Construction schedule plan wbs automatic generation method and system based on large model
By using a large language model for supervised fine-tuning and reinforcement learning of construction documents and BIM models, the efficiency and consistency issues in the generation of construction schedule plans (WBS) were resolved, enabling intelligent and digital management of engineering projects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING URBAN CONSTR GROUP
- Filing Date
- 2026-02-24
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies are inefficient and inconsistent in generating work breakdown structures (WBS) for construction schedules, making it difficult to adapt to different project types and project management requirements, and lacking intelligence and adaptability.
A large language model is used to supervise and fine-tune construction documents and BIM model data and perform reinforcement learning to generate a WBS that conforms to engineering expert standards and industry specifications. Through multimodal information fusion, construction task items and process levels are automatically identified, and the granularity is dynamically adjusted to reduce the amount of manual maintenance work.
It improves the efficiency and consistency of construction schedule (WBS) generation, adapts to different project types, reduces manual intervention, and realizes digital management and intelligent generation of construction projects.
Smart Images
Figure CN122114501A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of engineering construction project management technology, and in particular to a method and system for automatically generating a construction schedule plan (WBS) based on a large model. Background Technology
[0002] In the construction process of building engineering projects, the Work Breakdown Structure (WBS) is the foundation of construction schedule planning and resource management, and its quality directly affects the accuracy of plan preparation and execution effectiveness.
[0003] Traditionally, the generation of the Work Breakdown Structure (WBS) has relied primarily on project planners manually reading construction drawings, construction organization designs, and standards, then breaking them down based on experience. This method has its shortcomings: 1. Low compilation efficiency: Manually analyzing construction drawings, BIM models (Building Information Modeling), engineering contracts, construction task lists, specifications and standards, etc., and then sorting out the sub-items and processes from the data, the planning and compilation cycle is long; 2. Poor consistency of results: Differences in the experience and comprehension of different planners lead to inconsistent WBS structures, affecting subsequent plan collaboration. Summary of the Invention
[0004] In view of this, the purpose of this application is to provide a method and system for automatically generating a construction schedule WBS based on a large model, which can integrate multimodal information, automatically generate a construction schedule WBS, improve the efficiency of planning, and promote the digital management of construction projects.
[0005] Firstly, this application provides a method for automatically generating a construction schedule plan (WBS) based on a large model, the method comprising: Supervised fine-tuning of the basic large language model yields an intermediate large language model for identifying construction task items at a specified task granularity and process levels from construction documents and / or BIM model data of engineering construction projects. The intermediate large language model is trained by reinforcement learning to obtain a target large language model whose output conforms to the standard expression of engineering experts and industry norms. The construction organization design document, construction plan document, contract terms document, bill of quantities document and / or BIM model data of the target project, as well as prompt words containing the granularity of the task to be identified, are input into the target large language model to identify the first construction task item and process level that match the prompt words; the construction drawings of the target project are input into the multimodal large model to identify the second construction task item, and the first construction task item is supplemented based on the second construction task item; The terminology of the supplemented first construction task item is standardized. Based on the task granularity level to which the standardized construction task item belongs and the process level, a hierarchical WBS tree containing task node ID, parent-child relationship and task description information is generated.
[0006] In one possible implementation, the supervised fine-tuning of the basic large language model to obtain an intermediate large language model for identifying construction task items at a specified task granularity and process levels from construction documents and / or BIM model data of an engineering construction project includes: The first engineering corpus required for model training and inference is collected from multi-source engineering data. The first engineering corpus includes at least construction organization design documents, construction plan documents, contract terms documents, bill of quantities documents, BIM model data containing component information, system information, spatial location, historical WBS plans, and process template library. The first engineering corpus is preprocessed, and the preprocessing includes cleaning, noise reduction, and format unification. The preprocessed first engineering corpus is annotated using a dialogue format, which includes task entities, sub-projects, and process levels. Sub-projects are organized into sub-projects, and sub-projects are organized into unit projects. The annotation basis is given, and a dialogue format annotation sample is constructed. Using the standard cross-entropy loss function as the learning algorithm loss function, the labeled samples are used to supervise and fine-tune the basic large language model to obtain the intermediate large language model; wherein, the standard cross-entropy loss function is: ; This represents the loss value during the monitoring and fine-tuning phase; This represents the summation of the loss at each position in the sequence; T represents the sequence length; i represents the position index in the sequence, and... The t in the text corresponds to; This means that the model predicts the current word given the input and preceding context. The logarithmic probability; The parameter is The probability distribution output by the model; This represents the actual word at position t, i.e., the label; x represents an input suggestion. This indicates all generated words preceding t.
[0007] In one possible implementation, the step of training the intermediate large language model through reinforcement learning to obtain a target large language model whose output conforms to engineering expert standard expressions and industry specifications includes: The second engineering corpus and prompt words containing the granularity of the task to be recognized are input into the intermediate large language model after supervision and fine-tuning. By adjusting the model parameters of the intermediate large language model, two hierarchical WBS trees are output for the same input data. By judging whether the construction task items contained in the hierarchical WBS tree conform to the standard expression of engineering experts, industry specifications and WBS hierarchical logic, the preferred answer and the inferior answer are distinguished from the two output hierarchical WBS trees, and the pair preference training data is constructed. Using the preference-aligned contrastive loss function and the paired preference training data, the intermediate large language model is trained using DPO reinforcement learning to obtain the target large language model; wherein, the preference-aligned contrastive loss function is: ; This represents the summation of N pairs of preference training data, where N represents the number of preference training data pairs. Each preference training data pair contains an input cue x and a preferred answer. A poor choice answer ; Represents the Sigmoid function; This indicates temperature hyperparameters; Indicates that the current parameter is The probability that the trained model generates the answer y; This represents the probability that the reference model generates the answer y; This represents the log-probability ratio between the current model and the reference model; This represents the difference between the logarithmic probabilities of the preferred and unpreferred answers; the training objective is to maximize... .
[0008] In one possible implementation, the terminology standardization process for the supplemented first construction task item includes: Using a semantic vector model, the first construction task item is transformed into a first semantic vector; Obtain the second semantic vector of the standard sub-item and sub-section construction task item; Based on the distance between the first semantic vector and the second semantic vector, the sub-item and sub-section construction task item most similar to the first construction task item is determined and used as the standardized construction task item.
[0009] In one possible implementation, the terminology standardization process for the supplemented first construction task item includes: If the first construction task item does not match any of the standard sub-item construction task items, then based on the distance between the first semantic vector of the first construction task item and the second semantic vector of each standard sub-item construction task item, the sub-item construction task item most similar to the first construction task item is determined and used as the standardized construction task item.
[0010] In one possible implementation, the method further includes: Obtain the corrected data of the hierarchical WBS tree of the target project, and write the corrected data into the iterative training database of the target large language model; Using the standard cross-entropy loss function as the learning algorithm loss function, the target large language model is periodically supervised and fine-tuned using the iterative training database to achieve iterative optimization of the target large language model.
[0011] In one possible implementation, the multimodal large model is trained based on construction drawings and corresponding drawing understanding text.
[0012] Secondly, this application provides an automatic construction schedule plan (WBS) generation system based on a large model, the system comprising: The supervised fine-tuning module is used to supervise and fine-tune the basic large language model to obtain an intermediate large language model for identifying construction task items at a specified task granularity and process level from construction documents and / or BIM model data of engineering construction projects. The reinforcement learning module is used to train the intermediate large language model through reinforcement learning to obtain a target large language model whose output conforms to the standard expression of engineering experts and industry specifications. The construction task item identification module is used to input the construction organization design document, construction plan document, contract terms document, bill of quantities document and / or BIM model data of the target project, as well as prompt words containing the granularity of the task to be identified, into the target large language model, identify the first construction task item and process level that matches the prompt words; input the construction drawings of the target project into the multimodal large model, identify the second construction task item, and supplement the first construction task item based on the second construction task item; The hierarchical WBS tree generation module is used to standardize the terminology of the supplemented first construction task item, and generate a hierarchical WBS tree containing task node ID, parent-child relationship and task description information based on the task granularity level to which the standardized construction task item belongs and the process level.
[0013] Thirdly, this application provides an electronic device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the automatic generation method for a construction schedule plan (WBS) based on a large model as described in any of the first aspects.
[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the automatic generation method for a construction schedule plan (WBS) based on a large model as described in any one of the first aspects.
[0015] The method and system for automatically generating construction schedule WBS based on a large model provided in this application can integrate multimodal information and automatically generate a construction schedule WBS, thereby improving the efficiency of planning and promoting the digital management of construction projects. Furthermore, it utilizes the transfer learning and few-shot learning capabilities of the large model to achieve rapid adaptation to new scenarios.
[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart of a method for automatically generating a construction schedule plan (WBS) based on a large model, provided in an embodiment of this application, is shown. Figure 2 This document illustrates a flowchart of an iterative optimization method for a large language model, as provided in an embodiment of this application. Figure 3 This illustration shows a structural diagram of a construction schedule plan (WBS) automatic generation system based on a large model, provided in an embodiment of this application. Figure 4 A schematic diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0020] In the construction process of building engineering projects, the Work Breakdown Structure (WBS) is the foundation of construction schedule planning and resource management, and its quality directly affects the accuracy of plan preparation and execution effectiveness. Traditionally, WBS generation has relied primarily on project planners manually reading construction drawings, construction organization designs, specifications, and standards, then manually breaking down the data based on experience. This method has its shortcomings: 1. Low compilation efficiency: The process involves manually analyzing construction drawings, BIM models, engineering contracts, construction task lists, specifications and standards, and then sorting out the sub-items and procedures from the data, resulting in a long planning cycle; 2. Poor consistency of results: Differences in the experience and comprehension of different planners lead to inconsistent WBS structures, affecting subsequent plan collaboration.
[0021] Currently, with the continuous advancement of technology, there are already some methods for automatically generating WBS. The following is an explanation of the existing methods and their advantages and disadvantages.
[0022] Method 1: A method, apparatus, equipment, and medium for generating Work Breakdown Structures (WBS) based on a decision tree model. This technical solution proposes an automatic method for generating WBS based on a machine learning decision tree model. It acquires project information (especially parameters and conditions for pumped storage projects), extracts and quantifies the project information, inputs the feature vectors into a trained decision tree model for reasoning, and outputs multi-level classification results such as project category, sub-category, section, item, and process. Based on this, it generates a standardized multi-level WBS code (including project code, unit project code, section project code, item project code, unit project code, and process type code). Finally, the generated results are stored in a database and can be used for real-time monitoring or project management platforms. The solution also includes training and optimizing the decision tree model using historical project data and provides corresponding apparatus, computing equipment, and storage media to support deployment and application. However, this technical solution has certain shortcomings: 1. This scheme is designed for specific types of projects (such as pumped storage projects), and its scope of application is limited, making it difficult to extend to other types of projects such as residential buildings, subways, and bridges; 2. This technology typically relies on structured parameter input, requiring manual formulation of feature extraction and quantification rules. When project information contains unstructured text, drawings, or BIM models, it cannot be processed directly, resulting in insufficient automation. 3. The WBS decomposition hierarchy and coding structure of this scheme are fixed, making it difficult to flexibly adjust according to the management requirements of different engineering projects; 4. Decision tree models are prone to decreased inference accuracy when faced with missing parameters, new construction techniques, or complex cross-domain projects, and lack adaptive learning and continuous optimization capabilities. 5. The solution lacks an understanding of construction drawings, BIM and other data. The solution does not have the ability to perform in-depth analysis of drawings or BIM models. It relies entirely on parameter input and feature coding, and cannot capture details that are clear on the drawings but not quantified (such as the excavation sequence of a certain area or the special characteristics of local structures).
[0023] The aforementioned shortcomings mean that this automatic WBS generation method still has room for improvement in terms of versatility, intelligence, and robustness.
[0024] Method 2: Construction Planning Method Based on Building Information Modeling. This technical solution proposes a construction planning and Work Breakdown Structure (WBS) automatic generation method based on Building Information Modeling (BIM). This solution uses the BIM model as the core input for construction planning. First, it extracts the basic, system, and spatial information of all physical components from the BIM model. Then, it combines pre-established knowledge bases and mapping relationship libraries, including a knowledge base of sub-items systems, a WBS generation strategy knowledge base, a sub-item mapping relationship library (the correspondence between modeling software systems and sub-items), a large machinery and equipment process library, and a process constraint rule library. Through mapping and decomposition of BIM elements, it automatically generates the project's sub-item structure and corresponding WBS. This method further associates the generated WBS with process data and establishes a three-dimensional and time-dimensional linkage with the BIM model, thereby achieving 4D construction simulation and visualization. It can output the construction time of each process and the total project duration, providing auxiliary support for construction plan preparation. The overall technical approach of this solution is as follows: using the BIM model as the data source, components are identified and classified through mapping relationships and a predefined strategy library to generate corresponding WBS and work processes, and the connection with BIM is established to output construction plans and visualization results. However, the above-mentioned BIM-based construction planning and automatic WBS generation method still has certain limitations: 1. This solution relies heavily on pre-built knowledge bases, mapping relationship databases, and strategy rules, resulting in a large amount of preparatory work. Furthermore, when construction techniques, component types, or coding rules change, these databases need to be manually updated and maintained, leading to poor flexibility.
[0025] 2. This solution mainly processes the structured information of the BIM model. It is difficult to directly utilize unstructured data such as design specifications, standard texts, and construction drawing annotations, resulting in limited information coverage.
[0026] 3. Its generation process is more based on rule-based mapping and matching, with limited intelligence, making it difficult to handle newly emerging component types or complex cross-disciplinary scenarios.
[0027] 4. Although the solution can output process and schedule information, it lacks adaptive learning and automatic optimization capabilities, and cannot automatically improve the quality and accuracy of the generated data as historical projects accumulate. Its versatility and scalability are insufficient.
[0028] 5. Reliance on complete BIM and preprocessing. Using BIM as the sole or primary information source, and requiring "preprocessing / adding a large amount of information" to BIM, necessitates establishing a mapping relationship library and ensuring that the modeling software's built-in system corresponds to the knowledge base of individual components. This significantly reduces automation capabilities when BIM is incomplete, no modeling is performed, or the project is still in its early stages and only drawings exist.
[0029] 6. Rule-based approach with weak semantic generalization ability. This scheme does not utilize high-level semantics to automatically correct inconsistencies or semantic errors, and therefore cannot solve the problem of inconsistencies in the recognition of transcription catalogs.
[0030] 7. Insufficient automation coverage and strategies for handling unmodeled items. "Unmodeled objects" require manual or rule-based mapping, making the method essentially semi-automatic. Automatic coverage depends on the quality of BIM modeling.
[0031] Therefore, there is an urgent need for a method that can integrate multimodal information and automatically generate a construction schedule plan (WBS) to improve the efficiency of planning, ensure the integrity and consistency of results, and promote the digital management of construction projects.
[0032] To address the aforementioned issues, this application provides a method and system for automatically generating a Work Breakdown Structure (WBS) based on a large-scale model. This system directly receives and understands multimodal engineering data, including BIM model data, construction drawings, design specifications, and standard documents, comprehensively covering key project information. It automatically extracts key information from unstructured or semi-structured engineering documents, design drawings, and BIM model data. Utilizing the knowledge and reasoning capabilities of a pre-trained large-scale model, it automatically identifies component types, engineering divisions, and sub-items, dynamically generating a WBS structure that conforms to standards, reducing manual maintenance workload. Through the contextual understanding and multi-round reasoning capabilities of the large-scale model, it ensures that a reasonable WBS result can still be output even when information is incomplete or expressions are diverse. This enhances the intelligence and adaptability of WBS generation.
[0033] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0034] 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.
[0035] To facilitate understanding of this embodiment, a detailed description of the automatic generation method for construction schedule plan (WBS) based on a large model disclosed in this application embodiment will be provided first.
[0036] Example 1: See Figure 1 As shown, Figure 1 A flowchart illustrating an automatic construction schedule plan (WBS) generation method based on a large model, provided for embodiments of this application, includes the following steps: S101. Supervised fine-tuning of the basic large language model to obtain an intermediate large language model for identifying construction task items and process levels at a specified task granularity from construction documents and / or BIM model data of engineering construction projects.
[0037] Step S101 performs supervised fine-tuning (SFT) on the Large Language Model (LLM). To distinguish it from the trained Large Language Model, the Large Language Model with the original parameters is called the basic Large Language Model (e.g., the Qwen series of Tongyi Qianwen Large Language Model and the DeepSeek Large Language Model). The Large Language Model after supervised fine-tuning is called the intermediate Large Language Model.
[0038] Construction project documentation includes construction organization design documents, construction plan documents, contract terms documents, and bill of quantities documents. BIM model data contains component information, system information, spatial location information, etc. Construction tasks are categorized into three levels of granularity: sub-item tasks, sub-section tasks, and unit project tasks. Work sequence level refers to the sequential relationship between the various construction tasks.
[0039] The input to the large language model is the construction documents of the engineering project, or BIM model data, or both. The large language model is used to identify construction task items and the work sequence hierarchy between these tasks from the construction documents and BIM model data.
[0040] Specifically, the basic large language model is supervised and fine-tuned using the following methods: S1011. Collect the first engineering corpus required for model training and inference from multi-source engineering data. The first engineering corpus includes at least construction organization design documents, construction plan documents, contract terms documents, bill of quantities documents, BIM model data containing component information, system information, and spatial location, historical WBS plans, and process template library.
[0041] Here, the engineering corpus used to train the basic large language model is referred to as the first engineering corpus. The engineering corpus includes construction documents, BIM model data, and historical WBS plans for the same engineering project. The process template library pre-stores the processes for implementing existing construction tasks.
[0042] S1012. Preprocess the first engineering corpus, the preprocessing including cleaning, noise reduction, and format unification.
[0043] S1013. Use a dialogue format to annotate the preprocessed first engineering corpus, annotating task entities, sub-projects, and process levels. Organize sub-projects into sub-projects, and sub-projects into unit projects. Provide the annotation basis and construct a dialogue format annotation sample.
[0044] Construct annotation samples in ChatML or other dialogue formats. ChatML (Chat Markup Language) is a dialogue markup language designed specifically for large language models. The specific annotation process is as follows: Professional data annotators, based on their understanding of construction organization design documents, construction plan documents, contract terms documents, bill of quantities documents, and BIM model data, manually organize task items at the task entity and sub-item level. They then organize the sub-items into WBS task items in the format of sub-projects and unit projects, providing the basis for compiling each task. For example, based on the structural type description "the main structure is reinforced concrete" in the construction organization plan, it can be determined that the main structure sub-project includes tasks related to reinforcement, formwork, and concrete.
[0045] S1014. Using the standard cross-entropy loss function as the learning algorithm loss function, the labeled samples are used to perform supervised fine-tuning of the basic large language model to obtain the intermediate large language model; wherein, the standard cross-entropy loss function is: ; This represents the loss value during the supervised fine-tuning phase, used to measure the model's performance on the current task. The smaller the value, the more accurate the model's predictions. This represents the summation of the loss at each position in the sequence; T represents the sequence length (i.e., the number of words generated); i represents the position index in the sequence, usually related to... The t in the text corresponds to; This means that the model predicts the current word given the input and preceding context. The logarithmic probability; The parameter is The probability distribution output by the model. The actual word at position t is the label, and x represents the input prompt (input text or command). This indicates all generated words preceding t.
[0046] When training the basic large language model, the model uses the data annotated in step S1013 to learn how to generate the WBS task structure using construction documents such as construction organization design documents and BIM model data.
[0047] S102. The intermediate large language model is trained by reinforcement learning to obtain a target large language model whose output conforms to the standard expression of engineering experts and industry norms.
[0048] Step S102 uses DPO (Direct Preference Optimization) reinforcement learning training to align the output of the large language model with human preferences, ensuring that the model output conforms to the standard expressions of engineering experts and industry norms, and ensuring that the generated WBS is professional and applicable.
[0049] Specifically, the intermediate large language model is trained using DPO reinforcement learning in the following manner: S1021. Input the second engineering corpus and prompt words containing the granularity of the task to be identified into the intermediate large language model after supervision and fine-tuning. By adjusting the model parameters of the intermediate large language model, output two hierarchical WBS trees for the same input data.
[0050] S1022. By judging whether the construction task items contained in the hierarchical WBS tree conform to the engineering expert standard expression, industry specifications and WBS hierarchical logic, the preferred answer and the inferior answer are distinguished from the two output hierarchical WBS trees, and the pair preference training data is constructed.
[0051] Steps S1021 and S1022 prepare preference data pairs for DPO reinforcement learning training of the intermediate large language model. For the same question, by setting the sampling parameters of the intermediate large language model after supervised fine-tuning, two WBS answers are generated. Professional data labelers judge which WBS is better based on evaluation dimensions such as whether the WBS answers meet the professionalism of engineering standards and specifications, and whether they conform to the WBS hierarchical logic (e.g., whether they are output in the format of individual units, parts, and items). The better WBS answer is selected as the preferred answer, and the worse WBS answer is selected as the inferior answer, thus constructing reinforcement learning pairwise preference training data.
[0052] It should be noted that the second engineering corpus is similar in content to the first engineering corpus, but the purpose is different. The first engineering corpus is used for model supervision and fine-tuning, while the second engineering corpus is used for model reinforcement learning training.
[0053] S1023. Using the preference-aligned contrastive loss function and the paired preference training data, perform DPO reinforcement learning training on the intermediate large language model to obtain the target large language model; wherein, the preference-aligned contrastive loss function is: , This represents the target loss function for DPO training, which should be as small as possible; the negative sign "-" indicates that the loss is calculated by taking the negative logarithm of the probability to maximize the preference probability. This represents the summation of N preference training data pairs, where N represents the number of preference pairs in the training dataset. Each preference pair contains: an input x and a preferred answer. A poor choice answer ; This represents the Sigmoid function, which maps real numbers to the interval (0,1); This indicates the temperature hyperparameter, controlling the degree to which the model deviates from the reference model; Indicates the current training model (parameters are) The probability of generating the answer y; This represents the probability that the reference model (usually the model after SFT, which remains unchanged) will generate the answer y. This represents the log-probability ratio between the current model and the reference model, measuring the current model's preference for a particular answer relative to the reference model. A positive value indicates that the current model is more likely to generate that answer than the reference model. This represents the difference between the log probabilities of the preferred and unpreferred answers. The larger this difference, the more biased the model is. Instead ; x represents the input prompt, and the training objective is to maximize This makes the model's relative preference for the best answer (compared to the reference model) greater than its relative preference for the worst answer. The large language model trained through reinforcement learning is called the target large language model.
[0054] S103. Input the construction organization design document, construction plan document, contract terms document, bill of quantities document and / or BIM model data of the target project, as well as prompt words containing the granularity of the task to be identified, into the target large language model, identify the first construction task item and process level that are compatible with the prompt words; input the construction drawings of the target project into the multimodal large model, identify the second construction task item, and supplement the first construction task item based on the second construction task item.
[0055] In actual operation, the input to the large language model is construction documents related to the project, BIM model data, or a combination of both. The process of calling the large language model to identify sub-items involves the large language model automatically identifying sub-sections, sub-items, and work processes from the semi-structured data of the construction documents or BIM model, generating a preliminary index table. After training through steps S1011-S1014, the large language model has the ability to understand the content of construction documents such as construction organization design documents, and can extract construction tasks based on prompts. For example, after converting the construction organization design document into text format, the model can be prompted with the following question: "Please extract the sub-section and sub-item level construction tasks from the document." The model will output the sub-section and sub-item projects contained in the construction organization design document. The construction task item output by the large language model is called the first construction task item.
[0056] Leveraging the knowledge and reasoning capabilities of a large language model, it automatically identifies component types, engineering divisions, and sub-items, dynamically generating a WBS structure that conforms to specifications, reducing manual maintenance workload. It overcomes the instability of reasoning in decision trees or rule matching when parameters are missing or information is incomplete. Through the contextual understanding and multi-turn reasoning capabilities of the large speech model, it ensures that reasonable WBS results are still output even with incomplete information or diverse expressions, improving intelligence and robustness.
[0057] Multimodal large models generate corresponding construction tasks by understanding construction drawings. The input to a multimodal large model is the construction drawings, specifically used to extract construction task items (referred to as second construction task items). These second construction task items supplement construction tasks not mentioned in the construction documents and BIM model data. A standard multimodal large model possesses a certain ability to understand construction drawings and extract construction task items. To further enhance the capabilities of a standard model, it can be trained using construction drawings and corresponding drawing-understanding text.
[0058] The large language model enables control over task granularity: it dynamically adjusts the WBS granularity according to actual needs, generating overall construction schedules down to sub-section tasks, or refining them to unit project schedules and sub-item project schedules. This dynamic granularity adjustment is primarily achieved through prompts, which require the model to extract construction task items at the sub-section or sub-item level. The construction task items output by the large language model match the task granularity contained in the input prompts. In practical applications, the decomposition depth and granularity of the WBS can be dynamically adjusted according to the management requirements of different projects, achieving finer or coarser task decomposition.
[0059] S104. Perform terminology standardization processing on the supplemented first construction task item, and generate a hierarchical WBS tree containing task node ID, parent-child relationship, and task description information based on the task granularity level to which the standardized construction task item belongs and the process level.
[0060] After extracting the construction task items, the model leverages the construction industry knowledge of a large language model to standardize the terminology and classify the task granularity of all supplemented first construction task items by comparing the professional terminology of the sub-items of the project using semantic similarity criteria. The final model outputs a structured hierarchical WBS tree, which includes node IDs, parent-child relationships, and task description information.
[0061] The terminology standardization process involves using a semantic vector model to transform the supplemented first construction task item into a semantic vector (called the first semantic vector). Simultaneously, the standard sub-item construction tasks are also transformed into vectors (called the second semantic vectors). The distance between the first semantic vector and each second semantic vector is calculated. The standard sub-item construction task corresponding to the second semantic vector with the closest distance is taken as the standard sub-item construction task most semantically similar to the first construction task item, thus completing the standardization of the construction task items. Standardization also classifies the task granularity, determining whether each task belongs to a sub-section or sub-item project. The above terminology standardization method requires standardization for each first construction task item. However, some first construction task items are already standard sub-item construction task items and do not require further standardization. Therefore, another standardization method is provided: first, the first construction task item is matched with each standard sub-item construction task item. If no match is found, the terminology standardization process is then performed as described above, improving standardization efficiency.
[0062] Further, see Figure 2 As shown, Figure 2 A flowchart for iterative optimization of a large language model provided in this application includes the following processing steps: S201. Obtain the corrected data of the hierarchical WBS tree of the target project, and write the corrected data into the iterative training database of the target large language model.
[0063] S202. Using the standard cross-entropy loss function as the learning algorithm loss function, the target large language model is periodically supervised and fine-tuned using the iterative training database to achieve iterative optimization of the target large language model.
[0064] To ensure consistently high-quality output across different projects and enterprise standards, this application supports writing the corrected WBS back to the training data in response to user adjustments and feedback on the WBS task generation results. Specifically, it collects WBS task item data generated by users during the process and uses it as training data for the next model training. Leveraging the transfer learning and few-shot learning capabilities of large models, it achieves rapid adaptation to new scenarios.
[0065] This application embodiment supports user correction of the generated results and confirmation of whether they agree to use the generated and corrected WBS task item data as training data. If they agree, the data is added to the large speech model training data in step S1014, achieving continuous iterative optimization of the large language model to adapt it to different project types and enterprise standards. It supports periodic fine-tuning training (SFT) using timed model training scripts, utilizing new data accumulated within the cycle, incorporating user feedback into the training loop, and continuously improving model performance without retraining from scratch. This application embodiment can periodically calculate user satisfaction indicators such as the modification rate of the model-generated WBS tasks, automatically triggering optimization tasks and forming a closed-loop iterative process of "generation → feedback → optimization → regeneration".
[0066] The method for automatically generating a construction schedule plan (WBS) based on a large model provided in this application has the following advantages: 1. High degree of automation and reduced human intervention: The WBS is generated directly from engineering documents, BIM model data and historical data through a large language model. It no longer relies on a large number of manually constructed mapping libraries and rule bases, which significantly reduces the cost of human participation and maintenance, and improves the generation efficiency and reliability.
[0067] 2. High adaptability, meeting the needs of various engineering projects: Due to the reasoning ability of the large language model, it can adapt to engineering projects of different types, scales and complexities, including building construction, rail transit, municipal and pumped storage projects. It can dynamically adjust the granularity of WBS and process logic according to input data and enterprise standards, avoiding the inapplicability problem caused by the limitation of the rule base.
[0068] 3. Accurate construction logic and feasible project plans: Through training and alignment, the model can automatically identify the dependencies and constraints between construction tasks. The generated WBS and schedule plan conform to the engineering construction logic, improve the scientific nature and feasibility of construction organization, and reduce construction delays and resource conflicts caused by unreasonable planning.
[0069] 4. Continuous learning and iterative optimization capabilities of the model: It supports writing back the user-corrected WBS and construction plan to the training data to form a closed-loop self-learning mechanism, enabling the model to continuously optimize the generation effect, adapt to changes in project standards and the needs of new engineering types, and improve the long-term applicability and intelligence level of the model.
[0070] In summary, the embodiments of this application, by combining the natural language understanding and reasoning capabilities of a large language model, achieve efficient, intelligent, and dynamic generation of engineering WBS, solving the current problems of low automation, strong rule dependence, poor adaptability, and high maintenance costs, and significantly improving the efficiency of engineering project management and the scientific nature of construction plans.
[0071] Example 2: Based on the same inventive concept, this application also provides an automatic construction schedule WBS generation system based on a large model, which corresponds to the automatic generation method of construction schedule WBS based on a large model. Since the principle of the system in this application is similar to the automatic generation method of construction schedule WBS based on a large model described above, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be described again.
[0072] See Figure 3 As shown, Figure 3 A schematic diagram of a construction schedule planning (WBS) automatic generation system based on a large model, provided in this application embodiment, is shown. The system includes: The supervisory fine-tuning module 301 is used to supervise and fine-tune the basic large language model to obtain an intermediate large language model for identifying construction task items at a specified task granularity and process level from construction documents and / or BIM model data of engineering construction projects. Reinforcement learning module 302 is used to perform reinforcement learning training on the intermediate large language model to obtain a target large language model whose output conforms to the standard expression of engineering experts and industry norms. The construction task item identification module 303 is used to input the construction organization design document, construction plan document, contract terms document, bill of quantities document and / or BIM model data of the target project, as well as prompt words containing the granularity of the task to be identified, into the target large language model, identify the first construction task item and process level that matches the prompt words; input the construction drawings of the target project into the multimodal large model, identify the second construction task item, and supplement the first construction task item based on the second construction task item; The hierarchical WBS tree generation module 304 is used to perform terminology standardization processing on the supplemented first construction task item, and generate a hierarchical WBS tree containing task node ID, parent-child relationship and task description information based on the task granularity level to which the standardized construction task item belongs and the process level.
[0073] In one possible implementation, the supervisory fine-tuning module 301, when performing supervisory fine-tuning on the basic large language model to obtain an intermediate large language model for identifying construction task items at a specified task granularity and process levels from construction documents and / or BIM model data of an engineering construction project, includes: The first engineering corpus required for model training and inference is collected from multi-source engineering data. The first engineering corpus includes at least construction organization design documents, construction plan documents, contract terms documents, bill of quantities documents, BIM model data containing component information, system information, spatial location, historical WBS plans, and process template library. The first engineering corpus is preprocessed, and the preprocessing includes cleaning, noise reduction, and format unification. The preprocessed first engineering corpus is annotated using a dialogue format, which includes task entities, sub-projects, and process levels. Sub-projects are organized into sub-projects, and sub-projects are organized into unit projects. The annotation basis is given, and a dialogue format annotation sample is constructed. Using the standard cross-entropy loss function as the learning algorithm loss function, the labeled samples are used to supervise and fine-tune the basic large language model to obtain the intermediate large language model; wherein, the standard cross-entropy loss function is: ; This represents the loss value during the monitoring and fine-tuning phase; This represents the summation of the loss at each position in the sequence; T represents the sequence length; i represents the position index in the sequence, and... The t in the text corresponds to; This means that the model predicts the current word given the input and preceding context. The logarithmic probability; The parameter is The probability distribution output by the model; This represents the actual word at position t, i.e., the label; x represents an input suggestion. This indicates all generated words preceding t.
[0074] In one possible implementation, the reinforcement learning module 302, when performing reinforcement learning training on the intermediate large language model to obtain a target large language model whose output conforms to engineering expert standard expressions and industry specifications, includes: The second engineering corpus and prompt words containing the granularity of the task to be recognized are input into the intermediate large language model after supervision and fine-tuning. By adjusting the model parameters of the intermediate large language model, two hierarchical WBS trees are output for the same input data. By judging whether the construction task items contained in the hierarchical WBS tree conform to the standard expression of engineering experts, industry specifications and WBS hierarchical logic, the preferred answer and the inferior answer are distinguished from the two output hierarchical WBS trees, and the pair preference training data is constructed. Using the preference-aligned contrastive loss function and the paired preference training data, the intermediate large language model is trained using DPO reinforcement learning to obtain the target large language model; wherein, the preference-aligned contrastive loss function is: ; This represents the summation of N pairs of preference training data, where N represents the number of preference training data pairs. Each preference training data pair contains an input cue x and a preferred answer. A poor choice answer ; Represents the Sigmoid function; This indicates temperature hyperparameters; Indicates that the current parameter is The probability that the trained model generates the answer y; This represents the probability that the reference model generates the answer y; This represents the log-probability ratio between the current model and the reference model; This represents the difference between the logarithmic probabilities of the preferred and unpreferred answers; the training objective is to maximize... .
[0075] In one possible implementation, the hierarchical WBS tree generation module 304, when performing terminology standardization processing on the supplemented first construction task item, includes: Using a semantic vector model, the first construction task item is transformed into a first semantic vector; Obtain the second semantic vector of the standard sub-item and sub-section construction task item; Based on the distance between the first semantic vector and the second semantic vector, the sub-item and sub-section construction task item most similar to the first construction task item is determined and used as the standardized construction task item.
[0076] In one possible implementation, the hierarchical WBS tree generation module 304, when performing terminology standardization processing on the supplemented first construction task item, includes: If the first construction task item does not match any of the standard sub-item construction task items, then based on the distance between the first semantic vector of the first construction task item and the second semantic vector of each standard sub-item construction task item, the sub-item construction task item most similar to the first construction task item is determined and used as the standardized construction task item.
[0077] In one possible implementation, the system further includes: The data correction writing module is used to obtain the correction data of the hierarchical WBS tree of the target project and write the correction data into the iterative training database of the target large language model; The model iterative optimization module is used to periodically supervise and fine-tune the target large language model using the standard cross-entropy loss function as the learning algorithm loss function and the iterative training database, so as to achieve iterative optimization of the target large language model.
[0078] In one possible implementation, the multimodal large model in the construction task identification module 303 is trained based on construction drawings and corresponding drawing understanding text.
[0079] Example 3: See Figure 4 As shown, Figure 4 This is a schematic diagram of an electronic device provided in an embodiment of this application. The electronic device 400 includes a processor 401, a memory 402, and a bus 403. The memory 402 stores machine-readable instructions that can be executed by the processor 401. When the electronic device is running, the processor 401 communicates with the memory 402 through the bus 403. The processor 401 executes the machine-readable instructions to perform the steps of the above-described automatic generation method for construction schedule plan (WBS) based on a large model.
[0080] Specifically, the memory 402 and processor 401 mentioned above can be general-purpose memory and processor, without any specific limitations. When the processor 401 runs the computer program stored in the memory 402, it can execute the above-mentioned automatic generation method of construction schedule plan (WBS) based on large model.
[0081] Example 4: Corresponding to the above-described method for automatically generating a construction schedule WBS based on a large model, this application embodiment also provides a computer-readable storage medium storing a computer program, which, when run by a processor, executes the steps of the above-described method for automatically generating a construction schedule WBS based on a large model.
[0082] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division; in actual implementation, there may be other division methods. Furthermore, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or modules may be electrical, mechanical, or other forms.
[0083] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.
[0084] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.
[0085] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0086] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for automatically generating a Work Breakdown Structure (WBS) based on a large model, characterized in that, The method includes: Supervised fine-tuning of the basic large language model yields an intermediate large language model for identifying construction task items at a specified task granularity and process levels from construction documents and / or BIM model data of engineering construction projects. The intermediate large language model is trained by reinforcement learning to obtain a target large language model whose output conforms to the standard expression of engineering experts and industry norms. The construction organization design document, construction plan document, contract terms document, bill of quantities document and / or BIM model data of the target project, as well as prompt words containing the granularity of the task to be identified, are input into the target large language model to identify the first construction task item and process level that match the prompt words; the construction drawings of the target project are input into the multimodal large model to identify the second construction task item, and the first construction task item is supplemented based on the second construction task item; The terminology of the supplemented first construction task item is standardized. Based on the task granularity level to which the standardized construction task item belongs and the process level, a hierarchical WBS tree containing task node ID, parent-child relationship and task description information is generated.
2. The method for automatically generating a construction schedule plan (WBS) based on a large model according to claim 1, characterized in that, The supervised fine-tuning of the basic large language model yields an intermediate large language model used to identify construction task items at a specified task granularity and process levels from construction documents and / or BIM model data of engineering construction projects, including: The first engineering corpus required for model training and inference is collected from multi-source engineering data. The first engineering corpus includes at least construction organization design documents, construction plan documents, contract terms documents, bill of quantities documents, BIM model data containing component information, system information, spatial location, historical WBS plans, and process template library. The first engineering corpus is preprocessed, and the preprocessing includes cleaning, noise reduction, and format unification. The preprocessed first engineering corpus is annotated using a dialogue format, which includes task entities, sub-projects, and process levels. Sub-projects are organized into sub-projects, and sub-projects are organized into unit projects. The annotation basis is given, and a dialogue format annotation sample is constructed. Using the standard cross-entropy loss function as the learning algorithm loss function, the labeled samples are used to supervise and fine-tune the basic large language model to obtain the intermediate large language model; wherein, the standard cross-entropy loss function is: ; This represents the loss value during the monitoring and fine-tuning phase; This represents the summation of the loss at each position in the sequence; T represents the sequence length; i represents the position index in the sequence, and... The t in the text corresponds to; This means that the model predicts the current word given the input and preceding context. The logarithmic probability; The parameter is The probability distribution output by the model; This represents the actual word at position t, i.e., the label; x represents an input suggestion. This indicates all generated words preceding t.
3. The method for automatically generating a construction schedule plan (WBS) based on a large model according to claim 2, characterized in that, The process of training the intermediate large language model through reinforcement learning to obtain a target large language model whose output conforms to engineering expert standard expressions and industry norms includes: The second engineering corpus and prompt words containing the granularity of the task to be recognized are input into the intermediate large language model after supervision and fine-tuning. By adjusting the model parameters of the intermediate large language model, two hierarchical WBS trees are output for the same input data. By judging whether the construction task items contained in the hierarchical WBS tree conform to the standard expression of engineering experts, industry specifications and WBS hierarchical logic, the preferred answer and the inferior answer are distinguished from the two output hierarchical WBS trees, and the pair preference training data is constructed. Using the preference-aligned contrastive loss function and the paired preference training data, the intermediate large language model is trained using DPO reinforcement learning to obtain the target large language model; wherein, the preference-aligned contrastive loss function is: ; This represents the summation of N pairs of preference training data, where N represents the number of preference training data pairs. Each preference training data pair contains an input cue x and a preferred answer. A poor choice answer ; Represents the Sigmoid function; This indicates temperature hyperparameters; Indicates that the current parameter is The probability that the trained model generates the answer y; This represents the probability that the reference model generates the answer y; This represents the log-probability ratio between the current model and the reference model; This represents the difference between the logarithmic probabilities of the preferred and unpreferred answers; the training objective is to maximize... .
4. The method for automatically generating a construction schedule plan (WBS) based on a large model according to claim 1, characterized in that, The terminology standardization process for the supplemented first construction task item includes: Using a semantic vector model, the first construction task item is transformed into a first semantic vector; Obtain the second semantic vector of the standard sub-item and sub-section construction task item; Based on the distance between the first semantic vector and the second semantic vector, the sub-item and sub-section construction task item most similar to the first construction task item is determined and used as the standardized construction task item.
5. The method for automatically generating a construction schedule plan (WBS) based on a large model according to claim 1, characterized in that, The terminology standardization process for the supplemented first construction task item includes: If the first construction task item does not match any of the standard sub-item construction task items, then based on the distance between the first semantic vector of the first construction task item and the second semantic vector of each standard sub-item construction task item, the sub-item construction task item most similar to the first construction task item is determined and used as the standardized construction task item.
6. The method for automatically generating a construction schedule WBS based on a large model according to claim 2, characterized in that, The method further includes: Obtain the corrected data of the hierarchical WBS tree of the target project, and write the corrected data into the iterative training database of the target large language model; Using the standard cross-entropy loss function as the learning algorithm loss function, the target large language model is periodically supervised and fine-tuned using the iterative training database to achieve iterative optimization of the target large language model.
7. The method for automatically generating a construction schedule plan (WBS) based on a large model according to claim 1, characterized in that, The multimodal large model is trained based on construction drawings and corresponding drawing understanding text.
8. A construction schedule planning (WBS) automatic generation system based on a large model, characterized in that, The system includes: The supervised fine-tuning module is used to supervise and fine-tune the basic large language model to obtain an intermediate large language model for identifying construction task items at a specified task granularity and process level from construction documents and / or BIM model data of engineering construction projects. The reinforcement learning module is used to train the intermediate large language model through reinforcement learning to obtain a target large language model whose output conforms to the standard expression of engineering experts and industry specifications. The construction task item identification module is used to input the construction organization design document, construction plan document, contract terms document, bill of quantities document and / or BIM model data of the target project, as well as prompt words containing the granularity of the task to be identified, into the target large language model, identify the first construction task item and process level that matches the prompt words; input the construction drawings of the target project into the multimodal large model, identify the second construction task item, and supplement the first construction task item based on the second construction task item; The hierarchical WBS tree generation module is used to standardize the terminology of the supplemented first construction task item, and generate a hierarchical WBS tree containing task node ID, parent-child relationship and task description information based on the task granularity level to which the standardized construction task item belongs and the process level.
9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the automatic generation method for a construction schedule plan (WBS) based on a large model as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the automatic generation method for a construction schedule plan (WBS) based on a large model as described in any one of claims 1 to 7.