Building entity-bim model deviation detection method, system, device and medium

Abstract

This invention discloses a method, system, equipment, and medium for detecting deviations between building entities and BIM models, relating to the field of building engineering technology. The method includes: collecting and fusing BIM model data and multimodal site data; inputting the fused data into an improved YOLOv8-3D model to detect geometric deviation parameters between the building entity and the BIM model; and detecting logical deviations between the building entity and the BIM model, or within the BIM model, based on the semantic attributes of the BIM data, a pre-defined specification knowledge graph, and the geometric deviation parameters. This invention utilizes channel segmentation and differentiated processing mechanisms to detect geometric deviations and logical deviations based on them, thereby significantly improving the detection accuracy of geometric deviation parameters while controlling detection overhead.

Description

Technical Field

[0001] This invention relates to the technical field of building engineering, and in particular to methods, systems, equipment and media for detecting deviations between building entities and BIM models. Background Technology

[0002] In the field of architectural engineering, the consistency inspection between the actual building and the BIM (Building Information Modeling) model is a core aspect of ensuring construction quality. Among these aspects, the accurate measurement of geometric deviations—that is, detecting the differences between the position and dimensions of the actual structure (such as beams, columns, and walls) and the design model—is fundamental to ensuring that the building meets design requirements and avoids potential structural safety hazards.

[0003] Currently, automated detection technologies for geometric deviations mainly revolve around comparing 3D laser scanning with BIM models, and their mainstream technical approaches are as follows: 1. Automated Comparison Technology of 3D Laser Scanning and BIM: Point cloud data of the construction site is acquired through 3D laser scanning to characterize the surface geometry of the building entity. Point cloud registration algorithms (such as ICP Iterative Closest Point Algorithm and NDT Normal Distribution Transform Algorithm) are then used to align the point cloud data with the BIM model. Based on the registration, the geometric deviation of each point is quantified by calculating the shortest distance from the point cloud surface to the BIM model surface. For example, in the inspection of complex structures such as pump gates and flow channels, the professional software BuidIT Construction achieves "best fit alignment" between the point cloud and the BIM model, outputting a deviation heatmap. Its advantages lie in non-contact measurement and improved accuracy compared to traditional manual measurement (average point cloud stitching error ±2.3mm), supporting geometric deviation analysis of complex structures (such as irregularly shaped flow channels). 2. Static Collision Detection Technology Based on Knowledge Rules: During the BIM design phase, collision detection is performed using predefined rules (such as pipeline spacing thresholds), generating a conflict report. For example, the random inspection of BIM application models in Shenzhen uses a component attribute rule library (such as duct size standards) for automated verification. Its advantage is that it can detect design conflicts in advance (such as interference between pipes and structural columns) and reduce the construction rework rate.

[0004] The aforementioned technical approaches have been applied to some extent in engineering practice. Among them, when processing fused multimodal building data, the general 3D convolutional neural network performs global and indiscriminate convolution operations on the input feature map in the backbone. Although this method can extract three-dimensional spatial features, it has a large computational load and insufficient sensitivity to features of various types of information. In particular, when processing feature maps that fuse point cloud geometric information and RGB-D image texture information, the feature maps contain both basic spatial structural features that represent the outline and boundary of components, as well as deep semantic features that reflect material, color and subtle depth changes. Under the condition of large detection overhead, the detection effect has the problem of low accuracy. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a method, system, equipment, and medium for detecting deviations between building entities and BIM models, thereby solving the problems in the prior art.

[0006] This invention specifically provides the following technical solution: a method for detecting deviations between a building entity and a BIM model, comprising the following steps: Collect BIM model data and multimodal site data, and then fuse the BIM model data and multimodal site data; The fused data is input into an improved YOLOv8-3D model to detect geometric deviation parameters between the building entity and the BIM model. The improved YOLOv8-3D model is based on the YOLOv8 model, replacing the 2D convolutional layers in the backbone with 3D convolutional layers to segment the feature map of the fused data by channel dimension. One part retains its original features, while the other part undergoes depth feature extraction and dimensional transformation. The processed data from both parts is then concatenated and compressed to obtain the final feature map. The 2D convolution and upsampling operations in the FPN feature pyramid network and PAN path aggregation network in the neck section are replaced with 3D versions to perform 3D convolution processing on the final feature map for top-down and bottom-up paths. The 3D convolution processing results are output as the head section of the YOLOv8 model to obtain the geometric deviation parameters to be predicted for each preset anchor point or each grid. Based on the semantic attributes of the BIM model data, the preset standard knowledge graph, and the geometric deviation parameters, the logical deviation between the building entity and the BIM model or within the BIM model is obtained.

[0007] Preferably, the fusion of BIM model data and multimodal site data specifically involves: Based on the point cloud in the multimodal field data and the geometric surface image in the BIM model data, feature matching is performed to obtain the initial spatial transformation relationship, and the initial transformation matrix is ​​generated based on the initial spatial transformation relationship. An optimizer is introduced to dynamically adjust the step size of the ICP algorithm, and the initial transformation matrix is ​​further registered using the improved ICP algorithm. Based on further registration results, the registration error between the registered point cloud and the target point cloud is obtained through Hausdorff distance, and the point cloud with the smallest registration error is used as the fused data under a unified coordinate system.

[0008] Preferably, before performing feature matching to obtain the initial spatial transformation relationship, the method further includes: The point cloud in the multimodal field data and the geometric surface image in the BIM model data are aligned and fused through calibration parameters to generate an enhanced point cloud with color or texture information. Geometric features are extracted from the augmented point cloud, and key points of the geometric surface are sampled from the BIM model as the basis data for feature matching.

[0009] Preferably, based on the semantic attributes of the BIM model data, a preset standard knowledge graph, and geometric deviation parameters, the logical deviation between the building entity and the BIM model, or within the BIM model, is obtained, specifically as follows: Constructing or invoking a canonical knowledge graph; The BIM components to be inspected and their spatial relationships are converted into query sub-graphs; The query subgraph is matched with a predefined conflict subgraph in the standard knowledge graph. If the match is successful, it is determined that there is a logical deviation.

[0010] Preferably, the query subgraph is matched with a predefined conflict subgraph in the canonical knowledge graph, specifically as follows: Obtain subgraph intersection rules, specifically including: pipe penetration through beams rule: requiring detection of spatial penetration between MEP pipes and structural beams; electromechanical collision rule: detecting violations of minimum spacing between pipelines of different systems; construction conflict rule: obtaining 4D progress data combining three-dimensional space and time dimensions, and detecting spatiotemporal conflicts between temporary supports and permanent components based on the 4D progress data. An R-tree is used to establish a query index for the spatial location of components. Conflicting component pairs are filtered out by finding spatial relationships. Triangle-facet level intersection calculations are performed on the filtered candidate components. The conflict level is divided according to the intersection volume and the importance dimension of the components after the intersection calculation. Based on the conflict level, dynamic weight allocation is performed on the intersection of conflict subgraphs according to preset rules. The query subgraph is matched with the predefined conflict subgraph in the standard knowledge graph with the corresponding weights.

[0011] Preferably, the standardized knowledge graph specifically comprises: multiple rules extracted using preset standards, and supplementary component connection logic from industry standards, stored in a knowledge graph, wherein the nodes of the knowledge graph are component type-attribute-rule, and the edges of the knowledge graph are dependency relationships.

[0012] Preferably, the other part of the deep feature extraction and dimensionality transformation is implemented by n cascaded Bottleneck3D units, where each Bottleneck3D typically consists of multiple 3D convolutional layers, including convolutional dimensionality reduction, feature transformation, and convolutional dimensionality enhancement.

[0013] This invention provides a building entity-BIM model deviation detection system, comprising: The data acquisition module is used to collect BIM model data and multimodal site data, and to fuse the BIM model data and multimodal site data. The geometric deviation module is used to input the fused data into an improved YOLOv8-3D model to detect geometric deviation parameters between the building entity and the BIM model. The improved YOLOv8-3D model is based on the YOLOv8 model, replacing the 2D convolutional layers in the backbone with 3D convolutional layers. This is used to segment the feature map of the fused data by channel dimension. One part retains its original features, while the other part undergoes depth feature extraction and dimensional transformation. The processed data from both parts is then concatenated and compressed to obtain the final feature map. The 2D convolution and upsampling operations in the FPN feature pyramid network and PAN path aggregation network in the neck section are replaced with 3D versions. These are used to perform 3D convolution processing on the final feature map for top-down and bottom-up paths. The 3D convolution processing results are output as the head section of the YOLOv8 model to obtain the geometric deviation parameters to be predicted for each preset anchor point or each grid. The logical deviation module is used to obtain the logical deviation between the building entity and the BIM model or within the BIM model based on the semantic attributes of the BIM model data, the preset standard knowledge graph, and the geometric deviation parameters.

[0014] The present invention provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-described method for detecting deviations in a building entity-BIM model.

[0015] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method for detecting deviations in a building entity-BIM model.

[0016] Compared with the prior art, the present invention has the following significant advantages: This invention introduces an improved YOLOv8-3D model, innovatively employing a channel segmentation and differential processing mechanism in the backbone network. After segmenting the feature map along the channel dimension, one part retains the original basic spatial structural features (such as component outlines and boundaries), while the other part undergoes dedicated depth feature extraction and dimensional transformation to capture subtle texture and depth variations. This decoupled processing avoids redundant global computation, significantly reducing the model's computational complexity. Simultaneously, customized processing of features at different levels enhances the model's sensitivity to key information, ensuring that after fusing point cloud geometric information and RGB-D image texture information, it can accurately identify macroscopic geometric contours while capturing microscopic material and depth differences. Finally, through feature splicing compression and full 3D path aggregation in the neck network, a set deviation parameter is obtained. Based on this deviation parameter, the final logical deviation is derived, thereby significantly improving the detection accuracy of the geometric deviation parameter while controlling detection overhead. Attached Figure Description

[0017] Figure 1 The present invention provides a flowchart of a method for detecting deviations in a building entity-BIM model. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0019] By dissecting the existing technical solutions, the following common defects can be found: (1) Insufficient data registration accuracy: The registration error between point cloud and BIM model is > ±5mm, and the local error of irregular structure is > ±9mm. It relies on a single registration algorithm (such as ICP) and does not integrate multimodal data (RGB-D image) to enhance feature descriptors; (2) Single deviation detection type: It only supports geometric deviation (displacement, size), and the missed detection rate of logical deviation (standard conflict) is > 30%. It has not built a computable standard knowledge graph and lacks reasoning mechanisms such as subgraph matching; (3) Low degree of automation: The manual intervention accounts for 80%, including point cloud splicing, rule threshold setting, and report interpretation. It has not realized the end-to-end AI process, and the detection and decision modules are separated; (4) Low computational efficiency: Point cloud processing of complex structures (such as pump gate flow channels) takes several hours, and single component detection takes > 30 seconds. The computational load of the algorithm has not been optimized (such as global point cloud matching), and distributed computing acceleration has not been introduced.

[0020] Compared with the existing technology, the advantages of this invention are reflected in the step-by-step solution of the above defects and the improvement of registration accuracy through multimodal fusion. Specifically: (1) Improve the ICP algorithm and fuse FPFH feature descriptors, combine laser point cloud (±1mm accuracy) and RGB-D image to control the registration error within ±1mm. (2) Geometric + logical dual-dimensional detection: Based on Neo4j, construct a standard knowledge graph (GB / T 51269 and 2400+ rules), and detect logical deviations such as pipeline collisions through subgraph matching. (3) End-to-end automated decision chain: Integrate YOLOv8-3D pose detection, PointNet++ matching, and LLaMA-7B repair suggestion generation, reducing manual intervention by 80%. (4) Algorithm optimization to accelerate calculation: By decomposing the pose transformation matrix by SVD and performing meshed local calculation (replacing global matching), the detection time of a single component is ≤5 seconds.

[0021] The following is combined Figure 1 The following describes specific embodiments of the present invention in further detail. These embodiments are merely for illustrating the technical solutions of the present invention more clearly and should not be construed as limiting the scope of protection of the present invention.

[0022] This invention proposes a technical solution comprising the following steps: Step 1: Collect BIM model data and multimodal site data.

[0023] (1) BIM model data: Format: IFC (Industry Foundation Classes) standard format, containing component geometric information (coordinates, dimensions, rotation angles) and semantic attributes (type, material, specification number); Parsing tool: Use the IfcOpenShell library to extract component attributes and store them as a structured database (such as SQLite).

[0024] (2) Multimodal field data: 1) Laser point cloud: Faro Focus scanner (accuracy ±1mm), single-station scanning time ≤2 minutes, point density ≥100 points / m². 2) RGB-D image: Intel RealSense D455 camera (resolution 1280×720, depth error ±1%), multi-view images acquired through UAV oblique photography. 3) Auxiliary sensors: GPS (positioning accuracy ±5cm), IMU (attitude angle error ±0.5°).

[0025] Data preprocessing: (1) Point cloud denoising and registration: The statistical outlier removal (SOR) algorithm (a commonly used filtering algorithm in point cloud processing, which will not be elaborated here) is used to remove noise points and retain 95% of the effective point cloud. Multi-site point cloud registration: The NDT (Normal Distributions Transform) algorithm (a conventional point cloud registration algorithm that optimizes the point cloud alignment effect through a probability model and is suitable for processing sparse or locally inconsistent point cloud data) is used for initial alignment, with a registration error ≤ ±2mm. (2) Image enhancement: Illumination normalization (CLAHE algorithm - setting a threshold for the histogram distribution, distributing the distribution exceeding the threshold evenly to the probability density distribution, thereby limiting the amplification of the transformation function and improving the image contrast) and distortion correction (Brown-Conrady model - eliminating image distortion caused by lens optical characteristics) are performed on the RGB-D image.

[0026] Step 2: Spatially align the BIM model data and multimodal site data to output fused data in a unified coordinate system.

[0027] The collected multimodal field data needs to undergo fusion preprocessing: the point cloud from the multimodal field data is aligned and fused with the RGB-D geometric surface image from the BIM model data using calibration parameters to generate an enhanced point cloud with color or texture information. Subsequently, geometric features (FPFH descriptors) are extracted from the enhanced point cloud, and key points of the geometric surface are sampled from the BIM model as the basis for feature matching. The generation of the initial transformation matrix mainly depends on the matching of these two sets of 3D geometric features, while multimodal information (such as color) can provide additional context for the robustness of feature matching or be used for optimization in subsequent fine registration stages.

[0028] In the registration process, the improved ICP registration algorithm is as follows: (1) Technical process: 1) Coarse Registration: Based on the Fast Point Feature Histogram (FPFH) feature descriptor, point clouds (i.e., point clouds in multimodal field data) are matched with key points on the BIM model surface (i.e., geometric surface images in BIM model data). Feature matching is performed to obtain the initial spatial transformation relationship, and an initial transformation matrix is ​​generated based on the initial spatial transformation relationship (loading point cloud data - using relevant MATLAB functions to read point cloud files, extracting FPFH features - calculating FPFH feature vectors for the neighborhood of each point, matching point pairs - comparing the similarity of FPFH features of two points to match point pairs with high similarity, calculating the initial transformation matrix - estimating the initial rotation matrix and translation vector based on the matched point pairs to achieve coarse registration. However, due to the high computational complexity and long time of FPFH, it is only used for coarse registration here).

[0029] 2) Fine registration: An improved ICP (Iterative Closest Point) algorithm is adopted, which introduces the Levenberg-Marquardt optimizer to dynamically adjust the step size of the ICP algorithm to avoid local optima. The improved ICP algorithm is used to further register the initial transformation matrix.

[0030] 3) Accuracy verification: Based on further registration results, the registration error between the registered point cloud and the target point cloud is obtained through Hausdorff distance and controlled within ±1mm (±5mm for traditional methods). The point cloud with the smallest registration error is used as the fused data under a unified coordinate system.

[0031] The objective function is reconstructed as a nonlinear least squares problem, and the damping factor of the LM algorithm is used to dynamically adjust the optimization step size. Its error function can be expressed as: By adjusting the Hessian matrix The damping coefficient λ is used to balance gradient descent and iteration step size. The algorithm flow is as follows:

[0032] Initialization: Set the initial transformation T0 (identity matrix) and LM parameters (initial λ, maximum number of iterations).

[0033] Iterative optimization: Corresponding point matching: For the transformed source point cloud P', search for the nearest neighbor pair in the target point cloud Q; Error and Jacobian calculation: Calculate the residual vector and Jacobian matrix J; Solving the LM matrix: Construct... Solve for the increment Transformation update: Apply the increment Δx to update the transformation matrix T, and adjust λ according to the error change.

[0034] Termination condition: Stop when the error change is less than the threshold or when the maximum number of iterations is reached.

[0035] (2) Experimental data: In the testing of steel structure workshops, the registration time averaged 8 seconds per component (traditional ICP requires 15 seconds), and the error distribution is shown in Table 1 below: Table 1 Error Distribution Coordinate system one, specifically: Coordinate transformation chain: Camera coordinate system → World coordinate system: Construct translation matrix TTT and rotation matrix RRR using IMU / GPS data (formula: Pworld=R⋅Pcamera+TP_{\text{world}}=R\cdot P_{\text{camera}}+TPworld=R⋅Pcamera+T). World coordinate system → BIM coordinate system: Achieve coordinate mapping through rigid body transformation with the BIM model origin as the reference.

[0036] Step 3: Input the fused data into the improved YOLOv8-3D model to detect the geometric deviation parameters between the building entity and the BIM model. The improved YOLOv8-3D model is based on the YOLOv8 model. The 2D convolutional layers in the backbone are replaced with 3D convolutional layers to perform channel dimension segmentation on the feature map of the fused data. After segmentation, one part retains the original features, while the other part undergoes depth feature extraction and dimensional transformation. The processed data of the two parts are then concatenated and compressed to obtain the final feature map. The 2D convolution and upsampling operations in the FPN feature pyramid network and PAN path aggregation network in the neck part are replaced with 3D versions to perform 3D convolution processing on the final feature map for top-down and bottom-up paths. The 3D convolution processing results are output as the head part of the YOLOv8 model to obtain the geometric deviation parameters that need to be predicted for each preset anchor point or each grid.

[0037] Specifically, the 3D object detection (YOLOv8-3D architecture) is as follows: (1) Model improvement: YOLOv8 itself is a single-stage detector designed for 2D images, and 3D adaptation is required to process point cloud or RGB-D data of building entities.

[0038] 1) Input Adjustment: The input data consists of features fused from multimodal data, which may include a voxel grid of point cloud projections, RGB-D image sequences, or a single depth-enhanced image. For example, point clouds can be converted to voxel grids (dimensions [Batchsize, Channels, Depth, Height, Width], where Depth represents time (number of consecutive frames), or RGB-D images can be used as 4-channel inputs (RGB + depth). If time-series data (such as video clips) is used, the input needs to be adjusted to [Batchsize, Channels, Time, Height, Width] (Time being the number of frames), and spatiotemporal features are extracted through 3D convolution.

[0039] 2) Output Data: The ultimate goal of the detection head is to predict the 6-DOF pose of an object in 3D space. This includes: center coordinates (x, y, z): the object's position in 3D space; dimensions (w, h, d): the object's width, height, and depth; rotation angles (θx, θy, θz): the object's rotation angles around each coordinate axis. Additionally, the confidence score and class probability for each bounding box are output.

[0040] 3) Network structure transformation (3D): a) Backbone: Retains CSPDarknet as the backbone, but replaces 2D convolutional layers with 3D convolutional layers to capture spatial depth features (e.g., a basic 3×3 convolutional kernel becomes a 3×3×3 cubic convolutional kernel). This enables the network to process voxelized point cloud data or RGB-D video sequences; the 3D convolutional kernels slide simultaneously in both spatial (height, width) and depth dimensions, thereby extracting features with spatiotemporal correlation.

[0041] The input feature map is first processed by a 1×1×1 Conv3d algorithm for channel adjustment and preliminary feature integration. Subsequently, the feature map is split into two parts along the channel dimension; one part enters the main branch for complex processing, while the other part is passed directly as a sub-branch to preserve the original gradient flow and shallow feature information.

[0042] The main branch performs deep feature extraction and dimensionality transformation using n concatenated Bottleneck3D units (the number is configurable). Each Bottleneck3D unit typically consists of multiple 3D convolutional layers, such as 1×1×1 convolutions for dimensionality reduction, 3×3×3 convolutions for feature transformation, followed by another 1×1×1 convolution for dimensionality increase, and may include residual connections. The features processed by the main branch are concatenated with the original features from the sub-branches. The concatenated features are then compressed through a 1×1×1 Conv3D unit to integrate the information and output the final feature map.

[0043] b) Neck: Upgrades the 2D convolution and upsampling operations in traditional FPN (Feature Pyramid Network) and PAN (Path Aggregation Network) to 3D versions (upsampling operations use trilinear interpolation, and convolution uses Conv3d).

[0044] Top-down Path (FPN): The deepest (smallest) feature map is upsampled using 3D transposed convolution or 3D interpolation to match its size with the feature map of the next layer. The upsampled features are then element-wise added or concatenated with the feature maps from the corresponding layers in the Backbone. This process is performed layer by layer, passing high-level semantic information to lower layers.

[0045] Bottom-up Path (PAN): The fused low-level features are downsampled again using 3D convolution and then fused with features from higher levels. This step propagates detailed information from the lower levels upwards, further optimizing feature representation.

[0046] c) Head: The detection head is responsible for predicting the final 3D bounding box parameters based on the fused features. The data comes from the multi-scale 3D feature map fused from the Neck, and is ultimately output through multiple 3D convolutional layers, specifying the parameters to be predicted for each preset anchor point or each grid cell, as follows:

[0047] 3D center coordinates (x, y, z): the object's position in 3D space; 3D dimensions (w, h, d): the object's width, height, and depth (or length); three rotation angles (θx, θy, θz): the object's rotation angles around the X, Y, and Z axes. Sometimes, for simplicity, only a yaw angle may be predicted, or other rotation representations such as quaternions may be used; confidence score (obj): the probability that an object exists within the bounding box; class probability (cls1, cls2, ... clsn): the probability that the object belongs to each class. The differences in this invention are shown in Table 2 below.

[0048] Table 2 Differences in Improvements (2) Training and performance: Dataset: 100,000 self-made samples (annotated with 6-DoF poses) + ScanNet public dataset. Accuracy indicators: mAP@0.5 reached 95.3%, displacement error ±1.2mm, rotation angle error ±0.8° (test set: steel structure beam-column joints).

[0049] Point cloud-BIM model matching: (1) Technical process: Feature extraction: PointNet++ samples point clouds and BIM model surface points to generate 1024-dimensional feature vectors. SVD pose calculation: The optimal transformation matrix is ​​calculated through singular value decomposition (SVD) to minimize the distance between the point cloud and the model surface. Deviation quantization: The translation vector (Δx, Δy, Δz) and rotation matrix deviation (Δθ) are output.

[0050] Step 4: Based on the semantic attributes of the BIM model data, the preset standard knowledge graph, and the geometric deviation parameters, obtain the logical deviation between the building entity and the BIM model or within the BIM model.

[0051] The semantic attributes of BIM model data refer to the non-geometric information used to describe building components or systems in terms of function, behavior, specifications, and life cycle, including component type, physical attributes, etc.

[0052] Constructing and invoking a canonical knowledge graph, specifically: (1) Rule base source: 2400+ rules extracted from national standards (i.e., preset standards) (GB / T 51269-2017) (such as "pipe spacing ≥ 1.2m"), and supplemented by industry specifications (ISO 16739) to provide component connection logic (such as "beam-column joints must meet seismic resistance level").

[0053] (2) Knowledge graph storage: Neo4j graph database. The nodes of the knowledge graph are component type-attribute-rule, and the edges of the knowledge graph are dependencies. Example: (pipeline)-[spacing constraint]->(fire protection code).

[0054] Subgraph matching algorithms (used to find subgraphs isomorphic to smaller graphs within a given larger graph, widely used in anomaly detection and other fields) are as follows: 1) Convert existing BIM model components (such as beams, columns, and pipelines) into knowledge graph nodes. Attributes include geometric dimensions, spatial coordinates, etc., and relationships include semantic links such as "adjacent" and "intersecting".

[0055] 2) Define conflict subgraph templates (such as pipeline intersection templates containing two pipeline nodes and "spatial overlap" relationship edges), and establish a rule base using the subgraph generation method of the AGDISTIS algorithm.

[0056] The BIM components to be inspected and their spatial relationships are converted into query sub-graphs, specifically: 1) Subgraph Generation Basis: Starting with the target entity and related entities in its context, a directed subgraph is generated within the knowledge graph. This process leverages the structural relationships within the knowledge graph, expanding the subgraph's scope through a breadth-first search (BFS) strategy.

[0057] 2) Expanded Graph Construction: Initialize the expanded graph: For each entity ci in the candidate entity set C, initialize an empty graph G0=(V0,E0), where V0={ci} and E0=Φ; Breadth-first expansion: Perform d expansion operations ρ. Each expansion merges the vertex set Vi of the current subgraph with the adjacent vertices and edges in the knowledge graph to form a new subgraph Gi+1. Specifically, Vi+1= Vi∪{y | (x,y)∈EKG, x∈Vi}, Ei+1= Ei∪{(x,y) | (x,y)∈EKG, x∈Vi}.

[0058] The query subgraph is matched with predefined conflict subgraphs in the canonical knowledge graph. If a match is found, a logical deviation is determined. Specifically: 3) Obtain subgraph intersection rules (conflict rule definition), specifically including: pipe through beam rule: requires detection of spatial penetration between MEP pipes and structural beams; electromechanical collision rule: detects violations of minimum spacing between pipelines of different systems (such as electrical cable trays and ventilation ducts); construction conflict rule: obtain 4D progress data combining three-dimensional space and time dimensions, and detect spatiotemporal conflicts between temporary supports and permanent components based on the 4D progress data.

[0059] 4) The specific process for subgraph intersection detection is as follows: a) Spatial index construction: R-trees are used to build a fast query index for the spatial location of components; b) Candidate set screening: Component pairs that may conflict are initially screened based on the spatial relationships found (such as bounding box intersections); c) Precise geometric verification: Triangle-facet level precise intersection calculations are performed on the screened candidate components; d) Conflict classification: Conflict levels are classified based on dimensions such as the intersection volume after intersection calculation and the importance of components.

[0060] 5) Based on the conflict level, dynamic weights are dynamically assigned to the intersection of conflicting subgraphs according to preset rules, and the query subgraph is matched with the predefined conflicting subgraphs in the standard knowledge graph with the corresponding weights; the dynamic weight calculation mechanism is as follows: Weighting rules: Traditional methods rely on HITS or PageRank algorithms to calculate node weights, but AGDISTIS improves this by calculating weights based on subgraph intersections, avoiding weight invalidation issues caused by zero in-degree. This example mainly defines the following rules:

[0061] Geometric weight: based on the proportion of intersecting volumes (0-1 normalization); Time weight: considers the overlap of construction stages (such as the conflict between the concrete pouring period and the pipeline installation period); Cost weight: the cost of modifying related components (such as the cost of reworking prefabricated steel structures).

[0062] Experimental results: In the detection of electromechanical pipelines, the logic deviation detection rate was 92.7%, and the false alarm rate was ≤3%.

[0063] Large-scale models are used to assist decision-making, specifically vertical industry large-scale models (LLaMA-7B+LoRA): (1) Training configuration: Base model: LLaMA-7B, fine-tuning data: 50,000 historical bias correction records. Fine-tuning method: LoRA (Low-Rank Adaptation), rank=8, training time 12 hours (4×A100).

[0064] (2) Repair suggestion generation: Input: Deviation type + parameter (e.g., "pipe offset 23mm"). Output: Natural language suggestion (e.g., "adjustment using flange connection, support and hanger position needs to be checked").

[0065] This invention employs VR interactive verification, specifically as follows: (1) Digital Twin Platform Integration: The BIM model is loaded using the Unity engine and presented as a VR environment via HTC Vive Pro. Users can adjust the position of components, and the system calculates the deviation correction amount in real time (response latency ≤50ms). The performance verification of this invention is shown in Table 3.

[0066] Table 3 Performance Verification The improved LM optimizer parameters for ICP are: initial step size μ=0.01, damping factor λ=10, and iteration limit of 100 (convergence speed improved by 60%). The YOLOv8-3D attention module is specifically positioned as follows: CBAM is inserted into layers 3 / 6 / 9 of the backbone network, focusing on small-scale defects. The knowledge graph rule priority is as follows: safety-related rules (such as fire separation distance) have a weight of 0.9, and non-critical rules have a weight of 0.5.

[0067] This invention achieves a breakthrough improvement in the accuracy, efficiency, and intelligence level of deviation detection between building entities and BIM models through three core technologies: multimodal data fusion, geometric-logical dual-dimensional detection, and large-model closed-loop decision-making. Compared with the closest existing technologies (such as 3D laser scanning comparison and static collision detection), this invention achieves significant optimizations in detection accuracy, coverage dimensions, automation level, and decision support, as detailed in Table 4 below:

[0068] Table 4 Evaluation Form The present invention has the following beneficial effects: 1. Improved accuracy and comprehensiveness: (1) Geometric deviation detection: The ICP algorithm was improved by integrating FPFH feature descriptors and dynamically adjusting the step size (initial μ=0.01, λ=10) with the Levenberg-Marquardt optimizer, reducing the registration error between point cloud and BIM model from ±5mm to ±0.9mm. The YOLOv8-3D model introduced the CBAM attention mechanism (layers 3 / 6 / 9) to improve the recognition of occluded components, achieving a displacement detection accuracy of ±1.2mm and a rotation angle error of ±0.8° (trained with 100,000 self-made samples). (2) Logical deviation coverage: A Neo4j knowledge graph (2400+ rules) was constructed based on standards such as GB / T 51269-2017, and non-geometric problems such as pipeline collisions and fireproof spacing violations were detected through subgraph matching. In the electromechanical pipeline test, the logical deviation detection rate was 92.7%, and the false alarm rate was ≤3% (the traditional method had a false alarm rate of >30%).

[0069] 2. Breakthroughs in efficiency and automation: (1) End-to-end process optimization: In the data alignment stage, SVD decomposition of pose transformation matrix is ​​used to replace traditional global point cloud matching, reducing the computational complexity from O(n²) to O(n log n), and the processing time of a single component is ≤5 seconds (traditional method >30 seconds). Distributed computing accelerates point cloud processing (e.g., 76.4% downsampling of 457 million point clouds), and the time consumption is only 1 / 6 of the traditional method. (2) Large model decision-making closed loop: The vertical industry large model (LLaMA-7B+LoRA fine-tuning) is trained based on 50,000 historical repair records to generate natural language suggestions (e.g., "pipeline offset 23mm, use flange connection adjustment"), with an adoption rate of 85%. VR interactive verification (HTC Vive Pro+Unity engine) realizes real-time preview of repair solutions, with a response latency of ≤50ms, reducing manual review time by 90%.

[0070] 3. Intelligence and Scalability: (1) Dynamic Knowledge Fusion: The knowledge graph supports incremental updates (such as adding GB / T51431-2020 specifications), and dynamically loads rules through the Cypher query language to adapt to different project needs. LoRA fine-tuning technology (rank r=8, lora_dropout=0.1) enables the model to quickly adapt to new tasks, and the training memory requirement is only 16GB (RTX 4080 can run it). (2) Multi-Scenario Compatibility: The modular design of the system supports the extension to the detection of irregular structures such as bridges and tunnels. In the pump gate flow channel test, the registration error of irregular curved surfaces is ≤±2.1mm (traditional methods are >±9mm). The technical advantages of this invention are compared with the existing technologies as shown in Table 5.

[0071] Table 5. Comparison of the technical advantages of this invention with existing technologies Existing technologies can lead to engineering risks, specifically: (1) Quality and safety risks: Geometric deviations (such as column offsets) that are not corrected in time may cause structural safety hazards; logical deviations (such as insufficient spacing of fire pipes) directly violate mandatory specifications. (2) Cost and schedule losses: Manual inspection is inefficient, and the rework rate for large projects is as high as 25%. For example, the pump gate flow channel is difficult to inspect due to its irregular shape, which leads to schedule delays. (3) Limited application of digital twins: Deviations between the entity and the model hinder the accurate simulation of the digital twin platform. For example, equipment maintenance fails due to model distortion during the operation and maintenance phase.

[0072] The root causes of the aforementioned project risks include: (1) Algorithm limitations: Traditional registration algorithms (such as ICP) are sensitive to the initial position and do not combine the adaptive feature extraction capabilities of deep learning. (2) Lack of data fusion: The lack of coordination of multimodal data (point cloud + image + BIM) leads to a lack of basis for detecting logical deviations. (3) System closed loop break: The detection results are disconnected from decision support, and there is a lack of natural language generation and interactive verification (such as VR) modules.

[0073] The core parameters of this invention are as follows: 1) Improved ICP optimizer parameters: Levenberg-Marquardt initial step size μ=0.01, damping factor λ=10, iteration limit 100 (convergence speed improved by 60%). FPFH feature subsampling rate 0.05, balancing accuracy and computational efficiency (time consumption only 53% of traditional ICP). 2) YOLOv8-3D attention mechanism deployment: CBAM module inserted into layers 3 / 6 / 9 of the backbone network, focusing on small-scale defects (recall improved by 12%). 3) LoRA fine-tuning configuration: rank r=8, target module ["q_proj", "v_proj"], retaining complete training of lm_head (medical question answering accuracy +23%).

[0074] Economic benefit analysis: 1) Direct Cost Savings: Manual Inspection Costs: Traditional large-scale projects require 20 person-days (approximately ¥80,000), while the automated inspection of this invention requires only 0.5 person-days (¥2,000), saving 97.5%. Reduced Rework Losses: The rate of missed inspections due to logical deviations is reduced by 22.7%, avoiding a rework rate greater than 25% (e.g., saving ¥1,200,000 in a pump gate project). 2) Indirect Benefit Improvements: Project Duration: End-to-end inspection speed is increased by 5 times, reducing the inspection cycle for a 100,000㎡ project from 7 days to 1.5 days. Digital Twin Value-Added: The system output interface is compatible with BIM platforms (such as Autodesk Revit), driving a 40% improvement in model accuracy during the operation and maintenance phase and extending the platform's lifespan.

[0075] This invention achieves comprehensive improvement in Jiangdu through geometric and logical deviations, achieves breakthroughs in efficiency and automation through end-to-end process optimization, and integrates dynamic knowledge to adapt to different project needs and is suitable for multiple scenarios.

[0076] Based on the above method, the present invention also provides a building entity-BIM model deviation detection system, comprising: The data acquisition module collects BIM model data and multimodal site data, and fuses them. The geometric deviation module inputs the fused data into an improved YOLOv8-3D model to detect geometric deviation parameters between the building entity and the BIM model. The improved YOLOv8-3D model, based on the YOLOv8 model, replaces the 2D convolutional layers in the backbone with 3D convolutional layers to segment the feature maps of the fused data by channel dimension. One part retains its original features, while the other part undergoes depth feature extraction and dimensional transformation. The processed data is then further processed... The features are concatenated and compressed to obtain the final feature map. The 2D convolution and upsampling operations in the FPN feature pyramid network and PAN path aggregation network in the neck part are replaced with 3D versions for 3D convolution processing of the final feature map from top to bottom and from bottom to top. The 3D convolution processing results are output as the head part of the YOLOv8 model to obtain the geometric deviation parameters that need to be predicted for each preset anchor point or each grid. The logical deviation module is used to obtain the logical deviation between the building entity and the BIM model or within the BIM model based on the semantic attributes of the BIM model data, the preset specification knowledge graph, and the geometric deviation parameters.

[0077] The present invention provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-described method for detecting deviations in a building entity-BIM model.

[0078] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).

[0079] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method for detecting deviations in a building entity-BIM model.

[0080] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0081] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for detecting deviations between a building entity and a BIM model, characterized in that, Including the following steps: Collect BIM model data and multimodal site data, and then fuse the BIM model data and multimodal site data; The fused data is input into an improved YOLOv8-3D model to detect geometric deviation parameters between the building entity and the BIM model. The improved YOLOv8-3D model is based on the YOLOv8 model, replacing the 2D convolutional layers in the backbone with 3D convolutional layers to segment the feature map of the fused data by channel dimension. One part retains its original features, while the other part undergoes depth feature extraction and dimensional transformation. The processed data from both parts is then concatenated and compressed to obtain the final feature map. The 2D convolution and upsampling operations in the FPN feature pyramid network and PAN path aggregation network in the neck section are replaced with 3D versions to perform 3D convolution processing on the final feature map for top-down and bottom-up paths. The 3D convolution processing results are output as the head section of the YOLOv8 model to obtain the geometric deviation parameters to be predicted for each preset anchor point or each grid. Based on the semantic attributes of the BIM model data, the preset standard knowledge graph, and the geometric deviation parameters, the logical deviation between the building entity and the BIM model or within the BIM model is obtained.

2. The method for detecting deviations between a building entity and a BIM model as described in claim 1, characterized in that, The fusion of BIM model data and multimodal site data specifically involves: Based on the point cloud in the multimodal field data and the geometric surface image in the BIM model data, feature matching is performed to obtain the initial spatial transformation relationship, and the initial transformation matrix is ​​generated based on the initial spatial transformation relationship. An optimizer is introduced to dynamically adjust the step size of the ICP algorithm, and the initial transformation matrix is ​​further registered using the improved ICP algorithm. Based on further registration results, the registration error between the registered point cloud and the target point cloud is obtained through Hausdorff distance, and the point cloud with the smallest registration error is used as the fused data under a unified coordinate system.

3. The method for detecting deviations between a building entity and a BIM model as described in claim 2, characterized in that, Before performing feature matching to obtain the initial spatial transformation relationship, the method further includes: The point cloud in the multimodal field data and the geometric surface image in the BIM model data are aligned and fused through calibration parameters to generate an enhanced point cloud with color or texture information. Geometric features are extracted from the augmented point cloud, and key points of the geometric surface are sampled from the BIM model as the basis data for feature matching.

4. The method for detecting deviations between a building entity and a BIM model as described in claim 1, characterized in that, Based on the semantic attributes of the BIM model data, the preset standard knowledge graph, and geometric deviation parameters, the logical deviations between the building entity and the BIM model, or within the BIM model, are obtained, specifically: Construct and invoke the canonical knowledge graph; The BIM components to be inspected and their spatial relationships are converted into query sub-graphs; The query subgraph is matched with a predefined conflict subgraph in the standard knowledge graph. If the match is successful, it is determined that there is a logical deviation.

5. The method for detecting deviations between a building entity and a BIM model as described in claim 4, characterized in that, The query subgraph is matched with predefined conflict subgraphs in the canonical knowledge graph, specifically as follows: Obtain subgraph intersection rules, specifically including: pipe penetration through beams rule: requiring detection of spatial penetration between MEP pipes and structural beams; electromechanical collision rule: detecting violations of minimum spacing between pipelines of different systems; construction conflict rule: obtaining 4D progress data combining three-dimensional space and time dimensions, and detecting spatiotemporal conflicts between temporary supports and permanent components based on the 4D progress data. An R-tree is used to establish a query index for the spatial location of components. Conflicting component pairs are filtered out by finding spatial relationships. Triangle-facet level intersection calculations are performed on the filtered candidate components. The conflict level is divided according to the intersection volume and the importance dimension of the components after the intersection calculation. Based on the conflict level, dynamic weight allocation is performed on the intersection of conflict subgraphs according to preset rules. The query subgraph is matched with the predefined conflict subgraph in the standard knowledge graph with the corresponding weights.

6. The method for detecting deviations between a building entity and a BIM model as described in claim 4, characterized in that, The specified knowledge graph is specifically composed of multiple rules extracted using preset standards, as well as industry standard supplementary component connection logic, and stored through a knowledge graph. The nodes of the knowledge graph are component type-attribute-rule, and the edges of the knowledge graph are dependency relationships.

7. The method for detecting deviations between a building entity and a BIM model as described in claim 1, characterized in that, The other part, which performs deep feature extraction and dimensionality transformation, is implemented through n cascaded Bottleneck3D units, where each Bottleneck3D typically consists of multiple 3D convolutional layers, including convolutional dimensionality reduction, feature transformation, and convolutional dimensionality enhancement.

8. A deviation detection system for building entity-BIM model, characterized in that, include: The data acquisition module is used to collect BIM model data and multimodal site data, and to fuse the BIM model data and multimodal site data. The geometric deviation module is used to input the fused data into an improved YOLOv8-3D model to detect geometric deviation parameters between the building entity and the BIM model. The improved YOLOv8-3D model is based on the YOLOv8 model, replacing the 2D convolutional layers in the backbone with 3D convolutional layers. This is used to segment the feature map of the fused data by channel dimension. One part retains its original features, while the other part undergoes depth feature extraction and dimensional transformation. The processed data from both parts is then concatenated and compressed to obtain the final feature map. The 2D convolution and upsampling operations in the FPN feature pyramid network and PAN path aggregation network in the neck section are replaced with 3D versions. These are used to perform 3D convolution processing on the final feature map for top-down and bottom-up paths. The 3D convolution processing results are output as the head section of the YOLOv8 model to obtain the geometric deviation parameters to be predicted for each preset anchor point or each grid. The logical deviation module is used to obtain the logical deviation between the building entity and the BIM model or within the BIM model based on the semantic attributes of the BIM model data, the preset standard knowledge graph, and the geometric deviation parameters.

9. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores a program that, when executed by the processor, causes the processor to perform the steps of the method for detecting deviations in a building entity-BIM model as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for detecting deviations in a building entity-BIM model according to any one of claims 1 to 7.

Images