Industrial heritage building structure disease identification method based on point cloud data

Abstract

The present application relates to a kind of industrial heritages building structure disease identification method based on point cloud data, it is related to industrial heritages building structure disease detection field, its steps include: obtaining global three-dimensional laser point cloud data, extracts geometric deformation feature and material reflection feature after pre-processing;Two-stage identification mode of combining initial screening with point cloud semantic segmentation fine classification using rule threshold, identify crack, concrete spalling, component deformation and steel corrosion and other disease types from point cloud;The disease identified is positioned in three-dimensional space, and according to disease type, length, width, depth, deformation, corrosion rate and other quantitative parameters are directly calculated from point cloud;Disease type, spatial location and quantitative parameter are integrated and output as standardization achievement.The present application aims to solve the problems that existing technology relies on image quality, it is difficult to effectively identify non-surface texture disease and cannot directly quantify the whole space based on point cloud.

Description

Technical Field

[0001] This invention relates to the field of structural defects detection in industrial heritage buildings, and specifically to a method for identifying structural defects in industrial heritage buildings based on point cloud data. Background Technology

[0002] Industrial heritage buildings generally have long service lives, with structural components exhibiting varying degrees of aging and damage. Furthermore, many original building drawings are missing, making structural defects highly concealed. Traditional structural defect detection relies heavily on manual on-site inspection and measurement, which has three core drawbacks: First, the detection coverage is limited, making it difficult to reach high-altitude areas and hidden nodes, resulting in a high rate of missed defects and the potential for structural safety hazards. Second, the quantitative accuracy of defects is insufficient; parameters such as crack width, spalling depth, and deformation rely on manual estimation, leading to large errors and failing to provide accurate data support for reinforcement design. Third, the detection efficiency is low, and manual inspection can easily cause secondary damage to historical components through contact.

[0003] Current image-based defect identification methods can only acquire surface visual information and cannot achieve three-dimensional spatial positioning and depth quantification. They are difficult to adapt to the detection needs of large spaces and complex components in industrial heritage buildings and cannot support full-dimensional structural safety assessment. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying structural defects in industrial heritage buildings based on point cloud data, in order to solve the technical problems of high false negative rate, low quantification accuracy, difficulty in detecting hidden defects, low detection efficiency, and easy damage to historical components in existing industrial heritage building defect detection methods.

[0005] The present invention is implemented as follows: A method for identifying structural defects in industrial heritage buildings based on point cloud data includes the following steps.

[0006] Step S1: Obtain full-area 3D laser point cloud data of the industrial heritage building to be inspected.

[0007] Step S2: Preprocess the global 3D laser point cloud data and extract multi-dimensional features from the preprocessed point cloud data. The multi-dimensional features include geometric deformation features and material reflection features.

[0008] Step S3: Based on the multi-dimensional features, a two-level identification method combining rule threshold initial screening and point cloud semantic segmentation fine classification is adopted to identify structural defects, and the defect types include cracks, concrete spalling, component deformation and steel corrosion are obtained.

[0009] Step S4: The identified defects are located in three-dimensional space, and parameters are quantified according to the defect type to obtain quantitative parameters of defects including cracks, concrete spalling, component deformation and steel corrosion.

[0010] Step S5: Integrate the disease type, spatial location information, and quantitative parameters into a standardized output.

[0011] Furthermore, the preprocessing in step S2 includes: removing outliers from the point cloud through statistical filtering, and compressing the point cloud data volume while preserving the structural contours and detailed features through voxel downsampling.

[0012] Furthermore, the extraction of geometric deformation features in step S2 includes: calculating the normal vector and curvature of each point cloud using principal component analysis; fitting the reference ideal surface of the component using moving least squares method, calculating the vertical deviation value of each point cloud from the reference surface, and extracting the elevation difference, local curvature abrupt change rate, and flatness deviation as geometric deformation features.

[0013] Furthermore, the extraction of material reflection features in step S2 includes: extracting the laser reflection intensity value inherent in the point cloud, completing intensity normalization processing through distance correction and incident angle correction, and obtaining standardized material reflection features, which are used to distinguish material change-related defects.

[0014] Furthermore, the judgment rules for the initial screening of the rule threshold in step S3 include: areas with a local curvature change rate greater than 0.2 mm / m and a surface vertical deviation greater than 3 mm are judged as suspected crack areas; areas with a surface vertical deviation greater than 5 mm are judged as suspected concrete spalling areas; areas with a reflection intensity value lower than 30% of the average value of intact components of the same type are judged as suspected corrosion areas; and component sections with an overall deviation exceeding the allowable value of the specification are judged as suspected deformation areas.

[0015] Furthermore, the point cloud semantic segmentation fine classification in step S3 adopts the improved PointNet++ point cloud semantic segmentation model to perform fine classification on the suspected areas obtained from the initial screening by rule thresholds, and outputs the identification results of four typical defects: cracks, concrete spalling, component deformation and steel corrosion.

[0016] Furthermore, the improved PointNet++ point cloud semantic segmentation model is pre-trained using an annotated industrial heritage building defects point cloud dataset. The training set covers industrial building defects samples from different eras and with different structural forms, including trusses, frames, and roof trusses.

[0017] Furthermore, the parameter quantification calculation of the crack in step S4 includes: extracting the central skeleton line of the point cloud of the crack region, calculating the total length of the skeleton line as the crack length through spatial polyline integration; taking multiple cross sections at equal intervals along the normal of the skeleton line, calculating the span value of the crack point cloud in each cross section, taking the average value of the span values ​​of all cross sections as the average crack width, and taking the maximum value as the maximum crack width.

[0018] Furthermore, the parameter quantification calculation for concrete spalling in step S4 includes: fitting a reference plane to the intact surface around the spalled area, calculating the vertical distance from each point in the spalled area to the reference plane, taking the average value as the average spalling depth, and taking the maximum value as the maximum spalling depth; and calculating the planar projected area of ​​the spalled area using the projected area method.

[0019] Furthermore, the parameter quantification calculation of component deformation in step S4 includes: registering and comparing the measured point cloud of the component to be tested with the theoretical model of the standard component obtained by reverse modeling, calculating the actual deviation value of each point, and outputting the maximum deformation, average deformation and deformation distribution range of the component.

[0020] Furthermore, step S4 includes the following steps: quantifying the parameters of corrosion, calculating the ratio of the corrosion area to the corresponding component surface area based on the point cloud distribution range of the corrosion area, and outputting the spatial coordinate range and distribution pattern of the corrosion area.

[0021] Furthermore, the standardized results described in step S5 include: classifying all identified defects by component, integrating defect types, spatial coordinates, quantitative parameters and corresponding component numbers to generate a structured digital defect ledger; mapping the defect classification results to the building's three-dimensional digital model to generate a visualized defect distribution map with spatial positioning; and outputting data in a format compatible with structural design software and BIM platforms.

[0022] This invention extracts multi-dimensional features directly from full-domain 3D laser point cloud data and adopts a two-level recognition architecture of rule threshold initial screening and point cloud semantic segmentation for fine classification. It can effectively identify non-visual texture defects such as cracks, peeling, deformation and corrosion by utilizing geometric deformation features and material reflection features, while improving processing efficiency while maintaining high recognition accuracy, resulting in high overall recognition accuracy.

[0023] Based on the three-dimensional spatial positioning and parameter quantization calculation of point cloud data under a unified engineering coordinate system, all quantized parameters are directly obtained from the point cloud, avoiding errors in manual estimation, and the measurement accuracy reaches the millimeter level, which can provide accurate data support for reinforcement design.

[0024] The non-contact, full-area point cloud acquisition method eliminates the need for personnel to approach high altitudes or hidden locations, thus eliminating blind spots in the detection process, significantly reducing the missed detection rate, and preventing damage to historical components during the detection process.

[0025] The output of standardized digital disease ledgers and 3D disease distribution maps can be directly connected to structural design software and BIM platforms, opening up the data link between inspection, design and construction, and meeting the application needs of the entire process of digital transformation of industrial heritage. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method for identifying structural defects in industrial heritage buildings based on point cloud data, according to the present invention. Detailed Implementation

[0027] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] like Figure 1 The diagram shown is a flowchart of the method for identifying structural defects in industrial heritage buildings based on point cloud data according to the present invention.

[0029] In this embodiment, the implementation process is described in detail using a single-story frame structure industrial plant as an example. The plant includes typical components such as concrete columns, steel roof trusses, and walls, and suffers from four common defects: cracks, concrete spalling, component deformation, and steel corrosion.

[0030] Step S1: Acquire full-area 3D laser point cloud data of the industrial heritage building to be inspected. A terrestrial 3D laser scanner is used to perform multi-station scans of the factory's interior and exterior to acquire high-density point cloud data. For blind spots such as roof trusses and roofs, supplementary point clouds are obtained using UAV oblique photography. All point cloud data are registered to a unified engineering coordinate system using targets and control points, forming full-area 3D laser point cloud data covering all areas including the building's facade, internal structure, and roof truss nodes. The spacing between single-station scan points is no greater than 2 millimeters to ensure the accuracy of subsequent feature extraction.

[0031] Step S2 involves preprocessing the global 3D laser point cloud data and extracting multi-dimensional features from the preprocessed point cloud data, including geometric deformation features and material reflection features.

[0032] The specific preprocessing steps are as follows: First, a statistical filtering algorithm is used to perform statistical analysis on the neighborhood of each point, calculating the average distance from each point to its K nearest neighbors. Outliers with average distances exceeding a set standard deviation (e.g., 2 times) of the global average distance are removed to eliminate random errors during the scanning process. Then, voxel downsampling is performed, dividing the point cloud space into cubic voxels with sides of 5 mm. Within each voxel, the point closest to the voxel center is retained. While preserving structural contours, edges, and detailed features, the point cloud data volume is compressed to approximately 30% of the original data volume, significantly improving subsequent computational efficiency.

[0033] The geometric deformation features are extracted as follows: Principal Component Analysis (PCA) is used to calculate the local neighborhood covariance matrix of each point cloud, solving for eigenvalues ​​and eigenvectors. The eigenvector corresponding to the smallest eigenvalue is the normal vector of that point, and the curvature value of that point is calculated using the ratio of eigenvalues. Then, Moving Least Squares (MLS) is used to perform local surface fitting on the point cloud of the component surface, serving as the reference ideal surface for the component. The vertical deviation of each point cloud from this reference surface is calculated. Based on this, three types of geometric deformation features are extracted: elevation difference, i.e., the difference between the actual elevation of the point cloud and the elevation of the reference surface; local curvature abrupt change rate, defined as the relative change in the curvature value of that point compared to the average curvature value of its neighborhood; and flatness deviation, i.e., the standard deviation of the local area point cloud from the fitted plane. These features can describe the changes in the surface shape of the component from different dimensions, providing a basis for identifying defects such as cracks, spalling, and deformation.

[0034] This embodiment solves the problem of difficulty in quantifying three-dimensional spatial defects such as component deformation and spalling depth by directly performing geometric calculations on the three-dimensional point cloud.

[0035] The extraction method for material reflection features is as follows: The laser reflection intensity value inherent in the point cloud is extracted, as this value is directly related to the reflection characteristics of the target material. Since scanning distance and laser incident angle cause nonlinear attenuation of the intensity value, normalization correction is necessary. This embodiment uses a distance correction function and a Lambert cosine correction model to correct the reflection intensity value of each point to the equivalent intensity value under standard distance and perpendicular incident conditions, eliminating differences in scanning conditions. After correction, standardized material reflection features are obtained. The intensity values ​​of intact concrete areas are usually concentrated in a certain range, while the intensity values ​​of corroded steel and weathered concrete deviate significantly from this range, enabling the differentiation of material change-related defects.

[0036] It should be noted that the extraction of geometric deformation features is not limited to the combination of PCA and MLS described above. In another feasible embodiment, a method based on point cloud segmentation and 3D model registration can be used to extract geometric deformation features. This involves first segmenting the point cloud into regions, constructing an ideal parametric 3D model (such as a BIM model) for each component, and then performing nearest-point iteration (ICP) registration between the measured point cloud and the ideal model. The deviation value of each point from the model surface is directly calculated as the geometric deformation feature. Similarly, the normalization of material reflection features can also employ an intensity correction method based on on-site calibration using a reference whiteboard. By scanning a standard whiteboard with known reflectivity at the same testing station, a linear mapping relationship between intensity values ​​and reflectivity is established, thus completing the correction. All of the above alternative methods can achieve the objectives of this invention and fall within the scope of protection of this invention.

[0037] Step S3: Based on the extracted multi-dimensional features, a two-level identification method combining rule threshold initial screening and point cloud semantic segmentation fine classification is adopted to identify structural diseases and obtain disease types.

[0038] The first level is the initial screening using rule thresholds. Its purpose is to quickly identify suspected defect areas, eliminate a large number of defect-free areas, and improve overall detection efficiency. The judgment rules set in this embodiment are as follows: For each point cloud area, if its local curvature abrupt change rate is greater than 0.2 mm / m and its surface vertical deviation is greater than 3 mm, it is judged as a suspected crack area; if the surface vertical deviation is greater than 5 mm and exhibits a concave shape, it is judged as a suspected concrete spalling area; if the standardized reflection intensity value is lower than 70% of the average strength of intact component areas of the same type (i.e., more than 30% lower than the average), it is judged as a suspected corrosion area; if the overall deviation of a component section exceeds the deformation limit allowed by specifications such as the "Technical Standard for Building Structure Inspection," it is judged as a suspected deformation area. These thresholds are set based on the statistical characteristics of common defects in industrial heritage buildings and can be fine-tuned according to component type and detection accuracy requirements in practical applications. Through the above rules, defect-free areas, such as smooth walls and intact coatings, accounting for more than 90%, can be quickly filtered out.

[0039] The second stage is point cloud semantic segmentation and fine classification, employing an improved PointNet++ point cloud semantic segmentation model. This model, based on the original PointNet++'s multi-scale grouping and feature propagation structure, adds a feature fusion layer targeting geometric deformation and material reflection features. The geometric and intensity feature vectors extracted in step S2 are merged with the original point cloud coordinates as input, enabling the network to simultaneously learn spatial geometry and material properties. The suspected area point cloud obtained from the first stage's initial screening is input into the model, which performs classification prediction for each point, outputting four labels: cracks, concrete spalling, component deformation, and steel corrosion, while removing a small number of false positives from the initial screening. The model is pre-trained using a large dataset of labeled point cloud data on industrial heritage building defects. The training set contains defect samples from industrial buildings with different structural forms, such as trusses, frames, and roof trusses, constructed between the 1950s and 1990s, totaling over 50 million labeled points, covering scenes under different lighting and scanning resolutions, ensuring strong adaptability of the model to industrial heritage scenes.

[0040] Step S4: The identified diseases are located in three-dimensional space, and the parameters are quantified according to the disease type to obtain the quantitative parameters of the disease.

[0041] All positioning and calculations are performed based on point cloud data in a unified engineering coordinate system, eliminating the need for manual measurement.

[0042] For the quantification of crack defects: First, extract the central skeleton line of the point cloud of the crack area, ensuring that the skeleton line runs through the crack area and maintains topological connectivity; then, calculate the total length of the skeleton line through spatial polyline integration, which is the crack length, with a calculation accuracy of up to 1 mm. Along the skeleton line direction, extract normal sections at equal intervals (e.g., 10 mm). Within each section, calculate the span value of the crack point cloud along the normal direction of the skeleton line; this span value is the crack width at that section. Traverse all sections, take the arithmetic mean of all section width values ​​as the average crack width, and take the maximum value as the maximum crack width.

[0043] For the quantification of concrete spalling defects: First, the point cloud of intact concrete surfaces surrounding the spalling area is fitted to a plane. After removing points in the spalling area using the RANSAC algorithm, a reference plane is obtained. Then, the vertical distance from each point in the spalling area to the reference plane is calculated. Positive values ​​indicate depressions, and the average of all distances is taken as the average spalling depth, with the maximum value being the maximum spalling depth. Simultaneously, the projected area method is used to project the point cloud of the spalling area along the normal direction of the reference plane, calculate the area of ​​the projected polygon, and obtain the planar projected area of ​​the spalling area.

[0044] For the quantification of component deformation defects: First, a theoretical 3D model of the standard component is constructed using reverse modeling based on the component's design dimensions or the point cloud of intact sections. For example, for a steel roof truss chord, the standard I-beam section and ideal axis are fitted by extracting the point clouds of the flanges and webs of several intact sections. The measured point cloud of the component to be tested is then registered with the theoretical model of the standard component using ICP. After eliminating the overall positional deviation, the actual deviation value of each point is calculated, and the maximum deformation, average deformation, and deformation distribution range along the component's length are output.

[0045] For quantifying corrosion damage: Based on the distribution range of the point cloud of the corrosion area, the Alpha Shape algorithm is used to extract the boundary, calculate the actual surface area of ​​the corrosion area, and then divide it by the theoretical surface area of ​​the corresponding intact area of ​​the component to obtain the corrosion rate. Simultaneously, the spatial coordinate range and polygonal distribution pattern of the corrosion area are output. For example, the corrosion rate of a steel roof truss node area is 15.8%, with corrosion mainly distributed at the edge of the node plate.

[0046] Step S5: Integrate the disease type, spatial location information, and quantitative parameters into a standardized output.

[0047] Specifically, all identified defects are categorized according to their respective components, generating a defect record for each component. This record includes the defect type (cracks / stripping / deformation / corrosion), spatial coordinates (e.g., minimum bounding box corner coordinates), quantitative parameters (length, width, depth, area, corrosion rate, etc.), corresponding component number (e.g., column C12, beam B07), and inspection timestamp, forming a structured digital defect ledger. Simultaneously, the defect classification results are mapped onto the building's 3D digital model, using different colors and symbols to label various defects, generating a spatially located visual defect distribution map. The output defect ledger is in CSV and JSON formats, while the defect distribution map uses the IFC standard format. This data can be directly imported into mainstream BIM platforms such as Revit and Tekla, as well as structural design software such as PKPM and SAP2000, enabling seamless data transfer from inspection to reinforcement design and construction management.

[0048] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying structural defects in industrial heritage buildings based on point cloud data, characterized in that, Includes the following steps: Step S1: Obtain full-area 3D laser point cloud data of the industrial heritage building to be inspected; Step S2: Preprocess the global three-dimensional laser point cloud data and extract multi-dimensional features from the preprocessed point cloud data. The multi-dimensional features include geometric deformation features and material reflection features. Step S3: Based on the multi-dimensional features, a two-level identification method combining rule threshold initial screening and point cloud semantic segmentation fine classification is adopted to identify structural defects, and the defect types including cracks, concrete spalling, component deformation and steel corrosion are obtained. Step S4: The identified defects are located in three-dimensional space, and the parameters are quantitatively calculated according to the defect type to obtain the quantitative parameters of defects including cracks, concrete spalling, component deformation and steel corrosion. Step S5: Integrate the disease type, spatial location information, and quantitative parameters into a standardized output.

2. The method for identifying structural defects in industrial heritage buildings based on point cloud data according to claim 1, characterized in that, The extraction of geometric deformation features in step S2 includes: calculating the normal vector and curvature of each point cloud using principal component analysis; fitting the reference ideal surface of the component using moving least squares method, calculating the vertical deviation value of each point cloud from the reference surface, and extracting the elevation difference, local curvature abrupt change rate, and flatness deviation as geometric deformation features.

3. The method for identifying structural defects in industrial heritage buildings based on point cloud data according to claim 1, characterized in that, The extraction of material reflection features in step S2 includes: extracting the laser reflection intensity value from the point cloud, completing intensity normalization processing through distance correction and incident angle correction, and obtaining standardized material reflection features, which are used to distinguish material change-related defects.

4. The method for identifying structural defects in industrial heritage buildings based on point cloud data according to claim 1, characterized in that, The judgment rules for the initial screening of the rule threshold in step S3 include: areas with a local curvature change rate greater than 0.2 mm / m and a surface vertical deviation greater than 3 mm are judged as suspected crack areas; areas with a surface vertical deviation greater than 5 mm are judged as suspected concrete spalling areas; areas with a reflection intensity value lower than 30% of the average value of intact components of the same type are judged as suspected corrosion areas; and component sections with an overall deviation exceeding the allowable value of the specification are judged as suspected deformation areas.

5. The method for identifying structural defects in industrial heritage buildings based on point cloud data according to claim 4, characterized in that, The point cloud semantic segmentation and fine classification in step S3 adopts the improved PointNet++ point cloud semantic segmentation model to perform fine classification on the suspected areas obtained from the initial screening by rule thresholds, and output the identification results of four typical defects: cracks, concrete spalling, component deformation and steel corrosion.

6. The method for identifying structural defects in industrial heritage buildings based on point cloud data according to claim 1, characterized in that, The parameter quantification calculation of the crack in step S4 includes: extracting the central skeleton line of the point cloud of the crack region, calculating the total length of the skeleton line as the crack length through spatial polyline integration; taking multiple cross sections at equal intervals along the normal of the skeleton line, calculating the span value of the crack point cloud in each cross section, taking the average value of all cross section span values ​​as the average crack width, and taking the maximum value as the maximum crack width.

7. The method for identifying structural defects in industrial heritage buildings based on point cloud data according to claim 1, characterized in that, The parameter quantification calculation for concrete spalling in step S4 includes: fitting a reference plane to the intact surface around the spalled area, calculating the vertical distance from each point in the spalled area to the reference plane, taking the average value as the average spalling depth, and taking the maximum value as the maximum spalling depth; and calculating the planar projected area of ​​the spalled area using the projected area method.

8. The method for identifying structural defects in industrial heritage buildings based on point cloud data according to claim 1, characterized in that, The parameter quantification calculation of component deformation in step S4 includes: registering and comparing the measured point cloud of the component to be tested with the theoretical model of the standard component obtained by reverse modeling, calculating the actual deviation value of each point, and outputting the maximum deformation, average deformation and deformation distribution range of the component.

9. The method for identifying structural defects in industrial heritage buildings based on point cloud data according to claim 1, characterized in that, Step S4 involves the following steps: quantifying the parameters of corrosion, calculating the ratio of the corrosion area to the corresponding component surface area based on the point cloud distribution range of the corrosion area, and outputting the spatial coordinate range and distribution pattern of the corrosion area.

10. The method for identifying structural defects in industrial heritage buildings based on point cloud data according to claim 1, characterized in that, The standardized deliverables in step S5 include: classifying all identified defects by component, integrating defect types, spatial coordinates, quantitative parameters and corresponding component numbers to generate a structured digital defect ledger; mapping the defect classification results to the building's three-dimensional digital model to generate a visualized defect distribution map with spatial positioning; and outputting data in a format compatible with structural design software and BIM platforms.

Images