BIM model automatic examination method and system based on visual large model image analysis and standard knowledge base coupling

Abstract

The invention discloses a BIM model automatic examination method and system based on visual large model image analysis and standard knowledge base coupling. The scheme comprises the following steps of training data collection and labeling, professional image review model training, review roaming path definition, profile image acquisition and processing, image compliance analysis, standard knowledge base matching, optimization suggestion generation and report output. By training an image review model in a professional field and combining with a dynamic specification knowledge base, detail problems which are difficult to find by a traditional method in a BIM model can be automatically recognized, optimization suggestions are provided by matching with real-time updated building specifications, a full-closed-loop review process from image analysis to specification verification is realized, review efficiency and accuracy are remarkably improved, and the review efficiency and accuracy are improved. And labor dependence is reduced.

Description

Technical Field

[0001] This invention relates to information technology in building engineering, specifically to automated review technology for BIM models. Background Technology

[0002] BIM, or Building Information Modeling, is a core technology for the digital transformation of the construction industry. It is based on various relevant information and data of a construction project and uses digital information to simulate the real information of the building.

[0003] To ensure the smooth progress of construction projects and reduce risks, the review of BIM models for construction projects is an important step. Through BIM model review, potential problems can be identified, designs can be optimized, construction efficiency and quality can be improved, and support can be provided for the long-term operation of construction projects.

[0004] Currently, BIM model review mainly relies on the following two types of technologies:

[0005] (1) Rule-based structured review: Using structured data such as IFC, the rule engine can detect problems such as component size and attribute information, but it cannot identify model details errors (such as local collisions, missing textures, mismatched connections, etc.) through visual information, and it is easy to miss minor problems such as misalignment of pipeline ends.

[0006] (2) Manual intervention review: This model relies on manual layer-by-layer inspection, which is time-consuming and labor-intensive and limited by the reviewer's experience, making it difficult to guarantee the comprehensiveness of the inspection.

[0007] As can be seen from the above, the current BIM model review scheme mainly has the following problems:

[0008] (1) Missed details: The limited coverage of structured data makes it impossible to directly perceive the visualization details of the model;

[0009] (2) Lagging standard updates: Using a static rule base, it is difficult to match the latest revisions of building codes and industry standards in a timely manner. Summary of the Invention

[0010] This invention provides an automated BIM model review scheme based on the coupling of visual large model image analysis and standard knowledge base, aiming to solve the problems existing in the current BIM model review scheme.

[0011] To achieve the above objectives, this invention provides an automated BIM model review method based on the coupling of visual large model image analysis and a standardized knowledge base. The automated BIM model review method includes the following steps:

[0012] (1) Training data collection and labeling;

[0013] Two-dimensional cross-sectional images of BIM models and corresponding review comments were collected from historical review projects. By annotating the issues in the images, training data of "image-issue type" was formed.

[0014] (2) Training of professional image review models;

[0015] Based on the constructed training data, an improved semantic segmentation network is trained to recognize BIM profile images and generate a professional image review model.

[0016] (3) Review the roaming path definition;

[0017] The review roaming path is defined based on the key review points. The defined review roaming path is used to guide the subsequent automatic profile image acquisition and processing.

[0018] (4) Profile image acquisition and processing;

[0019] Based on the defined roaming path, standardized profile snapshots of the BIM model are dynamically captured along the roaming path, and the collected profile snapshot images of the BIM model are subjected to illumination normalization and noise filtering.

[0020] (5) Image compliance analysis;

[0021] Based on a trained professional image review model, the standardized cross-sectional images of the collected and preprocessed BIM models are analyzed in real time, and problem detection diagrams, compliance verification results and problem type labels are output.

[0022] (6) Standardize knowledge base matching;

[0023] First, keywords were extracted from the image compliance analysis results to identify problematic keywords;

[0024] Next, based on the extracted problem keywords, relevant normative suggestions with the highest relevance to the problem keywords are retrieved from the normative knowledge base;

[0025] (7) Generate optimization suggestions and reports;

[0026] Based on the matching results of the normative knowledge base given in step S(6), specific optimization suggestions are generated, and combined with the problem types and normative clauses identified in the compliance analysis of step S(5), a detailed review report is output through the text generation module.

[0027] Furthermore, when collecting and labeling training data, the automated review method first needs to collect historical BIM review reports. After collection, it extracts BIM problem screenshots and BIM problem types from the reports to form "image-problem type" training data.

[0028] Furthermore, when training a professional image review model, the automated review method embeds a standard knowledge retrieval interface at the end of the model, so that when the model outputs a question, it synchronously associates the relevant standard entries.

[0029] Furthermore, the automated review method specifies key points in the BIM model visualization interface and then generates a smooth review roaming path using B-spline curve interpolation.

[0030] Furthermore, the automated review method dynamically associates question types with normative clauses based on a knowledge graph-based normative knowledge base, and generates optimization suggestions through a text generation module accordingly.

[0031] To achieve the above objectives, this invention provides an automated BIM model review system based on the coupling of visual large model image analysis and standard knowledge base. The review system is composed of a training data collection and annotation module, a professional image review model training module, a review roaming path definition module, a profile image acquisition and processing module, an image compliance analysis module, a standard knowledge base matching module, and an optimization suggestion and report generation module.

[0032] The training data collection and annotation module is configured to collect two-dimensional cross-sectional images of BIM models and corresponding review opinions from historical review projects. By annotating the problems in the images, training data of "image-problem type" is formed.

[0033] The professional image review model training module is configured to interact with the training data collection and annotation module, and can train the improved semantic segmentation network to generate a professional image review model based on the constructed training data and BIM profile image recognition.

[0034] The review roaming path definition module is configured to define the review roaming path based on review key points;

[0035] The surface image acquisition and processing module is configured to interact with the review roaming path definition module, and can dynamically capture standardized cross-sectional snapshots of the BIM model along the defined roaming path, and perform illumination normalization and noise filtering on the acquired cross-sectional snapshot images of the BIM model.

[0036] The image compliance analysis module is configured to interact with the surface image acquisition and processing module, and can perform real-time analysis on the standardized cross-sectional images of the acquired and preprocessed BIM model based on a trained professional image review model, and output problem detection diagrams, compliance verification results and problem type labels.

[0037] The standard knowledge base matching module is configured to first extract keywords from the image compliance analysis results to identify problem keywords; then, based on the extracted problem keywords, retrieve the relevant standard recommendations with the highest relevance to the problem keywords from the standard knowledge base.

[0038] The optimization suggestion and report generation module is configured to generate specific optimization suggestions based on the standard knowledge base matching results generated by the standard knowledge base matching module, and, in conjunction with the problem types and standard clauses generated by the image compliance analysis module, output a detailed review report through the text generation module.

[0039] Furthermore, during the training of the professional image review model, the training module embeds a standard knowledge retrieval interface at the end of the model, so that when the model outputs a question, it synchronously associates with relevant standard entries.

[0040] Furthermore, the review roaming path definition module generates a smooth review roaming path by specifying key points in the BIM model visualization interface and then using B-spline curve interpolation.

[0041] Furthermore, the optimization suggestion and report generation module dynamically associates problem types and normative clauses based on the normative knowledge base of the knowledge graph, and generates optimization suggestion information through the text generation module accordingly.

[0042] The BIM model automated review scheme provided by this invention trains a professional image review model and combines it with a dynamic specification knowledge base. It can automatically identify details in BIM models that are difficult to find by traditional methods (such as small-scale collisions and accuracy deviations), and provide optimization suggestions by matching them with real-time updated building specifications. This realizes a closed-loop review process from "image analysis" to "specification verification", which significantly improves review efficiency and accuracy and reduces reliance on manual labor. Attached Figure Description

[0043] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0044] Figure 1 This is a flowchart of the automated BIM model review method in this invention;

[0045] Figure 2 This is a schematic diagram illustrating the structure of the automated BIM model review system of this invention.

[0046] Figure 3 This is a flowchart illustrating the automated review process for BIM models in an example of the present invention. Detailed Implementation

[0047] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0048] This invention addresses the issue of BIM model review by coupling image analysis with a regulatory knowledge base to automate the review of BIM models. It can automatically identify detailed problems in BIM models that are difficult to find using traditional methods, and match the identified problems with real-time updated building codes to provide optimization suggestions in a synchronous manner. This achieves a closed-loop review process from "image analysis" to "regulatory verification", significantly improving review efficiency and accuracy and reducing reliance on manual labor.

[0049] Accordingly, this invention provides an automated BIM model review method based on the coupling of visual large model image analysis and standard knowledge base, combined with... Figure 1 As shown, this automated BIM model review method consists of the following steps:

[0050] (1) Training data collection and labeling;

[0051] Two-dimensional cross-sectional images of BIM models and corresponding review comments are collected from historical review projects. By annotating the issues in the images, training data of "image-issue type" is formed.

[0052] The training data here includes image data and labeled question types.

[0053] The image data mainly consists of two-dimensional cross-sectional images extracted from historical BIM projects (such as building, bridge, and pipeline designs), including floor plans and bridge cross-sections.

[0054] The types of annotation issues include problematic areas in manually or semi-automatically annotated images. Specific types include: misaligned pipeline ends (such as insufficient spacing between ventilation ducts and cable trays), missing components (such as bridge decks lacking guardrails), accuracy deviations (such as component dimensional errors exceeding the allowable range), missing textures (such as concrete surfaces not being labeled with materials according to specifications), insufficient clearance (such as equipment installation space not meeting specification requirements), and collisions between different disciplines (such as collisions between various pipelines), etc.

[0055] (2) Training of professional image review models;

[0056] Based on the constructed training data, an improved semantic segmentation network is trained to recognize BIM profile images and generate a professional image review model.

[0057] Once trained, the professional image review model can automatically detect problem areas based on BIM model screenshots. In other words, after inputting a BIM model screenshot, the professional image review model can directly mark the problem areas with rectangles based on the screenshot without manual intervention.

[0058] Once trained, the professional image review model can automatically identify problem types, such as pipeline collisions, missing components, and insufficient clearance, after the problem area is labeled.

[0059] Once trained, the professional image review model can automatically associate with regulatory clauses and provide corresponding suggestions. Based on the problem types identified in the previous analysis, the model can automatically associate with the corresponding regulatory clauses and, in conjunction with the regulatory knowledge base, return appropriate adjustment suggestions.

[0060] (3) Review the roaming path definition;

[0061] Here, a corresponding review roaming path is defined based on the key review points. This defined review roaming path is used to facilitate the subsequent automatic acquisition and processing of cross-sectional images. After the review roaming path is defined, cameras can be used to capture model cross-sectional images at equal intervals (e.g., 10m per step) along the roaming path.

[0062] (4) Profile image acquisition and processing;

[0063] Based on the defined roaming path, standardized section snapshots of the BIM model are dynamically captured along the roaming path, and the collected section snapshot images of the BIM model are preprocessed to eliminate false detections caused by differences in model rendering.

[0064] (5) Image compliance analysis;

[0065] The trained professional image review model performs real-time analysis on the standardized cross-sectional images of the collected and preprocessed BIM models, and outputs problem detection maps, compliance verification results and problem type labels.

[0066] The trained professional image review model can identify visual detail issues when analyzing standardized profile images in real time. These issues include small-scale collisions, accuracy deviations, missing textures / materials, and insufficient clearance. The model outputs issue type labels (collision / missing / insufficient clearance, etc.) and confidence scores (threshold ≥ 0.7 triggers judgment). High-confidence issues (≥ 0.9) are judged directly, while medium-to-low-confidence issues (0.7-0.9) require secondary manual verification.

[0067] Accordingly, the entire process of image compliance analysis can be automated, from image acquisition to problem identification, and the review efficiency can be increased to 5 times that of manual work (which traditionally takes several hours to several days). Furthermore, it can improve the quality of review. Through visual large model review, the entire scope of the project can be automatically reviewed, replacing the BIM model sampling method. Problems that are difficult to find with traditional review methods can be improved, thereby improving the quality of review.

[0068] (6) Standardize knowledge base matching;

[0069] This step first involves extracting keywords from the image compliance analysis results to identify problematic keywords;

[0070] Next, based on the extracted problem keywords, the relevant standard recommendations with the highest relevance to the problem keywords are retrieved from the standard knowledge base. These retrieved standard recommendations serve as the basis for generating subsequent optimization suggestions.

[0071] Furthermore, the RAG (Retrieval-Augmented Generation) technique can be used here to retrieve the relevant normative recommendations that are most relevant to the problem keywords from the normative knowledge base.

[0072] Commonly used retrieval methods include text-based retrieval (such as TF-IDF, BM25) and embedding-based retrieval (such as FAISS, Annoy).

[0073] (7) Generate optimization suggestions and reports;

[0074] Based on the matching results of the normative knowledge base given in step S(6), this step generates specific optimization suggestions and, combined with the problem types and normative clauses identified in the compliance analysis of step S(5), outputs a detailed review report through the text generation module.

[0075] Furthermore, the optimization suggestions here include specific recommendations and referenced standard clauses, such as "Adjust pipeline elevation to ensure spacing ≥ 300mm (based on Appendix F of the 'Code for Acceptance of Construction Quality of Building Electrical Engineering')".

[0076] The review report includes raw issue data (such as issue detection charts and compliance analysis results), optimization suggestions, and issue grading heatmaps (such as generating red / yellow / orange graded heatmaps based on issue severity).

[0077] This invention provides specific implementation schemes for the corresponding steps in the automated review method for BIM models.

[0078] In the training data collection and annotation stage, the present invention can use manual annotation or semi-automatic annotation tools to annotate problems in the image, such as collision region selection and model missing annotation.

[0079] As further explanation, this scheme can generate "image-question type" training data through the following steps:

[0080] (1.1) Determine the scope of data collection:

[0081] Extract typical cross-sectional images (such as floor plans and bridge construction drawings) from historical BIM projects (such as architectural and bridge designs) and associate them with corresponding review reports. These review reports include manually annotated locations of collisions and missing elements.

[0082] Based on this data collection scope, the overall performance of the solution can be improved in several ways, mainly in the following aspects:

[0083] 1) Improve the model's generalization ability and applicability: By collecting typical cross-sectional images from BIM projects of different types (civil buildings, infrastructure, etc.) and scales, the diversity and representativeness of the training dataset can be ensured, which helps to improve the model's generalization ability and adaptability.

[0084] 2) Improve the accuracy of problem identification: The associated review report includes manually annotated collision and missing problem locations and standard references. This information can be used as part of the training data to help the model learn how to identify and locate these problems.

[0085] 3) Enhance the model's ability to understand regulations: The regulatory basis in the review report can help the model understand different types of building codes and standards, thereby more accurately judging whether they meet the regulatory requirements in actual reviews.

[0086] (1.2) Problem labeling:

[0087] For the data collected in step (1.1), problems in the images can be labeled using manual mode or semi-automatic annotation tools.

[0088] Specifically, during manual annotation: reviewers use rectangles to mark problem areas, such as pipe collision locations, on the received BIM project section images. For the marked problem areas, they add problem type labels, such as "structural collision" or "insufficient clearance".

[0089] When performing semi-automatic annotation: the model rendering engine automatically captures the disputed areas and generates initial annotations, which are then manually verified for correctness. For example, a screenshot of a component marked "conflict" can be generated after Navisworks collision checking.

[0090] (1.3) Data cleaning:

[0091] For the image data annotated in step (1.2), remove duplicate images or images with a resolution lower than 1080p, and filter out sample image data without valid review comments, such as images containing only format annotations.

[0092] (1.4) Generate training data:

[0093] After completing the data collection and processing, the BIM problem screenshots and BIM problem types in the report are extracted to form "image-problem type" training data.

[0094] In the "image-problem type" training data generation stage, this invention tightly integrates images and problem types to form a complete data unit. This structured training data can more effectively guide the model to learn and understand problem types. Furthermore, during the data annotation process, problem types are classified and labeled in detail, such as structural collisions and insufficient clearance, which enables the model to more accurately identify and classify different types of problems.

[0095] In the training phase of the professional image review model, the present invention introduces U-Net into the semantic segmentation network to improve the semantic segmentation network, thereby providing the subsequent trained model with the ability to recognize BIM profile images.

[0096] To further enhance U-Net's ability to extract fine-grained features from BIM profile images, the present invention further improves the introduced U-Net.

[0097] The present invention integrates a Feature Pyramid Network (FPN) into U-Net.

[0098] Specifically, the present invention addresses the encoder portion of U-Net by adding an FPN layer after each convolutional block to generate multi-scale feature maps. This allows the U-Net encoder portion to acquire information from these multi-scale feature maps, enhancing its ability to detect small feature targets on BIM profile images.

[0099] The present invention introduces an attention mechanism into U-Net.

[0100] Specifically, the present invention introduces a self-attention mechanism in the decoder part of U-Net, enabling the network to accurately identify and distinguish different components in complex BIM profile images, allowing the network to focus on important areas in the image.

[0101] The present invention introduces residual connections in U-Net.

[0102] Specifically, the present invention adds residual connections between the encoder and decoder of U-Net. These residual connections help the network learn deep features more effectively, thereby accelerating the training process and improving the performance of the model.

[0103] The present invention introduces a fine-grained feature extraction module into U-Net.

[0104] Specifically, the present invention sets a smaller convolutional kernel (e.g., 3x3) in the corresponding stage of the U-Net decoder to capture more detailed features, thereby enhancing the detection capability of small-scale features.

[0105] Based on this, the preferred method of the present invention is to train a professional image review model through the following training strategy.

[0106] (1) Perform data augmentation; use data augmentation techniques (such as rotation, scaling, flipping, etc.) to expand the training dataset to cope with the diversity of BIM profile images.

[0107] (2) Use loss functions; use multi-class cross-entropy loss function to optimize the model to ensure that each class can achieve good segmentation results, and combine Dice loss or Jaccard loss to further improve segmentation accuracy.

[0108] (3) Perform transfer learning; use pre-trained models (models pre-trained on ImageNet) as initial weights to accelerate the training process and improve model performance.

[0109] As a further explanation, the present invention further improves U-Net by integrating a Feature Pyramid Network (FPN), introducing an attention mechanism, adding residual connections, and designing a fine-grained feature extraction module, thereby enhancing the performance of U-Net in processing BIM profile images.

[0110] Here, an integrated Feature Pyramid Network (FPN) is used to generate multi-scale feature maps in the encoder part of the U-Net, thereby helping the model to better capture features at different scales, especially when dealing with BIM profile images with complex structures and multi-scale features.

[0111] Furthermore, a self-attention mechanism is introduced into the decoder part of U-Net, which enables the model to pay more attention to important regions in the image, thereby improving the ability to identify different components in complex BIM profile images.

[0112] Based on the improved semantic segmentation network and training strategy described above, the present invention trains a professional image review model and further optimizes the model:

[0113] (1) Identification of detailed issues;

[0114] An adaptive dilated convolution module is added after the traditional convolutional layer to improve the detection capability for small-scale collisions (such as misalignment of pipeline ends). Here, the adaptive dilated convolution can dynamically adjust the sampling rate of the convolution kernel according to the input features, thereby expanding the receptive field without adding additional parameters and better capturing detailed information in the image.

[0115] (2) Multi-label classification output;

[0116] Output problem type labels (such as collision, missing, insufficient clearance, etc.) and confidence scores. This enables the model not only to identify problems in an image, but also to classify the type of problem and provide a confidence score to quantify the certainty of the model's judgment.

[0117] (3) Dynamic knowledge base association;

[0118] Embedding a specification knowledge retrieval interface at the end of the model, such as Faiss-based vector similarity matching technology, allows the model to synchronously associate relevant specification entries when outputting a problem. This enables reviewers to directly obtain the specification basis corresponding to the detected problem, thus facilitating faster verification and correction.

[0119] The professional image review model trained based on the above scheme can be subsequently adapted for automated review design of BIM models. It can achieve detailed problem identification, multi-label classification, dynamic knowledge base association, image compliance analysis, optimization suggestion generation, and report output. The model employs an improved semantic segmentation network, combined with a Feature Pyramid Network (FPN), attention mechanism, and residual connections, enhancing its ability to detect small-scale features. Furthermore, through data augmentation, transfer learning, and canonical knowledge base matching, the model achieves an efficient automated review process and can generate detailed review reports.

[0120] In practical implementation, the model can identify and locate detailed issues in the BIM model with fine granularity, and simultaneously associate specific specification items when outputting issues, providing immediate specification basis. Furthermore, the model employs an optimized training strategy, combining data augmentation, loss functions, and transfer learning, specifically optimized for the characteristics of BIM cross-sectional images, significantly improving review efficiency and accuracy.

[0121] The present invention can employ either manual or automatic mode during the roaming path definition review stage.

[0122] If the review walkthrough path is defined in manual mode, key points are specified manually in the BIM model visualization interface, and a smooth review walkthrough path is generated for the specified key points using the B-spline curve interpolation method.

[0123] Compared to traditional BIM reviews that rely on manual spot checks and thus fail to cover the entire project scope, this invention's solution, by defining a roaming path for scanning, can cover the entire project area while improving review efficiency and quality.

[0124] To further improve the accuracy of the defined review roaming path, the present invention introduces a dynamic density adjustment mechanism to optimize the defined review roaming path.

[0125] The dynamic density adjustment mechanism here targets dense areas: the path spacing is reduced to 0.5 meters, and pitch angle sampling is increased in the vertical direction (to capture collisions between upper and lower levels).

[0126] Specifically, this solution allows for setting the sampling density of a specific segment within the inspection path. For example, if the overall sampling path spacing is 2m, the sampling density of a particular segment, such as a pipeline comprehensive inspection path, can be increased by reducing the path sampling spacing.

[0127] In the profile image acquisition and processing stage, the present invention dynamically captures standardized profile snapshots of the BIM model based on a defined roaming path. The corresponding implementation process is as follows:

[0128] (1) Define the path;

[0129] First, define the review walk path. Key points can be specified in the BIM model visualization interface, and the system will generate a smooth review walk path by interpolating B-spline curves.

[0130] (2) Path sampling;

[0131] The sampling step size can be set, such as setting a sampling point every 10m; for areas with dense components, the starting and ending nodes of the dense area can be selected to set the sampling step size separately, such as setting a sampling point every 1m, thereby increasing the sampling density.

[0132] (3) Image preprocessing

[0133] The captured cross-sectional images undergo preprocessing, including illumination normalization, noise filtering, and standardized cropping. Illumination normalization uses histogram equalization to eliminate differences in lighting sources between the model and the scene; noise filtering uses a nonlocal mean denoising algorithm to eliminate rendering jagged edges; standardized cropping uses the component center as an anchor point to capture a local image containing the context region, avoiding the loss of boundary information.

[0134] Such cross-sectional image acquisition and processing has the following technical characteristics:

[0135] (1) Dynamic path sampling: It can dynamically extract profile images based on user-defined paths or automatically calculated paths to ensure the comprehensiveness and relevance of the review.

[0136] (2) Standardization processing: Through unified illumination normalization, noise filtering and cropping processing, the quality and consistency of each cross-sectional image are ensured, which facilitates subsequent image analysis and problem detection.

[0137] (3) Efficient preprocessing: The preprocessing techniques used, such as nonlocal mean denoising algorithm, can effectively remove noise while maintaining image details and improve image quality.

[0138] As further explanation, the preferred standard for acquiring BIM model section images is a fixed resolution of 2048×2048 pixels, with a vertical section viewpoint, and the rendering engine is forced to enable shadow and wireframe overlay mode (to enhance detail visibility). This acquisition and processing method has the following technical characteristics for this solution:

[0139] (1) High resolution ensures detail capture: The 2048×2048 pixel resolution ensures that the profile image has sufficient detail, which is crucial for identifying and analyzing small-scale collisions, accuracy deviations, and other issues in the BIM model. High-resolution images can provide clear visual information, helping the model review system to detect potential problems more accurately.

[0140] (2) Standardization of vertical profile perspective: Using a vertical profile perspective ensures that each acquired image has a uniform perspective and scale, which is beneficial for the standardization of the model and subsequent analysis. Standardized perspective can reduce the analytical difficulty caused by changes in perspective and improve the consistency and comparability of the review.

[0141] (3) Enhanced Detail Visibility with Shadow and Wireframe Overlay: Enabling shadow and wireframe overlay mode can highlight the three-dimensional structure and details of the model in the image, making the model's outline and internal structure clearer. Shadow effects can enhance the three-dimensionality and depth of an object, while wireframe mode can clearly show the model's borders and connection points. The combination of the two helps reviewers or automated systems more accurately identify the model's geometric features and potential problems.

[0142] As further explanation, the preprocessing process for the acquired profile images is as follows:

[0143] (4.1) First, perform illumination normalization: use histogram equalization to eliminate differences in light sources between the model and the scene;

[0144] (4.2) Next, noise filtering is performed on the profile image after illumination normalization: rendering jagged edges are eliminated based on the non-local means denoising algorithm.

[0145] (4.3) Finally, the cross-sectional image with noise filtering is standardized and cropped: the local image containing the context region is cropped with the center of the component as the anchor point, so as to avoid the loss of boundary information.

[0146] In the image compliance analysis stage, the present invention uses a trained professional image review model to perform compliance analysis on the acquired cross-sectional images. The corresponding implementation process is as follows:

[0147] (1) Problem areas can be automatically detected based on screenshots of the BIM model;

[0148] After inputting a screenshot of the BIM model, the trained model can directly mark the problem area with a rectangle based on the screenshot without manual intervention.

[0149] (2) Automatic problem type identification

[0150] After labeling the problem areas, the system can automatically identify the problem type, such as pipeline collisions, missing components, insufficient clearance, etc., and output the model's problem confidence level. The model outputs problem type labels and confidence levels (≥0.7). For high-confidence problems (≥0.9), they are directly judged as violations; for medium-to-low-confidence problems (0.7-0.9), BIM model screenshots are used for manual secondary verification, such as checking whether the component coordinates actually overlap.

[0151] (3) Automatically associate with standard clauses and provide corresponding suggestions

[0152] Based on the problem types identified in the previous analysis, the system automatically associates them with the corresponding regulatory clauses and, in conjunction with the regulatory knowledge base, returns appropriate adjustment suggestions.

[0153] This image compliance analysis solution, compared to existing solutions, has the following technical characteristics.

[0154] (1) Fine-grained feature extraction:

[0155] By introducing Feature Pyramid Network (FPN) and attention mechanism, the model is able to more accurately identify and locate detailed problems in BIM models, especially in terms of small-scale collisions and accuracy deviations.

[0156] (2) Dynamic knowledge base association

[0157] When the model outputs a question, it can simultaneously link relevant building code entries, providing immediate code reference. This feature is relatively rare in existing technologies and significantly improves the accuracy and efficiency of the review process.

[0158] (3) Multi-label classification and confidence output:

[0159] The model not only identifies problems, but also classifies them and provides confidence scores, increasing the richness of the model's output and the transparency of the decision.

[0160] Furthermore, based on the above image compliance analysis and processing scheme, the following results are output after real-time analysis of the cross-sectional image:

[0161] 1. Generate a problem detection map: Highlight problem areas on the problem detection map. For example, red can represent structural collisions, and yellow can represent insufficient clearance for equipment.

[0162] 2. Problem detection logic:

[0163] (1) Main problem identification: The model outputs the problem type label and confidence score (≥0.7);

[0164] (2) Cross-validation: For high-confidence questions (≥0.9), the decision is made directly;

[0165] (3) Low to medium confidence level problem (0.7-0.9) calling BIM model screenshot Figure 2 Secondary verification (such as checking whether the component coordinates actually overlap).

[0166] 4. After completing the problem detection, link it to the specification knowledge base.

[0167] The above-mentioned image compliance analysis and judgment is used to determine whether it violates the design specifications. First, the problems are found through problem detection and compliance verification results. Then, the problems found are compared with the specifications to determine whether they violate the specifications. If they do, a compliance judgment is then made.

[0168] In the standard knowledge base matching stage, the present invention first constructs a corresponding standard knowledge base. The design specifications are converted into a vector knowledge base using a text embedding model (such as BertBERT, RoBERTa, etc.), which is then used for knowledge retrieval in large models.

[0169] Furthermore, this specification knowledge base includes a dynamic update module, enabling it to support dynamic updates. Specifically, this dynamic update module incorporates new specification versions in real time through incremental learning.

[0170] Furthermore, this specification's knowledge base includes a text embedding module, which uses pre-trained text embedding models (such as BERT, RoBERTa, etc.) to convert queries and documents into vector representations so that similarity calculations can be performed in the vector space.

[0171] Specifically, the similarity calculation formula is as follows:

[0172]

[0173] Here, A·B is the dot product of vectors A and B, and ||A|| and ||B|| are the moduli of vectors A and B, respectively. The higher the cosine similarity, the more similar the two vectors are.

[0174] This invention utilizes a pre-trained text embedding model in a specification knowledge base to convert specification clauses into vector representations, enabling rapid retrieval of specification clauses relevant to the question, thus achieving efficient retrieval. Furthermore, the pre-trained text embedding model captures semantic information from the text; through semantic understanding, the model identifies the specification clauses most relevant to the question, reducing the possibility of false matches and making the retrieval results more accurate and relevant, enhancing matching accuracy. Moreover, it supports dynamic updates to the specification knowledge base, using vector similarity matching to associate the latest specification clauses in real time, thereby achieving dynamic matching of the specification knowledge base and ensuring the timeliness of review standards.

[0175] By introducing similarity calculation, the model can determine the similarity between problem keywords and regulatory clauses. This allows the model to identify the most relevant regulatory provisions, providing a basis for generating optimization suggestions. Based on the similarity calculation results, the model can generate personalized optimization suggestions for specific problems, ensuring the suggestions' relevance and practicality. When generating review reports, the model can cite specific regulatory clauses, enhancing the report's professionalism and authority.

[0176] As further explained, when using a pre-trained text embedding model for semantic matching in this invention, the first step is to extract problem keywords from the output of the trained BIM image review model (such as "ventilation duct," "cable tray," and "insufficient spacing"). Next, based on the extracted problem keywords, the RAG technology is used to retrieve information from the knowledge base, and the retrieval engine returns the three most relevant specification recommendations. This retrieval engine is responsible for retrieving information relevant to the current task from an external knowledge base. Commonly used retrieval methods include text-based retrieval (such as TF-IDF and BM25) and embedding-based retrieval (such as FAISS and Annoy). The retrieval engine outputs the documents or fragments most relevant to the input query.

[0177] Here, semantic matching is used to precisely associate the problems identified by the model with specific clauses in the normative knowledge base, ensuring the accuracy and standardization of the review results. At the same time, the automated semantic matching process reduces the time spent manually searching and comparing norms, improving review efficiency. By matching specific normative clauses, the review results are more interpretable, easier to understand and implement.

[0178] This invention, through a standard knowledge base matching stage, uses the problem type tags output by the model after image compliance analysis as query keywords. These tags are then compared with clauses in the standard knowledge base using semantic matching technology. Algorithms such as cosine similarity are then used to calculate similarity and select the most relevant standard recommendations. Based on the matched standard clauses, specific optimization suggestions are generated, ensuring the recommendations' relevance and authority. This also allows for the citation of matched standard clauses when generating review reports, enhancing the reports' professionalism and readability.

[0179] As a further explanation, the process of achieving specification knowledge base matching in this invention will be described here in conjunction with the aforementioned stages:

[0180] 1. Embed a standardized knowledge retrieval interface at the end of the model:

[0181] An interface for retrieving normative knowledge is embedded at the end of the professional image review model to synchronously associate relevant building code entries when the model outputs a question.

[0182] 2. Problem output matching with specifications:

[0183] When the model performs compliance analysis on the profile image and outputs issues, it simultaneously outputs issue type labels and confidence levels. Through the normative knowledge retrieval interface, the model uses the issue type labels as query keywords to retrieve relevant normative entries in the normative knowledge base.

[0184] 3. Similarity calculation:

[0185] The query keywords and canonical entries are converted into vector representations using pre-trained text embedding models (such as BERT, RoBERTa, etc.), and the canonical entries most relevant to the query keywords are found by calculating cosine similarity.

[0186] 4. Related specification items:

[0187] Based on the similarity calculation results, select the most similar normative items and associate them with the problem.

[0188] In the optimization suggestion and report generation stage, the present invention dynamically associates the problem types found in the compliance analysis stage of the image with the normative knowledge base based on the matching knowledge graph, and then outputs optimization suggestions through the text generation module.

[0189] As an example, based on the matching of the standard knowledge base, the types of problems found in the image compliance analysis steps are dynamically associated with the standard clauses in the matching standard knowledge base. For example, for the problem of cable trays colliding with other pipelines, it is mapped to Appendix F of the "Code for Acceptance of Construction Quality of Building Electrical Engineering" (GB50303-2015).

[0190] Based on this, and taking into account the problem type and the corresponding normative clauses, suggestions are generated and output through the text generation module.

[0191] The optimization suggestion generation logic process is as follows:

[0192] The system retrieves normative clauses from the normative database and matches them with question types to construct enhanced suggestion templates. The large language model receives the normative clauses retrieved from the normative database and the original question, and generates optimization suggestions.

[0193] For example, if the problem type is a collision between a ventilation duct and a cable tray, the matching standard provision is: Appendix F of the "Code for Acceptance of Construction Quality of Building Electrical Engineering" GB 50303-2015, which specifies the minimum clearance requirements between busbars, cable trays, and duct boxes and pipes. The problem type and the standard provision are then input into the large language model as prompts. The large language model generates optimization suggestions based on these prompts, such as "Adjust the pipeline elevation to ensure a spacing ≥ 300mm (according to Appendix F of the "Code for Acceptance of Construction Quality of Building Electrical Engineering" GB 50303-2015)".

[0194] In addition to generating optimization suggestions, the present invention also generates a review report.

[0195] The generated review report here includes raw issue data (issue detection graph, compliance analysis results), optimization suggestions, and issue grading heatmaps (generated as red / yellow / orange graded heatmaps based on issue severity).

[0196] The original problem data and optimization suggestions are obtained from the problem detection, compliance analysis and optimization suggestion generation modules, while the problem classification heatmap is determined based on the mandatory nature of the violated regulations.

[0197] If the matching regulatory clause includes keywords such as "should not", "must not", or "must", it is considered a mandatory clause, and its severity is marked in red (highest severity).

[0198] If the matching normative clause contains keywords such as "recommendation" or "advisable", the severity level is orange.

[0199] If the matching specification clause contains keywords such as "can" or "not required", the severity is yellow (the lowest severity).

[0200] The BIM model automated review method based on the coupling of visual large model image analysis and standard knowledge base proposed in this invention can be configured into a corresponding software program to form a corresponding BIM model automated review system. When running, this software program will execute the aforementioned smooth curve construction method and store the results in a corresponding storage medium for the processor to retrieve and execute.

[0201] See Figure 2The resulting BIM model automated review system 100, based on the coupling of visual large model image analysis and standard knowledge base, mainly consists of the following modules: training data collection and annotation module 110, professional image review model training module 120, review roaming path definition module 130, profile image acquisition and processing module 140, image compliance analysis module 150, standard knowledge base matching module 160, and optimization suggestion and report generation module 170.

[0202] The training data collection and annotation module 110 in this system is configured to collect two-dimensional cross-sectional images of BIM models and corresponding review opinions from historical review projects. By annotating the issues in the images, training data of "image-issue type" is formed.

[0203] The professional image review model training module 120 in this system is configured to interact with the training data collection and annotation module 110, and can train the improved semantic segmentation network for BIM profile image recognition based on the constructed training data.

[0204] The review roaming path definition module 130 in this system is configured to formulate corresponding review roaming paths based on review key points.

[0205] The surface image acquisition and processing module 140 in this system is configured to interact with the review roaming path definition module 130. Based on the defined roaming path, it can dynamically capture standardized cross-sectional snapshots of the BIM model along the roaming path and preprocess the acquired cross-sectional snapshot images of the BIM model to eliminate false detections caused by differences in model rendering.

[0206] The image compliance analysis module 150 in this system is configured to interact with the surface image acquisition and processing module 140 and the professional image review model training module 120. It can retrieve the professional image review model trained by the professional image review model training module 120 to perform real-time analysis on the standardized cross-sectional images of the acquired and preprocessed BIM model.

[0207] The standard knowledge base matching module 160 in this system is configured to interact with the image compliance analysis module 150. It can extract keywords from the image compliance analysis results generated by the image compliance analysis module 150, extract problem keywords, and retrieve the relevant standard recommendations with the highest relevance to the problem keywords from the standard knowledge base based on the extracted problem keywords.

[0208] The optimization suggestion and report generation module 170 in this system is configured to interact with the image compliance analysis module 150 and the standard knowledge base matching module 160. Based on the standard knowledge base matching results given by the standard knowledge base matching module 160, it can generate specific optimization suggestions and, combined with the problem types and standard clauses analyzed by the image compliance analysis module 150, output a detailed review report through the text generation module.

[0209] This invention provides specific implementation schemes for each functional module in the BIM model automated review system 100.

[0210] Specifically, the composition and implementation scheme of the training data collection and annotation module 110 in this system are described in the aforementioned training data collection and annotation phase scheme, and will not be repeated here.

[0211] The composition and implementation scheme of the professional image review model training module 120 in this system are described in the aforementioned professional image review model training scheme, and will not be repeated here.

[0212] The review roaming path definition module 130 in this system can be configured to work with manual operation mode to generate a smooth review roaming path by specifying key points in the BIM model visualization interface and then using B-spline curve interpolation.

[0213] The composition and implementation scheme of the image compliance analysis module 150 in this system are described in the aforementioned image compliance analysis scheme, and will not be repeated here.

[0214] The composition and implementation scheme of the standard knowledge base matching module 160 in this system are described in the aforementioned standard knowledge base matching scheme, and will not be repeated here.

[0215] The composition and implementation scheme of the optimization suggestion and report generation module 170 in this system are described in the aforementioned optimization suggestion and report generation scheme, and will not be repeated here.

[0216] As a further example, the implementation process of automated BIM model review will be described here based on the BIM model automated review system 100 based on the coupling of visual large model image analysis and standard knowledge base given in this invention.

[0217] In this example, the BIM model automated review system is first built based on the aforementioned system solution.

[0218] Combination Figure 3 As shown, the BIM model automated review system, based on the constructed system, automatically reviews the compliance of BIM models in two stages: the model training stage and the automatic review stage.

[0219] The professional image review model training process is as follows:

[0220] (1) Training data collection;

[0221] Extract typical cross-sectional images (such as floor plans and bridge construction drawings) from historical BIM projects (such as building and bridge designs) and associate them with corresponding review reports. These review reports include manually marked collision and missing problem locations and the relevant standards.

[0222] (2) Data labeling and cleaning;

[0223] For the collected data, issues in the images can be annotated using manual mode or semi-automatic annotation tools; and for the annotated image data, duplicate images or images with a resolution lower than 1080p are removed, and sample image data without effective review comments, such as images containing only format annotations, are filtered out.

[0224] (3) Training and optimization of professional image review models;

[0225] Based on the constructed training data, an improved semantic segmentation network is trained to recognize BIM profile images and generate a professional image review model.

[0226] After completing the training of the professional image review model, the automatic review stage begins, and the relevant process is as follows:

[0227] (4) Define the review roaming path for the BIM model to be reviewed;

[0228] The review walkthrough path can be defined in manual mode. Key points can be specified directly in the BIM model visualization interface, and a smooth review walkthrough path can be generated for the specified key points using the B-spline curve interpolation method.

[0229] (5) Based on the defined review roaming path, collect the cross-sectional images of the BIM model to be reviewed, and preprocess the collected images.

[0230] (6) Using a trained professional image review model, perform real-time image compliance analysis on the collected and preprocessed cross-sectional images of the BIM model to determine whether there are any compliance issues:

[0231] If there are no compliance issues, the review task is completed; if there are compliance issues, proceed to step (7).

[0232] (7) For the compliance issues discovered, analyze the issue type and dynamically associate the corresponding normative clauses in the normative knowledge base according to the issue type.

[0233] (8) Based on the problem type and the mapped normative clauses, generate and output suggestions through the text generation module.

[0234] (9) Based on the generated optimization suggestions, a review report is further generated.

[0235] As can be seen from the above, the present invention, by training an image review model in a professional field and combining it with a dynamic specification knowledge base, can automatically identify details in BIM models that are difficult to detect by traditional methods (such as small-scale collisions and accuracy deviations), and provide optimization suggestions by matching them with real-time updated building specifications. This achieves a closed-loop review process from "image analysis" to "specification verification," significantly improving review efficiency and accuracy and reducing reliance on manual labor.

[0236] Based on the above-mentioned automated BIM model review scheme based on the coupling of visual large model image analysis and standard knowledge base, this embodiment of the invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the above-mentioned automated BIM model review method.

[0237] This invention also provides a processor for running a program, wherein the program executes the steps of the above-described automated BIM model review method during runtime.

[0238] This invention also provides a terminal device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. The program code is loaded and executed by the processor to implement the steps of the above-described automated BIM model review method.

[0239] The present invention also provides a computer program product, which, when executed on a data processing device, is adapted to perform the steps of the above-described automated review method for BIM models.

[0240] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0241] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0242] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0243] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0244] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0245] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0246] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0247] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0248] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0249] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0250] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0251] The method, specific system unit, or part thereof of the present invention described above is a pure software architecture. It can be deployed via program code on physical media, such as hard disks, optical discs, or any electronic device (such as smartphones or computer-readable storage media). When a machine loads and executes the program code (e.g., a smartphone loads and executes it), the machine becomes an apparatus for implementing the present invention. The method and apparatus of the present invention can also be transmitted in program code form via transmission media, such as cables, optical fibers, or any transmission method. When the program code is received, loaded, and executed by a machine (e.g., a smartphone), the machine becomes an apparatus for implementing the present invention.

[0252] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for automated review of BIM models based on the coupling of visual large-scale model image analysis and a standardized knowledge base, characterized in that, The automated BIM model review method includes the following steps: (1) Training data collection and labeling; Two-dimensional cross-sectional images of BIM models and corresponding review comments were collected from historical review projects. By annotating the issues in the images, training data of "image-issue type" was formed. (2) Training of professional image review models; Based on the constructed training data, an improved semantic segmentation network is trained to recognize BIM profile images and generate a professional image review model. (3) Review the roaming path definition; The review roaming path is defined based on the key review points. The defined review roaming path is used to guide the subsequent automatic profile image acquisition and processing. (4) Profile image acquisition and processing; Based on the defined roaming path, standardized profile snapshots of the BIM model are dynamically captured along the roaming path, and the collected profile snapshot images of the BIM model are subjected to illumination normalization and noise filtering. (5) Image compliance analysis; Based on a trained professional image review model, the standardized cross-sectional images of the collected and preprocessed BIM models are analyzed in real time, and problem detection diagrams, compliance verification results and problem type labels are output. (6) Standardize knowledge base matching; First, keywords were extracted from the image compliance analysis results to identify problematic keywords; Next, based on the extracted problem keywords, relevant normative suggestions with the highest relevance to the problem keywords are retrieved from the normative knowledge base; (7) Generate optimization suggestions and reports; Based on the matching results of the normative knowledge base given in step S(6), specific optimization suggestions are generated, and combined with the problem types and normative clauses identified in the compliance analysis of step S(5), a detailed review report is output through the text generation module.

2. The automated review method for BIM models according to claim 1, characterized in that, When collecting and labeling training data, the automated review method first needs to collect historical BIM review reports. After collection, it extracts BIM problem screenshots and BIM problem types from the reports to form "image-problem type" training data.

3. The automated review method for BIM models according to claim 1, characterized in that, The automated review method embeds a standard knowledge retrieval interface at the end of the model when training a professional image review model, so that the model can synchronously associate relevant standard entries when outputting questions.

4. The automated review method for BIM models according to claim 1, characterized in that, The automated review method specifies key points in the BIM model visualization interface and then generates a smooth review roaming path using B-spline curve interpolation.

5. The automated review method for BIM models according to claim 1, characterized in that, The automated review method dynamically associates question types with normative clauses based on a knowledge graph-based normative knowledge base, and generates optimization suggestions through a text generation module accordingly.

6. An automated BIM model review system based on the coupling of visual large-scale model image analysis and a standardized knowledge base, characterized in that, The review system consists of a training data collection and annotation module, a professional image review model training module, a review roaming path definition module, a profile image acquisition and processing module, an image compliance analysis module, a standard knowledge base matching module, and an optimization suggestion and report generation module. The training data collection and annotation module is configured to collect two-dimensional cross-sectional images of BIM models and corresponding review opinions from historical review projects. By annotating the issues in the images, training data of "image-issue type" is formed. The professional image review model training module is configured to interact with the training data collection and annotation module, and can train the improved semantic segmentation network to generate a professional image review model based on the constructed training data and BIM profile image recognition. The review roaming path definition module is configured to define the review roaming path based on review key points; The surface image acquisition and processing module is configured to interact with the review roaming path definition module, and can dynamically capture standardized cross-sectional snapshots of the BIM model along the defined roaming path, and perform illumination normalization and noise filtering on the acquired cross-sectional snapshot images of the BIM model. The image compliance analysis module is configured to interact with the surface image acquisition and processing module, and can perform real-time analysis on the standardized cross-sectional images of the acquired and preprocessed BIM model based on a trained professional image review model, and output problem detection diagrams, compliance verification results and problem type labels. The specification knowledge base matching module is configured to first extract keywords from the image compliance analysis results to identify problem keywords; then, based on the extracted problem keywords, retrieve the relevant specification suggestions with the highest relevance to the problem keywords from the specification knowledge base. The optimization suggestion and report generation module is configured to generate specific optimization suggestions based on the standard knowledge base matching results generated by the standard knowledge base matching module, and, in conjunction with the problem types and standard clauses generated by the image compliance analysis module, output a detailed review report through the text generation module.

7. The BIM model automated review system according to claim 6, characterized in that, The professional image review model training module embeds a standard knowledge retrieval interface at the end of the model during training, so that when the model outputs a question, it synchronously associates it with relevant standard entries.

8. The BIM model automated review system according to claim 6, characterized in that, The review roaming path definition module generates a smooth review roaming path by specifying key points in the BIM model visualization interface and then using B-spline curve interpolation.

9. The BIM model automated review system according to claim 6, characterized in that, The optimization suggestion and report generation module dynamically associates problem types and normative clauses based on the normative knowledge base of the knowledge graph, and generates optimization suggestion information through the text generation module accordingly.

Images