Ground penetrating radar 3D point cloud disease detection method, device, equipment and medium

Abstract

The application discloses a ground penetrating radar 3D point cloud disease detection method, device, equipment and medium, and relates to the technical field of road engineering; the method comprises the following steps: stacking and converting a plurality of radar scanning profile data according to a distance to obtain a target space-time echo matrix; performing physical reconstruction and amplitude correction to obtain an amplitude correction three-dimensional point cloud; performing gradient expansion and discretization processing on the amplitude correction three-dimensional point cloud to obtain a target five-dimensional sparse point cloud; performing overlapping cutting and screening based on a sliding window parameter to obtain a plurality of three-dimensional training blocks and training a disease prediction model, so as to calculate a target disease real three-dimensional physical volume according to a predicted target disease discrete point total number and a benchmark physical voxel parameter; since the application realizes the elimination of point clouds with sudden changes between roads through wave field physical reconstruction and amplitude correction, and overcomes the pain point that small diseases are easily submerged in a background grid through sliding window overlapping cutting, thereby realizing accurate detection of underground diseases and accurate quantification of disease three-dimensional physical volume.

Description

Technical Field

[0001] This application relates to the field of road engineering technology, and in particular to a method, device, equipment and medium for detecting defects in 3D point cloud ground penetrating radar. Background Technology

[0002] In the field of non-destructive testing of roads and underground engineering, ground-penetrating radar is often used to detect hidden defects such as cavities, voids, looseness, and water accumulation inside roads. With the increasing demand for refined road maintenance and early identification of underground defects, simply determining the existence of defects is no longer sufficient to meet the requirements of engineering applications. Actual testing also needs to further obtain the three-dimensional distribution pattern of defects in underground space, the type of defects, and their actual physical volume, so as to provide a data foundation for road defect classification, maintenance plan formulation, and engineering quantity calculation.

[0003] Existing ground-penetrating radar (GPR) defect detection methods typically rely on manual interpretation of two-dimensional radar scan profile images, or inputting the two-dimensional scan profiles into two-dimensional target detection networks or two-dimensional image segmentation networks for defect identification, outputting the bounding box or segmented region of the defect on a single profile. Some three-dimensional detection methods network multiple two-dimensional survey lines into three-dimensional data through linear interpolation, spline interpolation, etc., then perform regular voxelization on the three-dimensional data and input it into a three-dimensional convolutional network for defect identification and segmentation.

[0004] The aforementioned two-dimensional detection methods disrupt the continuity of road defects in three-dimensional space, making it difficult to accurately represent the three-dimensional morphology and physical volume of the defects. Furthermore, the aforementioned three-dimensional detection methods typically rely on mathematical interpolation to stitch together data from multiple survey lines, which can easily violate the laws of electromagnetic wave propagation and scattering, leading to breaks in the hyperbolic features between survey lines. Simultaneously, in ground-penetrating radar simulations or semi-physical data, spatial misalignment can easily exist between the original defect labels and strongly reflective electromagnetic features. In cases of extreme imbalance between defect points and background points, even minute defect points can easily be submerged by a large number of background points during voxelization and network training, resulting in insufficient accuracy in defect identification and volume quantization. Summary of the Invention

[0005] The main purpose of this application is to provide a method, device, equipment and medium for detecting defects in 3D point cloud ground penetrating radar, which aims to solve the technical problems in the existing technology of 3D defect detection by ground penetrating radar, such as the characteristic break between survey lines, spatial misalignment of defect labels and the fact that small defect points are easily submerged by background points, which makes it difficult to accurately detect defects and quantify the real three-dimensional physical volume of the underground.

[0006] To achieve the above objectives, this application provides a method for detecting defects in 3D point clouds using ground-penetrating radar, the method comprising the following steps: Based on the physical survey line spacing, multiple parallel ground-penetrating radar scanning profile data are spatially stacked and depth-converted to obtain the target's three-dimensional spatiotemporal echo matrix. The target three-dimensional spatiotemporal echo matrix is ​​physically reconstructed and its amplitude corrected, and the amplitude-corrected target three-dimensional spatiotemporal echo matrix is ​​converted into an amplitude-corrected three-dimensional point cloud; The amplitude-corrected three-dimensional point cloud is subjected to physical gradient expansion and discretization to obtain the target five-dimensional sparse point cloud. Based on preset sliding window parameters, the target five-dimensional sparse point cloud is overlapped, cut, and extracted to obtain multiple three-dimensional training blocks. Based on the three-dimensional training block, a pre-constructed disease prediction model is trained. The ground-penetrating radar data to be tested is input into the trained disease prediction model to predict and statistically analyze the disease category to obtain the total number of discrete points of the target disease. Based on the reference physical voxel parameters and the total number of discrete points of the target disease, the actual three-dimensional physical volume of the target disease underground is calculated.

[0007] Furthermore, to achieve the above objectives, this application also proposes a ground-penetrating radar 3D point cloud defect detection device that applies the ground-penetrating radar 3D point cloud defect detection method described above, the ground-penetrating radar 3D point cloud defect detection device comprising: The matrix construction module is used to spatially stack and depth-convert multiple parallel ground-penetrating radar scanning profile data according to the physical survey line spacing to obtain the target's three-dimensional spatiotemporal echo matrix. The reconstruction and correction module is used to physically reconstruct and correct the amplitude of the target three-dimensional spatiotemporal echo matrix, and convert the amplitude-corrected target three-dimensional spatiotemporal echo matrix into an amplitude-corrected three-dimensional point cloud. An extended discretization module is used to perform physical gradient expansion and discretization processing on the amplitude-corrected three-dimensional point cloud to obtain the target five-dimensional sparse point cloud. The training generation module is used to overlap and cut out the target five-dimensional sparse point cloud based on preset sliding window parameters, and filter to obtain multiple three-dimensional training blocks. The disease detection module is used to train a pre-built disease prediction model based on the three-dimensional training block, input the ground-penetrating radar data to be tested into the trained disease prediction model, perform disease category prediction statistics to obtain the total number of discrete points of the target disease, and calculate the actual three-dimensional physical volume of the target disease underground based on the reference physical voxel parameters and the total number of discrete points of the target disease.

[0008] In addition, to achieve the above objectives, this application also proposes a ground-penetrating radar 3D point cloud defect detection device, the device comprising: a memory, a processor, and a ground-penetrating radar 3D point cloud defect detection program stored in the memory, the processor being used to run the ground-penetrating radar 3D point cloud defect detection program, the computer program being configured to implement the steps of the ground-penetrating radar 3D point cloud defect detection method as described above.

[0009] In addition, to achieve the above objectives, this application also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the ground-penetrating radar 3D point cloud defect detection method described above.

[0010] This application spatially stacks and transforms the data from multiple ground-penetrating radar (GPR) scanning profiles according to the physical survey line spacing, thus mapping the echo data originally scattered across the two-dimensional profiles to the underground physical space. By physically reconstructing and correcting the amplitude of the target's three-dimensional spatiotemporal echo matrix, the echo breaks and energy abrupt changes caused by sampling intervals between survey lines are made continuous, reducing spatial distortion caused by simple mathematical interpolation. By physically expanding and discretizing the amplitude-corrected three-dimensional point cloud, the medium abrupt change characteristics of the diseased area are expanded from a single amplitude expression to a joint representation of amplitude and depth gradients, and sparsification reduces the interference of invalid background points on the disease characteristics. By using overlapping sliding windows to extract and filter 3D training blocks, minor defects are covered multiple times in different window positions, preventing defect points from being submerged by massive background points during the full-map voxelization process. Based on the 3D training blocks, a defect prediction model is trained, and the volume is calculated by combining the total number of discrete points of the target defect and the baseline physical voxel parameters. This allows the defect identification results to be directly converted into the actual 3D physical volume of the underground. Thus, in ground-penetrating radar scenarios with sparse survey lines, easily misaligned labels, and a high proportion of background points, defect detection is achieved, which combines 3D connectivity, minor defect identification, and quantifiable physical volume. This significantly improves the accuracy of defect detection and the precision of defect volume quantification on real roads. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the first embodiment of the ground-penetrating radar 3D point cloud defect detection method of this application; Figure 2 This is a flowchart illustrating the second embodiment of the ground-penetrating radar 3D point cloud defect detection method of this application; Figure 3 This is a flowchart illustrating the third embodiment of the ground-penetrating radar 3D point cloud defect detection method of this application; Figure 4This is a schematic diagram of the module structure of the ground-penetrating radar 3D point cloud defect detection device according to an embodiment of this application; Figure 5 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the ground-penetrating radar 3D point cloud defect detection method in the embodiments of this application.

[0014] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0016] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0017] It should be noted that the executing entity of this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a computer, server, etc., or an electronic device capable of realizing the above functions. The following uses a ground-penetrating radar 3D point cloud defect detection device (hereinafter referred to as defect detection device) as an example to describe this embodiment and the following embodiments.

[0018] The ground-penetrating radar 3D point cloud defect detection method proposed in this application aims to address the following shortcomings of the prior art: Physical distortion of multi-line network interpolation: Traditional methods use simple mathematical interpolation to splice parallel survey lines, which violates the physical law of spherical diffusion and scattering of electromagnetic waves. This results in severe breaks in the hyperbolic characteristics of the defects between tracks (perpendicular to the survey line direction), and extremely poor 3D connectivity.

[0019] Spatial misalignment between simulated labels and physical features: In 3D GPR (Ground Penetrating Radar) forward simulations or semi-physical models, there is an inherent spatial offset between the geometry of the actual physical defect labels and the strong reflective electromagnetic features (such as hyperbola vertices) generated on the radar echo. Forcibly using the original labels for training will segment out misaligned data containing "false positive" backgrounds, interfering with network learning.

[0020] The "black hole effect" of gridding and the submersion of minor defects: In real road surface inspection, the voxel ratio of healthy matrix to defects often reaches 1:10000 or even more extreme. Traditional methods input the entire point cloud of the image into the network for coarse voxelization. A very small number of tiny defect points are "swallowed" by the massive background points around them and forcibly smoothed into the background, resulting in an extremely high false negative rate and failing to directly provide the 3D physical absolute volume indicators required for engineering.

[0021] To address the pain points in ground-penetrating radar (GPR) intelligent detection, such as inter-channel feature breaks, label spatial misalignment, and the easy obscuring of minute defects under extreme imbalance, this application proposes a 3D point cloud defect detection method for GPR. The key solutions are how to generate a high-fidelity, zero-misalignment 3D radar point cloud training set, and how to accurately extract defects from hundreds of millions of background points and directly output their true 3D physical volume using a sliding window strategy in conjunction with a 3D sparse network.

[0022] Based on this, the first embodiment of this application provides a method for detecting defects in 3D point clouds using ground-penetrating radar, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the ground-penetrating radar 3D point cloud defect detection method of this application.

[0023] In this embodiment, the ground-penetrating radar 3D point cloud defect detection method includes: Step S10: Based on the physical survey line spacing, the ground-penetrating radar scanning profile data collected in multiple parallel lines are spatially stacked and depth-converted to obtain the target's three-dimensional spatiotemporal echo matrix.

[0024] It should be noted that the physical survey line spacing refers to the actual spatial distance between two adjacent ground-penetrating radar survey lines in the transverse direction of the road or in the direction of the survey line arrangement. For example, the interval between adjacent survey lines is a fixed transverse sampling distance along the road.

[0025] Ground penetrating radar (GPR) scanning profile data refers to a two-dimensional amplitude profile (B-scan) image formed after the GPR moves along a single survey line, emits electromagnetic waves, receives reflected signals, and samples them. The horizontal axis represents the distance of the survey line, and the vertical axis represents the two-way travel time of the electromagnetic waves. It is a type of unsupervised image data.

[0026] The target three-dimensional spatiotemporal echo matrix refers to the three-dimensional echo image data corresponding to the underground space of the road after being formed by stacking multiple ground-penetrating radar scan profile data according to their actual spatial locations and undergoing depth conversion.

[0027] In some embodiments, spatial stacking can be achieved by arranging multiple two-dimensional profile images sequentially in a direction perpendicular to the survey lines according to the survey line spacing, thereby constructing a three-dimensional image sequence. Depth conversion can be achieved by utilizing the propagation speed of electromagnetic waves in the underground medium to convert the time axis of the two-dimensional image into the depth axis, thus obtaining a three-dimensional image in the depth domain.

[0028] Understandably, this step maps echo data that was originally scattered across different two-dimensional survey lines to a unified three-dimensional coordinate space, so that subsequent disease analysis no longer depends on the local information of a single profile, and provides physical scale constraints for the generation of three-dimensional point clouds.

[0029] In some embodiments, the disease detection equipment acquires multiple parallel ground-penetrating radar (GPR) scan profiles along the same road segment, reads the actual spacing between each survey line, assigns survey line alignment direction coordinates to each scan profile data according to the survey line acquisition order, determines the spatial lateral coordinates based on the scan mileage or sampling point number, and determines the travel time coordinates based on the echo sampling time. Then, the multiple scan profile data are stacked in three dimensions according to the spatial lateral coordinates, survey line alignment direction coordinates, and travel time coordinates. The travel time coordinates are then converted to depth coordinates by combining the average electromagnetic wave velocity of the underground medium, thereby obtaining the target three-dimensional spatiotemporal echo matrix. For example, in a road cavity detection scenario, the profile data acquired from multiple parallel lane survey lines can be stacked according to the road's lateral intervals, allowing suspected hyperbolic echoes from different survey lines to form an analyzable data volume in the same three-dimensional space.

[0030] Step S20: Physically reconstruct and amplitude correct the target three-dimensional spatiotemporal echo matrix, and convert the amplitude-corrected target three-dimensional spatiotemporal echo matrix into an amplitude-corrected three-dimensional point cloud.

[0031] It should be noted that amplitude-corrected 3D point cloud refers to point cloud image data formed based on 3D spatial coordinates and corrected echo amplitude.

[0032] Understandably, physical reconstruction refers to the spatial compensation and continuous reconstruction of the inter-line data in the three-dimensional spatiotemporal echo matrix of a target, based on the propagation characteristics of ground-penetrating radar electromagnetic waves. Amplitude correction refers to adjusting the echo intensity differences between different survey line points, interpolation points, or spatially neighboring points after reconstruction.

[0033] In some embodiments, the disease detection device can first convert the target three-dimensional spatiotemporal echo matrix to the frequency wavenumber domain, perform wavenumber expansion or spatial compensation in the direction of the measurement line arrangement, and then convert it back to the three-dimensional spatial domain to obtain continuous echo data between measurement lines; then set different weights according to the type differences between the original measurement line points and the newly added spatial points, and perform weighted amplitude correction in the three-dimensional local neighborhood to make the amplitude of the newly added spatial points consistent with the surrounding original measurement line points.

[0034] After reconstruction and correction, the disease detection equipment saves the spatial lateral coordinates, survey line arrangement direction coordinates, depth coordinates, and corrected amplitude corresponding to each effective spatial location as point cloud points, thereby obtaining an amplitude-corrected three-dimensional point cloud. For example, for delamination defects spanning multiple survey lines, physical reconstruction can fill in the missing echo areas between survey lines, and amplitude correction can reduce the intensity discontinuity between adjacent survey lines.

[0035] Step S30: Perform physical gradient expansion and discretization on the amplitude-corrected three-dimensional point cloud to obtain the target five-dimensional sparse point cloud.

[0036] It should be noted that the target five-dimensional sparse point cloud refers to a sparse point cloud composed of spatial horizontal coordinates, coordinates of survey line arrangement direction, depth coordinates, amplitude characteristics, and depth direction gradient characteristics.

[0037] Understandably, physical gradient augmentation refers to introducing gradient features that reflect changes in amplitude along the depth direction, in addition to the original spatial coordinates and amplitude features of the point cloud. Discretization refers to converting a continuous 3D point cloud into a sparse point cloud structure suitable for model input, based on spatial sampling or effective reflection conditions.

[0038] In some embodiments, the disease detection device extracts the spatial lateral coordinates, survey line arrangement direction coordinates, depth coordinates, and amplitude of each point from the amplitude-corrected 3D point cloud. It compares the amplitude differences between adjacent points or neighboring depth layers along the depth direction to obtain the physical gradient in the depth direction. The amplitude and depth-direction physical gradient are then subjected to amplitude limiting, normalization, or standardization to reduce the impact of outliers on model training. Low-intensity background points are removed based on amplitude thresholds, gradient thresholds, or effective reflection conditions, and the remaining points are organized into a target 5D sparse point cloud. For example, in the detection of road waterlogging defects, the dielectric difference between the water body and the roadbed material makes the amplitude changes in the depth direction more pronounced; gradient features can serve as a supplement to amplitude features.

[0039] Step S40: Based on preset sliding window parameters, overlap and cut the target five-dimensional sparse point cloud to obtain multiple three-dimensional training blocks.

[0040] It should be noted that the preset sliding window parameters can be the 3D window size, window movement step size, window overlap ratio, and training block selection criteria. A 3D training block refers to a local 3D point cloud image sample block used to train the disease prediction model, and can include disease training blocks and background training blocks.

[0041] It is understandable that in this step, overlapping cutting and truncation refers to truncating multiple spatially overlapping local point cloud blocks in the target five-dimensional sparse point cloud according to the window movement path.

[0042] In some embodiments, the disease detection device sets a three-dimensional window size based on the spatial range of the target five-dimensional sparse point cloud, and sets a sliding step size smaller than the window size to create overlapping areas between adjacent windows. The system extracts local point cloud blocks according to the sliding window traversal path and filters candidate point cloud blocks based on disease labels, point cloud amplitude distribution, or background stability. Candidate point cloud blocks containing disease areas can be used as disease training blocks, while candidate point cloud blocks that do not contain disease areas and have stable echo changes can be used as background training blocks. Then, coordinate normalization and feature standardization are performed on each three-dimensional training block to ensure that training blocks extracted from different spatial locations have a uniform input scale. For example, when the disease is located at the edge of the window, the overlapping window can incorporate the disease area into the center of the window in the next extraction, thereby forming training samples with different coverage states.

[0043] Step S50: Train the pre-constructed disease prediction model based on the three-dimensional training block, input the ground-penetrating radar data to be tested into the trained disease prediction model, perform disease category prediction statistics to obtain the total number of discrete points of the target disease, and calculate the actual three-dimensional physical volume of the target disease underground based on the reference physical voxel parameters and the total number of discrete points of the target disease.

[0044] It should be noted that the disease prediction model refers to a machine learning model used to predict the disease category of each discrete point in a 3D point cloud. For example, it can be a 3D sparse convolutional network, a point cloud image segmentation network, or other neural networks that can process sparse 3D point cloud images.

[0045] It is understandable that in this step, disease category prediction refers to determining the category to which the discrete points in the point cloud belong, such as background, voids, voids, looseness, or water accumulation.

[0046] It should be noted that the total number of discrete points for the target disease refers to the number of point cloud points predicted to be of the specified disease category. The baseline physical voxel parameter refers to the actual spatial size parameter corresponding to a single point or voxel in the discrete point cloud.

[0047] In some embodiments, the disease detection device trains a disease prediction model based on three-dimensional training blocks and corresponding labels. The model input is the five-dimensional features in a local three-dimensional point cloud block, and the model output is the category probability corresponding to each point.

[0048] After training, the ground-penetrating radar data to be tested is converted into three-dimensional point cloud blocks according to the same data processing procedure as the training data. The three-dimensional point cloud blocks to be tested are then input into the trained disease prediction model to obtain the point-by-point disease category prediction results.

[0049] Finally, the number of discrete points belonging to the target disease category is counted, and the physical volume of the unit point cloud is determined according to the spatial lateral dimension, survey line arrangement direction dimension and depth dimension corresponding to each discrete point. The physical volume of the unit point cloud is then calculated in correspondence with the total number of discrete points of the target disease to obtain the actual three-dimensional physical volume of the target disease underground.

[0050] This embodiment establishes a three-dimensional physical space data foundation by converting the physical survey line spacing and depth. It improves the echo continuity between survey lines through physical reconstruction and amplitude correction, enhances the abrupt change characteristics of the disease medium through physical gradient expansion and sparsification, generates local training blocks by overlapping sliding windows and reduces the proportion of background points, and converts the point-by-point category prediction results into three-dimensional volume results through the disease prediction model and physical voxel mapping. This effectively solves the problems of poor connectivity and inaccurate labels in 3D radar point clouds. Thus, in ground-penetrating radar scenarios with sparse survey lines, easy label misalignment, and a very high proportion of background points, it achieves disease detection with three-dimensional connectivity, identifiable minor diseases, and quantifiable physical volume, which greatly improves the accuracy of disease detection and the precision of disease volume quantification on real roads.

[0051] Based on the first embodiment described above, in the second embodiment of this application, the same or similar content as the first embodiment can be referred to the above description, and will not be repeated hereafter.

[0052] Based on this, please refer to Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the ground-penetrating radar 3D point cloud defect detection method of this application. In this embodiment, step S10 further includes: Step S101: The ground-penetrating radar scanning profile data are equidistantly stacked in the longitudinal coordinate axis direction, and a discrete initial three-dimensional spatiotemporal echo matrix is ​​constructed based on the spatial stacking result.

[0053] It should be noted that the longitudinal coordinate axis direction can be the direction of multiple parallel survey lines, such as the direction of a road cross section or a spatial direction perpendicular to the extension direction of a single survey line. The initial three-dimensional spatiotemporal echo matrix refers to the three-dimensional echo data structure before air layer removal, phase compensation, and depth transformation are performed.

[0054] In some embodiments, the disease detection device stacks N parallel B-scan data lines at equal intervals along the Y-axis according to the actual physical survey line spacing, forming an initial discrete three-dimensional spatiotemporal echo matrix. .

[0055] Step S102: Calculate the absolute envelope amplitude sequence of each echo signal in the time domain in the initial three-dimensional spatiotemporal echo matrix, and extract the maximum amplitude within the sliding window based on the absolute envelope amplitude sequence.

[0056] It should be noted that the absolute envelope amplitude sequence refers to the time-domain amplitude sequence obtained by extracting the envelope of a single-channel echo signal and taking its absolute value. The echo signal refers to the time-sampled waveform corresponding to a fixed spatial position in a three-dimensional matrix. The sliding window refers to the local analysis interval that moves along the time sampling axis.

[0057] It should be understood that an air gap exists between the ground-penetrating radar antenna and the ground, causing early echoes to contain strong direct wave interference. This is addressed by calculating the absolute envelope amplitude sequence of each trace in the time domain and setting a sliding window.

[0058] Step S103: Use the preset ratio of the maximum amplitude as the adaptive peak threshold, and determine the air-ground interface refraction point of each echo signal according to the adaptive peak threshold.

[0059] It should be noted that the adaptive peak threshold is a threshold calculated based on the local maximum amplitude of a single-channel echo signal. The air-ground interface refraction point refers to the sampling point in the echo signal corresponding to the interface between the air layer and the ground surface or road surface medium. The preset ratio can be set according to the equipment type, antenna height above the ground, and on-site noise level; for example, it can be set to 0.6 times the maximum amplitude.

[0060] In practice, the disease detection equipment extracts the maximum amplitude within the sliding window and sets a specific ratio (such as 0.6 times) as the adaptive peak threshold. When the absolute envelope amplitude reaches or exceeds the adaptive peak threshold for the first time, the corresponding sampling point is determined as the refraction point of the air-ground interface. When there are multiple candidate points, the earliest candidate point that meets the continuity condition can be selected.

[0061] Step S104: Perform air layer removal on the initial three-dimensional spatiotemporal echo matrix according to the refraction point at the air-ground interface, and align the first wave based on the preset phase compensation parameters to generate a phase-compensated three-dimensional spatiotemporal echo matrix.

[0062] It should be noted that air gap removal refers to deleting or shielding the early echo sampling segment corresponding to the air gap between the ground-penetrating radar antenna and the ground. Preset phase compensation parameters can be system delay compensation, antenna zero-time correction, sampling point shift, or phase correction coefficients. The first wave refers to the earliest arriving waveform in each echo that is valid for time base calibration.

[0063] It should be understood that this embodiment reduces the interference of direct waves and air gap reflections on underground echoes by removing air layers; and reduces the initial phase difference between different survey lines and different traces by aligning the first wave, thereby improving the consistency of the three-dimensional matrix on the time or depth axis.

[0064] In practice, the disease detection equipment uses the air-to-ground interface refraction point of each echo signal as a reference to delete, zero out, or realign the early air layer sampling segment of that echo. Then, based on preset phase compensation parameters, the sampling points of each echo are shifted or the phase is corrected to ensure that the first wave positions of different echo signals remain consistent. After completing air layer removal and first wave alignment, a phase-compensated three-dimensional spatiotemporal echo matrix is ​​generated.

[0065] Step S105: Based on the pre-determined average electromagnetic wave velocity of the underground medium, the travel time axis of the phase-compensated three-dimensional spatiotemporal echo matrix is ​​converted into the real physical depth axis to obtain the target three-dimensional spatiotemporal echo matrix.

[0066] It should be noted that the average electromagnetic wave velocity in the underground medium refers to the average speed at which electromagnetic waves propagate in the underground medium being tested. This velocity can be obtained through calibration, empirical parameters, core sampling verification, back-calculation from targets at known burial depths, or estimation of the dielectric constant. The travel time axis refers to the time sampling axis corresponding to the transmission, reflection, and reception of the echo.

[0067] In some embodiments, the disease detection equipment performs air layer removal on the full 3D data, introducing phase compensation parameters to align the first wave. Subsequently, using the measured average electromagnetic wave velocity v of the subsurface medium, the travel time axis t is converted into the true physical depth axis. ), to obtain a 3D matrix that strictly corresponds to the underground physical space. .

[0068] This embodiment establishes a three-dimensional spatiotemporal framework for multiple survey lines by equidistant spatial stacking, enhances the stability of early strong reflection identification by absolute envelope amplitude and sliding window, determines the location of the air-ground interface by adaptive peak threshold, reduces direct wave interference and unifies the first wave time base by air layer removal and phase compensation, and completes the scale conversion from travel time to physical depth by average electromagnetic wave velocity, thereby generating a target three-dimensional spatiotemporal echo matrix with consistency in spatial coordinates, time starting point, and depth scale.

[0069] Based on the above embodiments, referring to Figure 3 , Figure 3 This is a flowchart illustrating the third embodiment of the ground-penetrating radar 3D point cloud defect detection method of this application. In this embodiment, step S20 further includes: Step S201: Perform a three-dimensional Fourier transform on the target three-dimensional spatiotemporal echo matrix to convert the target three-dimensional spatiotemporal echo matrix to the three-dimensional frequency wavenumber domain to obtain a frequency wavenumber domain matrix.

[0070] Understandably, because the spacing between GPR survey lines is much larger than the track spacing within a single survey line, directly using a 3D neural network would lead to severe spatial anisotropy. Disease detection equipment can utilize GPU tensors to accelerate the execution of 3D frequency-wavenumber domain (F-K Domain) physical interpolation. The FK domain (Frequency-Wavenumber Domain) refers to the frequency-wavenumber domain, the domain in which wavefield data resides after undergoing multidimensional Fourier transforms in time and space, and is commonly used for the offset relocation and reconstruction of physical wavefields.

[0071] In practical implementation, the disease detection equipment can perform a three-dimensional fast Fourier transform on the target's three-dimensional spatiotemporal echo matrix to obtain the corresponding frequency wavenumber domain matrix, as shown in the following formula: To mitigate the truncation effect at matrix boundaries, window function processing or boundary padding can be performed before the transformation to reduce the impact of spectral leakage on subsequent reconstruction.

[0072] Step S202: Perform resolution encryption processing on the frequency wavenumber domain matrix based on the zero-filling algorithm in the direction of the measurement line arrangement to obtain the extended wavenumber matrix.

[0073] It should be noted that the zero-filling algorithm refers to inserting zero-value spectral sampling points into a specified dimension of the frequency wavenumber domain to extend the sampling length of that dimension.

[0074] It should be noted that the survey line arrangement direction refers to the spatial arrangement direction between multiple parallel survey lines. The extended wavenumber matrix refers to the frequency wavenumber domain matrix after wavenumber sampling and encryption according to the survey line arrangement direction.

[0075] Understandably, the disease detection equipment applies zero-padding in the direction of the measurement line arrangement (ky axis) to increase its spatial resolution by a specific factor, thereby expanding the high-frequency physical wavenumber domain space.

[0076] Step S203: Perform a three-dimensional inverse Fourier transform on the extended wavenumber matrix to obtain the interpolated three-dimensional spatial echo matrix.

[0077] In practice, the disease detection equipment performs a three-dimensional inverse fast Fourier transform on the extended wavenumber matrix to restore the echo data in the extended frequency wavenumber domain to the three-dimensional spatial domain. After the inverse transform, the number of sampling points in the direction of the survey lines increases, resulting in an interpolated three-dimensional spatial echo matrix that includes the original survey line positions and the newly added inter-survey line positions.

[0078] Step S204: Perform spatial smoothing on the interpolated three-dimensional spatial echo matrix based on the three-dimensional convolution kernel to obtain the physically reconstructed three-dimensional echo matrix.

[0079] It should be noted that a 3D convolution kernel refers to a convolution template that performs weighted calculations on the echo amplitude within a 3D spatial neighborhood; it can be a Gaussian kernel, a mean kernel, or a kernel simulating a point spread function. The physically reconstructed 3D echo matrix refers to the 3D echo matrix formed after frequency domain encryption and spatial smoothing.

[0080] It should be understood that this embodiment uses a three-dimensional convolution kernel to spatially smooth the interpolation results, which can simulate the spatial diffusion effect of electromagnetic waves in the underground medium, reduce abrupt changes and isolated noise points between survey lines, and improve the connectivity of the disease hyperbola in the three-dimensional direction.

[0081] In the specific implementation, after the inverse Fourier transform returns to the spatial domain, a three-dimensional convolution kernel is applied along the three-dimensional space to simulate the beam broadening effect of electromagnetic waves radiating in the underground three-dimensional space, so that the originally broken hyperbola can be physically connected between the survey lines.

[0082] Step S205: Construct a dynamic type weight matrix based on the original survey points and interpolated new points in the physically reconstructed three-dimensional echo matrix.

[0083] It should be noted that the original survey points refer to spatial sampling points from actual ground-penetrating radar (GPR) survey lines. Interpolated new points refer to inter-survey point sampling points generated during frequency domain densification and spatial reconstruction. The dynamic type weight matrix is ​​a data matrix used to distinguish between original survey points and interpolated new points, and whose weights are adjusted during the iteration process.

[0084] It is understandable that, due to the energy discontinuity between the newly added survey points and the original survey points after FK domain interpolation, this embodiment can maintain the constraint effect of the original sampling data in the early stage of amplitude correction by setting different weights for points from different sources, and gradually increase the degree of participation of the newly added interpolated points in the fitting, thereby reducing the excessive disturbance of the real sampling information by the newly added points.

[0085] Step S206: Based on the preset physical search radius and the dynamic type weight matrix, perform tensor weighted convolution iteration on the physical reconstruction three-dimensional echo matrix to obtain the iteratively corrected three-dimensional echo matrix.

[0086] It is understood that this embodiment uses physical distance neighborhood, spatial weight and point type weight to jointly constrain amplitude updates, which can reduce the energy discontinuity between the original point and the interpolated new point, while maintaining the spatial consistency of the local echo structure.

[0087] Step S207: If the amplitude mean square error of the iteratively corrected three-dimensional echo matrix is ​​lower than the preset amplitude mean square error threshold, the iteratively corrected three-dimensional echo matrix is ​​converted into an amplitude-corrected three-dimensional point cloud.

[0088] In its implementation, the disease detection equipment first sets up a three-dimensional local search kernel: a physical search radius R (e.g., 0.2m) is set within the fitted mesh, the number of meshes required for each direction is calculated, and a shape-matching 3D Gaussian kernel is generated. Secondly, construct a dynamic type weight matrix. : Mark the "original point" and "interpolated new point" in the label space. The weight of the original point is constant. Interpolation adds point weights Initially, a very low value is assigned, which then increases linearly with the number of iterations k until a stable value is reached.

[0089] 3D Tensor Weighted Convolution Fitting: For any point P(x,y,z) in space, the formula for calculating its corrected amplitude is: Then, in-place memory reuse optimization is performed: to prevent large tensors from causing GPU OOM (Out of Memory), the calculated weighted sum is only divided and replaced in-place in the effective region of the mask (i.e., the region where the weighted sum is greater than 1 × 10^{-12}) until the overall mean square error of amplitude (MSE) converges to a specified threshold (e.g., ...). This completely eliminates interchannel abrupt changes, reduces energy discontinuities between the original point and the interpolated new point, and maintains the spatial consistency of the local echo structure.

[0090] This embodiment transforms the echo matrix to the frequency-wavenumber domain using a three-dimensional Fourier transform, improves the reconstruction resolution between survey lines by zero-filling in the survey line arrangement direction, obtains the spatial interpolation matrix by a three-dimensional inverse Fourier transform, enhances the connectivity of the disease reflection structure by smoothing with a three-dimensional convolution kernel, distinguishes between original points and newly added points by dynamic type weights, corrects amplitude discontinuities by tensor weighted convolution iteration, and controls convergence by amplitude mean square error threshold, thereby obtaining three-dimensional point cloud data with stronger spatial connectivity and more stable amplitude distribution.

[0091] Based on the above embodiments, in the fourth embodiment of the ground-penetrating radar 3D point cloud defect detection method of this application, step S30 further includes: Step S301: Extract the spatial lateral coordinates, survey line arrangement direction coordinates, depth coordinates and amplitude from the amplitude-corrected three-dimensional point cloud to obtain a three-dimensional amplitude point set.

[0092] It should be noted that a three-dimensional amplitude point set refers to a set of points composed of three-dimensional spatial coordinates and corresponding echo amplitudes. The spatial lateral coordinates can correspond to the direction of travel or lateral sampling of a single survey line, the survey line arrangement direction coordinates can correspond to the arrangement direction between multiple parallel survey lines, and the depth coordinates can correspond to the underground physical depth.

[0093] Understandably, in order to conform to the input specifications of 3D sparse convolutional networks, this embodiment needs to discretize the continuous three-dimensional matrix.

[0094] Step S302: Perform adjacent depth difference processing on the amplitude of the three-dimensional amplitude point set along the depth direction to obtain the physical gradient in the depth direction.

[0095] It should be noted that the physical gradient in the depth direction refers to the local variation of the echo amplitude along the underground depth direction, which can be represented by the ratio of the amplitude difference between adjacent depth sampling points to the depth interval. Adjacent depth differential processing refers to performing differential calculations on adjacent depth points under the same or similar lateral coordinates and survey line arrangement direction coordinates.

[0096] It is understandable that defects (such as cavities and water accumulation) are essentially regions of abrupt changes in dielectric constant, where electromagnetic waves undergo phase reversal. Based on coordinates (X,Y,Z) and amplitude A, the physical gradient along the Z-axis is calculated using numerical difference. .

[0097] Step S303: Based on the preset quantile limiting parameters, the amplitude and the physical gradient in the depth direction are respectively limited to obtain the limited amplitude and the limited physical gradient in the depth direction.

[0098] It should be noted that the preset quantile limiting parameter refers to a quantile threshold parameter set based on the amplitude or gradient numerical distribution, such as the 95th quantile. Limiting processing refers to restricting abnormally large or small values ​​exceeding the quantile threshold range to within that threshold range.

[0099] Step S304: Normalize the limiting amplitude and the limiting depth direction physical gradient respectively to obtain normalized amplitude and normalized depth direction physical gradient.

[0100] Understandably, this embodiment applies a 95th percentile safety clamp (Clamp) to amplitude A and gradient Grad_Z to prevent interference from extreme singular values, and then normalizes them respectively.

[0101] Step S305: Filter invalid points from the three-dimensional amplitude point set based on a preset absolute amplitude threshold to obtain an effective reflection point set.

[0102] It should be noted that the preset absolute amplitude threshold refers to the amplitude threshold used to determine whether a point cloud point has an effective reflection response. Invalid points can be low-intensity background points, points in clean medium regions, or noise points that do not possess effective reflection information. The effective reflection point set refers to the set of points that are retained and whose absolute amplitude meets the threshold condition.

[0103] Understandably, this step reduces the density of the point cloud by filtering low-reflectivity background points, thereby reducing the input burden on the subsequent 3D sparse network and increasing the relative proportion of disease-related reflection points in the input data.

[0104] Step S306: Based on the normalized amplitude, the normalized depth direction physical gradient, and the spatial lateral coordinates, survey line arrangement direction coordinates, and depth coordinates of the effective reflection point set, generate a target five-dimensional sparse point cloud.

[0105] It should be noted that the five dimensions in this step refer to the data dimensions composed of three-dimensional spatial coordinates and two types of echo features, namely, spatial lateral coordinates, survey line arrangement direction coordinates, depth coordinates, normalized amplitude, and normalized depth direction physical gradient.

[0106] Understandably, this embodiment applies an absolute amplitude threshold to directly remove the pure background mesh without reflection, discretizing the originally dense three-dimensional matrix into a sparse (X,Y,Z,A,Grad_Z) 5D point cloud that retains only the effective reflection signal, and forcibly sets the Z-axis ratio to 1.0 during export to ensure 1:1 distortion-free physical space reconstruction.

[0107] This embodiment establishes basic point cloud features by extracting a three-dimensional amplitude point set, introduces abrupt change information of the disease interface through depth difference, suppresses outliers by quantile amplitude limiting, unifies feature scale by normalization, filters weakly reflective background points by absolute amplitude thresholding, and generates sparse point cloud input by five-dimensional feature stitching. This reduces background redundancy while retaining disease-related reflection intensity and depth gradient information, thereby improving the feature expression ability of point cloud samples.

[0108] Based on the above embodiments, in the fifth embodiment of the ground-penetrating radar 3D point cloud defect detection method of this application, step S40 further includes: Step S401: Obtain the original disease labels and preset sliding window parameters.

[0109] It should be noted that the original disease label refers to the spatial labeling information of the disease obtained in advance through manual or simulation labeling.

[0110] The preset sliding window parameters include the 3D window size, sliding step size, and window overlap ratio. The 3D window size refers to the length, width, and height of each sliding window in 3D space. The sliding step size refers to the distance moved between the starting positions of adjacent windows. The window overlap ratio refers to the degree of overlap between the spatial ranges of adjacent windows, such as a 50% overlap ratio.

[0111] Step S402: Perform dynamic label alignment based on the original disease labels and the target five-dimensional sparse point cloud to obtain aligned disease labels.

[0112] In its implementation, the disease detection equipment maps the original disease labels to the coordinate system of the target five-dimensional sparse point cloud and searches for strong reflection point cloud regions near the labels. Based on the spatial location, amplitude intensity, and offset of the strong reflection points from the original labels, the label correction direction is determined, and the original disease labels are translated or matched at the point level to obtain aligned disease labels. This reduces the spatial misalignment between the original geometric labels and radar reflection features, and reduces the possibility of erroneous backgrounds or missing reflection regions being mixed into the disease training blocks.

[0113] Furthermore, to address the misalignment between the GPR simulation label and the electromagnetic scattering characteristics, step S402 above may include: Step S4021: Determine the three-dimensional boundary volume of the disease label based on the original disease label, and determine the center coordinates and top depth of the label based on the three-dimensional boundary volume of the disease label.

[0114] It should be noted that the 3D boundary volume of the disease label refers to the three-dimensional geometric area surrounding the original disease label, which can be a cuboid bounding box, a voxel set boundary, or a spatial envelope formed by multiple annotation points. The label center coordinates refer to the center position of the 3D boundary volume of the disease label in the horizontal plane or in three-dimensional space. The label apex depth refers to the depth position of the side of the 3D boundary volume of the disease label closest to the ground surface.

[0115] Step S4022: Based on the preset local search range, the coordinates of the label center, and the depth of the label top, extract local candidate point clouds from the target five-dimensional sparse point cloud.

[0116] It should be noted that the preset local search range refers to a three-dimensional search area defined around the center coordinates of the label and the depth of the label's top. For example, the range in the horizontal and survey line arrangement directions is ±0.4m, and the range in the depth direction is ±0.15m. The local candidate point cloud refers to the set of points extracted from the target five-dimensional sparse point cloud that are located within this search area.

[0117] It should be understood that this step, by limiting the search space to the vicinity of the original labels, can reduce interference from irrelevant strong reflection points and reduce the computational load of subsequent clustering and matching.

[0118] In some embodiments, the disease detection device constructs a three-dimensional local search box using the label center coordinates as the horizontal center and the label top depth as the depth reference. The device traverses the target's five-dimensional sparse point cloud, filtering points whose spatial coordinates fall within the search box to form a local candidate point cloud. For different disease sizes, the preset local search range can be adaptively enlarged or reduced according to the disease label boundary size.

[0119] Step S4023: Extract strong reflection points from the local candidate point cloud based on the preset strong reflection sorting range, and perform density space clustering on the strong reflection points to obtain multiple strong reflection candidate clusters.

[0120] It should be noted that the preset strong reflection sorting range refers to the proportion or number of strong reflection points selected based on absolute amplitude, such as selecting the top 1% of points in terms of absolute amplitude. Strong reflection points refer to points with high absolute amplitude that may correspond to disease interfaces or areas of strong scattering. Strong reflection candidate clusters refer to point cloud clusters composed of spatially adjacent points with high reflection intensity.

[0121] It should be noted that density spatial clustering can be a density-based spatial clustering algorithm (Density-Based Spatial Clustering of Applications with Noise, DBSCAN), which can discover clusters of arbitrary shapes and effectively filter out noise points.

[0122] In some embodiments, the disease detection device sorts the points in the local candidate point cloud from high to low absolute amplitude and extracts strong reflection points according to a preset strong reflection sorting range. The disease detection device inputs the strong reflection points into a density space clustering algorithm, identifies multiple clusters based on the neighborhood radius and minimum number of points, and filters out isolated noise points. Each retained cluster is treated as a strong reflection candidate block.

[0123] Step S4024: Determine the matching score of each strong reflection candidate cluster based on the cluster center, average amplitude, and spatial offset relative to the three-dimensional boundary volume of the disease label.

[0124] It should be noted that the cluster center refers to the central position of the spatial coordinates of all points within the strong reflection candidate cluster. The average amplitude refers to the mean or absolute mean of the amplitude of all points within the candidate cluster. The spatial offset refers to the offset distance of the candidate cluster center relative to the center, top depth, or boundary position of the three-dimensional boundary volume of the disease label.

[0125] It should be understood that this embodiment incorporates reflection intensity, spatial position, and offset distance into asymmetric scoring, which can avoid mismatches caused by matching based solely on amplitude strength or distance.

[0126] Step S4025: Extract the label translation vector based on the strong reflection candidate cluster with the highest matching score, and perform a unique match on the original disease label based on the label translation vector to obtain the aligned disease label.

[0127] It should be noted that the label translation vector refers to the three-dimensional displacement required to move the original disease label from its original position to the location of the target highly reflective cluster. Unique matching means that each label point in the original label corresponds to only one nearest and valid point cloud target point after translation, avoiding multiple label points being repeatedly mapped to the same point.

[0128] It is understood that this embodiment obtains the label geometric reference by extracting the three-dimensional boundary volume of the disease label, reduces the interference of irrelevant points by limiting the local search range, extracts candidate scattering clusters by strong reflection sorting and density clustering, filters the clusters most likely to correspond to disease reflection by multi-factor matching scoring, and achieves label correction by label translation vector and unique matching. This makes the original disease label more accurately fit the geometric features of radar strong reflection and reduces the interference of misaligned labels on model training.

[0129] In one example, the disease detection equipment locates the boundary of the three-dimensional cuboid corresponding to the original label and calculates its center ( , ) and top depth Set the local search radius (e.g., XY range ±0.4m, Z range pm±0.15m).

[0130] Within the search range, the top 1% of extremely strong reflection points in terms of absolute amplitude are extracted, and then input into the DBSCAN algorithm for unsupervised clustering to extract multiple effective strong reflection candidate clusters. }

[0131] For each block Extract its point cloud center ( , , ), average amplitude Applying an asymmetric scoring formula: The weighting is based on the physical characteristics of the radar. It is the power-law enhancement factor for intensity, used to widen the fractional difference between highly reflective clusters.

[0132] Greedy Unique Matching: Select the highest-scoring cluster to extract the translation vector Vshift, and limit the maximum allowed translation distance. Apply cKDTree to find the unique and nearest valid point cloud target point for each point of the original label after translation, and accurately "attach" the misaligned label to the highly reflective geometry.

[0133] Step S403: Determine the sliding window traversal path according to the preset sliding window parameters, and perform overlapping cutting and truncating on the target five-dimensional sparse point cloud with the aligned disease label according to the sliding window traversal path to obtain candidate three-dimensional point cloud blocks.

[0134] It is understandable that this embodiment generates local point cloud samples covering the entire image by overlapping cuts, so that the disease structure located at the edge of the window has the opportunity to be captured more completely in the adjacent window, thereby enhancing the coverage of training data.

[0135] In its implementation, the disease detection equipment generates a starting window based on the spatial boundary and 3D window size of the target 5D sparse point cloud, and then moves the window sequentially along the horizontal coordinate, the coordinate of the survey line arrangement direction, and the depth coordinate according to the sliding step size. After each movement, the disease detection equipment captures the point cloud points within the current window range and their corresponding aligned disease labels to form candidate 3D point cloud blocks.

[0136] Step S404: Determine the candidate 3D point cloud block containing the aligned disease label as the disease training block.

[0137] It should be noted that a disease training block refers to a candidate 3D point cloud block that contains at least one disease label point, disease label area, or disease category marker.

[0138] Step S405: Based on the amplitude variance and disease boundary safety distance of the candidate 3D point cloud blocks that do not contain the aligned disease labels, background training blocks are selected.

[0139] It should be noted that amplitude variance refers to the dispersion of amplitude values ​​within a candidate 3D point cloud block. The defect boundary safety distance refers to the minimum spatial distance between a candidate 3D point cloud block and the nearest defect label boundary. The safety distance condition can be that the minimum distance is greater than a preset distance threshold, such as greater than a sliding step size or greater than a certain number of voxel sizes.

[0140] It is understood that this embodiment filters background blocks with stable signals by amplitude variance and excludes suspected transition blocks close to the disease area by the safe distance of the disease boundary, thereby reducing the risk of background samples being contaminated by the disease edge.

[0141] In its implementation, the disease detection equipment calculates the amplitude variance of candidate 3D point cloud blocks that do not contain aligned disease labels, and calculates the minimum spatial distance between the candidate block and the boundary of the nearest aligned disease label. When the amplitude variance of a candidate block is lower than a preset variance threshold and the safe distance from the disease boundary is greater than a preset distance threshold, the candidate block is identified as a background training block. The number of background training blocks can be maintained in a preset ratio with the number of disease training blocks to control the balance between positive and negative samples.

[0142] Step S406: Perform coordinate normalization and feature standardization on the disease training block and the background training block to obtain multiple three-dimensional training blocks.

[0143] In its implementation, the disease detection device uses the origin, center, or minimum coordinates of each training block's window as a reference to translate and scale the point coordinates within the training block. The device standardizes the point cloud features using training set statistical parameters and retains the corresponding disease labels. After processing, multiple 3D training blocks that can be input into the disease prediction model are obtained.

[0144] In one example, for an extreme imbalance where the ratio of underground background to disease points is as high as 10000:1, the entire process was abandoned. Figure 1 One-time voxelized input, using sliding window cropping to dynamically generate high-quality training blocks.

[0145] 1. Window Parameter Settings: Set a globally uniform 3D window size. Set the sliding step size to half the window size (e.g., 0.25m) to achieve a 50% physical overlap.

[0146] 2. Natural Data Augmentation Truncation: The sliding window traverses the entire image step by step. Any truncation hyperbola located at the edge of the previous window will inevitably be completely enclosed in the center of the window in the next slide. The resulting "complete hyperbola" and "incomplete edge hyperbola" together constitute an excellent data augmentation sample.

[0147] 3. Pure background block selection and mining: Window containing disease labels is marked as disease block; for disease-free point blocks, the amplitude variance of its internal point cloud is calculated, and blocks with extremely low variance values ​​(extremely stable signal) are added to the candidate background pool. Combined with the minimum safe distance to the nearest disease boundary, pure negative samples in proportion to the number of disease blocks are strictly screened.

[0148] This embodiment determines the basis for sample generation by obtaining the original labels and window parameters, reduces the misalignment between labels and echo features by dynamic label alignment, generates candidate blocks covering the entire image by overlapping sliding windows, identifies disease training blocks by label intersection, filters background training blocks by amplitude variance and safety distance, and unifies the input scale by coordinate normalization and feature standardization, thereby forming a point cloud training set with more accurate label positions, higher quality positive and negative samples, and suitable for training three-dimensional sparse networks.

[0149] Based on the above embodiments, in the sixth embodiment of the ground-penetrating radar 3D point cloud defect detection method of this application, the above step S50 further includes: Step S501: Construct a disease prediction model based on a three-dimensional sparse convolutional network, and construct a joint loss function based on focal loss and region overlap loss.

[0150] It should be noted that focus loss is a loss function that mitigates class imbalance by reducing the weight of easily classified samples and increasing the influence of difficult-to-classify samples. Region overlap loss is used to constrain the degree of overlap between the predicted region and the true labeled region. The joint loss function is the training objective function formed by combining the focus loss and region overlap loss according to their weights.

[0151] Understandably, for a target with a fraction as small as one ten-thousandth, a joint loss function is used, as shown in the following formula: in, Indicates joint loss, Indicates focal loss. Indicates the region overlap loss. and These represent the weights of each loss, for example... .

[0152] Configure aggressive settings in the FocalLoss parameter group. Array allocation (e.g., setting a very low penalty weight of 0.02 for the background and a very high penalty weight of 2.0~2.5 for voids / holes) forces the model to focus on small hyperbolas.

[0153] In some embodiments, the disease detection device constructs a disease prediction model, which may include a sparse convolutional input layer, an encoder sparse convolution module, a downsampling module, a decoder upsampling module, a skip connection module, and a point-level classification output layer. The input layer receives point cloud coordinates and point features, which may include normalized amplitude and normalized depth direction physical gradient. A weighted combination of focus loss and region overlap loss is used; for example, a joint loss function can be constructed using 0.8 times Softmax Focal Loss and 0.2 times Dice Loss. For categories such as background, voids, and holes, different category weights can be set to increase the training focus on minor disease categories.

[0154] Step S502: Train the disease prediction model based on the three-dimensional training block and the joint loss function to obtain the trained disease prediction model.

[0155] In some embodiments, the disease detection device divides multiple 3D training blocks into training and validation sets, and inputs the point cloud coordinates, point features, and labels from the training blocks into the disease prediction model. A joint loss function is calculated based on the model output and training labels, and the model parameters are updated using an optimizer. During training, batch training, learning rate decay, early stopping strategies, or saving the optimal weights from the validation set can be employed to obtain the trained disease prediction model.

[0156] Step S503: Convert the ground-penetrating radar data to be tested into point cloud blocks to obtain three-dimensional point cloud blocks to be tested, and input the three-dimensional point cloud blocks to be tested into the trained disease prediction model to predict the disease category point by point, and obtain the prediction result point by point.

[0157] In some embodiments, the disease detection equipment sequentially performs multi-line spatial stacking, depth conversion, physical reconstruction, amplitude correction, five-dimensional sparse point cloud generation, and sliding window cutting on the ground-penetrating radar data to be tested, resulting in a three-dimensional point cloud block to be tested. Each three-dimensional point cloud block to be tested is input into the trained disease prediction model, and the model outputs the probability that each point belongs to the background, cavity, void, water accumulation, or other target categories, and obtains the point-by-point category prediction result based on the maximum probability or a preset confidence threshold.

[0158] Step S504: Based on the point-by-point category prediction results, count the number of discrete points corresponding to the target disease category to obtain the total number of discrete points of the target disease; Step S505: Determine the physical volume of the unit point cloud based on the benchmark physical voxel parameters, and calculate the actual three-dimensional physical volume of the target disease underground based on the physical volume of the unit point cloud and the total number of discrete points of the target disease.

[0159] It should be noted that the physical volume of a unit point cloud refers to the physical volume represented by a single discrete point or a single voxel in the actual underground space. The reference physical voxel parameters may include the lateral voxel size dx, the voxel size dy in the survey line alignment direction, and the voxel size dz in the depth direction.

[0160] Understandably, during the inference phase, the disease prediction model outputs point-by-point category prediction results on the test set. Based on the uniformly calibrated fine grid size (grid_size, e.g., 0.02m) during model voxelization, the physical reference volume of a unit point cloud in three-dimensional space is calculated. The statistical prediction calculates the total number of discrete points Ndisease belonging to the target disease (such as cavity), and directly outputs the actual physical volume of the disease underground. This embodiment constructs a disease prediction model adapted to sparse point clouds through a three-dimensional sparse convolutional network. It uses focal loss and region overlap loss to jointly constrain minority class identification and region segmentation morphology. It completes model parameter learning through three-dimensional training blocks, realizes the inference of the test data through the transformation of uniform point cloud blocks, obtains the number of discrete points of the target disease through point-by-point category statistics, and completes volume mapping through the benchmark physical voxel parameters, thereby realizing point-level identification of underground diseases and direct quantization of three-dimensional physical volume.

[0161] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the ground-penetrating radar 3D point cloud defect detection method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0162] This application also provides a ground-penetrating radar 3D point cloud defect detection device, please refer to... Figure 4 , Figure 4 This is a schematic diagram of the module structure of the ground-penetrating radar 3D point cloud defect detection device according to an embodiment of this application. The ground-penetrating radar 3D point cloud defect detection device includes: The matrix construction module 10 is used to spatially stack and depth-convert multiple parallel ground-penetrating radar scanning profile data according to the physical survey line spacing to obtain the target's three-dimensional spatiotemporal echo matrix. The reconstruction and correction module 20 is used to physically reconstruct and correct the amplitude of the target three-dimensional spatiotemporal echo matrix, and convert the amplitude-corrected target three-dimensional spatiotemporal echo matrix into an amplitude-corrected three-dimensional point cloud. The expanded discrete module 30 is used to perform physical gradient expansion and discretization processing on the amplitude-corrected three-dimensional point cloud to obtain the target five-dimensional sparse point cloud. The training generation module 40 is used to overlap and cut the target five-dimensional sparse point cloud based on preset sliding window parameters, and filter to obtain multiple three-dimensional training blocks. The disease detection module 50 is used to train a pre-built disease prediction model based on the three-dimensional training block, input the ground-penetrating radar data to be tested into the trained disease prediction model, perform disease category prediction statistics to obtain the total number of discrete points of the target disease, and calculate the actual three-dimensional physical volume of the target disease underground based on the reference physical voxel parameters and the total number of discrete points of the target disease.

[0163] The ground-penetrating radar 3D point cloud defect detection device provided in this application, employing the ground-penetrating radar 3D point cloud defect detection method described in the above embodiments, can solve the technical problems to be addressed by the aforementioned ground-penetrating radar 3D point cloud defect detection method. Compared with the prior art, the beneficial effects of the ground-penetrating radar 3D point cloud defect detection device provided in this application are the same as those of the ground-penetrating radar 3D point cloud defect detection method provided in the above embodiments, and other technical features in the ground-penetrating radar 3D point cloud defect detection device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0164] This application provides a ground-penetrating radar 3D point cloud defect detection device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the ground-penetrating radar 3D point cloud defect detection method in the first embodiment described above.

[0165] The following is for reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing the 3D point cloud defect detection device of ground penetrating radar in the embodiments of this application. The 3D point cloud defect detection device of ground penetrating radar in the embodiments of this application may include, but is not limited to, road inspection vehicle-mounted industrial computers, edge computing terminals, portable inspection workstations, desktop computers, graphics processing workstations, servers, and cloud computing devices. Figure 5 The ground-penetrating radar 3D point cloud disease detection device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0166] like Figure 5 As shown, the ground-penetrating radar 3D point cloud defect detection device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to the program stored in ROM 1002 (Read Only Memory) or the program loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the ground-penetrating radar 3D point cloud defect detection device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the ground-penetrating radar 3D point cloud defect detection equipment to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a ground-penetrating radar 3D point cloud defect detection equipment with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0167] The ground-penetrating radar 3D point cloud defect detection device provided in this application, employing the ground-penetrating radar 3D point cloud defect detection method described in the above embodiments, can solve the technical problems to be addressed by the aforementioned ground-penetrating radar 3D point cloud defect detection method. Compared with the prior art, the beneficial effects of the ground-penetrating radar 3D point cloud defect detection device provided in this application are the same as those of the ground-penetrating radar 3D point cloud defect detection method provided in the above embodiments, and other technical features in this ground-penetrating radar 3D point cloud defect detection device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0168] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0169] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0170] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the ground-penetrating radar 3D point cloud defect detection method in the above embodiments.

[0171] The computer-readable storage medium provided in this application can be a tangible medium capable of storing computer programs, such as a hard disk, random access memory, read-only memory, flash memory, optical storage device, or magnetic storage device. When the computer program is executed by a processor, it implements the above-described ground-penetrating radar 3D point cloud defect detection method.

[0172] The aforementioned computer-readable storage medium may be included in the ground-penetrating radar 3D point cloud defect detection equipment; or it may exist independently and not be assembled into the ground-penetrating radar 3D point cloud defect detection equipment.

[0173] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the ground-penetrating radar 3D point cloud defect detection device, the device performs spatial stacking and depth conversion on multiple parallel ground-penetrating radar scanning profile data according to the physical survey line spacing to obtain a target three-dimensional spatiotemporal echo matrix. The device then performs physical reconstruction and amplitude correction on the target three-dimensional spatiotemporal echo matrix, converting the amplitude-corrected matrix into an amplitude-corrected three-dimensional point cloud. The amplitude-corrected three-dimensional point cloud undergoes physical gradient expansion and discretization to obtain a target five-dimensional sparse point cloud. Based on preset sliding window parameters, the target five-dimensional sparse point cloud is overlapped and cut to obtain multiple three-dimensional training blocks. A pre-constructed defect prediction model is trained based on the three-dimensional training blocks. The ground-penetrating radar data to be tested is input into the trained defect prediction model to perform defect category prediction statistics to obtain the total number of discrete points of the target defect. Based on the baseline physical voxel parameters and the total number of discrete points of the target defect, the actual three-dimensional physical volume of the target defect underground is calculated.

[0174] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0175] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0176] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described ground-penetrating radar 3D point cloud defect detection method, and is capable of solving the technical problems to be solved by the above-described ground-penetrating radar 3D point cloud defect detection method. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the ground-penetrating radar 3D point cloud defect detection method provided in the above embodiments, and will not be repeated here.

[0177] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for detecting defects in 3D point clouds using ground-penetrating radar, characterized in that, The ground-penetrating radar 3D point cloud defect detection method includes: Based on the physical survey line spacing, multiple parallel ground-penetrating radar scanning profile data are spatially stacked and depth-converted to obtain the target's three-dimensional spatiotemporal echo matrix. The target three-dimensional spatiotemporal echo matrix is ​​physically reconstructed and its amplitude corrected, and the amplitude-corrected target three-dimensional spatiotemporal echo matrix is ​​converted into an amplitude-corrected three-dimensional point cloud; The amplitude-corrected three-dimensional point cloud is subjected to physical gradient expansion and discretization to obtain the target five-dimensional sparse point cloud. Based on preset sliding window parameters, the target five-dimensional sparse point cloud is overlapped, cut, and extracted to obtain multiple three-dimensional training blocks. Based on the three-dimensional training block, a pre-constructed disease prediction model is trained. The ground-penetrating radar data to be tested is input into the trained disease prediction model to predict and statistically analyze the disease category to obtain the total number of discrete points of the target disease. Based on the reference physical voxel parameters and the total number of discrete points of the target disease, the actual three-dimensional physical volume of the target disease underground is calculated.

2. The ground-penetrating radar 3D point cloud defect detection method as described in claim 1, characterized in that, The process involves spatially stacking and depth-converting multiple parallel ground-penetrating radar scan profiles based on the physical survey line spacing to obtain the target's three-dimensional spatiotemporal echo matrix, including: The ground-penetrating radar scanning profile data are spatially stacked at equal intervals along the longitudinal coordinate axis, and a discrete initial three-dimensional spatiotemporal echo matrix is ​​constructed based on the spatial stacking result. Calculate the absolute envelope amplitude sequence of each echo signal in the time domain in the initial three-dimensional spatiotemporal echo matrix, and extract the maximum amplitude within the sliding window based on the absolute envelope amplitude sequence; The preset proportion of the maximum amplitude is used as the adaptive peak threshold, and the air-ground interface refraction point of each echo signal is determined according to the adaptive peak threshold. Based on the refraction point at the air-ground interface, the initial three-dimensional spatiotemporal echo matrix is ​​subjected to air layer removal, and the first wave is aligned based on preset phase compensation parameters to generate a phase-compensated three-dimensional spatiotemporal echo matrix. Based on the pre-determined average electromagnetic wave velocity of the underground medium, the travel time axis of the phase-compensated three-dimensional spatiotemporal echo matrix is ​​converted into the actual physical depth axis to obtain the target three-dimensional spatiotemporal echo matrix.

3. The ground-penetrating radar 3D point cloud defect detection method as described in claim 1, characterized in that, The step of physically reconstructing and amplitude correcting the target three-dimensional spatiotemporal echo matrix, and converting the amplitude-corrected target three-dimensional spatiotemporal echo matrix into an amplitude-corrected three-dimensional point cloud, includes: A three-dimensional Fourier transform is performed on the target three-dimensional spatiotemporal echo matrix to transform the target three-dimensional spatiotemporal echo matrix into the three-dimensional frequency wavenumber domain, thereby obtaining a frequency wavenumber domain matrix. The frequency wavenumber domain matrix is ​​subjected to resolution encryption processing based on the zero-filling algorithm in the direction of the measurement line arrangement to obtain the extended wavenumber matrix. Perform a three-dimensional inverse Fourier transform on the extended wavenumber matrix to obtain the interpolated three-dimensional spatial echo matrix; The interpolated three-dimensional spatial echo matrix is ​​spatially smoothed based on a three-dimensional convolution kernel to obtain a physically reconstructed three-dimensional echo matrix. A dynamic type weight matrix is ​​constructed based on the original survey points and interpolated new points in the physically reconstructed three-dimensional echo matrix. Based on the preset physical search radius and the dynamic type weight matrix, tensor weighted convolution iteration is performed on the physically reconstructed three-dimensional echo matrix to obtain the iteratively corrected three-dimensional echo matrix. If the amplitude mean square error of the iteratively corrected three-dimensional echo matrix is ​​lower than a preset amplitude mean square error threshold, the iteratively corrected three-dimensional echo matrix is ​​converted into an amplitude-corrected three-dimensional point cloud.

4. The ground-penetrating radar 3D point cloud defect detection method as described in claim 1, characterized in that, The physical gradient augmentation and discretization of the amplitude-corrected three-dimensional point cloud to obtain the target five-dimensional sparse point cloud includes: Spatial lateral coordinates, survey line arrangement direction coordinates, depth coordinates, and amplitude are extracted from the amplitude-corrected three-dimensional point cloud to obtain a three-dimensional amplitude point set. The amplitudes in the three-dimensional amplitude point set are subjected to adjacent depth difference processing along the depth direction to obtain the physical gradient in the depth direction; Based on preset quantile limiting parameters, the amplitude and the physical gradient in the depth direction are respectively limited to obtain the limited amplitude and the limited physical gradient in the depth direction. The amplitude and the physical gradient in the depth direction of the amplitude limiting are normalized respectively to obtain the normalized amplitude and the normalized physical gradient in the depth direction. The set of three-dimensional amplitude points is filtered for invalid points based on a preset absolute amplitude threshold to obtain a set of valid reflection points; Based on the normalized amplitude, the normalized depth direction physical gradient, and the spatial lateral coordinates, survey line arrangement direction coordinates, and depth coordinates of the effective reflection point set, a five-dimensional sparse point cloud of the target is generated.

5. The ground-penetrating radar 3D point cloud defect detection method as described in claim 1, characterized in that, The method involves overlapping and cutting the target five-dimensional sparse point cloud based on preset sliding window parameters to obtain multiple three-dimensional training blocks, including: Obtain the original disease labels and preset sliding window parameters, wherein the preset sliding window parameters include the three-dimensional window size, sliding step size and window overlap ratio; Dynamic label alignment is performed based on the original disease labels and the target five-dimensional sparse point cloud to obtain aligned disease labels. The sliding window traversal path is determined according to the preset sliding window parameters, and the target five-dimensional sparse point cloud with the aligned disease label is overlapped and cut off according to the sliding window traversal path to obtain candidate three-dimensional point cloud blocks. The candidate 3D point cloud block containing the aligned disease label is determined as the disease training block; Based on the amplitude variance and disease boundary safety distance of candidate 3D point cloud blocks that do not contain the aligned disease labels, background training blocks are selected. The disease training block and the background training block are normalized in coordinates and standardized in features to obtain multiple three-dimensional training blocks.

6. The ground-penetrating radar 3D point cloud defect detection method as described in claim 5, characterized in that, The process of dynamically aligning disease labels based on the original disease labels and the target five-dimensional sparse point cloud to obtain aligned disease labels includes: The three-dimensional boundary volume of the disease label is determined based on the original disease label, and the center coordinates and top depth of the label are determined based on the three-dimensional boundary volume of the disease label. Based on a preset local search range, the coordinates of the label center, and the depth of the label top, local candidate point clouds are extracted from the target five-dimensional sparse point cloud; Strong reflection points are extracted from the local candidate point cloud based on a preset strong reflection sorting range, and density spatial clustering is performed on the strong reflection points to obtain multiple strong reflection candidate clusters. A matching score for each strong reflection candidate cluster is determined based on the cluster center, average amplitude, and spatial offset relative to the three-dimensional boundary volume of the disease label. The label translation vector is extracted based on the candidate cluster with the highest matching score and strong reflection. The original disease label is then uniquely matched based on the label translation vector to obtain the aligned disease label.

7. The ground-penetrating radar 3D point cloud defect detection method as described in claim 1, characterized in that, The process involves training a pre-built disease prediction model based on the three-dimensional training blocks, inputting the ground-penetrating radar data to be measured into the trained disease prediction model, performing disease category prediction statistics to obtain the total number of discrete points of the target disease, and calculating the actual three-dimensional physical volume of the target disease underground based on the baseline physical voxel parameters and the total number of discrete points of the target disease, including: A disease prediction model is constructed based on a three-dimensional sparse convolutional network, and a joint loss function is constructed based on focal loss and region overlap loss. The disease prediction model is trained based on the three-dimensional training block and the joint loss function to obtain the trained disease prediction model. The ground-penetrating radar data to be tested is converted into point cloud blocks to obtain a three-dimensional point cloud block to be tested. The three-dimensional point cloud block to be tested is then input into the trained disease prediction model to predict the disease category point by point and obtain the point-by-point category prediction result. Based on the point-by-point category prediction results, the number of discrete points corresponding to the target disease category is counted to obtain the total number of discrete points of the target disease. The physical volume of a unit point cloud is determined based on the benchmark physical voxel parameters, and the actual three-dimensional physical volume of the target disease underground is calculated based on the physical volume of the unit point cloud and the total number of discrete points of the target disease.

8. A ground-penetrating radar 3D point cloud defect detection device using the ground-penetrating radar 3D point cloud defect detection method according to any one of claims 1 to 7, characterized in that, The device includes: The matrix construction module is used to spatially stack and depth-convert multiple parallel ground-penetrating radar scanning profile data according to the physical survey line spacing to obtain the target's three-dimensional spatiotemporal echo matrix. The reconstruction and correction module is used to physically reconstruct and correct the amplitude of the target three-dimensional spatiotemporal echo matrix, and convert the amplitude-corrected target three-dimensional spatiotemporal echo matrix into an amplitude-corrected three-dimensional point cloud. An extended discretization module is used to perform physical gradient expansion and discretization processing on the amplitude-corrected three-dimensional point cloud to obtain the target five-dimensional sparse point cloud. The training generation module is used to overlap and cut out the target five-dimensional sparse point cloud based on preset sliding window parameters, and filter to obtain multiple three-dimensional training blocks. The disease detection module is used to train a pre-built disease prediction model based on the three-dimensional training block, input the ground-penetrating radar data to be tested into the trained disease prediction model, perform disease category prediction statistics to obtain the total number of discrete points of the target disease, and calculate the actual three-dimensional physical volume of the target disease underground based on the reference physical voxel parameters and the total number of discrete points of the target disease.

9. A ground-penetrating radar 3D point cloud defect detection device, characterized in that, The ground-penetrating radar 3D point cloud defect detection device includes: a memory, a processor, and a ground-penetrating radar 3D point cloud defect detection program stored in the memory. The processor is used to run the ground-penetrating radar 3D point cloud defect detection program, which is configured to implement the ground-penetrating radar 3D point cloud defect detection method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a ground-penetrating radar 3D point cloud defect detection program, which, when executed by a processor, implements the ground-penetrating radar 3D point cloud defect detection method as described in any one of claims 1 to 7.

Images