Real model consistency review system and method based on large language model

Abstract

The invention discloses a real model consistency review system and method based on a large language model, according to the scheme, through integration of multi-source heterogeneous data, standardized preprocessing of real scene cloud or images is completed, and through combination of deep learning recognition and multi-scale feature analysis, accurate correspondence and intelligent comparison of a real model and a BIM component are achieved; performing semantic reasoning and logic verification by using a large language model, automatically detecting differences such as dimensional deviation, positioning deviation, missing and redundant components and the like, and generating structured difference data; furthermore, difference information is visually presented through graphical annotation and natural language description, and batch screening and multi-scale interaction are supported. Based on a difference analysis result, a standardized review report is automatically compiled, and statistical analysis, change trend tracking and multi-format export functions are provided. According to the scheme, the consistency review efficiency and reliability in a large-scale data environment can be guaranteed, and the quality management level of building construction and operation and maintenance stages can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a real-model consistency review technology based on a large language model. Background Art

[0002] With the continuous improvement of the informatization and digitalization level of the construction industry, building information modeling technology has been widely used in the entire process of project design, construction, and operation and maintenance. Through BIM models, the geometric information, component properties, and construction procedures of buildings can be intuitively expressed in a virtual environment, effectively improving the standardization and efficiency of project management. However, BIM models usually remain in the planning and design stages. During the actual construction process, due to factors such as construction process deviations, material changes, environmental impacts, and human operational errors, there are often varying degrees of inconsistency between the actual built object and the BIM model, such as dimensional errors, positioning offsets, missing components, and additional components.

[0003] Traditional consistency reviews mainly rely on manual on-site verification and comparison, which is not only inefficient and labor-intensive, but also easily subject to subjective judgment, resulting in frequent missed detections and false detections, making it difficult to meet the high requirements of accuracy and real-time performance for large and complex projects.

[0004] In recent years, with the development of computer vision and artificial intelligence technologies, automated review methods based on image processing and point cloud analysis have gradually emerged. However, existing methods often have limitations such as poor adaptability, insufficient environmental robustness, and the ability to only handle simple rule differences. In addition, they lack an in-depth understanding of the logical relationships of buildings, resulting in limited accuracy in detection results and an inability to fully support refined construction quality management. Summary of the Invention

[0005] In response to the problems of existing real-model consistency review schemes in building information modeling in terms of review efficiency and accuracy, the purpose of the present invention is to provide a real-model consistency review scheme based on a large language model. This scheme is based on the recognition and reasoning capabilities of the large language model, combined with computer vision technology, to compare the consistency of photos taken on site or real-site point cloud data with the pre-built BIM model, automatically detect the differences, and achieve high-efficiency and high-precision difference detection between the real model and the BIM model.

[0006] In order to achieve the above-mentioned object, the present invention provides a real model consistency review system based on a large language model, the real model consistency review system comprising:

[0007] A data integration and preprocessing unit, which is used to obtain image data and / or three-dimensional point cloud data of the on-site building, as well as the BIM model file corresponding to the on-site building, preprocess the obtained image data and / or three-dimensional point cloud data to form real model data, and perform structured analysis on the BIM model to disassemble the entire BIM model into several component units and extract attribute information;

[0008] A real model identification and comparison analysis unit, which uses deep learning visual recognition technology to perform component-level identification and semantic segmentation on the preprocessed real model data. For the identified real model components, the unit extracts their key attributes, performs a preliminary match based on the extracted key attributes with the corresponding component unit entries in the BIM database, and uses a large language model to perform semantic reasoning and logical verification on the preliminary matching results based on constraint conditions. It determines whether there are any differences between the successfully matched real model components and the corresponding component units in the BIM database according to the set detection rules;

[0009] a difference marking and rendering unit, configured to present the difference results detected by the real model recognition and comparison analysis unit in a visual manner;

[0010] The report generation and statistical analysis unit is used to generate a complete and standardized report file in a structured and standardized manner based on the analysis results output by the difference annotation and rendering unit, and to perform multi-dimensional statistical analysis on the overall difference situation.

[0011] Furthermore, the data integration and preprocessing unit includes a data acquisition module, a data preprocessing module, and a BIM database construction module.

[0012] The data acquisition module is capable of acquiring image data and / or three-dimensional point cloud data of on-site buildings, and importing BIM model files corresponding to on-site buildings;

[0013] The data preprocessing module is configured to interact with the data acquisition module, perform standardization processing on the image data, and perform preliminary geo-registration according to the spatial coordinate system set on site; perform coordinate system conversion, point density homogenization and smoothing filtering on the three-dimensional point cloud data;

[0014] The BIM database construction module is configured to interact with the data acquisition module, perform structured analysis on the imported BIM model file, decompose the overall BIM model into several component units, extract attribute information, and store it as standardized database entries to form a BIM database.

[0015] Furthermore, the data integration and preprocessing unit further includes an abnormal data detection module, and the abnormal data detection module is configured to perform data interaction with the data acquisition module and automatically identify abnormal data in the acquired data.

[0016] Furthermore, the real model identification and comparison analysis unit includes a real model component identification module, a preliminary matching module, a matching verification module, and an anomaly detection module.

[0017] The real model component recognition module performs component-level recognition and semantic segmentation on the pre-processed real model data based on deep learning visual recognition technology. The recognition results include component type, size parameters, spatial position and direction attributes;

[0018] The preliminary matching module is configured to interact with the real model component identification module data, extract key attributes of the identified real model component through a feature extraction algorithm, and perform preliminary matching of the extracted key attributes with corresponding component unit entries in the BIM database;

[0019] The matching verification module is configured to interact with the preliminary matching module data, and use the large language model to perform semantic reasoning and logical verification on the preliminary matching results based on the constraints to determine whether the actual model components are correct with the corresponding component units in the preliminary matching BIM database;

[0020] The anomaly detection module is configured to interact with the matching verification module data, calculate the size difference, positioning deviation, and morphological consistency index between each pair of successfully matched real model components and the corresponding component units in the BIM database, and determine whether there is a difference based on the set detection rules.

[0021] Furthermore, the difference annotation and rendering unit includes a difference item sorting module and an annotation production module.

[0022] The difference item sorting module is used to classify and sort the detected difference data according to component category, difference type and severity, and construct a difference item list;

[0023] The annotation production module interacts with the difference item sorting module in data, and automatically generates graphical annotations and / or natural language descriptions based on the spatial positions corresponding to the differences in the real model image or three-dimensional point cloud model formed by the data integration and preprocessing unit.

[0024] Furthermore, the report generation and statistics unit includes a difference data extraction module, an audit report production module, and a difference data analysis module.

[0025] The difference data extraction module is used to extract all difference data from the difference marking and rendering unit and archive and organize them according to preset dimensions;

[0026] The review report production module is configured to interact with the difference data extraction module to automatically compile and generate a review report based on a preset report template for the difference data collated by the difference data extraction module;

[0027] The difference data analysis module is configured to interact with the difference data extraction module to perform multi-dimensional mining on the difference data collated by the difference data extraction module and to visually present the analysis results in the form of charts.

[0028] Furthermore, the real model consistency review system also includes a system management and operation interface unit, which is configured to interact with the data integration and preprocessing unit, the real model identification and comparison analysis unit, the difference annotation and rendering unit, and the report generation and statistical analysis unit to complete the data flow, authority control, task scheduling and user interaction of the real model consistency review system for the consistency review process.

[0029] In order to achieve the above-mentioned object, the present invention provides a real model consistency review method based on a large language model, the real model consistency review method comprising:

[0030] The on-site building corresponds to the BIM model file, and the BIM model is structurally analyzed, the overall BIM model is disassembled into several component units, and attribute information is extracted to form a BIM database;

[0031] Acquire image data and / or three-dimensional point cloud data of the on-site building, and pre-process the acquired image data and / or three-dimensional point cloud data to form real model data;

[0032] Based on deep learning visual recognition technology, the pre-processed real model data is identified and semantically segmented at the component level. For the identified real model components, their key attributes are extracted and preliminarily matched with the corresponding component unit entries in the BIM database based on the extracted key attributes.

[0033] Use a large language model to perform semantic reasoning and logical verification on the initial matching results based on constraints;

[0034] Determine whether there are any differences between the successfully matched real model components and the corresponding component units in the BIM database based on the set detection rules;

[0035] Output and display the detected differences.

[0036] Furthermore, when pre-processing the acquired image data and / or three-dimensional point cloud data, the real model consistency review method performs standardization processing on the image data and performs preliminary geographic alignment based on the spatial coordinate system set on site; and performs coordinate system conversion, point density homogenization and smoothing filtering processing on the three-dimensional point cloud data.

[0037] Furthermore, in the real model consistency review method, the size difference, positioning deviation, and morphological consistency index are calculated between each pair of successfully matched real model components and the corresponding component units in the BIM database, and whether there is a difference is determined according to the set detection rules.

[0038] The solution provided by the present invention is based on the recognition and reasoning capabilities of a large language model, combined with computer vision technology, to compare the consistency of photos taken on-site or on-site point cloud data with the pre-built BIM model, automatically detect the differences, and comprehensively display the difference areas in the form of graphic annotations and natural language descriptions, and automatically generate statistical analysis and difference analysis reports. The process is efficient and automated, effectively realizing high-precision, full-scene intelligent difference detection and investigation between the real model and the BIM model.

[0039] The solution provided by the present invention can adopt a modular design in application, support multi-project management, version control and permission allocation, and be equipped with a responsive operation interface and intelligent reminder mechanism to ensure the efficiency and reliability of consistency review in a large-scale data environment.

[0040] The solution provided by the present invention can effectively improve the quality management level in the construction and operation and maintenance stages, and provide key technical support for the realization of digital twins and intelligent construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0042] Figure 1 This is a diagram illustrating an example of the structure of a real model consistency review system based on a large language model in the present invention;

[0043] Figure 2 This is an example diagram of the structure of the data integration and pre-processing unit in the present invention;

[0044] Figure 3 This is a diagram illustrating an example of the structure of a real model identification and comparison analysis unit in the present invention;

[0045] Figure 4 This is an example diagram of the structure of the difference marking and rendering unit in the present invention;

[0046] Figure 5 This is an example diagram of the report generation and statistical analysis structure in the present invention. DETAILED DESCRIPTION

[0047] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below with reference to specific illustrations.

[0048] This invention conducts in-depth research and analysis on the data characteristics of real model consistency review in building information modeling, innovatively introduces a large language model and integrates computer vision technology to construct a real model consistency review solution that integrates the reasoning advantages of the large language model and the accuracy of visual recognition. It can efficiently and accurately realize the difference detection, annotation and report generation between the real model and the BIM model, and promote the intelligent transformation and digital twin construction of the construction industry.

[0049] See also Figure 1 , which shows an example of the structure of the real model consistency review system 100 based on the large language model given by the present invention.

[0050] Based on the diagram, the real model consistency review system 100 based on the large language model is mainly composed of a data integration and preprocessing unit 110, a real model recognition and comparison analysis unit 120, a difference annotation and rendering unit 130, and a report generation and statistical analysis unit 140.

[0051] The data integration and preprocessing unit 110 in this system is primarily used to collect, clean, standardize, and organize and store relevant data. Specifically, this unit is used to obtain image data and / or 3D point cloud data of on-site buildings, as well as the corresponding BIM model files of on-site buildings, preprocess the acquired image data and / or 3D point cloud data to form real model data, perform structured analysis on the BIM model, disassemble the overall BIM model into several component units, and extract attribute information to form a BIM database.

[0052] The real model recognition and comparison analysis unit 120 in this system interacts with the data integration and preprocessing unit 110 to perform component-level recognition and semantic segmentation on the preprocessed real model data based on deep learning visual recognition technology. For the identified real model components, their key attributes are extracted, and a preliminary match is performed with the corresponding component unit entries in the BIM database based on the extracted key attributes. The large language model is further used to perform semantic reasoning and logical verification on the preliminary matching results based on constraint conditions to determine whether the real model components and the corresponding component units in the BIM database are successfully matched; for the successfully matched real model components and the corresponding component units in the BIM database, whether there are differences is determined based on the set detection rules.

[0053] The difference annotation and rendering unit 130 in this system interacts with the real model recognition and comparison analysis unit 120 to present the difference results detected by the real model recognition and comparison analysis unit 120 in a visual manner.

[0054] The report generation and statistical analysis unit 140 in this system interacts with the difference annotation and rendering unit 130 to form a complete and standardized report file in a structured and standardized manner based on the analysis results output by the difference annotation and rendering unit 130, while performing multi-dimensional statistical analysis on the overall difference situation.

[0055] The composition of each functional unit in the large language model-based real model consistency review system 100 and the structures that may be involved are further described in detail below.

[0056] See also Figure 2 The data integration and preprocessing unit 110 in this system is mainly composed of a data acquisition module 111, a data preprocessing module 112, and a BIM database construction module 113 that cooperate with each other.

[0057] The data acquisition module 111 constitutes the data source input interface of the entire system and supports access to multi-source heterogeneous data.

[0058] Specifically, the data acquisition module 111 can obtain image data of on-site buildings, where the image data includes multi-angle photos and videos taken at the construction site.

[0059] The data acquisition module 111 can also acquire three-dimensional point cloud data of on-site buildings, such as three-dimensional point cloud data of buildings acquired through laser scanning, photogrammetry, etc.

[0060] To ensure the reliability and accuracy of subsequent data processing, the image data of the on-site buildings acquired by the data acquisition module 111 needs to be high-definition, unobstructed, and fully covered. It is preferred to use a camera device with calibration parameters for multi-angle acquisition.

[0061] The point cloud data acquired by the data acquisition module 111 is required to have uniform density, low noise, and spatial coordinate information. It is preferably acquired using a laser scanner with millimeter-level accuracy or high-precision drone photogrammetry to ensure the accuracy of subsequent processing.

[0062] Furthermore, the data acquisition module 111 can also simultaneously import BIM model files, specifically supporting formats such as IFC, RVT, DGN, etc.

[0063] Imported BIM model files should be generated by professional modeling software, with complete component information, a clear hierarchical structure, and a unified coordinate system and naming conventions. Importing requires a pre-processing process, including format standardization, semantic parsing, and component decomposition, to ensure accurate matching of the model data with the actual model data. Mainstream formats such as IFC, RVT, and DGN are supported, with the open standard IFC being preferred for enhanced system compatibility.

[0064] The data preprocessing module 112 in this unit is configured to perform data interaction with the data acquisition module 111 and to preprocess the data collected or imported by the data acquisition module 111 so as to adapt it to subsequent recognition and comparison processing.

[0065] Specifically, the data preprocessing module 112 performs corresponding preprocessing on different categories of data collected by the data collection module 111 .

[0066] The data preprocessing module 112 first performs a data quality check on the real model data (ie, image data and / or three-dimensional point cloud data of the on-site building) collected by the data collection module 111 to ensure the integrity and accuracy of the data.

[0067] The data preprocessing module 112 automatically screens low-quality data such as blurred images, underexposure, and severe occlusion for the image data of the collected on-site buildings; the data preprocessing module 112 removes noise and outliers for the point cloud data of the collected on-site buildings, thereby ensuring that the input data has the integrity and accuracy required for subsequent processing.

[0068] Furthermore, when performing quality detection on image data, the data preprocessing module 112 automatically identifies blur, exposure anomalies and occlusion areas in the image by introducing an image quality assessment model based on a deep convolutional neural network, and performs multi-index comprehensive screening by combining image clarity scoring, histogram distribution analysis and occlusion detection algorithm.

[0069] When processing point cloud data, the data preprocessing module 112 uses statistical filtering and density-based clustering algorithm DBSCAN to remove noise and outliers, and combines normal vector consistency detection to improve the smoothness and continuity of the point cloud.

[0070] Here, intelligent image analysis is innovatively combined with spatial structure recognition to achieve adaptive quality control of multimodal data and improve the automation and robustness of preprocessing.

[0071] After completing the data quality detection and screening, the data pre-processing module 112 further performs standardization processing on the detected and screened image data, and performs preliminary geo-registration according to the spatial coordinate system set on site.

[0072] This data preprocessing module 112 incorporates image enhancement technologies such as automatic exposure correction, color balancing, and distortion correction into image standardization to ensure uniformity in brightness, color, and geometry across multiple sources. It also combines embedded EXIF ​​metadata with on-site calibration targets to automatically georeference the images within a unified spatial coordinate system.

[0073] By integrating an automatic calibration algorithm based on image feature point matching with known station locations, we achieve spatial alignment between the image and the BIM model, significantly improving the accuracy and efficiency of subsequent component identification and spatial comparison. This processing solution is effective for rapid alignment of indoor spaces in GPS-free environments.

[0074] Furthermore, the data preprocessing module 112 performs coordinate system conversion, point density homogenization and smoothing filtering on the three-dimensional point cloud data after noise removal and outlier removal, thereby improving the consistency and accuracy of the spatial data.

[0075] When processing three-dimensional point cloud data, this data preprocessing module 112 introduces a multi-level coordinate system conversion mechanism, supports automatic conversion from the local measurement coordinate system to the global coordinate system consistent with the BIM model, and combines the ICP iterative closest point algorithm and feature alignment technology to improve the registration accuracy; when uniformizing the point density, voxel grid downsampling combined with an adaptive resampling algorithm is used to ensure consistent point spacing in different scanning areas, which is conducive to subsequent component identification; when performing smoothing filtering, a bilateral filtering method based on curvature estimation is used to remove local noise while retaining boundary details.

[0076] This data preprocessing module 112 significantly improves the structural consistency and spatial stability of point cloud data by synergistically integrating multiple algorithms, taking into account both accuracy and efficiency, and adapting to the high-quality modeling needs in complex construction environments.

[0077] The BIM database construction module 113 in this unit is configured to interact with the data acquisition module 111 for data exchange, and is used to perform structured analysis on the BIM model text imported by the data acquisition module 111, decompose the overall BIM model into several component units, and extract attribute information, such as component name, number, type, size, positioning coordinates, material properties, etc., and store them as standardized database entries to form a BIM database, which is convenient for subsequent rapid matching and retrieval.

[0078] Specifically, the BIM database construction module 113 calls component-level geometry and semantic data by parsing the native file structure of the BIM model (such as IFC or RVT format), automatically traverses the model using the IFC entity tree structure or Revit API, and decomposes the entire BIM into component units.

[0079] On this basis, key fields such as the unique identifier, category, geometric parameters, positioning information, and material properties are extracted for each component node. Semantic mapping and regularization filtering are used in this attribute extraction process to avoid inconsistent model naming.

[0080] Finally, the extracted results are stored in a structured format (such as JSON or relational database table) to build a standardized BIM database.

[0081] This processing of the BIM database construction module 113 effectively realizes the semantic alignment basis of BIM data and heterogeneous real model data, significantly improves the efficiency of subsequent automatic matching and comparison, and supports cross-platform and cross-format data consistency processing.

[0082] In terms of data organization, this BIM database construction module 113 establishes a unified data indexing mechanism, associates each real model unit with the BIM component unit through a unique ID, and records metadata such as data source, collection time, location information, processing version, etc., to achieve data traceability and version control.

[0083] Specifically, the BIM database construction module 113 adopts a unified index generation mechanism to generate a globally unique ID (UID) for each component during the BIM database construction process. The UID consists of the component category, model source identifier and spatial location code to ensure uniqueness and interpretability.

[0084] Based on this, during the real model recognition process, the corresponding BIM component UID can be bound to the identified real model unit based on the results of spatial registration and component semantic matching, and metadata such as acquisition time, equipment number, and data version can be simultaneously recorded. All index information is stored in a relational database and a multidimensional index structure (such as spatial index + time index) is established to enable rapid retrieval, comparison, and updating.

[0085] Based on this mechanism, the accuracy of data association between the real model and BIM is improved, data traceability, reproducibility and version management are achieved, and technical support is provided for subsequent difference tracking and change analysis.

[0086] To improve processing efficiency, this BIM database construction module 113 introduces data slicing and partitioning management strategies, logically divides the real model and BIM data according to geographical location, floor area or construction stage, and supports local loading, parallel processing and incremental updates.

[0087] Specifically, the BIM database construction module 113 establishes slicing rules based on the spatial structure hierarchy (e.g., floors, partitions, and component systems) within the BIM model, logically dividing the overall BIM model into several sub-blocks. Simultaneously, the spatial mapping relationship between the real model and the BIM is established, combining the geographic location of the real model data with the photographed / scanned calibration information. On this basis, a spatial partitioning tree is used to rapidly locate and index component data. The processing supports loading sub-models by region, performing comparison analysis only on modified blocks, enabling incremental updates and parallel computing.

[0088] This strategy significantly reduces the memory consumption and computing time of large-scale data processing, improves the system response speed and scalability, and adapts to the efficient computing needs in complex scenarios.

[0089] As a further explanation, the data integration and preprocessing unit 110 may also be configured with an abnormal data detection module 114, which can interact with the data acquisition module 111 for data identification of the real model data (i.e., image data and / or three-dimensional point cloud data of the on-site building) collected by the data acquisition module 111, i.e., identifying abnormalities in the real model data, such as missing areas, scanning blind spots or data discontinuity, and generating an abnormality report to prompt the user to re-sample or correct.

[0090] Specifically, the abnormal data detection module 114 identifies sparse or empty areas of the point cloud based on acquisition trajectory analysis and point cloud coverage density detection, combined with a spatial grid division algorithm; at the same time, it detects blind spots for image data through perspective coverage analysis and image overlap calculation.

[0091] The abnormal data detection module 114 also combines the spatial component distribution of the BIM model to compare which areas have missing data.

[0092] The detection results generated by the abnormal data detection module 114 prompt the user to re-sample in the form of graphic annotations and reports.

[0093] The abnormal data detection module 114 can effectively enhance the spatial completeness of the collected data and improve the accuracy and stability of subsequent processing.

[0094] The data integration and preprocessing unit 110 thus formed can effectively ensure the accuracy, stability and efficiency of the subsequent identification, comparison and analysis processing links of the system during operation, and provide a reliable data basis for the accuracy of the actual model consistency review results.

[0095] The real model recognition and comparison analysis unit 120 in this system is the main functional unit for realizing accurate matching and difference detection between real scene data and BIM models in this system solution.

[0096] See also Figure 3 The real model recognition and comparison analysis unit 120 in this system is mainly composed of a real model component recognition module 121, a preliminary matching module 122, a matching verification module 123 and an anomaly detection module 124 in specific implementation.

[0097] The real model component recognition module 121 in this unit performs component-level recognition and semantic segmentation on the pre-processed real model data based on deep learning visual recognition technology. The recognition results include component type, size parameters, spatial position and direction attributes.

[0098] Specifically, the real-model component recognition module 121 utilizes Mask R-CNN, a deep semantic segmentation model based on a convolutional neural network architecture, to perform end-to-end training and inference on image or point cloud data to achieve component-level recognition. For the image recognition path, a multi-scale feature extraction network is used to extract component edge and morphological features, and a spatial attention mechanism is incorporated to enhance the detection of small or occluded components. For the point cloud path, models such as PointNet++ are used to classify and aggregate component point clouds in three-dimensional space.

[0099] The output of the real model component identification module 121 includes component type labels, size parameters, spatial position and orientation attributes.

[0100] This real model component recognition module 121 can significantly improve the integrity and accuracy of component recognition in complex construction scenarios, providing a reliable basis for subsequent comparison and difference analysis.

[0101] The component types identified here mainly include beams, columns, walls, and floor slabs.

[0102] Furthermore, in order to improve the recognition accuracy in complex scenarios, a multi-scale feature fusion network and attention mechanism are introduced into the real model component recognition module 121 to effectively enhance the detection capability of small components and occluded components.

[0103] Specifically, the multi-scale feature fusion network in this real-model component recognition module 121 extracts image or point cloud features at different resolutions by constructing a feature pyramid (FPN) or using an encoder-decoder structure (U-Net), thereby fusing shallow details with deep semantics. Each level of features is upsampled and concatenated channel by channel to enhance the recognition of components with large size variations, such as small embedded parts, doors, and windows. Regarding the attention mechanism, the channel attention SE module and the spatial attention CBAM module are introduced to dynamically adjust the feature weights of different regions and channels, enhancing the perception of occluded or blurred components.

[0104] Based on this, the detection robustness and refined segmentation capability of the real model component recognition module 121 in a complex construction site environment can be significantly improved, ensuring more comprehensive and accurate component recognition.

[0105] The preliminary matching module 122 in this unit is configured to interact with the real model component identification module 121 for data. For the real model components identified by the real model component identification module 121, the key attributes thereof are extracted through a feature extraction algorithm, and the extracted key attributes are preliminarily matched with the corresponding component entries in the constructed BIM database.

[0106] The preliminary matching here is roughly screened based on spatial approximate position, type consistency and size similarity, and multiple tolerance thresholds are set to accommodate construction errors.

[0107] Specifically, the preliminary matching module 122 extracts key attributes from the component recognition results, including spatial position (X, Y, Z coordinates), component type (e.g., beam, column, wall), dimensional parameters (length, width, height), and orientation information. Spatial position is calculated using the point cloud centroid or image space projection, dimensional parameters are extracted from the component bounding box, and type is output by the semantic segmentation model.

[0108] When performing preliminary matching, the preliminary matching module 122 searches the BIM database for candidate components that are of the same type, close in position (set spatial tolerance, ±10 cm), and have high size similarity (set ratio threshold), to achieve rough matching screening.

[0109] As a preferred approach, the matching process uses a KD tree to accelerate spatial searches and ensure matching efficiency. By setting multiple tolerances, it can adapt to actual deviations caused by construction errors, improve matching tolerance and coverage, and lay the foundation for subsequent semantic verification.

[0110] Based on this method, the preliminary matching module 122 can balance matching speed and flexibility, significantly reduce false matches, and improve the accuracy and stability of the overall recognition and comparison process.

[0111] The matching verification module 123 in this unit is configured to interact with the preliminary matching module 122 in data, and use the large language model to perform semantic reasoning and logical verification on the preliminary matching results based on constraints to determine whether the actual model components are correct with the corresponding component units in the preliminary matching BIM database, thereby further screening out incorrect matches and significantly improving the matching accuracy.

[0112] The constraints here include construction specifications and building logic, such as beams must be connected to columns, floor slabs must be attached to beams or walls, etc.

[0113] Furthermore, this matching verification module 123 converts the preliminary matching results, including the correspondence between the real model components and their candidate BIM components, into structured text or coding prompts, and inputs them into the large language model through a predefined template for semantic reasoning. The large language model compares and infers the input component type, spatial relationship, connection logic and other information with the embedded architectural semantic rules (such as "beams must be connected to columns" and "floor slabs should be attached to beams or walls") to determine whether the current match conforms to the architectural structural logic and construction specifications. If an unreasonable spatial relationship or logical conflict is found, it is determined to be an incorrect match and is eliminated.

[0114] This matching verification module 123 breaks through the limitations of traditional rule matching on the processing ability of complex logical structures by integrating the context understanding and reasoning capabilities of the language model. It has strong robustness and scalability, and significantly improves the matching accuracy and engineering usability in heterogeneous components and changing scenarios. It is especially suitable for the identification of complex structures and interactive components.

[0115] The anomaly detection module 124 in this unit is configured to interact with the matching verification module 123 to calculate the size difference, positioning deviation, and morphological consistency index between each pair of successfully matched real model components and the corresponding component units in the BIM database, and determine whether there is a difference based on the set detection rules.

[0116] The differences mentioned here include excessive dimensional deviation, excessive positional offset, severe morphological deformation, etc.

[0117] As a further explanation, BIM components that cannot be matched in the actual model are judged as "missing", otherwise they are judged as "redundant".

[0118] For BIM components judged as "missing" in the inspection results, the system marks them as potential construction omissions or unfinished items, and automatically records their component numbers, expected locations and attribute information, generates difference entries, and pushes them to the difference annotation and reporting module for highlighting prompts and rectification suggestions; at the same time, it can trigger the task reminder mechanism to notify the relevant responsible persons to conduct on-site verification. For real-model components judged as "redundant", the system marks them as abnormal additions or construction deviations, conducts risk assessments based on their component types and the areas they are located, and attaches automatically generated processing suggestions to the report, such as "Recommendation to verify drawing changes" or "Recommendation to review construction logs." Such a processing mechanism can effectively achieve closed-loop management of difference information, facilitating problem tracing and quality control.

[0119] Furthermore, for each matched pair of real-model components and BIM components, the anomaly detection module 124 first extracts key data such as geometric dimensional parameters (e.g., length, width, height) and center point coordinates, and calculates three types of difference indicators. Dimensional differences are evaluated using a combination of absolute and relative errors, with a construction tolerance threshold (e.g., ±20mm) set for determination. Positioning deviations are calculated using the Euclidean distance of the component's centroid in three-dimensional space, combined with spatial tolerance limits to determine if there is an offset. Morphological consistency is assessed by aligning the point cloud contour with the BIM geometry boundary using the Hausdorff distance to assess their spatial fit. Any deviation beyond the set standard is considered a deformation.

[0120] Furthermore, this system supports customizable rules, allowing for different detection accuracies based on component types. This technology features the integration of multi-dimensional parameters—geometric, spatial, and topological—to achieve multi-granular, multi-threshold anomaly detection through algorithmic combination, resulting in high adaptability and precision control. Ultimately, it achieves comprehensive identification of component dimensional deviations, positional errors, and morphological variations, enhancing the sophistication and credibility of consistency reviews.

[0121] As a further explanation, the real model identification and comparison analysis unit 120 can cope with large-scale data processing scenarios, adopt a regional division and parallel processing mechanism, and slice the overall comparison task according to floors, areas or functional areas, thereby greatly improving processing efficiency and response speed; after the comparison and analysis is completed, all difference results are classified and sorted, and the component number, category, difference type, difference value and possible cause of each difference are marked, providing standardized input for subsequent difference marking and report generation.

[0122] In order to cope with large-scale data processing scenarios, this solution introduces regional division and parallel processing mechanisms in the real model identification and comparison analysis unit 120 to achieve an efficient and scalable data analysis process. Specifically, this solution first divides the overall model into multiple logical sub-blocks based on the floor structure, component spatial distribution and functional zoning in the BIM model through the spatial index algorithm R-Tree, and assigns a unique partition ID to each sub-block. Subsequently, based on the spatial coordinate information of the collected data, the system automatically classifies the real model image or point cloud data into the corresponding partition to form a set of subtasks to be processed. Within each subtask, the real model component identification module 121, the preliminary matching module 122, the matching verification module 123 and the anomaly detection module 124 run in parallel in a multi-threaded or distributed computing environment to independently complete component identification, matching and difference analysis. The task scheduling module performs dynamic resource allocation and scheduling management based on the current computing resource status, and supports task priority sorting and load balancing. After the analysis is completed, the difference results output by each partition are summarized and classified according to component number, component category, difference type, difference value and possible cause. At the same time, the corresponding spatial partition information and processing version number are recorded to provide structured and standardized data input for subsequent difference annotation and reporting modules.

[0123] This solution combines spatial partitioning with asynchronous computing, which not only significantly reduces system memory pressure and response time, but also significantly improves recognition and comparison efficiency. Its flexible data structure design and decoupling characteristics between modules make the system scalable and maintainable, adapting to the consistency review needs of large-scale data in various types of construction projects.

[0124] The real model recognition and comparison analysis unit 120 formed accordingly can realize accurate recognition and intelligent comparison, and can realize efficient and reliable detection of differences between the real model and the BIM model, providing strong support for quality control and problem tracing during the construction process.

[0125] The difference marking and rendering unit 130 in this system is specifically used to present the difference results detected by the real model recognition and comparison analysis unit 120 in a systematic and visual manner.

[0126] See also Figure 4 The difference annotation and rendering unit 130 is mainly composed of a difference item sorting module 131 and an annotation production module 132 that cooperate with each other.

[0127] The difference item sorting module 131 in this unit is specifically used to classify and sort the detected difference data according to component category, difference type and severity, and construct a difference item list.

[0128] Specifically, the difference item sorting module 131 parses and classifies the difference data output by the anomaly detection module 124 based on a standardized data structure.

[0129] Here, core fields are extracted for each difference record, including component ID, category (such as beam, column, wall), difference type (size, location, shape) and severity index (such as excess ratio or importance level); then a rule engine is used to automatically label and classify the difference type, and the severity is graded and ranked according to the set threshold (such as minor, general, severe); finally, the classification results are stored in a structured difference item form, and an index is built to support efficient retrieval, screening and statistical analysis.

[0130] Based on this, the book-specific item sorting module 131 can achieve semantic clustering and priority sorting of difference data, improve the pertinence and management efficiency of problem tracking, and support subsequent accurate labeling and report generation.

[0131] The annotation production module 132 in this unit is configured to interact with the difference item sorting module 131 in data, and can automatically generate graphical annotations and / or natural language descriptions based on the spatial positions corresponding to the differences in the real model image or three-dimensional point cloud model formed by the data integration and preprocessing module according to the difference item list constructed by the difference item sorting module 131.

[0132] Specifically, the annotation production module 132 reads the difference component position information and category labels provided by the difference item arrangement module 131 to realize automatic positioning and annotation generation in the real model image or the three-dimensional point cloud model.

[0133] Furthermore, this annotation production module 132 uses the geometric center or bounding box of the component in the spatial coordinate system to overlay a highlighted rectangular box or outline on the image, generating colored point clusters or transparent stereo boxes in the point cloud, and distinguishing the type of difference through color coding (e.g., red indicates severe deviation). As an example, the annotation layer uses WebGL rendering technology to achieve high-performance overlay, supporting zooming, rotation, and click interaction. This module also supports the generation of natural language prompts linked to component attributes to assist in explaining the causes and impacts of the differences.

[0134] The marking forms here include rectangular boxes, highlighted outlines, symbol overlays, etc., and different color codes are used to distinguish the difference categories, such as size deviation is marked in yellow, positioning deviation is marked in red, and missing or redundant components are marked in blue.

[0135] Furthermore, during the annotation process, the annotation production module 132 can support multi-scale adaptive adjustment, that is, dynamically adjust the size, thickness and transparency of the annotation graphics according to the user's current viewing zoom level, to ensure that the difference information can be clearly identified at different observation levels.

[0136] Specifically, when implementing multi-scale adaptive annotation, this annotation production module 132 monitors the user's current zoom ratio in real time through view level perception, and dynamically adjusts the graphic parameters of the annotation elements (such as line width, transparency, font size) according to the preset ratio threshold, and achieves a smooth transition through vector interpolation and scale mapping to avoid blocking details at high magnification or blurred annotations at low magnification.

[0137] In addition, this further anchors the annotation elements to the component's three-dimensional coordinates to ensure that there is no misalignment during scaling.

[0138] This implementation can effectively overcome the problem of traditional static annotation easily losing semantic focus, and significantly improve the clarity and interactive friendliness of difference recognition at multiple scales.

[0139] This annotation generation module 132 specifically combines the capabilities of a large language model to generate a corresponding natural language description for each difference.

[0140] The natural language description here includes component name, difference properties, numerical value, possible causes and impact analysis, and supports multi-language output.

[0141] Specifically, when generating a natural language description, the annotation generation module 132 first converts the structured data of the difference item (such as component name, difference type, value, location, severity, etc.) into a standard prompt template and inputs it into the large language model via a preset prompt. Based on its contextual reasoning capabilities, the large language model automatically generates a complete description text that includes an explanation of the nature of the difference, an analysis of possible causes, and an assessment of potential impacts. It supports output in multiple languages, including Chinese and English.

[0142] Based on this, this annotation production module 132 adopts a template + model hybrid drive mechanism to ensure that the generated content strikes a balance between controllability and flexibility. It can automatically generate semantically complete, logically clear, and professional description content, significantly reducing the burden of manual annotation, and improving the comprehensibility of difference information and cross-language communication capabilities, making it suitable for international engineering scenarios.

[0143] Furthermore, the annotation production module 132 simultaneously overlays all generated graphical annotations and language descriptions onto a unified real model display interface, so that when used, detailed information can be viewed by clicking or hovering over the difference annotations.

[0144] Specifically, this annotation production module 132 implements the simultaneous overlay of graphical annotations and natural language descriptions, supporting click / hover interactions. It constructs an integrated mapping relationship between "graphic annotations" and "attribute descriptions" based on component IDs. Combined with asynchronous rendering and event monitoring mechanisms, this allows each annotation element to be highly linked to its corresponding description, supporting real-time positioning and information pop-up from different viewing angles. It also incorporates occlusion avoidance and dynamic layout optimization algorithms to ensure that information is not lost when multiple annotations overlap.

[0145] Based on this implementation scheme, the annotation production module 132 can break through the limited interaction problem of the traditional static annotation interface and significantly improve the interactivity, readability and user experience during the actual model review process.

[0146] As a further explanation, the difference annotation and rendering unit 130 may further be provided with an annotation presentation and interaction module 133 as needed. The annotation presentation and interaction module 133 cooperates with the annotation production module 132 and can perform various forms of interactive viewing on the graphical annotations and / or natural language descriptions generated by the annotation production module 132.

[0147] Specifically, the annotation presentation and interaction module 133 supports batch annotation and screening for generated annotations, and supports rapid screening of interesting difference items for centralized viewing and processing based on conditions such as difference category, impact level, and component type.

[0148] In the process of realizing the synchronous superposition of graphical annotations and natural language descriptions and supporting interaction, this annotation presentation and interaction module 133 adopts the "annotation-description dual binding mechanism" driven by the component's unique ID to achieve real-time synchronous updates between graphic elements and language prompts; at the same time, it introduces an event proxy mechanism to support precise click or hover positioning in high-density annotation areas, and dynamically display the difference details window.

[0149] In addition, the annotation presentation and interaction module 133 can intelligently adjust the display priority and layout according to the user's operation behavior to avoid information obstruction and interface congestion.

[0150] The annotation presentation and interaction module 133 can significantly improve the interaction efficiency and user experience of the simultaneous display of multiple information, and is particularly suitable for visual review of differences in complex architectural scenes.

[0151] Furthermore, the annotation presentation and interaction module 133 also supports a time series comparison mode, which can track the difference change trends of components at the same position in the real model data taken at multiple time points, and intuitively present the change process in the form of animation.

[0152] When implementing the time series comparison mode, this annotation presentation and interaction module 133 constructs a three-dimensional "component-time-difference" data model by binding a time sequence identifier and difference status record to each component. It also introduces a time frame synchronization control mechanism to enable continuous loading and switching of real-model data at multiple time points. Furthermore, by combining slow motion interpolation and frame difference rendering techniques, the evolution of component differences is smoothly animated, supporting user-defined playback time periods and comparison speeds.

[0153] Compared with the traditional static comparison method, the implementation scheme adopted by this annotation presentation and interaction module 133 shows stronger dynamic interactivity and data expression in construction progress visualization, change traceability and trend analysis, and significantly improves the intuitiveness and intelligence level of project management.

[0154] Furthermore, the annotation presentation and interaction module 133 also supports exporting all difference annotations and rendering information into high-resolution images, interactive 3D models or standard report files for easy archiving, communication and review.

[0155] This annotation presentation and interaction module 133 supports exporting difference annotations and rendering information into high-resolution images, interactive 3D models, and standard report files. This module uses a unified data model to embed graphical annotations, language descriptions, and spatial geometry information in a structured manner within the exported files. Specifically, it enables linking difference annotations and attributes within interactive 3D models (e.g., highlighting components upon clicking them and displaying a difference description). Furthermore, exported reports can be embedded with interactive controls, enabling enhanced user experience, such as 3D rotation and jump viewing within the document.

[0156] Accordingly, the annotation presentation and interaction module 133 can effectively break through the traditional static export limitations, and improve the expressiveness, readability and collaborative review efficiency of the review documents.

[0157] The difference annotation and rendering unit 130 thus formed can effectively realize an efficient closed loop from detection to intuitive presentation of difference data, greatly improving the accuracy and operational convenience of consistency review.

[0158] See also Figure 5The report generation and statistics unit 140 in this system is mainly composed of a difference data extraction module 141, an audit report production module 142, and a difference data analysis module 143 that cooperate with each other.

[0159] The difference data extraction module 141 in this unit is used to extract all difference data and archive and organize them according to preset dimensions.

[0160] The difference data here is provided by the Difference Annotation and Rendering Unit. After completing the comparative analysis, this unit generates structured difference information and graphically annotates it in the actual model view. Based on this result, the Difference Data Extraction Module is called to extract the organized difference items, including component number, location, type, and difference nature, for further archiving, analysis, and report generation. The two form a front-and-back data processing relationship.

[0161] The difference data extraction module 141 can specifically archive and organize the difference data according to key dimensions such as project name, inspection time, inspection area, component type, difference category and severity to ensure data integrity and traceability.

[0162] Specifically, when implementing multi-dimensional data archiving and organization, this difference data extraction module 141 preferably employs an object-oriented data modeling approach. This approach encapsulates each difference record and its associated attributes, such as component ID, timestamp, and detection region, in a nested structure, and constructs a three-level tracking index: "project-component-difference." This structure not only supports efficient screening and cluster analysis, but also allows for tracing back the evolution of differences across any dimension.

[0163] Compared with traditional flat storage, the difference data extraction module 141 can significantly improve the query efficiency, maintainability and information traceability of the difference data throughout the project life cycle.

[0164] The review report production module 142 in this unit is configured to interact with the difference data extraction module 141 and automatically compile and generate a review report in a predetermined format based on a preset report template for the difference data collated by the difference data extraction module 141 .

[0165] This review report production module 142, in its basic implementation, uses a conventional document template engine (such as LaTeX, Docx4j, or an HTML-to-PDF conversion tool) to automatically populate the report content. Furthermore, it integrates the multi-dimensional structured data from the difference data extraction module 141 and, through configurable semantic rules and component category association templates, intelligently generates a highly readable report that includes mixed text and image layout, a dynamic table of contents, numbered cross-indexing, and component jump links. Furthermore, this review report production module 142 automatically generates difference analysis conclusions and recommendations by integrating large language model capabilities, enhancing the professionalism and quality of the report.

[0166] This review report production module 142 not only breaks through the traditional static template editing method, but also realizes the intelligent organization and dynamic optimization of the report content, significantly improving the automatic generation level and application value of the review results.

[0167] Furthermore, the content of the review report generated here includes an overview of the basic project information, a difference overview chart, various types of difference statistical charts (such as pie charts, bar charts), a display of typical difference examples, a detailed difference table, and an analysis summary generated based on a large language model.

[0168] The review report generated here can achieve the technical effect of comprehensive information presentation, highlighted key differences, and clear and intuitive structure, significantly improving the readability and decision-making assistance value of the review.

[0169] Furthermore, each difference in the review report generated here is assigned an independent number and is associated with a screenshot of the specific location, a detailed natural language description, and recommended treatment opinions.

[0170] This review report assigns an independent number to each difference and associates a location screenshot, natural language description and processing suggestions, which has the significant technical effect of improving the accuracy of difference tracing, accelerating problem understanding and assisting rectification decision-making.

[0171] Furthermore, this solution leverages a unique numbering mechanism to assign individual numbers to each difference, ensuring traceability and supporting batch comparison. It also automatically locates the spatial coordinates of the difference components, generating screenshots with overlaid graphical annotations. Furthermore, by integrating a large language model, it generates targeted descriptions and corrective suggestions, requiring input of difference attributes, component information, and contextual context.

[0172] Furthermore, when generating the review report, the review report production module 142 supports intelligent optimized layout, including automatic pagination, mixed text and image typesetting, directory generation and cross-indexing to improve readability, and supports multi-language version output to meet international needs.

[0173] When realizing the intelligent optimization typesetting function of the review report, this review report production module 142 constructs a semantic block structure based on the component difference information and image annotation content, automatically matches the optimal layout template, and realizes the linked typesetting of graphic and text content; and realizes cross-index jump of cross-page difference items through ID mapping and page number tracking mechanism.

[0174] On this basis, the 142 multilingual large model of the review report production module realizes accurate term conversion and context adaptation to ensure the semantic consistency and cultural adaptability of the report.

[0175] Furthermore, the review report production module 142 also supports the generation of interactive electronic reports, based on which the three-dimensional difference distribution can be dynamically viewed in the report interface, click to jump to the specific difference details page, or customize the difference display range according to the filtering conditions.

[0176] This review report production module 142 realizes the interactive electronic report function, establishes a two-way mapping relationship between the component unique ID and the difference attribute, and realizes the linkage between the three-dimensional model and the difference details in the report; uses the embedded view engine to dynamically render the difference distribution, and users can click on the model to jump to the difference item page to view details; the filtering module supports combined filtering by component type, difference category, time and other dimensions.

[0177] Based on this, the review report production module 142 can overcome the problem of weak interactivity of traditional static reports and greatly improve the intuitiveness, navigation efficiency and application flexibility of the report.

[0178] The difference data analysis module 143 in this unit is configured to interact with the difference data extraction module 141 , perform multi-dimensional mining on the difference data collated by the difference data extraction module 141 , and intuitively present the analysis results in the form of charts.

[0179] Specifically, the difference data analysis module 143 can analyze and mine by dimensions such as statistical difference frequency by component type, statistical difference evolution trend by construction stage, and statistical difference density by spatial region.

[0180] When implementing the multi-dimensional statistics and visualization functions of difference data, this difference data analysis module 143 uses the "component-time-space" three-dimensional indexing mechanism to build a data warehouse, and combines difference categories and project stage labels to achieve on-demand aggregation and trend analysis; on this basis, it further introduces a rule-driven data pivot algorithm to support user-defined dimension combinations and chart type switching, thereby improving analysis flexibility and depth.

[0181] Based on this, the difference data analysis module 143 not only enhances the insight of the difference data, but also provides an accurate quantitative basis for quality warning and construction improvement, which is significantly better than the traditional static statistical method.

[0182] Furthermore, the difference data analysis module 143 can also customize statistical dimensions and screening rules to form personalized analysis results.

[0183] This variance data analysis module 143 implements custom statistical dimensions and filtering rules by pre-building a multidimensional data model and introducing a rule-driven configuration engine. This allows users to customize analysis dimensions (such as component material, construction unit, and floor) and filtering conditions (such as variance type, severity, and time period) through a graphical interface. Users can generate customized charts or reports without programming, and the system refreshes views in real time and supports the storage and reuse of conditional combinations.

[0184] Based on this, the difference data analysis module 143 can significantly improve the usability, pertinence and adaptability of data analysis, and is suitable for the difference monitoring needs of multiple projects and multiple roles.

[0185] Furthermore, the difference data analysis module 143 also supports longitudinal comparative analysis of previous inspection data, generates a difference change trend chart, and assists in evaluating the effectiveness of construction quality improvement.

[0186] This difference data analysis module 143, when implementing longitudinal comparative analysis and trend chart generation for previous inspection data, specifically constructs a time series difference archive based on component ID, automatically aligns historical records by combining inspection batches with component spatial locations, and introduces a change attribution algorithm to identify the primary causes of increase or decrease in differences (such as construction adjustments or design changes). This allows trend charts to be aggregated and displayed by dimensions such as component, region, and difference type, supporting the calculation of difference change rates and the setting of warning thresholds. This effectively enhances the timeliness and data support capabilities of construction quality tracking and improvement effect evaluation.

[0187] The review report production module 142 and the difference data analysis module 143 in this report generation and statistics unit 140 respectively support exporting reports and statistical data into common file formats such as PDF, Word, Excel, etc., and support one-click uploading to the enterprise-level management platform or project archive system to realize the electronic management of the entire review report process.

[0188] This solution automatically generates multi-format documents through a template engine and data mapping module when exporting review reports and statistical data into formats such as PDF, Word, and Excel, and uploading them to the management platform with one click. It also maintains the consistency of graphics, text, tables, numbering, and cross-references during export. Furthermore, it automatically matches archiving paths through component-level metadata and project IDs, enabling seamless synchronization of batch review data to the BIM management platform or engineering archive system. This effectively overcomes the bottlenecks of traditional manual export and archiving processes, significantly improving report delivery efficiency and the level of electronic management throughout the entire lifecycle.

[0189] The report generation and statistics unit 140 thus formed can greatly improve the efficiency and quality of preparing the difference review report, and provide strong data support for project decision-making, problem rectification tracking and quality archiving.

[0190] On the basis of the above scheme, a system management and operation interface unit 150 can be further introduced into the real model consistency review system 100 according to needs. The system management and operation interface unit 150 is configured to interact with the data integration and preprocessing unit 110, the real model identification and comparison analysis unit 120, the difference annotation and rendering unit 130, and the report generation and statistical analysis unit 140 to complete the data flow, authority control, task scheduling and user interaction of the real model consistency review system for the consistency review process.

[0191] Specifically, the system management and operation interface unit 150 can implement functions such as user account management, role authority allocation, project data management, version control and log tracking by building a unified background management platform.

[0192] As an example, the role permission allocation allows different users to access or operate corresponding modules according to their permission levels. For example, ordinary users can view difference results, advanced users can review and approve differences, and administrators can manage system settings and user allocation.

[0193] The project data management supports parallel management of multiple projects, ensuring the independence of each project's data, models, and test records, and ensuring data security and privacy compliance.

[0194] The version management mechanism can record each data update, comparison execution and difference change, support historical version backtracking and difference comparison, and facilitate dynamic monitoring and problem tracking during the construction phase.

[0195] When implementing the parallel management and version control functions of multiple projects, the system management and operation interface unit 150 adopts a multi-layer data partitioning architecture of "project-model-detection batch" to ensure complete isolation of resources between projects and comply with engineering privacy and regulatory requirements; in terms of version management, a component-level difference tracking mechanism is introduced to automatically record the change details of each comparison or model update, and can visually present the difference content and related images between versions to achieve traceable quality control of the entire construction process.

[0196] Furthermore, when constructing the operation interface, the system management and operation interface unit 150 may adopt a responsive Web architecture that is compatible with both PC and mobile access.

[0197] Furthermore, the interface layout follows the principles of modularity, process-based operation, and simplicity and intuitiveness, and is primarily divided into a navigation area, a display area, and an operation area. The navigation area provides quick switching between function menus, while the display area presents real-time comparison views of the actual model and BIM, along with difference annotations and statistical charts. The operation area integrates common functions such as filtering, searching, annotation, and exporting.

[0198] In constructing its interface layout, the system management and operation interface unit 150 incorporates the specialized needs of BIM discrepancy review. The display area integrates synchronized views of the 3D real model and BIM model, along with a linked refresh mechanism for discrepancy graphic annotations and statistical charts. This allows users to compare and analyze structured data and image information within the same view. The filtering, annotation, and export functions within the operation area are highly integrated with the recognition results, supporting real-time dynamic filtering based on component attributes, discrepancy types, and other factors, while also enabling direct report generation or triggering of review processes. This resulting interface layout transcends the traditional single-data or image interface model, improving both human-computer interaction efficiency and the operability of engineering data.

[0199] Furthermore, the system management and operation interface unit 150 also introduces optimization function modules such as local refresh, delayed loading, intelligent search, batch operation, etc., and supports customized interface themes, layout adjustment and multi-language switching.

[0200] This system management and operation interface unit 150 combines component-level indexing mechanisms and regional division strategies to implement local refresh, delayed loading, intelligent search, and batch operations, enabling on-demand loading and dynamic refresh based on the user's current operational context, effectively reducing page load and response latency. The intelligent search module implements semantic-level retrieval based on difference tags and component attributes, while batch operations support unified processing and one-click export of multiple differences. Combined with multilingual and theme customization capabilities, this solution significantly improves usability, scalability, and cross-regional adaptability in complex engineering environments.

[0201] Furthermore, the system management and operation interface unit 150 also has a built-in intelligent reminder function, which automatically pushes notifications to the user interface or bound mailbox when new differences, task expiration or data anomalies are detected.

[0202] When implementing intelligent reminder functions, the system management and operation interface unit 150, based on a task scheduler and difference recognition logic, monitors key events such as data updates, new differences, and approval delays in real time. Once a set condition is triggered, it constructs structured reminder content (including component number, time, severity level, etc.) and accurately pushes it to the user interface or email. Compared to conventional reminder mechanisms, this intelligent reminder mechanism intelligently categorizes notifications based on user role and task type, supports multi-terminal synchronization and reminder priority management, and improves response efficiency and management closed-loop management in complex project collaboration.

[0203] Furthermore, the operation interface constructed by the system management and operation interface unit 150 supports interface connection with third-party systems (such as project management platform, construction log platform, etc.) to achieve automatic data synchronization and linkage processing.

[0204] When realizing the interface connection function between the operation interface and the third-party system, the system management and operation interface unit 150 realizes the automatic alignment of semantic mapping and data fields of BIM models, detection data, difference logs, etc. with external platforms by building a unified data middle platform and interface adaptation layer; at the same time, it supports event-triggered linkage processing mechanism, such as automatically synchronizing to the construction log or triggering a task work order when a difference occurs in a component.

[0205] For example, the designed third-party system interface docking solution greatly reduces the system docking cost, breaks through information silos, realizes the data closed loop throughout the project life cycle, and significantly improves the collaborative efficiency and data consistency management level.

[0206] Furthermore, the system management and operation interface unit 150 is deployed in a distributed architecture, supports elastic expansion and load balancing, and ensures a smooth and efficient response experience in a multi-user, high-concurrency environment.

[0207] The system management and operation interface unit 150 implements a distributed architecture deployment, supporting elastic expansion and load balancing. It dynamically schedules computing resources based on BIM model size, the number of difference detection tasks, and user access behavior, enabling concurrent processing and cache optimization by component region. Furthermore, it incorporates asynchronous message queues and data slicing mechanisms to ensure smooth response and stable operation of high-load modules such as difference identification, annotation rendering, and report generation. This distributed architecture deployment significantly improves the system's stability, scalability, and service continuity in complex building project scenarios.

[0208] The system management and operation interface unit 150 can ensure the efficiency, reliability and scalability of the overall operation of the entire system, while providing users with a convenient and intuitive operating environment, greatly improving the execution efficiency and management level of the real model consistency review process.

[0209] The real model consistency review system based on the large language model formed by the above solution integrates multi-source heterogeneous data, completes the standardized preprocessing of real-world point clouds or images, and combines deep learning recognition and multi-scale feature analysis to achieve accurate correspondence and intelligent comparison between real models and BIM components.

[0210] The system first uses a large language model to perform semantic reasoning and logical verification, automatically detecting differences such as size deviation, positioning deviation, missing and redundant components, and generating structured difference data;

[0211] On this basis, we further present difference information intuitively through graphical annotation and natural language description, supporting batch screening and multi-scale interaction;

[0212] Based on the results of the difference analysis, a standardized review report is automatically compiled, providing statistical analysis, change trend tracking and multi-format export functions.

[0213] The system also features a modular design, supporting multi-project management, version control, and permission allocation. It also features a responsive user interface and intelligent reminder mechanisms, ensuring efficient and reliable consistency reviews in large-scale data environments. This invention effectively improves quality management during construction and operation, providing key technical support for the realization of digital twins and intelligent construction.

[0214] The large language model-based real-model consistency review system solution provided by the present invention can be implemented as a corresponding software program in a specific application, forming a corresponding large language model-based real-model consistency review software system. When running, the software program will execute the above-mentioned large language model-based real-model consistency review function method and store it in a corresponding storage medium for processor access and execution.

[0215] The following describes the implementation process of the real model consistency review system solution based on the large language model for real model consistency review.

[0216] First, deploy a real-model consistency review system based on a large language model.

[0217] In practical applications, the real-model consistency review system based on the large language model can be deployed in the cloud, locally, or in a hybrid deployment mode. Before deploying the system, it is necessary to select appropriate hardware resources and network environment according to the scale of the project, and configure the server cluster, database, GPU acceleration unit, and container management platform (such as Kubernetes). The deployment process includes key steps such as model loading and microservice splitting, data middle platform configuration, BIM and point cloud interface adaptation, and permission system initialization. If cloud deployment is selected, access rights and encryption channels must be configured to ensure data security. The system supports parallel deployment and elastic expansion of multiple projects, and can dynamically adjust resources according to project progress to ensure an efficient and stable real-model review operating environment in complex construction scenarios after deployment.

[0218] After completing the deployment of the real-model consistency review system based on the large language model, the real-model consistency review system based on the large language model will be put into operation.

[0219] On this basis, the BIM model files corresponding to the on-site buildings that require real-model consistency review are imported into the system. The system then performs a structured analysis of the BIM model, breaks down the overall BIM model into several component units, and extracts attribute information to form a BIM database.

[0220] Next, image data and / or 3D point cloud data of the on-site building are acquired and input into the system, which pre-processes the acquired image data and / or 3D point cloud data to form real model data. During this pre-processing step, the image data is normalized and initially geo-referenced based on the spatial coordinate system established on-site. The 3D point cloud data is then subjected to coordinate system conversion, point density homogenization, and smoothing filtering.

[0221] Next, the system uses deep learning visual recognition technology to perform component-level recognition and semantic segmentation on the pre-processed real-model data, generating a building recognition result that includes component type (such as beams, columns, walls, and floor slabs), dimensional parameters, spatial position, and orientation attributes. Subsequently, a feature extraction algorithm is used to extract key attributes from the identified real-model components, and a preliminary match is performed with the corresponding component entries in the BIM database.

[0222] Next, a large language model is used to perform semantic reasoning and logical verification on the preliminary matching results, and combined with constraints such as construction specifications and architectural logical relationships (such as beams must be connected to columns and floor slabs must be attached to beams or walls) to further screen out incorrect matches.

[0223] Next, for each successfully matched pair of real-model components and BIM components, metrics such as dimensional differences, positioning deviations, and morphological consistency are calculated. Based on pre-set detection rules, anomalies such as excessive dimensional deviations, excessive positional offsets, or severe morphological deformations are determined. BIM components that fail to match the real-model components are deemed "missing," while those that do not are deemed "redundant."

[0224] Next, the differences detected by the identification and comparison analysis module are presented in a systematic and visual manner. This step first categorizes the detected difference data by component type, difference type, and severity, creating a list of differences. Subsequently, graphical annotations are automatically generated based on the spatial locations of the differences in the actual model image or 3D point cloud model.

[0225] Finally, the analysis results output by the difference detection and annotation module are compiled into a complete and standardized report file in a structured and standardized manner, and a multi-dimensional statistical analysis of the overall difference situation is also performed. In this step, all difference data is automatically extracted and archived according to key dimensions such as project name, inspection time, inspection area, component type, difference category, and severity to ensure data integrity and traceability. Subsequently, the system automatically compiles and generates a unified review report based on the preset report template.

[0226] From the above, it can be seen that the real model consistency review solution provided by the present invention integrates the advantages of large language model reasoning and visual recognition accuracy, and can efficiently and accurately realize the difference detection, labeling and report generation between the real model and the BIM model, thereby promoting the intelligent transformation and digital twin construction of the construction industry.

[0227] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0228] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0229] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0230] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A real model consistency review system based on a large language model, characterized by: The real model consistency review system includes: A data integration and preprocessing unit, which is used to obtain image data and / or three-dimensional point cloud data of the on-site building, as well as the BIM model file corresponding to the on-site building, preprocess the obtained image data and / or three-dimensional point cloud data to form real model data, and perform structured analysis on the BIM model to disassemble the entire BIM model into several component units and extract attribute information; A real model identification and comparison analysis unit, which uses deep learning visual recognition technology to perform component-level identification and semantic segmentation on the preprocessed real model data. For the identified real model components, the unit extracts their key attributes, performs a preliminary match based on the extracted key attributes with the corresponding component unit entries in the BIM database, and uses a large language model to perform semantic reasoning and logical verification on the preliminary matching results based on constraint conditions. It determines whether there are any differences between the successfully matched real model components and the corresponding component units in the BIM database according to the set detection rules; a difference marking and rendering unit, configured to present the difference results detected by the real model recognition and comparison analysis unit in a visual manner; The report generation and statistical analysis unit is used to generate a complete and standardized report file in a structured and standardized manner based on the analysis results output by the difference annotation and rendering unit, and to perform multi-dimensional statistical analysis on the overall difference situation.

2. The large language model-based real model consistency review system according to claim 1, characterized in that: The data integration and preprocessing unit includes a data acquisition module, a data preprocessing module, and a BIM database construction module. The data acquisition module is capable of acquiring image data and / or three-dimensional point cloud data of on-site buildings, and importing BIM model files corresponding to on-site buildings; The data preprocessing module is configured to interact with the data acquisition module, perform standardization processing on the image data, and perform preliminary geo-registration according to the spatial coordinate system set on site; Perform coordinate system conversion, point density homogenization and smoothing filtering on 3D point cloud data; The BIM database construction module is configured to interact with the data acquisition module, perform structured analysis on the imported BIM model file, decompose the overall BIM model into several component units, extract attribute information, and store it as standardized database entries to form a BIM database.

3. The large language model-based real model consistency review system according to claim 1, characterized in that: The data integration and preprocessing unit further includes an abnormal data detection module, which is configured to interact with the data acquisition module and automatically identify abnormal data in the acquired data.

4. The large language model-based real model consistency review system according to claim 1, characterized in that: The real model identification and comparison analysis unit includes a real model component identification module, a preliminary matching module, a matching verification module, and an anomaly detection module. The real model component recognition module performs component-level recognition and semantic segmentation on the pre-processed real model data based on deep learning visual recognition technology. The recognition results include component type, size parameters, spatial position and direction attributes; The preliminary matching module is configured to interact with the real model component identification module data, extract key attributes of the identified real model component through a feature extraction algorithm, and perform preliminary matching of the extracted key attributes with corresponding component unit entries in the BIM database; The matching verification module is configured to interact with the preliminary matching module data, and use the large language model to perform semantic reasoning and logical verification on the preliminary matching results based on the constraints to determine whether the actual model components are correct with the corresponding component units in the preliminary matching BIM database; The anomaly detection module is configured to interact with the matching verification module data, calculate the size difference, positioning deviation, and morphological consistency index between each pair of successfully matched real model components and the corresponding component units in the BIM database, and determine whether there is a difference based on the set detection rules.

5. The large language model-based model consistency review system according to claim 1, characterized in that: The difference annotation and rendering unit includes a difference item sorting module and an annotation production module. The difference item sorting module is used to classify and sort the detected difference data according to component category, difference type and severity, and construct a difference item list; The annotation production module interacts with the difference item sorting module in data, and automatically generates graphical annotations and / or natural language descriptions based on the spatial positions corresponding to the differences in the real model image or three-dimensional point cloud model formed by the data integration and preprocessing unit.

6. The large language model-based model consistency review system according to claim 1, characterized in that: The report generation and statistics unit includes a difference data extraction module, an audit report production module, and a difference data analysis module. The difference data extraction module is used to extract all difference data from the difference marking and rendering unit and archive and organize them according to preset dimensions; The review report production module is configured to interact with the difference data extraction module to automatically compile and generate a review report based on a preset report template for the difference data collated by the difference data extraction module; The difference data analysis module is configured to interact with the difference data extraction module to perform multi-dimensional mining on the difference data collated by the difference data extraction module and to visually present the analysis results in the form of charts.

7. The large language model-based real model consistency review system according to claim 1, characterized in that: The real model consistency review system also includes a system management and operation interface unit, which is configured to interact with the data integration and preprocessing unit, the real model identification and comparison analysis unit, the difference annotation and rendering unit, and the report generation and statistical analysis unit to complete the data flow, authority control, task scheduling and user interaction of the real model consistency review system for the consistency review process.

8. A method for checking consistency of a real model based on a large language model, characterized in that: The real model consistency review method includes: The on-site building corresponds to the BIM model file, and the BIM model is structurally analyzed, the overall BIM model is disassembled into several component units, and attribute information is extracted to form a BIM database; Acquire image data and / or three-dimensional point cloud data of the on-site building, and pre-process the acquired image data and / or three-dimensional point cloud data to form real model data; Based on deep learning visual recognition technology, the pre-processed real model data is identified and semantically segmented at the component level. For the identified real model components, their key attributes are extracted and preliminarily matched with the corresponding component unit entries in the BIM database based on the extracted key attributes. Use a large language model to perform semantic reasoning and logical verification on the initial matching results based on constraints; Determine whether there are any differences between the successfully matched real model components and the corresponding component units in the BIM database based on the set detection rules; Output and display the detected differences.

9. The method for checking consistency of a real model based on a large language model according to claim 8, characterized in that: The real model consistency review method performs standardization processing on the acquired image data and / or three-dimensional point cloud data when pre-processing the acquired image data and / or three-dimensional point cloud data, and performs preliminary geo-registration according to the spatial coordinate system set on site; Perform coordinate system conversion, point density homogenization and smoothing filtering on three-dimensional point cloud data.

10. The method for checking consistency of a real model based on a large language model according to claim 8, characterized in that: The real model consistency review method calculates the size difference, positioning deviation, and morphological consistency index between each pair of successfully matched real model components and the corresponding component units in the BIM database, and determines whether there is a difference based on the set detection rules.

Images