BIM model and design budget integration method based on AI
By integrating multi-source data and AI models, deep integration of BIM models and real-world data has been achieved, automatically identifying and correcting defects, and updating design estimates in real time. This solves the problem of the disconnect between BIM models and real-world data, and improves the accuracy and timeliness of the design phase.
Patent Information
- Application Number
- CN202511500646.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-21
- Publication Date
- 2026-01-13
AI Technical Summary
In existing technologies, BIM models are disconnected from real-world data, lacking automated defect identification and intelligent repair capabilities. This results in insufficient accuracy and timeliness of design estimates, reliance on manual defect identification, low level of intelligent cost correlation, and a disconnect between repair and cost updates.
By collecting and fusing multi-source data, and using AI models for defect identification and repair, including multi-task deep learning models and reinforcement learning algorithms, the BIM model and real-world data are deeply integrated to automatically identify equipment missing, redundant, and geometric distortions, and to update the design budget in real time.
It achieves high-precision alignment between the BIM model and the construction site, automatically identifies defects and updates the budget in real time, ensures consistency between the model and cost data, and improves the model quality and cost control accuracy in the design phase.
Smart Images
Figure CN121328320A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of building construction technology, specifically to a method for integrating AI-based BIM models with design estimates. Background Technology
[0002] Building Information Modeling (BIM), as an integrated digital representation of buildings, has been widely used in the design, construction, and operation and maintenance phases of architecture. BIM models not only contain the geometric information of the building but also integrate multi-dimensional data such as component attributes, quantities, and costs, providing crucial support for project management throughout the entire process. As a key link in early-stage cost control, the accuracy of design estimates directly impacts project investment decisions and subsequent cost management during construction. Traditionally, the preparation of design estimates relies primarily on design drawings, bills of quantities, and the experience-based judgment of cost estimators, which suffers from high subjectivity, low efficiency, and susceptibility to errors.
[0003] In recent years, some studies have attempted to link BIM models with cost data to automate cost estimates during the design phase. For example, component information can be extracted from the BIM model using IFC standards and matched with unit price information from a cost database to generate a preliminary bill of quantities and cost estimate. However, these methods still have the following shortcomings:
[0004] 1. Disconnect between model and real-world data: BIM models are mostly theoretical models from the design phase, which differ from the actual conditions of the construction site; the lack of an effective mechanism for collecting and integrating real-world data makes it difficult for the model to accurately reflect the geometric features, equipment layout and environmental conditions of the building entity, thus affecting the accuracy of the budget estimate.
[0005] 2. Defect identification relies on manual labor: Most existing methods rely on manual inspection of BIM models for issues such as missing or redundant equipment or geometric distortions, which is inefficient and prone to omissions.
[0006] 3. Low level of intelligence in cost association: Although some systems can achieve simple mapping between components and cost items, they lack the ability to intelligently parse unstructured text in design documents; the extraction of information such as engineering quantity keywords and cost rules is still mainly done manually, making it difficult to achieve dynamic and accurate component-cost association, resulting in the budget update lagging behind design changes.
[0007] 4. Disconnect between repair and cost update: The model repair process and cost adjustment are often separated. The model changes after repair cannot be reflected in the budget in real time, resulting in inconsistency between cost data and model status; the lack of a closed-loop optimization mechanism makes it difficult to achieve optimal cost while ensuring model compliance.
[0008] Therefore, there is an urgent need in this field for an integrated method that can achieve deep integration of BIM models and real-world data, possess automated defect identification and intelligent repair capabilities, and can update design estimates in real time, so as to improve model quality and cost control accuracy during the design phase. Summary of the Invention
[0009] To address the shortcomings of existing technologies, the purpose of this invention is to provide an AI-based method for integrating BIM models and design estimates, which enables deep fusion of BIM models and real-world data, automated defect identification and intelligent repair capabilities, and real-time synchronous updates of design estimates.
[0010] The technical solution adopted by this invention to solve its technical problem is:
[0011] A method for integrating BIM models and design estimates based on AI includes the following steps:
[0012] Step 1, Multi-source Data Acquisition and Fusion: Texture data of the building entity is acquired using a panoramic camera, geometric data of the building entity is acquired using a laser scanner to generate a laser scan point cloud, and environmental data of the building entity is acquired using IoT sensors. The texture data, laser scan point cloud, and environmental data are combined to construct a multimodal dataset. The original BIM model constructed using BIM design software in the early stages of the project is obtained and converted into an original BIM model point cloud. The multimodal dataset and the original BIM model point cloud are virtually fused to generate an enhanced BIM model with spatial coordinates.
[0013] Step 2, AI model training and defect identification:
[0014] S201: AI Model Training: Train a multi-task deep learning model, which includes a defect detection sub-model and a cost association sub-model. The defect detection sub-model is built based on the YOLO algorithm or Mask R-CNN algorithm and is used to identify equipment missing, equipment redundancy, and geometric distortion defects in the BIM model during the enhancement and subsequent repair process. The cost association sub-model uses natural language processing technology to parse the design documents, extract the engineering quantity information and cost keywords in the documents, and construct the component-cost mapping relationship between components and cost items in the BIM model.
[0015] S202: Input the enhanced BIM model, the defect detection sub-model of the multi-task deep learning model identifies defects in the enhanced BIM model and outputs the defect type, defect location and defect severity;
[0016] Step 3, Dynamic Repair and Cost Synchronization: Based on the defect type, defect location and defect severity output in Step 2, call the predefined repair strategy or generate new components through generative AI to repair and enhance the BIM model to obtain the repaired BIM model; call the cost association sub-model in the multi-task deep learning model to parse the repaired BIM model in real time, update the bill of quantities and unit price library in linkage, and generate a revised budget report.
[0017] Step 4, Difference Assessment and Closed-Loop Optimization:
[0018] S401: Calculate the difference index between the enhanced BIM model and the repaired BIM model, including spatial deviation ΔL and equipment quantity deviation ΔN, and combine it with the preset cost tolerance threshold to determine whether secondary optimization is triggered;
[0019] S402: If secondary optimization is required, a dynamic decision-making model is trained based on reinforcement learning algorithm to automatically select the repair scheme with the lowest cost and compliance with building codes. Repeat steps S301-S302 until the difference index and cost data meet the preset requirements.
[0020] Step 5, Visualization and Collaborative Management: Integrate difference heatmaps and cost fluctuation curve visualization tools into the BIM platform to support multi-role collaborative decision-making; output standardized reports, which include remediation logs, cost change details, and compliance verification results.
[0021] As a preferred embodiment, a further technical solution of the present invention is:
[0022] Preferably, in step 1, the specific process of virtual fusion includes,
[0023] S101: Point cloud registration. The iterative nearest point algorithm is used to register the laser scanned point cloud with the original BIM model point cloud. The point cloud registration uses the ICP algorithm. Through the iterative nearest point algorithm, the optimal spatial transformation matrix between the laser scanned point cloud and the original BIM model point cloud is calculated, so that the points in the laser scanned point cloud and the original BIM model point cloud representing the same building component or the same spatial location are aligned in three-dimensional space.
[0024] S102: Image semantic segmentation. The texture image of the building entity is classified at the pixel level using the U-Net network. That is, each pixel in the image is classified, and the pixel regions belonging to walls, floors, doors and windows, and electromechanical equipment are automatically identified and extracted. The semantic information of the building elements is obtained based on the extracted pixel regions. At the same time, the two-dimensional image coordinates of the elements corresponding to the semantic information of the building elements are obtained. Then, combined with the three-dimensional spatial coordinates provided by the laser scanning point cloud, the two-dimensional image coordinates of the elements are converted into three-dimensional semantic information with three-dimensional spatial positioning. The three-dimensional semantic information is associated and matched with the spatial coordinates of the corresponding components in the original BIM model to obtain the enhanced BIM model.
[0025] Preferably, in S102, obtaining semantic information of building elements based on the extracted pixel regions specifically includes: obtaining the outline range and material texture features of the wall in the image based on the extracted wall pixel regions; obtaining the planar position and thickness visual features of the floor slab in the image based on the extracted floor slab pixel regions; obtaining the installation position and size ratio features of the doors and windows in the image based on the extracted door and window pixel regions; and obtaining the shape outline and installation position features of the electromechanical equipment in the image based on the extracted electromechanical equipment pixel regions; wherein, each extracted feature is the semantic information of building elements.
[0026] Preferably, in step 201, the training process of the cost-related sub-model includes,
[0027] S2011, Data Preprocessing: Clean unstructured data such as design documents and contract texts, remove redundant information, and extract engineering quantity keywords and cost keywords through word segmentation and part-of-speech tagging. Among them, engineering quantity keywords include component size, component quantity, and component material, and cost keywords include component unit price, project total price, and billing standard.
[0028] S2012, Constructing a Knowledge Graph: Using components in the original BIM model as nodes, setting cost items, with cost items as node attributes, including engineering quantity calculation rules, component unit price, and cost accounting formula, establishing semantic association between components and cost items, supporting automatic mapping from engineering quantity keywords and cost keywords to corresponding cost data, forming an initial component-cost mapping relationship;
[0029] S2013, Model Training: The cost association sub-model is initialized using the BERT pre-trained model. The cost association sub-model is trained by importing component-cost matching data from historical projects. Training stops when the F1-Score of the cost association sub-model on the validation set is greater than or equal to the preset standard. The final cost association sub-model is then output. When the final cost association sub-model takes component parameters and keywords as input, it directly outputs the final component-cost mapping relationship.
[0030] Preferably, step 3 specifically includes:
[0031] S301, Model Repair: Call the defect type, defect location, and defect severity information output by the defect detection sub-model in step S202, and perform corresponding repair operations for different defect types; if it is a missing equipment defect, complete the equipment according to the parameters of the surrounding components and the building's functional requirements; if it is a redundant equipment defect, delete the redundant equipment and adjust the pipeline connection relationship associated with the redundant equipment; if it is a geometric distortion defect, correct the size and spatial position of the defective component based on the geometric data of the enhanced BIM model, and complete the BIM model repair; obtain the repaired BIM model;
[0032] S302, Budget Update: The component-cost mapping relationship output from the cost-related sub-model in step S201 is invoked to analyze the changes in the repaired BIM model in S301 in real time. When new components are added to the repaired BIM model, the unit price data and quantity calculation rules corresponding to the new components are matched through the component-cost mapping relationship to calculate the cost of the new components. When the dimensions of components in the repaired BIM model change, the quantity of the changed components is recalculated according to the quantity calculation rules in the component-cost mapping relationship, and the component cost is adjusted based on the corresponding unit price data. When the repaired BIM model involves the addition or removal of equipment, the unit price of the added or removed equipment is extracted based on the component-cost mapping relationship, and the cost difference caused by the addition or removal of equipment is calculated. Based on the calculation results, the bill of quantities and unit price library are updated in conjunction, and a revised budget report is generated.
[0033] Preferably, in step S301, the repair process for geometric distortion defects includes,
[0034] S3011: Extract the geometric data of the enhanced BIM model corresponding to the defect location output in step S202, including the component vertex coordinates and outline dimensions;
[0035] S3012: Compare the geometric parameters of the defective component in the original BIM model with the geometric data of the enhanced model, and calculate the geometric deviation value;
[0036] S3013: Based on the allowable range of component size deviation in the building code, adjust the vertex coordinates and outline dimensions of the defective component so that the deviation between the geometric parameters of the repaired component and the geometric data of the enhanced model is ≤0.03m, thus completing the repair of geometric distortion defects.
[0037] Preferably, in step S401, the spatial deviation for,
[0038] = ;
[0039] Equipment quantity deviation for,
[0040] ;
[0041] The difference index is calculated for components with identified defects in the enhanced BIM model and their associated components that are affected in terms of physical connection or function. >0.1m or When the value is greater than 1, the model deviation is determined to be excessive, triggering a second optimization.
[0042] Preferably, in step S402, the training process of the reinforcement learning decision model includes,
[0043] S4021: Define the state space as the current model deviation index and cost data, and the action space as the set of repair solutions, which includes component size adjustment, adding new components, and adding or removing equipment.
[0044] S4022: Define the reward function
[0045] ;
[0046] Where α and β are weighting coefficients, α takes the value of 0.6, β takes the value of 0.4, and ΔLmax is the maximum allowable spatial deviation, ΔLmax takes the value of 0.1m;
[0047] S4023: The model is trained iteratively using the PPO algorithm. In each iteration, a repair plan is selected and executed based on the current state. The reward value is calculated and the model parameters are optimized. When the average reward value of 100 consecutive iterations is ≥0.85, the model training is complete and the optimal repair plan can be automatically output.
[0048] Preferably, generative AI employs a Transformer-based generative adversarial network or diffusion model to automatically generate new component models that conform to engineering logic based on the defect location context and building codes.
[0049] Preferably, the environmental data collected by the IoT sensor includes at least one of temperature, humidity, light intensity, and personnel flow data, which is used in step S202 to assist in verifying the authenticity of the equipment redundancy or missing defects identified by the defect detection sub-model, and in step S402 as auxiliary input to the state space of the reinforcement learning decision model to optimize the repair scheme.
[0050] The present invention, which adopts the above technical solution, has the following prominent features compared with the prior art:
[0051] This invention achieves intelligent management of the entire design process, from data acquisition and model optimization to cost control, significantly improving the accuracy of BIM models and the precision and timeliness of design estimates. Specifically:
[0052] 1. By virtually integrating multi-source real-world data (laser point cloud, texture, IoT) with the BIM model, high-precision alignment between the design model and the actual state of the construction site was achieved, resulting in a more realistic and reliable enhanced BIM model.
[0053] 2. Through a deep learning-based multi-task AI model, the system achieves automated identification of defects such as missing equipment, redundancy, and geometric distortion in BIM models, as well as intelligent analysis of engineering quantity and cost information in design documents.
[0054] 3. By using the cost-related sub-model to perform real-time analysis of the repaired model, any changes to the model can be linked to updates of the bill of quantities and cost estimates, ensuring real-time synchronization and consistency of the "model-cost estimate" data.
[0055] 4. By using a closed-loop optimization mechanism based on difference index and reinforcement learning, the system can automatically find and execute the most cost-effective repair solution while meeting building codes, thus achieving dual control over quality and cost. Attached Figure Description
[0056] Figure 1 This is a flowchart illustrating the method for integrating AI-based BIM models and design estimates in an embodiment of the present invention.
[0057] Figure 2 This is another flowchart illustrating the method for integrating AI-based BIM models and design estimates in an embodiment of the present invention. Detailed Implementation
[0058] The present invention will be further illustrated below with reference to specific embodiments. The purpose of this illustration is solely to provide a better understanding of the invention. Therefore, the examples given do not limit the scope of protection of the present invention.
[0059] like Figures 1 to 2 As shown in the figure, this embodiment presents a method for integrating BIM models and design estimates based on AI, characterized by the following steps:
[0060] Step 0, Building the basic dataset:
[0061] BIM Model Training Dataset: Collected defect samples of BIM models from 100 building projects (including 3000 samples of missing equipment, 2500 samples of redundant equipment, and 3500 samples of geometric distortion). Each sample is labeled with "defect type - defect location (XYZ coordinates) - defect severity (level 1~5, level 1 is minor, level 5 is severe)" and divided into training set, validation set and test set in a 7:2:1 ratio.
[0062] Component-Cost Mapping Dataset: Historical data from 50 completed projects were compiled, including component parameters (e.g., C30 concrete frame column: cross-sectional dimensions 500mm×500mm, height 3m), corresponding cost items (material cost 380 yuan / m³, labor cost 120 yuan / m³, machinery cost 80 yuan / m³), and engineering quantity calculation rules (frame column volume = cross-sectional area × height), and a structured database was constructed (using MySQL 8.0 for storage).
[0063] Step 1, Multi-source data acquisition and fusion: Texture data of the building entity is acquired through a panoramic camera, geometric data of the building entity is acquired through a laser scanner to generate a laser scanning point cloud, and environmental data of the building entity is acquired through IoT sensors. The texture data, laser scanning point cloud, and environmental data are combined to construct a multimodal dataset; the original BIM model constructed by BIM design software in the early stage of the project is obtained and converted into a BIM original model point cloud; the multimodal dataset and the BIM original model point cloud are virtually fused to generate an enhanced BIM model with spatial coordinates.
[0064] The laser scanner used is the Faro Focus S70 (point cloud accuracy ±2mm@10m), the panoramic camera is the Ricoh Theta X (11K resolution, supports 360° panoramic acquisition), and the IoT sensor is a LoRa wireless sensor (sampling frequency 1 time / minute, temperature measurement range -20℃~60℃, humidity measurement accuracy ±3%RH). The BIM design software is Autodesk Revit 2024; the point cloud processing software is CloudCompare 2.13.3; the deep learning framework is PyTorch 2.0; and the BIM platform integration uses Autodesk Navisworks 2024.
[0065] The specific process of virtual fusion includes,
[0066] S101: Point cloud registration. The iterative nearest point algorithm is used to register the laser scanned point cloud with the original BIM model point cloud. The point cloud registration uses the ICP algorithm. Through the iterative nearest point algorithm, the optimal spatial transformation matrix between the laser scanned point cloud and the original BIM model point cloud is calculated, so that the points in the laser scanned point cloud and the original BIM model point cloud representing the same building component or the same spatial location are aligned in three-dimensional space.
[0067] S102: Image semantic segmentation. The texture image of the building entity is classified at the pixel level using the U-Net network. That is, each pixel in the image is classified, and the pixel regions belonging to walls, floors, doors and windows, and electromechanical equipment are automatically identified and extracted. The semantic information of the building elements is obtained based on the extracted pixel regions. At the same time, the two-dimensional image coordinates of the elements corresponding to the semantic information of the building elements are obtained. Then, combined with the three-dimensional spatial coordinates provided by the laser scanning point cloud, the two-dimensional image coordinates of the elements are converted into three-dimensional semantic information with three-dimensional spatial positioning. The three-dimensional semantic information is associated and matched with the spatial coordinates of the corresponding components in the original BIM model to obtain the enhanced BIM model.
[0068] Specifically, obtaining semantic information of building elements based on extracted pixel regions includes: obtaining the outline range and material texture features of the wall in the image based on the extracted wall pixel regions; obtaining the planar position and thickness visual features of the floor slab in the image based on the extracted floor slab pixel regions; obtaining the installation position and size ratio features of the doors and windows in the image based on the extracted door and window pixel regions; and obtaining the shape outline and installation position features of the electromechanical equipment in the image based on the extracted electromechanical equipment pixel regions. The extracted features constitute the semantic information of building elements.
[0069] Step 2, AI model training and defect identification:
[0070] S201: AI Model Training: Training a multi-task deep learning model, which includes a defect detection sub-model and a cost-related sub-model. The defect detection sub-model, built based on the YOLO or Mask R-CNN algorithm, is used to identify defects such as missing equipment, redundant equipment, and geometric distortion in the enhanced BIM model and subsequent repair processes. During the training of the defect detection sub-model, the BIM model training dataset serves as the core source of training samples.
[0071] The cost association sub-model uses natural language processing technology to parse the design documents, extract the engineering quantity information and cost keywords in the documents, and construct the component-cost mapping relationship between components and cost items in the BIM model.
[0072] The training process for the cost-related sub-model includes,
[0073] S2011, Data Preprocessing: Clean unstructured data such as design documents and contract texts, remove redundant information, and extract engineering quantity keywords and cost keywords through word segmentation and part-of-speech tagging. Among them, engineering quantity keywords include component size, component quantity, and component material, while cost keywords include component unit price, project total price, and billing standard.
[0074] S2012, Constructing a Knowledge Graph: Using components in the original BIM model as nodes, cost items are set as node attributes, including quantity calculation rules, component unit prices, and cost accounting formulas. Semantic relationships between components and cost items are established, supporting automatic mapping from quantity keywords and cost keywords to corresponding cost data, forming an initial component-cost mapping relationship. The component-cost mapping dataset serves as the data source for knowledge graph construction.
[0075] S2013, Model Training: The cost association sub-model was initialized using a BERT pre-trained model. Component-cost matching data from historical projects was imported to train the cost association sub-model. Training stopped when the F1-Score of the cost association sub-model on the validation set was greater than or equal to a preset standard, and the final cost association sub-model was output. When the final cost association sub-model takes component parameters and keywords as input, it directly outputs the final component-cost mapping relationship. In the S2013 model training phase, 80% of the data in the component-cost mapping dataset was used as the training set, and 20% was used as the validation set.
[0076] The defect detection sub-model (based on YOLOv8) has the following parameters: input image size of 640×640, batch size of 16, initial learning rate of 0.01 (cosine annealing learning rate scheduling, decaying to 0.8 every 50 epochs), 100 training epochs, and loss function of CIoULoss (for bounding box regression) + FocalLoss (to address sample imbalance, α=0.25, γ=2.0). The defect recognition accuracy on the validation set must reach at least 92% (equipment missing recognition accuracy ≥93%, equipment redundancy recognition accuracy ≥91%, geometric distortion recognition accuracy ≥92%); otherwise, additional samples are required for retraining. YOLOv8 can improve inference speed while maintaining a 92% recognition accuracy, meeting the real-time defect detection needs in engineering projects.
[0077] Cost-related sub-model (based on BERT): The BERT-base-chinese pre-trained model is used, with a sequence length of 512, a batch size of 32, a learning rate of 2e-5 (using the AdamW optimizer with a weight decay of 0.01), 50 training epochs, and cross-entropy loss (used for matching and classifying components and cost items). The preset F1-Score standard is 0.9 (training stops when the F1-Score on the validation set is consistently ≥0.9 for 5 consecutive epochs). If the standard is not met, the knowledge graph needs to be optimized (supplementing missing component-cost association rules, such as adding a mapping relationship between "fire door" and "fire door installation fee").
[0078] S202: Input the enhanced BIM model, and the defect detection sub-model of the multi-task deep learning model identifies defects in the enhanced BIM model and outputs the defect type, defect location, and defect severity.
[0079] Step 3, Dynamic Repair and Cost Synchronization:
[0080] S301, Model Repair: Call the defect type, defect location, and defect severity information output by the defect detection sub-model in step S202, call the predefined repair strategy or generate new components through generative AI; specifically, if it is a missing equipment defect, complete the equipment according to the parameters of the surrounding components and the building's functional requirements; if it is a redundant equipment defect, delete the redundant equipment and adjust the pipeline connection relationship associated with the redundant equipment; if it is a geometric distortion defect, correct the size and spatial position of the defective component based on the geometric data of the enhanced BIM model to complete the BIM model repair; obtain the repaired BIM model.
[0081] The repair process for geometric distortion defects includes,
[0082] S3011: Extract the geometric data of the enhanced BIM model corresponding to the defect location output in step S202, including the component vertex coordinates and outline dimensions;
[0083] S3012: Compare the geometric parameters of the defective component in the original BIM model with the geometric data of the enhanced model, and calculate the geometric deviation value;
[0084] S3013: Based on the allowable range of component size deviation in the building code, adjust the vertex coordinates and outline dimensions of the defective component so that the deviation between the geometric parameters of the repaired component and the geometric data of the enhanced model is ≤0.03m, thus completing the repair of geometric distortion defects.
[0085] The generative AI employs a Transformer-based generative adversarial network or diffusion model to automatically generate new component models that conform to engineering logic based on the defect location context and building codes. The diffusion model generates component models that conform to building codes through progressive denoising, and in this invention, it is used for the automatic generation of new components when equipment is missing, avoiding errors from manual modeling.
[0086] S302, Budget Update: The component-cost mapping relationship output from the cost-related sub-model in step S201 is invoked to analyze the changes in the repaired BIM model in S301 in real time. When new components are added to the repaired BIM model, the unit price data and quantity calculation rules corresponding to the new components are matched through the component-cost mapping relationship to calculate the cost of the new components. When the dimensions of components in the repaired BIM model change, the quantity of the changed components is recalculated according to the quantity calculation rules in the component-cost mapping relationship, and the component cost is adjusted based on the corresponding unit price data. When the repaired BIM model involves the addition or removal of equipment, the unit price of the added or removed equipment is extracted based on the component-cost mapping relationship, and the cost difference caused by the addition or removal of equipment is calculated. Based on the calculation results, the bill of quantities and unit price library are updated in conjunction, and a revised budget report is generated.
[0087] Step 4, Difference Assessment and Closed-Loop Optimization:
[0088] S401: Calculate the difference index between the enhanced BIM model and the repaired BIM model, including spatial deviation ΔL and equipment quantity deviation ΔN. Combined with the preset cost tolerance threshold, determine whether to trigger secondary optimization. The preset cost tolerance threshold is: the deviation between the total cost after repair and the target cost is ≤3%, and the change in cost of a single component is ≤5%. This threshold is set according to the cost deviation requirements of the "Construction Engineering Quantity List Pricing Specification" GB 50500.
[0089] Among them, spatial deviation for,
[0090] = ;
[0091] Equipment quantity deviation for,
[0092] ;
[0093] The difference index is calculated for components with identified defects in the enhanced BIM model and their associated components that are affected in terms of physical connection or function. >0.1m or When the value is greater than 1, the model deviation is determined to be excessive, triggering a second optimization.
[0094] S402: If secondary optimization is required, a dynamic decision-making model is trained based on reinforcement learning algorithm to automatically select the repair scheme with the lowest cost and compliance with building codes. Steps S301-S302 are repeated until the difference index and cost data meet the preset requirements.
[0095] The training process of the reinforcement learning decision-making model includes,
[0096] S4021: Define the state space as the current model deviation index and cost data, and the action space as the set of repair solutions, which includes component size adjustment, adding new components, and adding or removing equipment.
[0097] S4022: Define the reward function
[0098] ;
[0099] Where α and β are weighting coefficients, α takes the value of 0.6, β takes the value of 0.4, and ΔLmax is the maximum allowable spatial deviation, ΔLmax takes the value of 0.1m;
[0100] S4023: The model is trained iteratively using the PPO (Proximal Policy Optimization) algorithm. In each iteration, a repair plan is selected based on the current state, the reward value is calculated, and the model parameters are optimized. When the average reward value of 100 consecutive iterations is ≥0.85, the model training is complete, and the optimal repair plan can be automatically output. The PPO algorithm avoids training oscillations by limiting the policy update amplitude, making it suitable for dynamic decision optimization of repair plans in this invention.
[0101] Step 5, Visualization and Collaborative Management: Integrate difference heatmaps and cost fluctuation curve visualization tools into the BIM platform to support multi-role collaborative decision-making; output standardized reports, including remediation logs, cost change details, and compliance verification results.
[0102] The environmental data collected by the IoT sensor includes at least one of temperature, humidity, light intensity, and personnel flow data, which is used in step S202 to assist in verifying the authenticity of the equipment redundancy or missing defects identified by the defect detection sub-model, and in step S402 as auxiliary input to the state space of the reinforcement learning decision model to optimize the repair scheme.
[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention. All equivalent changes made based on the description and drawings of the present invention are included within the scope of the present invention.
Claims
1. A method for integrating BIM models and design estimates based on AI, characterized in that, Includes the following steps: Step 1, Multi-source Data Acquisition and Fusion: Texture data of the building entity is acquired using a panoramic camera, geometric data of the building entity is acquired using a laser scanner to generate a laser scan point cloud, and environmental data of the building entity is acquired using IoT sensors. The texture data, laser scan point cloud, and environmental data are combined to construct a multimodal dataset. The original BIM model constructed using BIM design software in the early stages of the project is obtained and converted into an original BIM model point cloud. The multimodal dataset and the original BIM model point cloud are virtually fused to generate an enhanced BIM model with spatial coordinates. Step 2, AI model training and defect identification: S201: AI Model Training: Train a multi-task deep learning model, which includes a defect detection sub-model and a cost association sub-model. The defect detection sub-model is built based on the YOLO algorithm or Mask R-CNN algorithm and is used to identify equipment missing, equipment redundancy, and geometric distortion defects in the BIM model during the enhancement and subsequent repair process. The cost association sub-model uses natural language processing technology to parse the design documents, extract the engineering quantity information and cost keywords in the documents, and construct the component-cost mapping relationship between components and cost items in the BIM model. S202: Input the enhanced BIM model, the defect detection sub-model of the multi-task deep learning model identifies defects in the enhanced BIM model and outputs the defect type, defect location and defect severity; Step 3, Dynamic Repair and Cost Synchronization: Based on the defect type, defect location and defect severity output in Step 2, call the predefined repair strategy or generate new components through generative AI to repair and enhance the BIM model to obtain the repaired BIM model; call the cost association sub-model in the multi-task deep learning model to parse the repaired BIM model in real time, update the bill of quantities and unit price library in linkage, and generate a revised budget report. Step 4, Difference Assessment and Closed-Loop Optimization: S401: Calculate the difference index between the enhanced BIM model and the repaired BIM model, including spatial deviation ΔL and equipment quantity deviation ΔN, and combine it with the preset cost tolerance threshold to determine whether secondary optimization is triggered; S402: If secondary optimization is required, a dynamic decision-making model is trained based on reinforcement learning algorithm to automatically select the repair scheme with the lowest cost and compliance with building codes. Repeat steps S301-S302 until the difference index and cost data meet the preset requirements. Step 5, Visualization and Collaborative Management: Integrate difference heatmaps and cost fluctuation curve visualization tools into the BIM platform to support multi-role collaborative decision-making; output standardized reports, which include remediation logs, cost change details, and compliance verification results.
2. The method for integrating AI-based BIM models and design estimates according to claim 1, characterized in that: In step 1, the specific process of virtual fusion includes, S101: Point cloud registration. The iterative nearest point algorithm is used to register the laser scanned point cloud with the original BIM model point cloud. The point cloud registration uses the ICP algorithm. Through the iterative nearest point algorithm, the optimal spatial transformation matrix between the laser scanned point cloud and the original BIM model point cloud is calculated, so that the points in the laser scanned point cloud and the original BIM model point cloud representing the same building component or the same spatial location are aligned in three-dimensional space. S102: Image semantic segmentation. The texture image of the building entity is classified at the pixel level using the U-Net network. That is, each pixel in the image is classified, and the pixel regions belonging to walls, floors, doors and windows, and electromechanical equipment are automatically identified and extracted. The semantic information of the building elements is obtained based on the extracted pixel regions. At the same time, the two-dimensional image coordinates of the elements corresponding to the semantic information of the building elements are obtained. Then, combined with the three-dimensional spatial coordinates provided by the laser scanning point cloud, the two-dimensional image coordinates of the elements are converted into three-dimensional semantic information with three-dimensional spatial positioning. The three-dimensional semantic information is associated and matched with the spatial coordinates of the corresponding components in the original BIM model to obtain the enhanced BIM model.
3. The method for integrating AI-based BIM models and design estimates according to claim 2, characterized in that: In S102, obtaining semantic information of building elements based on the extracted pixel regions specifically includes: obtaining the outline range and material texture features of the wall in the image based on the extracted wall pixel regions; obtaining the planar position and thickness visual features of the floor slab in the image based on the extracted floor slab pixel regions; and obtaining the installation position and size ratio features of the doors and windows in the image based on the extracted door and window pixel regions. Based on the extracted pixel regions of electromechanical equipment, the shape contour and installation location features of the electromechanical equipment in the image are obtained; among them, the extracted features are the semantic information of building elements.
4. The method for integrating AI-based BIM models and design estimates according to claim 1, characterized in that: In step 201, the training process of the cost-related sub-model includes, S2011, Data Preprocessing: Clean unstructured data such as design documents and contract texts, remove redundant information, and extract engineering quantity keywords and cost keywords through word segmentation and part-of-speech tagging. Among them, engineering quantity keywords include component size, component quantity, and component material, and cost keywords include component unit price, project total price, and billing standard. S2012, Constructing a Knowledge Graph: Using components in the original BIM model as nodes, setting cost items, with cost items as node attributes, including engineering quantity calculation rules, component unit price, and cost accounting formula, establishing semantic association between components and cost items, supporting automatic mapping from engineering quantity keywords and cost keywords to corresponding cost data, forming an initial component-cost mapping relationship; S2013, Model Training: The cost association sub-model is initialized using the BERT pre-trained model. The cost association sub-model is trained by importing component-cost matching data from historical projects. Training stops when the F1-Score of the cost association sub-model on the validation set is greater than or equal to the preset standard. The final cost association sub-model is then output. When the final cost association sub-model takes component parameters and keywords as input, it directly outputs the final component-cost mapping relationship.
5. The method for integrating AI-based BIM models and design estimates according to claim 1, characterized in that: Step 3 specifically includes: S301, Model Repair: Call the defect type, defect location, and defect severity information output by the defect detection sub-model in step S202, and perform corresponding repair operations for different defect types; if it is a missing equipment defect, complete the equipment according to the parameters of the surrounding components and the building's functional requirements; if it is a redundant equipment defect, delete the redundant equipment and adjust the pipeline connection relationship associated with the redundant equipment; if it is a geometric distortion defect, correct the size and spatial position of the defective component based on the geometric data of the enhanced BIM model, and complete the BIM model repair; obtain the repaired BIM model; S302, Budget Update: The component-cost mapping relationship output from the cost-related sub-model in step S201 is invoked to analyze the changes in the repaired BIM model in S301 in real time. When new components are added to the repaired BIM model, the unit price data and quantity calculation rules corresponding to the new components are matched through the component-cost mapping relationship to calculate the cost of the new components. When the dimensions of components in the repaired BIM model change, the quantity of the changed components is recalculated according to the quantity calculation rules in the component-cost mapping relationship, and the component cost is adjusted based on the corresponding unit price data. When the repaired BIM model involves the addition or removal of equipment, the unit price of the added or removed equipment is extracted based on the component-cost mapping relationship, and the cost difference caused by the addition or removal of equipment is calculated. Based on the calculation results, the bill of quantities and unit price library are updated in conjunction, and a revised budget report is generated.
6. The method for integrating AI-based BIM models and design estimates according to claim 5, characterized in that: In step S301, the repair process for geometric distortion defects includes, S3011: Extract the geometric data of the enhanced BIM model corresponding to the defect location output in step S202, including the component vertex coordinates and outline dimensions; S3012: Compare the geometric parameters of the defective component in the original BIM model with the geometric data of the enhanced model, and calculate the geometric deviation value; S3013: Based on the allowable range of component size deviation in the building code, adjust the vertex coordinates and outline dimensions of the defective component so that the deviation between the geometric parameters of the repaired component and the geometric data of the enhanced model is ≤0.03m, thus completing the repair of geometric distortion defects.
7. The method for integrating AI-based BIM models and design estimates according to claim 1, characterized in that: In step S401, spatial deviation for, = ； Equipment quantity deviation for, ； The difference index is calculated for components with identified defects in the enhanced BIM model and their associated components that are affected in terms of physical connection or function. >0.1m or When the value is greater than 1, the model deviation is determined to be excessive, triggering a second optimization.
8. The method for integrating AI-based BIM models and design estimates according to claim 1, characterized in that: In step S402, the training process of the reinforcement learning decision model includes, S4021: Define the state space as the current model deviation index and cost data, and the action space as the set of repair solutions, which includes component size adjustment, adding new components, and adding or removing equipment. S4022: Define the reward function ； Where α and β are weighting coefficients, α takes the value of 0.6, β takes the value of 0.4, and ΔLmax is the maximum allowable spatial deviation, ΔLmax takes the value of 0.1m; S4023: The model is trained iteratively using the PPO algorithm. In each iteration, a repair plan is selected and executed based on the current state. The reward value is calculated and the model parameters are optimized. When the average reward value of 100 consecutive iterations is ≥0.85, the model training is complete and the optimal repair plan can be automatically output.
9. The method for integrating AI-based BIM models and design estimates according to claim 1, characterized in that: Generative AI employs a Transformer-based generative adversarial network or diffusion model to automatically generate new component models that conform to engineering logic based on the defect location context and building codes.
10. The method for integrating AI-based BIM models and design estimates according to claim 1, characterized in that: The environmental data collected by the IoT sensor includes at least one of temperature, humidity, light intensity, and personnel flow data. It is used in step S202 to assist in verifying the authenticity of the equipment redundancy or missing defects identified by the defect detection sub-model, and in step S402, it is used as auxiliary input to the state space of the reinforcement learning decision model to optimize the repair plan.