Underground engineering lining disease detection system based on point cloud

Abstract

The invention discloses an underground engineering lining disease detection system based on point cloud, and belongs to the technical field of underground engineering detection. In order to solve the technical problems that an existing underground engineering lining disease detection method is low in detection precision, low in automation degree and the like, underground engineering point cloud data to be detected and corresponding position information are collected, and an improved PointNet + + model is adopted for disease recognition. According to the method, original three-dimensional coordinates of a point cloud are expanded into seven-dimensional point cloud data containing coordinates, normal vectors and reflection intensity, the normal vector standard deviation of points in a neighborhood of each candidate point is calculated to serve as local geometric complexity, a local geometric complexity index is fused into sampling distance measurement, then a multi-scale local neighborhood is constructed by combining sphere query, and therefore the multi-scale local neighborhood is obtained. And extracting geometric features, texture features and deformation features by using a PCA feature dimension reduction technology, carrying out feature fusion based on a normal vector weighting mechanism, finally obtaining disease type classification based on a network model, and calculating the size and position of the disease.

Description

Technical Field

[0001] This invention belongs to the field of underground engineering detection technology, specifically relating to a system and method for detecting defects in the lining of underground engineering. Background Technology

[0002] The 21st century is an era of booming development and utilization of underground engineering, with countries worldwide increasingly emphasizing the development and utilization of underground space resources. In recent years, with the acceleration of economic development and urbanization, my country has achieved remarkable success in underground engineering construction. In hydropower engineering, according to incomplete statistics, the total length of completed underground hydraulic engineering projects in my country is nearly 1000 km, including 120 underground powerhouses. In transportation engineering, by the end of 2020, the cumulative operating mileage of railways, highways, and subways reached 19630, 21999.3, and 6280.8 km, respectively. my country has become the country with the largest number of underground engineering projects, the fastest development speed, and the most complex geological and structural forms in the world. With the acceleration of infrastructure construction, high-altitude, high-intensity earthquake zones, and deep-buried projects will become increasingly common. From the perspective of resource extraction, deep resource extraction will become the norm in the future, which places higher demands on the safety monitoring of underground engineering projects. During long-term service, underground engineering projects are affected by multiple factors such as complex geological environments, groundwater erosion, temperature changes, and traffic loads, inevitably leading to various defects such as cracks and water leakage in the lining structure. Statistics show that more than 80% of underground projects begin to show varying degrees of lining defects after 5 years of operation. These defects not only affect the normal use of the project, but may also lead to major safety accidents such as structural instability.

[0003] Traditional methods for detecting geological defects mainly rely on manual visual inspection, supplemented by technologies such as fiber optic sensors or ground-penetrating radar. However, in complex underground environments, these methods suffer from low efficiency, high subjectivity, high safety risks, and limited accuracy. Laser point cloud technology, as a high-precision three-dimensional measurement method, can accurately acquire surface information of underground structures through laser scanning and generate high-quality point cloud data, providing a new technical approach for defect detection.

[0004] However, existing laser scanning-based methods for detecting defects in underground engineering still suffer from problems such as relatively simple data processing algorithms, lack of systematic feature extraction, insufficient positioning accuracy, and low level of intelligence. Summary of the Invention

[0005] This invention aims to address the technical problems of low detection accuracy and low automation in current methods for detecting defects in underground engineering linings.

[0006] A point cloud-based system for detecting defects in underground engineering linings includes:

[0007] Data acquisition module: used to collect point cloud data and corresponding location information of the underground engineering to be inspected;

[0008] Data preprocessing module: performs noise filtering, coordinate registration, and data format conversion on the 3D point cloud data;

[0009] Disease Feature Extraction Module: This module expands the original 3D coordinates of the point cloud into 7D point cloud data containing 3D coordinates, 3D normal vector information, and reflection intensity. It then uses a feature extraction network to extract features, which includes:

[0010] The first layer, Set Abstraction Unit, samples the augmented point cloud containing coordinates and normal vectors. During sampling, the standard deviation of the normal vectors of the neighborhood points of each candidate point is calculated as the local geometric complexity, and sampling is performed based on the sampling distance incorporating the local geometric complexity. Subsequently, based on the sampled center points, PointNet++'s Set Abstraction architecture is used to extract local feature vectors for each point at multiple scales. A normal vector weighting mechanism is introduced during feature aggregation at each resolution branch. Distance weight, normal vector consistency weight, and normal vector variance weight are calculated for the points in the local neighborhood of each resolution branch, and the three weights are multiplied to obtain the surrounding weights of each point in the neighborhood. Then, the weights of the surrounding neighborhood points of each center point are normalized. This normalized weight matrix is ​​used to perform weighted summation of the local feature vectors of the neighborhood points to obtain the local aggregated features of the center point at that scale. Then, the local aggregated features of the three scale branches are concatenated and fed into a one-dimensional convolutional layer to reduce the dimensionality and obtain the concatenated and reduced aggregated features. The first layer, Set Abstraction Unit, finally outputs the coordinates, normal vectors, and corresponding concatenated and reduced aggregated feature vectors of N1 sampled points.

[0011] Disease Attention Unit: Based on the sampling points finally output by the first layer set abstract unit, the crack confidence score and the leakage confidence score are obtained. Then, the two confidence scores are weighted and combined to obtain the comprehensive score of the center point, and then multiplied element-wise with the spliced ​​dimension reduction aggregated feature vector to obtain the enhanced feature.

[0012] The second layer abstract unit takes the N1 sampling points and their enhanced features output by the disease attention unit as input. First, the spatial coordinates of these N1 sampling points are sampled from the farthest point to select N2 points as new center points. Then, a spherical third neighborhood radius is set for each center point. For each center point, the features of the N3 neighboring points within the spherical neighborhood of its three-neighbor radius are aggregated: each neighboring point provides a multi-dimensional vector as input data. The dimension of the multi-dimensional vector is the dimension of the enhanced feature + the 3-dimensional relative coordinates of the neighboring point relative to the center point. The input is extracted into M1-dimensional features by the fourth network model. Then, the M1-dimensional features of the N3 neighboring points are weighted and pooled through the normal vector weighting mechanism to obtain the final M1-dimensional feature representation of the center point. Finally, N2 high-level semantic points and their M1-dimensional feature description vectors are output.

[0013] Deep feature dimensionality reduction unit: Principal component analysis is performed on the M1-dimensional feature description vector output by the second-layer set abstraction unit to reduce the dimensionality. Principal components are selected by eigenvalue decomposition. For N2 points, the M1-dimensional feature description vector of each point is reduced to M2 dimensions by PCA, i.e., multimodal feature vector.

[0014] Disease identification module: Receives the multimodal feature vectors output by the disease feature extraction module and classifies the disease types.

[0015] Furthermore, improvements to the input point cloud are achieved through an information augmentation unit. The process by which the information augmentation unit expands the original 3D coordinates of the point cloud includes:

[0016] By normalizing the coordinates and calculating the normal vectors of the input tunnel lining point cloud data, and constructing the covariance matrix of the k-neighborhood point set of each point and performing eigenvalue decomposition, the normal vector information representing the local surface direction is obtained. The normal vector of the local surface direction is used as the normal vector of the corresponding point on the local surface. Then, the original three-dimensional coordinates are expanded into seven-dimensional point cloud data containing coordinates, normal vectors and reflection intensity.

[0017] Furthermore, the first-level set abstraction unit calculates the standard deviation of the normal vector of each candidate point's neighborhood as the local geometric complexity. The sampling process based on the sampling distance incorporating the local geometric complexity includes:

[0018] The standard deviation of the normal vectors of the points in the neighborhood of each candidate point is calculated as the local geometric complexity, and this local geometric complexity index is incorporated into the sampling distance metric to obtain the comprehensive score of the distance metric. , where d represents the minimum Euclidean distance from the selected point set to the candidate point i, and i represents the standard deviation of the normal vector in the neighborhood of point i, which represents the local geometric complexity; This represents the complexity weighting coefficient; point cloud sampling is performed based on a comprehensive score derived from distance metrics.

[0019] Furthermore, the process of using PointNet++'s Set Abstraction architecture to perform multi-scale feature extraction to obtain the local feature vector for each point includes:

[0020] Three parallel branches are constructed around each center point. The small-scale branch sets a spherical neighborhood with a radius of 0.05 meters containing 16 neighboring points to capture subtle texture features. The medium-scale branch sets a radius of 0.1 meters containing 32 neighboring points to identify the main shape of the crack. The large-scale branch sets a radius of 0.2 meters containing 64 neighboring points to perceive surrounding contextual information. Each scale branch maps the seven-dimensional input data, which is composed of coordinates, normal vector, and reflection intensity, into a 64-dimensional local feature vector through a standard PointNet module.

[0021] Furthermore, the process of calculating the distance weight, normal vector consistency weight, and normal vector variance weight includes:

[0022] Distance weights: Based on the distance between points, distance weights are calculated using a Gaussian kernel function of the distance.

[0023] Normal vector consistency weight: Calculates the cosine similarity between the normal vectors of neighboring points and the normal vector of the center point. , The angle between the two normal vectors is used; then, the consistency weight of the normal vectors is calculated based on the cosine similarity. ,in For coefficients;

[0024] Normal vector variance weight: The standard deviation of the normal vector within the neighborhood is used as the normal vector variance weight. ,in It is the set of normal vectors within the domain.

[0025] Furthermore, in the disease attention unit, the process of obtaining the crack confidence score and leakage confidence score based on the sampling points finally output by the first-level set abstraction unit includes:

[0026] S221. Generate the crack confidence score for each sampling point:

[0027] Based on the N1 sampling points finally output by the first layer set abstract unit, a spherical first neighborhood radius is set around each sampling point. Neighboring points within this neighborhood are selected, and the standard deviation of the normal vectors of these neighboring points is calculated. The standard deviation of the normal vectors corresponding to the neighboring points is input into the first network model and processed by the Sigmoid activation function to generate the crack confidence score of each sampling point.

[0028] S222. Generate the leakage confidence score for each point:

[0029] The reflection intensity anomaly α2 = max(0,TI) / T is calculated for each point based on the reflection intensity I, where T is the reflection intensity threshold. A second spherical neighborhood radius is set around each suspected leakage point. The principal direction vector v is obtained by principal component analysis for the spatial coordinates of the points in the neighborhood. The angle θ between the principal direction and the gravity direction g is calculated, and the directional score β2 = |cos(θ2)| is defined. The two features, reflection intensity anomaly α2 and directional score β2, are concatenated into a two-dimensional vector, input into the second network model and activated by Sigmoid to generate the leakage confidence score for each point.

[0030] S223. Generate the overall score for each point:

[0031] The crack confidence score and leakage confidence score of the center point and each point in the surrounding area are averaged separately and used as the crack confidence score and leakage confidence score of the center point. Then, the crack confidence score and leakage confidence score of the center point are concatenated with the dimensionality-reduced aggregated feature vector of the points obtained from the first layer set abstract unit. Then, the vector is input into the third network model and normalized by Softmax to obtain two dynamic weights. Finally, the two confidence scores are weighted and combined to obtain the comprehensive score of the center point.

[0032] Furthermore, the radius of the first neighborhood is 0.1 meters, and the radius of the second spherical neighborhood is 0.15 meters.

[0033] Furthermore, the radius of the third neighborhood is 0.2 meters.

[0034] Furthermore, the disease identification module's process of classifying disease types includes:

[0035] Global Feature Aggregation Unit: For each element in the multimodal feature vector output by the disease feature extraction module, max pooling is performed on point An and its neighborhood points to obtain the value of the corresponding element of An, thereby obtaining the feature vector corresponding to point An with the same dimension as the multimodal feature vector, denoted as the global feature vector.

[0036] Multi-layer classifier unit: The fifth network model is used to perform non-linear mapping on the global feature vector. The classifier output is converted into the probability distribution of each disease category through the Softmax function, and then the category identification result is obtained.

[0037] Furthermore, the point cloud-based underground engineering lining defect detection system also includes a defect calculation module; the defect calculation module is used to calculate the geometric dimensions and distribution location of the identified defects.

[0038] Beneficial effects:

[0039] This invention provides a point cloud-based underground engineering defect detection system, which utilizes intelligent algorithms to achieve automated and accurate defect detection, improves the real-time monitoring and safety management capabilities of underground engineering, and provides technical support for ensuring the safe operation of underground engineering. Attached Figure Description

[0040] Figure 1 Block diagram of a point cloud-based underground engineering lining defect detection system;

[0041] Figure 2 Flowchart for detecting defects in underground engineering lining based on a detection system;

[0042] Figure 3 Schematic diagram of the controller;

[0043] Figure 4 Accuracy curves for different methods;

[0044] Figure 5 This is a diagram showing the detection results of the present invention. Detailed Implementation

[0045] The purpose of this invention is to provide a point cloud-based underground engineering lining defect detection system to solve the technical problems of low detection accuracy, poor efficiency and low degree of automation in the prior art, and to achieve high-precision, automated detection and quantitative analysis of underground engineering lining defects.

[0046] Combination Figure 1 and Figure 2 This embodiment describes a point cloud-based system for detecting defects in underground engineering linings, which mainly includes:

[0047] Data acquisition module: includes an inspection vehicle equipped with a 3D laser scanner, a UWB ultra-wideband positioning system, a controller, and a data transmission submodule.

[0048] A 3D laser scanner is used to emit laser beams and receive reflected signals to obtain spatial coordinate information of the lining surface;

[0049] Inspection vehicles are used to carry laser scanning equipment and move it autonomously or remotely within underground engineering projects.

[0050] The UWB ultra-wideband positioning system, including UWB base stations and tags, is used to determine the precise location coordinates of inspection vehicles in underground engineering projects in real time, with positioning accuracy down to the centimeter level.

[0051] The controller includes a point cloud information acquisition unit, a UWB ultra-wideband positioning information acquisition unit, and an inspection and control unit, such as... Figure 3 As shown, the point cloud information acquisition unit is used for point cloud acquisition, the UWB ultra-wideband positioning information acquisition unit is used for acquiring location information, and the inspection control unit is used for the movement control of the inspection vehicle.

[0052] The information acquisition unit includes a 3D laser scanner, an inspection vehicle, and a UWB ultra-wideband positioning system.

[0053] The data transmission submodule is used for data transmission.

[0054] The data acquisition module's processing steps include:

[0055] Three-dimensional point cloud data of the lining surface of underground engineering is obtained by laser scanning equipment mounted on the inspection vehicle.

[0056] The scanning parameters of the laser scanning equipment are set, and the inspection vehicle equipped with the laser scanning equipment is controlled to move within the underground project at a fixed speed (V=0.5-1m / s) along a predetermined path. This ensures the density and quality of the point cloud data acquisition. The onboard control system monitors the operational status in real time to guarantee uniform linear motion. Simultaneously, the UWB positioning system is used to obtain the precise location information of the inspection vehicle in real time, performing a full-coverage scan of the lining surface to acquire complete point cloud data containing spatial coordinates, reflection intensity, and color information.

[0057] During the real-time acquisition of the inspection vehicle's precise location information using the UWB positioning system, the CPU acquires the UWB ultra-wideband positioning system's location information at a frequency of 100Hz in real time, establishing a timestamp synchronization mechanism to ensure that the location information is synchronized with the laser scanning data. The specific formula is as follows: ,in This indicates the time it takes for the ultra-wideband positioning system to acquire the location information of the inspection vehicle. This indicates the time it takes for the lidar to complete one frame of scanning and record point cloud data. This represents the fixed time difference between the two.

[0058] Inspection vehicles equipped with laser scanners move along a predetermined path within the underground engineering project.

[0059] Data preprocessing module: performs noise filtering, coordinate registration, and data format conversion on the 3D point cloud data.

[0060] The data preprocessing module's processing steps include:

[0061] First, point cloud data is registered and stitched together. Coarse registration is performed based on UWB location information, and fine registration is performed using the ICP algorithm to obtain the final transformation matrix.

[0062] Then, noise filtering and quality control are performed. Statistical filters are used to remove outliers, a standard deviation multiple threshold is set, points that meet the conditions are retained, and dynamic filtering parameters are adjusted based on the running speed.

[0063] Disease feature extraction module: Feature extraction is performed based on an improved PointNet++ deep learning model, including:

[0064] Information enhancement unit: By normalizing the coordinates and calculating the normal vectors of the input tunnel lining point cloud data, constructing the covariance matrix of the k-neighborhood point set of each point and performing eigenvalue decomposition, the normal vector information representing the local surface direction is obtained. The normal vector of the local surface direction is used as the normal vector (three-dimensional vector) of the corresponding point on the local surface. The original three-dimensional coordinates are expanded into seven-dimensional point cloud data containing coordinates, normal vectors and reflection intensity.

[0065] First-level set abstraction (sa level) unit:

[0066] Progressive feature abstraction is performed on augmented point clouds containing coordinates and normal vectors. First, 1024 representative points are selected as center points using an improved farthest point sampling strategy. This improved strategy calculates the standard deviation of the normal vectors of the neighboring points of each candidate point as the local geometric complexity, and incorporates this local geometric complexity index into the sampling distance metric to construct a comprehensive scoring function. , where d represents the minimum Euclidean distance from the selected point set to the candidate point i, and i represents the standard deviation of the normal vector in the neighborhood of point i, which represents the local geometric complexity; This represents the complexity weighting coefficient. This prioritizes regions with significant geometric features when selecting the farthest point; generally, when geometric features are significant, the standard deviation of the normal vector will be >0.15. This results in higher sampling density in geometrically significant regions such as crack edges, increasing the sampling density in critical regions by approximately 2.5 times compared to the standard FPS method.

[0067] Subsequently, based on the sampled center points, multi-scale feature extraction is performed using the PointNet++ Set Abstraction architecture: three parallel branches are constructed around each center point. The small-scale branch sets a spherical neighborhood with a radius of 0.05 meters containing 16 neighboring points to capture subtle texture features; the medium-scale branch sets a radius of 0.1 meters containing 32 neighboring points to identify the main crack shape; and the large-scale branch sets a radius of 0.2 meters containing 64 neighboring points to perceive surrounding contextual information. Each scale branch maps the seven-dimensional input data (coordinates, normal vector, and reflection intensity) into a 64-dimensional local feature vector through a standard PointNet module (shared MLP: 7→32→32→64 + Max Pooling).

[0068] A normal vector weighting mechanism is introduced during the feature aggregation process of each resolution branch. Three weights are calculated for points in the local neighborhood of each resolution branch and then combined:

[0069] Distance weights: Gaussian kernel function is used. Ensure that spatially nearest points receive higher weight;

[0070] Normal vector consistency weight: Calculates the cosine similarity between the normal vectors of neighboring points and the normal vector of the center point. , Let be the angle between the two normal vectors; the specific calculation is as follows: for the normal vector nc of the center point c and the normal vector np of the neighboring point p: Then, the consistency weight of the normal vector is calculated based on cosine similarity. This is to ensure that points with similar surface orientations contribute more; among which is a coefficient.

[0071] Normal vector variance weight: The standard deviation of the normal vector within the neighborhood is used as the normal vector variance weight. To identify regions with high geometric complexity; among which It is the set of normal vectors within the domain;

[0072] Based on distance weight, normal vector consistency weight, and normal vector variance weight, the comprehensive weight of each neighboring point is obtained by element-wise multiplication as follows: For each point in the neighborhood, three weights are obtained, and then the three weights are multiplied to obtain the surrounding weights of each point in the neighborhood; then the weights of the surrounding neighboring points of each center point (a total of 1024) are normalized (so that their sum is 1). The normalized weight matrix is ​​used to perform weighted summation of the 64-dimensional local feature vectors of the neighboring points, thereby obtaining the 64-dimensional local aggregated features of the center point at this scale. This makes geometrically similar points obtain higher contribution weights in feature extraction.

[0073] The output features of the three scale branches (64-dimensional local aggregated features of each scale branch) are fused into a 192-dimensional feature vector through a concatenation operation, and then reduced to 64-dimensionality through a one-dimensional convolutional layer.

[0074] The first-level abstract unit ultimately outputs the coordinates, normal vectors, and corresponding 64-dimensional feature vectors of 1024 sampling points.

[0075] Disease Attention Unit:

[0076] S221. Generate the crack confidence score for each sampling point:

[0077] Based on the 1024 sampling points output by the first-layer ensemble abstraction unit, a spherical neighborhood with a radius of 0.1 meters is set around each sampling point. The standard deviation of the normal vectors of the 32 nearest neighbors within this neighborhood is calculated to reflect the degree of drastic change in the surface normal vector. At the crack edge, the normal vector changes rapidly due to abrupt changes in surface geometry, resulting in a larger standard deviation; while the normal vectors of a normal surface are evenly distributed, resulting in a smaller standard deviation. The standard deviations of the normal vectors corresponding to the 32 nearest neighbors are input into a three-layer multilayer perceptron network (hidden layer dimensions 32→16→1) and processed by a sigmoid activation function to generate a crack confidence score for each sampling point.

[0078] S222. Generate the leakage confidence score for each point:

[0079] Because moisture absorbs laser energy, the reflection intensity in the leaking area is significantly lower than that on a dry concrete surface. Statistical analysis shows that the reflection intensity in the normal area is between 80 and 100, while that in the leaking area is typically below 50. The anomaly degree of reflection intensity α2 = max(0,TI) / T is defined to measure the degree of intensity deviation at each point. When the reflection intensity I at a point is lower than the threshold T (set to 60), the anomaly degree α2 is greater than 0; the lower the reflection intensity, the higher the anomaly degree.

[0080] Directional constraint: Leaking water flows vertically under the influence of gravity, forming a vertical strip-like distribution pattern. A spherical neighborhood with a radius of 0.15 meters is set around each suspected leakage point. The spatial coordinates of the points within the neighborhood are used to obtain the principal direction vector v through principal component analysis. The angle θ between this principal direction and the direction of gravity g is calculated, and the directional score β2 = |cos(θ2)| is defined. The larger the score, the more it conforms to the vertical flow characteristics of water.

[0081] The two features, reflection intensity anomaly α2 and directionality score β2, are concatenated into a two-dimensional vector, input into a two-layer multilayer perceptron network (2→8→1) and activated by Sigmoid to generate the leakage confidence score for each point.

[0082] S223. Generate the overall score for each point:

[0083] The crack confidence score and leakage confidence score of the center point and each point within its vicinity are averaged to obtain the crack confidence score and leakage confidence score of the center point. Then, the crack confidence score and leakage confidence score of the center point are concatenated with the 64-dimensional feature vectors of the points obtained from the first layer's ensemble abstraction unit. These are then input into a weighted generation network (using a network model; in this embodiment, an MLP is employed) and normalized using Softmax to obtain two dynamic weights. Finally, these two weighted confidence scores are combined to obtain the comprehensive score of the center point.

[0084] The final composite score is element-wise multiplied with the 64-dimensional feature vector to obtain the enhanced 64-dimensional feature. This operation ensures that the feature response of the diseased areas is amplified, while the features of the normal areas are relatively suppressed. The enhanced feature is then passed to the second-level set abstraction unit for further high-level semantic abstraction.

[0085] Second-level abstract unit:

[0086] The 1024 sampling points output by the disease attention unit and their enhanced 64-dimensional features are used as input. First, the spatial coordinates of these 1024 sampling points are sampled from the farthest point to select 512 points as new center points. Then, a spherical neighborhood with a radius of 0.2 meters is set for each center point, and at most 64 neighboring points are found within each of these 512 points. For each center point, the features of its 64 neighboring points are aggregated: each neighboring point provides 67-dimensional input data (its own 64-dimensional features + its 3-dimensional relative coordinates with respect to the center point). This 67-dimensional input is extracted into 128-dimensional features by an MLP network (containing two 64-dimensional hidden layers). Then, the 128-dimensional features of the 64 neighboring points are weighted and pooled using a normal vector weighting mechanism to obtain the final 128-dimensional feature representation of the center point. The normal vector weighting mechanism continues to play a role in this layer to ensure that geometric consistency is maintained in multi-scale feature extraction. The final output consists of 512 high-level semantic points and their 128-dimensional feature description vectors. These feature description vectors implicitly encode multi-scale geometric information, normal vector distribution patterns, local roughness variations, and density features.

[0087] Deep feature dimensionality reduction unit:

[0088] Principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional latent feature vectors output by the second-level set abstraction unit. Principal components with a cumulative variance contribution rate of 95% are selected by eigenvalue decomposition. For 512 points, the 128-dimensional feature vectors of each point are reduced to 64 dimensions by PCA, which is a multimodal feature vector. This improves the efficiency of subsequent calculations while maintaining the feature discrimination ability.

[0089] The entire "Disease Feature Extraction Module" is essentially the SA layer (Sample Abstraction Layer) of the PointNet++ model. It increases dimensionality by reducing the number of points. The first layer is the first layer of abstraction units mentioned above, and the second layer is the second layer of abstraction units; both are internal components of the PointNet++ model. Only the first layer's input is the original point cloud (N points × 7 dimensions). Subsequent layers input features from the previous layer; for example, the second layer's input is 1024 points (128 dimensions) from the first layer's output. Then, the PointNet++ network's FP layer increases the number of points, and the high-dimensional features are fed back to the original point cloud to complete classification and segmentation.

[0090] Disease identification module: Receives the multimodal feature vector output from the disease feature extraction module, and classifies disease types using supervised learning methods, including:

[0091] Global Feature Aggregation Unit: For each element in the multimodal feature vector output by the disease feature extraction module, max pooling is performed on point An and its neighborhood points (32 points are selected in this implementation) to obtain the value of the corresponding element of An, thereby obtaining the feature vector corresponding to point An with the same dimension as the multimodal feature vector (actually a 64-dimensional feature vector), which is denoted as the global feature vector.

[0092] Multi-layer classifier unit: A three-layer fully connected neural network is used to perform non-linear mapping on the global feature vector. The first layer maps the input features to 256 dimensions and activates them with ReLU and regularization with Dropout. The second layer maps them to 128 dimensions. The output of the third layer is the dimension of the disease category. Probability output unit: The classifier output is converted into the probability distribution of each disease category through the Softmax function. The category with the highest probability is selected as the recognition result, and the confidence score is output at the same time.

[0093] The recognition effect of this invention compared to other methods is as follows: Figure 4 As shown, the recognition effect of the present invention is superior.

[0094] Disease calculation module: Calculates the geometric dimensions, distribution location, and severity of identified diseases;

[0095] (1) Crack length calculation:

[0096] Extracting point cloud sets of crack regions ,in Indicates coordinates as point, The index of the point. This represents the number of points in the point cloud set.

[0097] Calculate the centroid of the point cloud: ;

[0098] Construct the covariance matrix: ;

[0099] Perform eigenvalue decomposition on the covariance matrix: ,in For eigenvalues, The eigenvectors are the first principal components; the direction of the first principal component (the eigenvector corresponding to the largest eigenvalue) is the principal direction. ;

[0100] The crack length can then be obtained: .

[0101] (2) Calculation of leakage area:

[0102] The leakage point cloud is rasterized, and the number of pixels N in the leakage area is counted; the actual area is calculated as: S = N × pixel area unit.

[0103] The results output stage generates data including disease type, 3D coordinates, and geometric dimensions.

[0104] Disease display module: Displays detection results, including disease type, location, size, and severity level, on a computer terminal. The detection effect of this invention is as follows: Figure 5 As shown.

[0105] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A point cloud-based system for detecting defects in underground engineering lining, characterized in that, include: Data acquisition module: used to collect point cloud data and corresponding location information of the underground engineering to be inspected; Data preprocessing module: performs noise filtering, coordinate registration, and data format conversion on the 3D point cloud data; Disease Feature Extraction Module: This module expands the original 3D coordinates of the point cloud into 7D point cloud data containing 3D coordinates, 3D normal vector information, and reflection intensity. It then uses a feature extraction network to extract features, which includes: The first layer, Set Abstraction Unit, samples the augmented point cloud containing coordinates and normal vectors. During sampling, the standard deviation of the normal vectors of the neighborhood points of each candidate point is calculated as the local geometric complexity, and sampling is performed based on the sampling distance incorporating the local geometric complexity. Subsequently, based on the sampled center points, PointNet++'s Set Abstraction architecture is used to extract local feature vectors for each point at multiple scales. A normal vector weighting mechanism is introduced during feature aggregation at each resolution branch. Distance weight, normal vector consistency weight, and normal vector variance weight are calculated for the points in the local neighborhood of each resolution branch, and the three weights are multiplied to obtain the surrounding weights of each point in the neighborhood. Then, the weights of the surrounding neighborhood points of each center point are normalized. This normalized weight matrix is ​​used to perform weighted summation of the local feature vectors of the neighborhood points to obtain the local aggregated features of the center point at that scale. Then, the local aggregated features of the three scale branches are concatenated and fed into a one-dimensional convolutional layer to reduce the dimensionality and obtain the concatenated and reduced aggregated features. The first layer, Set Abstraction Unit, finally outputs the coordinates, normal vectors, and corresponding concatenated and reduced aggregated feature vectors of N1 sampled points. Disease Attention Unit: Based on the sampling points finally output by the first layer set abstract unit, the crack confidence score and the leakage confidence score are obtained. Then, the two confidence scores are weighted and combined to obtain the comprehensive score of the center point, and then multiplied element-wise with the spliced ​​dimension reduction aggregated feature vector to obtain the enhanced feature. The second layer abstract unit takes the N1 sampling points and their enhanced features output by the disease attention unit as input. First, the spatial coordinates of these N1 sampling points are sampled from the farthest point to select N2 points as new center points. Then, a spherical third neighborhood radius is set for each center point. For each center point, the features of the N3 neighboring points within the spherical neighborhood of its three-neighbor radius are aggregated: each neighboring point provides a multi-dimensional vector as input data. The dimension of the multi-dimensional vector is the dimension of the enhanced feature + the 3-dimensional relative coordinates of the neighboring point relative to the center point. The input is extracted into M1-dimensional features by the fourth network model. Then, the M1-dimensional features of the N3 neighboring points are weighted and pooled through the normal vector weighting mechanism to obtain the final M1-dimensional feature representation of the center point. Finally, N2 high-level semantic points and their M1-dimensional feature description vectors are output. Deep feature dimensionality reduction unit: Principal component analysis is performed on the M1-dimensional feature description vector output by the second-layer set abstraction unit to reduce the dimensionality. Principal components are selected by eigenvalue decomposition. For N2 points, the M1-dimensional feature description vector of each point is reduced to M2 dimensions by PCA, i.e., multimodal feature vector. Disease identification module: Receives the multimodal feature vectors output by the disease feature extraction module and classifies the disease types.

2. The point cloud-based underground engineering lining defect detection system according to claim 1, characterized in that, The improvement to the input point cloud is achieved through an information augmentation unit. The process by which the information augmentation unit expands the original 3D coordinates of the point cloud includes: By normalizing the coordinates and calculating the normal vectors of the input tunnel lining point cloud data, and constructing the covariance matrix of the k-neighborhood point set of each point and performing eigenvalue decomposition, the normal vector information representing the local surface direction is obtained. The normal vector of the local surface direction is used as the normal vector of the corresponding point on the local surface. Then, the original three-dimensional coordinates are expanded into seven-dimensional point cloud data containing coordinates, normal vectors and reflection intensity.

3. The point cloud-based underground engineering lining defect detection system according to claim 1, characterized in that, The first-level set abstraction unit calculates the standard deviation of the normal vector of points in the neighborhood of each candidate point as the local geometric complexity. The sampling process based on the sampling distance incorporating the local geometric complexity includes: The standard deviation of the normal vectors of the points in the neighborhood of each candidate point is calculated as the local geometric complexity, and this local geometric complexity index is incorporated into the sampling distance metric to obtain the comprehensive score of the distance metric. , where d represents the minimum Euclidean distance from the selected point set to the candidate point i, and i represents the standard deviation of the normal vector in the neighborhood of point i, which represents the local geometric complexity; This represents the complexity weighting coefficient; point cloud sampling is performed based on a comprehensive score derived from distance metrics.

4. The point cloud-based underground engineering lining defect detection system according to claim 1, characterized in that, The process of obtaining the local feature vector of each point by multi-scale feature extraction using the Set Abstraction architecture of PointNet++ includes: Three parallel branches are constructed around each center point. The small-scale branch sets a spherical neighborhood with a radius of 0.05 meters containing 16 neighboring points to capture subtle texture features. The medium-scale branch sets a radius of 0.1 meters containing 32 neighboring points to identify the main shape of the crack. The large-scale branch sets a radius of 0.2 meters containing 64 neighboring points to perceive surrounding contextual information. Each scale branch maps the seven-dimensional input data, which is composed of coordinates, normal vector, and reflection intensity, into a 64-dimensional local feature vector through a standard PointNet module.

5. The point cloud-based underground engineering lining defect detection system according to claim 1, characterized in that, The process of calculating the distance weight, normal vector consistency weight, and normal vector variance weight includes: Distance weights: Based on the distance between points, distance weights are calculated using a Gaussian kernel function of the distance. Normal vector consistency weight: Calculates the cosine similarity between the normal vectors of neighboring points and the normal vector of the center point. , The angle between the two normal vectors is used; then, the consistency weight of the normal vectors is calculated based on the cosine similarity. ,in For coefficients; Normal vector variance weight: The standard deviation of the normal vector within the neighborhood is used as the normal vector variance weight. ,in It is the set of normal vectors within the domain.

6. The point cloud-based underground engineering lining defect detection system according to claim 1, characterized in that, In the disease attention unit, the process of obtaining the crack confidence score and leakage confidence score based on the sampling points finally output by the first-level set abstraction unit includes: S221. Generate the crack confidence score for each sampling point: Based on the N1 sampling points finally output by the first layer set abstract unit, a spherical first neighborhood radius is set around each sampling point. Neighboring points within this neighborhood are selected, and the standard deviation of the normal vectors of these neighboring points is calculated. The standard deviation of the normal vectors corresponding to the neighboring points is input into the first network model and processed by the Sigmoid activation function to generate the crack confidence score of each sampling point. S222. Generate the leakage confidence score for each point: The reflection intensity anomaly α2 = max(0,TI) / T is calculated for each point based on the reflection intensity I, where T is the reflection intensity threshold. A second spherical neighborhood radius is set around each suspected leakage point. The principal direction vector v is obtained by principal component analysis for the spatial coordinates of the points in the neighborhood. The angle θ between the principal direction and the gravity direction g is calculated, and the directional score β2 = |cos(θ2)| is defined. The two features, reflection intensity anomaly α2 and directional score β2, are concatenated into a two-dimensional vector, input into the second network model and activated by Sigmoid to generate the leakage confidence score for each point. S223. Generate the overall score for each point: The crack confidence score and leakage confidence score of the center point and each point in the surrounding area are averaged separately and used as the crack confidence score and leakage confidence score of the center point. Then, the crack confidence score and leakage confidence score of the center point are concatenated with the dimensionality-reduced aggregated feature vector of the points obtained from the first layer set abstract unit. Then, the vector is input into the third network model and normalized by Softmax to obtain two dynamic weights. Finally, the two confidence scores are weighted and combined to obtain the comprehensive score of the center point.

7. The point cloud-based underground engineering lining defect detection system according to claim 6, characterized in that, The first neighborhood has a radius of 0.1 meters, and the second spherical neighborhood has a radius of 0.15 meters.

8. The point cloud-based underground engineering lining defect detection system according to claim 7, characterized in that, The radius of the third neighborhood is 0.2 meters.

9. The point cloud-based underground engineering lining defect detection system according to claim 1, characterized in that, The disease identification module classifies disease types, including: Global Feature Aggregation Unit: For each element in the multimodal feature vector output by the disease feature extraction module, max pooling is performed on point An and its neighborhood points to obtain the value of the corresponding element of An, thereby obtaining the feature vector corresponding to point An with the same dimension as the multimodal feature vector, denoted as the global feature vector. Multi-layer classifier unit: The fifth network model is used to perform non-linear mapping on the global feature vector. The classifier output is converted into the probability distribution of each disease category through the Softmax function, and then the category identification result is obtained.

10. A point cloud-based underground engineering lining defect detection system according to any one of claims 1 to 9, characterized in that, The system also includes a disease calculation module; the disease calculation module is used to calculate the geometric dimensions and distribution location of the identified diseases.

Images