Ground penetrating radar-laser point cloud tunnel lining three-dimensional reconstruction method, device and medium based on deep learning

Abstract

The application relates to the technical field of tunnel engineering health monitoring and intelligent diagnosis, and discloses a ground penetrating radar-laser point cloud tunnel lining three-dimensional reconstruction method, equipment and medium based on deep learning, which comprises the following steps: data acquisition, synchronous control of uniform movement of a detection vehicle, synchronous acquisition of double-frequency ground penetrating radar signals and submillimeter laser point clouds; data preprocessing, dynamic correction of a lining reference surface for the point clouds, and time-varying gain compensation for the radar signals; feature extraction and fusion, extraction of radar features, point cloud crack depth and curvature distribution, etc., and layered fusion with material aging weights; three-dimensional reconstruction and decision making, generation of a quantitative report conforming to tunnel health evaluation specifications. The application innovatively embeds a physical mechanism into deep learning, combines multi-source data collaborative calibration technology, overcomes three technical bottlenecks of high deep cavity missing detection rate, large curved crack depth error and difficult interlayer defect identification of a traditional method, and provides a full-chain detection and analysis scheme for tunnel intelligent maintenance.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering health monitoring and intelligent diagnosis technology, and in particular to a method, equipment and medium for three-dimensional reconstruction of tunnel lining based on deep learning ground penetrating radar-laser point cloud. Background Technology

[0002] Traditional tunnel lining inspection methods mainly rely on localized destructive methods such as manual tapping and core drilling, which suffer from low efficiency, insufficient coverage, and a high rate of missed detection of hidden defects. In recent years, while ground-penetrating radar (GPR) and laser scanning technologies have been applied to non-destructive testing, their multi-source data coordination is poor: GPR signals suffer severe attenuation at depths (signal-to-noise ratio drops by 8-10 dB at depths >40 cm), laser point cloud surface deformation correction errors are large (the depth of surface cracks can be underestimated by up to 300%), and the spatial registration accuracy between radar and point cloud data is insufficient (error >5 mm). This makes it difficult to achieve accurate three-dimensional visualization diagnosis of key defects such as internal voids, interlayer delamination, and steel corrosion in the lining. Existing deep learning reconstruction methods mostly rely on end-to-end black-box training, ignoring physical mechanisms such as concrete dielectric frequency variation and crack stiffness attenuation. They exhibit poor generalization ability under complex working conditions and cannot meet the quantitative assessment requirements of highway and railway tunnel maintenance standards. Summary of the Invention

[0003] The purpose of this invention is to propose a method, equipment, and medium for three-dimensional reconstruction of tunnel lining based on deep learning ground-penetrating radar-laser point cloud, which solves the technical problem that existing three-dimensional reconstruction methods for tunnel lining have poor generalization ability under complex working conditions and cannot meet the requirements of quantitative evaluation in highway and railway tunnel maintenance specifications.

[0004] Specifically, this invention provides a deep learning-based method for three-dimensional reconstruction of tunnel lining using ground-penetrating radar-laser point cloud, comprising the following steps: S1. Equipment Co-calibration: A multi-band ground-penetrating radar antenna array and a laser scanner are rigidly installed on the tunnel inspection vehicle. A spatiotemporal reference is established through the GNSS-INS integrated navigation system to calibrate the spatial attitude parameters of the multi-band ground-penetrating radar antenna array and the laser scanner. S2. Multi-source data acquisition: Control the detection vehicle to move at a preset speed at a constant speed, and simultaneously acquire dual-band electromagnetic reflection signals from ground penetrating radar and sub-millimeter-level laser point clouds; S3. Dynamic reference preprocessing: Perform dynamic reference surface correction on the laser point cloud, calculate the lining deformation based on the design axis equation and real-time point cloud, and establish an adaptive coordinate system; implement time-varying gain compensation based on the frequency-varying dielectric properties of concrete for the radar signal. S4. Multimodal feature extraction: Radar features are extracted from multi-band radar signals using a pre-trained dual-branch network, including dielectric anomalies, rebar reflection waveforms and dispersion characteristics. Point cloud features are extracted from the corrected point cloud, including crack depth and curvature distribution. S5. Structure-oriented fusion: Through the lining layered attention module, radar features and point cloud features are fused layer by layer according to the design thickness, and material aging weights characterized by dispersion characteristics are injected to obtain fused features; S6. Probabilistic Reconstruction: Input the fused features into the 3D generative adversarial network and output a 3D voxel model containing thickness error distribution, void probability and steel corrosion risk level. S7. Engineering Decision Support: Generate quantitative reports that conform to tunnel health assessment standards based on three-dimensional voxel models, and mark high-risk areas and maintenance priorities.

[0005] A storage device that stores instructions and data for implementing a deep learning-based three-dimensional reconstruction method for tunnel lining using ground-penetrating radar-laser point cloud.

[0006] A deep learning-based ground-penetrating radar-laser point cloud tunnel lining 3D reconstruction device includes: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a deep learning-based ground-penetrating radar-laser point cloud tunnel lining 3D reconstruction method.

[0007] The beneficial effects provided by this invention are: Breakthrough in multi-source data collaborative calibration accuracy: Sub-millimeter-level spatial registration is achieved through spatiotemporal reference unification technology. GNSS-INS fusion positioning combined with multi-sensor joint adjustment algorithms controls the spatial pose residuals of radar and laser scanners to within 0.18mm, improving positioning accuracy by 27 times. A quantum clock synchronization mechanism eliminates transmission delay, with a time deviation ≤40ns (displacement error <0.05mm), laying the foundation for deep signal analysis.

[0008] A revolutionary feature extraction method driven by physical mechanisms: By modeling dielectric and mechanical fields across domains and embedding the dielectric anomaly enhancement kernel into the Herveside step function physical model, the generalization problem of traditional black box models is solved, and the detection rate of thin-layer voids is improved; by using stiffness-weighted geodesic field fusion curvature tensor, the error of surface crack depth is reduced, and the efficiency of geometric distortion correction is improved.

[0009] Enhanced intelligent diagnostic capability for interlayer defects: The structure-guided fusion mechanism enables accurate capture of hidden defects. It utilizes dispersion-stiffness degradation mapping to calculate material aging weights, reducing errors and accurately predicting support life. Interlayer coupling tensor quantifies strain transmission continuity, detecting smaller interlayer peelings and significantly improving the sensitivity of layered defect identification. Attached Figure Description

[0010] Figure 1 This is a simplified schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the hardware device used in this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0012] Before formally describing the present invention, a general description of the solution of the present invention will be given first to facilitate understanding.

[0013] Example 1 Please refer to Figure 1 , Figure 1 This is a schematic diagram of the method flow of the present invention; This invention provides a deep learning-based method for three-dimensional reconstruction of tunnel lining using ground-penetrating radar-laser point cloud, comprising the following steps: S1. Equipment Co-calibration: A multi-band ground-penetrating radar antenna array and a laser scanner are rigidly installed on the tunnel inspection vehicle. A spatiotemporal reference is established through the GNSS-INS integrated navigation system to calibrate the spatial attitude parameters of the multi-band ground-penetrating radar antenna array and the laser scanner. It should be noted that the spatiotemporal reference mentioned in step S1 is established based on the original position and attitude data output by the GNSS-INS integrated navigation system.

[0014] The calibration described in step S1 requires measuring the spatial positional deviation between the phase center of the ground-penetrating radar antenna array and the optical center of the laser scanner. Specifically, a multi-sensor joint adjustment algorithm is used for calibration to calculate the three-dimensional translation and rotation parameters of the radar antenna and laser scanner in the rigid coordinate system of the inspection vehicle, as shown in the following formula:

[0015] The formula is calculated to obtain To minimize the spatial positional deviation between the optical center of the laser scanner and the phase center of the radar antenna, mm², where: N For N sets of data, achieve sub-millimeter spatial registration of multiple sensors (measured residual ≤0.05mm); The coordinates of the optical center of the laser scanner at the i-th calibration point The unit is mm; The coordinates of the phase center of the ground-penetrating radar antenna at the i-th calibration point. The unit is mm; Rc This is a dimensionless rotation matrix from the radar coordinate system to the laser coordinate system. The matrix is: Orthogonal matrix; The translation vector from the radar coordinate system to the laser coordinate system , mm.

[0016] In addition, residuals are calibrated using pose parameters. As a control indicator for calibration accuracy. , where is the root mean square threshold for spatial calibration error, in mm; N The number of calibration points must be greater than or equal to 4, and this value must be an integer. for and The difference is the spatial position deviation vector of the i-th calibration point. It can accommodate any non-collinear calibration points, and can be solved with N≥4. The measured accuracy is optimal when N=6, and it can resist environmental vibration interference. Finally, the PPS second pulse signal from GNSS-INS is used to trigger the hardware clock alignment of all sensors, eliminating transmission delay errors.

[0017] S2. Multi-source data acquisition: Control the detection vehicle to move at a preset speed at a constant speed, and simultaneously acquire dual-band electromagnetic reflection signals from ground penetrating radar and sub-millimeter-level laser point clouds; Specifically, the control and detection vehicle moves at a constant speed of ≤5km / h, simultaneously collecting electromagnetic reflection signals from dual-band (500MHz / 1GHz) ground-penetrating radar and sub-millimeter-level laser point clouds.

[0018] S3. Dynamic reference preprocessing: Perform dynamic reference surface correction on the laser point cloud, calculate the lining deformation based on the design axis equation and real-time point cloud, and establish an adaptive coordinate system; implement time-varying gain compensation based on the frequency-varying dielectric properties of concrete for the radar signal. It should be noted that in step S3, the dynamic reference surface correction of the laser point cloud uses the least squares method to fit the design axis equation, and the error range of the lining deformation calculated in real time is within ±0.5 mm; the technical principle of the least squares method for fitting the design axis equation is as follows: Where s is the arc length parameter, given a measured point cloud, find the optimal parameters to minimize the objective function:

[0019] in for The projection point parameters on the design axis are controlled by using an iterative weighted least squares method to suppress outliers and manage errors. The curve length (mm) is calculated along the axis from the starting point. N This represents the total number of sampling points in the point cloud. The 3D coordinates of the i-th point cloud, in mm. The requirements of step 3 are met. Real-time output of deformation cloud map (lining deformation): , The normal deformation at position s of the design axis is in mm, and the normal offset of the lining surface relative to the design axis is also given. s To determine the arc length position on the design axis, in mm, the curve length measured along the axis from the tunnel starting point will be used as the data. m The number of neighboring points, the number of nearby measured points used for interpolation. The arc length of the j-th nearest point, in mm, is the projection position of the measured point on the design axis. Gaussian kernel bandwidth parameter, mm, is a parameter that controls the influence range of neighboring points; The measured normal deformation of the j-th neighboring point, in mm, is the normal distance after the point cloud is projected onto the axis. Time-varying gain compensation for radar signals is implemented based on a piecewise calibration curve derived from the frequency-varying characteristics of the dielectric constant of concrete. Radar gain compensation involves first establishing a compensation model based on the frequency-varying dielectric characteristics of concrete, calculating the static dielectric constant, then using an attenuation compensation function to calculate the time-varying gain, and finally multiplying the time-varying gain by the original result to obtain the compensated radar signal.

[0020] Dielectric frequency-varying characteristic model: The complex dielectric constant of concrete is , The complex permittivity is the dielectric constant that characterizes the material at a given frequency. f Electromagnetic response characteristics under these conditions f Hertz (Hz) is the frequency of electromagnetic waves, and the operating frequency of ground-penetrating radar (e.g., 500MHz / 1GHz). The real part of the dielectric constant characterizes the energy storage capacity of a material: the larger the value, the slower the electromagnetic wave propagation speed. Typical values ​​for concrete are 4.0~9.0. The imaginary part of the dielectric constant characterizes material loss: the larger the value, the stronger the electromagnetic wave attenuation; typical values ​​for concrete are 0.1~0.8. j The imaginary unit, .

[0021] Dispersion model:

[0022] in ε s It is the static dielectric constant; ε ∞ To calculate the static dielectric constant, which is the optical frequency dielectric constant, it is necessary to extract the low-frequency (500MHz) signal.

[0023] in c For the speed of light ( s); d represents the time delay difference between the direct wave and the reflected wave (ns); d is the known thickness of the structural layer (m).

[0024] In the construction of the attenuation compensation function, the electromagnetic wave attenuation coefficient is:

[0025] ) is the electromagnetic wave attenuation coefficient, which characterizes the energy loss rate of electromagnetic waves propagating in concrete, expressed in dB / m (decibels per meter). f Electromagnetic wave frequency, ground penetrating radar operating frequency, Hz (Hertz). z Propagation depth, m, is the depth to which the electromagnetic wave penetrates from the lining surface into the interior; c is the speed of light in vacuum, a constant value. Time-varying gain function:

[0026] t Electromagnetic wave propagation time, ns, is the time from radar transmission to reception (two-way travel time). The attenuation coefficient is related to frequency and depth. From the previous step, we can calculate that the electromagnetic wave at depth... ζ Energy loss rate at the point, dB / m (decibels per meter); Time-depth transformation function, m, time t The corresponding electromagnetic wave propagation depth.

[0027] Signal compensation: Compensated signal: ,in The original measured signal is the original voltage signal received by the ground penetrating radar, in V.

[0028] Segmented calibration implementation: When frequency band 1 (300-600MHz). When frequency band 2 (800-1200MHz) is used. .

[0029] S4. Multimodal feature extraction: Radar features are extracted from multi-band radar signals using a pre-trained dual-branch network, including dielectric anomalies, rebar reflection waveforms and dispersion characteristics. Point cloud features are extracted from the corrected point cloud, including crack depth and curvature distribution. It should be noted that the multimodal feature extraction uses a dual-branch network to process radar signal waveforms (including dielectric anomalies and rebar reflection features (rebar reflection features are the aforementioned rebar reflection waveforms and dispersion characteristics)) and point cloud (geometric) features (including crack depth l and curvature distribution).

[0030] The extraction of dual physical constraint features (i.e., dielectric anomaly and rebar reflection features) of radar signals utilizes a dielectric anomaly enhancement kernel to construct a physical sensing convolution kernel based on the Herveside step function:

[0031] This formula simulates the abrupt change in dielectric constant at the concrete-void interface, suppresses noise through Gaussian attenuation, and utilizes a physical sensing kernel function. It is a core operator in radar signal processing. The time axis, representing the time delay variable and the electromagnetic wave reflection signal, is expressed using the time parameter of convolution operation, in ns; ) Here, the Herveside step function is used to simulate a sudden change in the dielectric constant. ; Here, represents the Gaussian distribution normalization coefficient (ensuring the kernel function integral is 1), and the normalization factor of the probability density function, in nanoseconds (ns). - ¹; It is an adaptive kernel width parameter, which is adaptively adjusted by the abrupt change in the dielectric properties of concrete.

[0032] Reinforcing bar reflection dispersion correction model; The velocity of electromagnetic waves in concrete (m / ns) is specified in the code as being between 0.10 and 0.12 m / ns. The tuning factor is for concrete type, which can improve the signal-to-noise ratio of thin-layer voids.

[0033] This uses a physics-driven design, incorporating a dielectric catastrophe physics model ( H ( τ The novel method of embedding concrete wave velocity into convolution kernels breaks through the frequency band limitation of traditional fixed kernel functions (such as Ricker wavelets) and improves the detection rate of voids, especially for thin-layer voids, overcoming the problems of weak signals and easy noise submersion of thin-layer voids.

[0034] The matched filter that performs dispersion compensation in the frequency domain: } For the rebar reflection signal after dispersion compensation, this formula outputs a high-fidelity rebar reflection waveform, eliminating the phase distortion caused by concrete dispersion, V; s ( t The signal is the original radar time-domain signal, used as the input signal, and its unit is V; F Fourier transform operator, which is the operator that transforms time-domain signals... s ( t Convert to the frequency domain, V / Hz; It is an ideal steel bar reflection model, that is, a physical reflection waveform template of steel bar (time domain). This is a dispersion phase compensation term, which is dimensionless. jTo construct orthogonal components (such as sin and cos, which need to be uniformly described by complex numbers), the amplitude and phase information of the signal are fully captured, and the dispersion effect of electromagnetic waves in concrete is accurately modeled.

[0035] Phase compensation term function , rad; This is the frequency-varying delay. Due to dispersion delay Determine, ns; For phase second derivative compensation, it is introduced into the tunneling ground-penetrating radar signal processing for use.

[0036] Specifically, the process of extracting crack depth and curvature distribution in step S4 is as follows: Point cloud stiffness-geometry fusion features in crack depth stiffness-weighted geodesy:

[0037] in Stiffness-weighted geodesic distance for crack depth (true 3D depth along the path of minimum stiffness), mm; Integration path (from the crack origin) to the border (the curve); For the local stiffness tensor, N / ; Path differential vector (tangential element along path γ), mm; crack initiation point Coordinates ( mm; Stiffness Tensor S Generated from local curvature features:

[0038] Principal curvature, ; n It is the surface normal vector; For tensor outer product, T This is the transpose operator.

[0039] Along the geodetic path γ Integrating the curvature stiffness tensor accurately quantifies the true depth of cracks in curved linings.

[0040] S5. Structure-oriented fusion: Through the lining layered attention module, radar features and point cloud features are fused layer by layer according to the design thickness, and material aging weights characterized by dispersion characteristics are injected to obtain fused features; It should be noted that in step S5, the lining layer attention module of the structure-guided fusion is layered according to the design thickness, the layer features are extracted, the dispersion characteristics are converted into material aging weight factors, the results are integrated into the radar, and the radar features and point cloud features are also fused according to the thickness layer.

[0041] For the layered tensor fusion model of the lining, the interlayer strain continuity constraint is used to define the interlayer connection tensor:

[0042] in S This is the interlayer stiffness transfer matrix for concrete. σ The dimensional parameters of material heterogeneity are determined by the elastic modulus of concrete. The dielectric anomaly, steel bar reflection, and other features calculated and output according to claim 5 represent the radar feature vector of the Lth layer. Innovation in Fusion Equations:

[0043] The feature vectors fused in layer l are output to the 3D reconstruction network. The radar characteristics of the first layer (dielectric anomaly, steel reinforcement reflection). Features of the point cloud in layer l (crack depth, curvature distribution); The radar characteristic gradient difference calculation obtained according to claim 5, i.e. Normal vector The axial direction of the lining layer design; The exponential decay model of interlayer feature differences, according to the above formula, takes values ​​in the range of (0,1]. The larger the value, the stronger the feature continuity between the two layers. The interlaminar strain transfer efficiency strength coefficient is determined based on experimental calibration values, balancing the contributions of the two components.

[0044] Establish a dispersion-stiffness degradation physical model and map the dispersion coefficients to stiffness degradation:

[0045] Elastic modulus of aged concrete, GPa; The initial elastic modulus of unaged concrete, in GPa; Dispersion sensitivity coefficient, a gauged scalar; The degradation saturation threshold determines the maximum stiffness loss rate; it is a scalar. Let be the average dispersion coefficient of the l-th layer, and be a fixed scalar, in ns / m.

[0046] The physical essence of aging weight has been improved as follows:

[0047] The aging weight of layer l is a scalar (0~1). Stiffness loss threshold, GPa, is set according to the highway tunnel maintenance specification standard. The stiffness variation scale parameter is given in GPa, and the experimental calibration value is also given. By converting dispersion characteristics into stiffness, cross-domain modeling of electromagnetic parameters into mechanical properties is achieved. The crack depth in the point cloud is transformed into a layered stiffness correction, and the crack influence function is as follows:

[0048] The effect of cracks in layer l on the stiffness correction factor, in mm; m The total number of crack sampling points, the number of all crack points in layer l, an integer; As calculated in claim 5, represents the depth of the i-th crack point, the true three-dimensional depth perpendicular to the lining surface, in mm; Let be the scale parameter for the crack's influence range, be a scalar, and be greater than 0, based on an empirical formula. The distance, in mm; The coordinates of the point where the maximum crack depth is greatest (the coordinate point of the maximum crack depth within the layer) are the corresponding points located according to the curvature gradient positioning in claim 5, in mm; Stiffness weighting correction factor: The maximum allowable crack impact value for concrete is related to the material strength and is specified in the standard (mm). sgn Ensure the formula activates the correction only when the threshold is exceeded (to avoid undefined operations). When sgn = -1, y < 0; when sgn = 0, y = 0; when sgn = 1, y > 0. In practical engineering, constraints need to be applied to ensure... .

[0049] S6. Probabilistic Reconstruction: Input the fused features into the 3D generative adversarial network and output a 3D voxel model containing thickness error distribution, void probability and steel corrosion risk level. It should be noted that in step S6, the probabilistic reconstruction maps the fused features into a three-dimensional voxel model through a three-dimensional generative adversarial network. This model includes the error distribution representing the thickness uncertainty, the probability of voids, and the risk level of steel corrosion. The corrosion risk level is solved by a Markov random field, which can clearly define the risk level optimization objective.

[0050] First, we perform voxel space modeling and then establish the generator network:

[0051] Based on the calculation in step 5, the fused feature tensor is dimensionless, and its dimension is... , C L represents the number of feature channels, such as dielectric anomalies and crack depths; L represents the number of layers in the depth direction (corresponding to lining layering); H and W represent the dimensions in the height and width directions, respectively. Z represents the latent variable noise, which follows a normal distribution with respect to i. V is the output voxel tensor, dimensionless (probability value), with dimensions established in three different directions (x, y, z) to form a three-dimensional spatial mesh.

[0052] Modeling thickness error:

[0053] For thickness error, in mm; The mean error is expressed in mm. ,in 1 For linear mapping functions (neural network layers); The standard deviation of the error is in mm. It is a linear mapping function.

[0054] Hollow probability quantification:

[0055] The probability of the existence of a void is dimensionless (0~1). For depth-direction gradient operators, Position weights; The formula used for modeling the risk of steel corrosion is:

[0056] The corrosion risk level is represented by a value ranging from {0, 1, 2, 3, 4} (0 = no corrosion, 4 = severe corrosion). One of the radar features output according to claim 5 is the steel bar reflection intensity feature; i,j These are adjacent voxel pairs; α and β are weight coefficients, where α is the current voxel feature weight and β is the spatial continuity constraint weight, selected according to relevant regulations.

[0057] S7. Engineering Decision Support: Generate quantitative reports that conform to tunnel health assessment standards based on three-dimensional voxel models, and mark high-risk areas and maintenance priorities.

[0058] It should be noted that in step S7, the engineering decision support integrates the thickness error, void probability and rebar risk level information extracted from the three-dimensional voxel model to generate a quantitative report that conforms to the tunnel health assessment specifications.

[0059] Based on the voxel model established in step 6, a two-parameter statistical model can be performed according to the Thickness Quality Index (TQI) calculation principle:

[0060] Let be the thickness quality index of the i-th segment, with a value range of [0, 1], where the closer it is to 1, the better the quality. Same as in step 6, the value is the average thickness error, in mm; Standard deviation of thickness error, mm; The mean threshold is selected according to the specification, in mm; Standard deviation threshold, based on industry experience, in mm; (x) is a hyperbolic secant function, which decays sharply when x>1, thus increasing the risk of exceeding tolerances. Risk level classification:

[0061] Quantification of fractal dimensions of airborne corrosion risk: First, the fractal features of connected components are calculated using the following formula: The fractal dimension of the connected void domain Ωk; The radius of the measured scale is mm, and the radius of the sphere covering the cavity is (initial value is 1 / 2 of the cavity diameter). ) represents the number of voxel points of the cavity within the sphere, as can be seen from the voxel model established according to claim 7; For the void probability field, the probability tensor is calculated in step 6; Multi-parameter risk aggregation formula:

[0062] The risk coefficient of void Ωk, the larger the value, the higher the risk (>1.0 is high risk); max(Pvoid) Maximum probability value within the void domain Area Void projection area A0 standard critical area, According to the specifications, 0.04 m² is the minimum cavity required for repair. The area penalty index amplifies the risk of large voids; according to the formula, the larger the area, the greater the risk. The higher the value, the higher the risk coefficient and the greater the risk.

[0063] Example 2 To verify the engineering adaptability of this invention in real-world scenarios, the lining inspection project of the Jinniushan Tunnel on the Beijing-Shanghai High-Speed ​​Railway was used as the research object to illustrate the application process of this invention in a small-sample inspection and modeling scenario for complex tunnel structures.

[0064] I. Work Area Background The tunnel is named Jinniushan Tunnel (Jinan Section) of the Beijing-Shanghai High-Speed ​​Railway. The arch strata are Jurassic tuff (Class III surrounding rock, with well-developed joints, RQD=65%). The sidewall area is strongly weathered granite (Class IV surrounding rock, softening when exposed to water). The invert foundation is clay mixed with gravel (bearing capacity 120kPa, prone to water seepage). The project is 1.8km long, with chainage ranging from K12+300 to K14+100. The designed lining thickness is 500mm for the arch crown, 450mm for the sidewalls, and 550mm for the invert. C35 concrete is used for the lining, and a double-layer steel reinforcement network of Φ22@150mm is laid for the arch crown.

[0065] Inspection restrictions: The inspection window is from 0:00 to 4:00 every day (only 4 hours of operation time). Due to airspace restrictions, the effective working height under the overhead contact line is <3.5m.

[0066] First, the detection equipment is configured: a dual-frequency ground-penetrating radar is installed on a rigid platform in the middle of the vehicle body, a laser scanner is installed on a rotating gimbal on the roof with an accuracy of 0.3mm, and a GNSS-INS integrated navigation system is installed at the center of gravity of the vehicle with an attitude accuracy of 0.005°. A customized PPS pulse trigger (jitter <1ns) is used in the equipment compartment synchronous controller.

[0067] II. Implementation Process: (1) Step 1, Equipment Co-calibration In this embodiment, the Jinniushan Tunnel (Jinan section) of the Beijing-Shanghai High-Speed ​​Railway is selected as the target case. The system is based on the raw position and attitude data output by the GNSS-INS integrated navigation system. The calibration object requires measuring the spatial position deviation between the phase center of the ground-penetrating radar antenna array and the optical center of the laser scanner. The three-dimensional translation and rotation parameters of the radar antenna and laser scanner in the rigid coordinate system of the inspection vehicle are calculated using a multi-sensor joint adjustment algorithm. Based on the measurements, six sets of measured coordinates of calibration points that meet the requirements are selected. Using these six sets of joint observation data, the following formula is used: Then set the data in , Based on this parameter, the following can be calculated using the pose parameter calibration residual formula: If the data meets the requirements, then N=6 and greater than or equal to 4, which meets the requirements and reaches the optimal level. The PPS second pulse signal of GNSS-INS is used to trigger the hardware clock alignment of all sensors to eliminate transmission delay error.

[0068] (2) Step 2, multi-source data acquisition During the detection process, the detection vehicle was controlled to move at a constant speed of 4 km / h, and the ground penetrating radar was simultaneously collected at two frequency bands: 500 MHz and 1 GHz, to collect electromagnetic reflection signals and sub-millimeter level laser point clouds.

[0069] (3) Step 3, dynamic reference preprocessing The dynamic reference surface correction of the laser point cloud uses the least squares method to fit the design axis equation, and the error range of the lining deformation calculated in real time is within ±0.5 mm; according to the technical principle, the least squares method fits the design axis equation as follows: The convergence value S1 = 0.835m is calculated. Based on the calculation result, the optimal parameters are solved to minimize the objective function:

[0070] The calculation results meet the requirements, and the deformation cloud map is output in real time: , Time-varying gain compensation for radar signals is implemented based on a piecewise calibration curve derived from the frequency-varying dielectric constant of concrete. Radar gain compensation involves first establishing a compensation model based on the frequency-varying dielectric characteristics of concrete, calculating the static dielectric constant, then using an attenuation compensation function to calculate the time-varying gain, and finally multiplying the time-varying gain by the original result to obtain the compensated radar signal. Based on the dielectric frequency-varying characteristic model, the operating frequency of the ground-penetrating radar is selected as 500MHz. Since the concrete is C35... The complex dielectric constant of concrete is , where c= s; depth z=0.55m, time-depth conversion calculation yields: In the construction of the attenuation compensation function, the electromagnetic wave attenuation coefficient is calculated as follows:

[0071] Time-varying gain function:

[0072] Signal compensation: The compensated signal is

[0073] (4) Step S4, multimodal feature extraction A dual-branch network is employed to process radar signal waveforms (including dielectric anomalies and rebar reflection features) and point cloud geometric features (including crack depth and curvature distribution) respectively. For radar signal dual-physical constraint feature extraction, a dielectric anomaly enhancement kernel is used to construct a physical perception convolution kernel based on the Hervyside step function.

[0074] To adaptively adjust based on the abrupt change in concrete dielectric scale: In the formula for the reinforcement reflection dispersion correction model, among which According to the specifications, this value should be between 0.10 and 0.12 m / ns, and this value meets the requirements; among which... Calculated according to the formula The original hole signal amplitude is 0.08V, so the amplitude calculated after convolution is: .

[0075] The formula for matched filtering that performs dispersion compensation in the frequency domain is: } Due to dispersion delay Phase compensation term function , According to calculation At a frequency of 1.0 GHz, the phase distortion before compensation is -0.25, and after compensation it is -0.17, thus reducing the phase distortion. .

[0076] Point cloud stiffness-geometry fusion features in crack depth stiffness-weighted geodesy:

[0077]

[0078] when When the coordinates of discrete points p1 = (120, 50, -10), p2 = (120.3, 50.2, -10.1), then calculate .

[0079] (5) Step S5, structure-guided fusion The structure-guided fusion lining layered attention module is layered according to the design thickness, performs layered feature extraction, and transforms the dispersion characteristics into material aging weight factors. The results are then integrated into the radar, and radar features and point cloud features are also fused according to thickness layering.

[0080] Taking the second and third floors as examples, then It is the characteristic vector of dielectric anomaly, rebar reflection, etc., calculated and output according to step 4; Perform matrix transformations as follows For the lining layer tensor fusion model, the interlayer strain continuity constraint is used to define the interlayer connectivity tensor:

[0081] According to the fusion equation:

[0082] A value >0.95 indicates good interlayer adhesion, while a value <0.9 indicates a risk of peeling (in this example, a value of 0.998 indicates no risk).

[0083] The point cloud crack depth is transformed into a layered stiffness-corrected crack influence function:

[0084] When the number of cracks in the third layer of the arch is m=8; Based on the calculations in step 4, The coordinates of the corresponding point are located based on the curvature gradient in step 4. = (125.3, 48.1, -11.2) is the location of the maximum crack u; According to empirical formulas Calculate the distance, taking the first point as an example. Then perform weight calculation. After weighted summation, the result is calculated according to the formula. .

[0085] Based on the stiffness weighting correction factor: < =0.2mm, in practical engineering, constraints need to be imposed on the project to ensure... ,so =0.

[0086] (6) Step S6, probabilistic reconstruction Probabilistic reconstruction maps fused features to a 3D voxel model using a 3D generative adversarial network. This model includes an error distribution representing thickness uncertainty, the probability of voids, and the risk level of steel corrosion. The corrosion risk level is solved using a Markov random field, allowing for a clearly defined risk level optimization objective. First, voxel space modeling is performed, establishing the generator network:

[0087] Based on the calculation in step 5, the fused feature tensor is dimensionless, and its dimension is... , C The number of characteristic channels, such as dielectric anomalies, crack depth, etc. L This represents the number of layers in the depth direction (corresponding to the lining layers). H and W : These represent the dimensions in the height and width directions, respectively. V Dimensions: 50×50×50 (3D spatial grid).

[0088] Modeling thickness error:

[0089] The corresponding location is the vault; calculation The lining thickness is 1.63mm thinner than the design value, with a standard deviation of 0.86mm (permissible construction error ±5mm).

[0090] Hollow probability quantification formula: gradient calculation Where V[25,30,19]=0.92, V[25,30,21]=0.12, Perform a weighted summation operation. When K=1, the weights... w =[0.2,1.0,0.2], then the calculation yields =0.74>0.7 indicates a high-risk cavity, which requires drilling verification.

[0091] The formula used for modeling the risk of steel corrosion is:

[0092] The value range is {0, 1, 2, 3, 4} (0 = no corrosion, 4 = severe corrosion). =0.91 is one of the radar features output in step 4; weighting coefficients α=0.7, β=0.3, selected according to relevant regulations; Corrosion level of neighboring units Calculate the objective function; calculate according to the formula. If the value is the minimum, it indicates moderate corrosion, requiring a protective coating.

[0093] (7) Step S7, Engineering Decision Support The engineering decision support system integrates information on thickness error, void probability, and rebar risk level extracted from the 3D voxel model to generate a quantitative report that conforms to tunnel health assessment standards. Based on the voxel model established in step 6, a two-parameter statistical model can be performed according to the Thickness Quality Index (TQI) calculation principle.

[0094] mm; =0.6mm; The mean threshold is selected according to the specifications; Standard deviation threshold, based on industry experience; w1 =0.6, w2 =0.4; (1.8 / 5) = 0.896. Let be the thickness quality index of the i-th segment, with a value range of [0,1], where the closer to 1, the better the quality; then calculate... Therefore, it is in a state of alert.

[0095] Risk level classification:

[0096] Quantification of fractal dimensions for airborne corrosion risk: First, calculate the fractal features of connected components. The area penalty index amplifies the risk of large cavities; A0 = 0.04 m² is the standard critical area, which, according to regulations, corresponds to the minimum cavity requiring repair; measurement scale. Number of voxels at 0.01m N ( )=62; then calculate according to the formula =0.896; According to the voxel model established in step 6; For the void probability field, the probability tensor is calculated in step 6; Area =0.045m² is the projected area of ​​the cavity; max ( Pvoid The maximum probability value within the void region is 0.92, calculated using a multi-parameter risk aggregation formula. The risk coefficient of void Ωk is as follows: the larger the value, the higher the risk. A result of <1.0 indicates a medium risk, which requires quarterly inspection and does not require immediate repair.

[0097] Example 3 Please see Figure 2 , Figure 2 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a three-dimensional reconstruction device 401 for a ground-penetrating radar-laser point cloud tunnel lining based on deep learning, a processor 402, and a storage device 403.

[0098] A deep learning-based ground-penetrating radar-laser point cloud tunnel lining three-dimensional reconstruction device 401: The deep learning-based ground-penetrating radar-laser point cloud tunnel lining three-dimensional reconstruction device 401 implements the deep learning-based ground-penetrating radar-laser point cloud tunnel lining three-dimensional reconstruction method.

[0099] Processor 402: The processor 402 loads and executes the instructions and data in the storage device 403 to implement the deep learning-based ground-penetrating radar-laser point cloud tunnel lining three-dimensional reconstruction method.

[0100] Storage device 403: The storage device 403 stores instructions and data; the storage device 403 is used to implement the deep learning-based ground-penetrating radar-laser point cloud tunnel lining three-dimensional reconstruction method.

[0101] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for three-dimensional reconstruction of tunnel lining based on deep learning ground-penetrating radar-laser point cloud, characterized in that: include: The method includes the following steps: S1. Equipment Co-calibration: A multi-band ground-penetrating radar antenna array and a laser scanner are rigidly installed on the tunnel inspection vehicle. A spatiotemporal reference is established through the GNSS-INS integrated navigation system to calibrate the spatial attitude parameters of the multi-band ground-penetrating radar antenna array and the laser scanner. S2. Multi-source data acquisition: Control the detection vehicle to move at a preset speed at a constant speed, and simultaneously acquire dual-band electromagnetic reflection signals from ground penetrating radar and sub-millimeter-level laser point clouds; S3. Dynamic reference preprocessing: Perform dynamic reference surface correction on the laser point cloud, calculate the lining deformation based on the design axis equation and real-time point cloud, and establish an adaptive coordinate system; implement time-varying gain compensation based on the frequency-varying dielectric properties of concrete for the radar signal. S4. Multimodal feature extraction: Radar features are extracted from multi-band radar signals using a pre-trained dual-branch network, including dielectric anomalies, rebar reflection waveforms and dispersion characteristics. Point cloud features are extracted from the corrected point cloud, including crack depth and curvature distribution. S5. Structure-oriented fusion: Through the lining layered attention module, radar features and point cloud features are fused layer by layer according to the design thickness, and material aging weights characterized by dispersion characteristics are injected to obtain fused features; S6. Probabilistic Reconstruction: Input the fused features into the 3D generative adversarial network and output a 3D voxel model containing thickness error distribution, void probability and steel corrosion risk level. S7. Engineering Decision Support: Generate quantitative reports that conform to tunnel health assessment standards based on three-dimensional voxel models, and mark high-risk areas and maintenance priorities.

2. The method for three-dimensional reconstruction of tunnel lining based on deep learning using ground-penetrating radar-laser point cloud as described in claim 1, characterized in that: In step S1, a multi-sensor joint adjustment algorithm is used to calibrate the spatial pose parameters of the multi-band ground-penetrating radar antenna array and the laser scanner, as shown in the following formula: To minimize the spatial positional deviation between the optical center of the laser scanner and the phase center of the radar antenna, mm², where N This represents the total number of sampling points in the point cloud. The coordinates of the optical center of the laser scanner at the i-th calibration point The unit is mm; The coordinates of the phase center of the ground-penetrating radar antenna at the i-th calibration point. The unit is mm; R C This is a dimensionless rotation matrix from the radar coordinate system to the laser coordinate system. The matrix is: Orthogonal matrix; The translation vector from the radar coordinate system to the laser coordinate system The unit is mm.

3. The method for three-dimensional reconstruction of tunnel lining based on deep learning using ground-penetrating radar-laser point cloud as described in claim 1, characterized in that: In step S3, the laser point cloud is dynamically corrected using a reference surface, and the least squares method is used to fit the design axis equation. Where s is the arc length parameter, given the measured point cloud, find the optimal parameters to minimize the objective function: in for The projection point parameters on the design axis are controlled by using an iterative weighted least squares method to suppress outliers and manage errors. The curve length (mm) is calculated along the axis from the starting point. N This represents the total number of sampling points in the point cloud. The 3D coordinates of the i-th point cloud, in mm.

4. The method for three-dimensional reconstruction of tunnel lining based on deep learning using ground-penetrating radar-laser point cloud as described in claim 3, characterized in that: The lining deformation in step S3 is as follows: , The normal deformation at position s of the design axis is in mm, and the normal offset of the lining surface relative to the design axis is also given. s To determine the arc length position on the design axis, in mm, the curve length measured along the axis from the tunnel starting point will be used as the data. m The number of neighboring points, the number of nearby measured points used for interpolation. The arc length of the j-th nearest point, in mm, represents the projected position of the measured point on the design axis. The Gaussian kernel bandwidth parameter, in mm, represents the parameter that controls the influence range of neighboring points; The measured value of the normal deformation of the j-th neighboring point, in mm, represents the normal distance after the point cloud is projected onto the axis.

5. The method for three-dimensional reconstruction of tunnel lining based on deep learning ground-penetrating radar-laser point cloud as described in claim 4, characterized in that: Step S3 involves performing time-varying gain compensation on the radar signal based on the frequency-varying dielectric properties of concrete, specifically as follows: Time-varying gain function: t Electromagnetic wave propagation time, in nanoseconds, is the time from radar transmission to reception. The attenuation coefficient is frequency- and depth-dependent. Time-depth transformation function, unit is m, time t The corresponding electromagnetic wave propagation depth; Signal compensation: Compensated signal: ,in This is the original measured signal, the original voltage signal received by the ground penetrating radar, in V.

6. The method for three-dimensional reconstruction of tunnel lining based on deep learning using ground-penetrating radar-laser point cloud as described in claim 1, characterized in that: In step S4, the extraction of dielectric anomalies employs a dielectric anomaly enhancement kernel, constructing a physics-aware convolution kernel based on the Herveside step function: This formula simulates the abrupt change in dielectric constant at the concrete-void interface, suppresses noise through Gaussian attenuation, and utilizes a physical sensing kernel function. It is a core operator in radar signal processing. The time axis, representing the time delay variable and the electromagnetic wave reflection signal, is expressed using the time parameter of convolution operation, in ns; ) Here, the Herveside step function is used to simulate a sudden change in the dielectric constant. ; Here, represents the Gaussian distribution normalization coefficient, which is the standardization factor of the probability density function, with units of ns. - ¹; It is an adaptive kernel width parameter, which is adaptively adjusted by the abrupt change in the dielectric properties of concrete. The propagation speed of electromagnetic waves in concrete (m / ns) The tuning factor is the type of concrete. The characteristics of the reflected waveform and dispersion properties of the reinforcing bars in step S4 are as follows: } For the rebar reflection signal after dispersion compensation, this formula outputs a high-fidelity rebar reflection waveform, eliminating the phase distortion caused by concrete dispersion, V; s ( t The signal is the original radar time-domain signal, used as the input signal, and its unit is V; F Fourier transform operator, which is the operator that transforms time-domain signals... s ( t Convert to the frequency domain, V / Hz; It is an ideal steel bar reflection model, that is, a physical reflection waveform template of steel bar (time domain). This is a dispersion phase compensation term, which is dimensionless. j For the constructed orthogonal components; The process of extracting crack depth and curvature distribution in step S4 is as follows: Using point cloud stiffness-geometry fusion features in crack depth stiffness-weighted geodesy: in Stiffness-weighted geodesic distance for crack depth, in mm; Integration path; For the local stiffness tensor, N / ; Path differential vector, mm; crack initiation point Coordinates ( mm; Stiffness Tensor S Generated from local curvature features: Principal curvature, ; n It is the surface normal vector; For tensor outer product, T This is the transpose operator. Along the geodetic path γ Integrating the curvature stiffness tensor accurately quantifies the true depth of cracks in curved linings.

7. The method for three-dimensional reconstruction of tunnel lining based on deep learning using ground-penetrating radar-laser point cloud as described in claim 1, characterized in that: The fusion features of step S5 are as follows: The feature vectors fused in the l-th layer of the layered attention module are then output to the 3D reconstruction network. The characteristics of the first layer of radar. Features of the first layer of point cloud; For radar characteristic gradient difference calculation, i.e. Normal vector The axial direction is designed for the layered lining; The exponential decay model for the difference in features between layers has a value range of (0,1], and the larger the value, the stronger the feature continuity between the two layers. The interlaminar strain transfer efficiency strength coefficient is determined based on experimental calibration values, balancing the two contributions. The aging weight of layer l, Stiffness loss threshold, GPa, is set according to the highway tunnel maintenance specification standard. , where is the stiffness variation scale parameter, GPa, and the experimental calibration value.

8. A storage device, characterized in that: The storage device stores instructions and data to implement the three-dimensional reconstruction method for tunnel lining based on deep learning, as described in any one of claims 1 to 7.

9. A three-dimensional reconstruction device for tunnel lining based on deep learning ground-penetrating radar-laser point cloud, characterized in that: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the deep learning-based ground-penetrating radar-laser point cloud tunnel lining three-dimensional reconstruction method as described in any one of claims 1 to 7.

Images