Method for testing vertical flatness and gradient based on laser measurement

Abstract

The invention discloses a method for testing vertical flatness and gradient based on laser measurement, which comprises the following steps of: acquiring three-dimensional space coordinate data of a measured surface by utilizing laser measurement, and establishing a unified coordinate system; preprocessing the three-dimensional space coordinate data; inputting the standardized point cloud data into a PointNet feature extraction model, and generating feature data including local geometric features, global geometric features and point-level confidence; the feature data and the standardized point cloud data are combined and input into a weighted random consistency algorithm, and fitting optimization is executed according to the point-level confidence coefficient; inputting the parameters into an improved deep embedded clustering algorithm, and executing joint training by adopting shared feature representation; and generating a detection report and displaying on the terminal. According to the invention, accurate detection of the vertical flatness and gradient of the detected surface and abnormal area identification are realized, and the automation degree and data reliability of the detection process are improved.

Description

Technical Field

[0001] This invention relates to the field of engineering inspection and intelligent measurement technology, and in particular to a test method for measuring vertical flatness and slope based on laser measurement. Background Technology

[0002] With the increasing demands for construction quality inspection and equipment installation accuracy, laser-based surface flatness and structural slope detection technologies are gradually becoming important tools in engineering inspection. Existing measurement methods typically rely on laser levels, total stations, or single-point laser scanning equipment to acquire the geometric information of the surface being measured, and then calculate surface flatness and slope through simple coordinate transformations and linear fitting. However, the following problems commonly exist in practical applications:

[0003] First, existing methods mostly employ traditional point cloud preprocessing and linear fitting, which are easily affected by noise points and outliers, leading to significant deviations in the flatness and slope calculation results. Second, conventional random consistency algorithms primarily rely on equal-weighted sampling, failing to dynamically adjust the discrimination threshold using point cloud feature information, resulting in inaccurate inlier evaluation and insufficient fitting stability. Third, existing clustering methods are mostly based on traditional algorithms such as K-means or density clustering, unable to incorporate deep neural networks for embedding optimization and classification assistance, resulting in limited accuracy in identifying abnormal regions. Finally, the generated detection results are usually only numerical values ​​or simple graphs, lacking complete distribution maps and deviation annotations, failing to intuitively present the abnormal distribution of the detected area, thus limiting the value of data traceability and engineering applications.

[0004] Therefore, how to provide a laser-based testing method for measuring vertical flatness and slope is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a testing method for vertical flatness and slope based on laser measurement. This invention fully utilizes the technical steps of laser measurement, point cloud data processing, PointNet feature extraction, weighted random consistency algorithm, and improved deep embedded clustering combined with convolutional neural network classification. It describes in detail the entire process of automatically acquiring three-dimensional spatial coordinate data, performing point cloud preprocessing, extracting geometric features, fitting and calculating flatness and slope parameters, identifying abnormal areas, and generating a detection report. It has the advantages of high detection accuracy, strong anomaly identification capability, and high degree of result visualization.

[0006] A laser-based method for measuring vertical flatness and slope according to an embodiment of the present invention includes the following steps: Laser measurement is used to obtain three-dimensional spatial coordinate data of the surface being measured, and a unified coordinate system is established to determine the vertical reference direction and the horizontal reference. Preprocess the three-dimensional spatial coordinate data to generate standardized point cloud data; Standardized point cloud data is input into the PointNet feature extraction model to generate feature data containing local geometric features, global geometric features, and point-level confidence. The feature data and standardized point cloud data are combined and input into a weighted random consistency algorithm. Fitting optimization is performed based on point-level confidence, and the vertical flatness parameters and slope parameters are output. The vertical flatness parameters and slope parameters are input into the improved deep embedded clustering algorithm. Shared feature representations are used between the clustering embedding layer and the convolutional neural network classification module to perform joint training and output the final abnormal region. The system generates an inspection report based on vertical flatness parameters, slope parameters, and the final abnormal areas, and displays the inspection report on the terminal. The inspection report includes parameter values, distribution maps, and deviation annotations.

[0007] Optionally, the step of acquiring three-dimensional spatial coordinate data of the measured surface using laser measurement, establishing a unified coordinate system, and determining the vertical reference direction and horizontal reference specifically includes: The laser measuring device is used to set the scanning range in the area to be measured, and the laser emitting unit is controlled to emit laser beams sequentially according to the set path to perform a comprehensive scan of the surface to be measured. The system receives the reflection signal of the laser beam from the surface being measured, converts the reflection signal into distance data that can be used for calculation, and records the azimuth angle information and measurement time information for each measurement. By combining distance data and azimuth angle information, the three-dimensional coordinate values ​​of each measurement point are calculated to obtain the original coordinate point data containing spatial location; The measurement process is repeated to collect the three-dimensional coordinate values ​​of multiple measurement points, forming an initial coordinate set covering the measured surface. The initial coordinate set is represented and stored in the form of a point cloud. The initial coordinate set is sorted and structured, and a standardized point sequence is generated according to the measurement point number and spatial position relationship, and stored as a point cloud file; In the point cloud file, a reference point is selected as the calibration starting point. Coordinate alignment processing is performed on the overall point cloud data. The reference point is placed at the origin of the coordinate system by translation operation, and the point cloud direction is adjusted to be consistent with the reference reference direction by rotation operation. A unified coordinate system is established based on the point cloud data after coordinate alignment, defining three mutually perpendicular axial directions, where the X and Y axes are used to represent the horizontal extension direction, and the Z axis is used to represent the vertical direction; In the unified coordinate system, the Z-axis direction is defined as the vertical reference direction, and the plane containing the X-axis and Y-axis is defined as the horizontal reference.

[0008] Optionally, the preprocessing of the three-dimensional spatial coordinate data to generate standardized point cloud data specifically includes: The original point cloud set is established by collecting the three-dimensional spatial coordinate data in the sampling order; Multi-level screening is performed on the original point cloud set. First, points exceeding the limit are removed by a global threshold. Then, isolated points are identified based on local neighborhood relationships to form the first processed data. In the first processed data, a spatial segmentation method based on grid division is introduced to divide the point cloud into multiple sub-regions, and point density equalization processing is performed in each sub-region to obtain the second processed data. In the second processed data, local curvature information is generated, curvature is calculated through the geometric relationship between the point and its neighboring points, and the weights of each point are redistributed according to the curvature results to output weighted point cloud data. The weighted point cloud data is normalized and oriented to be consistent with the reference origin and coordinate range under a unified scale. The orientation of the point cloud is adjusted to be consistent with the unified coordinate system to obtain the third processed data. Multi-level formatting is performed on the third processed data to standardize and store the coordinate values, weight values, and curvature values ​​of each point, and output them as standardized point cloud data.

[0009] Optionally, the step of inputting standardized point cloud data into the PointNet feature extraction model to generate feature data containing local geometric features, global geometric features, and point-level confidence scores specifically includes: Standardized point cloud data is input as a point-level input sequence into the PointNet feature extraction model. Each point contains three-dimensional coordinate information. The point-level input sequence is then processed by sequential shuffling to obtain the first point-level input data. In the first-level input data, position encoding information is generated for each point. The position encoding information is combined with the three-dimensional coordinate information to obtain the second-level input data. Perform a neighborhood index construction operation on the second-level input data, build a neighborhood index table based on the spatial distance between points, and append the neighborhood index information to the corresponding point-level data to obtain the third-level input data; The third-level input data is input into the multilayer sensing unit. Point-by-point feature calculation is performed on each point, and the corresponding primary point-level feature vector is output. All primary point-level feature vectors are combined to form a primary point-level feature set. Based on the primary point-level feature set and the neighborhood index table, a local feature aggregation operation is performed to pool the features in the neighborhood of each point to generate a set of local feature vectors. Each point corresponds to a local feature vector, thus obtaining the first feature combination data. Perform a global aggregation operation on the first feature combination data, perform global pooling on the local feature vectors of all points to obtain a global feature vector, and combine it with the primary point-level feature set and the local feature vector set to generate the second feature combination data; In the second feature combination data, a composite feature vector is established for each point. The composite feature vector is formed by concatenating the corresponding primary point-level features, local feature vectors and global feature vectors, and outputs a composite feature set. The composite feature set is input into the uncertainty estimation module, the confidence value of each point is calculated, and the confidence value is appended to the corresponding composite feature vector to obtain the feature set with confidence. The confidence-based feature set is structured and output by separating the local geometric feature set, the global geometric feature vector, and the point-level confidence sequence to generate the third feature combination data. The third feature combination data is output as feature data, which includes a set of local geometric features, a global geometric feature vector, and a point-level confidence sequence.

[0010] Optionally, the step of combining feature data with standardized point cloud data into a weighted random consistency algorithm, performing fitting optimization based on point-level confidence, and outputting vertical flatness parameters and slope parameters specifically includes: Receive feature data and standardized point cloud data, and combine the local geometric feature set, global geometric feature vector, and point-level confidence sequence with the standardized point cloud coordinate point set to form the fitting input data; In fitting the input data, the sampling probability is determined based on the point-level confidence sequence, and weighted random sampling is performed on the standardized point cloud coordinate point set to generate the first sample point data, which contains the points with the highest confidence weight. Plane parameters are calculated using the first sample point data to generate first plane parameter data, and the first plane parameter data is combined with the global geometric feature vector to form first fitting result data; The first fitting result data is applied to the standardized point cloud coordinate point set, the geometric distance between each point and the first plane parameter data is calculated, and weighted discrimination is performed based on the point-level confidence sequence to generate the first interior point set data; The first set of interior points is combined with the set of local geometric features to update the fitted input data. The weighted random sampling and plane parameter calculation steps are repeated in the updated fitted input data to generate the second plane parameter data and the second fitting result data. The second fitting result data is applied again to the standardized point cloud coordinate point set, the geometric distance between each point and the second plane parameter data is calculated, and the interior point is evaluated based on the point-level confidence sequence. The second interior point set data is output, and multiple iterations are performed on this basis until the preset iteration termination condition is reached. After the iteration is completed, the final plane parameter data is extracted from the fitting result data output by the final iteration, and the final plane parameter data is decomposed into vertical flatness parameters and slope parameters in a unified coordinate system.

[0011] Optionally, the step of inputting vertical flatness parameters and slope parameters into an improved deep embedded clustering algorithm, using shared feature representations between the clustering embedding layer and the convolutional neural network classification module, performing joint training, and outputting the final abnormal region specifically includes: The vertical flatness parameters and slope parameters are combined into an input feature vector according to the order of the detection area number. The input feature vector is then standardized and batch normalized to generate clustered input data. The clustering input data is input into the encoder part of the improved deep embedded clustering algorithm. The encoder is composed of multiple fully connected operation units and nonlinear activation units in sequence. It performs dimensionality compression on the clustering input data, outputs embedded feature vectors, and establishes initial cluster centers in the embedding space to form the clustering data of the first improved deep embedded clustering algorithm. In the clustering data of the first improved deep embedded clustering algorithm, the assignment result is calculated based on the distance between the embedded feature vector and the initial cluster center. The soft assignment matrix is ​​generated using the Student distribution function, and the target distribution is constructed. The cluster center position is updated by minimizing the Kullback-Leibler divergence, and the clustering data of the second improved deep embedded clustering algorithm is output. The embedded feature vectors in the clustering data of the second improved deep embedded clustering algorithm are simultaneously input into the clustering embedding layer and the convolutional neural network classification module of the improved deep embedded clustering algorithm. The cluster center assignment is further calculated in the clustering embedding layer, and convolution, pooling and fully connected operations are performed in the convolutional neural network classification module. The classification result data of the convolutional neural network classification module is output. A shared feature representation is established between the clustering embedding layer and the convolutional neural network classification module of the improved deep embedded clustering algorithm. The embedded feature vector is used as a joint input, and the soft assignment matrix and classification result data are combined to form joint training data. The parameters of the encoder, clustering embedding layer and convolutional neural network classification module are trained end-to-end using a joint loss function that includes clustering loss term and classification loss term, and the clustering data of the third improved deep embedded clustering algorithm are output. In the clustering data of the third improved deep embedded clustering algorithm, suspected abnormal regions are marked according to the classification results of the convolutional neural network classification module, and the suspected abnormal regions are filtered in combination with the updated cluster center allocation results to generate candidate data of abnormal regions. The candidate data of abnormal regions are processed in a structured manner to output the final abnormal region data, which includes region number, boundary coordinates and abnormal label.

[0012] Optionally, the step of generating a test report from the vertical flatness parameters, slope parameters, and final abnormal areas, and displaying the test report on the terminal, includes parameter values, distribution maps, and deviation annotations, specifically including: The vertical flatness parameters and slope parameters are arranged in order of the inspection area number, and the final abnormal area data is marked according to the area index to generate the inspection report numerical data. Graphical processing is performed on the numerical data in the inspection report. The vertical flatness parameter and slope parameter are plotted as parameter distribution maps, the final abnormal area data is mapped as an abnormal area map, and the parameter distribution map and the abnormal area map are merged to generate graphical data of the inspection report. Add deviation annotations to the graphic data of the inspection report, mark the vertical flatness parameters and slope parameters that exceed the preset threshold at the corresponding parameter distribution map positions, and mark the boundaries of abnormal areas as deviation areas on the abnormal area map to generate inspection report annotation data; The data labeled in the test report is formatted by adding a test title, test area number, timestamp, and unified display label to generate formatted test report data. The test report format data is transmitted to the display terminal, and the test report containing parameter values, parameter distribution map, abnormal area map and deviation annotation is output on the terminal interface, generating test report display data; The inspection report data is stored as an inspection report file, and the index correspondence between the vertical flatness parameters, slope parameters, and final abnormal area data is maintained.

[0013] The beneficial effects of this invention are: This invention introduces intelligent point cloud processing and deep learning methods into traditional laser measurement, forming a complete vertical flatness and slope detection process. Point cloud preprocessing effectively removes outliers and isolated points, and curvature calculation and weight allocation improve data quality. The PointNet feature extraction model acquires local and global geometric features and point-level confidence scores, which are then combined with a weighted random consensus algorithm for fitting optimization, making the calculation of flatness and slope parameters more stable and reliable. Furthermore, an improved deep embedded clustering algorithm is used, employing shared feature representations and joint training between the clustering embedding layer and the convolutional neural network classification module, achieving high-precision identification of abnormal regions. The final detection report not only includes parameter values ​​but also intuitively presents distribution maps and deviation annotations, facilitating rapid location of problem areas. Compared with existing methods relying on single fitting or traditional clustering, this invention significantly improves detection accuracy and robustness, enhances the ability to identify abnormal regions, and improves the visualization and traceability of detection results, possessing high engineering application value. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0015] Figure 1 This is a flowchart of a laser-based test method for measuring vertical flatness and slope, as proposed in this invention. Figure 2 This is a schematic diagram of a laser-based test method for measuring vertical flatness and slope proposed in this invention. Figure 3 This is a schematic diagram of an improved deep embedded clustering algorithm in a laser-based test method for measuring vertical flatness and slope proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figure 1-3 A test method for measuring vertical flatness and slope based on laser measurement includes the following steps: Laser measurement is used to obtain three-dimensional spatial coordinate data of the surface being measured, and a unified coordinate system is established to determine the vertical reference direction and the horizontal reference. Preprocess the three-dimensional spatial coordinate data to generate standardized point cloud data; Standardized point cloud data is input into the PointNet feature extraction model to generate feature data containing local geometric features, global geometric features, and point-level confidence. The feature data and standardized point cloud data are combined and input into a weighted random consistency algorithm. Fitting optimization is performed based on point-level confidence, and the vertical flatness parameters and slope parameters are output. The vertical flatness parameters and slope parameters are input into the improved deep embedded clustering algorithm. Shared feature representations are used between the clustering embedding layer and the convolutional neural network classification module to perform joint training and output the final abnormal region. The system generates an inspection report based on vertical flatness parameters, slope parameters, and the final abnormal areas, and displays the inspection report on the terminal. The inspection report includes parameter values, distribution maps, and deviation annotations.

[0018] In this embodiment, the step of using laser measurement to acquire three-dimensional spatial coordinate data of the measured surface, establishing a unified coordinate system, and determining the vertical reference direction and horizontal reference specifically includes: The laser measuring device is used to set the scanning range in the area to be measured, and the laser emitting unit is controlled to emit laser beams sequentially according to the set path to perform a comprehensive scan of the surface to be measured. The system receives the reflection signal of the laser beam from the surface being measured, converts the reflection signal into distance data that can be used for calculation, and records the azimuth angle information and measurement time information for each measurement. By combining distance data and azimuth angle information, the three-dimensional coordinate values ​​of each measurement point are calculated to obtain the original coordinate point data containing spatial location; The measurement process is repeated to collect the three-dimensional coordinate values ​​of multiple measurement points, forming an initial coordinate set covering the measured surface. The initial coordinate set is represented and stored in the form of a point cloud. The initial coordinate set is sorted and structured, and a standardized point sequence is generated according to the measurement point number and spatial position relationship, and stored as a point cloud file; In the point cloud file, a reference point is selected as the calibration starting point. Coordinate alignment processing is performed on the overall point cloud data. The reference point is placed at the origin of the coordinate system by translation operation, and the point cloud direction is adjusted to be consistent with the reference reference direction by rotation operation. A unified coordinate system is established based on the point cloud data after coordinate alignment, defining three mutually perpendicular axial directions, where the X and Y axes are used to represent the horizontal extension direction, and the Z axis is used to represent the vertical direction; In the unified coordinate system, the Z-axis direction is defined as the vertical reference direction, and the plane containing the X-axis and Y-axis is defined as the horizontal reference.

[0019] This implementation method calculates three-dimensional coordinate points in the measured area using laser emission and reflection signals, and combines coordinate alignment and the establishment of a unified coordinate system. This effectively ensures the integrity of data acquisition and the consistency of the benchmark, reduces errors caused by uneven distribution of measurement points or differences in coordinate systems, and provides a high-precision foundation for subsequent point cloud processing.

[0020] In this embodiment, the preprocessing of the three-dimensional spatial coordinate data to generate standardized point cloud data specifically includes: The original point cloud set is established by collecting the three-dimensional spatial coordinate data in the sampling order; Multi-level screening is performed on the original point cloud set. First, points exceeding the limit are removed by a global threshold. Then, isolated points are identified based on local neighborhood relationships to form the first processed data. In the first processed data, a spatial segmentation method based on grid division is introduced to divide the point cloud into multiple sub-regions, and point density equalization processing is performed in each sub-region to obtain the second processed data. In the second processed data, local curvature information is generated, curvature is calculated through the geometric relationship between the point and its neighboring points, and the weights of each point are redistributed according to the curvature results to output weighted point cloud data. The weighted point cloud data is normalized and oriented to be consistent with the reference origin and coordinate range under a unified scale. The orientation of the point cloud is adjusted to be consistent with the unified coordinate system to obtain the third processed data. Multi-level formatting is performed on the third processed data to standardize and store the coordinate values, weight values, and curvature values ​​of each point, and output them as standardized point cloud data.

[0021] This implementation method, by screening, density equalizing, curvature calculating, and weighting the original point cloud, can significantly reduce the interference of noise points and isolated points on subsequent analysis, improve the continuity and geometric expression accuracy of point cloud data, and provide more stable data support for vertical flatness and slope calculation.

[0022] In this embodiment, the step of inputting standardized point cloud data into the PointNet feature extraction model to generate feature data containing local geometric features, global geometric features, and point-level confidence scores specifically includes: Standardized point cloud data is input as a point-level input sequence into the PointNet feature extraction model. Each point contains three-dimensional coordinate information. The point-level input sequence is then processed by sequential shuffling to obtain the first point-level input data. In the first-level input data, position encoding information is generated for each point. The position encoding information is combined with the three-dimensional coordinate information to obtain the second-level input data. Perform a neighborhood index construction operation on the second-level input data, build a neighborhood index table based on the spatial distance between points, and append the neighborhood index information to the corresponding point-level data to obtain the third-level input data; The third-level input data is input into the multilayer sensing unit. Point-by-point feature calculation is performed on each point, and the corresponding primary point-level feature vector is output. All primary point-level feature vectors are combined to form a primary point-level feature set. Based on the primary point-level feature set and the neighborhood index table, a local feature aggregation operation is performed to pool the features in the neighborhood of each point to generate a set of local feature vectors. Each point corresponds to a local feature vector, thus obtaining the first feature combination data. Perform a global aggregation operation on the first feature combination data, perform global pooling on the local feature vectors of all points to obtain a global feature vector, and combine it with the primary point-level feature set and the local feature vector set to generate the second feature combination data; In the second feature combination data, a composite feature vector is established for each point. The composite feature vector is formed by concatenating the corresponding primary point-level features, local feature vectors and global feature vectors, and outputs a composite feature set. The composite feature set is input into the uncertainty estimation module, the confidence value of each point is calculated, and the confidence value is appended to the corresponding composite feature vector to obtain the feature set with confidence. The confidence-based feature set is structured and output by separating the local geometric feature set, the global geometric feature vector, and the point-level confidence sequence to generate the third feature combination data. The third feature combination data is output as feature data, which includes a set of local geometric features, a global geometric feature vector, and a point-level confidence sequence.

[0023] This implementation introduces location encoding, neighborhood indexing, and uncertainty estimation through the PointNet feature extraction model, which can simultaneously obtain geometric features at both global and local scales and assign a confidence value to each point, thereby improving feature representation capabilities and enhancing the adaptability of subsequent fitting optimization to complex point cloud structures.

[0024] In this embodiment, the step of combining feature data with standardized point cloud data into a weighted random consistency algorithm, performing fitting optimization based on point-level confidence, and outputting vertical flatness parameters and slope parameters specifically includes: Receive feature data and standardized point cloud data, and combine the local geometric feature set, global geometric feature vector, and point-level confidence sequence with the standardized point cloud coordinate point set to form the fitting input data; In fitting the input data, the sampling probability is determined based on the point-level confidence sequence, and weighted random sampling is performed on the standardized point cloud coordinate point set to generate the first sample point data, which contains the points with the highest confidence weight. Plane parameters are calculated using the first sample point data to generate first plane parameter data, and the first plane parameter data is combined with the global geometric feature vector to form first fitting result data; The first fitting result data is applied to the standardized point cloud coordinate point set, the geometric distance between each point and the first plane parameter data is calculated, and weighted discrimination is performed based on the point-level confidence sequence to generate the first interior point set data; The first set of interior points is combined with the set of local geometric features to update the fitted input data. The weighted random sampling and plane parameter calculation steps are repeated in the updated fitted input data to generate the second plane parameter data and the second fitting result data. The second fitting result data is applied again to the standardized point cloud coordinate point set, the geometric distance between each point and the second plane parameter data is calculated, and the interior point is evaluated based on the point-level confidence sequence. The second interior point set data is output, and multiple iterations are performed on this basis until the preset iteration termination condition is reached. After the iteration is completed, the final plane parameter data is extracted from the fitting result data output by the final iteration, and the final plane parameter data is decomposed into vertical flatness parameters and slope parameters in a unified coordinate system.

[0025] This implementation combines feature data with standardized point cloud data into a weighted random consistency algorithm, and performs weighted sampling and iterative updates based on point-level confidence. This enables more accurate extraction of planar parameters in noisy environments, significantly improving the fitting accuracy and stability of vertical flatness and slope parameters.

[0026] In this embodiment, the step of inputting vertical flatness parameters and slope parameters into an improved deep embedded clustering algorithm, using shared feature representations between the clustering embedding layer and the convolutional neural network classification module, performing joint training, and outputting the final abnormal region specifically includes: The vertical flatness parameters and slope parameters are combined into an input feature vector according to the order of the detection area number. The input feature vector is then standardized and batch normalized to generate clustered input data. The clustering input data is input into the encoder part of the improved deep embedded clustering algorithm. The encoder is composed of multiple fully connected operation units and nonlinear activation units in sequence. It performs dimensionality compression on the clustering input data, outputs embedded feature vectors, and establishes initial cluster centers in the embedding space to form the clustering data of the first improved deep embedded clustering algorithm. In the clustering data of the first improved deep embedded clustering algorithm, the assignment result is calculated based on the distance between the embedded feature vector and the initial cluster center. The soft assignment matrix is ​​generated using the Student distribution function, and the target distribution is constructed. The cluster center position is updated by minimizing the Kullback-Leibler divergence, and the clustering data of the second improved deep embedded clustering algorithm is output. The embedded feature vectors in the clustering data of the second improved deep embedded clustering algorithm are simultaneously input into the clustering embedding layer and the convolutional neural network classification module of the improved deep embedded clustering algorithm. The cluster center assignment is further calculated in the clustering embedding layer, and convolution, pooling and fully connected operations are performed in the convolutional neural network classification module. The classification result data of the convolutional neural network classification module is output. A shared feature representation is established between the clustering embedding layer and the convolutional neural network classification module of the improved deep embedded clustering algorithm. The embedded feature vector is used as a joint input, and the soft assignment matrix and classification result data are combined to form joint training data. An end-to-end training is performed on the parameters of the encoder, clustering embedding layer and convolutional neural network classification module using a joint loss function that includes clustering loss term and classification loss term. The clustering loss term is constructed by comparing the difference between the soft assignment matrix and the target distribution, and the classification loss term is constructed by comparing the difference between the classification result data and the preset classification label. The joint loss function is the sum of the clustering loss term and the classification loss term, and the output is the clustering data of the third improved deep embedded clustering algorithm. In the clustering data of the third improved deep embedded clustering algorithm, suspected abnormal regions are marked according to the classification results of the convolutional neural network classification module, and the suspected abnormal regions are filtered in combination with the updated cluster center allocation results to generate candidate data of abnormal regions. The candidate data of abnormal regions are processed in a structured manner to output the final abnormal region data, which includes region number, boundary coordinates and abnormal label.

[0027] This implementation method, through the shared feature representation and joint training of the improved deep embedded clustering algorithm and the convolutional neural network classification module, can improve the classification judgment ability while maintaining the rationality of the clustering structure, achieve high-precision identification of abnormal regions, and avoid the problem of misclassification by single clustering methods.

[0028] In this embodiment, the process of generating a test report from the vertical flatness parameters, slope parameters, and final abnormal areas, and displaying the test report on the terminal, includes parameter values, distribution maps, and deviation annotations, specifically including: The vertical flatness parameters and slope parameters are arranged in order of the inspection area number, and the final abnormal area data is marked according to the area index to generate the inspection report numerical data. Graphical processing is performed on the numerical data in the inspection report. The vertical flatness parameter and slope parameter are plotted as parameter distribution maps, the final abnormal area data is mapped as an abnormal area map, and the parameter distribution map and the abnormal area map are merged to generate graphical data of the inspection report. Add deviation annotations to the graphic data of the inspection report, mark the vertical flatness parameters and slope parameters that exceed the preset threshold at the corresponding parameter distribution map positions, and mark the boundaries of abnormal areas as deviation areas on the abnormal area map to generate inspection report annotation data; The data labeled in the test report is formatted by adding a test title, test area number, timestamp, and unified display label to generate formatted test report data. The test report format data is transmitted to the display terminal, and the test report containing parameter values, parameter distribution map, abnormal area map and deviation annotation is output on the terminal interface, generating test report display data; The inspection report data is stored as an inspection report file, and the index correspondence between the vertical flatness parameters, slope parameters, and final abnormal area data is maintained.

[0029] This implementation method generates an intuitive inspection report by graphically processing and marking deviations of vertical flatness parameters, slope parameters, and final abnormal areas, and displays it on the terminal. This enables the digitization, visualization, and traceability of inspection results, improving the transparency and usability of the inspection process.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to the vertical flatness and slope detection of the exterior wall facade at a high-rise building construction site. In this scenario, the construction team needs to understand the overall flatness and local slope deviations of the exterior wall before completing large-scale tiling and decoration work, in order to correct construction deviations and control quality. Traditional detection methods mainly rely on manual string lines, straightedges, and total stations, which are time-consuming and labor-intensive, and due to the limited number of sampling points, they often fail to accurately reflect the overall condition of a large surface area. This invention, through a full-process detection based on laser measurement and intelligent algorithms, can complete large-scale facade scanning, data analysis, and result visualization in a short time.

[0031] During the on-site operation, a laser measurement device was first used to perform a full-coverage scan of the exterior facade, acquiring a large number of spatial measurement points. The measurement area covered a 30-meter-long and 60-meter-high area of ​​the building facade, with a point density controlled at approximately 500 points per square meter, totaling about 900,000 spatial coordinate points. After acquisition, the raw coordinate data was imported into the data processing system. The system first performed point cloud preprocessing operations, including outlier removal, noise point removal, point density equalization under grid division, and curvature-based weighted correction. After processing, the number of point clouds was reduced to approximately 800,000 points, but the point cloud distribution was more uniform, and the geometric representation was more stable.

[0032] Subsequently, the preprocessed point cloud data is input into the PointNet feature extraction model. The model automatically generates local geometric features, global geometric features, and point-level confidence scores for each point. Through feature extraction, the system can identify protruding brick joints, window frame edges, and local irregular areas in the exterior wall plane, and assign different confidence weights at the point level. Confidence scores are distributed between 0 and 1, with high-confidence points mainly concentrated in large planar areas, while low-confidence points are distributed at edges or in areas affected by noise.

[0033] In the fitting calculation stage, the system combines feature data with standardized point cloud data and inputs it into a weighted random consensus algorithm. Unlike traditional RANSAC, this invention introduces point-level confidence as a weighting factor during the sampling process, ensuring that high-confidence points are preferentially selected in the sampling. After hundreds of iterations, the system obtains complete flatness and slope parameter results. For example, in this inspection, the average vertical flatness of the exterior wall was 1.25 mm, with a maximum deviation of 4.83 mm; the average slope was 0.68°, with a maximum deviation of 2.15°. Comparing these data with the allowable deviation limits in the construction acceptance specifications shows that some areas significantly exceed the tolerances.

[0034] To further identify specific anomalous areas, this invention employs an improved deep embedded clustering algorithm and establishes shared feature representations between the clustering embedding layer and the convolutional neural network classification module. Through joint training of the clustering embedding and classification modules, the system can distinguish suspected anomalous areas. Taking the area at the bottom of the exterior wall and above the windows as an example, after inputting the flatness and slope parameters into the clustering algorithm, the system outputs multiple candidate anomalous areas and performs secondary confirmation based on the convolutional neural network classification results. Ultimately, approximately 30,000 points out of 800,000 points were identified as anomalous areas, accounting for about 3.7%. These anomalous areas are concentrated at the concrete joints at the bottom of the exterior wall and at individual formwork splicing locations. The system automatically outputs the area number, boundary coordinates, and anomalous label.

[0035] In the results output stage, the system comprehensively processes the vertical flatness parameters, slope parameters, and final abnormal area data to generate an inspection report. The inspection report is displayed in the form of numerical values, distribution maps, and deviation annotations. Users can intuitively view the average flatness, average slope, and locations of deviations exceeding limits for each area on their terminal screen. The report includes a traceable Excel spreadsheet; some data examples are shown in Table 1.

[0036] Table 1 Statistical Table of Exterior Wall Inspection Results

[0037] Analysis of the data in Table 1 reveals that the average flatness of most areas is controlled between 1.0 and 1.3 mm, and the average slope is controlled between 0.6° and 0.7°, meeting the construction specifications. However, some areas, such as A-02 and B-07, have a maximum flatness deviation exceeding 3.5 mm and a maximum slope deviation exceeding 2°. These areas also have a significantly higher number of anomalies than other areas, indicating localized construction quality issues. Based on the inspection report generated by this invention, the construction team can quickly locate the anomalies and arrange for manual correction, avoiding rework during subsequent large-scale decoration work.

[0038] The following benefits of the present invention are demonstrated by the embodiments: By introducing an optimization process that combines PointNet feature extraction with a weighted random consensus algorithm, the calculation of flatness and slope parameters becomes more accurate; by sharing feature representation and joint training of an improved deep embedded clustering algorithm and a convolutional neural network classification module, high-precision identification of abnormal areas is achieved; by combining graphical inspection reports with Excel numerical data, the inspection results are both intuitive and traceable, greatly improving the efficiency and reliability of construction quality inspection.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for measuring vertical flatness and slope based on laser, characterized in that, The method comprises the following steps: acquiring three-dimensional spatial coordinate data of the measured surface by laser measurement, establishing a unified coordinate system, determining a vertical reference direction and a horizontal reference, and preprocessing the three-dimensional spatial coordinate data to generate standardized point cloud data; inputting the standardized point cloud data into a PointNet feature extraction model to generate feature data containing local geometric features, global geometric features and point-level confidence; combining the feature data and the standardized point cloud data to input a weighted random consistency algorithm, performing fitting optimization according to the point-level confidence, and outputting a vertical flatness parameter and a slope parameter; inputting the vertical flatness parameter and the slope parameter into an improved deep embedding clustering algorithm, adopting shared feature representation between the clustering embedding layer and the convolutional neural network classification module, performing joint training, and outputting a final abnormal area; generating a detection report including the parameter values, distribution map and deviation label based on the vertical flatness parameter, slope parameter and final abnormal area, and displaying the detection report on a terminal. The method of acquiring three-dimensional spatial coordinate data of the measured surface by laser measurement, establishing a unified coordinate system, determining a vertical reference direction and a horizontal reference, specifically comprises:

2. The method of claim 1, wherein the method further comprises: setting a scanning range in the measured area by a laser measurement device, and controlling a laser emission unit to emit laser beams in sequence according to the set path to perform covering scanning on the measured surface; receiving the reflection signals of the laser beams on the measured surface, converting the reflection signals into distance data that can be used for calculation, and recording the azimuth angle information and measurement time information of each measurement; combining the distance data and the azimuth angle information to calculate the three-dimensional coordinate values of each measurement point to obtain original coordinate point data containing spatial positions; repeating the measurement process to collect three-dimensional coordinate values of multiple measurement points to form an initial coordinate set covering the measured surface, which is represented and stored in the form of point cloud; sorting and structuring the initial coordinate set, generating a standardized point sequence according to the measurement point number and spatial position relationship, and storing it as a point cloud file; selecting a reference point as a calibration starting point in the point cloud file, performing coordinate alignment processing on the overall point cloud data, placing the reference point at the origin position by translation operation, and adjusting the point cloud direction to be consistent with the reference reference direction by rotation operation; establishing a unified coordinate system based on the point cloud data after completing the coordinate alignment, defining three mutually perpendicular axial directions, wherein the X-axis and Y-axis are used to represent the horizontal extension direction, and the Z-axis is used to represent the vertical direction; in the unified coordinate system, the Z-axis direction is determined as the vertical reference direction, and the plane where the X-axis and Y-axis are located is determined as the horizontal reference. The method of preprocessing the three-dimensional spatial coordinate data to generate standardized point cloud data specifically comprises:

3. The method of claim 1, wherein the method further comprises: establishing an original point cloud set according to the sampling order of the three-dimensional spatial coordinate data; performing multi-level screening on the original point cloud set, first removing out-of-limit points by a global range threshold, and then identifying isolated points based on local neighborhood relationship to form first processing data; ​ In the first processing data, a spatial segmentation method based on grid division is introduced, the point cloud is divided into multiple sub-regions, and point density equalization processing is performed in each sub-region to obtain second processing data; In the second processing data, local curvature information is generated, the curvature is calculated through the geometric relationship between the point and the neighborhood point, and the weight of each point is redistributed according to the curvature result, and the weighted point cloud data is output; The weighted point cloud data is subjected to coordinate normalization and direction correction, unified to the coordinate range under the reference origin and unified scale, and the point cloud orientation is adjusted to be consistent with the unified coordinate system to obtain third processing data; In the third processing data, a multi-level formatting operation is performed, the coordinate value, weight value and curvature value of each point are standardized and stored, and standardized point cloud data is output.

4. The method of claim 1, wherein the method further comprises: The standardized point cloud data is input into the PointNet feature extraction model to generate feature data containing local geometric features, global geometric features and point-level confidence, specifically including: The standardized point cloud data is input into the PointNet feature extraction model as a point-level input sequence, each point contains three-dimensional coordinate information, and the point-level input sequence is subjected to order scattering processing to obtain first point-level input data; In the first point-level input data, position encoding information is generated for each point, and the position encoding information and the three-dimensional coordinate information are combined to obtain second point-level input data; The second point-level input data is subjected to a neighborhood index construction operation, a neighborhood index table is established based on the spatial distance between points, and the neighborhood index information is attached to the corresponding point-level data to obtain third point-level input data; The third point-level input data is input into a multi-layer perception unit, and a point-by-point feature calculation is performed on each point to output a corresponding primary point-level feature vector, and all primary point-level feature vectors are combined to form a primary point-level feature set; Based on the primary point-level feature set and the neighborhood index table, a local feature aggregation operation is performed, the features in the neighborhood of each point are pooled to generate a local feature vector set, each point corresponds to a local feature vector, and first feature combination data is obtained; The first feature combination data is subjected to a global aggregation operation, the local feature vectors of all points are globally pooled to obtain a global feature vector, and the primary point-level feature set, the local feature vector set and the global feature vector are combined to generate second feature combination data; In the second feature combination data, a composite feature vector is established for each point, the composite feature vector is formed by splicing the corresponding primary point-level feature, local feature vector and global feature vector, and a composite feature set is output; The composite feature set is input into an uncertainty estimation module to calculate the confidence value of each point, and the confidence value is attached to the corresponding composite feature vector to obtain a feature set with confidence; The feature set with confidence is subjected to structured processing to separate and output a local geometric feature set, a global geometric feature vector and a point-level confidence sequence to generate third feature combination data; The third feature combination data is output as feature data, and the feature data includes a local geometric feature set, a global geometric feature vector and a point-level confidence sequence.

5. The method of claim 1, wherein the method further comprises: The feature data and the standardized point cloud data are combined and input into a weighted random consistency algorithm, fitting optimization is performed according to point-level confidence, and vertical flatness parameters and slope parameters are output, specifically comprising: Receiving feature data and standardized point cloud data, combining local geometric feature sets, global geometric feature vectors, and point-level confidence sequences with standardized point cloud coordinate point sets to form fitting input data; In the fitting input data, the sampling probability is determined according to the point-level confidence sequence, the weighted random sampling is performed on the standardized point cloud coordinate point set, the first sample point data is generated, and the first sample point data contains points with the highest confidence weight; The first plane parameter data is calculated using the first sample point data, and the first plane parameter data is combined with the global geometric feature vector to form the first fitting result data; The first fitting result data is applied to the standardized point cloud coordinate point set, the geometric distance of each point to the first plane parameter data is calculated, and the weighted judgment is performed according to the point-level confidence sequence to generate the first inlier set data; The first inlier set data is combined with the local geometric feature set to update the fitting input data, and the weighted random sampling and plane parameter calculation steps are repeatedly performed in the updated fitting input data to generate second plane parameter data and second fitting result data; The second fitting result data is applied to the standardized point cloud coordinate point set again, the geometric distance of each point to the second plane parameter data is calculated, and the inlier is evaluated according to the point-level confidence sequence to output the second inlier set data, and multiple iterations are performed on this basis until a preset iteration termination condition is reached; After the iteration ends, the final plane parameter data is extracted from the fitting result data output in the final iteration, and the final plane parameter data is decomposed into vertical flatness parameters and slope parameters in a unified coordinate system.

6. The method of claim 1, wherein the method further comprises: The vertical flatness parameters and the slope parameters are input into the improved deep embedded clustering algorithm, a shared feature representation is adopted between the clustering embedding layer and the convolutional neural network classification module, joint training is performed, and the final abnormal area is output, specifically comprising: The vertical flatness parameters and the slope parameters are combined into an input feature vector in the order of detection region numbers, standardization and batch normalization processing is performed on the input feature vector to generate clustering input data; The clustering input data is input into the encoder part of the improved deep embedded clustering algorithm, the encoder is composed of multiple fully connected operation units and nonlinear activation units in sequence, the dimension of the clustering input data is compressed to output an embedded feature vector, and an initial clustering center is established in the embedding space to form the clustering data of the first improved deep embedded clustering algorithm; In the clustering data of the first improved deep embedded clustering algorithm, the distance between the embedded feature vector and the initial clustering center is calculated to generate a soft assignment matrix using a student distribution function, and a target distribution is constructed, the clustering center position is updated by minimizing the Kullback-Leibler divergence, and the clustering data of the second improved deep embedded clustering algorithm is output; The embedding feature vector in the clustering data of the second improved deep embedding clustering algorithm is input into the clustering embedding layer and the convolutional neural network classification module of the improved deep embedding clustering algorithm, the clustering center assignment is continued to be calculated in the clustering embedding layer, the convolution operation, the pooling operation and the full connection operation are executed in the convolutional neural network classification module, and the classification result data of the convolutional neural network classification module is output; The shared feature representation is established between the clustering embedding layer and the convolutional neural network classification module of the improved deep embedding clustering algorithm, the embedding feature vector is input as a joint input, the soft assignment matrix and the classification result data are combined to form joint training data, and the joint loss function including the clustering loss term and the classification loss term is used to perform end-to-end training on the parameters of the encoder, the clustering embedding layer and the convolutional neural network classification module, and the clustering data of the third improved deep embedding clustering algorithm is output; In the clustering data of the third improved deep embedding clustering algorithm, the suspected abnormal area is marked according to the classification result data of the convolutional neural network classification module, and the suspected abnormal area is screened according to the updated clustering center assignment result to generate abnormal area candidate data; The abnormal area candidate data is structured to output final abnormal area data, and the final abnormal area data includes area number, boundary coordinates and abnormal label.

7. The method of claim 1, wherein the method further comprises: The vertical flatness parameter, the slope parameter and the final abnormal area are used to generate a detection report, and the detection report is displayed on the terminal, and the detection report includes parameter values, distribution maps and deviation labels, specifically including: The vertical flatness parameter and the slope parameter are arranged in order of the detection area number, and the final abnormal area data is marked according to the area index to generate detection report value data; In the detection report value data, graphical processing is performed, the vertical flatness parameter and the slope parameter are drawn as parameter distribution maps, the final abnormal area data is mapped as an abnormal area map, and the parameter distribution maps and the abnormal area map are combined to generate detection report graphical data; In the detection report graphical data, the deviation label is added, the vertical flatness parameter and the slope parameter exceeding the preset threshold are marked in the corresponding parameter distribution map position, and the boundary of the abnormal area is marked as a deviation area on the abnormal area map to generate detection report labeling data; The detection report labeling data is formatted to add a detection title, a detection area number, a timestamp and a unified display label to generate detection report formatted data; The detection report formatted data is transmitted to the display terminal, and the detection report containing parameter values, parameter distribution maps, abnormal area maps and deviation labels is output on the terminal interface to generate detection report display data; The detection report display data is stored as a detection report file, and the index correspondence relationship with the vertical flatness parameter, the slope parameter and the final abnormal area data is maintained.

Images