A tunnel lining structure disease three-dimensional model construction method based on sparse survey lines

Abstract

This invention discloses a method for constructing a 3D model of tunnel lining structure defects based on sparse survey lines. First, a multi-survey line feature interaction imaging network based on a semi-supervised mean teacher framework is constructed. Bidirectional convolutional gated recurrent units are used to mine the spatiotemporal correlation between adjacent survey lines, transforming sparse survey line data into dense survey lines. Second, a voxelized spatial mapping method is used to map the dense survey line imaging results onto a 3D mesh, and trilinear interpolation is used to fill the gaps between survey line layers, achieving spatially continuous reconstruction of the 3D morphology of the defects. This invention effectively reduces the density requirements for on-site survey line layout and provides reliable 3D geometric information for the quantitative assessment of tunnel lining structure defects.

Description

Technical Field

[0001] This invention relates to the field of tunnel non-destructive testing technology, specifically to a method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines. Background Technology

[0002] During long-term operation, tunnel lining structures are affected by factors such as changes in geological conditions, construction quality defects, and groundwater erosion, which may lead to hidden defects such as voids, delamination, and lack of compaction, seriously threatening structural safety. Ground penetrating radar (GPR) can detect internal anomalies of the lining in a non-destructive manner by emitting high-frequency electromagnetic pulses into the lining and receiving the reflected echoes, and has become the mainstream technical means for detecting defects in operating tunnels.

[0003] By deploying multiple parallel survey lines along the longitudinal or circumferential direction of the tunnel for B-Scan scanning and spatially stacking the resulting two-dimensional profiles, three-dimensional structural information of the lining can be obtained. However, according to the spatial sampling theorem (i.e., a generalization of the Nyquist–Shannon sampling theorem in the spatial domain), in order to reconstruct the three-dimensional morphology of the target without distortion, the spacing between survey lines is crucial. Must meet:

[0004]

[0005] in The smallest feature scale of the target in the direction perpendicular to the survey line (subscript) (This refers to the feature). For typical volumetric defects in tunnel lining... Typically, the spacing between test lines is several centimeters to tens of centimeters, requiring the spacing to be at the centimeter level. However, in actual engineering testing, due to multiple constraints such as time, manpower, and safety, the spacing between test lines is usually much larger than this theoretical requirement, resulting in sparse sampling.

[0006] For 3D reconstruction under sparse sampling conditions, existing technologies have the following shortcomings: First, traditional linear or spline interpolation methods directly interpolate sparse survey lines, which easily leads to overly smooth reconstruction results and blurred disease boundaries; Second, although end-to-end deep learning methods can learn from prior data, they usually rely on large-scale, high-quality labeled samples, while the labeling cost of measured data for tunnel lining is high and difficult to obtain; Third, there are significant differences in data distribution between the source domain (simulation or indoor specimens) and the target domain (field measurements), and the generalization performance of directly transferring trained models is often insufficient; Fourth, existing methods do not fully explore the spatial correlation between adjacent survey lines, and the accuracy of disease inference in unobserved areas is limited.

[0007] Therefore, there is an urgent need for a method that can achieve high-precision and robust three-dimensional reconstruction of defects under engineering realities characterized by sparse survey lines, scarce annotations, and significant differences in cross-domain distribution. Summary of the Invention

[0008] The present invention aims to provide a method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines, in order to solve the above problems.

[0009] The technical solution of this invention is: a method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines, comprising the following steps:

[0010] S1, collecting sparse survey line GPR data; including sparsely laying parallel survey lines along the tunnel lining structure according to the survey line spacing and collecting ground penetrating radar B-Scan data to form a sparse survey line GPR data sequence;

[0011] S2, input the sparse survey line GPR data sequence into the multi-survey line feature interactive imaging network composed of the feature extraction module and the spatiotemporal interaction module based on Bi-ConvGRU;

[0012] The feature extraction module uses an encoder-decoder type convolutional neural network to extract deep semantic feature maps in parallel for each B-Scan data, forming a multi-line feature sequence; the spatiotemporal interaction module performs bidirectional spatiotemporal correlation aggregation on the multi-line feature sequence, infers the disease distribution characteristics at the virtual intermediate line position, and outputs a dense line dielectric constant distribution map sequence.

[0013] S3, Semi-supervised Adversarial Consistency Training; includes employing a semi-supervised learning strategy based on the mean teacher framework during the training phase, dividing the network into student networks. Teacher Network The student network updates parameters through gradient backpropagation. The teacher network obtains parameters by updating the historical parameters of the student network using an exponential moving average. Simultaneously, a source-target domain discriminator is introduced. Perturbation Consistency Discriminator The dual discriminator structure is used for adversarial consistency training, and a joint consistency regularization constraint in the spatial and frequency domains is applied. The training loss of the student network is composed of the source domain supervision loss, the joint spatial-frequency domain consistency regularization loss, and the adversarial loss, and is alternately optimized with the discriminator loss under the minimax game framework. A sequence of dense test line dielectric constant distribution maps is generated.

[0014] S4 will train the teacher network. The output sequence of dielectric constant distribution maps of dense survey lines is mapped to a three-dimensional voxel mesh according to the spatial coordinates of each survey line using a voxelization spatial mapping method. Establish a discrete voxel representation of the disease target;

[0015] S5 uses trilinear interpolation to continuously fill the gaps between survey lines in the three-dimensional voxel grid, thus obtaining a three-dimensional continuous model of tunnel lining structural defects.

[0016] Preferably, in S2, the feature extraction module adopts a U-shaped architecture, which includes a four-downsampled encoder, a four-upsampled decoder, and skip connections between corresponding levels of the encoder and decoder.

[0017] The encoder consists of 5 convolutional blocks cascaded with 4 max-pooling layers; Each convolutional block contains two Convolutional layers, each followed by instance normalization and ReLU activation; the number of channels in the output feature map of the convolutional block is... ,satisfy , ;No. Each pooling layer uses a step size of 2. Max pooling halves the feature map space size.

[0018] The decoder includes 4 The transposed convolutional layer alternates with 4 convolutional blocks; the first The transposed convolutional layer doubles the spatial size of the feature map and halves the number of channels; its output is the same as that of the encoder. The outputs of the convolutional blocks are concatenated along the channel dimension using skip connections, and used as the inputs to the corresponding convolutional blocks of the decoder;

[0019] pass The convolutional layer maps multi-channel features into a prediction map with the number of channels equal to the number of categories, and then adjusts it to the same spatial scale as the label through bilinear interpolation;

[0020] Applying the feature extraction module to multi-line GPR data sequences For each data point, multi-line feature sequences are output in parallel. ,in , , , These represent the height, width, and number of channels of the feature map, respectively, with subscripts... , , These correspond to height, width, and channel, respectively.

[0021] Preferably, in S2, the Bi-ConvGRU module includes a forward ConvGRU layer, a backward ConvGRU layer, and a feature fusion convolutional layer;

[0022] The forward ConvGRU layer processes the feature sequence along the increasing direction of the survey line and outputs the forward hidden state sequence. (subscript) (meaning forward);

[0023] The backward ConvGRU layer processes the feature sequence along the decreasing direction of the survey line and outputs the backward hidden state sequence. (subscript) (indicates backward)

[0024] By splicing the two along the channel dimension, we obtain Then, bidirectional context-enhanced features are obtained through a feature fusion convolutional layer;

[0025] Subsequently Feature fusion convolutional layers perform adaptive weighting and nonlinear transformations to obtain the corresponding feature. Two-way contextual enhancement features for strip survey lines; all The enhanced characteristics of the test lines constitute a sequence of dielectric constant distribution maps of dense test lines, denoted as... ,in This indicates the imaging output.

[0026] Preferably, in S2, the ConvGRU unit is in the first... The update equation for each time step is:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] in To update the door, subscript Indicates update gate; To reset the door, subscript Indicates resetgate; In the candidate hidden state, Currently in a hidden state. For the first Input feature maps at each time step; For learnable convolutional kernel weights, the first letter of the subscript indicates the input source. Indicates input features, This indicates the hidden state at the previous moment, and the second letter of the subscript indicates the object being applied to. Indicates updating the door, Indicates resetting the door, Indicates the candidate hidden state; For the corresponding bias; This represents the convolution operation. This represents the Hadamard element-wise product. This represents the Sigmoid activation function.

[0033] Preferably, in S3, the parameter update rule for the teacher network is as follows:

[0034]

[0035] in and They represent the first Step and the first Step-by-step teacher network parameters, Indicates the first Step student network parameters, The smoothing coefficient of the exponential moving average, subscript Indicates the training iteration step index. Values ;

[0036] In S3, the dual discriminator adopts a multi-task architecture with a shared feature extraction backbone and dual-task discriminator heads: the first three convolutional layers form a shared backbone to extract domain-independent general discriminative features, and at the output of the third layer, it splits into two independent discriminator heads, namely the source domain discriminator and the target domain discriminator. Perturbation Consistency Discrimination Head .

[0037] Preferably, in S3, the overall loss function of the student network It consists of three weighted parts:

[0038]

[0039] in For source domain monitoring loss, For consistency regularization loss, To combat loss in generators; and These are the weighting coefficients for the consistency loss and the adversarial loss, respectively; the supervision loss is determined by the cross-entropy loss. With Dice loss The sum of them constitutes the whole.

[0040] Preferably, source domain supervision loss Cross-entropy loss With Dice loss composition:

[0041]

[0042]

[0043] in For the first The predicted probability of each pixel. To correspond to the actual label value, Total number of pixels To prevent division by zero smoothing constant;

[0044] Consistency regularization loss Using a combined spatial and frequency domain approach:

[0045]

[0046] in and These are the predicted values ​​under undisturbed and disturbed conditions, respectively. It is a two-dimensional discrete Fourier transform. These are the frequency domain constraint strength coefficients;

[0047] The generator adversarial loss is implemented using a least-squares GAN approach.

[0048]

[0049] in Represents the mathematical expectation. This represents the distribution of imaging results in the undisturbed target domain.

[0050] Preferably, S3 further includes:

[0051] The discriminator loss, used to train the discriminator network, enables it to distinguish imaging results from different sources or states, thereby providing the generator with effective adversarial training signals. These are as follows:

[0052]

[0053]

[0054] in This represents the source domain imaging results;

[0055] The overall optimization objective is a minimax game, defining the game relationships and final optimization objective throughout the training process, unifying the training of the generator and the two discriminators within a minimax framework:

[0056]

[0057] For the consistency term weight coefficient, For the weighting coefficients of the adversarial terms;

[0058] The training phase alternates between two steps: the first step fixes the discriminator parameters, and then... Optimize student network; the second step is to fix student network parameters, according to... and Optimize the discriminator;

[0059] The teacher network is updated synchronously according to the EMA rules after each student network update; only the teacher network is used during the inference phase. By inputting the sequence of sparse test lines to be tested, the corresponding sequence of dielectric constant distribution maps of dense test lines can be output.

[0060] Preferably, in S4, the voxelization mapping is established using... The axis is the direction of detection depth. The axis represents the direction of the survey line. A three-dimensional voxel mesh with the axis being the horizontal section:

[0061]

[0062] in, This represents a voxel unit located at voxel coordinates. , , The total number of voxels in the three coordinate axes are respectively; the first Dense survey lines Axis physical coordinates according to Confirmed, among which This is the starting position of the survey line. The virtual survey line spacing set during network reconstruction;

[0063] The two-dimensional imaging results of the survey line Each pixel in Mapping to 3D voxels :

[0064]

[0065] Subscript Indicates pixel. This indicates the voxel layer index corresponding to the survey line. This indicates a floor operation; after all dense survey lines are mapped, an initial three-dimensional voxel grid is obtained, which carries complete two-dimensional imaging information on the survey line layer, while the interlayer regions remain initially blank.

[0066] Preferably, trilinear interpolation is performed along... , , The axes are sequentially subjected to three linear weightings, with the normalized distance weights for the three directions being as follows: , and ,in For the physical coordinates of the target point, The physical coordinates of the smallest index among the 8 known voxels surrounding the target point; specifically including:

[0067] S51, along Axial linear interpolation yields four intermediate values:

[0068]

[0069]

[0070]

[0071]

[0072] in, For the target point at Normalized distance weights along the axis, The physical reference point Coordinates; the intermediate value subscripts "00", "01", "10", and "11" respectively represent Offset combination, first bit corresponding Axis offset, second position corresponding Axis offset;

[0073] S52, along Axial linear interpolation yields two intermediate values:

[0074]

[0075]

[0076] in, For the target point at Normalized distance weights along the axis, The physical reference point Coordinates; the subscripts "0" and "1" for intermediate values ​​represent... Is the axis offset?

[0077] S53, along Linear interpolation is used to obtain the final interpolation result:

[0078]

[0079] in, For the target point at Normalized distance weights along the axis, subscript With weighted terms Distinguish this Frequency domain weights in consistency loss ; The physical reference point coordinate.

[0080] The beneficial effects of this invention are as follows:

[0081] First, by mining the bidirectional spatial continuity constraints of adjacent survey lines through the Bi-ConvGRU spatiotemporal interaction module, the distribution of defects at unobserved locations can be inferred under sparse sampling conditions, significantly relaxing the requirements for field survey line density.

[0082] Second, based on the Mean-Teacher semi-supervised framework, cross-domain transfer can be achieved with only a small number of target domain annotations, significantly reducing annotation costs.

[0083] Third, the dual-discriminator adversarial structure simultaneously achieves two types of constraints: "global domain distribution alignment" and "local perturbation robustness," which is more comprehensive than the single-discriminator scheme.

[0084] Fourth, the spatial-frequency domain joint consistency loss explicitly protects high-frequency information such as disease boundaries, avoiding excessive smoothing that can easily occur during adversarial training.

[0085] Fifth, the geometric-physical synergy between voxelization and trilinear interpolation enables the final 3D model to have both pixel-level local accuracy and spatial continuity. Attached Figure Description

[0086] Figure 1 A flowchart illustrating a method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines, provided in an embodiment of the present invention;

[0087] Figure 2 A flowchart outlining the teacher-student collaborative transfer network provided in this embodiment of the invention;

[0088] Figure 3 A schematic diagram of a semi-supervised imaging architecture provided in an embodiment of the present invention;

[0089] Figure 4 This is a schematic diagram of the feature extraction module architecture provided in an embodiment of the present invention;

[0090] Figure 5 This is a schematic diagram of the spatiotemporal interaction module architecture provided in an embodiment of the present invention. Detailed Implementation

[0091] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. The embodiments of the present invention are not limited thereto.

[0092] Example 1

[0093] Step 1: Sparse Line GPR Data Acquisition

[0094] Layout along the longitudinal (or circumferential) direction of the tunnel There are 3 parallel survey lines, and the distance between adjacent survey lines is denoted as . (Unit: m). Along each survey line, the ground-penetrating radar antenna moves at a constant speed along the survey line direction, according to spatial sampling intervals. (Unit: m) Trigger transmit-receive, each sampling channel in the two-way travel direction at time sampling intervals. (Unit: nanoseconds) Record signal. (Number of ns) The two-dimensional B-Scan data obtained from the survey line are denoted as ,in This represents the total number of time samples per channel. This represents the total number of traces along the spatial direction of the survey line; superscript. Indicates the survey line sequence index. All survey line data together constitute the sparse survey line GPR data sequence. .

[0095] To facilitate network training and cross-condition migration, the data is divided into a source domain and a target domain: source domain data Derived from simulation or standard specimens, with sufficient annotation information and subscripts. Indicates source; target domain data Sourced from actual field conditions, containing only a small number of labeled samples (subscripts) (representing target) and a large number of unlabeled samples (subscript) (Indicates unlabeled).

[0096] Step 2: Multi-line Feature Interactive Imaging Network Structure

[0097] The overall network consists of a feature extraction module and a spatiotemporal interaction module connected in series. The overall process is shown in Figure 2. In the structure shown in the figure, the multi-line feature interaction imaging network consists of a feature extraction module and a spatiotemporal interaction module based on Bi-ConvGRU connected in series.

[0098] The feature extraction module adopts an encoder-decoder type convolutional neural network (U-shaped architecture). The encoder extracts the deep semantic features of each B-Scan data in the input sparse survey line GPR data sequence through four downsampling steps. The decoder restores the spatial resolution through four upsampling steps combined with skip connections and outputs a multi-survey line deep semantic feature map sequence in parallel.

[0099] The spatiotemporal interaction module receives the multi-line feature sequence output by the feature extraction module. It contains a forward ConvGRU layer, a backward ConvGRU layer, and a feature fusion convolutional layer. The forward ConvGRU layer aggregates context information along the increasing direction of the test lines, and the backward ConvGRU layer aggregates context information along the decreasing direction of the test lines. The bidirectional hidden states output by the two are concatenated along the channel dimension and then adaptively weighted and fused by the feature fusion convolutional layer to finally output a dense test line dielectric constant distribution map sequence.

[0100] The semi-supervised training architecture is shown in Figure 3. In the architecture shown, the semi-supervised adversarial consistency training architecture is built based on the mean teacher framework and includes a student network G. stu Teacher Network G tea And the source and target domain discriminator D su And perturbation consistency discriminator D uu The dual discriminator structure is constructed from ′.

[0101] The student network Gstu updates the parameters θ through gradient backpropagation. stu Teacher Network G tea The parameter θ tea By synchronously updating the historical parameters of the student network using an exponential moving average (EMA), they do not participate in gradient backpropagation.

[0102] In the training data flow, labeled data from the source domain is processed by the student network to generate source domain imaging results, which are then compared with the real labels to calculate the supervised loss L. S (Including cross-entropy loss and Dice loss); Unlabeled data in the target domain are input into the student network in both unperturbed and perturbed versions to generate corresponding target domain imaging results. The spatial-frequency domain joint consistency regularization loss L is calculated between the two. cons .

[0103] The two discriminators share a main feature extraction backbone, which then splits into two independent discriminator heads: Discriminator D. su Discriminator D is used to distinguish between source domain imaging results and undisturbed target domain imaging results. uu ′ is used to distinguish between undisturbed imaging results and perturbed imaging results in the target domain; the student network uses adversarial loss L GD It forms a minimax game with the two discriminators.

[0104] During training, the student network is optimized with fixed discriminator parameters, and the discriminator is optimized with fixed student network parameters, alternating between the two methods. The discriminator loss is denoted as L. Dsu With L Duu ′. Figure 3 It is also shown that the reasoning phase only uses the teacher network G. tea Input the sparse test line sequence to be tested and the corresponding dense test line dielectric constant distribution map sequence will be output. The dual discriminator and consistency regularization module do not participate in the forward calculation during the inference stage.

[0105] 2.1 Feature Extraction Module

[0106] The feature extraction module adopts a U-shaped encoder-decoder architecture, as shown in Figure 4. The module input is a single measurement line B-Scan data, and the output is the corresponding deep semantic feature map.

[0107] The encoder consists of cascaded 5 double convolutional blocks and 4 max pooling layers. Convolutional blocks ( ) consisting of two cascaded The system consists of convolutional layers, each followed by Instance Normalization (IN) and ReLU activation functions; the number of channels in the output feature map of this convolutional block is... ,satisfy , (That is, 64, 128, 256, 512, 1024 in sequence). The... Pooling layers ( Using a step size of 2 Max pooling halves the size of the feature map space.

[0108] The decoder is symmetrical to the encoder and consists of 4 The transpose convolutional layer is composed of four alternating convolutional blocks. layer( Transposed convolution doubles the feature map spatial size and halves the number of channels; its output is the same as the encoder's output. The outputs of the convolutional blocks are concatenated along the channel dimension using skip connections, and used as the inputs to the corresponding convolutional blocks of the decoder.

[0109] Finally passed The convolutional layer maps multi-channel features into a prediction map with the number of channels equal to the number of categories (for binary classification of disease area and background, the number of output channels is 2), and then adjusts it to the same spatial scale as the label through bilinear interpolation.

[0110] Apply the above modules to multi-line GPR data sequences For each data point, multi-line feature sequences are output in parallel. ,in , , , These represent the height, width, and number of channels of the feature map, respectively (subscripts). , , These correspond to height, width, and channel, respectively.

[0111] 2.2 Spatiotemporal Interaction Module Based on Bi-ConvGRU

[0112] To fully exploit the spatial continuity between adjacent survey lines, the module uses characteristic sequences. For input, enter the survey line number. Treated as a sequential index (equivalent to a "time step" in the ConvGRU framework). The structure is shown in Figure 5.

[0113] The core update equation for the ConvGRU cell has been given in claim 4, and the meaning of each variable is further explained here: For the first The time step (i.e., the ) The input feature map of the test line, For the first The hidden state of the step, and These are the update door and the reset door, respectively. In the candidate hidden state, Currently hidden. Gating weight. The subscript uses the two-character convention of "input source - target": Indicates the current input feature source, Indicates the source of historical hidden states; , , These correspond to the update gate, reset gate, and candidate state, respectively. This is a learnable bias term.

[0114] This module contains two bidirectional ConvGRU computation streams:

[0115] (a) Forward ConvGRU along the increasing direction of the survey line ( Process the feature sequence to obtain the forward hidden state sequence. (arrow With subscript (Together indicating the forward direction);

[0116] (b) Reverse ConvGRU along the decreasing direction of the survey line ( Process the feature sequence to obtain the backward hidden state sequence. (arrow With subscript (Together representing the backward direction);

[0117] The two are spliced ​​along the channel dimension:

[0118]

[0119] Subsequently Feature fusion convolutional layers perform adaptive weighting and nonlinear transformations to obtain the corresponding feature. Two-way contextual enhancement features for strip survey lines. All The enhanced characteristics of the test lines constitute a sequence of dielectric constant distribution maps of dense test lines, denoted as... ,in This represents the imaging output (dielectric constant distribution map).

[0120] Step 3: Semi-supervised adversarial consistency training strategy

[0121] 3.1 Mean-Teacher Framework

[0122] Build student networks with identical structures Teacher Network The parameters of the two are denoted as follows: and The student network is updated via gradient backpropagation; the teacher network is updated via EMA rules.

[0123]

[0124] in For the first Teacher parameters after one iteration Smoothing coefficient (typical value) ), subscript This represents the training iteration step index.

[0125] 3.2 Dual Discriminator Adversarial Structure

[0126] Design a multi-task discriminator with a shared backbone and dual discriminator heads:

[0127] (a) The shared backbone consists of 3 convolutional layers, each of which is The number of channels are 64, 128, and 256 respectively (of which... Indicates the kernel size. Indicates step size stride, (Indicates padding);

[0128] (b) Source-target domain discrimination head : Connected in series after sharing the backbone (512 channels) and (1 channel) Output the binary discrimination probability of "source domain / target domain", index This indicates a source-unlabeled-target discrimination task;

[0129] (c) Perturbation Consistency Discrimination Head : Connected in series after sharing the backbone (512 channels) and (1 channel) Outputs the binary probability of "unperturbed / perturbed", with the index... This indicates the unlabeled–unlabeled-perturbed discrimination task.

[0130] For unlabeled target domain data Apply random perturbation (subscript) (Indicates the perturbation type index, including additive Gaussian noise, random occlusion, amplitude jitter, etc.) to obtain the perturbation version. ;Will Input teacher network to get ,Will Input teacher network to get .

[0131] 3.3 Loss Function

[0132] The total network loss for students is:

[0133]

[0134] (1) Monitoring loss Cross-entropy loss With Dice loss composition:

[0135]

[0136]

[0137] in For the first The predicted probability of each pixel. To correspond to the actual label value, Total number of pixels To prevent division by zero, a smoothing constant (typical value) ).

[0138] (2) Consistency loss Using a combined spatial and frequency domain approach:

[0139]

[0140] in and These are the predicted values ​​under undisturbed and disturbed conditions, respectively. It is a two-dimensional discrete Fourier transform. Frequency domain constraint strength coefficient (typical values) ).

[0141] (3) The adversarial loss adopts the least squares GAN form:

[0142]

[0143] in Represents the mathematical expectation. This represents the distribution of imaging results in the undisturbed target domain.

[0144] The discriminator losses are as follows:

[0145]

[0146]

[0147] in This indicates the source domain imaging results.

[0148] The overall optimization objective is a minimax game:

[0149]

[0150] Weighting coefficient (Consistency item weight) (Weight of adversarial terms) The subscript numbers are only for distinguishing ordinal values; typical values ​​are... , .

[0151] 3.4 Training and Reasoning Process

[0152] The training phase alternates between two steps: the first step fixes the discriminator parameters, and then... Optimize student network; the second step is to fix student network parameters, according to... and Optimize the discriminator. The teacher network is synchronized with the student network according to the EMA rules after each step of the student network update. Only the teacher network is used during the inference phase. By inputting the sequence of sparse test lines to be tested, the corresponding sequence of dielectric constant distribution maps of dense test lines can be output.

[0153] Step 4: Voxelization and 3D Spatial Mapping

[0154] Build with The axis is the direction of detection depth. The axis represents the direction of the survey line. The axis is a three-dimensional Cartesian coordinate system with the horizontal direction of the cross-section. The three-dimensional voxel mesh is defined as follows:

[0155]

[0156] in coordinates voxel units at the location, , , These represent the number of voxels along the three coordinate axes (subscripts correspond to the coordinate axes). Voxel physical dimensions in each direction. , , The vertical resolution of the GPR system, the spacing between virtual survey lines reconstructed by the network, and the scanning sampling interval are determined respectively.

[0157] For the Dense survey lines ( ),That The physical coordinates of the axis are:

[0158]

[0159] in These are the coordinates of the starting position of the survey line. This represents the virtual survey line spacing. The two-dimensional imaging result of this survey line... Each pixel in Mapping to 3D voxels :

[0160]

[0161] Subscript Indicates pixel. This indicates the voxel layer index (subscript) corresponding to the measurement line. (representing voxel layer) This indicates a floor operation. After mapping all dense survey lines, an initial three-dimensional voxel mesh is obtained, which carries complete two-dimensional imaging information on the survey line layer, while the interlayer regions remain initially blank.

[0162] Step 5: Trilinear Interpolation Continuous Reconstruction

[0163] To fill the interlayer gaps into a spatially continuous three-dimensional shape, trilinear interpolation is performed on the initial voxel mesh. For any voxel to be interpolated in the target mesh... Its physical coordinates are Locate the eight known voxels surrounding this position, with indices as follows: , , , , , , , subscript This indicates the reference starting point (i.e., the corner point with the smallest index in the enclosing cube).

[0164] First step along Axial linear interpolation yields four intermediate values:

[0165]

[0166]

[0167]

[0168]

[0169] in For the target point at Normalized distance weights along the axis, The physical reference point Coordinates; the intermediate value subscripts "00", "01", "10", and "11" respectively represent Offset combination (first digit corresponds) Axis offset, second position corresponding (axis offset).

[0170] Step Two Axial linear interpolation yields two intermediate values:

[0171]

[0172]

[0173] in For the target point at Normalized distance weights along the axis, The physical reference point Coordinates; the subscripts "0" and "1" for intermediate values ​​represent... Is the axis offset?

[0174] Step 3 Linear interpolation is used to obtain the final interpolation result:

[0175]

[0176] in For the target point at Normalized distance weights along the axis (subscript) With weighted terms Distinguish this Frequency domain weights in consistency loss ), The physical reference point coordinate.

[0177] By traversing all voxels to be interpolated in the target mesh, a spatially continuous three-dimensional voxel model is obtained. Finally, a disease determination threshold is set. (The subscript means threshold) By binarizing the voxel values ​​and separating the diseased area from the background area, a three-dimensional geometric model of the tunnel lining structure disease can be output, which can then support the calculation of quantitative evaluation indicators such as disease volume, maximum depth, and spatial orientation.

[0178] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines, characterized in that, Includes the following steps: S1, collecting sparse survey line GPR data; including sparsely laying parallel survey lines along the tunnel lining structure according to the survey line spacing and collecting ground penetrating radar B-Scan data to form a sparse survey line GPR data sequence; S2, input the sparse survey line GPR data sequence into the multi-survey line feature interactive imaging network composed of the feature extraction module and the spatiotemporal interaction module based on Bi-ConvGRU; The feature extraction module uses an encoder-decoder type convolutional neural network to extract deep semantic feature maps in parallel for each B-Scan data, forming a multi-line feature sequence; the spatiotemporal interaction module performs bidirectional spatiotemporal correlation aggregation on the multi-line feature sequence, infers the disease distribution characteristics at the virtual intermediate line position, and outputs a dense line dielectric constant distribution map sequence. S3, Semi-supervised Adversarial Consistency Training; includes employing a semi-supervised learning strategy based on the mean teacher framework during the training phase, dividing the network into student networks. Teacher Network The student network updates parameters through gradient backpropagation. The teacher network obtains parameters by updating the historical parameters of the student network using an exponential moving average. Simultaneously, a source-target domain discriminator is introduced. Perturbation Consistency Discriminator The dual discriminator structure is used for adversarial consistency training, and a joint consistency regularization constraint in the spatial and frequency domains is applied. The training loss of the student network is composed of the source domain supervision loss, the joint spatial-frequency domain consistency regularization loss, and the adversarial loss, and is alternately optimized with the discriminator loss under the minimax game framework. A sequence of dense test line dielectric constant distribution maps is generated. S4 will train the teacher network. The output sequence of dielectric constant distribution maps of dense survey lines is mapped to a three-dimensional voxel mesh according to the spatial coordinates of each survey line using a voxelization spatial mapping method. Establish a discrete voxel representation of the disease target; S5 uses trilinear interpolation to continuously fill the gaps between survey lines in the three-dimensional voxel grid, thus obtaining a three-dimensional continuous model of tunnel lining structural defects.

2. The method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines according to claim 1, characterized in that, In S2, the feature extraction module adopts a U-shaped architecture, which includes a four-downsampled encoder, a four-upsampled decoder, and skip connections between corresponding levels of the encoder and decoder. The encoder consists of 5 convolutional blocks cascaded with 4 max-pooling layers; Each convolutional block contains two Convolutional layers, each followed by instance normalization and ReLU activation; the number of channels in the output feature map of the convolutional block is... ,satisfy , ;No. Each pooling layer uses a step size of 2. Max pooling halves the feature map space size. The decoder includes 4 The transposed convolutional layer alternates with 4 convolutional blocks; the first The transposed convolutional layer doubles the spatial size of the feature map and halves the number of channels; its output is the same as that of the encoder. The outputs of the convolutional blocks are concatenated along the channel dimension using skip connections, and used as the inputs to the corresponding convolutional blocks of the decoder; pass The convolutional layer maps multi-channel features into a prediction map with the number of channels equal to the number of categories, and then adjusts it to the same spatial scale as the label through bilinear interpolation; Applying the feature extraction module to multi-line GPR data sequences For each data point, multi-line feature sequences are output in parallel. ,in , , , These represent the height, width, and number of channels of the feature map, respectively, with subscripts... , , These correspond to height, width, and channel, respectively.

3. The method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines according to claim 1, characterized in that, In S2, the Bi-ConvGRU module includes a forward ConvGRU layer, a backward ConvGRU layer, and a feature fusion convolutional layer; The forward ConvGRU layer processes the feature sequence along the increasing direction of the survey line and outputs the forward hidden state sequence. (subscript) (meaning forward); The backward ConvGRU layer processes the feature sequence along the decreasing direction of the survey line and outputs the backward hidden state sequence. (subscript) (indicates backward) By splicing the two along the channel dimension, we obtain Then, bidirectional context-enhanced features are obtained through a feature fusion convolutional layer; Subsequently Feature fusion convolutional layers perform adaptive weighting and nonlinear transformations to obtain the corresponding feature. Two-way contextual enhancement features for strip survey lines; all The enhanced characteristics of the test lines constitute a sequence of dielectric constant distribution maps of dense test lines, denoted as... ,in This indicates the imaging output.

4. The method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines according to claim 3, characterized in that, In S2, the ConvGRU unit is in the... The update equation for each time step is: in To update the door, subscript Indicates update gate; To reset the door, subscript This indicates a reset gate; In the candidate hidden state, Currently in a hidden state. For the first Input feature maps at each time step; For learnable convolutional kernel weights, the first letter of the subscript indicates the input source. Indicates input features, This indicates the hidden state at the previous moment, and the second letter of the subscript indicates the object being applied to. Indicates updating the door, Indicates resetting the door, Indicates the candidate hidden state; For the corresponding bias; This represents the convolution operation. This represents the Hadamard element-wise product. This represents the Sigmoid activation function.

5. The method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines according to claim 1, characterized in that, In S3, the parameter update rules for the teacher network are as follows: in and They represent the first Step and the first Step-by-step teacher network parameters, Indicates the first Step student network parameters, The smoothing coefficient of the exponential moving average, subscript Indicates the training iteration step index. Values ; In S3, the dual discriminator adopts a multi-task architecture with a shared feature extraction backbone and dual-task discriminator heads: the first three convolutional layers form a shared backbone to extract domain-independent general discriminative features, and at the output of the third layer, it splits into two independent discriminator heads, namely the source domain discriminator and the target domain discriminator. Perturbation Consistency Discrimination Head .

6. The method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines according to claim 1, characterized in that, In S3, the overall loss function of the student network Weighted by three parts composition: in For source domain monitoring loss, For consistency regularization loss, To combat loss in generators; and These are the weighting coefficients for the consistency loss and the adversarial loss, respectively; the supervision loss is determined by the cross-entropy loss. With Dice loss The sum of them constitutes the whole.

7. The method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines according to claim 6, characterized in that, Source domain monitoring loss Cross-entropy loss With Dice loss composition: in For the first The predicted probability of each pixel. To correspond to the actual label value, Total number of pixels To prevent division by zero smoothing constant; Consistency regularization loss Using a combined spatial and frequency domain approach: in and These are the predicted values ​​under undisturbed and disturbed conditions, respectively. It is a two-dimensional discrete Fourier transform. These are the frequency domain constraint strength coefficients; The generator adversarial loss is implemented using a least-squares GAN approach. in Represents the mathematical expectation. This represents the distribution of imaging results in the undisturbed target domain.

8. The method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines according to claim 6, characterized in that, S3 also includes: The discriminator loss, used to train the discriminator network, enables it to distinguish imaging results from different sources or states, thereby providing the generator with effective adversarial training signals. These are as follows: in This represents the source domain imaging results; The overall optimization objective is a minimax game, defining the game relationships and final optimization objective throughout the training process, unifying the training of the generator and the two discriminators within a minimax framework: For the consistency term weight coefficient, For the weighting coefficients of the adversarial terms; The training phase alternates between two steps: the first step fixes the discriminator parameters, and then... Optimize student network; the second step is to fix student network parameters, according to... and Optimize the discriminator; The teacher network is updated synchronously according to the EMA rules after each student network update; only the teacher network is used during the inference phase. By inputting the sequence of sparse test lines to be tested, the corresponding sequence of dielectric constant distribution maps of dense test lines can be output.

9. The method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines according to claim 1, characterized in that, In S4, voxelization mapping is established based on The axis is the direction of detection depth. The axis represents the direction of the survey line. A three-dimensional voxel mesh with the axis being the horizontal section: in, This represents a voxel unit located at voxel coordinates. , , The total number of voxels in the three coordinate axes are respectively; the first Dense survey lines Axis physical coordinates according to Confirmed, among which This is the starting position of the survey line. The virtual survey line spacing set during network reconstruction; The two-dimensional imaging results of the survey line Each pixel in Mapping to 3D voxels : Subscript Indicates pixel. This indicates the voxel layer index corresponding to the survey line. This indicates a floor operation; after all dense survey lines are mapped, an initial three-dimensional voxel grid is obtained, which carries complete two-dimensional imaging information on the survey line layer, while the interlayer regions remain initially blank.

10. The method for constructing a three-dimensional model of tunnel lining structure defects based on sparse survey lines according to claim 1, characterized in that, Trilinear interpolation is performed along... , , The axes are sequentially subjected to three linear weightings, with the normalized distance weights for the three directions being as follows: , and ,in For the physical coordinates of the target point, The physical coordinates of the smallest index among the 8 known voxels surrounding the target point; specifically including: S51, along Axial linear interpolation yields four intermediate values: in, For the target point at Normalized distance weights along the axis, The physical reference point Coordinates; the intermediate value subscripts "00", "01", "10", and "11" respectively represent Offset combination, first bit corresponding Axis offset, second position corresponding Axis offset; S52, along Axial linear interpolation yields two intermediate values: in, For the target point at Normalized distance weights along the axis, The physical reference point Coordinates; the subscripts "0" and "1" for intermediate values ​​represent... Is the axis offset? S53, along Linear interpolation is used to obtain the final interpolation result: in, For the target point at Normalized distance weights along the axis, subscript With weighted terms Distinguish this Frequency domain weights in consistency loss ; The physical reference point coordinate.

Images