in the construction process by matching the knowledge base with the received construction requirement data, such as the construction process of raft foundation combined with aluminum formwork construction.

[0036] In this embodiment, one approach, if the knowledge base is structured data and the process matching logic is clear, is to employ a knowledge graph and a semantic similarity model. The knowledge graph stores process knowledge and specifically contains numerous nodes and edges. Each node stores parameters and process steps, while each edge stores conditional dependencies. The semantic similarity model is used to calculate the matching degree between the input parameters and the "conditional parameters" in the knowledge base, quickly locating the most relevant process rules. This semantic similarity model can be the BERT model.

[0037] In some embodiments, if the knowledge base consists of unstructured data (such as historical process documents and case texts), and if process knowledge is scattered across documents and cases (without explicit rules, requiring logical extraction from the text), then it is necessary to learn the implicit patterns of "parameters to processes" from the text. A text generation model can be used to output the corresponding process step text based on input parameters (such as "project type = bridge, material = steel structure, environment = high temperature"). The text generation model can be, for example, BART or T5.

[0038] Furthermore, a combination of classification and retrieval can be used. First, a classification model (such as XGBoost or BERT) is used to determine the type of process scenario based on the received parameters, such as the process scenario of "concrete construction of high-rise buildings". Then, the most matching process details are retrieved from the historical cases of this scenario.

[0039] The classification model can be XGBoost or BERT. Retrieval can employ a combination of keyword and dense retrieval. Dense retrieval models can include Sentence-BERT and CoSENT. By converting parameter text (e.g., "frame column C40 summer construction") and document fragments into vectors, vector similarity is calculated to find the most matching document.

[0040] When using keywords, the BERT-KPE model can be employed to quickly filter irrelevant documents. First, keyword retrieval narrows the scope, then a dense retrieval model precisely matches the most relevant case documents for subsequent model generation.

[0041] Resource allocation, such as generating labor force curves, machine schedules, and project schedules, such as generating Gantt charts.

[0042] In this embodiment, taking the automatic generation of Gantt charts as an example, the core of a Gantt chart is the sequential relationship of "process-time" (e.g., process A is on days 1-5, process B is on days 6-10). Essentially, it's a constrained sequence generation problem, which can be addressed using machine learning models capable of handling structured constraints and sequential dependencies, such as the Transformer model. The Transformer model learns the mapping between "input parameters and schedule / resource plan" from historical project data, making it particularly suitable for scenarios with relatively flexible processes and rules that are difficult to exhaustively enumerate. The Transformer model can capture long-distance process dependencies (e.g., "process C must start after A and B are completed"), taking project parameters (workload, time requirements, etc.) as input, and outputting the start / end time series of processes.

[0043] In some embodiments, taking labor curves and machine schedules as examples, the labor curves and machine schedules need to consider total resource constraints (such as "no more than 50 workers per day"), which are constrained optimization problems. They can be adjusted using a combination of machine learning prediction and optimization algorithms. First, gradient boosting trees (XGBoost / LightGBM) or neural networks can be used to predict the resource requirements (such as "how many workers / machines are needed for process A") and time consumption for a single process. Then, optimization algorithms (such as genetic algorithms or simulated annealing) are used to adjust the allocation based on the prediction results and resource constraints (total resource limits, time conflicts), outputting a conflict-free resource plan.

[0044] For intelligent typesetting, tools like Aspose.Words can be used, or a document structure understanding model and a typesetting generation model can be employed to automate formatting. The document structure parsing model can utilize the LayoutLM model, while the typesetting generation model can employ a Transformer-based format generation model.

[0045] In some embodiments, the automatic generation model for construction plan documents includes at least a proximal strategy optimization model, wherein the proximal strategy optimization model is a machine learning algorithm. The proximal strategy optimization model includes a reward function, the calculation formula of which is as follows:

[0046] R_total =α×R_logic +β×R_consistency +γ×R_compliance +δ×R_quality.

[0047] R_logic is a multi-dimensional logical association reward, where the multi-dimensional factors include schedule matching degree, material usage deviation degree, and cost.

[0048] R_consistency is the cross-paragraph consistency reward;

[0049] R_compliance is the specification compliance reward;

[0050] R_quality is a reward for text quality and diversity.

[0051] α is the multidimensional logical association reward weight;

[0052] β is the cross-segment consistency reward weight within the scheme;

[0053] γ is the compliance reward weight;

[0054] δ represents the reward weight for text quality and diversity. α, β, γ, and δ can be adjusted according to different engineering types (such as civil engineering, mechanical and electrical engineering, and decoration).

[0055] The technical solution of this application utilizes an automatic construction plan document generation model to automatically generate construction plan documents. This significantly improves the generation efficiency of construction plan documents and enhances their quality from multiple dimensions. Notably, by employing a near-end strategy optimization model, a reward function is set within the model: R_total = α×R_logic + β×R_consistency + γ×R_compliance + δ×R_quality. R_logic represents a multi-dimensional logical association reward, where the multiple dimensions include schedule matching degree, material usage deviation degree, and cost. R_consistency represents a cross-paragraph consistency reward. R_compliance represents a specification compliance reward. R_quality represents a text quality and diversity reward. The reward function of this application includes multi-dimensional logical association rewards, cross-paragraph consistency rewards, specification compliance rewards, and text quality and diversity rewards, each adjusted with its corresponding weights. The multi-dimensional logical association reward allows for tighter binding of information from different dimensions within the document, while the cross-paragraph consistency reward reduces inconsistencies within the document. Maintaining consistency in the document's stance, information, and viewpoints makes it more coherent to read. Compliance with standards and norms rewards: This ensures the document adheres more closely to pre-defined rules or standards. Text quality and diversity rewards: On one hand, it improves basic text fluency and accuracy, reducing grammatical errors and ambiguity; on the other hand, it avoids monotonous and repetitive content, for example, by using more diverse examples and argumentation methods, avoiding the repetition of the same expressions or cases, thus increasing document readability. This is beneficial for improving the quality of construction documents.

[0056] In some embodiments, the above method further includes: updating the specifications stored in the database in a timely manner.

[0057] Specifically, Scrapy crawlers are used for real-time monitoring of updates to national and industry standards. After a new standard is released, revised clauses are automatically parsed, outdated content in the knowledge base is marked, and update logs are generated for reference in the solution, thus achieving intelligent synchronization.

[0058] After the specifications stored in the database are updated, the updated specifications in the updated database are used to retrain the automatic generation model of the above-mentioned construction plan documents in order to improve the generation accuracy.

[0059] In some embodiments, R_logic = w1×f_d + w2×f_m + w3×f_c;

[0060] Where f_d represents the project schedule matching degree;

[0061] f_m represents the deviation in material usage;

[0062] f_c represents the accuracy of cost calculation;

[0063] w1 is the weight of the project schedule matching degree;

[0064] w2 is the weighting value for the deviation in material usage;

[0065] w3 is the weight for the accuracy of cost calculation.

[0066] In this embodiment, the dynamic knowledge base query and calculation of this sub-reward is used to evaluate the inherent logical rationality among key engineering parameters in the solution. This goes beyond preset static rules, performing dynamic calculations by querying an external knowledge base.

[0067] The calculation process is as follows: Entity and numerical extraction: Using the Named Entity Recognition (NER) model and regular expressions, key entities and their numerical values ​​are extracted from the generated scheme text S, such as the quantity set V={v1:v_num1, v2:v_num2,...}, the construction period D generated by the trained construction scheme document automatic generation model, the equipment and labor set E={e1:e_num1, e2:e_num2,...}, and the material set M={m1:m_num1, m2:m_num2,...}.

[0068] Knowledge base query: Use the extracted entities (such as "pumping" and "C30 concrete") as keywords to query the engineering knowledge graph or professional database to obtain their dynamic attributes, such as equipment, daily labor capacity P_e, material content R_m, unit cost C_m, and efficiency adjustment coefficients Weather and Season (considering dynamic factors such as seasons and weather).

[0069] In some embodiments, f_d = sigmoid(-k|D-(D_min+D_max) / 2| / (D_max-

[0070] D_min) / 2);

[0071] Where k is the sensitivity coefficient;

[0072] D represents the construction period generated by the automatically generated model of the trained construction plan document;

[0073] D_min is the theoretical minimum construction period;

[0074] D_min=(Σ(P_v×num_m)) / (Σ(P_e×num_e))×Weather×Season;

[0075] Where P_v is the predicted quantity of work corresponding to a certain sub-item of the project;

[0076] num_m represents the unit material quantity corresponding to a certain project quantity;

[0077] P_e represents the average daily productivity of equipment and labor.

[0078] num_e represents the number of units of equipment or labor corresponding to a certain material;

[0079] Weather is a weather adjustment factor;

[0080] Season is a seasonal adjustment factor;

[0081] D_max is the theoretical maximum construction period.

[0082] In some embodiments, f_m = max(0, 1 - |M_extracted - M_calc| / M_calc);

[0083] Where M_extracted represents the material usage generated by the trained construction plan document automatic generation model.

[0084] M_calc represents the theoretical material usage;

[0085] M_calc = P_v×R_m;

[0086] f_c = ;

[0087] in, To calculate the total cost; This represents the theoretical total cost.

[0088] Cost_calc =Σ(M×C_m)+Σ(equipment shift cost);

[0089] Where M represents the theoretical quantity of a certain sub-item of the project;

[0090] C_m represents the unit cost of a certain amount of work;

[0091] In some embodiments, R_consistency = average(Sim_ij) - λ×(Num_contraction);

[0092] Sim_ij = cosine_sim(F_i, F_j);

[0093] Where λ is the penalty coefficient; it can be preset.

[0094] Num_contraction represents the number of inconsistent paragraphs;

[0095] F_i is the core engineering feature vector of the i-th paragraph;

[0096] F_j is the core engineering feature vector of the j-th paragraph;

[0097] Sim_ij is the cosine similarity between F_i and F_j;

[0098] average(Sim_ij) is the average cosine similarity of the core engineering feature vectors of any two paragraphs in all paragraphs.

[0099] In this embodiment, semantic vector consistency verification is performed. This sub-reward is used to detect semantic contradictions between the preceding and following text, and between the summary and details, thus resolving the common "self-conflict" problem in generated text.

[0100] Calculation process: paragraph segmentation and feature extraction: the generated scheme S is divided into multiple paragraphs [sec1, sec2, ..., secN] according to chapters or semantics.

[0101] Engineering Feature Identification: Using a fine-tuned model or prompt word engineering, extract a core engineering feature vector F_i for each paragraph (e.g., {"Structure Type": "Frame Shear Wall", "Construction Period": "180 Days", "Main Material": "Ready-Made Concrete"}). This feature vector is a high-level summary of the paragraph.

[0102] Consistency check: Traverse all paragraph pairs (i, j).

[0103] Calculate the cosine similarity of the feature vectors: Sim_ij = cosine_sim(F_i, F_j). If the cosine similarity Sim_ij is below the threshold T_low, it is determined that the two descriptions may refer to different projects, triggering a contradiction detection: input the two original texts into the discriminant model to determine whether there is a factual contradiction (e.g., one says to use process A, and the other says to use process B). If a contradiction exists, a significant penalty is imposed. Simultaneously, check the consistency of key parameter propagation: for example, the "building area" determined in the "Project Overview" section should not be changed in the "Materials Plan" section.

[0104] In some embodiments, each core paragraph and its engineering feature vector are determined;

[0105] Determine the embedding vector of the core paragraph;

[0106] Using the engineering feature vectors, suitable clauses are retrieved from the standard clause library to obtain the first clause set;

[0107] Calculate the similarity between the embedded vector and all the text vectors in the standard text library, and obtain the texts with the highest similarity to obtain the second text set.

[0108] The first set of articles and the second set of articles are merged, duplicate articles are removed, and a set of standard articles applicable to the core paragraph is constructed.

[0109] A pre-trained semantic contradiction discrimination model is used to pair the core paragraph with each normative provision in the set of normative provisions to determine the number of provisions that semantically contradict the core paragraph.

[0110] The compliance reward R_compliance is determined based on the number of clauses that contradict the semantics of the core paragraph.

[0111] R_compliance calculation: R_compliance = - (number of clauses that contradict the semantics of the core paragraph × severity weight).

[0112] The severity score is weighted between 0.1 and 0.25.

[0113] In some embodiments, R_quality = w4×f_readability + w5×f_terminology;

[0114] Where f_readability is the readability coefficient;

[0115] w4 is the readability coefficient weight;

[0116] f_terminology = 1 - (NUM_error terms / NUM_total terms);

[0117] Where NUM_error terms is the number of error terms;

[0118] NUM_Total Terms refers to the total number of terms;

[0119] f_terminology incorrect terminology reverse weighting;

[0120] w5 represents the reverse weighting of incorrect terms.

[0121] In this embodiment, the sub-reward inherits the traditional text generation evaluation metrics, but is optimized for engineering text.

[0122] Calculation content: Readability score: Evaluates whether the text is easy to understand using formulas such as Flesch.

[0123] Domain terminology accuracy: By comparing with a domain terminology dictionary, the correct use of professional terms is rewarded, and neologisms or incorrect word usage is penalized.

[0124] Formatting compliance: Check whether necessary chapters are included and whether the numbering of figures and tables is consistent.

[0125] Readability and standardization check: Substitute into the formula RE = 206.835 - 1.015 × (number of words / number of sentences) - 84.6 × (number of syllables / number of words).

[0126] The score is compared with the target range (e.g., 50-60) to obtain the sub-score f_readability.

[0127] Specifically, f_readability: If RE = 55 (perfect hit), then f_readability = 1.0.

[0128] If RE = 60 or 50 (at the target boundary), then f_readability = 0.5.

[0129] If RE = 65 or 45 (out of range), then f_readability = 0.0.

[0130] If RE goes further out of range, the score remains 0.

[0131] Terminology accuracy check. Check the text for errors in the domain stop word list or inappropriate terminology usage.

[0132] In some embodiments, after the construction plan document is automatically generated and output as a model, the method may further include the following steps: reviewing the construction plan document, specifically including the following steps:

[0133] Submit the proposal to the system for review.

[0134] The construction plan document is divided into several semantically complete paragraphs using prompt words.

[0135] In this embodiment, the number of words in each paragraph can be controlled between 2,000 and 5,000.

[0136] Each section of content is characterized.

[0137] In this embodiment, the feature identifiers include building area, structural type, and construction process.

[0138] For each paragraph, further semantic analysis is performed, using vector similarity and engineering feature identifiers to find similar specification requirements.

[0139] Use a large language model to determine whether there is a contradiction between the semantics of the construction plan document and the standard semantics.

[0140] Perform similarity matching on the feature data represented by all paragraphs, analyze data consistency, and check for any contradictions.

[0141] In this embodiment, the retrieved original text of the specifications and the original text of the scheme paragraphs are used as prompt words and analyzed by the large model.

[0142] If the content described in the original specification differs from the content in the original scheme paragraph in terms of data scope and semantic conformity, the large model will determine and provide contradictory data feedback.

[0143] For example: The original design document states that the roof slope should be 10%, while the standard requires that the roof slope should not exceed 5%. The large model will then determine that there is a contradiction and return the following result:

[0144] {“Original text: The roof slope is 10%;}

[0145] "Standard": The slope of the roof should not exceed 5%;

[0146] "Contradiction": Value out of range}.

[0147] In some embodiments, the further semantic analysis of each paragraph, using vector similarity and engineering feature identifiers to find similar specification requirements, may further include the following steps:

[0148] For any paragraph, calculate the original text vector in that paragraph.

[0149] In this embodiment, word2vec or BERT models can be used to convert the original text in the above paragraph into original text vectors.

[0150] Calculate the cosine similarity between the original text vector and the data vectors in the vector library.

[0151] Identify data in the vector library that meet the cosine similarity threshold.

[0152] In this embodiment, the cosine similarity threshold can be flexibly set, for example, to 0.85. If the cosine similarity between the original text vector and the data vector in the vector library is greater than 0.85, the data in the vector library is determined to be similar data. The data in the aforementioned vector library is then identified.

[0153] The feature identifiers of the original text are extracted using a large model.

[0154] The database is searched based on the feature identifiers of the original text to obtain standardized data that conforms to the feature identifiers.

[0155] The standardized data that conforms to the feature identifier and the data in the vector library that conforms to the cosine similarity threshold are merged.

[0156] Return the data after the above merging process.

[0157] In this embodiment, the returned result is as follows:

[0158] [{“Standard Name”: Concrete Construction Quality Acceptance Standard;

[0159] "Standard Number": XXXXXX;

[0160] "Original text of the standard": XXXXXXX};

[0161] {“Standard Name”: Standard for Acceptance of Construction Quality of Masonry Engineering}

[0162] "Standard Number": XXXXXX;

[0163] “Original text of the standard”: XXXXXXX}).

[0164] In some embodiments, step S106, training the automatic generation model of the construction plan document, may further include the following steps:

[0165] Acquire and preprocess experience data, specifically including:

[0166] Using web scraping technology, we regularly download various engineering specifications, historical plans, and design drawings from the official website of the Ministry of Housing and Urban-Rural Development.

[0167] In this embodiment, a crawler can be written using Scrapy to periodically access the official website of the Ministry of Housing and Urban-Rural Development and download more than 200 current standards, such as GB50204 "Code for Acceptance of Construction Quality of Concrete Structures".

[0168] In this embodiment, information is extracted from the design drawings. The CAD drawings are parsed using the Autodesk Forge API to extract parameters such as "foundation type", "story height" and "reinforcement specifications", and then converted into JSON format.

[0169] In some embodiments, data preprocessing and data fetching during specification updates employ the following steps:

[0170] 1. Start the service regularly.

[0171] 2. Download the HTML file for the designated website.

[0172] 3. Use the Scrapy open-source library to analyze web page content and obtain the name and version information of the specifications within the website.

[0173] 4. Compare the website's version with the specifications and versions in the database. If the website version is newer, analyze the download address.

[0174] 5. Download the specifications to your local machine, call the data entry module, and enter the specifications data into the database.

[0175] 6. User-uploaded solution data is directly stored using the data entry module.

[0176] Data preprocessing and data entry can further include the following steps:

[0177] 1. Receive raw data and determine the data type: text, doc, pdf, drawings, images, BIM model data.

[0178] 2. Based on different types, call the corresponding module to store the entered data structure and original files in the database.

[0179] 3. Text Input: Query engineering feature identifier dimension data from the database (e.g., structure type, process type, construction period, etc.), integrate it with the prompt word template to form prompt words, call the large model for recognition, and return a dictionary: {"Original paragraph 1": [tag 1, tag 2, ...], "Original paragraph 2": ......}

[0180] 4. doc input: Convert doc to text and call the text input module for input.

[0181] 5. PDF Input: The process involves calling a multimodal large model to convert PDFs into text, treating images and tables within the PDF as images, calculating image hash values, recording the image hash values ​​and multimodal semantic descriptions separately, recording the relationship information between images and their corresponding paragraphs (e.g., a dictionary: {"image hash value", "paragraph content"}), calling the text input module, and returning the text data and image relationship information in a unified manner.

[0182] 6. Drawing Input: Convert drawings into PDF files and call the PDF input module.

[0183] 7. Image Input: Query engineering feature identification dimension data from the database (such as structure type, process type, construction period length, etc.), integrate it with the prompt word template to form prompt words, call multimodal recognition, and return a dictionary: {"Original paragraph 1": [tag 1, tag 2, ...], "Original paragraph 2": ......}.

[0184] 8. BIM model entry: Query engineering feature identification dimension data from the database, such as structural type, process type, construction period length, etc., call the large model to analyze the data, extract the data that meets the requirements, and return it to the dictionary.

[0185] The historical construction plans are cleaned to remove garbled characters and duplicate paragraphs from the text, and key information is extracted and stored in the database system.

[0186] In this embodiment, a Python script is used to remove garbled characters and repeated paragraphs from the solution text and extract key information to store in a MongoDB database (such as project name, construction process, and parameter table).

[0187] Extract the key parameters from the design drawings and convert them into JSON format.

[0188] The database system includes a MySQL subsystem and a MinIO subsystem. The MySQL subsystem is used to store relational data, and the MinIO subsystem is used to store raw files.

[0189] Depending on the data type of the received data, the received data is stored in different subsystems.

[0190] In this embodiment, the aforementioned database system is used to store multiple types of data. In specific implementation, the following steps can be adopted:

[0191] 1. Build a database system: including MySQL and MinIO. MySQL is used to store relational data, and MinIO is used to store raw files.

[0192] 2. Implement data distribution service: Determine the data type returned by the data preprocessing module, and store different data types in different data. For example, dictionary and relational data types are stored in MySQL relational tables, while the original files are stored in MinIO.

[0193] 3. Implement query services: Provide query interfaces for MySQL and MinIO.

[0194] In some embodiments, the process of training the above-described model includes the following steps:

[0195] 1. Start the training function periodically or manually.

[0196] 2. Use the identification results of the data entry work as input information and the original text content as output information to form a dataset.

[0197] 3. Merge similar content in the dataset.

[0198] 4. Divide the training set and the test set in a 7:3 ratio.

[0199] 5. Use a large model training framework to train on the training set.

[0200] 6. Iterate through the training until the number of iterations or the loss metric converges to the required level.

[0201] 7. Use the test set to further validate the training results. If the loss metric fails to meet the requirements, return to iterative training until the requirements are met.

[0202] In some embodiments, the training steps of the above model are illustrated below:

[0203] (1) Pre-training: Use historical schemes to pre-train the language model and learn the text patterns in the construction field.

[0204] (2) Fine-tuning: Select 500 high-quality plans that have been approved by experts, mark key nodes such as “construction process” and “technical measures”, and supervise and learn from them.

[0205] (3) Reinforcement learning: Set the reward function:

[0206] R_total = α×R_logic +β×R_consistency + γ×R_compliance + δ×R_quality;

[0207] R_logic is a multi-dimensional logical association reward, where the multi-dimensional factors include schedule matching degree, material usage deviation degree, and cost.

[0208] R_consistency is the cross-paragraph consistency reward;

[0209] R_compliance is the specification compliance reward;

[0210] R_quality is a reward for text quality and diversity.

[0211] α is the multidimensional logical association reward weight;

[0212] β is the cross-segment consistency reward weight within the scheme;

[0213] γ is the compliance reward weight;

[0214] δ represents the reward weight for text quality and diversity.

[0215] The following is an example of a solution generation method:

[0216] (1) Input parameters: "A high-rise residential building, frame-shear wall structure, 30 floors above ground and 2 floors below ground, with a building area of ​​35,000 square meters and a seismic fortification intensity of 7 degrees.

[0217] (2) Generated content:

[0218] Project Overview: Automatically extract parameters such as structural type and building area, and match them with "Key Points for Construction of High-Rise Residential Buildings in Seismic Fortification Areas of 7 Degrees";

[0219] Construction deployment: The recommended vertical transportation solution is "tower crane + construction elevator", with a peak workforce of 180 people calculated based on the construction period.

[0220] Technical measures: Generate a "temperature control scheme for large-volume concrete" (pre-embedded temperature measuring pipes, temperature difference control ≤25℃), referencing the latest clauses of GB50666-2011.

[0221] This application provides an intelligent method for compiling construction plans that integrates AI technology. Through data-driven and model optimization, it achieves "rapid generation, automatic verification, and dynamic updating" of plans, solving the efficiency and quality pain points of traditional manual compilation and improving the level of intelligence in the compilation of building construction plans.

[0222] This application achieves the following technical effects:

[0223] (1) An AI generation model for construction-specific plans was constructed, which integrates historical plan data, standard specifications and engineering parameters to achieve automatic plan generation.

[0224] (2) A dynamic specification update mechanism was designed to synchronize industry standards in real time and solve the problem of specification lag in traditional solutions.

[0225] (3) Integrate semantic analysis and logic verification modules to improve the technical rationality and format standardization of the solution through human-machine collaboration.

[0226] (4) Reinforcement learning is adopted to optimize the generation strategy and the model is continuously iterated based on the scheme review feedback to improve the practicality of the generated scheme.

[0227] (5) The entire process from inputting engineering parameters to outputting a complete solution (including text, charts, and references to standards) is automated.

[0228] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of this disclosure. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0229] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for AI-assisted generation and compilation of construction plans, characterized in that, include: Obtain construction requirement data; Establish an automatic model for generating construction plan documents; The automatic generation model for the construction plan document is trained. The construction requirements data are input into a trained construction plan document automatic generation model, which then outputs a construction plan document. The automatic generation model for the construction plan document includes at least a near-end strategy optimization model; The near-end policy optimization model includes a reward function, which is calculated using the following formula: R_total = α×R_logic +β×R_consistency + γ×R_compliance + δ×R_quality; R_logic is a multi-dimensional logical association reward, where the multi-dimensional factors include schedule matching degree, material usage deviation degree, and cost. R_consistency is the cross-paragraph consistency reward; R_compliance is the specification compliance reward; R_quality is a reward for text quality and diversity. α is the multi-dimensional logical association reward weight; β is the cross-segment consistency reward weight within the scheme; γ is the compliance reward weight; δ represents the reward weight for text quality and diversity; R_logic = w1×f_d + w2×f_m + w3×f_c; Where f_d represents the project schedule matching degree; f_m represents the deviation in material usage; f_c represents the accuracy of cost calculation; w1 is the weight of the project schedule matching degree; w2 is the weighting value for the deviation in material usage; w3 is the weight for the accuracy of cost calculation; f_d = sigmoid(-k|D - (D_min+D_max) / 2| / (D_max-D_min) / 2); Where k is the sensitivity coefficient; D represents the construction period generated by the automatically generated model of the trained construction plan document; D_min is the theoretical minimum construction period; D_max is the theoretical maximum construction period; f_m = max(0, 1 - |M_extracted - M_calc| / M_calc); Where M_extracted represents the material usage generated by the trained construction plan document automatic generation model; M_calc represents the theoretical material usage; f_c = ; in, To calculate the total cost; This represents the theoretical total cost.

2. The construction plan AI-assisted generation and compilation method according to claim 1, characterized in that, R_consistency =average(Sim_ij)-λ×(Num_contraction); Sim_ij = cosine_sim(F_i, F_j); Where λ is the penalty coefficient; Num_contraction represents the number of inconsistent paragraphs; F_i is the core engineering feature vector of the i-th paragraph; F_j is the core engineering feature vector of the j-th paragraph; Sim_ij is the cosine similarity between F_i and F_j; average(Sim_ij) is the average cosine similarity of the core engineering feature vectors of any two paragraphs in all paragraphs.

3. The construction plan AI-assisted generation and compilation method according to claim 1, characterized in that, Identify each core paragraph and its engineering feature vector; Determine the embedding vector of the core paragraph; Using the engineering feature vectors, suitable clauses are retrieved from the standard clause library to obtain the first clause set; Calculate the similarity between the embedded vector and all the text vectors in the standard text library, and obtain the texts with the highest similarity to obtain the second text set; The first set of articles and the second set of articles are merged, duplicate articles are removed, and a set of standard articles applicable to the core paragraph is constructed. The core paragraph is paired with each normative provision in the set of normative provisions using a pre-trained semantic contradiction discrimination model to determine the number of provisions that semantically contradict the core paragraph. The compliance reward R_compliance is determined based on the number of clauses that contradict the semantics of the core paragraph.

4. The construction plan AI-assisted generation and compilation method according to claim 1, characterized in that, R_quality = w4×f_readability + w5×f_terminology; Where f_readability is the readability coefficient; w4 is the readability coefficient weight; f_terminology = 1 - (NUM_error terms / NUM_total terms); f_terminology represents the reverse weighting of erroneous terms; NUM_ErrorTerms represents the number of error terms; NUM_Total Terms refers to the total number of terms; w5 represents the reverse weighting of incorrect terms.

5. The construction plan AI-assisted generation and compilation method according to claim 1, characterized in that, The method further includes: reviewing the construction plan document, specifically including: The construction plan document is divided into several semantically complete paragraphs using prompt words; Each section of content is characterized; For each paragraph, further semantic analysis is performed, and similar specification requirements are found using vector similarity and engineering feature identifiers. Use a large language model to determine whether there are any contradictions between the semantics of the construction plan document and the specification semantics; Perform similarity matching on the feature data represented by all paragraphs, analyze data consistency, and check for any contradictions.

6. The construction plan AI-assisted generation method according to claim 5, characterized in that, The further semantic analysis of each paragraph, using vector similarity and engineering feature identifiers to find similar specification requirements, includes: For any paragraph, calculate the original text vector in the paragraph; Calculate the cosine similarity between the original text vector and the data vectors in the vector library; Identify data in the vector library that meet the cosine similarity threshold; The feature identifiers of the original text are extracted using a large model; The database is searched based on the feature identifiers of the original text to obtain standardized data that conforms to the feature identifiers; The standardized data that conforms to the feature identifier and the data in the vector library that conforms to the cosine similarity threshold are merged. Return the merged data.

7. The construction plan AI-assisted generation method according to claim 1, characterized in that, The training of the automatic generation model for the construction plan document includes: Acquire and preprocess experience data, specifically including: Using web crawling technology, we regularly download various engineering specifications, historical plans and design drawings from the official website of the Ministry of Housing and Urban-Rural Development. The historical construction plans are cleaned to remove garbled characters and duplicate paragraphs from the text, and key information is extracted and stored in the database system. Extract the key parameters from the design drawings and convert them into JSON format; The database system includes a MySQL subsystem and a MinIO subsystem, wherein the MySQL subsystem is used to store relational data and the MinIO subsystem is used to store raw files; Based on the data type of the received data, the received data is stored in different subsystems; Data collection and preprocessing specifically include: Using web crawling technology, we regularly download various engineering specifications, historical plans and design drawings from the official website of the Ministry of Housing and Urban-Rural Development. The historical construction plans are cleaned to remove garbled characters and duplicate paragraphs from the plan text, and key information is extracted and stored in the database. Extract the key parameters from the design drawings and convert them into JSON format.

Images