Bridge crack time sequence evolution analysis method and system based on dynamic point cloud registration

Abstract

The invention relates to the technical field of bridge support damage identification, in particular to a bridge crack time sequence evolution analysis method and system based on dynamic point cloud registration, and the method comprises the steps: synchronously collecting the three-dimensional original point cloud data of a target region where a to-be-detected bridge support is located through a plurality of laser radar sensors; on the basis of the three-dimensional original point cloud data, preprocessing the point cloud data, and extracting a region of interest of the to-be-detected bridge support; constructing a plurality of local geometric features of the point cloud based on an improved multi-point geometric operator; predicting label categories of the points based on a graph attention network; obtaining a segmentation result of each part of the support by using a hierarchical segmentation strategy; according to the method, geometric analysis is carried out on each part of the support, a support part segmentation result and a support damage evaluation result are output, and a hierarchical point cloud segmentation strategy can effectively consider the overall structure and local details of the support, so that the segmentation accuracy is improved, and a more reliable basis is provided for subsequent damage identification.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge crack evolution analysis, and more specifically, to a method and system for analyzing the temporal evolution of bridge cracks based on dynamic point cloud registration. Background Art

[0002] With the rapid development of infrastructure construction, bridges, as important transportation hubs, their safety and durability have attracted increasing attention. Bridge cracks are one of the important factors affecting structural safety. Timely and accurate monitoring and analysis of the development of cracks are crucial for bridge maintenance and management. Traditional bridge crack detection methods mainly rely on manual visual inspection, which is not only time-consuming and laborious, but also easily affected by subjective factors, making it difficult to ensure the accuracy and consistency of detection.

[0003] In recent years, with the progress of computer vision and 3D reconstruction technologies, crack detection methods based on image processing have gradually become a research hotspot. These methods collect high-resolution images of the bridge surface and use image processing algorithms to identify and quantify cracks. However, methods based on 2D images have limitations in dealing with complex 3D structures and are difficult to accurately reflect the depth information and spatial distribution characteristics of cracks.

[0004] To overcome the limitations of 2D images, some researchers have begun to attempt to use 3D laser scanning technology to collect point cloud data of bridges. Point cloud data can more comprehensively describe the geometric characteristics of bridges and provide new possibilities for crack detection. However, existing point cloud-based crack detection methods still have some problems. First, the registration accuracy of point cloud data directly affects the accuracy of crack identification, but current registration algorithms still face challenges in dealing with large-scale and high-precision bridge point cloud data. Second, how to effectively extract crack features from massive point cloud data and accurately distinguish cracks from other surface defects is still an unsolved problem. In addition, most existing methods focus on crack detection at a single moment and lack in-depth analysis of the temporal evolution process of cracks, making it difficult to provide strong support for the long-term health monitoring of bridges. Summary of the Invention

[0005] The present invention aims to solve the above technical problems and proposes a method and system for analyzing the temporal evolution of bridge cracks based on dynamic point cloud registration. This method realizes high-precision point cloud registration, accurate crack identification, and prediction of crack development trends through an innovative combination of an improved ICP algorithm, density clustering, and deep learning technology.

[0006] The present invention provides a method for analyzing the temporal evolution of bridge cracks based on dynamic point cloud registration, including:

[0007] An acquisition step, including:

[0008] Collect the surface laser point cloud data of the bridge;

[0009] Construct a multi-temporal three-dimensional image sample library;

[0010] Processing steps, including:

[0011] Based on the surface laser point cloud data, calculate the normal vector and curvature, and identify the cracks in the point cloud;

[0012] Use the improved ICP algorithm to register the bridge surface point clouds at different time series;

[0013] Output steps, including:

[0014] Based on the registered point cloud, use the point cloud convolutional neural network and attention mechanism to analyze the development trend of bridge cracks, and obtain the crack propagation prediction and crack damage analysis diagram.

[0015] Preferably, the identification of cracks in the point cloud specifically includes:

[0016] Calculate the normal vector and curvature within the neighborhood of each point in the three-dimensional surface laser point cloud;

[0017] Use the density clustering method to cluster all the point clouds on the bridge, and record each clustering area as a crack.

[0018] Preferably, the construction of the multi-temporal three-dimensional image sample library specifically includes:

[0019] Use a drone equipped with a lidar to collect the three-dimensional surface data of the bridge above the bridge;

[0020] Use a portable data recorder to record the point cloud data, the actual position, length, and width data of the cracks at time t of the three-dimensional point cloud data;

[0021] Input the collected point cloud data into the point cloud recognition system to identify cracks, and output the bridge surface point cloud crack recognition results, including the position, length, and width information of the cracks;

[0022] Match and analyze the crack information on the bridge surface and the crack recognition information on the point cloud, establish a bridge crack analysis model for multi-temporal point cloud registration, and record the evolution law and change trend of the cracks.

[0023] Preferably, the use of the point cloud convolutional neural network and attention mechanism to analyze the development trend of bridge cracks specifically includes:

[0024] Establish a crack evolution analysis model based on the point cloud convolutional neural network and attention mechanism;

[0025] Using the registered data point cloud and the crack recognition results, train and test the crack time-series evolution model to obtain an optimized model;

[0026] Input the multi-temporal point cloud into the optimized model to output the crack propagation prediction and the crack damage analysis diagram.

[0027] Preferably, the improved ICP algorithm specifically includes:

[0028] Perform preprocessing of denoising, downsampling, resampling, and plane extraction on the collected point cloud data, and screen the feature points;

[0029] Calculate the normal vector and curvature feature information of the feature points;

[0030] Using the ICP algorithm, iteratively calculate the minimum value of the distance between corresponding points, obtain the transformation relationship of the point cloud, and complete the point cloud registration.

[0031] Preferably, the density clustering method specifically includes:

[0032] Use the Euclidean formula to calculate the number of neighbor points of each point p(x, y, z) in the point cloud;

[0033] Mark the points with the calculated point cloud density greater than the density threshold as the points in the crack area;

[0034] Set the number of samples of the minimum clustering cluster and the density of the minimum clustering cluster, use the DBSCAN clustering algorithm to cluster the point cloud, and mark all the clustered regions as cracks.

[0035] Preferably, the improved ICP algorithm further includes:

[0036] Use principal component analysis to find the point planes of the two groups of point clouds, obtain the plane normal vector of the point cloud from point cloud A, and use this normal vector as the initialization direction for registration;

[0037] Adjust the direction of the normal vector to the z-axis direction, and use the normal vector combined with the SVD decomposition method to obtain the rotation transformation matrix;

[0038] Calculate the overlapping region of the two groups of point clouds through the registration transformation;

[0039] Solve the least squares method of the overlapping region to minimize the distance between corresponding points and solve the parameters of the registration transformation.

[0040] Preferably, the crack evolution analysis model based on the point cloud convolutional neural network and the attention mechanism specifically includes:

[0041] Use the point cloud convolutional neural network algorithm to extract the features of the registered point cloud by using a multi-layer convolutional neural network for the registered point cloud;

[0042] An LSTM neural network is used to construct a temporal evolution model of cracks for crack identification based on the crack features on point clouds at different time series;

[0043] The attention mechanism is used to weight and filter the crack features at different time series to obtain the evolution law and change trend of cracks.

[0044] Preferably, it also includes setting the change of the crack area in three registration periods to determine the growth rate of cracks.

[0045] A bridge crack temporal evolution analysis system based on dynamic point cloud registration for executing the above method, the system includes:

[0046] A point cloud registration module, which is used to roughly register the binocular three-dimensional point cloud data with the standard three-dimensional point cloud data, and then use the ICP iterative algorithm to accurately register the binocular three-dimensional point cloud data with the standard three-dimensional point cloud data to obtain the registered three-dimensional point cloud data;

[0047] A crack extraction module, which is used to obtain the crack area of the registered three-dimensional point cloud data;

[0048] A crack analysis module, which is used to analyze the crack width, crack length, crack depth, and crack propagation trend of the real crack area;

[0049] A growth rate trend evaluation module, which is used to set the change of the crack area in three registration periods.

[0050] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0051] First of all, the improved ICP algorithm proposed by the present invention significantly improves the accuracy and efficiency of point cloud registration. By introducing the distance metric from point to plane and the adaptive threshold strategy, this algorithm can better handle the complex geometric structure of the bridge surface, effectively reduce the problem of local optimal solutions, and ensure the precise alignment of point cloud data at different times. This lays a solid foundation for subsequent crack identification and evolution analysis.

[0052] Secondly, the density clustering method adopted by the present invention combines local geometric features and global spatial distribution information, greatly improving the accuracy and robustness of crack identification. This method can effectively distinguish cracks from other surface defects, such as textures, stains, etc., reducing the false detection rate. At the same time, by reasonably setting the clustering parameters, this method can also identify tiny cracks, improving the detection sensitivity.

[0053] Furthermore, the present invention innovatively introduces a point cloud convolutional neural network and an attention mechanism to achieve an in-depth analysis of the temporal evolution process of cracks. This method can not only capture the spatial features of cracks but also effectively model the variation law of cracks over time. Through the joint analysis of multi-temporal point cloud data, this method can accurately predict the development trend of cracks, providing an important basis for the preventive maintenance of bridges.

[0054] Finally, the method proposed by the present invention has good scalability and adaptability. By adjusting relevant parameters, this method can be applied to bridge structures of different types and scales. At the same time, the modular design of this method enables each link to be independently optimized and upgraded according to actual needs, ensuring the long-term effectiveness of the system.

[0055] In summary, the present invention not only solves the problems in the prior art such as insufficient point cloud registration accuracy, low crack recognition accuracy, and lack of temporal evolution analysis but also realizes the automation and intelligence of bridge crack detection and prediction. This is of great significance for improving the efficiency and accuracy of bridge detection, extending the service life of bridges, and reducing maintenance costs, providing a new technical solution for the field of bridge health monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is the flowchart of the method of the present invention.

[0057] Figure 2 is the logic block diagram of the point cloud registration module of the present invention.

[0058] Figure 3 is the logic block diagram of the crack extraction module of the present invention.

[0059] Figure 4 is the logic block diagram of the crack analysis module of the present invention.

[0060] Figure 5 is the logic block diagram of the growth rate trend evaluation module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] Please refer to Figures 1-3 , the present invention provides a method and system for analyzing the temporal evolution of bridge cracks based on dynamic point cloud registration. This method first collects the surface laser point cloud data of the bridge and constructs a multi-temporal three-dimensional image sample library. Then, based on the surface laser point cloud data, the normal vector and curvature are calculated to identify cracks in the point cloud. Next, the improved ICP algorithm is used to register the bridge surface point clouds at different time series. Finally, based on the registered point clouds, a point cloud convolutional neural network and an attention mechanism are used to analyze the development trend of bridge cracks, obtaining a crack expansion prediction and a crack damage analysis diagram.

[0062] In the step of identifying cracks in the point cloud, the method of the present invention first calculates the normal vector and curvature within the neighborhood of each point in the three-dimensional surface laser point cloud. Preferably, an appropriate neighborhood radius, such as 5 cm - 10 cm, can be selected to ensure that sufficient local geometric information can be captured. The normal vector can be calculated using the principal component analysis (PCA) method, and the curvature can be calculated based on the fitted local surface.

[0063] The present invention uses the PCA method to calculate the normal vector. For point p and its neighborhood point set {p 1 , p 2 ,..., p K}, the covariance matrix is constructed:

[0064]

[0065] where is the mean of the neighborhood points. Perform eigenvalue decomposition on C, and the eigenvector corresponding to the minimum eigenvalue is the normal vector.

[0066] Based on the normal vector, the curvature of point p can be estimated:

[0067]

[0068] where: λ 1 , λ 2 , λ 3 are the eigenvalues of the covariance matrix C, and λ 1 ≥λ 2 ≥λ 3

[0069] Next, the method of the present invention uses the density clustering method to cluster all the point clouds on the bridge, and each clustering region is recorded as a crack. Here, DBSCAN (Density-Based Spatial Clustering of Applications with Noise) can be used to achieve density clustering. The two key parameters of the DBSCAN algorithm are eps (neighborhood radius) and minPts (minimum number of points). For bridge crack identification, according to experience, eps can be selected as 2 mm - 5 mm, and minPts can be selected as 5 - 10 points. Such parameter selection can effectively identify small cracks and avoid misjudging noise points as cracks.

[0070] The core of the DBSCAN algorithm is to expand the points that are density-reachable. For point p, its ∈ neighborhood is defined as:

[0071] N ∈ (p) = {q ∈ D | dist(p, q) ≤ ∈}

[0072] where D is the point set, dist(p, q) is the distance between point p and q,

[0073] The condition for point p to be a core point is:

[0074] |N ∈ (p)| ≥ MinPts

[0075] where MinPts is the minimum number of points threshold.

[0076] In the step of constructing the multi-temporal three-dimensional image sample library, the method of the present invention first uses an unmanned aerial vehicle to carry a lidar to collect the three-dimensional surface data of the bridge over the bridge. Preferably, a high-precision lidar can be selected, such as Velodyne HDL-64E, whose point cloud accuracy can reach ±2 cm and the scanning frequency can reach 20 Hz, and it can quickly and accurately collect the three-dimensional data of the bridge surface.

[0077] At the same time, the method of the present invention uses a portable data recorder to record the point cloud data, the actual position, length, and width data of the cracks at time t of the three-dimensional point cloud data. The portable data recorder here can be a portable system integrated with a high-precision GPS, IMU (inertial measurement unit), and a data storage device. Preferably, the GPS accuracy should reach the centimeter level, and the angle accuracy of the IMU should be better than 0.1° to ensure the accuracy of the data.

[0078] Next, the method of the present invention inputs the collected point cloud data into a point cloud recognition system to identify cracks, and outputs the recognition result of the point cloud cracks on the bridge surface, including the position, length, and width information of the cracks. The point cloud recognition system here can be based on deep learning methods, such as PointNet++ or DGCNN (Dynamic Graph Convolutional Neural Network), and these methods can directly process three-dimensional point cloud data without converting the point cloud into a two-dimensional image.

[0079] The present invention adopts the PointNet++ architecture, and its core operation is the SetAbstraction (SA) layer. The output of the SA layer can be expressed as:

[0080]

[0081] where x i , f i are the coordinates and features of the input points, x′ i , f′ i are the coordinates and features of the output points, N is the number of input points, and the SA layer internally includes three steps: sampling, grouping, and feature extraction. Feature extraction uses a multi-layer perceptron (MLP):

[0082]

[0083] where is the neighborhood of point i.

[0084] The detailed description of these algorithms ensures the feasibility of the method of the present invention, providing a solid technical foundation for the analysis of the temporal evolution of bridge cracks.

[0085] Finally, the method of the present invention performs matching analysis on the crack information on the bridge surface and the crack identification information on the point cloud, establishes a bridge crack analysis model for multi-temporal point cloud registration, and records the evolution law and change trend of the cracks. In this step, the ICP (Iterative Closest Point) algorithm can be used for point cloud registration, and then time series analysis methods such as ARIMA (Autoregressive Integrated Moving Average Model) or LSTM (Long Short-Term Memory Network) can be used to model the evolution law of the cracks.

[0086] Through the above steps, the method of the present invention can comprehensively and accurately capture the temporal evolution process of bridge cracks, providing an important basis for the health monitoring and maintenance of bridges. This method based on dynamic point cloud registration has higher accuracy and stronger robustness compared with traditional two-dimensional image analysis methods, and can better cope with the influence of complex bridge structures and various environmental factors. When the method of the present invention analyzes the development trend of bridge cracks using a point cloud convolutional neural network and an attention mechanism, a crack evolution analysis model based on a point cloud convolutional neural network and an attention mechanism is first established. This model combines the spatial characteristics of point cloud data and the temporal characteristics of crack evolution, and can effectively capture the development law of cracks.

[0087] Preferably, the present invention uses PointNet++ as the basic architecture of the point cloud convolutional neural network. PointNet++ can effectively process local features of different scales through hierarchical sampling and grouping operations, and is suitable for complex bridge structures. In the design of the network, 3 - 5 SetAbstraction layers can be selected, and the number of sampled points in each layer can be gradually reduced from 2048 to 256 to capture sufficient detailed information while maintaining computational efficiency.

[0088] Then, this method uses the registered data point cloud and the crack identification results to train and test the crack temporal evolution model to obtain an optimized model. During the training process, the Adam optimizer can be used, the initial learning rate is set to 0.001, and a learning rate decay strategy is used to halve the learning rate every 50 epochs. To improve the generalization ability of the model, data augmentation techniques such as random rotation, translation, and scaling are also introduced in the preferred embodiment of the present invention to simulate point cloud data at different angles and distances.

[0089] Finally, the method of the present invention inputs multi-temporal point clouds into an optimization model, and outputs crack propagation predictions and crack damage analysis diagrams. In this step, the attention mechanism plays an important role. The present method uses a self-attention mechanism, enabling the model to adaptively focus on important features at different time points and spatial positions. Specifically, a multi-head attention mechanism can be used, and the number of heads can be selected as 8 or 16 to capture feature correlations in different subspaces.

[0090] In terms of the improved ICP algorithm, the method of the present invention first preprocesses the collected point cloud data. This includes steps such as denoising, downsampling, resampling, and plane extraction. For denoising, the Statistical Outlier Removal method can be used, setting the neighborhood point number to 50 and the standard deviation multiple to 2.0 to effectively remove outliers. For downsampling, voxel grid filtering can be used, and the voxel size can be set to 0.01m to significantly reduce the data volume while maintaining the main features of the point cloud.

[0091] In the improved ICP algorithm of the present invention, a new error metric method is introduced, which combines the point-to-point and point-to-plane distances to improve the accuracy and robustness of registration. The error function can be expressed as:

[0092]

[0093] where E is the total error, p i is a point in the source point cloud, q i is the corresponding point in the target point cloud, n i is the normal vector of the corresponding point in the target point cloud, R is the rotation matrix, t is the translation vector, α is the weight coefficient controlling the proportion of the point-to-point distance and the point-to-plane distance, N is the number of corresponding point pairs, and α can be set to 0.5 to balance the contributions of the two distance metrics.

[0094] To solve this optimization problem, the present invention adopts an iterative optimization method based on Lie algebra. Define the transformation matrix T = [R|t], and its Lie algebra representation is Through the exponential mapping, the Lie algebra can be converted into a transformation matrix:

[0095]

[0096] where: ω∧ is the skew-symmetric matrix of ω,

[0097]

[0098] Using this representation method, the optimization problem can be transformed into a minimization problem regarding ξ. The Gauss-Newton method is used for iterative solution, and the increment Δξ of ξ is updated in each iteration:

[0099] Δξ = -(JT WJ) -1 J T Wr

[0100] Among them, J is the Jacobian matrix, W is the weight matrix, and r is the residual vector.

[0101] This optimization method based on Lie algebra can avoid the singularity problem of the rotation matrix and improve the stability and convergence speed of the algorithm.

[0102] In terms of point cloud feature extraction, the present invention proposes a multi-scale feature descriptor that combines local geometric information and global context information. The descriptor can be expressed as:

[0103] F(p) = [F l (p), F g (p)]

[0104] Among them, F(p) is the feature descriptor of point p, F l (p) is the local feature, and F g (p) is the global feature;

[0105] The local feature F l (p) is calculated using an improved PCA method:

[0106] F l (p) = [λ 1 , λ 2 , λ 3 , θ 1 , θ 2 , θ 3

[0107] Among them, λ 1 , λ 2 , λ 3 are the eigenvalues of PCA, and θ 1 , θ 2 , θ 3 are the angles between the eigenvectors and the coordinate axes.

[0108] The global feature F g (p) is obtained through an adaptive spatial pyramid pooling method:

[0109]

[0110] Among them, B k (p) represents the k-th layer spatial pyramid centered at p, f(q) is the local feature of point q, K is the number of pyramid layers, preferably set to 3. This multi-scale feature descriptor can effectively capture the local details and global structure of the point cloud and improve the accuracy of crack recognition.

[0111] ​In the analysis of crack evolution trends, the present invention proposes a spatio-temporal model based on the Graph Attention Network (GAT). This model can be expressed as:

[0112]

[0113] Where, represents the feature of node i at the l-th layer, represents the set of neighbors of node i, and W (l) is the weight matrix at the ι-th layer, and σ is the activation function α ij is the attention coefficient, which is calculated as follows:

[0114]

[0115] Where, a is the learnable attention vector, and ∥ represents the concatenation operation.

[0116] By stacking multiple layers of GAT, the model can effectively capture the spatial relationships and temporal evolution characteristics between cracks, thereby accurately predicting the development trend of cracks.

[0117] The introduction of these algorithms has greatly improved the performance and accuracy of the present invention in the time-series evolution analysis of bridge cracks, providing strong technical support for bridge health monitoring.

[0118] Next, this method calculates the normal vector and curvature feature information of the feature points. The normal vector calculation can use the PCA method, and the number of neighborhood points is selected as 30. The curvature feature can be estimated by fitting a quadratic surface, and usually the same neighborhood size as the normal vector calculation is selected.

[0119] Finally, the method of the present invention uses the ICP algorithm to iteratively calculate the minimum value of the distance between corresponding points, obtain the transformation relationship of the point cloud, and complete the point cloud registration. Based on the standard ICP algorithm, some improvement strategies are introduced in this method. For example, the distance metric from point to plane is used instead of the distance from point to point, which can better handle the bridge surface with more plane structures. In addition, an adaptive threshold strategy is also introduced. The initial threshold can be set to 1% of the diagonal length of the point cloud bounding box, and then gradually decreased with the iteration to balance between rough registration and fine registration.

[0120] In terms of the density clustering method, the method of the present invention first calculates the number of neighbor points of each point p(x, y, z) in the point cloud using the Euclidean formula. Here, the neighbor is defined as the points within the radius r, and r can be determined according to the density of the point cloud. Usually, 2 - 3 times the average point spacing of the point cloud can be selected. Preferably, the k-d tree can be used to accelerate the nearest neighbor search and improve the calculation efficiency.

[0121] Next, the method marks the points with a point cloud density greater than the density threshold as points in the crack area. The selection of the density threshold has an important impact on the results and can be determined through statistical analysis. For example, the upper quartile of the density distribution can be selected as the threshold, which can effectively identify the areas with abnormal density, that is, the possible crack positions.

[0122] Finally, the method of the present invention sets the number of samples in the minimum clustering cluster and the density of the minimum clustering cluster, and uses the DBSCAN clustering algorithm to cluster the point cloud, and marks all the clustered areas as cracks. The two key parameters of the DBSCAN algorithm are eps (neighborhood radius) and minPts (minimum number of points). For bridge crack identification, eps can be selected as 2mm - 5mm, and minPts can be set as 5 - 10 points. The selection of these parameters needs to be adjusted according to the specific point cloud density and the expected crack size.

[0123] Through the above steps, the method of the present invention can effectively identify and extract crack information from the point cloud data, providing reliable basic data for subsequent crack evolution analysis. This density-based clustering method can better process three-dimensional point cloud data, adapt to complex bridge surface structures, and improve the accuracy and robustness of crack identification compared with traditional image processing methods. In the improved ICP algorithm of the present invention, the principal component analysis method is used to find the point planes of the two groups of point clouds, and this step is crucial for improving the registration efficiency. Specifically, the method obtains the plane normal vector of the point cloud from point cloud A and uses this normal vector as the initialization direction for registration. This method can quickly determine the main direction of the point cloud, providing a good starting point for subsequent fine registration.

[0124] Preferably, the method of the present invention adjusts the direction of the normal vector to the z-axis direction. This step is achieved through coordinate transformation, which can significantly simplify the subsequent calculation process. Then, the method uses the normal vector combined with the SVD (singular value decomposition) method to obtain the rotation transformation matrix. SVD decomposition can effectively handle noise and outliers, improving the robustness of registration. In practical applications, truncated SVD can be used to improve the calculation efficiency, and usually retaining the first 3 singular values can meet the accuracy requirements.

[0125] The method of the present invention then calculates the overlapping area of the two groups of point clouds through the registration transformation. This step is crucial for accurately evaluating the registration quality. The calculation of the overlapping area can use the kd-tree structure to accelerate the nearest neighbor search, and the threshold can be set as 2 - 3 times the average point spacing of the point cloud. Preferably, the method also introduces a dynamic threshold strategy, gradually reducing the threshold as the number of iterations increases to balance between coarse registration and fine registration.

[0126] Finally, the method of the present invention solves the least squares method for overlapping regions to minimize the distance between corresponding points and solves the parameters of the registration transformation. Here, the point-to-plane error metric is adopted, which can better handle the bridge surface with rich planar structures compared to the point-to-point error metric. Preferably, the present method also introduces the Huber loss function to reduce the influence of outliers and improve the robustness of registration.

[0127] In the crack evolution analysis model based on point cloud convolutional neural network and attention mechanism, the method of the present invention first uses the point cloud convolutional neural network algorithm to extract the features of the registered point cloud by using a multi-layer convolutional neural network for the registered point cloud. Here, the PointNet++ architecture is used, which can effectively process local features of different scales through hierarchical sampling and grouping operations. In the network design, the present method uses 4 SetAbstraction layers, and the number of sampled points gradually decreases from 8192 to 256. Each layer uses 3 MLP layers, and the number of channels is [32, 32, 64], [64, 64, 128], [128, 128, 256], [256, 256, 512] in sequence.

[0128] Next, the method of the present invention uses an LSTM (Long Short-Term Memory) neural network to construct a temporal evolution model of cracks for crack recognition based on the crack features on point clouds at different time series. The hidden layer dimension of the LSTM network is set to 256, and the number of layers is 2, which can effectively capture the long-term dependence relationship of crack evolution. Preferably, the present method also introduces a bidirectional LSTM structure to simultaneously consider the forward and backward temporal information and improve the prediction accuracy of the model.

[0129] Finally, the method of the present invention uses the attention mechanism to weight and filter the crack features at different time series to obtain the evolution law and change trend of cracks. Here, the multi-head self-attention mechanism is used, and the number of heads is set to 8, and the dimension of each head is 32. Through the attention mechanism, the model can adaptively focus on the important features at different time points and spatial positions, improving the ability to capture the crack evolution trend.

[0130] The method of the present invention also includes setting the change of the crack area in three registration periods to determine the growth rate of the cracks. This step is crucial for evaluating the development speed of the cracks. Specifically, the present method first calculates the crack area in each registration period, and then compares the area changes between adjacent periods. Preferably, the relative change rate can be used to measure the growth rate of the cracks, that is, (A2 - A1) / A1 and (A3 - A2) / A2, where A1, A2, and A3 respectively represent the crack areas in three consecutive registration periods.

[0131] Based on the above method, the present invention also proposes a system for analyzing the temporal evolution of bridge cracks based on dynamic point cloud registration. The system includes a point cloud registration module 1, a crack extraction module 2, a crack analysis module 3, and a growth rate trend evaluation module 4.

[0132] The point cloud registration module 1 is used to roughly register the binocular three-dimensional point cloud data with the standard three-dimensional point cloud data, and then use the ICP iterative algorithm to accurately register the binocular three-dimensional point cloud data with the standard three-dimensional point cloud data to obtain the registered three-dimensional point cloud data. In practical applications, the point cloud registration module 1 can be further divided into a rough registration sub-module 11 and a fine registration sub-module 12. The rough registration sub-module 11 adopts a feature-based registration method, such as FPFH (Fast Point Feature Histograms) feature matching, to quickly obtain the initial registration result. The fine registration sub-module 12 then uses an improved ICP algorithm, as described above, to achieve high-precision point cloud registration.

[0133] The crack extraction module 2 is used to obtain the crack area of the registered three-dimensional point cloud data. This module can be further divided into a feature extraction sub-module 21 and a crack identification sub-module 22. The feature extraction sub-module 21 calculates the local geometric features of the point cloud, such as the normal vector and curvature. The crack identification sub-module 22 then uses a density clustering method to identify potential crack areas based on these features.

[0134] The crack analysis module 3 is used to analyze the crack width, crack length, crack depth, and crack propagation trend of the real crack area. This module can be further divided into a feature measurement sub-module 31 and a trend analysis sub-module 32. The feature measurement sub-module 31 is responsible for accurately measuring the geometric parameters of the crack, while the trend analysis sub-module 32 analyzes the development trend of the crack based on the point cloud convolutional neural network and the attention mechanism.

[0135] The growth rate trend evaluation module 4 is used to set the change in the crack area for three registration periods. This module includes an area calculation sub-module 41 and a speed evaluation sub-module 42. The area calculation sub-module 41 is responsible for accurately calculating the crack area in each registration period, while the speed evaluation sub-module 42 calculates the growth rate of the crack based on the area change and evaluates its development trend.

[0136] Through the collaborative work of these modules, the system of the present invention can comprehensively and accurately analyze the temporal evolution process of bridge cracks, providing an important basis for the health monitoring and maintenance decision-making of bridges.

[0137] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A bridge crack temporal evolution analysis method based on dynamic point cloud registration, characterized in that: include: The acquisition steps include: Collect surface laser point cloud data of the bridge; Construct a multi-temporal three-dimensional image sample library; Processing steps include: Based on the surface laser point cloud data, normal vectors and curvature are calculated to identify cracks in the point cloud; The improved ICP algorithm is used to register the bridge surface point clouds of different time series; Output steps include: Based on the registered point cloud, the point cloud convolutional neural network and attention mechanism are used to analyze the development trend of bridge cracks, and the crack extension prediction and crack damage analysis diagram are obtained.

2. The method according to claim 1, characterized in that The identifying cracks in the point cloud specifically includes: Calculate the normal vector and curvature in the neighborhood of each point in the 3D surface laser point cloud; All point clouds on the bridge are clustered using the density clustering method, and each cluster area is recorded as a crack.

3. The method according to claim 1, characterized in that: The construction of a multi-phase three-dimensional image sample library specifically includes: Use a drone mounted with a laser radar to collect three-dimensional surface data of the bridge from above; A portable data recorder is used to record the point cloud data of the three-dimensional point cloud data at time t, the actual position, length and width of the crack; Input the collected point cloud data into the point cloud recognition system to identify cracks, and output the point cloud crack recognition results of the bridge surface, including the location, length and width of the cracks; The crack information on the bridge surface and the crack identification information on the point cloud are matched and analyzed, and a bridge crack analysis model with multi-temporal point cloud registration is established to record the evolution law and change trend of the cracks.

4. The method according to claim 1, characterized in that: The use of point cloud convolutional neural network and attention mechanism to analyze the development trend of bridge cracks specifically includes: Establish a crack evolution analysis model based on point cloud convolutional neural network and attention mechanism; The registered data point cloud and crack identification results are used to train and test the crack time series evolution model to obtain the optimized model; The multi-time point cloud is input into the optimization model, and the crack propagation prediction and crack damage analysis diagram are output.

5. The method according to claim 1, characterized in that The improved ICP algorithm specifically includes: Perform pre-processing of the collected point cloud data by denoising, downsampling, resampling, and plane extraction to screen feature points; Calculate the normal vector and curvature feature information of the feature points; Using the ICP algorithm, the minimum distance between corresponding points is iteratively calculated to obtain the transformation relationship of the point cloud and complete the point cloud registration.

6. The method according to claim 2, characterized in that The density clustering method specifically includes: Use the Euclidean formula to calculate the number of neighboring points of each point p(x,y,z) in the point cloud; The points whose calculated point cloud density is greater than the density threshold are marked as points in the crack area; The number of samples and the density of the minimum cluster are set, the DBSCAN clustering algorithm is used to cluster the point cloud, and all the clustered areas are marked as cracks.

7. The method according to claim 5, characterized in that The improved ICP algorithm also includes: The principal component analysis is used to find the point planes of the two groups of point clouds, the plane normal vector of the point cloud is obtained from point cloud A, and the normal vector is used as the initialization direction for registration; Adjust the direction of the normal vector to the z-axis direction, and use the normal vector combined with the SVD decomposition method to obtain the rotation transformation matrix; Calculate the overlapping area of ​​two sets of point clouds through registration transformation; Solve the least squares method in the overlapping area to minimize the distance between corresponding points and solve the parameters of the registration transformation.

8. The method according to claim 4, characterized in that The crack evolution analysis model based on point cloud convolutional neural network and attention mechanism specifically includes: The point cloud convolutional neural network algorithm is used to extract the features of the registered point cloud using a multi-layer convolutional neural network. The LSTM neural network is used to construct a crack temporal evolution model based on the crack features on point clouds of different time series to identify cracks. The attention mechanism is used to weight and filter the crack features of different time series to obtain the evolution law and change trend of the cracks.

9. The method according to any one of claims 1 to 8, characterized in that: It also includes setting three registration cycles to determine the growth rate of the cracks based on the changes in the crack area.

10. A bridge crack temporal evolution analysis system based on dynamic point cloud registration that executes the method of claim 9, characterized in that: The system includes: The point cloud registration module is used to roughly register the binocular 3D point cloud data with the standard 3D point cloud data, and then use the ICP iterative algorithm to accurately register the binocular 3D point cloud data with the standard 3D point cloud data to obtain the registered 3D point cloud data; Crack extraction module, used to obtain the crack area of ​​the registered 3D point cloud data; Crack analysis module, used to analyze the crack width, crack length, crack depth, and crack expansion trend of the real crack area; Growth rate trend evaluation module is used to set the changes of crack area in three registration cycles.

Images