A shield tunnel hole crack identification method driven by a cross-modal feature deep fusion model

Abstract

The application provides a cross-modal feature deep fusion model driving tunnel bolt hole seam identification method, wherein the method comprises the following steps: first, a shield tunnel mobile LiDAR point cloud and dimensionality reduction projection cross-modal data feature fusion method is designed, and a conformal mapping model of three-dimensional shield tunnel point cloud and two-dimensional image dimensionality reduction data is established; second, a shield segment classification method driven by a double template considering local morphological features is designed, and the shield segment is accurately classified according to the bidirectional driving of the 'T' and 'O' templates. Finally, a shield segment target accurate extraction method fusing multi-modal feature offset correction is designed, and the shield segment joint information is accurately extracted after the spatial correction of the model positioning result.

Description

Technical Field

[0001] This invention relates to the field of point cloud target recognition technology, specifically to a method for identifying gaps in shield tunnels driven by a cross-modal feature deep fusion model. Background Technology

[0002] As an important emerging transportation method for social development, shield tunnels have become an indispensable part of the urbanization process, and their construction plays a vital supporting role in the sustainable economic development of cities. The existing morphological characteristics of shield tunnels mainly consist of bolt holes for fixing assembled blocks and joints between these blocks. Locating longitudinal joints can effectively reconstruct tunnel mileage and correct tunnel scanning deformation data, while locating transverse joints and bolt holes plays a crucial role in the segmentation and 3D reconstruction of individual shield tunnel segments. Due to the long service life of shield tunnels and the local deformation caused by geological compression, traditional measurement methods for detecting shield tunnel geometric parameters may be affected by interference from tunnel ancillary facilities, easily leading to missing information in the measurement targets. This has become a challenge in the operation and maintenance of shield tunnels. In recent years, Mobile Laser Scanning (MLS) technology, compared with traditional tunnel deformation monitoring methods such as total stations and convergence meters, can efficiently acquire high-precision, high-density 3D point clouds of tunnels, providing rich 3D spatial information, and has gradually become an important means of operation and maintenance inspection of rail transit tunnels. Therefore, exploring efficient methods for locating gaps in shield tunnels using MLS technology has significant scientific and practical value.

[0003] Methods for target identification from point clouds in tunnel boring machines (TBMs) both domestically and internationally can be broadly categorized into two types: ① Direct methods: These methods directly locate key target features of TBM segments in three-dimensional space by comprehensively analyzing the spatial relationships of point cloud features. This type of method effectively avoids data distortion during projection and fully utilizes the multi-dimensional feature information of the point cloud. However, the large volume of point cloud data limits computational efficiency, and it is difficult to overcome the "hole" problem in the point cloud. ② Indirect methods: These methods reduce the dimensionality of the three-dimensional point cloud and project it to generate a two-dimensional image, then comprehensively utilize deep learning and image processing algorithms to identify TBM segments. This type of method extracts targets from two-dimensional images. Dimensionality reduction improves computational efficiency, but it also loses multi-dimensional feature information and is susceptible to interference from auxiliary facilities. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a cross-modal feature deep fusion model-driven method for identifying the gaps in shield tunnels, so as to solve the problems in the prior art such as the inability to identify key gap targets in shield tunnels due to obstruction by nearby auxiliary facilities, and the omission and misposition of shield tunnel segment targets.

[0005] Therefore, the present invention adopts the following technical solution:

[0006] A method for identifying gaps in shield tunnels driven by a cross-modal feature deep fusion model includes the following steps:

[0007] S1, Feature fusion of cross-modal data from moving LiDAR point cloud and dimensionality-reduced projection in shield tunnel:

[0008] S11, the moving LiDAR point cloud of the tunnel is fitted using the RANSAC algorithm, and the bolt hole feature point cloud that meets the conditions is obtained according to the following formula.

[0009] |PF1|+|PF2|>2a

[0010] Where F1 and F2 represent the two foci of the fitted elliptical cross-section of the shield tunnel; a represents the major semi-axis of the elliptical cross-section; P represents the laser point of the bolt hole feature cloud; |PF1| represents the distance from the laser point of the feature cloud to the focal point F1; |PF2| represents the distance from the laser point of the feature cloud to the focal point F2.

[0011] S12, convert the 3D tunnel-moving LiDAR point cloud into a 2D projected tile map and a 2D coordinate set according to the following formula:

[0012]

[0013] In the formula, P represents the laser point on the c-th cross-section point cloud. c The corresponding X-axis coordinate in the grayscale image; P represents the laser point on the c-th cross-section point cloud. c The corresponding Y-axis coordinate in the grayscale image; P represents the laser point on the c-th cross-section point cloud. c The cumulative distance between the current survey line and the starting survey line; H represents the set horizontal resolution; P represents the laser point on the c-th cross-section point cloud. c The angle mapped onto the image; R represents the radius of the projected cylinder; V represents the set vertical resolution. express The pixel value at the location, where c = 1, 2, ..., k, k is the total number of point clouds in the cross section, and Intensity represents the image intensity value;

[0014] The two-dimensional coordinate set uses a cylindrical model composed of near-circular tunnel cross-sections as the projection surface and scanning survey lines as the units;

[0015] S2, after classifying the tunnel segments using a dual-template method, coarse positioning of bolt holes and splice joints is performed:

[0016] S21, the two-dimensional projection tiled map obtained in S1 is divided into single-ring shield two-dimensional projection tiled maps, and a coordinate system is established on it with the pixel coordinates of the upper left corner of each single-ring shield two-dimensional projection tiled map as the origin.

[0017] S22, determine the main template and secondary template of the shield tunnel segment, wherein the main template conforms to the bolt hole distribution characteristics of the capping block, and the secondary template conforms to the bolt hole distribution characteristics after splicing with adjacent blocks and standard blocks;

[0018] S23, using the main template and secondary template determined in S22, the bolt holes and splicing seams of the two-dimensional projection tiling are positioned to obtain a coarse positioning result, which is the set of bolt hole boxes R and the set of straight lines L of splicing seam positions;

[0019] S3, Correct the coarse positioning results:

[0020] S31, obtain the geometric center coordinate set M of the bolt hole frame according to the bolt hole frame set R obtained in S23, and obtain the slope dataset k and intercept dataset b of the splice seam according to the line set L obtained in S23.

[0021] S32, cluster the bolt hole feature point cloud obtained in S11 to obtain the clustered center points, and then use the method in S12 to two-dimensionalize the center points to obtain the bolt hole geometric center coordinate set N{N1(x1,y1),N2(x2,y2)……N i (x i ,y i )}, where i represents the number of the i-th bolt hole, and the set of geometric center coordinates N of the bolt holes is used for precise positioning of the bolt holes;

[0022] S33, using the geometric center coordinate set M of the bolt hole frame as a reference, find the corresponding center coordinates at the corresponding positions in the geometric center coordinate set N of the bolt holes. When the corresponding center coordinates can be found, it is as shown in the following formula:

[0023]

[0024] In the formula, Δd i express Point and The distance between them, i.e., the corrected distance.

[0025] When the corresponding center coordinates cannot be found, the following formula applies:

[0026]

[0027] In the formula, Δs i This represents the relative distance between the current bolt hole and the adjacent bolt hole; Δd i Indicates the correction distance; Δdi±1 Indicates the correction distance between adjacent bolt holes;

[0028] The bolt hole positions are obtained by correcting the bolt hole frame set R using the correction distance;

[0029] S34, find the line with the smallest total pixel sum in the two-dimensional projection tiling and use this line as the final position of the stitching seam. Establish a buffer zone within the range of the stitching seam with a radius of K pixels, record the dataset of the optimal buffer radius Δr, correct the intercept dataset b according to the following formula and save it as dataset b' to obtain the corrected stitching seam position.

[0030] b' j =b j +Δr j

[0031] In the formula, b j Δr represents the intercept of the j-th line in the intercept dataset. j b' represents the optimal buffer radius for the j-th line in the intercept dataset; j This represents the adjusted intercept of the j-th line in the intercept dataset;

[0032] The corrected splice seam position and the corrected bolt hole position are output as the precise positioning result.

[0033] Preferably, the main template number is "T" and the secondary template number is "O".

[0034] Preferably, when the template designed in S22 is used to match and locate key segments of the single-ring shield tunnel 2D projection tile image segmented in S21, if the bolt hole features of the single-ring shield tunnel 2D projection tile image match the features of the capping block, it is classified as a type A or type B single-ring shield tunnel. If the single-ring shield tunnel 2D projection tile image is classified as a type C single-ring shield tunnel because the image is blurred or occluded and can only be classified by feature classification through the splicing of the standard block and adjacent blocks.

[0035] Preferably, for the dual-template classification in S2, the normalized correlation coefficient localization method is used to evaluate the similarity between the template and the target image, and the similarity calculation is shown in the following formula:

[0036]

[0037] In the formula: T(x′,y′) represents the template image; I(x,y) represents the image to be located; w and h represent the width and height of the template image; R(x,y) represents the function describing the similarity.

[0038] Preferably, the method for dividing the two-dimensional projection tiled image of a single-ring shield in S2 is as follows: search for the line with the lowest longitudinal gray level accumulation in the longitudinal seam in the two-dimensional projection tiled image as the final longitudinal seam position, and use the final longitudinal seam to divide the shield ring to obtain the two-dimensional projection tiled image of the single-ring shield.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The method of the present invention utilizes the fusion of data features between the three-dimensional point cloud of the shield tunnel and the dimension-reduced projection two-dimensional image, which can effectively improve the accuracy of hole and gap identification.

[0041] 2. The method of the present invention is a shield tunnel segment classification method driven by dual templates that takes into account local morphological features, so as to achieve accurate classification of tunnel segments within the shield ring.

[0042] 3. The method of the present invention can accurately identify the gaps in the tunnel lining even when there is obstruction or interference from auxiliary facilities, and also takes into account the spatial relationship of the tunnel lining segments. It has a higher recognition rate and accuracy, and has better applicability and robustness. Attached Figure Description

[0043] Figure 1 This is a flowchart of the method of the present invention;

[0044] Figure 2 This is a flowchart of the cross-modal feature fusion process in this invention;

[0045] Figure 3 This is a diagram of the shield tunnel ring structure;

[0046] Figure 4 The distribution pattern of bolt holes in shield tunnels;

[0047] Figure 5 This is a schematic diagram of K-pixel buffer offset correction;

[0048] Figure 6 This is a schematic diagram illustrating the classification of shield tunnel segments using the dual-template driven method of the present invention.

[0049] Figure 7 This is a schematic diagram of the coarse positioning process in this invention;

[0050] Figure 8 This is a schematic diagram of the precise positioning process in this invention;

[0051] Figure 9 This is a comparison chart showing the identification results of longitudinal seams in tunnel segments using existing methods and the method of this invention.

[0052] Figure 10 This is a comparison diagram of the positioning results of bolt holes, straight and oblique transverse seams by the present invention and existing identification methods. Detailed Implementation

[0053] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0054] like Figure 1 As shown, this invention provides a method for identifying gaps in shield tunnels driven by a cross-modal feature deep fusion model, comprising the following steps:

[0055] First, a method for feature fusion of cross-modal data of moving LiDAR point cloud and dimensionality-reduced projection of shield tunnel is designed, and a mapping model between 3D shield tunnel point cloud and 2D image dimensionality-reduced data is established.

[0056] Secondly, a shield tunnel segment classification method driven by dual templates that takes into account local morphological characteristics is designed, and shield tunnel segments are accurately classified according to the bidirectional driving of "T" and "O" templates;

[0057] Finally, a method for accurately extracting shield tunnel segment targets by integrating multimodal feature offset correction is designed, and the joint information of shield tunnel segments is accurately extracted by spatial correction of the model positioning results.

[0058] S1, Feature fusion of cross-modal data from moving LiDAR point cloud and dimensionality-reduced projection in shield tunnel:

[0059] Extracting features from shield tunnels in a single dimension has limitations. For example, feature extraction from point clouds involves a large amount of data, resulting in low efficiency. Features extracted from two-dimensional images also suffer from insufficient utilization of three-dimensional information due to the loss of spatial information. The method of this invention fuses the moving LiDAR point cloud of the shield tunnel with the cross-modal data features of dimensionality reduction projection, which can effectively integrate spatial information from various dimensions and improve the accuracy of subsequent feature recognition.

[0060] The flowchart of cross-modal data feature fusion is as follows: Figure 2 As shown.

[0061] The cross-section of a shield tunnel can be considered as an ellipse close to a standard circle. The tunnel point cloud is fitted using the RANSAC algorithm, the initial parameters of the ellipse are calculated, and the bolt hole positions are deduced according to equation (1):

[0062] |PF1|+|PF2|>2a (1)

[0063] In the formula: F1 and F2 represent the two foci of the fitted ellipse; a represents the length of the semi-major axis of the ellipse; P represents the feature point cloud of the bolt hole; |PF1| represents the distance from the feature point to the focus F1; |PF2| represents the distance from the feature point to the focus F2.

[0064] Based on the characteristics of shield tunnels being linearly distributed and having a nearly circular cross-section, a cylindrical model is used as the projection surface. This cylindrical projection surface is then unfolded into a two-dimensional plane, and projection is performed line by line using scanning survey lines as units to achieve modal information conversion. The conversion mapping can be calculated using equation (2):

[0065]

[0066] In the formula, P represents the laser point on the c-th cross-section point cloud. c The X coordinate corresponding to the grayscale image; P represents the laser point on the c-th cross-section point cloud. c The corresponding Y-coordinate in the grayscale image; P represents the laser point on the c-th cross-section point cloud. c The cumulative distance between the current survey line and the starting survey line; H represents the set horizontal resolution; P represents the laser point on the c-th cross-section point cloud. c The angle to be mapped onto the image; R represents the radius of the projected cylinder; V represents the set vertical resolution. express The pixel value at the location, where c = 1, 2, ..., k, k is the total number of point clouds in the cross section, and Intensity represents the image intensity value.

[0067] S2, after classifying the tunnel segments using a dual-template method, coarse positioning of bolt holes and splice joints is performed:

[0068] like Figure 3 As shown, the shield tunnel structure consists of different types of segments, including one K-type segment (capping block), two B-type segments (adjacent blocks), and several A-type segments (standard blocks). Figure 4 As shown, there are generally two ways to connect bolt holes in shield tunnels: ① riveting continuous shield rings (horizontal bolt holes in the figure); ② riveting different types of tunnel segments inside the ring (vertical bolt holes in the figure).

[0069] In the two-dimensional image of a shield tunnel after projection, there are many interfering factors such as cables and pipelines, making it difficult to accurately classify the shield segments. To address this problem, this invention segments the two-dimensional image of the shield tunnel based on the location of the longitudinal joint. Regarding the question of how to accurately extract the longitudinal joint, Xu proposed using Canny edge detection and Hough transform to obtain the vertical line with the lowest gray-level accumulation in the upper column of the two-dimensional image of the shield tunnel as the central longitudinal joint, and then shifting the longitudinal joint to both sides according to the width of the segment to obtain the longitudinal joint. Figure 5 As shown in (a). However, considering the problem of inconsistent local deformation after projection, this invention improves upon Xu's method by employing a K-pixel buffer offset correction method to search for the straight line with the lowest longitudinal grayscale accumulation at each temporary longitudinal seam as the final corrected position, as shown in (a). Figure 5 As shown in (b) and (c), the single-ring shield tunnel is segmented based on the specific location of the longitudinal joint in the image, as follows. Figure 6 As shown in (a).

[0070] To achieve efficient and accurate classification of single-ring shield tunnels, a template positioning concept was introduced, and a dual-template driven shield segment classification method considering local morphological characteristics was designed. Typically, shield tunnels are classified according to... Figure 6 (c) For the A and B type single-ring shield tunnels, the present invention utilizes K-type segments to create the main template for these two types of single-ring shield tunnels. Since the template positioning method may lead to mispositioning or missed positioning when the target object is blurred or non-existent in the image during the image feature region traversal search, for the C type single-ring shield tunnels that are obscured by equipment, a secondary template is created using the bolt hole distribution characteristics after the adjacent blocks and standard blocks are spliced ​​together. Figure 6 As shown in (b).

[0071] This invention uses the normalized correlation coefficient localization method to evaluate the similarity between a template and a target image. The similarity is calculated as follows:

[0072]

[0073] In the formula: T(x′,y′) represents the template image; I(x,y) represents the image to be located; w and h represent the width and height of the template image; R(x,y) represents the function describing the similarity, i.e. the localization result.

[0074] like Figure 7 As shown, the present invention first deduces the characteristics of single-ring shield tunnel type, establishes a coordinate system with the pixel coordinates of the upper left corner of the two-dimensional image of the segmented shield ring as the origin, establishes a corresponding single-ring shield tunnel model according to the distribution characteristics of bolt holes and transverse seams on the image, and creates a set of boxes R to mark the approximate position of bolt holes and a set of lines L to mark the approximate position of transverse seams of the assembled blocks.

[0075] S3, Correct the coarse positioning results:

[0076] like Figure 8 As shown, in order to address the problem of inaccurate model positioning caused by the deformation of two-dimensional images of shield tunnels due to internal and external forces, and to further realize the precise positioning of bolt holes and transverse joints by making full use of three-dimensional point cloud and two-dimensional image information, this invention simplifies the bolt hole and transverse joint models of three single-ring shield types. The geometric center coordinate set M of the bolt hole box is obtained according to the box set R, and the splice joint slope dataset k and intercept dataset b are obtained according to the line set L.

[0077] To address the problem of inaccurate bolt hole positioning, this invention utilizes the center point coordinate set N{N1(x1,y1),N2(x2,y2)……N... after clustering the bolt hole point cloud. i (x i ,y iThe geometric center coordinate set M of the bolt hole frame is corrected using the bolt hole as the geometric center of the 2D image of the shield tunnel, thereby achieving precise positioning of the bolt holes. Addressing the issues of cluster center coordinate offset caused by partial occlusion of the bolt hole point cloud by the equipment and feature loss caused by complete occlusion, resulting in the inability to correspond between the two center coordinate sets M and N, this invention determines the position of the occluded bolt holes more accurately based on the relative positional relationship between the bolt holes.

[0078] Specifically, using the center coordinate set of M as a reference, find the corresponding center coordinates at the corresponding positions in the center coordinate set of N. When the corresponding center coordinates can be found, Equation 4 applies:

[0079]

[0080] In the formula, Δd i express Point and The distance between them is the corrected distance.

[0081] When the corresponding center coordinates cannot be found, Equation 5 applies:

[0082]

[0083] In the formula, Δs i This represents the relative distance between the current bolt hole and the adjacent bolt hole; Δd i Indicates the correction distance; Δd i±1 This indicates the correction distance between adjacent bolt holes.

[0084] The bolt hole frame set R is corrected using a correction distance.

[0085] To address the problem of inaccurate seam positioning, this invention designs a horizontal seam positioning method based on image gradient changes. A buffer zone with a radius of K pixels is established within both oblique and straight horizontal seams. The straight line with the smallest total pixel count is found and used as the final position of the seam. The dataset of the optimal buffer radius Δr is recorded. The intercept dataset b is finely adjusted according to Equation 6 and saved as dataset b', thereby achieving the purpose of precise seam positioning.

[0086] b' j =b j +Δr j (6)

[0087] In the formula, b j Δr represents the intercept of the j-th line in the intercept dataset. j b' represents the optimal buffer radius for the j-th line in the intercept dataset; j This represents the adjusted intercept of the j-th line in the intercept dataset.

[0088] The self-propelled 3D scanning system for rail transit, independently developed by China Railway Design Corporation, is equipped with a Z+F 9012 scanner, with a moving speed of 2 km / h. The Z+F scanner has a line frequency of 100 Hz and a point frequency of 1016 kHz, with 10,160 scanning points per cross-section and a line spacing of approximately 5.6 mm. The laser point spacing at the waist section of the tunnel on both sides is approximately 3.3 mm.

[0089] Different extraction methods were used for different elements in comparative experiments to verify the accuracy and performance of the present invention in identifying different tunnel targets. Several representative methods were selected for comparison:

[0090] Comparative literature [1] (Xu L, Gong J, Na J, et al.Shield Tunnel Convergence DiameterDetection Based on Self-Driven Mobile Laser Scanning[J].Remote Sensing, 2022, 14(3):767.);

[0091] Comparative literature [2] (Liu X, Chen Y, Liu X. Laser Scanning-based rapid detection of deformation of shield tunnel section [J]. Journal of Traffic and Transportation Engineering. 2021, 21(2):10.);

[0092] Compare with reference [3] (Cui H, Ren X, Mao Q, et al. Shield subway tunnel deformation detection based on mobile laser scanning [J]. Automation in Construction, 2019, 106: 102889.).

[0093] The comparative analysis data is shown in the table below:

[0094] Table 1 Comparison of Accuracy and Efficiency of Longitudinal Seam Recognition

[0095]

[0096] Table 2 Comparison of Bolt Hole Identification Accuracy and Efficiency

[0097]

[0098] Table 3 Comparison of transverse seam b recognition accuracy

[0099]

[0100] Table 4 Comparison of accuracy and efficiency in identifying diagonal and transverse seams (k is 0 for straight and transverse seams)

[0101]

[0102] Figure 9 , Figure 10 The diagrams show a comparison of the identification results of longitudinal seams in tunnel segments and the positioning results of bolt holes, oblique seams, and straight transverse seams using existing methods and the method of this invention. From... Figure 9 and Figure 10 As can be clearly seen from the data, this method can effectively achieve accurate identification of holes and gaps, with a higher recognition rate and accuracy compared to similar methods, and has better applicability and robustness.

[0103] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for identifying gaps in shield tunnels driven by a cross-modal feature deep fusion model, characterized in that, Includes the following steps: S1, Feature fusion of cross-modal data from moving LiDAR point cloud and dimensionality-reduced projection in shield tunnel: S11, the moving LiDAR point cloud of the tunnel is fitted using the RANSAC algorithm, and the bolt hole feature point cloud that meets the conditions is obtained according to the following formula. |PF1|+|PF2|>2a Where F1 and F2 represent the two foci of the fitted elliptical cross-section of the shield tunnel; a represents the major semi-axis of the elliptical cross-section; P represents the laser point of the bolt hole feature cloud; |PF1| represents the distance from the laser point of the feature cloud to the focal point F1; |PF2| represents the distance from the laser point of the feature cloud to the focal point F2. S12, convert the 3D tunnel-moving LiDAR point cloud into a 2D projected tile map and a 2D coordinate set according to the following formula: In the formula, P represents the laser point on the c-th cross-section point cloud. c The corresponding X-axis coordinate in the grayscale image; P represents the laser point on the c-th cross-section point cloud. c The corresponding Y-axis coordinate in the grayscale image; P represents the laser point on the c-th cross-section point cloud. c The cumulative distance between the current survey line and the starting survey line; H represents the set horizontal resolution; P represents the laser point on the c-th cross-section point cloud. c The angle mapped onto the image; R represents the radius of the projected cylinder; V represents the set vertical resolution. express The pixel value at the location, where c = 1, 2, ..., k, k is the total number of point clouds in the cross section, and Intensity represents the image intensity value; The two-dimensional coordinate set uses a cylindrical model composed of near-circular tunnel cross-sections as the projection surface and scanning survey lines as the units; S2, after classifying the tunnel segments using a dual-template method, coarse positioning of bolt holes and splice joints is performed: S21, the two-dimensional projection tiled map obtained in S1 is divided into single-ring shield two-dimensional projection tiled maps, and a coordinate system is established on it with the pixel coordinates of the upper left corner of each single-ring shield two-dimensional projection tiled map as the origin. S22, determine the main template and secondary template of the shield tunnel segment, wherein the main template conforms to the bolt hole distribution characteristics of the capping block, and the secondary template conforms to the bolt hole distribution characteristics after splicing with adjacent blocks and standard blocks; S23, using the main template and secondary template determined in S22, the bolt holes and splicing seams of the two-dimensional projection tiling are positioned to obtain a coarse positioning result, which is the set of bolt hole boxes R and the set of straight lines L of splicing seam positions; S3, Correct the coarse positioning results: S31, obtain the geometric center coordinate set M of the bolt hole frame according to the bolt hole frame set R obtained in S23, and obtain the slope dataset k and intercept dataset b of the splice seam according to the line set L obtained in S23. S32, cluster the bolt hole feature point cloud obtained in S11 to obtain the clustered center points, and then use the method in S12 to two-dimensionalize the center points to obtain the bolt hole geometric center coordinate set N{N1(x1,y1),N2(x2,y2)……N i (x i ,y i )}, where i represents the number of the i-th bolt hole, and the set of geometric center coordinates N of the bolt holes is used for precise positioning of the bolt holes; S33, using the geometric center coordinate set M of the bolt hole frame as a reference, find the corresponding center coordinates at the corresponding positions in the geometric center coordinate set N of the bolt holes. When the corresponding center coordinates can be found, it is as shown in the following formula: In the formula, Δd i express Point and The distance between them, i.e., the corrected distance. When the corresponding center coordinates cannot be found, the following formula applies: In the formula, Δs i This represents the relative distance between the current bolt hole and the adjacent bolt hole; Δd i Indicates the correction distance; Δd i±1 Indicates the correction distance between adjacent bolt holes; The bolt hole positions are obtained by correcting the bolt hole frame set R using the correction distance; S34, find the line with the smallest total pixel sum in the two-dimensional projection tiling and use this line as the final position of the stitching seam. Establish a buffer zone within the range of the stitching seam with a radius of K pixels, record the dataset of the optimal buffer radius Δr, correct the intercept dataset b according to the following formula and save it as dataset b' to obtain the corrected stitching seam position. b' j =b j +Δr j In the formula, b j Δr represents the intercept of the j-th line in the intercept dataset. j b' represents the optimal buffer radius for the j-th line in the intercept dataset; j This represents the adjusted intercept of the j-th line in the intercept dataset; The corrected splice seam position and the corrected bolt hole position are output as the precise positioning result.

2. The cross-modal feature deep fusion model-driven shield tunnel gap identification method as described in claim 1, characterized in that, The main template number is "T", and the secondary template number is "O".

3. The cross-modal feature deep fusion model-driven shield tunnel gap identification method as described in claim 1, characterized in that, When using the template designed in S22 to match and locate key segments in the two-dimensional projection tiled image of the single-ring shield tunnel segmented in S21, if the bolt hole features of the two-dimensional projection tiled image of the single-ring shield tunnel match the features of the capping block, it is classified as a type A or type B single-ring shield tunnel. If the two-dimensional projection tiled image of the single-ring shield tunnel is blurred or obscured and can only be classified by feature classification through the splicing of the standard block and adjacent blocks, it is classified as a type C single-ring shield tunnel.

4. The cross-modal feature deep fusion model-driven shield tunnel gap identification method as described in claim 1, characterized in that, For the dual-template classification in S2, the normalized correlation coefficient localization method is used to evaluate the similarity between the template and the target image. The similarity calculation is shown in the following formula: In the formula: T(x′,y′) represents the template image; I(x,y) represents the image to be located; w and h represent the width and height of the template image; R(x,y) represents the function describing the similarity.

5. The cross-modal feature deep fusion model-driven shield tunnel gap identification method as described in claim 1, characterized in that, The method for dividing the two-dimensional projection tiled image of a single-ring shield in S2 is as follows: search for the line with the lowest longitudinal gray level accumulation in the longitudinal joint in the two-dimensional projection tiled image as the final longitudinal joint position, and use the final longitudinal joint to divide the shield ring to obtain the two-dimensional projection tiled image of the single-ring shield.

Images