Existing building structure deformation detection method and system based on three-dimensional laser scanning

Abstract

The present application relates to the technical field of intelligent sensing system, and more particularly to a kind of existing building structure deformation detection method and system based on three-dimensional laser scanning, comprising: obtaining building initial point cloud model by three-dimensional laser sensor, extracting geometric and physical reflection characteristics to generate semantic label and aggregate entity component using intelligent sensing algorithm;Based on semantic label and adjacency matrix, topological law is mined;Positioning topological conflict area, calculate the occlusion probability in combination with point cloud density, generate virtual component to repair the blind area of perception according to topological law when the completion condition is met, complete initial point cloud model;Obtain deformation reference to calculate component deformation, extract new structure deformation, and conduct safety trend evaluation combined with semantic label.The application realizes semantic analysis and automatic completion of component, and improves the accuracy of deformation detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing system technology, specifically to a method and system for detecting deformation of existing building structures based on three-dimensional laser scanning. Background Technology

[0002] As urbanization enters the stock renewal phase, existing buildings experience gradual structural performance degradation due to material aging, foundation settlement, and changes in usage. This necessitates efficient and accurate deformation detection methods to assess their safety status. Traditional manual measurement methods (such as levels and total stations) suffer from drawbacks such as sparse measurement points, low efficiency, and difficulty in covering complex spaces. Therefore, 3D laser scanning technology, with its advantages of non-contact operation, high density, and full-scene coverage, is gradually being introduced into the field of building structural deformation detection.

[0003] However, existing methods for detecting deformation in existing buildings based on 3D laser scanning still suffer from the following major technical shortcomings: First, they lack component-level semantic information parsing. Existing methods typically only perform simple geometric coordinate subtraction calculations on the point cloud sets from two scans to obtain displacement or settlement, completely ignoring the semantic attributes of the specific building components (such as load-bearing columns, decorative beams, and non-load-bearing partitions) represented by the point cloud. This makes it impossible for the system to distinguish whether the deformation originates from critical load-bearing components or insignificant decorative components, thus failing to accurately and scientifically assess the impact of local deformation on the overall structural stability of the building. Second, they struggle to handle scenarios with missing or occluded point clouds. When there are obstructions inside the existing building, or when the point cloud scan of the structural surface is incomplete due to aging renovations, existing methods cannot automatically identify and infer these complex situations. Inspectors must rely on manual methods, spending a significant amount of time performing tedious manual comparisons and subjective inferences against the original design drawings. This not only severely reduces the efficiency of automated inspection but also easily introduces human comparison errors. Summary of the Invention

[0004] To address the low accuracy of existing three-dimensional laser scanning-based methods for detecting deformation in existing buildings, this invention aims to provide a method and system for detecting structural deformation in existing buildings based on three-dimensional laser scanning. The specific technical solution adopted is as follows:

[0005] One embodiment of the present invention provides a method for detecting deformation of existing building structures based on three-dimensional laser scanning, the method comprising the following steps:

[0006] Obtain the initial point cloud model of the existing building;

[0007] The initial point cloud model is segmented to obtain individual point cloud blocks; based on the geometric shape features and physical reflection features of each point cloud block, a semantic classification label is obtained for each point cloud block, and the point cloud blocks are merged according to the semantic classification labels to obtain several entity components.

[0008] Based on the semantic classification labels of the entity components and the adjacency matrix constructed from the spatial distribution of each entity component, the association rule algorithm is used to mine the frequent itemsets of the component arrangement to generate building topology patterns.

[0009] The building topology rules are used to verify the current physical components, locate the three-dimensional spatial region where topological conflicts occur, determine the occlusion probability by combining the three-dimensional local point cloud density of the three-dimensional spatial region, and generate virtual components in the three-dimensional spatial region according to the building topology rules when the occlusion probability meets the completion condition to complete the initial point cloud model.

[0010] Obtain a deformation reference benchmark, calculate the deformation of each component based on the completed current point cloud model, filter construction deviations by combining a preset tolerance threshold, extract the newly added structural deformation, and conduct a structural safety trend assessment by combining the semantic classification labels of the corresponding components.

[0011] Another embodiment of the present invention provides a deformation detection system for existing building structures based on three-dimensional laser scanning, the system comprising:

[0012] The point cloud acquisition and preprocessing module is used to acquire the initial point cloud model of the existing building;

[0013] The component acquisition module is used to segment the initial point cloud model to obtain individual point cloud blocks; obtain semantic classification labels for each point cloud block based on its geometric shape features and physical reflection features; and merge the point cloud blocks based on the semantic classification labels to obtain several entity components.

[0014] The topology pattern generation module is used to generate building topology patterns by constructing an adjacency matrix based on the semantic classification labels of the entity components and the spatial distribution between the entity components, and by using an association rule algorithm to mine frequent itemsets of the component arrangement.

[0015] The topology mining and completion module is used to verify the current number of entity components using the building topology rules, locate the three-dimensional spatial region where topological conflicts occur, determine the occlusion probability by combining the three-dimensional local point cloud density of the three-dimensional spatial region, and generate virtual components in the three-dimensional spatial region according to the building topology rules when the occlusion probability meets the completion condition, so as to complete the initial point cloud model.

[0016] The deformation analysis and safety assessment module is used to obtain deformation reference benchmarks, calculate the deformation of each component based on the completed current point cloud model, filter construction deviations by combining preset tolerance thresholds, extract newly added structural deformation, and conduct structural safety trend assessment by combining the semantic classification labels of the corresponding components.

[0017] Another embodiment of the present invention provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a method for detecting deformation of existing building structures based on three-dimensional laser scanning.

[0018] The present invention has the following beneficial effects:

[0019] This invention provides a method and system for detecting deformation of existing building structures based on 3D laser scanning. First, the method achieves semantic parsing at the component level. By extracting geometric shape and physical reflection features to assign semantic classification labels to components, and performing safety assessments based on these labels, it can accurately distinguish the deformation of load-bearing components from decorative components, and realistically and scientifically assess the impact of local deformation on the overall stability of the building. Second, it overcomes the defects of missing point clouds and physical occlusion. Based on the topological patterns of the building from the solid components, it locates topological conflicts and calculates the occlusion probability by combining local point cloud density, automatically generating virtual component completion models in missing areas. This, to a certain extent, eliminates the reliance on subjective inference from manual comparison with drawings, improves the efficiency of automated detection, and eliminates human comparison errors. Finally, it accurately distinguishes between construction deviations and actual deformation. A preset tolerance threshold is introduced when calculating deformation, effectively filtering out initial construction deviations naturally inherited during the construction phase, accurately extracting the actual new structural deformation that occurs during the service life, and significantly reducing the false alarm rate of safety warnings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the steps of a method for detecting deformation of existing building structures based on three-dimensional laser scanning, according to an embodiment of the present invention.

[0022] Figure 2 This is a flowchart illustrating the process of obtaining building topology patterns in an embodiment of the present invention. Detailed Implementation

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] One embodiment of the present invention provides a method for detecting deformation of existing building structures based on three-dimensional laser scanning, such as... Figure 1 As shown, it includes the following steps:

[0025] S1, obtain the initial point cloud model of the existing building.

[0026] Traditional single-point measurement methods have sparse measurement points and blind spots, failing to comprehensively reflect the overall structural state of a building. Obtaining an initial point cloud model aims to establish a digital geometric foundation containing dense 3D coordinates and reflection intensity information through non-contact panoramic scanning. This provides raw data support for subsequent extraction of semantic topological relationships of components and comparative calculation of newly added structural deformations. The specific steps for obtaining an initial point cloud model of an existing building include:

[0027] The first step is station deployment and multi-source data collection.

[0028] Multiple 3D laser scanning stations are deployed inside and around the existing building. To ensure the accuracy and coherence of subsequent point cloud stitching, scanning paths need to be planned, and the overlap rate of the scanning areas of adjacent stations needs to be controlled to be above 30%. The scanner (3D laser sensor) is activated to acquire the original local point cloud data of each station. Each set of data not only contains the 3D coordinates of the points, but also records the surface reflection intensity, forming multiple sets of unstitched local point cloud sets.

[0029] The second step is voxel downsampling and statistical noise reduction processing.

[0030] Due to the large volume of original scan data and the presence of environmental noise (such as dust and moving objects), voxel downsampling was first performed on each local point cloud group (voxel side length v=0.05m) to homogenize the point cloud density and significantly reduce computational load while preserving the building's geometric features. Subsequently, statistical outlier removal was performed: the average distance between each point and its K nearest neighbors was calculated, and points with a distance greater than a preset distance threshold were considered noise and removed based on the Gaussian distribution principle. The preset distance threshold can be set as the average distance plus n times the standard deviation. To balance denoising effectiveness with the preservation of component edge features, the standard deviation factor n is set to a preset empirical value of 2 or 3.

[0031] The third step is geometric feature calculation and descriptor extraction.

[0032] For the denoised local point cloud, its spatial geometric descriptive features are calculated for stitching. First, principal component analysis (PCA) of the local neighborhood is performed on the point cloud (with a neighborhood search radius of r=0.1m) to calculate the surface normal vector of each point. Based on this, the Fast Point Feature Histogram (FPFH) descriptor of each point is calculated (with a feature calculation radius set). =0.2m), this descriptor transforms the differences in normal vectors around a point into a high-dimensional histogram, which is used to uniquely characterize the local three-dimensional geometric structure of that point.

[0033] The fourth step is feature matching and rigid body transformation matrix calculation.

[0034] Taking two adjacent local point clouds as an example, a KD-tree data structure is constructed to perform efficient nearest neighbor search on the FPFH descriptors of the two point clouds. The Euclidean distance in the feature space is calculated, and a predetermined number of point pairs with the smallest distance (e.g., randomly selecting 2000 candidate pairs) are selected as the initial matching pairs. Since there are many similar structures in buildings (such as identical doors, windows, and columns), the initial matching often includes false matches. Therefore, the RANSAC (Random Sample Consensus) algorithm is introduced, and a distance tolerance threshold is set. =0.05m and a maximum number of iterations of 5000. In the iteration, incorrect matching pairs are eliminated, and finally the optimal rigid body transformation matrix that maximizes the number of interior points (correct matching points) is calculated, which includes the rotation matrix R and the translation vector q.

[0035] Step 5: Global coordinate system integration with the model.

[0036] Using the optimal rigid body transformation matrix output by the RANSAC algorithm, adjacent local point clouds are transformed one by one and mapped to a unified global coordinate system. The point cloud data of all stations are traversed sequentially to complete coordinate transformation and stitching fusion, eliminating redundant points in overlapping areas, and finally outputting the initial point cloud model of the existing building as a whole.

[0037] It should be noted that the voxel downsampling, statistical outlier removal, principal component analysis normal vector calculation, fast point feature histogram descriptor extraction, KD tree nearest neighbor search, and random sampling consensus algorithm involved in the above point cloud preprocessing process are all existing mature algorithms in the fields of computer vision and 3D point cloud processing. They are only used as basic technical means to acquire and process initial point cloud data and are not within the scope of protection of this invention. Therefore, they will not be described in detail here.

[0038] S2, the initial point cloud model is segmented to obtain individual point cloud blocks; based on the geometric shape features and physical reflection features of each point cloud block, the semantic classification label of each point cloud block is obtained, and the point cloud blocks are merged according to the semantic classification label to obtain several entity components.

[0039] The initial point cloud model is essentially an unordered and continuous set of three-dimensional coordinate points, lacking component boundary information. Since it's impossible to directly identify building components from the overall data, this embodiment divides the overall point cloud into small, continuous regions to extract local point cloud features. Using point cloud blocks as the basic computational unit, local flatness, cylindricity, and reflection intensity can be calculated independently, thus providing a basis for subsequent semantic labeling.

[0040] In this embodiment, a region growing algorithm can be used to divide the initial point cloud model into multiple small, continuous point cloud blocks. Each point cloud block may correspond to a component or a part of a component.

[0041] Basic point cloud segmentation algorithms are prone to oversegmentation, and are susceptible to occlusion or uneven point cloud density, potentially dividing the same building component into multiple fragmented point cloud blocks. Relying solely on spatial distance is insufficient to accurately determine whether different point cloud blocks belong to the same building component. This embodiment introduces semantic classification labels as constraints, reassembling spatially adjacent point cloud blocks with consistent labels (e.g., both made of concrete and columnar in shape). This eliminates oversegmentation errors, restoring fragmented blocks to complete solid components. It is worth noting that obtaining complete solid components is a prerequisite for extracting the component's true center coordinates and actual dimensions, and is also the foundation for subsequently establishing building topology rules and calculating structural deformation.

[0042] As an exemplary implementation, semantic classification labels for each point cloud block are obtained based on their geometric shape and physical reflection features. Then, point cloud blocks are merged based on these semantic classification labels to obtain several entity components, including:

[0043] The first step is to fit a surface model of multiple geometric primitives to each point cloud block. By calculating the residual characteristics of the data points deviating from the surface model, the component geometry type of the point cloud block is determined.

[0044] First, a planar model is fitted to each point cloud block to obtain a fitted plane. The root mean square error of the distance from the data points in the point cloud block to the fitted plane is calculated as the flatness residual.

[0045] In this embodiment, for each point cloud block, the least squares method is first used to perform plane fitting on the point cloud block to obtain a fitted plane. Then, the distance from each data point in the point cloud block to the fitted plane is calculated. Subsequently, the root mean square error is calculated for all distances corresponding to the obtained point cloud block. The obtained root mean square error is used as the flatness residual of the corresponding point cloud block, denoted as . .

[0046] Secondly, a cylindrical model is fitted to each point cloud block to determine the cylindricity residual of the point cloud block.

[0047] In this embodiment, for each point cloud block, the least squares method can be used to fit the 3D data points in the point cloud block into an optimal cylindrical model, thereby determining the spatial position and orientation of the central axis of the cylindrical model, as well as the radius of the cylinder, denoted as the fitting radius. Then, the vertical distance from each data point in the point cloud block to the central axis is calculated. Next, the absolute value of the difference between each vertical distance and the fitting radius is calculated, reflecting the degree of deviation of a single data point from the ideal cylindrical surface. Finally, the average of the absolute values ​​of the differences corresponding to all data points in the current point cloud block is calculated, and this average value is taken as the cylindricity residual of the point cloud block, denoted as . .

[0048] Finally, the component geometry type of the point cloud block is determined by comparing the flatness residual and cylindricity residual of the point cloud block with their corresponding preset thresholds.

[0049] In this embodiment, a preset flatness threshold is set. and preset cylindricity threshold Taking any point cloud patch as an example, the calculated... and Each value is compared with its corresponding threshold, and classification is performed by combining directional features:

[0050] when and At that time, the normal vector of the fitted plane of the point cloud block is extracted. If the angle between the normal vector and the vertical Z-axis is less than the preset angle threshold of 15° or greater than 165°, the geometric type of the component is determined to be a plate-shaped component; if the angle is in the range of 75° to 105°, the geometric type is determined to be a wall-shaped component.

[0051] when and When the centerline direction vector of the fitted cylinder is extracted, if the angle between this vector and the vertical Z-axis is less than a preset angle threshold of 15° or greater than 165°, the geometric type of the component is determined to be a columnar component; if the angle is within the range of 75° to 105°, the geometric type is determined to be a beam-like component.

[0052] when and At that time, compared and The numerical value is used to initially classify the point cloud blocks into a category with smaller residual values, and further determine them as slabs, walls, columns or beams according to the angle conditions of the normal vector or the central axis vector mentioned above.

[0053] when and When determining that the component's geometric type is a complex component, subsequent topological rule construction steps are eliminated.

[0054] It is worth noting that plate-like members and wall-like members are classified as surface-like members, while column-like members and beam-like members are classified as linear-like members.

[0055] For the preset flatness threshold and preset cylindricity threshold Because the nominal ranging accuracy of a 3D laser scanner is superimposed on the natural roughness of the building component surface, in order to avoid misjudging normal scanning noise and surface texture as non-geometric surfaces, the threshold must be greater than the sum of the device's ranging error and the surface roughness. Based on this, the empirical range of the preset flatness threshold in this embodiment can be 5mm to 15mm, and the empirical range of the preset cylindricity threshold is 10mm to 15mm.

[0056] The second step is to statistically analyze the distribution characteristics of the reflection intensity of all data points in each point cloud block and compare them with a pre-set material reflection attribute library to determine the component material properties of the point cloud block.

[0057] Taking any point cloud block as an example, the reflection intensity values ​​of all data points in the point cloud block are extracted, and the mean and variance of the reflection intensity are calculated as the distribution characteristics of the point cloud block. A pre-built material reflection attribute library is constructed, which records the standard reflection intensity mean and variance of various known building materials. For example, the standard characteristics of concrete are set as mean 120 and variance 20; the standard characteristics of steel are set as mean 155 and variance 10; and the standard characteristics of wood are set as mean 90 and variance 40. The calculated mean and variance are compared with the standard characteristics of each material in the attribute library. The specific method is as follows:

[0058] Calculate the weighted Euclidean distance between the distribution characteristics of the point cloud block and the standard characteristics of each material in the attribute library. For the preset weights, the mean is set to dominate; for example, the weights of the mean and variance are empirically set to 0.7 and 0.3, respectively. Select the material corresponding to the minimum weighted Euclidean distance as the component material attribute of the point cloud block. If the minimum distance obtained from comparing all materials is greater than the preset feature distance threshold, the material attribute of the point cloud block is marked as "unknown material" and will not participate in subsequent material-based specification verification.

[0059] The process begins by calculating the weighted Euclidean distance between any two standard features of materials in a pre-set material reflectance attribute library, and extracting the minimum distance. Then, one-third to one-half of this minimum distance is used as the baseline recognition radius to ensure that the classification spaces of different materials do not overlap. Finally, the baseline recognition radius is added to the laser scanner's preset reflectance intensity drift error to obtain the final feature distance threshold. Based on the set standard features and weights (mean weight 0.7, variance weight 0.3), the closest materials in the library are concrete and wood, with a weighted Euclidean distance of approximately 27. One-third of this distance, or 9, is taken as the baseline recognition radius. Combined with the empirical value of 2 for the inherent drift error of general scanner reflectance intensity, the empirical value of the feature distance threshold can be set to 11.

[0060] The third step is to combine the component geometry type and component material properties of the same point cloud block to obtain the semantic classification label of each point cloud block.

[0061] In this embodiment, the component material properties and component geometry types of the same point cloud block are concatenated and combined in a text format of "material property-geometry type" to generate a semantic classification label for the point cloud block. For example, if the material property of the point cloud block is "concrete" and the geometry type is "column component", the generated semantic classification label is "concrete-column component"; if the material property is "steel" and the geometry type is "beam component", the label is "steel-beam component".

[0062] It is particularly important to note that if the material property of a point cloud block is "unknown material" or the geometry type is "complex component," the generated semantic classification label will be recorded as "invalid label." Point cloud blocks with invalid labels will be directly filtered out and will not participate in the subsequent construction of the adjacency matrix and the mining of frequent itemsets, in order to prevent abnormal components from interfering with the extraction of building topology patterns.

[0063] The fourth step is to merge point cloud blocks according to semantic classification labels to obtain several entity components.

[0064] The first sub-step is to obtain the spatial bounding box of each point cloud block.

[0065] In this embodiment, the maximum and minimum coordinates of all data points in each point cloud block on the X, Y, and Z axes are calculated, and a three-dimensional axial bounding box for the corresponding point cloud block is constructed as a spatial bounding box.

[0066] The second sub-step involves merging multiple point cloud blocks with the same semantic classification label based on their component geometry type.

[0067] When the component geometry is linear, the fitted geometric centerline of each point cloud block is extracted, and point cloud blocks whose fitted geometric centerlines meet the preset collinearity condition and whose spatial bounding box spacing meets the preset linear connection threshold are merged and combined into a solid component.

[0068] In this embodiment, the direction vector of the fitted cylinder's central axis and the coordinates of the center point of two point cloud blocks are extracted. The angle between the direction vectors of the two point cloud blocks is calculated, and the perpendicular distance from one center point to the other central axis is calculated. If the angle is less than a preset axis angle threshold and the perpendicular distance is less than a preset coaxial distance threshold, the preset collinearity condition is satisfied. The shortest spatial distance between the two spatial bounding boxes is calculated. If the preset collinearity condition is satisfied and the shortest spatial distance is less than a preset linear connection threshold, the two point cloud blocks are stitched together into a single solid component.

[0069] Among them, the preset axis angle threshold is used to tolerate point cloud fitting errors, and the empirical value is 5 degrees; the preset coaxial distance threshold is used to prevent the side-by-side independent columns and beams from being mistakenly merged, and it needs to be much smaller than the cross-sectional width of conventional components, and the empirical value is 0.1 meters; the preset linear connection threshold is used to cross point cloud break zones caused by obstruction by debris, and the empirical value is 0.5 meters.

[0070] When the component geometry is planar, the fitting plane normal vector of each point cloud block is extracted, and point cloud blocks whose fitting plane normal vectors are parallel, whose normal distance meets the preset coplanar condition, and whose spatial bounding box spacing meets the preset planar connection threshold are merged and combined into a solid component.

[0071] In this embodiment, the fitting plane normal vectors and center point coordinates of two point cloud blocks are extracted. The angle between the normal vectors of the two point cloud blocks is calculated, and the normal distance from one center point to the other fitting plane is calculated. If the angle is less than a preset normal angle threshold and the normal distance is less than a preset normal distance threshold, the preset coplanarity condition is satisfied. The shortest spatial distance between the two bounding boxes is calculated. If the preset coplanarity condition is satisfied and the shortest spatial distance is less than a preset planar connection threshold, the two point cloud blocks are stitched together into a single solid component.

[0072] Among them, the preset normal angle threshold is empirically set at 5 degrees; the preset normal distance threshold is used to prevent parallel multi-story slabs or parallel walls of different rooms from being mistakenly merged, and should be much smaller than the floor height or room span, with an empirical value of 0.1 meters; since the wall panel surface area is large, the breakage blank caused by furniture obstruction is usually larger than that of linear components, and the preset surface connection threshold is empirically set at 1.0 meters.

[0073] S3, based on the semantic classification labels of entity components and the adjacency matrix constructed from the spatial distribution of each entity component, uses association rule algorithms to mine frequent itemsets of component arrangements to generate building topological patterns.

[0074] The spatial connections between building components (such as beam-column joints and slab-beam overlaps) are not random combinations, but rather exhibit clear repetition, symmetry, and regularity under the control of structural logic. For example, the column spacing on the same axis remains equal, and the top of the beam and the bottom of the slab are at the same elevation. In the point cloud data dimension, this physical connection regularity manifests as a high-frequency co-occurrence pattern of component semantic labels and their spatial adjacency relationships.

[0075] To extract and quantify the aforementioned patterns, a transaction mapping mechanism is introduced to establish analysis samples: each entity component identified in the point cloud is transformed into an independent analysis transaction. The transaction content consists of the semantic classification label of the entity component itself, and the spatial orientation and semantic labels of adjacent entity components extracted by reading the component adjacency matrix. Based on this sample, an association rule algorithm is used to calculate a combination of support and confidence for a specific itemset. When both the support and confidence of an itemset exceed a preset threshold, the high-frequency data association pattern is confirmed as an objectively existing structural topological pattern in the building. Subsequently, based on such summarized topological patterns, when some components are completely missing due to occlusion or scanning blind spots in the point cloud, this can be used as a basis for inferring missing components. Among them, support is used to filter out component adjacency combination patterns widely distributed in the overall building, filtering out accidental fits caused by clutter or occlusion by non-structural components; confidence assesses the probability of the condition occurring and determines the inevitable tendency of subordinate connections.

[0076] As an example, the process of obtaining architectural topological patterns, such as Figure 2 As shown, it includes:

[0077] S31. Calculate the distance between the center points of any two solid components and the angle between the normals of the contact surfaces of the spatial bounding boxes. If the preset distance and preset angle conditions are met, it is determined that there is a spatial adjacency relationship between the two solid components, and a binary adjacency matrix is ​​constructed based on the spatial adjacency relationship between each solid component.

[0078] In building structures, components that transfer loads or are physically connected to each other are spatially adjacent and their contact surfaces are opposite. By extracting distance features to eliminate irrelevant components that are too far apart physically, and by extracting the normal direction features of the contact surfaces to eliminate non-load-bearing components that are close in distance but are stacked in parallel, the true physical connection relationships can be accurately identified.

[0079] In this embodiment, the three-dimensional Euclidean distance between the center points of the bounding boxes of any two solid components is calculated. The normal vectors of the closest contact surfaces of the spatial bounding boxes of the two solid components are calculated. It is determined whether the distance between the center points is less than a preset distance condition; the dot product of the two normal vectors is calculated, and it is determined whether the included angle satisfies a preset angle condition. The two solid components are determined to be adjacent if and only if both the distance and angle conditions are satisfied. An N×N square matrix is ​​constructed (N is the total number of solid components), and the values ​​of the corresponding row and column coordinates in the matrix are marked as 1, and the remaining non-adjacent values ​​are marked as 0, generating a binary adjacency matrix.

[0080] The preset distance condition is used to capture real physical contact. Since the distance between the center points of physical components is affected by the cross-sectional dimensions of the components themselves (such as column width and beam height), this threshold cannot be set to 0. The value needs to be slightly larger than half of the cross-sectional dimensions of conventional building components to accommodate the shrinkage of the bounding box caused by dimensional fluctuations and missing point clouds, while it should be smaller than the standard room span to avoid misjudging opposing walls as adjacent. The distance threshold is set to 0.5 meters.

[0081] For preset angle conditions, such as the included angle of the normal. Degree, i.e., dot product Used to confirm the alignment of the stress interface. The normal vectors of the contact surfaces of actual socketed or lapped components should be strictly opposite (i.e., 180 degrees). Considering that actual construction may involve misalignment and local missing points in the point cloud leading to deviations in normal vector calculation, the threshold is relaxed to 150 degrees to distinguish between non-structural contacts that are perpendicularly intersecting or parallelly stacked.

[0082] S32. For two solid components that are spatially adjacent, calculate the projection components of the center point coordinate difference vector on each coordinate axis, take the coordinate axis where the projection component with the largest absolute value is located as the main direction axis, and determine the three-dimensional relative direction between the two solid components according to the sign of the projection component with the largest absolute value.

[0083] The building topology possesses orthogonal anisotropy (e.g., floor slabs in the horizontal plane, columns in the vertical plane). The spatial vector formed by the center points of two components, and the length of its projection onto the three orthogonal coordinate axes X, Y, and Z, directly reflects the dominant direction of component assembly. Selecting the projection component with the largest absolute value can filter out installation misalignment deviations in secondary axes, and its sign accurately points to the absolute orientation in orthogonal space.

[0084] In this embodiment, the coordinates of the center points of two spatially adjacent solid components are extracted, and the coordinates of the source component are subtracted from the coordinates of the target component to obtain a three-dimensional coordinate difference vector. , , Take them separately. , , Find the component with the largest absolute value by comparing the absolute values ​​of the components. Maximum, with the Z-axis as the principal direction; when At that time, the relative direction is determined as "upper adjacent". It is "below adjacent". If Maximum, with the main direction axis being the X-axis; when When the relative direction is determined, it is defined as "right adjacent". This is "left-side adjacent". If... Maximum, with the Y-axis as the principal direction; when Δ At that time, the relative direction is determined as "front adjacent". This is for "back adjacency".

[0085] S33, for each entity component, combine the semantic classification label of the entity component, the semantic classification labels of other entity components that have spatial adjacency with the entity component as determined by the adjacency matrix, and the corresponding three-dimensional relative directions to generate transaction data.

[0086] Association rule mining requires standardized transaction flow input. By flattening the three-dimensional spatial network of the structure and encoding it into a one-dimensional set of text labels, subsequent mining algorithms can count the occurrence frequency of each set and calculate support by character matching.

[0087] In this embodiment, for the i-th entity component, its own semantic classification label is used as the first item of the transaction, for example, extracted as "concrete-column component". The i-th row of the adjacency matrix is ​​traversed to find the j-th entity component corresponding to a value of 1. The semantic classification label of the j-th entity component is extracted and combined with the three-dimensional relative direction calculated in step S32 to form a relation item. For example, it is concatenated as "upper adjacency_concrete-beam component". All relation items of the i-th component are combined with its own semantic label to generate a transaction data, for example: {concrete-column component, upper adjacency_concrete-beam component, lower adjacency_concrete-slab component}. All components are traversed to generate a transaction database.

[0088] S34. Using the association rule algorithm, extract frequent itemsets that meet the preset support threshold from each generated transaction data, and calculate the confidence of the association rules generated from the frequent itemsets. Association rules with confidence greater than the preset confidence threshold are taken as building topology rules.

[0089] Architectural design follows standard atlases and modular paradigms, where real topological patterns are frequently reproduced globally (high support) and exhibit rigorous conditional derivation relationships (high confidence, e.g., every beam end must have a column). Data mining algorithms are used to eliminate low-frequency, accidental construction distortions or incomplete noise, solidifying the correlation features that satisfy dual statistical thresholds into standard architectural rules.

[0090] In this embodiment, based on the generated transaction database, the Apriori algorithm is used to count the occurrence frequency of all itemsets, and the support is calculated by dividing by the total number of transactions. A preset support threshold of 0.3 is set. Itemets with support less than 0.3 are filtered out, and frequently occurring component combinations are retained to form frequent itemsets. Frequent itemsets are split into antecedents (component attributes) and consequents (adjacency relationship attributes) to form association rules, and the confidence of the rules is calculated. The confidence calculation formula is: the number of transactions containing the entire frequent itemset, divided by the number of transactions containing only the antecedent. A preset confidence threshold of 0.85 is set. All association rules with a confidence greater than 0.85 are extracted and output, for example, "concrete-beam-like component → adjacent below_concrete-column-like component", which are then fixed and stored in the building topology rule library.

[0091] The preset support threshold is used to establish the global universality of the pattern. Due to the large amount of modular repetition in architectural design, the recurrence rate of real patterns is usually high. Setting it to 0.3 means that the connection pattern appears in at least 30% of similar components throughout the building. If the threshold is too low, localized accidental modifications or overlapping errors will be regarded as patterns; if it is too high, secondary load-bearing patterns on non-primary axes will be missed.

[0092] A pre-set confidence threshold is used to ensure the determinism of logical inference. Confidence reflects the probability that if component A exists, component B must also exist. Building load-bearing structures have strong correlations; for example, the presence of columns supporting beams is a strong certainty. Setting an extremely high threshold of 0.85 ensures that the extracted topological patterns have sufficiently high causal constraints, guaranteeing the rigor and extremely low misjudgment rate when performing virtual completion inferences based on these patterns.

[0093] S4 uses the building topology rules to verify several current physical components, locates the three-dimensional spatial region where topological conflicts occur, determines the occlusion probability by combining the three-dimensional local point cloud density of the three-dimensional spatial region, and generates virtual components in the three-dimensional spatial region according to the building topology rules when the occlusion probability meets the completion condition, so as to complete the initial point cloud model.

[0094] In 3D laser scanning of buildings, local point clouds of components are often not captured due to on-site clutter, ceiling obstructions, or instrument blind spots. Traditional methods struggle to automatically distinguish between data loss caused by environmental occlusion and actual physical defects or damage to components at the algorithmic level. To overcome this technical bottleneck, this step introduces a priori logic verification and density mapping mechanisms. First, the discovered building topology patterns are used as objective constraints to perform a forward structural consistency check on the currently measured and extracted entity component network, accurately identifying local spaces that violate high-confidence association rules as topological conflict regions. Second, a quantitative evaluation model based on the density of reflection points in 3D space is established to address the causes of conflict regions. By comparing the point cloud density of the local space with the baseline density of the original structural surface, the probability of laser signals being truncated by environmental interference is quantified, and the occlusion probability is output.

[0095] As an exemplary implementation, the method utilizes architectural topology to verify several existing physical components, locates the three-dimensional spatial region where topological conflicts occur, and determines the occlusion probability by combining the three-dimensional local point cloud density of the three-dimensional spatial region, including:

[0096] The first step is to traverse each physical component and determine whether the actual spatial adjacency relationship of each physical component satisfies the building topology rules.

[0097] A forward structural consistency check is performed on the measured and extracted three-dimensional component network to verify whether the physical connection state around the physical component conforms to the summarized building standard mechanics and construction system logic.

[0098] In this embodiment, the semantic classification label of each entity component is extracted. All association rules using this semantic classification label as the "antecedent" are retrieved from the building topology rule library. The binary adjacency matrix records corresponding to the target entity component are traversed to check whether they fully contain the consequent attributes specified in the association rules, i.e., the semantics of adjacent components and their three-dimensional relative orientation. Secondly, if the actual spatial adjacency relationship of the target entity component lacks the association rules indicated in the building topology rules, it is determined that a expected component is missing.

[0099] The topological laws of buildings have inherent correlations at the structural mechanics level (e.g., beam ends must have supports). When a strongly correlated logical breakpoint appears in the measured topological network, the system determines that it is due to an explicit gap in the physical location of the expected component caused by incomplete data acquisition.

[0100] In this embodiment, if a strong association rule is found (e.g., the rule is: "concrete-beam-shaped component" has "below adjacent_concrete-column-shaped component"), and the target entity component does not match the corresponding component semantic label in the lower region identified in the adjacency matrix, then it is determined that the target entity component has a missing expected component in the indicated direction.

[0101] The third step is to extract the three-dimensional relative directions indicated in the association rules and the corresponding average spacing of similar components. The center coordinates of the target entity component are translated along the three-dimensional relative directions by the average spacing to obtain the expected center point. A bounding box of a preset size is then constructed using the expected center point as the three-dimensional spatial region where topological conflicts occur.

[0102] Based on the modular design principle that the spacing between similar components on the same axis is equal, the known spacing of the surrounding similar components is used as a span reference. Coordinate translation is performed along the associated vector to accurately locate the absolute spatial range where the missing entity should have been.

[0103] In this embodiment, the three-dimensional relative direction vector is extracted from the association rules that trigger the missing item alarm. In the extracted samples of similar patterns without missing items, the average spacing L of such components is calculated. The center coordinates (X, Y, Z) of the target entity component are translated a distance L along the three-dimensional relative direction vector on the spatial projection axis to calculate the expected center point. Using this expected center point as the geometric center, a three-dimensional cube bounding box with a spatial cross-sectional size of 1m × 1m × 1m is generated, and this bounding box is defined as the three-dimensional spatial region where topological conflicts occur.

[0104] The fourth step is to count the number of point clouds in the three-dimensional spatial region as the three-dimensional local point cloud density; select multiple reference entity components that have not experienced topological conflicts from each entity component, and count the average number of point clouds on the surface of all reference entity components within the same preset size as the reference point cloud density of the overall building surface.

[0105] To eliminate the influence of objective variables such as scanner resolution and measurement distance attenuation, a relative evaluation system for point cloud density needs to be established. This involves calibrating the device's baseline reflectivity on a known, intact native component surface, and then measuring the residual point cloud in areas of topological conflict, thus shielding measurement errors caused by the hardware itself.

[0106] In this embodiment, the total number of all 3D point cloud data points within the aforementioned 1m×1m×1m topological conflict bounding box is obtained and denoted as the 3D local point cloud density. (Unit: points / cubic meter). In the current initial point cloud model, 50 reference solid components that do not trigger any topological conflicts and have complete geometric features are randomly selected. A 1-square-meter mapping space is segmented on the surface of each reference solid component, and the number of point clouds within it is counted. The arithmetic mean of the point cloud counts is calculated and denoted as the reference point cloud density of the overall building surface. (To meet the ratio calculation requirements, this item is equivalent to the unit standard space capacity limit during data calculation). To prevent subsequent division overflow anomalies, set... The preset minimum lower limit value is 0.01.

[0107] For reference solid components with small surface dimensions that cannot meet the 1 square meter sampling condition, the ratio of the area of ​​the actual extracted surface to the 1 square meter calibration area is first calculated. Then, the number of locally measured point clouds is divided by this ratio to perform linear scaling, and the scaled values ​​are uniformly used in the summary calculation of the average number of point clouds.

[0108] The fifth step is to calculate the occlusion probability of the expected missing components based on the difference between the 3D local point cloud density and the reference point cloud density of the overall building surface.

[0109] The deviation of the ratio between the 3D local point cloud density and the overall surface reference density reflects the visibility status of the original components in the target space. When the ratio approaches 0, it indicates a lack of effective reflection, with the laser completely blocked by obstacles in front. When the ratio is much greater than 1, it indicates the presence of abnormally dense unstructured reflective objects (such as illegally piled high-density debris), meaning the original components are essentially also obscured. Only when the ratio approaches 1 does it represent a normal structural surface characteristic, truly indicating a physical absence under conditions of visual transparency. Therefore, by constructing a bidirectional threshold mapping formula based on absolute deviation, the abnormal deviation of density is quantified into a probability value indicating that physical obscuration is valid.

[0110] In this embodiment, the ratio of local density to reference density is calculated, and then the occlusion probability mapping formula is constructed:

[0111] In the formula, This represents the occlusion probability of a missing component within a topological conflict region; This represents the spatial point cloud density within a topological conflict region; represents the overall average point cloud density of the identified entity surface; min represents the minimum value function; 0.4 is set as the basic skepticism base when the system experiences topological conflicts.

[0112] This formula calculates the density offset using absolute values: when the local point cloud density is extremely low (i.e., in the scanning blind zone), The ratio is close to 0 or extremely high (e.g., when obstructed by dense stockpiles of miscellaneous materials). When ), the absolute value deviation term If the value is equal to or greater than 1, the maximum threshold of 1 is truncated by the min function, and a post-multiplication penalty term is taken to be 0.6, representing the occlusion probability. The calculation result reaches the upper limit of 1.0, thus determining that the completion condition is met and triggering the subsequent repair rule; only when the local density is close to the baseline density (i.e., the ratio is near 1), the absolute value deviation term approaches 0, and the penalty term becomes ineffective. The result fell back to the lower limit of 0.4. This indicates that there is normal density point cloud reflection within this 3D spatial region, and the physical components objectively exist rather than being missing. At this point, the system directly blocks the virtual data completion operation.

[0113] Furthermore, the calculated occlusion probability is compared with a preset probability threshold to determine whether there are missing components in the three-dimensional spatial region, i.e., whether the occlusion probability meets the completion condition.

[0114] Specifically, when the occlusion probability is greater than a preset probability threshold, the spatial point cloud data within the view frustum constructed by connecting the scanning station to the three-dimensional spatial region is further extracted; if there are discrete ray obstructing point clouds in the view frustum other than structural point clouds, it is determined that the data anomaly in the three-dimensional spatial region is caused by data loss due to physical occlusion, and the completion condition is met; if there are no obstructing point clouds in the view frustum, it is determined that the three-dimensional spatial region has experienced actual breakage, damage or loss, and the completion operation is refused and a structural anomaly alarm is marked.

[0115] In this embodiment, a preset probability threshold is set (an empirical value of 0.7 or 0.8), and the occlusion probability output by each topological conflict region is compared with this threshold one by one. If the occlusion probability is less than or equal to the preset probability threshold, it indicates that the point cloud reflection density in the region is normal, the component objectively exists, and the system refuses to perform the completion operation; if the occlusion probability is greater than the preset probability threshold, it indicates that there is an abnormal point cloud reflection (i.e., missing point cloud) in the topological conflict region. To distinguish between physical occlusion and actual damage, the system further extracts the spatial point cloud data within the view frustum constructed by connecting the scanning station to the three-dimensional spatial region. If there are discrete ray obstructing point clouds (i.e., debris point clouds) in the view frustum other than structural point clouds, it is determined that the data anomaly in the region is caused by data loss due to physical occlusion, the system meets the completion conditions, and enters the subsequent virtual reconstruction process; if there are no obstructing point clouds in the view frustum, it indicates that the line of sight ahead is clear but the point cloud reflection of the corresponding regional component is still not obtained, it is determined that the expected physical component has suffered actual breakage, damage, or loss, the system also refuses to perform the completion operation, and directly marks the spatial region location with a structural anomaly alarm. The preset probability threshold can be set by the implementer according to specific actual requirements, and no specific limitation is made here.

[0116] After summarizing the topological patterns of the building structure and assessing the occlusion probability of topological conflict areas, the structural topological patterns are further used as a priori expectations of component existence, with occlusion probability serving as the criterion. When the occlusion probability meets preset conditions, it is determined that a component in that spatial region is truly missing rather than occluded. Then, based on the topological patterns, the key attributes of the missing component are deduced in reverse, including semantic classification labels, spatial location, and geometric dimensions. Finally, virtual components conforming to mechanical construction logic and modular distribution characteristics are constructed in three-dimensional space and incorporated into the original point cloud model to achieve pattern-based missing component repair.

[0117] As an exemplary implementation, when the occlusion probability meets the completion condition, virtual components are generated in a three-dimensional spatial region according to the building topology to complete the initial point cloud model, including:

[0118] The first step is to extract the semantic classification label of the expected component from the association rules of the missing target entity component; and to determine the expected center point corresponding to the three-dimensional spatial region as the center coordinate of the virtual component.

[0119] In this embodiment, the frequent itemsets and high-confidence association rule base constructed during the topology pattern mining stage are queried. Using the semantic classification labels and three-dimensional relative directions of known components adjacent to the topological conflict region as antecedents, matching association rules are retrieved, and the semantic attributes of the consequent of these rules are extracted as the semantic classification labels of the missing expected components. For example, if the known component is "concrete-beam-like component," the missing orientation is "above," and the rule base contains a high-frequency rule "concrete-beam-like component-above adjacent → concrete-slab-like component," then "concrete-slab-like component" is directly extracted as the semantic classification label of the expected component. Simultaneously, the expected center point obtained by axial translation when the topological conflict occurs is directly retrieved and used as the absolute center coordinates of the virtual component to be generated in the global coordinate system.

[0120] The second step is to extract entity components from the initial point cloud model that have the same semantic classification labels as the expected components, and calculate their average length, width, and height dimensions as the three-dimensional geometric dimensions of the virtual components.

[0121] Load-bearing components with the same semantic classification label in the same building all follow modular design requirements in terms of cross-sectional dimensions and length, exhibiting a high degree of statistical consistency. Using actual components with successfully acquired features and identical labels in the initial point cloud model as a reference, their average size is the optimal estimate of the geometry at the missing location.

[0122] In this embodiment, the component set of the initial point cloud model is traversed, and all valid entity components that are completely consistent with the extracted expected component semantic classification labels are selected. The length, width, and height of the bounding box are extracted for each of the selected component groups. The average value of each dimension of this sample set is calculated, and the obtained average length, width, and height are directly assigned to the virtual component to be generated as its three-dimensional geometric dimensions. If the number of samples extracted for the same label is lower than a preset judgment threshold (e.g., less than 3), it indicates insufficient on-site data support, and the theoretical design dimensions of the corresponding type of component in the architectural standard modular drawing set are directly retrieved through semantic indexing for replacement.

[0123] The third step is to construct the bounding box of the virtual component based on its center coordinates, three-dimensional geometric dimensions, and semantic classification labels, thereby generating the virtual component.

[0124] In this embodiment, the extracted expected center point coordinates are used as the positioning reference, and the calculated three-dimensional geometric dimensions are used as the spatial boundary. Combined with semantic classification labels, parameters are assembled to parametrically construct a standard orthogonal spatial bounding box in a three-dimensional coordinate system. This bounding box fully defines the position, volume, and attribute affiliation of the missing component within the building network, thus completing the physical generation of the virtual component.

[0125] The fourth step is to incorporate the virtual components and their semantic classification labels into the component set of the initial point cloud model to complete the model completion.

[0126] In this embodiment, the generated virtual component, along with its associated semantic classification label, is injected into the global component set of the initial point cloud model, and the global component spatial index structure and binary adjacency matrix of the model are updated simultaneously. After the update loop is closed, the virtual component is given the same node status as the physical component extracted in the initial scan. The spatial data discontinuities and force information gaps caused by environmental occlusion in the original point cloud model are perfectly filled, and the entire system completes automated model repair and completion.

[0127] S5: Obtain deformation reference benchmark, calculate the deformation of each component based on the completed current point cloud model, filter construction deviations by combining preset tolerance threshold, extract the newly added structural deformation, and conduct structural safety trend assessment by combining the semantic classification labels of the corresponding components.

[0128] Based on the selected deformation reference type, differentiated deformation extraction and deviation filtering mechanisms are established. When using a historical scan point cloud model as the reference, since the historical model naturally incorporates the initial deviations of the construction phase, the current point cloud model after completion is precisely registered with it, and displacement and tilt angle changes are calculated for each component. The resulting deformation physically represents the new structural deformation between two scan cycles, without the need for additional deduction of construction tolerances. When using architectural structural design drawings as the reference, the drawings only represent the ideal design state and do not contain any construction errors. Therefore, a preset tolerance threshold from the national construction acceptance specifications needs to be introduced as a quantitative filtering boundary. By calculating the absolute difference between the actual spatial distribution parameters and the theoretical design parameters of each component, and comparing it with the tolerance threshold of the corresponding component semantic type, only the portion exceeding the tolerance threshold is identified as new structural deformation, while differences within the threshold are identified as construction-related deviations and filtered out.

[0129] Both of these approaches ultimately correlate the deformation to the identified and inferred components, including both physical and virtual components, providing accurate and unbiased deformation input data for subsequent structural safety assessments.

[0130] As an exemplary implementation, a deformation reference benchmark is obtained, the deformation of each component is calculated based on the completed current point cloud model, and the construction deviations are filtered out by combining a preset tolerance threshold to extract the newly added structural deformation.

[0131] Scenario 1: When the deformation reference is a historical scan point cloud model, extract the newly added structural deformation, including:

[0132] The first step is to spatially register and align the completed current point cloud model with the deformation reference datum to obtain the spatial registration and alignment result.

[0133] The accuracy of deformation calculation depends on the fact that the current point cloud data and historical point cloud data are in the same spatial coordinate system.

[0134] In this embodiment, preset stable control points or point cloud subsets corresponding to reference components identified as underlying bedrock or non-deformable areas are extracted from the current point cloud model and the historical scan point cloud model, respectively. Based on the extracted point cloud subsets from the current and historical scan point cloud models, the optimal rigid body transformation matrix (rotation matrix R and translation vector q) is calculated. The objective is to minimize the overall difference after registration of the extracted point cloud subsets. Since the point cloud subsets themselves are not deformed, the obtained transformation matrix only reflects the overall rigid displacement of the building between the two scan acquisition time points. The obtained transformation matrix is ​​applied to the entire current point cloud model. The non-deformable area is thus perfectly aligned with the historical scan point cloud model, while the relative displacement of the components in the deformable area is preserved in the same coordinate system, providing an accurate basis for subsequent component-by-component deformation calculation.

[0135] The second step is to match the point cloud regions with the same semantic classification labels and the smallest Euclidean distance between their center points in the deformation reference datum as historical corresponding components based on the spatial registration and alignment results. The coordinate displacement change and spatial attitude vector between the physical component and the historical corresponding component are calculated to obtain the deformation of the physical component.

[0136] After coordinate system alignment, the spatial position difference of the same component in the current point cloud model and the historical scanned point cloud model directly reflects its deformation. For solid components, the corresponding point cloud region can be located in the historical point cloud through the dual constraints of semantic tags and spatial proximity. By performing a geometric comparison between the two, the coordinate displacement and normal vector angle changes can be obtained. The former reflects the translation and settlement of the component, while the latter reflects the tilt deformation.

[0137] In this embodiment, each extracted entity component in the completed point cloud model is traversed. Taking the i-th entity component as an example, its semantic classification label and center coordinates are obtained. Point cloud regions with the same semantic classification label and the smallest Euclidean distance between their center points are retrieved from the historical scanned point cloud model and identified as the corresponding historical component. The coordinate displacement changes of the i-th entity component and its corresponding historical component in the X, Y, and Z directions are calculated respectively. And the angular change of the spatial attitude vector (normal vector or central axis direction vector). The deformation of the i-th solid component is obtained by combining the results. .

[0138] The third step is to obtain the deformation of multiple adjacent physical components that are spatially adjacent to the virtual component; calculate the spatial distance from the center coordinates of the virtual component to the center coordinates of each adjacent physical component, and use the spatial distance and deformation of each adjacent physical component to determine the deformation of the virtual component.

[0139] The generated virtual components have no corresponding entities in the historical point cloud, making it impossible to directly calculate deformation through point cloud comparison. However, based on the mechanical continuity of the building structure, the deformation of the virtual component's spatial location is correlated with the deformation of its adjacent components. Therefore, the deformation of the virtual component can be estimated by using spatial distance-weighted interpolation based on the known deformation of adjacent actual components; the closer the adjacent components are, the greater their contribution to the interpolation result.

[0140] In this embodiment, taking any virtual component as an example, the deformation amounts of multiple adjacent physical components that are spatially adjacent to the virtual component are obtained; the spatial Euclidean distances from the center coordinates of the virtual component to the center coordinates of each adjacent physical component are calculated; and the deformation amount of the virtual component is calculated by weighting the deformation amounts of each adjacent physical component using an inverse distance weighted average method. Specifically, the reciprocal of the distance from the center of the virtual component to the center of each adjacent physical component is calculated, and the reciprocal of the distance to a certain adjacent physical component is divided by the sum of the reciprocals of the distances to all adjacent physical components, which serves as the weighting weight for the deformation amount of the corresponding adjacent physical component.

[0141] The fourth step is to use the physical and virtual components as target components and determine the resulting deformation as the new structural deformation.

[0142] Since the historical scan point cloud model serves as a historical record of post-construction or early-stage scanning, the spatial positions of components already include construction deviations generated during the construction phase. Therefore, the component deformation directly calculated based on the registration difference of the historical scan point cloud model is physically equivalent to the net increase in deformation from the historical moment to the current moment, and does not need to be compared and deducted from the construction tolerance.

[0143] In this embodiment, all physical and virtual components involved in the calculation are uniformly defined as target components, and their corresponding deformation amounts are directly used as the new structural deformation amounts of the target components, and uniformly output to the subsequent structural safety assessment stage.

[0144] Scenario 2: When the deformation reference is the architectural structural design drawings, extract the newly added structural deformation, including:

[0145] The first step is to obtain the theoretical center three-dimensional coordinates and principal axis theoretical normal vectors of each component from the architectural structural design drawings, and then use semantic classification tags to match the corresponding preset position tolerance thresholds and angle tolerance thresholds in the preset specification database.

[0146] Since components of different semantic types (such as columns, beams, and slabs) have different allowable deviation values ​​in the construction and acceptance specifications, this embodiment directly extracts the ideal three-dimensional coordinate points and normal orientation of each component in the design drawings as absolute theoretical benchmarks and binds them with the corresponding specification tolerances, providing accurate judgment limits for subsequent three-dimensional deviation calculations at the individual unit level.

[0147] The second step is to align the completed point cloud model with the architectural structural design drawings in a global spatial coordinate system, and calculate the independent translational absolute deviation of the actual center three-dimensional coordinates and the theoretical center three-dimensional coordinates of each component in each axis, as well as the difference in the normal angle between the actual principal axis normal vector and the principal axis theoretical normal vector.

[0148] To overcome the limitation of macroscopic scalar ranging in decoupling single-axis displacement, this embodiment performs global alignment between the model and the drawing based on a non-deformable reference region. Based on this, the actual center 3D coordinates and actual principal axis normal vectors of each component in the current point cloud model are extracted. For each component, the independent absolute translational deviations between its actual center 3D coordinates and theoretical center 3D coordinates in the X, Y, and Z axes are calculated, and the difference in the normal angle between the actual principal axis normal vector and the theoretical principal axis normal vector is also calculated.

[0149] First, for the independent translational absolute deviation: calculate the difference between the actual center coordinate value and the theoretical center coordinate value in the X, Y, and Z axes respectively, and take the absolute value as the independent translational absolute deviation of each axis; simultaneously extract the original algebraic sign of the difference between the actual center coordinate value and the theoretical center coordinate value in each axis, and record it as the original displacement positive and negative direction in each axis respectively.

[0150] Secondly, regarding the difference in the normal angle: calculate the dot product between the actual principal axis normal vector and the theoretical principal axis normal vector of the component, divide the dot product value by the product of the magnitudes of the two normal vectors, and obtain the cosine value of the spatial angle between the two vectors; calculate the true spatial angle using the inverse cosine function, and determine the absolute value of the true spatial angle as the difference in the normal angle. Simultaneously, calculate the cross product of the two normal vectors, extract the direction of the vector obtained from the cross product, and use it as the original deflection direction of the component.

[0151] The third step involves comparing the absolute deviation of independent translation along each axis with the preset position tolerance threshold for each component, and comparing the difference in normal angle with the angle tolerance threshold to determine the amount of new structural deformation for each target component.

[0152] The tolerance thresholds specified in national construction acceptance standards represent the maximum allowable deviation range during the construction phase. If the actual difference from the design falls within the tolerance threshold, it can be reasonably attributed to construction deviations left over from the construction process and should not be used as a basis for structural safety warnings. Only when the difference exceeds the tolerance threshold does the excess constitute additional displacement that cannot be explained by construction errors and can be determined as newly occurring harmful deformation during the service life.

[0153] In this embodiment, an OR logic judgment is performed on each component for item-by-item comparison. Specifically, the system checks whether the absolute translational deviation of the component in the X, Y, and Z axes is greater than the preset position tolerance threshold, and checks whether the difference in its normal angle is greater than the matched angle tolerance threshold. If the difference index in any of the above dimensions exceeds the limit, it means that the component has undergone actual deformation beyond the allowable construction range in at least one specific direction, and is judged as exceeding the difference limit, and the corresponding component is marked as the target component. Conversely, if the differences in all dimensions are within the tolerance limit, it is determined that the component only has a construction baseline deviation and is filtered out.

[0154] For entities identified as target components, a multi-dimensional feature vector extraction and assembly operation is performed: First, for the translation dimension, only for specific axes where the absolute deviation exceeds a preset position tolerance threshold, the difference between the absolute deviation and the tolerance threshold is calculated as the net deformation, and this net deformation is combined with the original displacement positive and negative directions recorded during calculation to reassign polarity. For the angle dimension, if the limit is exceeded, the difference after deducting the angle tolerance threshold is also combined with the original deflection direction. Second, for the remaining translation or angle dimensions of the target component that do not exceed the corresponding tolerance threshold, the system considers the small fluctuations in these specific directions to be residual construction errors, and forces their deformation components to be set to zero to achieve directional noise reduction. Finally, the directional net deformation obtained by the above combination is merged with the forced zeroing components, and assembled into a multi-dimensional feature vector according to a unified rule dimension, such as the three axes X, Y, and Z and the angle. The system determines this multi-dimensional feature vector as the new structural deformation of the target component, which serves as a precise and decoupled three-dimensional deformation input parameter and is passed to the subsequent structural safety assessment stage.

[0155] As an exemplary implementation, for any target component, a structural safety trend assessment is performed based on the newly added structural deformation and the corresponding semantic classification label of the component, including:

[0156] The first step is to obtain the geometric type and material properties of the target component from the semantic classification label, retrieve its corresponding dimensionless deformation limit and reference parameters from the pre-set design specification database, calculate the comprehensive modulus of the translation vector in the newly added structural deformation and divide it by the reference parameters to obtain the dimensionless offset conversion coefficient.

[0157] To avoid dimensional conflicts between displacements with absolute length units and dimensionless proportional limits in the specifications, data standardization is performed first. In this embodiment, the corresponding deformation limits (such as the inter-story drift angle limit for concrete frame columns and the deflection-to-span ratio limit for steel beams) and the corresponding original design story height dimensions or span reference parameters are retrieved. The comprehensive modulus of the translation vector in the newly added structural deformation of the target component is calculated, divided by the aforementioned reference parameters, and converted to obtain the dimensionless attribute offset conversion coefficient.

[0158] The second step is to subtract the deviation conversion coefficient from the deviation conversion coefficient recorded at the previous historical time point, divide the difference by the number of days between the two data collections, and calculate the current deformation rate.

[0159] In this embodiment, the historical deformation record of the component is acquired. The current deviation conversion coefficient is subtracted from the deviation conversion coefficient recorded at the previous historical time node, and then divided by the number of days between the two scan acquisitions to obtain the dimensionless current deformation rate of the target component.

[0160] The third step is to compare the component's offset conversion coefficient with the set deformation limit, and combine this with a trend comparison between the current deformation rate and the historical deformation rate to classify the risk level.

[0161] The component's offset conversion coefficient is used as the equivalent additional deformation, and a direct dimensional comparison is performed with the set deformation limit. The risk level is then determined based on the rate trend.

[0162] Low risk is defined as an offset conversion coefficient less than 50% of the limit, and the current deformation rate less than or equal to the historical deformation rate; medium risk is defined as an offset conversion coefficient greater than or equal to 50% of the limit and less than 80% of the limit, or the current deformation rate greater than the historical deformation rate and less than or equal to 1.5 times the historical deformation rate; high risk is defined as an offset conversion coefficient greater than or equal to 80% of the limit, or the current deformation rate greater than 1.5 times the historical deformation rate. It is worth noting that when different trigger levels are applied, the higher level is used.

[0163] If the calculated current deformation rate is negative, it indicates that the actual deformation of the component is tending towards convergence and rebound. The system automatically implements negative polarity truncation: directly blocking the high-level penalty alarm caused by the historical rate of increase, and determining the final level only based on the static value range of the current deviation conversion coefficient. In addition, when the system finds that the component lacks historical deformation rate data records (such as when performing a comparison task for the first time), the system will automatically trigger a downgrade assessment strategy: skipping the trend comparison of deformation rate, and only using the deviation conversion coefficient as the assessment indicator for the three-level risk classification.

[0164] Another embodiment of the present invention provides a deformation detection system for existing building structures based on three-dimensional laser scanning, the system comprising:

[0165] The point cloud acquisition and preprocessing module is used to acquire the initial point cloud model of the existing building;

[0166] The component acquisition module is used to segment the initial point cloud model to obtain individual point cloud blocks; obtain semantic classification labels for each point cloud block based on its geometric shape features and physical reflection features; and merge the point cloud blocks based on the semantic classification labels to obtain several entity components.

[0167] The topology pattern generation module is used to generate building topology patterns by constructing an adjacency matrix based on the semantic classification labels of the entity components and the spatial distribution between the entity components, and by using an association rule algorithm to mine frequent itemsets of the component arrangement.

[0168] The topology mining and completion module is used to verify the current number of entity components using the building topology rules, locate the three-dimensional spatial region where topological conflicts occur, determine the occlusion probability by combining the three-dimensional local point cloud density of the three-dimensional spatial region, and generate virtual components in the three-dimensional spatial region according to the building topology rules when the occlusion probability meets the completion condition, so as to complete the initial point cloud model.

[0169] The deformation analysis and safety assessment module is used to obtain deformation reference benchmarks, calculate the deformation of each component based on the completed current point cloud model, filter construction deviations by combining preset tolerance thresholds, extract newly added structural deformation, and conduct structural safety trend assessment by combining the semantic classification labels of the corresponding components.

[0170] Another embodiment of the present invention provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a method for detecting deformation of existing building structures based on three-dimensional laser scanning.

Claims

1. A method for detecting deformation of existing building structures based on three-dimensional laser scanning, characterized in that, Includes the following steps: Obtain the initial point cloud model of the existing building; The initial point cloud model is segmented to obtain individual point cloud blocks; based on the geometric shape features and physical reflection features of each point cloud block, a semantic classification label is obtained for each point cloud block, and the point cloud blocks are merged according to the semantic classification labels to obtain several entity components. Based on the semantic classification labels of the entity components and the adjacency matrix constructed from the spatial distribution of each entity component, the association rule algorithm is used to mine the frequent itemsets of the component arrangement to generate building topology patterns. The building topology rules are used to verify the current physical components, locate the three-dimensional spatial region where topological conflicts occur, determine the occlusion probability by combining the three-dimensional local point cloud density of the three-dimensional spatial region, and generate virtual components in the three-dimensional spatial region according to the building topology rules when the occlusion probability meets the completion condition to complete the initial point cloud model; wherein, the occlusion probability meets the completion condition includes: the occlusion probability is greater than a preset probability threshold, and there are discrete ray blocking point clouds other than structural point clouds in the view frustum constructed by the line connecting the scanning station to the three-dimensional spatial region; Obtain a deformation reference benchmark, calculate the deformation of each component based on the completed current point cloud model, filter construction deviations by combining a preset tolerance threshold, extract the newly added structural deformation, and conduct a structural safety trend assessment by combining the semantic classification labels of the corresponding components.

2. The method for detecting deformation of existing building structures based on three-dimensional laser scanning according to claim 1, characterized in that, The step of obtaining semantic classification labels for each point cloud block based on its geometric shape and physical reflection features includes: For each point cloud block, a surface model is fitted using multiple types of geometric primitives. By calculating the residual characteristics of the data points deviating from the surface model, the component geometry type of the point cloud block is determined. The distribution characteristics of the reflection intensity of all data points in each point cloud block are statistically analyzed and compared with a pre-set material reflection attribute library to determine the component material properties of the point cloud block. By combining the component geometry type and component material properties of the same point cloud block, semantic classification labels for each point cloud block are obtained.

3. The method for detecting deformation of existing building structures based on three-dimensional laser scanning according to claim 2, characterized in that, For each point cloud block, a surface model is fitted using multiple types of geometric primitives. The component geometry type of the point cloud block is determined by calculating the residual characteristics of the data points deviating from the surface model. This includes: For each point cloud block, a planar model is fitted to obtain a fitted plane. The root mean square error of the distance from the data points in the point cloud block to the fitted plane is calculated as the flatness residual. For each cloud point, a cylindrical model is fitted to obtain the central axis and fitting radius of the fitted cylinder; Calculate the vertical distance from each data point in the point cloud block to the central axis, and calculate the absolute value of the difference between each vertical distance and the fitting radius; Calculate the average of the absolute values ​​of the differences corresponding to all data points, and use it as the cylindricity residual of the point cloud block; The component geometry type of the point cloud block is determined by comparing the flatness residual and cylindricity residual of the point cloud block with their corresponding preset thresholds.

4. The method for detecting deformation of existing building structures based on three-dimensional laser scanning according to claim 3, characterized in that, The component geometry types include linear and planar shapes. Several solid components are obtained by merging point cloud blocks based on semantic classification labels, including: Obtain the spatial bounding box of each point cloud block; For multiple point cloud blocks with the same semantic classification label, merge them according to the component geometry type: When the geometry of the component is linear, the fitted geometric centerline of each point cloud block is extracted, and point cloud blocks whose fitted geometric centerlines meet the preset collinearity condition and whose spatial bounding box spacing meets the preset linear connection threshold are merged and combined into a solid component. When the geometry of the component is planar, the fitting plane normal vector of each point cloud block is extracted, and point cloud blocks whose fitting plane normal vectors are parallel, whose normal distance satisfies the preset coplanar condition, and whose spatial bounding box spacing satisfies the preset planar connection threshold are merged and combined into a solid component.

5. The method for detecting deformation of existing building structures based on three-dimensional laser scanning according to claim 1, characterized in that, The adjacency matrix constructed based on the semantic classification labels of the entity components and the spatial distribution among the entity components is used to mine frequent itemsets of component arrangements using association rule algorithms to generate building topological patterns, including: Calculate the distance between the center points of any two solid components and the angle between the normals of the contact surfaces of the spatial bounding box. If the preset distance and preset angle conditions are met, it is determined that there is a spatial adjacency relationship between the two solid components, and a binary adjacency matrix is ​​constructed based on the spatial adjacency relationship between each solid component. For two physical components that are spatially adjacent, calculate the projection components of the center point coordinate difference vector on each coordinate axis, take the coordinate axis where the projection component with the largest absolute value is located as the principal direction axis, and determine the three-dimensional relative direction between the two physical components according to the sign of the projection component with the largest absolute value. For each entity component, the semantic classification label of the entity component, the semantic classification labels of other entity components that have spatial adjacency with the entity component as determined by the adjacency matrix, and the corresponding three-dimensional relative directions are combined to generate transaction data. The association rule algorithm is used to extract frequent itemsets that meet the preset support threshold from each generated transaction data, and the confidence of the association rules generated from the frequent itemsets is calculated. The association rules with a confidence greater than the preset confidence threshold are taken as the building topology rules.

6. The method for detecting deformation of existing building structures based on three-dimensional laser scanning according to claim 1, characterized in that, The process of using the building topology rules to verify the current number of physical components, locating the three-dimensional spatial region where topological conflicts occur, and determining the occlusion probability by combining the three-dimensional local point cloud density of the three-dimensional spatial region includes: Traverse each physical component and determine whether the actual spatial adjacency relationship of each physical component satisfies the building topology rules; If the actual spatial adjacency relationship of the target entity component lacks the association rules indicated in the building topology, it is determined that there is a missing expected component. Extract the three-dimensional relative direction indicated in the association rule and the corresponding average spacing of similar components, translate the center coordinates of the target entity component along the three-dimensional relative direction by the average spacing to obtain the expected center point, and construct a bounding box of a preset size with the expected center point as the three-dimensional spatial region where topological conflict occurs. The number of point clouds in the three-dimensional spatial region is counted as the three-dimensional local point cloud density; multiple reference entity components that do not have topological conflicts are selected from each entity component, and the average number of point clouds on the surface of all reference entity components within the same preset size is counted as the reference point cloud density of the overall building surface. Based on the difference between the three-dimensional local point cloud density and the reference point cloud density of the overall building surface, the occlusion probability of the expected missing component is calculated.

7. The method for detecting deformation of existing building structures based on three-dimensional laser scanning according to claim 6, characterized in that, When the occlusion probability meets the completion condition, virtual components are generated in the three-dimensional space region according to the building topology rules to complete the initial point cloud model, including: Extract the semantic classification label of the expected component from the association rules of the missing target entity component; determine the expected center point corresponding to the three-dimensional spatial region as the center coordinate of the virtual component; Extract entity components from the initial point cloud model that have the same semantic classification labels as the expected component, and calculate the average length, width, and height dimensions as the three-dimensional geometric dimensions of the virtual component; Based on the center coordinates, three-dimensional geometric dimensions, and semantic classification labels of the virtual component, a spatial bounding box of the virtual component is constructed to generate the virtual component. The virtual components and their semantic classification labels are incorporated into the component set of the initial point cloud model to complete the model.

8. The method for detecting deformation of existing building structures based on three-dimensional laser scanning according to claim 1, characterized in that, The process of obtaining the deformation reference benchmark involves calculating the deformation of each component based on the completed current point cloud model, filtering out construction-related deviations using a preset tolerance threshold, and extracting the newly added structural deformation, including: When the deformation reference datum is a historical scan point cloud model, the completed current point cloud model is spatially registered and aligned with the deformation reference datum to obtain the spatial registration and alignment result. Based on the spatial registration and alignment results, for the physical component, the point cloud region with the same semantic classification label and the smallest Euclidean distance between the center points is matched in the deformation reference datum as the historical corresponding component; the coordinate displacement change and spatial attitude vector between the physical component and the historical corresponding component are calculated to obtain the deformation amount of the physical component. Obtain the deformation of multiple adjacent physical components that are spatially adjacent to the virtual component; calculate the spatial distance from the center coordinates of the virtual component to the center coordinates of each adjacent physical component, and determine the deformation of the virtual component using the spatial distance and deformation of each adjacent physical component; The physical component and the virtual component are used as target components, and the resulting deformation is determined as the new structural deformation. When the deformation reference benchmark is the building structure design drawing, the new structural deformation of each target component is determined based on the building structure design drawing and the completed current point cloud model, combined with the preset tolerance threshold to filter out construction deviations.

9. The method for detecting deformation of existing building structures based on three-dimensional laser scanning according to claim 8, characterized in that, The process involves determining the additional structural deformation of each target component based on the architectural structural design drawings and the completed current point cloud model, combined with a preset tolerance threshold to filter out construction deviations. This includes: The theoretical center three-dimensional coordinates and principal axis theoretical normal vectors of each component are obtained from the architectural structural design drawings, and the corresponding preset position tolerance thresholds and angle tolerance thresholds are matched in the preset specification database using semantic classification tags. Align the completed current point cloud model with the building structure design drawings in the global spatial coordinate system, and calculate the independent translational absolute deviation of the actual center three-dimensional coordinates and the theoretical center three-dimensional coordinates of each component in each axis, as well as the difference in the normal angle between the actual principal axis normal vector and the principal axis theoretical normal vector. For each component, the absolute deviation of independent translation in each axis is compared with the preset position tolerance threshold, and the difference in normal angle is compared with the angle tolerance threshold. If the absolute deviation of independent translation in any axis is greater than the preset position tolerance threshold, or the difference in the normal angle is greater than the angle tolerance threshold, then the difference is determined to be excessive and the corresponding component is taken as the target component. For the target component, the difference exceeding the corresponding tolerance threshold in the translation dimension is combined with the original positive and negative displacement directions of that translation dimension, and the difference exceeding the corresponding tolerance threshold in the angular dimension is combined with the original deflection direction of that angular dimension. For translational or angular dimensions that do not exceed the corresponding tolerance threshold, the deformation component is set to zero; the multidimensional feature vector composed of the combined information and the zeroed component is determined as the new structural deformation of the target component.

10. A deformation detection system for existing building structures based on three-dimensional laser scanning, characterized in that, The system includes: The point cloud acquisition and preprocessing module is used to acquire the initial point cloud model of the existing building; The component acquisition module is used to segment the initial point cloud model to obtain individual point cloud blocks; obtain semantic classification labels for each point cloud block based on its geometric shape features and physical reflection features; and merge the point cloud blocks based on the semantic classification labels to obtain several entity components. The topology pattern generation module is used to generate building topology patterns by constructing an adjacency matrix based on the semantic classification labels of the entity components and the spatial distribution between the entity components, and by using an association rule algorithm to mine frequent itemsets of the component arrangement. The topology mining and completion module is used to verify the current number of entity components using the building topology rules, locate the three-dimensional spatial region where topological conflicts occur, determine the occlusion probability by combining the three-dimensional local point cloud density of the three-dimensional spatial region, and generate virtual components in the three-dimensional spatial region according to the building topology rules when the occlusion probability meets the completion conditions, so as to complete the initial point cloud model; wherein, the occlusion probability meets the completion conditions including: the occlusion probability is greater than a preset probability threshold, and there are discrete ray blocking point clouds other than structural point clouds in the view frustum constructed by the line connecting the scanning station to the three-dimensional spatial region; The deformation analysis and safety assessment module is used to obtain deformation reference benchmarks, calculate the deformation of each component based on the completed current point cloud model, filter construction deviations by combining preset tolerance thresholds, extract newly added structural deformation, and conduct structural safety trend assessment by combining the semantic classification labels of the corresponding components.

Images