A shield tunnel leakage automatic inspection system and method based on adaptive training optimization

Abstract

The application discloses a kind of based on adaptive training optimization's shield tunnel leakage automated inspection system and method, comprising: obtaining the image data of shield tunnel and pre-processing, reconstruct sparse three-dimensional point cloud based on the image data after pre-processing;Based on sparse three-dimensional point cloud obtains tunnel axis parameter;After segmentation to tunnel axis, section profile fitting is carried out, obtains shield tunnel dense point cloud;Based on shield tunnel dense point cloud obtains Gaussian model, iteratively optimizes Gaussian model by rendering image and real image loss function;Leakage area is detected by semantic segmentation network, and based on leakage area and iteratively optimized Gaussian model obtains the shield tunnel three-dimensional grid model containing leakage label;Based on shield tunnel three-dimensional grid model realizes the positioning and parameter quantization of leakage disease.The application significantly improves detection efficiency and precision, reduces artificial cost, and promotes the intelligent development of tunnel detection.

Description

Technical Field

[0001] This invention belongs to the field of digital detection technology for tunnel engineering, and particularly relates to an automated inspection system and method for leakage in shield tunnels based on adaptive training and optimization. Background Technology

[0002] With the continuous advancement of urbanization, the development and utilization of underground space has gradually become an important means of alleviating the shortage of surface space. Especially in the construction of critical underground infrastructure such as subway rail transit, the shield tunneling method has become a widely used construction method due to its advantages such as high construction safety, fast construction speed, and minimal impact on surface traffic. However, shield tunnels contain numerous circumferential joints and bolt holes. When affected by adverse factors such as settlement disturbances and the degradation of segment material performance, these locations are prone to leakage problems, and leakage is often a key indicator of the structural safety of shield tunnels. Related surveys show that leakage-related disasters account for 70% of all tunnel-related disasters. Therefore, from the construction phase to the operation phase of shield tunnels, leakage inspection has always been considered a crucial task for ensuring tunnel safety.

[0003] Leakage inspection, a complex system task involving identification, location, and measurement, has long been hampered by high labor costs and extremely low efficiency due to its reliance on manual inspection. In recent years, driven by artificial intelligence, neural networks have enabled the automatic detection and identification of leak areas in images. However, detection results based solely on two-dimensional digital images are insufficient to support decision-making in actual engineering projects. The spatial location and area of ​​leaks are often the most valuable information for inspectors, making the acquisition of three-dimensional structural information of shield tunnels equally crucial.

[0004] Currently, among the main methods for acquiring 3D tunnel structural information, relying on specialized equipment such as 3D laser scanners or LiDAR places a heavy burden on leakage inspection due to the high cost of the equipment, and the cumbersome data sources also pose a challenge to system integration. In contrast, photogrammetry based on image 3D reconstruction offers a low-cost solution and demonstrates certain accuracy advantages.

[0005] Undeniably, image-based 3D reconstruction helps establish a bridge between visual semantics and 3D structural information, enabling multi-dimensional detection and analysis of leak inspection tasks using only visual sensing equipment. Furthermore, semantic data, represented by visual images, has become an indispensable data type in many tasks of tunnel engineering. It not only has the advantage of being easily captured by numerous visual sensing portals, but also, driven by the development of computer vision, will support more possible intelligent analysis tasks. Therefore, under the trend of digitalization and intelligentization in tunnel engineering, the development of a digital leak inspection process based on image-based 3D reconstruction is an important research direction.

[0006] However, despite the vast potential of image-based 3D reconstruction, it still faces significant challenges in the inspection of leaks in tunnel boring machines (TBMs). Mainstream baseline workflows involve multiple stages, including SfM and MVS. For TBM scenarios requiring large-scale inspections, these methods impose significant limitations on computational efficiency, hindering timely feedback of leak conditions to inspectors. Furthermore, the 3D point cloud data generated by these methods is prone to information loss in sparse texture areas, impeding effective visual analysis of the surrounding leak area.

[0007] Some researchers have attempted to address these limitations by introducing process improvements. For example, some scholars have enhanced the dark details of images through histogram equalization and gamma transform, improving the feature matching ability of the SfM algorithm in dark regions and alleviating the problem of lost 3D structural information in cable tunnels. However, limitations in computational efficiency still exist. Other scholars have used the RANSAC algorithm to extract the tunnel's geometric contours and generate dense point clouds, achieving high-efficiency 3D reconstruction through the mapping of point clouds to pixels. However, this is limited to cylindrical shield tunnels, resulting in a loss of application robustness.

[0008] A fundamental challenge is the inherent discretization limitation of point clouds. Existing methods lack the ability to rapidly optimize point clouds to automatically converge to 3D reconstruction of the true tunnel structure. Therefore, these methods can only offer performance trade-offs and cannot provide a comprehensive and competitive approach for acquiring 3D structural information of shield tunnels to support systematic shield tunnel leakage inspection tasks. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention proposes an automated inspection system and method for leakage in shield tunnels based on adaptive training optimization, thereby resolving the issues present in the prior art.

[0010] To achieve the above objectives, this invention provides an automated inspection method for leakage in shield tunnels based on adaptive training and optimization, comprising:

[0011] Image data of the shield tunnel is acquired and preprocessed, and sparse 3D point cloud is reconstructed based on the preprocessed image data.

[0012] The sparse 3D point cloud is projected onto a vertical plane and fitted to obtain the tunnel centerline parameters; the tunnel centerline is then segmented and the cross-sectional profile is fitted to obtain the shield tunnel dense point cloud.

[0013] A Gaussian model was obtained based on the dense point cloud of the shield tunnel, and the Gaussian model was iteratively optimized by comparing the loss functions of the rendered image and the real image.

[0014] Leakage areas are detected by semantic segmentation network, and a three-dimensional mesh model of the shield tunnel containing leakage annotations is obtained based on the leakage areas and the iteratively optimized Gaussian model.

[0015] Segmented inspections are conducted based on a three-dimensional mesh model of the shield tunnel to locate and quantify leakage defects.

[0016] Optionally, the process of acquiring and preprocessing image data of the shield tunnel includes:

[0017] The drone flies at a constant speed along the central axis of the tunnel to collect video data of the shield tunnel from a shooting perspective facing the depth of the tunnel. The video data of the shield tunnel is converted into image data by frame extraction method, and the image data is enhanced by histogram equalization.

[0018] Optionally, the process of reconstructing a sparse 3D point cloud based on preprocessed image data includes:

[0019] The SfM algorithm is used to process the acquired image data, reconstruct sparse 3D point clouds, and estimate the camera pose corresponding to each image. Based on the ratio of the average spatial size of multiple segment ring widths in the sparse 3D point cloud to the designed ring width, the conversion relationship of physical scale is calculated.

[0020] Optionally, the process of projecting the sparse three-dimensional point cloud onto a vertical plane and fitting it to obtain the tunnel centerline parameters includes:

[0021] The sparse 3D point cloud is projected onto two vertical planes to obtain two sets of planar point clouds. The two sets of planar point clouds are then divided along the Z-axis at fixed intervals. The average of the maximum and minimum values ​​of each slice in the X-axis and Y-axis directions is calculated. The tunnel center point cloud is constructed by combining the Z-axis coordinates. The tunnel center point cloud is converted into actual physical dimensions using the conversion relationship of physical scales. Curve fitting is performed using a cubic polynomial fitting method to obtain the tunnel centerline parameters.

[0022] Optionally, the process of obtaining a dense point cloud of a shield tunnel includes:

[0023] The tunnel centerline is discretized into several segments. Within each segment, the point cloud is projected onto a vertical plane, and the contour feature points are calculated in a polar coordinate system. A polar coordinate system is defined with the center of the plane as the origin, and the median distance of the point set corresponding to each angle is calculated at a preset angle interval, which is used as the tunnel cross-section contour point. The tunnel cross-section contour points are extended along the negative space vector direction to obtain the tunnel cross-section contour point cloud. Based on the tunnel cross-section contour point cloud, a dense point cloud of the shield tunnel is automatically generated.

[0024] Optionally, the process of obtaining a Gaussian model based on the dense point cloud of the shield tunnel and iteratively optimizing the Gaussian model by comparing the loss function of the rendered image and the real image includes:

[0025] The point primitives in the dense point cloud of the shield tunnel are parameterized into Gaussian primitives with position, normal vector, tangent vector, and scale factor. The color parameters of the Gaussian units are represented using spherical harmonic functions to obtain a Gaussian model. Based on the camera pose corresponding to each image, the Gaussian model is projected onto the actual image shooting position through rasterization, and the rendered image corresponding to the camera position is output. The photometric loss and normal consistency loss between the rendered image and the real image are calculated. Based on the loss function and adaptive training strategy, the geometric parameters of the Gaussian primitives are iteratively adjusted to achieve iterative optimization of the Gaussian model.

[0026] Optionally, the process of detecting leakage areas using a semantic segmentation network and obtaining a 3D mesh model of the shield tunnel containing leakage annotations based on the leakage areas and an iteratively optimized Gaussian model includes:

[0027] A semantic segmentation network is used to identify leakage areas in real-time acquired images and generate a binary mask of the leakage area with preset color coding. The optimized 3D Gaussian model is rendered to generate a depth map. The binary mask of the leakage area and the depth map are fused into multimodal data to obtain RGB-D data. Based on the TSDF voxel fusion algorithm and the RGB-D data, a 3D mesh model of the shield tunnel containing leakage annotations is obtained.

[0028] Optionally, the process of segmented inspection based on a 3D mesh model of the shield tunnel to locate and quantify leakage defects includes:

[0029] Based on the design ring width along the tunnel's central axis, the three-dimensional mesh model of the shield tunnel is segmented. Leakage sections are screened by identifying preset color codes in the mesh vertices. The axial distance from the center point of the leakage section to the starting point is calculated to determine the tunnel station number of the leakage section. The mesh surface of the leakage area is extracted using a color threshold recognition method. The mesh of the leakage area is determined by setting the RGB channel threshold and the area is calculated.

[0030] This invention provides an automated inspection system for leakage in shield tunnels based on adaptive training and optimization, comprising:

[0031] The data acquisition module is used to acquire video data of the shield tunnel using a drone as an acquisition platform;

[0032] The data preprocessing module is used to convert video data into image data and perform enhancement processing;

[0033] A sparse point cloud construction module is used to reconstruct sparse 3D point clouds from enhanced image data using the SfM algorithm;

[0034] The dense point cloud generation module is used to automatically generate dense point clouds of shield tunnels through centerline extraction and cross-section fitting.

[0035] The Gaussian model optimization module is used to obtain a Gaussian model based on the dense point cloud of the shield tunnel, and to iteratively optimize the Gaussian model by comparing the loss function of the rendered image and the real image.

[0036] The leakage detection module is used to detect leakage areas through a semantic segmentation network;

[0037] The mesh model building module is used to obtain a 3D mesh model of the shield tunnel containing leakage annotations based on the leakage area and the iteratively optimized Gaussian model.

[0038] The spatial positioning and quantification module is used for segmented inspection based on the three-dimensional mesh model of the shield tunnel to locate and quantify leakage defects.

[0039] Compared with the prior art, the present invention has the following advantages and technical effects:

[0040] Based on the sparse point cloud obtained by mainstream image 3D reconstruction methods, this invention realizes the automatic generation of dense point cloud in tunnels through the extraction of the central axis and the fitting of discretized cross sections, which significantly improves the efficiency of densification of 3D structural information and also has a certain degree of adaptability in shield tunnels with curved sections and irregular cross-sectional contours.

[0041] This invention utilizes Gaussian sputtering technology to parametrically represent point cloud data, enabling adaptive training and optimization of the 3D structural information of shield tunnels. Compared with the current mainstream SfM+MVS baseline process, the geometric error of the constructed model is only 1%, the modeling time cost is reduced by nearly 30%, and the problem of missing 3D structural information is significantly improved.

[0042] This invention is based on the rendering depth information of the Gaussian model. It uses TSDF to realize the correspondence between image semantic information and three-dimensional structural information. Combined with the Mask R-CNN leakage semantic segmentation network, it expands the application dimension of the two-dimensional semantic segmentation network. Through automated segmented inspection and color threshold recognition, it realizes leakage spatial positioning and parameter quantization in digital three-dimensional space. In actual shield tunnel engineering applications, the average leakage quantization accuracy rate reaches 90.3%.

[0043] This invention uses drones as an image data acquisition platform, which reduces equipment costs and improves the efficiency and flexibility of data acquisition, making it suitable for environments with limited space inside shield tunnels.

[0044] The method of this invention significantly reduces the tunnel dwell time requirement for inspection personnel in shield tunnel leakage inspection tasks, requiring only drone flight photography at the inspection starting point, thereby simplifying the leakage inspection operation process and laying the foundation for future autonomous cruise photography.

[0045] This invention provides a systematic process for inspecting leakage in shield tunnels. Under the trend of digitalization and intelligentization, it promotes the transformation of leakage inspection methods towards automation and intelligence, and provides a basic platform for multi-task integration in tunnel engineering. Attached Figure Description

[0046] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0047] Figure 1 This is an overall flowchart of the method according to an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of shield tunnel image data acquisition according to an embodiment of the present invention;

[0049] Figure 3 This is a diagram showing the initial sparse point cloud construction result of an embodiment of the present invention;

[0050] Figure 4 These are schematic diagrams illustrating the acquisition of tunnel center point cloud and fitting of center axis in an embodiment of the present invention. (a) is a schematic diagram of the acquisition of tunnel center point cloud, and (b) is a schematic diagram of fitting tunnel center axis.

[0051] Figure 5 This is a schematic diagram illustrating the automatic generation of dense point clouds in tunnels according to an embodiment of the present invention.

[0052] Figure 6 This is a schematic diagram of point primitives being Gaussian parameterized according to an embodiment of the present invention;

[0053] Figure 7 This is a schematic diagram of the construction of a three-dimensional mesh model for a shield tunnel according to an embodiment of the present invention;

[0054] Figure 8 This is a schematic diagram illustrating the spatial location and parameter quantification of leakage in a shield tunnel according to an embodiment of the present invention.

[0055] Figure 9 This is a site map of an experimental shield tunnel project according to an embodiment of the present invention;

[0056] Figure 10 The following are the geometric accuracy evaluation results of the self-generated tunnel dense point cloud in an embodiment of the present invention: (a) is a distance heatmap, and (b) is a point distance distribution curve.

[0057] Figure 11 This is a graph showing the change in the total loss of the Gaussian model for a shield tunnel during the training process of an embodiment of the present invention.

[0058] Figure 12 A comparison image of the depth map rendered by the Gaussian model trained in different rounds of this invention and a reference depth map;

[0059] Figure 13 This is a diagram showing the geometric accuracy evaluation results of the shield tunnel mesh model according to an embodiment of the present invention;

[0060] Figure 14 This is a visual comparison of the shield tunnel mesh model of the method of this invention and the mainstream method;

[0061] Figure 15 This is a three-dimensional visualization of the leakage area based on a shield tunnel mesh model, according to an embodiment of the present invention.

[0062] Figure 16 This is a comparison chart of leakage detection results in an embodiment of the present invention. Detailed Implementation

[0063] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0064] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0065] Example 1

[0066] like Figure 1 As shown, this embodiment provides an automated inspection method for leakage in shield tunnels based on adaptive training and optimization, including:

[0067] I. Image Data Acquisition and Preprocessing;

[0068] Using drones as a data acquisition platform, the shield tunnel was filmed from a shooting perspective facing the tunnel's depth. The flight path was controlled near the tunnel's central axis, maintaining a constant speed and avoiding rapid maneuvers.

[0069] The video data is converted into image data by frame extraction and then enhanced using histogram equalization technology.

[0070] II. Acquisition of three-dimensional structural information of shield tunnels;

[0071] a) Initialize sparse point cloud construction:

[0072] The SfM algorithm is used to reconstruct sparse 3D point clouds of scene structure from image data and estimate camera pose.

[0073] The conversion relationship of physical scale is calculated based on the ratio of the average spatial dimension of multiple segment ring widths in the point cloud to the designed ring width.

[0074] b) Automatic generation of dense point clouds in shield tunnels:

[0075] In the global coordinate system, the sparse point cloud is projected onto two vertical planes, XOZ and YOZ, respectively, to obtain two sets of planar point clouds;

[0076] Divide the data along the Z-axis at intervals of Step, and extract the maximum and minimum values ​​of the point set at each Step along the other coordinate axis and calculate the average value.

[0077] By combining the average value of the two sets of planar point clouds with the Z-axis coordinates, the center point cloud of the tunnel is constructed.

[0078] The tunnel center point cloud was curve-fitted using a cubic polynomial fitting method to obtain the tunnel centerline parameters.

[0079] The tunnel's central axis is discretized into multiple straight line segments, and cross-sectional profiles are fitted for each segment.

[0080] The sparse point cloud within the straight line segment is projected onto the vertical plane at the end of the spatial vector. A polar coordinate system is defined with the center of the plane as the origin, and the median distance of the point set corresponding to each angle is calculated at the interval of angle θ, which is used as the tunnel section contour point.

[0081] The contour points are extended along the negative space vector direction to form the tunnel cross-section contour point cloud of the current straight line segment, and the dense point cloud of the shield tunnel is automatically generated.

[0082] c) Parameterization and training optimization based on Gaussian sputtering technology:

[0083] The point primitives in the dense point cloud are replaced with Gaussian splats with elliptical geometric distributions.

[0084] Each Gaussian element is parameterized, including its planar spatial location, normal vector, tangent vector, and corresponding scale factor.

[0085] Define color parameters for each Gaussian element and use spherical harmonic functions to represent the correspondence of color values ​​under different spatial perspectives;

[0086] The Gaussian model is projected onto the actual image capture position, and the rendered image corresponding to the camera position is output through rasterization processing.

[0087] Calculate the photometric loss and normal consistency loss between the rendered image and the real image, and use this to guide the automatic adjustment of the geometric parameters of splats in the Gaussian model;

[0088] Through iterative training, the geometric structure of the Gaussian model is optimized to make it more closely resemble the actual scene structure reflected in the image data.

[0089] III. Spatial location and parameter quantification of leakage problems in shield tunnels;

[0090] a) Correspondence between image semantic information and 3D structural information:

[0091] The Mask R-CNN semantic segmentation neural network was used to detect the leakage areas in the actual captured images, and the leakage pixel areas were extracted and assigned a uniform blue mask.

[0092] The depth map rendered by the Gaussian model of the shield tunnel is combined with the image after leakage segmentation to form RGB-D data;

[0093] By using TSDF multi-frame depth map fusion technology, the symbolic distance field in voxel space is updated to construct a three-dimensional mesh model of the shield tunnel containing leakage annotations.

[0094] b) Spatial location and parameter quantification of leakage defects:

[0095] Using the designed ring width of the tunnel segments as the step size, the shield tunnel mesh model is automatically segmented and detected along the tunnel centerline.

[0096] Extract the mesh model segment containing the blue vertex color parameter as the leakage segment mesh model;

[0097] Calculate the axial spatial distance from the leakage section to the starting position of the inspection, and determine the tunnel station number where the leakage section is located;

[0098] The color threshold recognition method is used to extract the mesh surface of the leakage area. By setting the RGB channel threshold, the mesh of the leakage area is determined and the area is calculated, so as to realize the parameter quantification of leakage disease.

[0099] Example 1:

[0100] Step 1: Image data acquisition and preprocessing;

[0101] like Figure 2 As shown, a DJI Mini 4 Pro drone was used as the data acquisition platform. Its camera had a field of view (FOV) of 82.1°, an equivalent focal length of 24mm, and an aperture of f / 1.7. The flight path followed the tunnel's central axis, with the flight direction facing the tunnel's depth. The acquired video was 2 minutes long and had a resolution of 1920×1080px. Frame extraction converted the video into a sequence of 426 images, which, after filtering out motion-blurred images, resulted in 340 images. Before data acquisition, the camera was calibrated, and distortion correction was performed on the image data based on the obtained distortion coefficients. The final image resolution used in the experiment was 1919×1079px.

[0102] Step 2: Obtaining the three-dimensional structural information of the shield tunnel;

[0103] (1) Initialize the sparse point cloud construction;

[0104] like Figure 3 As shown, the SfM algorithm is used to obtain the initial sparse point cloud of the shield tunnel. This initialization step is implemented using the open-source tool COLMAP, obtaining a sparse point cloud containing 114,267 points. The sparse 3D point cloud accurately reflects the overall structural outline of the shield tunnel and provides basic 3D structural information.

[0105] To achieve the conversion to physical scale, the average spatial dimension of multiple segment ring widths in the point cloud is calculated, and the physical scale is converted based on the designed ring width:

[0106] Actual Scale=(Ring Width(design) / Ring Width(point cloud))×SpatialScale;

[0107] Wherein, Actual Scale is the actual scale, Ring Width (design) is the actual design ring width, Ring Width (point cloud) is the spatial ring width of the point cloud, and Spatial Scale is the spatial scale.

[0108] (2) Automatic generation of dense point clouds in shield tunnels;

[0109] The tunnel orientation is defined along the Z-axis in the global coordinate system. The sparse point cloud is projected onto two vertical planes, XOZ and YOZ, to obtain two sets of planar point clouds. For each set of planar point clouds, the Z-axis is divided at intervals of Step. The maximum and minimum values ​​of the point set at each Step along the other coordinate axis are extracted and their averages are calculated. These averages of the two sets of planar point clouds are combined with the Z-axis coordinates of the Step division to form the tunnel center point cloud in three-dimensional space. Step determines the sampling interval, which is set to 1 / 5 of the Ring Width (point cloud) by default and can be adjusted according to actual conditions.

[0110] A cubic polynomial fitting method was used to fit the tunnel center point cloud to obtain the tunnel centerline. For the fitting process, based on the magnitude of each point Pi = (xi, yi, zi) in the Z-axis direction of the tunnel center point cloud, a parameter sequence ti ∈ (0, 1) was constructed and associated with each point Pi in the tunnel center point cloud to drive the independent variable of the fitting polynomial. The tunnel centerline polynomial C(t) is expressed as:

[0111]

[0112] The objective function is:

[0113]

[0114] The parameters of the tunnel centerline are obtained by solving the polynomial coefficients using the least squares criterion, such as... Figure 4 As shown.

[0115] The tunnel centerline is discretized into multiple straight line segments, and cross-sectional profile fitting is performed separately for each segment. The tunnel centerline is sampled at fixed intervals to form a point cloud, defining a spatial vector. The initial sparse point cloud within the straight line segment interval is projected onto the vertical plane at the end of the spatial vector. A polar coordinate system is defined with the center of the plane as the origin. The median distance of the point set corresponding to each angle is calculated at intervals of θ = 0.5 degrees (in the planar polar coordinate system, starting with the horizontal direction, the calculation proceeds counterclockwise, with the first angle being 0.5 degrees, the second 1 degree, and so on, calculating the median distance of the point set corresponding to each angle), serving as the tunnel cross-sectional profile points. These profile points are expanded along the negative spatial vector direction to form the tunnel cross-sectional profile point cloud for the current straight line segment. The number of expansion groups along the spatial vector for each cross-section is set to 10. A schematic diagram of the automatically generated dense tunnel point cloud is shown below. Figure 5 As shown.

[0116] (3) Parameterization and training optimization based on Gaussian sputtering technology;

[0117] like Figure 6As shown, Gaussian primitives with elliptical geometric distributions are used to replace point primitives in dense point clouds. In three-dimensional space, the three-dimensional coordinates of a point are taken as its spatial position on a plane, and the plane direction is determined by the normal vector t. w =t u ×t v It is determined that for a Gaussian distribution in a two-dimensional uv plane, its definition in three-dimensional space is expressed as:

[0118]

[0119] Where Pc represents the spatial coordinates of the point cloud, u and v represent the spatial coordinates of the plane containing the Gaussian distribution, and t u and t v Let S represent two mutually perpendicular direction vectors in the plane. u and S v H represents the corresponding scale, and H represents the transformation matrix.

[0120]

[0121] Where R represents the rotation matrix and S represents the scale matrix, the color parameters are defined using spherical harmonic functions to represent the correspondence between splat and color values ​​f(t) at different spatial viewpoints (θ, φ):

[0122]

[0123] Where l and m represent degree and order respectively, K is the normalization factor, P is the Associated Legendre Polynomials, and e imφ This represents a complex exponential function, which provides the periodicity of the azimuth angle φ, and c is the weighting coefficient.

[0124] The Gaussian model was trained 50,000 times. The optimizer used was Adam, with the coefficients for calculating the gradient and the square of the gradient set to 0.9 and 0.999, respectively. The color learning rate for splats was set to 0.0025, and the learning rates for scaling and rotation were set to 0.005 and 0.001, respectively.

[0125] Training loss includes photometric loss and normal consistency loss:

[0126] L c =(1-λ)L1+λL D-SSIM ;

[0127]

[0128] Where n represents the normal direction in the camera coordinate system, and N represents the actual surface normal (calculated based on the gradient of depth in the x and y directions).

[0129] Step 3: Spatial location and parameter quantification of leakage problems in shield tunnels;

[0130] (1) Image semantic information corresponds to three-dimensional structural information;

[0131] Leakage areas in actual captured images were detected using a Mask R-CNN semantic segmentation neural network. Leakage pixels were extracted and assigned a uniform blue mask. Rendered depth maps from a Gaussian model of the tunnel boring machine were combined one-to-one with these maps to form RGB-D data, and multi-frame depth map fusion was performed using TSDF technology. Figure 7 As shown.

[0132] During the fusion process, the computer's video memory space is discretized into multiple unit cubes (voxels). Positive, negative, and zero values ​​of the TSDF (Transformed Signed Distance Parameter) are used to represent the voxels' location outside, inside, and on the surface of the model, respectively. The first camera position is used as the global reference coordinate system, and subsequent RGB-D images are sequentially transformed to this global coordinate system. The TSDF value is then updated using image pixel values ​​and depth values. The TSDF value update formula is:

[0133]

[0134] Among them, TSDF(p) old and W(p) old This represents the TSDF value and weight value of the current voxel p, TSDF(p). new and W(p) new This represents the TSDF value and weight value of voxel p in the newly added RGB-D.

[0135] (2) Spatial location and parameter quantification of leakage defects;

[0136] like Figure 8 As shown, based on the actual shield tunnel conditions, a reasonable step size (5m in this case, but the step size can be adjusted according to the actual situation) is used to automatically segment and detect the shield tunnel mesh model along the tunnel's central axis. The mesh model segment containing the blue vertex color parameter is extracted separately as the leakage section mesh model, and the axial spatial distance from this position to the inspection start position is used to calculate the tunnel station number of the leakage section.

[0137] For the mesh model of the leakage section, the color thresholding method is used to extract the mesh surface of the leakage area. By setting the RGB channel threshold, the mesh of the leakage area is determined within the blue channel range and the area is calculated, thus realizing the parameter quantification of the leakage problem.

[0138] Example 2:

[0139] Using a shield tunnel in Shaoxing City, Zhejiang Province, China as the engineering background, a 140m long curved section from G2K1+280 to G2K1+420 of the main shield tunnel section was selected as the experimental site. Figure 9 As shown in the image, this section of the shield tunnel has an outer diameter of 8.8m, an inner diameter of 8.0m, and a ring width of 1.6m. Located below the bottom of the riverbed, it falls within a key area of ​​concern for leakage issues.

[0140] like Figure 15 , 16 As shown, the method of this invention was used for leakage detection, and three leakage areas were finally detected in this section of the shield tunnel, located at chainages G2K1+366, G2K1+377, and G2K1+402, respectively. The calculated areas of the three leakage areas are 4.25 m². 2 1.50m 2 and 3.19m 2 The actual leakage area, measured visually using professional equipment, was 4.94 m². 2 1.65m 2 and 3.29m 2 Compared to the previous method, the proposed method achieved area quantification accuracy of 86%, 91%, and 94% in the three leakage areas, with an average accuracy of 90.3%.

[0141] This example verifies the effectiveness and accuracy of the method of the present invention in practical engineering applications, and can meet the engineering requirements for detecting leakage defects in shield tunnels. The performance advantages of this method, especially in terms of computational efficiency and the integrity of geometric structural information, are of great significance to practical shield tunnel inspection procedures. The proposed method not only significantly reduces computational time costs (modeling time is reduced by approximately 30% compared to mainstream methods), but also provides more continuous and smooth three-dimensional structural information of shield tunnels. Figure 14 (For comparison of the modeling effects of this method with mainstream methods), it effectively makes up for the limitations of current mainstream methods in the task of inspecting leakage in shield tunnels. Figure 10 and Figure 13 The accuracy of the 3D structural information before and after Gaussian parameterization training is shown respectively. Figure 11 , 12 It shows the overall structural changes and depth information changes during the Gaussian training process.

[0142] This invention provides an automated inspection system for leakage in shield tunnels based on adaptive training and optimization, comprising:

[0143] The data acquisition module is used to acquire video data of the shield tunnel using a drone as an acquisition platform;

[0144] The data preprocessing module is used to convert video data into image data and perform enhancement processing;

[0145] A sparse point cloud construction module is used to reconstruct sparse 3D point clouds from enhanced image data using the SfM algorithm;

[0146] The dense point cloud generation module is used to automatically generate dense point clouds of shield tunnels through centerline extraction and cross-section fitting.

[0147] The Gaussian model optimization module is used to obtain a Gaussian model based on the dense point cloud of the shield tunnel, and to iteratively optimize the Gaussian model by comparing the loss function of the rendered image and the real image.

[0148] The leakage detection module is used to detect leakage areas through a semantic segmentation network;

[0149] The mesh model building module is used to obtain a 3D mesh model of the shield tunnel containing leakage annotations based on the leakage area and the iteratively optimized Gaussian model.

[0150] The spatial positioning and quantification module is used for segmented inspection based on the three-dimensional mesh model of the shield tunnel to locate and quantify leakage defects.

[0151] The dense point cloud generation module includes:

[0152] Point cloud projection unit is used to project sparse point clouds onto a vertical plane;

[0153] Central point cloud building unit, used to construct the tunnel central point cloud by calculating the average value through segmentation;

[0154] The centerline fitting unit is used to obtain the tunnel centerline parameters through cubic polynomial fitting.

[0155] The cross-section fitting unit is used to obtain the tunnel cross-section contour points by calculating the median distance in polar coordinates.

[0156] Point cloud extension units are used to extend contour points along spatial vector directions to generate tunnel-dense point clouds.

[0157] The Gaussian model optimization module includes:

[0158] The parameterization unit is used to convert point primitives into Gaussian primitives and define the relevant parameters;

[0159] The rendering unit is used to project the Gaussian model onto the camera position and generate a rendered image.

[0160] The loss calculation unit is used to calculate various losses between the rendered image and the real image;

[0161] The parameter optimization unit is used to automatically adjust the Gaussian parameters based on the loss.

[0162] The spatial positioning and quantization module includes:

[0163] The segmented detection unit is used to automatically segment and detect the mesh model along the tunnel's central axis.

[0164] The leakage extraction unit is used to extract the mesh surface of the leakage area using a color threshold recognition method.

[0165] The stationing calculation unit is used to calculate the stationing location of the seepage section in the tunnel.

[0166] Area quantization unit, used to calculate the area of ​​the grid surface of the leakage area.

[0167] Based on the sparse point cloud obtained by mainstream image 3D reconstruction methods, this invention realizes the automatic generation of dense point cloud in tunnels through the extraction of the central axis and the fitting of discretized cross sections, which significantly improves the efficiency of densification of 3D structural information and also has a certain degree of adaptability in shield tunnels with curved sections and irregular cross-sectional contours.

[0168] This invention utilizes Gaussian sputtering technology to parametrically represent point cloud data, enabling adaptive training and optimization of the 3D structural information of shield tunnels. Compared with the current mainstream SfM+MVS baseline process, the geometric error of the constructed model is only 1%, the modeling time cost is reduced by nearly 30%, and the problem of missing 3D structural information is significantly improved.

[0169] This invention is based on the rendering depth information of the Gaussian model. It uses TSDF to realize the correspondence between image semantic information and three-dimensional structural information. Combined with the Mask R-CNN leakage semantic segmentation network, it expands the application dimension of the two-dimensional semantic segmentation network. Through automated segmented inspection and color threshold recognition, it realizes leakage spatial positioning and parameter quantization in digital three-dimensional space. In actual shield tunnel engineering applications, the average leakage quantization accuracy rate reaches 90.3%.

[0170] This invention uses drones as an image data acquisition platform, which reduces equipment costs and improves the efficiency and flexibility of data acquisition, making it suitable for environments with limited space inside shield tunnels.

[0171] The method of this invention significantly reduces the tunnel dwell time requirement for inspection personnel in shield tunnel leakage inspection tasks, requiring only drone flight photography at the inspection starting point, thereby simplifying the leakage inspection operation process and laying the foundation for future autonomous cruise photography.

[0172] This invention provides a systematic process for inspecting leakage in shield tunnels. Under the trend of digitalization and intelligentization, it promotes the transformation of leakage inspection methods towards automation and intelligence, and provides a basic platform for multi-task integration in tunnel engineering.

[0173] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An automated inspection method for leakage in shield tunnels based on adaptive training and optimization, characterized in that, Includes the following steps: Image data of the shield tunnel is acquired and preprocessed, and sparse 3D point cloud is reconstructed based on the preprocessed image data. The sparse 3D point cloud is projected onto a vertical plane and fitted to obtain the tunnel centerline parameters; the tunnel centerline is then segmented and the cross-sectional profile is fitted to obtain the shield tunnel dense point cloud. A Gaussian model was obtained based on the dense point cloud of the shield tunnel, and the Gaussian model was iteratively optimized by comparing the loss functions of the rendered image and the real image. Leakage areas are detected by semantic segmentation network, and a three-dimensional mesh model of the shield tunnel containing leakage annotations is obtained based on the leakage areas and the iteratively optimized Gaussian model. Segmented inspections are conducted based on a three-dimensional mesh model of the shield tunnel to locate and quantify leakage defects.

2. The automated inspection method for leakage in shield tunnels based on adaptive training and optimization according to claim 1, characterized in that, The process of acquiring and preprocessing image data of a shield tunnel includes: The drone flies at a constant speed along the central axis of the tunnel to collect video data of the shield tunnel from a shooting perspective facing the depth of the tunnel. The video data of the shield tunnel is converted into image data by frame extraction method, and the image data is enhanced by histogram equalization.

3. The automated inspection method for leakage in shield tunnels based on adaptive training and optimization according to claim 1, characterized in that, The process of reconstructing sparse 3D point clouds based on preprocessed image data includes: The SfM algorithm is used to process the acquired image data, reconstruct sparse 3D point clouds, and estimate the camera pose corresponding to each image. Based on the ratio of the average spatial size of multiple segment ring widths in the sparse 3D point cloud to the designed ring width, the conversion relationship of physical scale is calculated.

4. The automated inspection method for leakage in shield tunnels based on adaptive training and optimization according to claim 3, characterized in that, The process of projecting the sparse three-dimensional point cloud onto a vertical plane and fitting it to obtain the tunnel centerline parameters includes: The sparse 3D point cloud is projected onto two vertical planes to obtain two sets of planar point clouds. The two sets of planar point clouds are then divided along the Z-axis at fixed intervals. The average of the maximum and minimum values ​​of each slice in the X-axis and Y-axis directions is calculated. The tunnel center point cloud is constructed by combining the Z-axis coordinates. The tunnel center point cloud is converted into actual physical dimensions using the conversion relationship of physical scales. Curve fitting is performed using a cubic polynomial fitting method to obtain the tunnel centerline parameters.

5. The automated inspection method for leakage in shield tunnels based on adaptive training and optimization according to claim 4, characterized in that, The process of obtaining a dense point cloud of a shield tunnel includes: The tunnel centerline is discretized into several segments. Within each segment, the point cloud is projected onto a vertical plane, and the contour feature points are calculated in a polar coordinate system. A polar coordinate system is defined with the center of the plane as the origin, and the median distance of the point set corresponding to each angle is calculated at a preset angle interval, which is used as the tunnel cross-section contour point. The tunnel cross-section contour points are extended along the negative space vector direction to obtain the tunnel cross-section contour point cloud. Based on the tunnel cross-section contour point cloud, a dense point cloud of the shield tunnel is automatically generated.

6. The automated inspection method for leakage in shield tunnels based on adaptive training optimization according to claim 3, characterized in that, The process of obtaining a Gaussian model based on dense point clouds of shield tunnels and iteratively optimizing the Gaussian model by comparing the loss functions of rendered images and real images includes: The point primitives in the dense point cloud of the shield tunnel are parameterized into Gaussian primitives with position, normal vector, tangent vector, and scale factor. The color parameters of the Gaussian units are represented using spherical harmonic functions to obtain a Gaussian model. Based on the camera pose corresponding to each image, the Gaussian model is projected onto the actual image shooting position through rasterization, and the rendered image corresponding to the camera position is output. The photometric loss and normal consistency loss between the rendered image and the real image are calculated. Based on the loss function and adaptive training strategy, the geometric parameters of the Gaussian primitives are iteratively adjusted to achieve iterative optimization of the Gaussian model.

7. The automated inspection method for leakage in shield tunnels based on adaptive training and optimization according to claim 1, characterized in that, The process of detecting leakage areas using a semantic segmentation network and obtaining a 3D mesh model of the shield tunnel containing leakage annotations based on the leakage areas and an iteratively optimized Gaussian model includes: A semantic segmentation network is used to identify leakage areas in real-time acquired images and generate a binary mask of the leakage area with preset color coding. The optimized 3D Gaussian model is rendered to generate a depth map. The binary mask of the leakage area and the depth map are fused into multimodal data to obtain RGB-D data. Based on the TSDF voxel fusion algorithm and the RGB-D data, a 3D mesh model of the shield tunnel containing leakage annotations is obtained.

8. The automated inspection method for leakage in shield tunnels based on adaptive training and optimization according to claim 1, characterized in that, The process of segmented inspection based on a 3D mesh model of a shield tunnel to locate and quantify leakage defects includes: Based on the design ring width along the tunnel's central axis, the three-dimensional mesh model of the shield tunnel is segmented. Leakage sections are screened by identifying preset color codes in the mesh vertices. The axial distance from the center point of the leakage section to the starting point is calculated to determine the tunnel station number of the leakage section. The mesh surface of the leakage area is extracted using a color threshold recognition method. The mesh of the leakage area is determined by setting the RGB channel threshold and the area is calculated.

9. An automated inspection system for leakage in shield tunnels based on adaptive training and optimization, characterized in that, include: The data acquisition module is used to acquire video data of the shield tunnel using a drone as an acquisition platform; The data preprocessing module is used to convert video data into image data and perform enhancement processing; A sparse point cloud construction module is used to reconstruct sparse 3D point clouds from enhanced image data using the SfM algorithm; The dense point cloud generation module is used to automatically generate dense point clouds of shield tunnels through centerline extraction and cross-section fitting. The Gaussian model optimization module is used to obtain a Gaussian model based on the dense point cloud of the shield tunnel, and to iteratively optimize the Gaussian model by comparing the loss function of the rendered image and the real image. The leakage detection module is used to detect leakage areas through a semantic segmentation network; The mesh model building module is used to obtain a 3D mesh model of the shield tunnel containing leakage annotations based on the leakage area and the iteratively optimized Gaussian model. The spatial positioning and quantification module is used for segmented inspection based on the three-dimensional mesh model of the shield tunnel to locate and quantify leakage defects.

Images