A segment misalignment detection method for a shield tunnel

Abstract

This invention discloses a method for detecting segment misalignment in shield tunnels, comprising: acquiring original multi-source datasets; constructing a ring-level data set indexed by ring number based on pre-stored propulsion encoder information; extracting local features of the ring joints; constructing self-calibration optimization data using the spatial deviation between the actual ring joints and the designed ring joints; jointly solving sensor extrinsic parameters and ring-level attitude correction quantities to generate a ring joint alignment result set containing precise spatial positions of the ring joints; mapping the ring joint alignment result set to the polar coordinate domain; extracting an attention-enhanced feature set through a pre-configured parallel patch perception network; and combining the geometric information in the ring joint alignment result set with the visual features in the attention-enhanced feature set to calculate the misalignment geometric vector decomposed into multi-directional components, generating a misalignment classification result set. This invention effectively solves the problem of sensor drift and difficulty in identifying minute misalignment features in long-distance tunneling, realizing fully automated and precise control of segment assembly quality throughout the entire process.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering inspection, and in particular, it is a method for detecting segment misalignment in shield tunnels. Background Technology

[0002] As the mainstream construction method for underground tunnels, the quality of tunnel boring machines (TBMs) directly depends on the assembly accuracy of the tunnel segments. Segment misalignment (step-faults), the most common quality defect, not only leads to tunnel waterproofing failure and leakage, but in severe cases, it can also cause segment cracking and damage, threatening the long-term safety of the tunnel structure. Therefore, achieving efficient, accurate, and automated detection of segment misalignment is of significant research importance and engineering value for ensuring the quality of underground engineering projects.

[0003] Currently, automated inspection technologies based on machine vision and 3D laser scanning have gradually replaced traditional manual inspection. Existing mainstream technologies typically employ laser profilometers and line-scan cameras mounted on tunnel boring machines or mobile trolleys to acquire data, and use encoder pulses for mileage positioning. For data processing, the standard Iterative Closest Point (ICP) algorithm is often used for rigid body registration of point clouds, and general-purpose convolutional neural networks (such as standard YOLO or ResNet) are used for feature extraction and defect identification from 2D images or depth maps.

[0004] However, in the long-distance and complex environment of shield tunneling, existing technologies still face challenges such as spatiotemporal alignment drift and the submersion of minute features. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting segment misalignment in shield tunnels, thereby solving the aforementioned problems in the existing technology.

[0006] Technical solution: A method for detecting segment misalignment in shield tunnels, comprising:

[0007] Obtain the original multi-source dataset containing 3D point cloud data and ring seam image data, and construct a ring-level data set indexed by ring number based on the pre-stored propulsion encoder information;

[0008] Local features of the ring gap are extracted from the ring-level data set. Self-calibration optimization data is constructed using the spatial deviation between the actual ring gap and the designed ring gap. Sensor extrinsic parameters and ring-level attitude correction are jointly solved to generate a ring gap alignment result set containing the precise spatial position of the ring gap.

[0009] The suture alignment result set is mapped to the polar coordinate domain, and the attention-enhanced feature set is extracted through a pre-configured parallel patch-aware network;

[0010] By combining the geometric information in the circumferential seam alignment result set with the visual features in the attention enhancement feature set, the misalignment geometric vectors decomposed into multi-directional components are calculated to generate a misalignment classification result set.

[0011] Beneficial effects: This invention effectively solves the problem of sensor drift and difficulty in identifying minor misalignment features during long-distance tunneling, and realizes full-process automated and precise control of segment assembly quality. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the steps of a method for detecting segment misalignment in shield tunnels, as provided in an embodiment of this application.

[0013] Figure 2 A flowchart illustrating the steps for jointly solving the sensor extrinsic parameters and ring-level attitude correction quantities provided in the embodiments of this application.

[0014] Figure 3 A flowchart illustrating the steps for constructing self-calibration optimization data provided in this application embodiment.

[0015] Figure 4 A flowchart illustrating the steps for extracting attention-enhanced feature sets provided in an embodiment of this application. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms include and have, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0018] The study found that existing methods typically assume fixed sensor extrinsic parameters, ignoring the minute physical deformations caused by tunnel boring machine (TBM) vibrations. Furthermore, ICP registration, which relies solely on surface geometry, lacks strong structural constraints, leading to a gradual deviation of the coordinate system reference during long-distance detection and failing to provide accurate geometric references for precision measurements. Additionally, the segment joints exhibit extremely narrow, arc-shaped curves. Existing deep learning networks, often employing square convolution kernels or general attention mechanisms, struggle to adapt to this unique geometric distribution, causing millimeter-level micro-misalignments to be easily masked by the texture noise of the concrete surface. Moreover, current detection methods only output a single misalignment value, unable to differentiate between intra-ring misalignments caused by assembly issues and inter-ring misalignments caused by TBM attitude problems based on geometric mechanisms, thus hindering targeted engineering rectification.

[0019] like Figure 1 As shown, a method for detecting segment misalignment in shield tunnels is proposed, including the following steps:

[0020] Obtain the original multi-source dataset containing 3D point cloud data and ring seam image data, and construct a ring-level data set indexed by ring number based on the pre-stored propulsion encoder information.

[0021] In other words, the original multi-source dataset is acquired and preprocessed to obtain the preprocessed multi-source dataset; the preprocessed multi-source dataset and the pre-stored propulsion encoder information are combined to construct a ring-level data packet set indexed by ring number; the original multi-source dataset contains 3D point cloud data, ring seam image data and attitude information.

[0022] In this embodiment, the original multi-source dataset serves as the input foundation for the entire detection process and can be acquired through a multi-sensor system mounted on the tunnel boring machine (TBM). Exemplarily, the multi-sensor system mainly includes a laser scanning device and a line array camera installed 3 to 4 rings behind the tail of the shield, a propulsion encoder installed on the stroke rod of the TBM's propulsion cylinder, and an attitude measurement device or automatic total station installed on the TBM body. The 3D point cloud data, acquired by the laser scanning device, characterizes the 3D geometry of the inner surface of the tunnel segments and is typically stored in the form of XYZ coordinate points, reflecting the spatial undulations and deformation of the segment surface. The annular joint image data is synchronously acquired by a line array camera or a high-resolution area array camera and is used to characterize the visual texture information of the segment surface, such as the color of the concrete surface, cracks, water stains, and joint line features. The attitude measurement device may specifically include a gyroscope or a laser guidance system to provide real-time attitude information such as the TBM's pitch angle, roll angle, and yaw angle.

[0023] Specifically, constructing a ring-level data packet indexed by ring number based on the propulsion encoder information refers to using the displacement pulse data generated by the propulsion encoder as a synchronization reference for time and space, and cutting and recombining continuously acquired time-series data into structured data units based on the segment ring number. In some implementations, each unit travel recorded by the propulsion encoder, such as 1.2 meters or 1.5 meters, corresponds to the completion of tunneling of one segment ring. The data processing system reads the pulse count value of the encoder, identifies the start and end timestamps of each tunneling ring, and defines the effective acquisition time window for that ring. All 3D point cloud data frames, ring seam image data frames, and corresponding attitude data whose timestamps fall within this time window are associated and packaged and stored as independent ring-level data packets. Each ring-level data packet is assigned a unique ring number index, thereby transforming massive streaming data into ordered discrete units, facilitating subsequent parallel processing or sliding window optimization by ring in subsequent algorithms.

[0024] In a preferred implementation, to eliminate the influence of construction environment noise during the construction of the ring-level data set, a density-based spatial clustering algorithm, namely the DBSCAN algorithm, can be used to preprocess the original point cloud data. Specifically, for each frame of point cloud, the system sets a neighborhood radius of, for example, 0.3 mm, and a minimum number of points, for example, 5. Each point in the point cloud is traversed, and the point density within its neighborhood is calculated. Points with a density below a threshold are marked as outliers and removed, while points in high-density areas are clustered into core point clusters. For each point cluster, its geometric center is calculated as the representative point at that location. This effectively filters out random noise caused by dust, water mist, or equipment vibration while preserving the true structural features of the pipe segment surface. This improves the signal-to-noise ratio of subsequent normal change detection and geometric fitting.

[0025] Local features of the ring gap are extracted from the ring-level data set. Self-calibration optimization data is constructed using the spatial deviation between the actual ring gap and the designed ring gap. Sensor extrinsic parameters and ring-level attitude correction are jointly solved to generate a ring gap alignment result set containing the precise spatial position of the ring gap.

[0026] In this embodiment, extracting local features of the circumferential joint refers to accurately locating and separating a subset of local data containing segment joint information from massive amounts of ring-level data. Since segment misalignment mainly occurs at joints, and joints have distinct geometric features, extracting these local features is a prerequisite for achieving high-precision alignment. Specifically, the system performs normal calculations on the point cloud data, identifying regions with drastic normal changes as the local point cloud subset of the circumferential joint; simultaneously, image processing techniques are used to identify lines with abrupt texture changes as the local image subset of the circumferential joint. The spatial deviation between the actual circumferential joint and the designed circumferential joint refers to the geometric difference between the extracted actual circumferential joint data position in the current coordinate system and the preset theoretical circumferential joint position in the tunnel design model, such as the distance from a point to a plane or the distance from a line to a line.

[0027] Specifically, constructing self-calibrating optimization data refers to organizing the aforementioned spatial deviations, currently estimated sensor extrinsic parameters, and tunnel boring machine (TBM) attitude information into an optimizable mathematical model input. Traditional measurement methods typically assume that sensor extrinsic parameters are constant; however, during long-distance tunneling, these parameters can drift slightly due to vibration and temperature. By introducing a self-calibrating mechanism, the optimization process not only corrects the TBM's attitude but also fine-tunes the sensor extrinsic parameters. The joint solution for sensor extrinsic parameters and ring-level attitude corrections involves constructing an objective function containing deviation and regularization terms, and using numerical optimization algorithms such as nonlinear least squares to find the optimal parameter combination that minimizes the overall deviation between the actual and designed ring joints.

[0028] Furthermore, to ensure the convergence of the optimization and the reliability of the results, an abnormal ring removal mechanism can be introduced. Before performing the joint solution, the system calculates the projection residual of each ring's data in the initial alignment state. For example, it calculates the average distance from the actual ring joint point cloud to the designed ring joint plane. If the average residual of a ring exceeds a preset threshold, such as 5 mm, or the point cloud coverage of the ring is lower than a preset ratio, such as 70%, then the ring is determined to be an abnormal ring. The data of abnormal rings will be marked and excluded from the input of the self-calibration optimization to prevent individual severely occluded or poorly constructed segment data from damaging the stability of the optimization algorithm. Generating a ring joint alignment result set containing the precise spatial location of the ring joints refers to using the optimal extrinsic parameters obtained after optimization convergence and the corrected pose to reproject the original point cloud and image onto a unified engineering coordinate system, obtaining the precise ring center coordinates, ring joint plane equation, and registered texture image for each ring.

[0029] In some alternative implementations, self-calibration optimization data can be constructed using the spatial deviation between the actual ring gap and the designed ring gap. This may also include comparing the advance distance recorded by the advance encoder with the designed ring width to assist in coarse time alignment, so as to determine the initial time offset of each ring data as the initial value for self-calibration optimization.

[0030] The suture alignment result set is mapped to the polar coordinate domain, and the attention-enhanced feature set is extracted through a pre-configured parallel patch-aware network.

[0031] In this embodiment, mapping the circumferential seam alignment result set to the polar coordinate domain means establishing a polar coordinate system with the radial distance r and the circumferential angle θ as the coordinate axes, using the precise center position of the ring as the origin. Since the tunnel segments are geometrically cylindrical and the seams are distributed along the circumferential or axial direction, they appear as complex curves in the Cartesian coordinate system, but transform into regular strips that are approximately straight lines in the polar coordinate system. This simplifies the learning difficulty of the subsequent feature extraction network, allowing the network to focus on identifying misalignment features distributed along a specific direction.

[0032] Specifically, extracting attention-enhanced feature sets through a parallel patch-aware network involves inputting polar-coordinate transformed image data into a pre-designed deep neural network. Unlike traditional global convolutional networks, this network employs a parallel branching structure to process different types of local regions. For example, based on the geometric location of the seam, the system divides the image into seam patches covering the seam, corner patches located at the corners of the segments, and other background patches. For each type of patch, the network extracts specific visual features through parallel convolutional branches. For instance, one branch extracts radial height abrupt changes, while another branch extracts circumferential texture continuity features. These features are then fused using attention-weighted fusion to form an enhanced feature map that is highly sensitive to misaligned areas.

[0033] By combining the geometric information in the circumferential seam alignment result set with the visual features in the attention enhancement feature set, the misalignment geometric vectors decomposed into multi-directional components are calculated to generate a misalignment classification result set.

[0034] In this embodiment, the misalignment geometric vector refers to vector data used to describe the spatial morphology of the segment misalignment. It is no longer a single numerical value, but is decomposed into components along three orthogonal directions: the tunnel axis, radial direction, and circumferential direction. The geometric information mainly comes from precise point cloud and pose data, used to calculate the physical dimensions of the misalignment; the visual features come from attention feature maps, used to provide visual confidence in the existence of the misalignment.

[0035] Specifically, calculating the misalignment geometric vector decomposed into multi-directional components involves establishing a local ring-level coordinate system at each suspected misalignment location, calculating the relative displacement of the pipe segments on both sides of the joint, and projecting this displacement onto the axial, radial, and circumferential axes. For example, the axial component reflects the magnitude of the inter-ring misalignment, while the radial component reflects the height difference between adjacent pipe segments within the ring. Generating a misalignment classification result set involves classifying the misalignment into intra-ring misalignment, inter-ring misalignment, or composite misalignment based on the numerical proportions of these components and visual feature verification, and outputting specific misalignment values ​​and classification labels. This provides construction personnel with more targeted rectification suggestions.

[0036] like Figure 3 As shown, in one possible implementation, the self-calibration optimization data is constructed, including:

[0037] Normal mutation detection is performed on the 3D point cloud data in the ring-level data package to extract the local point cloud subset of the ring seam containing the edge features of the ring seam.

[0038] In this embodiment, because there are usually tiny grooves or chamfers at the joints of the pipe segments, the direction of the point cloud normal vector in this area differs from the direction of the normal vector of the smooth surface of the pipe segment. In a preferred implementation, a subset of the local point cloud of the circumferential joint containing the edge features of the circumferential joint is extracted, including:

[0039] For each 3D point in the ring-level data packet set, a local geometric neighborhood is constructed, and the local normal vector of the 3D point is calculated using the spatial distribution within the local geometric neighborhood;

[0040] Calculate the angle change between the local normal vectors of adjacent 3D points, and mark the region with the angle change greater than the preset normal change threshold as a local point cloud subset of the circumferential seam; wherein, the normal change threshold is dynamically generated based on the statistical mean and standard deviation of the local point cloud normal angle distribution, or set as a fixed empirical value between 15 degrees and 30 degrees.

[0041] Specifically, the system constructs a local geometric neighborhood for each 3D point, for example, selecting a neighborhood with a radius of 5 mm or containing the 10 nearest points. Principal component analysis (PCA) is used to calculate the covariance matrix of this neighborhood; the eigenvector corresponding to the smallest eigenvalue of this matrix is ​​the local normal vector of that point. The angle between the normal vectors of adjacent points is calculated. Let the normal vector of point A be n. A The normal vector of point B is n. B Then the normal angle θ can be obtained using the inverse cosine function. A preset normal abrupt change threshold θ is used. th When the calculated angle θ is greater than the threshold, the two points are determined to be located in the region of abrupt change in normal direction, i.e., the suspected edge of the circumferential joint. The threshold for abrupt change in normal direction can be set as a fixed empirical value between 15 and 30 degrees, such as 20 degrees. In some preferred embodiments, the threshold can also be generated adaptively, specifically by calculating the statistical mean μ and standard deviation σ of the included normal angles of all points in the local region, and setting the threshold to μ + 3σ. This dynamic threshold can adapt to concrete surfaces with different roughness, improving the robustness of detection. The extracted set of points with high normal change rates constitutes a subset of the local point cloud of the circumferential joint.

[0042] Perform seam texture recognition on the seam image data in the ring-level data packet set, and extract a subset of local images of the seam containing the seam centerline features.

[0043] In this embodiment, seam texture recognition utilizes computer vision technology to locate seam positions from a two-dimensional image. For example, since seams typically appear as dark-colored lines of a certain width in an image, the system performs grayscale conversion and contrast enhancement on the image. Edge detection operators, such as the Canny or Sobel operators, or line structure detection algorithms, such as the Steger algorithm, are used to identify linear textures in the image. Based on preset seam width ranges and directional constraints, texture lines that conform to the characteristics of segment seams are selected, and their centerline pixel coordinates are extracted to form a subset of the local image of the circumferential seam.

[0044] The local geometric surface is fitted based on a subset of the local point cloud of the circumferential seam, and the centerline constraint correction of the local geometric surface is performed by using the texture position of the local image subset of the circumferential seam in the spatial projection, thereby generating the actual geometric parameter set of the circumferential seam.

[0045] In this embodiment, a single data source often has limitations. For example, the point cloud may contain noise at the joint groove, and the image lacks depth information. Complementary advantages can be achieved through joint fitting. Specifically, the system uses the least squares method to fit a subset of the local point cloud of the circumferential joint to obtain the local plane or surface equations on both sides of the joint. In some alternative implementations, to accommodate more complex nonlinear deformations of the pipe segment surface, deep learning networks such as PointNet++ can be used to perform feature encoding and surface regression on the local point cloud, replacing the traditional least squares fitting. Based on this, the joint centerline of the local image subset of the circumferential joint is projected into three-dimensional space using the current initial extrinsic parameters. If there is a deviation between the projected position and the geometric joint position fitted from the point cloud, the system will utilize the high-resolution characteristics of the image texture to fine-tune and correct the position of the geometric joint centerline. For example, the point cloud fitting center is weighted closer to the image texture center. The final generated set of actual circumferential joint geometric parameters includes the corrected spatial coordinates of the joint centerline, the joint plane normal, and high-precision geometric parameters such as the joint width.

[0046] The actual circumferential seam geometry parameter set is spatially paired with the preset design circumferential seam geometry parameter set, the geometric deviation between the two is calculated, and self-calibration optimization data containing the geometric deviation is constructed.

[0047] Specifically, the geometric parameters of the design circumferential joints are derived from the tunnel's BIM model or design drawings, encompassing the theoretical spatial location and shape of each segment joint. The system uses the ring number index to map the extracted parameters one-to-one with the design parameters. The geometric deviation can be specifically represented as the vertical distance from a point on the actual circumferential joint to the design circumferential joint plane. The self-calibration optimization data is a structured dataset that includes not only the aforementioned deviations but also the initial values ​​of the sensor extrinsic parameters at the current moment, the initial values ​​of the tunnel boring machine's attitude, and the confidence weights of each data point.

[0048] like Figure 2 As shown, in an exemplary embodiment, jointly solving for the sensor extrinsic parameters and the ring-level attitude correction includes:

[0049] The design annular seam plane and the actual annular seam point cloud subset contained in the self-calibration optimization data are read. The main objective is to minimize the sum of squares of the spatial distances from each point in the actual annular seam point cloud subset to the corresponding design annular seam plane after pose transformation. Combined with regularization terms that limit the variation of sensor extrinsic parameters and limit the numerical amplitude of the ring-level attitude correction, an alignment optimization objective function is constructed.

[0050] In this embodiment, by constructing an alignment optimization objective function, the optimal transformation parameters are found to ensure the best fit between the observed data and the theoretical model. Specifically, the objective function J can be expressed in the following form:

[0051] J(T,{ΔR i}) =Σ i，k Σ p∈Pik d(Π i k ΔR i ·R i ·T·p) 2 +λ1Σ i ||ΔR i || 2 +λ2||T -T0|| 2 ;

[0052] Where T represents the sensor extrinsic parameter transformation matrix to be solved; ΔR i Let R represent the attitude correction amount for the i-th ring, which is also the optimization variable to be solved; i The initial attitude of the i-th ring is provided by the attitude sensor; p represents a point in the actual ring gap point cloud set; Π i k P represents the plane of the k-th design ring joint of the i-th ring; d() represents the Euclidean distance function from the point to the plane; i k Let be the subset of the actual point cloud of the i-th ring and k-th seam; T0 is the initial extrinsic parameter. The first term of the function is the data fitting term, which minimizes the sum of squared distances from the point cloud to the design plane to ensure geometric alignment accuracy. The second term is a regularization constraint on the attitude correction, where λ1 is the corresponding weight coefficient. This term limits the magnitude of attitude correction, preventing drastic, non-physical attitude jumps during optimization and ensuring trajectory smoothness. The third term is a regularization constraint on extrinsic parameter changes, where λ2 is the corresponding weight coefficient. This term constrains the range of extrinsic parameter drift, ensuring that the changes conform to physical laws and preventing overfitting. By adjusting the values ​​of λ1 and λ2, the degree of data fitting and the stability of the model parameters can be balanced.

[0053] Based on the alignment optimization objective function, the sensor extrinsic parameters and the ring-level attitude correction amount of each ring are solved simultaneously through a numerical iterative algorithm.

[0054] In a preferred implementation, a numerical iterative algorithm is used to simultaneously solve for the sensor extrinsic parameters and the ring-level attitude correction for each ring, including:

[0055] Based on the pre-stored tunnel boring machine (TBM) excavation sequence, consecutive predetermined ring data are selected from the ring-level data packet set to construct a sliding window optimized data subset.

[0056] In this embodiment, to meet the requirements of continuity and real-time operation in tunnel boring machine (TBM) construction, global optimization of all data is not used; instead, a sliding window strategy is employed. Specifically, the system selects several recently generated consecutive rings of data, such as the most recent 20 rings, to construct a subset of data for sliding window optimization. The window size W is set to 20 to ensure that the window contains a sufficient number of ring joint constraints. Optionally, each ring contains 3 or more joints, and 20 rings can provide more than 60 geometric constraints, sufficient to solve for extrinsic parameters and attitude variables.

[0057] Substitute the optimized subset of data into the alignment optimization objective function and perform iterative optimization to solve the problem until the alignment optimization objective function value converges or reaches the preset number of iterations. Output the multi-source alignment parameter set corresponding to the current window. The multi-source alignment parameter set includes sensor extrinsic parameters and ring-level attitude corrections for each ring.

[0058] For example, the system uses the Levenberg-Marquardt (LM) algorithm or the Gauss-Newton method to perform a nonlinear iterative solution to the constructed objective function. In each iteration, the objective function is calculated with respect to the optimization variables T and ΔR. i The Jacobian matrix is ​​calculated, and variable values ​​are updated to reduce the total function error. The iterative process continues until the change in the objective function value between two consecutive iterations is less than a preset convergence threshold, such as 1e2. -6 Alternatively, it can reach the maximum number of iterations, such as 100. At this point, the output parameter set is the optimal solution for the current window, containing the latest estimates of sensor extrinsic parameters and the precise attitude of each loop within the window.

[0059] The pose information of each ring in the ring-level data packet set is updated using a multi-source alignment parameter set, and the sliding window moves forward as the tunnel boring machine advances to perform incremental alignment solution for the ring-level data packets added later.

[0060] In this embodiment, whenever the tunnel boring machine completes a new ring of excavation, the sliding window moves forward by one step (one ring). The system removes the oldest ring data from the window, adds the latest ring data, and uses the optimization result of the previous window as the initial value for the current window to perform optimization again. This allows the system to track the slow drift of sensor extrinsic parameters in real time, maintaining a high-precision alignment and achieving online dynamic self-calibration.

[0061] like Figure 4 As shown, according to one aspect of this application, an attention-enhancing feature set is extracted, including:

[0062] Using the precise ring center position in the ring seam alignment result set as the origin of polar coordinates, the image data in the ring seam alignment result set is mapped into a polar coordinate ring seam image set represented by circumferential and radial coordinates.

[0063] In this embodiment, since the tunnel lining segments are distributed cylindrically in physical space, their joints typically appear as complex curves or arcs in traditional Cartesian coordinate images, making it difficult for convolutional neural networks to extract features using fixed convolutional kernels. For example, by introducing the precise ring center position calculated using the SA-Align algorithm, the system establishes a polar coordinate system with this center as the origin, the tunnel radius direction as the radial coordinate r, and the circumferential angle as the circumferential coordinate θ. Using a bilinear interpolation algorithm, the original image pixels are resampled into this polar coordinate system. After this transformation, the originally curved ring joints are straightened into approximately horizontal or vertical stripes in the polar coordinate image, reducing the geometric complexity of subsequent feature extraction.

[0064] Based on the radial distribution of the precise spatial position of the annular seam in polar coordinates in the annular seam alignment result set, a narrow annular seam region mask covering the seam area is generated.

[0065] Specifically, the seam narrow band region mask is a binarized or probabilistic matrix used to indicate which regions in the image belong to the seam area of ​​interest. Since SA-Align already provides the precise spatial position equation for each seam, projecting it onto the polar coordinate image corresponds to a specific radial position curve. The system sets a band-shaped region centered at this location with a preset pixel value, such as 32 pixels or 64 pixels, marking the pixels within this region as foreground and the remaining regions as background. This guides the network to concentrate computational resources on the areas most likely to have misalignments, avoiding the computational waste and background false detections caused by full-image search.

[0066] The polar coordinate annular seam image set is divided into a predetermined number of regular image blocks. Based on the spatial overlap relationship between each regular image block and the annular seam narrow band region mask, the regular image blocks are classified and an annular seam patch set, a corner patch set located at the corner of the segment, and a background patch set are constructed respectively.

[0067] In this embodiment, the system divides the high-resolution polar coordinate image into fixed-size image blocks, or patches, such as small blocks of 16×16 pixels or 32×32 pixels. The intersection-over-union (IoU) value between each patch and the mask of the narrow band of the circumferential seam is calculated. Patches with an IoU value greater than a first preset threshold, such as 0.6, are classified as circumferential seam patches, which mainly contain seam textures. Patches located near the cross joints or T-joints of the pipe segments and with IoU values ​​within a specific range are classified as corner patches, which contain key corner geometric features. Patches with an IoU value less than a second preset threshold, such as 0.1, are classified as background patches. This provides a data foundation for subsequently employing different feature extraction logics for different regions.

[0068] By using the set of circumferential seam patches, the set of corner patches, and the set of background patches as input data for a parallel patch-aware network, an attention-enhanced feature set is obtained.

[0069] Specifically, unlike traditional networks that treat all regions equally, the parallel patch-aware network allows patches from different sets to flow into different processing branches, thereby enabling specialized customization of feature extraction.

[0070] In a further embodiment, extracting the attention-enhanced feature set further includes performing parallel feature extraction on each patch set, specifically:

[0071] For the circumferential patch set, a one-dimensional convolution kernel along the radial direction is used to extract the stagger response features reflecting the high abrupt changes, and a convolution kernel extending along the circumferential direction is used to extract the seam texture features reflecting the seam connectivity.

[0072] In this embodiment, the projection of the height difference caused by the misalignment onto the image can be captured through the misalignment response features. Since misalignment typically manifests as a step signal perpendicular to the seam direction, it appears as a radial grayscale abrupt change in polar coordinates. Therefore, this branch preferably uses a slender convolution kernel of size 1×k, where k is 5 or 7, and combines it with a Sobel operator or a high-pass filter to enhance radial gradient information. The continuity of the seam itself can be captured through seam texture features. This branch preferably uses a flat convolution kernel of size k×1. In some preferred embodiments, to accommodate the minute bending deformations that may exist in seams in actual engineering, this branch employs deformable convolution technology, allowing the sampling points of the convolution kernel to adaptively shift, thereby more accurately covering the seam texture and extracting robust seam connectivity features.

[0073] For the corner patch set, multi-scale convolution kernels and high-frequency enhancement operators are used to extract high-frequency corner features that reflect the edges of corner steps.

[0074] Specifically, the corners of the tunnel segments are high-risk areas for misalignment and damage, and their complex geometry includes edges in multiple directions. Therefore, this branch employs a multi-scale feature extraction strategy, such as using 3×3, 5×5, and 7×7 convolutional kernels in parallel to capture corner morphology under different receptive fields. Simultaneously, a Laplacian operator or a Difference of Gaussian operator is introduced as a high-frequency enhancement layer to highlight the sharp edges of the corners and the irregular boundaries formed by concrete spalling, generating high-frequency corner features highly sensitive to minor damage.

[0075] The misalignment response features, seam texture features, and corner high-frequency features are used as the basic feature components for generating the attention-enhanced feature set.

[0076] In this embodiment, the feature maps output by the three branches are uniformly adjusted to the same channel dimension, for example, by adjusting them to 256 channels through 1×1 convolution, and used as input material for the subsequent fusion module.

[0077] In a further embodiment, extracting the attention-enhanced feature set also includes calculating patch-level weights, specifically:

[0078] Global pooling is performed on the misalignment response features, seam texture features, and corner high-frequency features to obtain feature statistical vectors that reflect the intensity of each branch feature.

[0079] In this embodiment, to determine which feature is most important for a given patch, the system needs to compress and statistically analyze the features. Specifically, global average pooling (GAP) is used to compress the spatial feature map output by each branch into a one-dimensional feature statistical vector, which summarizes the overall response strength of the patch in the corresponding feature dimension.

[0080] The feature statistical vector is input into a preset weight mapping network, the attention response value corresponding to each branch is calculated, and the attention response value is normalized to generate a set of patch-level weight coefficients that satisfy local constraints.

[0081] Specifically, the weight mapping network can establish a non-linear mapping relationship between feature strength and importance weights. This network includes dimensionality reduction layers and dimensionality increase layers, where the compression ratio r of the dimensionality reduction layer is preferably set to 16 to reduce the number of parameters. Assuming the input feature dimension is C, the dimension after dimensionality reduction is C / 16. The system uses the Softmax function to normalize the output of the mapping network, ensuring that the sum of the three generated weight coefficients α1, α2, and α3 equals 1. This set of weight coefficients is the patch-level weight coefficient set, dynamically reflecting whether the current patch resembles a misalignment, seam, or corner.

[0082] In a further embodiment, extracting the attention-enhanced feature set further includes:

[0083] The patch fusion feature is obtained by weighting and summing the misalignment response feature, seam texture feature and corner high-frequency feature using the patch-level weighted coefficient set.

[0084] In this embodiment, the system performs a linear weighted combination of the feature vectors of the three branches based on the calculated weight coefficients. For example, if a patch is located at a seam and there is obvious misalignment, the weight α1 of the misalignment response feature will automatically increase, making the fused feature more prominent in terms of abrupt changes and suppressing background texture noise.

[0085] Based on the spatial location index of each patch in the polar coordinate domain, the patch fusion features are mapped back to the corresponding regions of the feature map, and normalization overlay processing is performed on the spatially overlapping regions to form a globally consistent attention enhancement feature set.

[0086] Specifically, since there may be spatial overlap during patch partitioning, for example, an overlap rate of 50%, direct splicing would cause the values ​​of the overlapping areas to multiply. Therefore, when writing the fused features back to the global feature map, the system records the number of times each pixel location is covered by a patch. The final feature value is equal to the sum of the feature values ​​of all patches covering that location divided by the number of coverages. After normalization and stacking, the patching effect is eliminated, resulting in a spatially continuous global attention-enhanced feature map with highly significant misaligned features.

[0087] This embodiment solves the technical problem of low detection rate of small targets caused by geometric distortion and background noise when processing annular tunnel textures in traditional computer vision methods.

[0088] In one embodiment of this application, calculating the staggered geometric vector decomposed into multi-directional components includes:

[0089] Using the ring center position and ring-level attitude information in the ring alignment result set, a local ring-level coordinate system containing axial, radial, and circumferential basis vectors is constructed for each misalignment detection position.

[0090] In this embodiment, the local ring-level coordinate system is a reference system independently established for each ring segment of the tunnel, different from the global geodetic coordinate system. For the i-th ring segment, the precise ring center O calculated using the SA-Align algorithm is... i The coordinate system is defined as follows: The axial basis vector Z points in the tunnel excavation direction, i.e., along the tunnel's central axis; the radial basis vector R points from the origin to the current misalignment detection point, i.e., perpendicular to the segment surface and outwards; and the circumferential basis vector T is defined as orthogonal to the axial and radial vectors, i.e., along the segment tangent direction. The establishment of this local coordinate system gives the subsequently calculated displacement components a clear engineering physical meaning.

[0091] In the local ring-level coordinate system, the spatial displacement between the locally fitted surfaces on both sides of the joint is calculated to obtain the geometric difference vector of the misalignment.

[0092] For example, the system selects point cloud data within a certain range, such as 50 mm to 100 mm, on both sides of the suspected misalignment location. The system then uses either the least squares method or the Random Sample Consensus (RANSAC) algorithm to fit the local planes on both sides. The system calculates the relative displacement vector between the two planes at the center of the joint. This vector connects the points on the left plane with the corresponding points on the right plane, representing the physical spatial difference between the two sides of the joint.

[0093] Project the misalignment geometric difference vector onto the basis vector direction of the local ring-level coordinate system to generate a misalignment geometric vector set containing the misalignment axial component, misalignment radial component, and misalignment circumferential component.

[0094] In this embodiment, the spatial displacement vector is decomposed into three scalar components through vector dot product operation. Among them, the misalignment axial component Δh... axial This characterizes the step along the tunnel axis, typically corresponding to inter-ring misalignment; the radial component Δh of the misalignment. radial It characterizes the height difference along the radial direction, usually corresponding to the unevenness of the segment assembly; the circumferential component of the misalignment Δh tangential It represents the amount of crack or opening along the circumferential direction.

[0095] In a further embodiment, generating a misclassification result set includes:

[0096] Read the axial, radial, and circumferential components of the misalignment at each detection location in the misalignment geometric vector set.

[0097] Specifically, the system extracts the specific values ​​of the three components from the generated vector set, usually in millimeters.

[0098] When the axial component of the misalignment in the geometric vector set is greater than the radial and circumferential components of the misalignment, the corresponding detection position is determined to be an inter-ring misalignment type.

[0099] In this embodiment, the determination logic is based on a preset engineering threshold. Specifically, an axial misalignment threshold T1 is set, for example, 2 mm. If the absolute value of the measured axial component is greater than T1, and the ratio of the axial component to the radial component is greater than a preset ratio, for example, 2.0, then it is determined that inter-ring misalignment has mainly occurred at that location. This means that there is an overall front-to-back displacement between two adjacent ring segments, which may require adjustment of the tunnel boring machine's propulsion parameters or grouting pressure.

[0100] When the radial or circumferential component of the misalignment in the geometric vector set is greater than the axial component of the misalignment, the corresponding detection position is determined to be an in-ring misalignment type.

[0101] For example, a radial or circumferential misalignment threshold T2 is set, such as 3 mm. If the absolute value of the measured radial or circumferential component is greater than T2, and its ratio to the axial component is greater than a preset ratio, such as 2.0, then it is determined that a misalignment mainly occurs within the ring. This means that there is a step in the assembly of adjacent segments within the same ring, and it may be necessary to check the segment selection or the lifting accuracy of the assembly machine. In some embodiments, if the above ratio is between 0.5 and 2.0, that is, the three components are of similar magnitude, then it is determined to be a composite misalignment type.

[0102] The judgment results are weighted by combining the local response intensity of the attention-enhanced feature set, and the output is a misalignment classification result set containing the misalignment type, dominant direction and misalignment quantity.

[0103] In this embodiment, to further improve the robustness of classification, visual features are introduced as verification. The local response intensity output by the HJ-PPA network is a value between 0 and 1, representing the probability of a visual misalignment. The system multiplies this response intensity as a confidence factor into the geometric classification result. For example, although geometric calculations show the presence of a minor misalignment, if the visual response intensity is extremely low, the system may mark it as a false detection or a low-confidence result to avoid false alarms. In addition, the system also has incremental learning capabilities. Whenever the tunnel boring machine completes 5 rings, the system collects the detection data distribution of these 5 rings and recalibrates the aforementioned thresholds T1 and T2 to achieve dynamic updates of the criteria.

[0104] This embodiment solves the problem that traditional methods have difficulty distinguishing between misaligned sections within and between rings, resulting in weak targeted maintenance decisions.

[0105] In one optional embodiment, the original multi-source dataset includes at least: three-dimensional point cloud data from a laser scanning device, used to characterize the spatial geometry of the tunnel segment surface; annular seam image data from an image acquisition device, used to characterize the visual texture information of the tunnel segment surface; displacement pulse data from a propulsion encoder, used to determine the correspondence between the tunnel boring machine's propulsion distance and the ring number; and attitude data from an attitude measurement device or a total station, used to provide an initial spatial attitude estimate of the tunnel boring machine body.

[0106] According to one aspect of this application, a segment misalignment detection device is provided, the hardware architecture of which includes a multi-source sensor assembly and an edge computing unit.

[0107] In this embodiment, the main structure of the segment misalignment detection device includes a rigid frame, which is fixed to the support beam of the shield machine's No. 1 trolley by bolts. A toothed circular track is mounted on the frame, and a sliding trolley that can move along the circumference is mounted on the track. The sliding trolley serves as a sensor carrier, integrating a high-precision linear array camera and a 3D laser profilometer. The segment misalignment detection device is installed at the 3rd to 4th ring behind the shield tail, where the segments have just exited the shield tail and are assembled, representing the optimal window for quality inspection. The edge computing unit uses a high-performance embedded processor, such as the NVIDIA Jetson series, to locally execute the real-time algorithm processing in the segment misalignment detection method for shield tunnels as described in any of the above embodiments.

[0108] In some preferred embodiments, considering that subway construction typically involves multiple sections and different construction units, the original engineering data has sensitive and confidentiality requirements. A federated learning architecture can be adopted, with the edge computing units on each tunnel boring machine acting as local nodes. The HJ-PPA network and the misalignment classification model are trained locally using locally collected segment data. After training, each node only uploads the model's gradient update parameters, i.e., the encrypted weight changes, to the cloud aggregation server, without uploading the original point cloud or image data. The cloud server performs weighted aggregation of the gradients from each node, updates the global model parameters, and then distributes them back to each local node. While ensuring the data privacy of each construction unit, the model's generalization ability to different geological conditions and segment types is improved by utilizing data from the entire network.

[0109] In some implementations, for fitting local geometric surfaces, in addition to employing the least squares method, the system also incorporates a deep learning fitting module based on PointNet++. This module can directly process unstructured point cloud inputs and regress the higher-order surface equations of the segment surface through hierarchical feature learning. Compared to traditional planar fitting, this embodiment can provide a geometric benchmark that more closely matches the real surface when dealing with local damage, grouting holes, or irregular deformations on the segment surface, further improving the accuracy of misalignment calculation.

[0110] As an optional implementation, a coarse alignment process based on design priors is performed before extracting local features of the ring joints from the ring-level data packet set. Specifically, the ring-level data packet set is read and the tunnel design model and design ring joint location data pre-stored in the design database are imported. Based on the ring number, theoretical ring center position, design ring joint number, and ring joint spatial distribution in the tunnel design model, the ring-level data packet set is traversed by ring number, and corresponding design ring center coordinates, design ring joint plane, or design ring joint centerline parameters are appended to each ring-level data packet to form the basic part of the initial alignment dataset. On this basis, an initial coordinate mapping relationship is established. Specifically, the attitude data and control point measurement data corresponding to each ring-level data packet are read from the basic part of the initial alignment dataset. According to the correspondence between the engineering control point coordinates and point cloud coordinates recorded in the control point measurement data, the initial extrinsic parameters from the laser scanning equipment coordinate system to the engineering coordinate system are calculated using a rigid body transformation solution method; at the same time, based on the shield attitude information recorded in the attitude data, the mapping relationship between the shield machine body coordinate system and the engineering coordinate system is derived. These parameters constitute the initial coordinate mapping parameters. Furthermore, coarse time alignment based on ring number and advancement information is performed. Multi-source time series data index information is read from the initial alignment dataset. Based on the shield advancement distance changes recorded in the encoder displacement pulse data, and combined with the designed ring length for each ring, a reference acquisition time window is determined for each ring-level data packet. The acquisition timestamps of the 3D point cloud data and ring seam image data are compared with the reference alignment timestamps. Point cloud frames and image frames whose acquisition times fall within the reference time window are assigned to the corresponding ring-level data packets, forming a coarse alignment time index for each ring. Using the initial coordinate mapping parameters, representative point cloud subsets and representative image subsets are projected onto the engineering coordinate system to generate multi-source coarse alignment data. This data includes the approximate ring center position, approximate ring seam position, and corresponding point cloud distribution for each ring, providing good initial iteration values ​​for subsequent self-calibration optimization and preventing optimization from getting trapped in local optima.

[0111] In one optional implementation, after extracting the attention-enhanced feature set through a parallel patch-aware network, the method further includes parsing the attention-enhanced feature set using an object detection head network. Specifically, the attention-enhanced feature set is input into a detection head network based on a YOLO series or similar architecture. This head network performs sliding window scanning or anchor box regression on the attention-enhanced feature map to detect and locate potential misalignment regions in and around the annular seam region. It outputs a misalignment candidate result set, which includes several image patches of candidate regions identified as misalignments, the spatial location (bounding box coordinates) of each candidate region in the engineering coordinate system, and the corresponding visual saliency score. The visual saliency score is a confidence value between 0 and 1, quantifying the probability that the region visually belongs to a misalignment. This result set retains the visual feature information enhanced by the HJ-PPA network and maintains a one-to-one correspondence with the geometric positions in the annular seam alignment result set, providing a target index for subsequent geometric decomposition.

[0112] In another alternative implementation, before calculating the misalignment geometric vectors decomposed into multi-directional components, the system needs to determine which ring and which ring seam each misalignment candidate region belongs to. Specifically, it reads the ring center position, ring-level attitude information, and ring seam spatial position from the ring seam alignment result set, while simultaneously reading the misalignment candidate result set. It calculates the Euclidean distance from the center point of each misalignment candidate region to each actual ring seam spatial curve, and associates the candidate region with the nearest ring seam. Through spatial association, a misalignment candidate association dataset is constructed. Each record in this dataset contains the ring number, ring seam number, spatial coordinates of the misalignment candidate region, and its corresponding saliency score. This ensures that when establishing a local ring-level coordinate system, the ring's attitude information can be accurately used as a reference, avoiding coordinate system confusion caused by cross-ring matching.

[0113] In another optional implementation, after generating the misalignment classification result set, the system summarizes the classification results of individual points into a macro-level quality report. Specifically, it reads the misalignment classification result set and combines it with the ring number information from the ring joint alignment result set. For each segment ring, it counts the number of intra-ring misalignments, the number of inter-ring misalignments, the maximum misalignment value, the average misalignment value, and the distribution of misalignments in the circumferential angle at all candidate positions within it (e.g., a 2mm intra-ring misalignment at the 3 o'clock position). These statistical indicators are organized into ring-level misalignment analysis items and further summarized to generate a structured result of the overall tunnel segment misalignment analysis. This structured result can be exported as an Excel report or a visualized quality distribution heatmap, directly used for project acceptance or quality assessment reports. In some optional implementations, this result is also used as a historical database for training and updating dynamic threshold prediction models in subsequent tunneling, achieving closed-loop utilization of data value.

[0114] In one exemplary embodiment, the raw multi-source dataset containing 3D point cloud data and circumferential seam image data can be acquired using a dedicated intelligent detection device for segment misalignment.

[0115] In this embodiment, the acquisition of the original multi-source dataset relies on a mechanical device. The main structure of this device includes a welded frame, for example, made of 50×50 square tubular material, which is fixed to the fulcrum of the tunnel boring machine trolley via adjustable locking supports. A toothed circular track is mounted on the frame, which engages with a sensor-equipped slider trolley via a transmission gear pair. A vision sensor is mounted sideways on the slider trolley to increase the effective scanning area, and is driven by a servo motor and reducer to move in a circular motion along the track.

[0116] In some specific implementations, to adapt to the confined space inside the tunnel boring machine, the sliding trolley and track components are preferably made of aluminum alloy to reduce weight. The servo motor motion control system supports the sliding trolley to run at a maximum speed of 600 mm / s, and the speed can be adjusted according to the actual data acquisition needs on site. One end of the moving cable chain is fixed to the center of the frame, while the other end reciprocates with the trolley, ensuring the stability of the sensor harness during movement and guaranteeing the spatial continuity and stability of the 3D point cloud data and annular seam image data during acquisition, providing high-quality physical input for subsequent algorithm processing.

[0117] Furthermore, when extracting attention-enhanced feature sets through a parallel patch-aware network, an improved deep neural network with an embedded parallelized patch-aware attention module is employed.

[0118] In this embodiment, the parallel patch-aware network does not exist independently, but is embedded as a core component within a larger target detection framework. Preferably, this detection framework is based on the YOLO11 algorithm model. This architecture includes a backbone network, a neck network, and a detection head. To address the problem of traditional models overwhelming defect features in edge regions, the system embeds the parallel patch-aware attention module, i.e., the HJ-PPA module, into the C2f module of the YOLO11 backbone network.

[0119] Specifically, when the feature map flows through the C2f module, the system utilizes the HJ-PPA module to perform polar coordinate domain patching and three-branch feature enhancement on the feature map. The embedded design leverages HJ-PPA's focusing ability on high-frequency misalignment areas such as segment joints and corners, improving the response intensity of small-sized misalignment features through adaptive weight allocation. Simultaneously, this network architecture may also incorporate the SPPFCSPC module to enhance multi-scale feature aggregation capabilities. The HJ-PPA-enhanced feature map is then fed into the decoupled detection head, where bounding box regression of misalignment locations and classification prediction of misalignment types are performed, improving the detection rate of minute misalignments while maintaining detection speed.

[0120] According to one aspect of this application, generating the ring joint alignment result set can also involve: reading the ring-level data package set, as well as the tunnel design model and design ring joint location data pre-stored in the design database. Based on the ring number, theoretical ring center position, design ring joint number, and ring joint spatial distribution in the tunnel design model, the ring-level data package set is traversed by ring number, and corresponding design ring center coordinates, design ring joint plane, or design ring joint centerline parameters are appended to each ring-level data package to form the basic part of the initial alignment dataset containing measured data and design priors. The attitude data and control point measurement data corresponding to each ring-level data package are read from the basic part of the initial alignment dataset, and the three-dimensional point cloud data and ring joint image data in the ring-level data packages are read simultaneously. Based on the correspondence between the engineering control point coordinates and point cloud coordinates recorded in the control point measurement data, the initial extrinsic parameters from the laser scanning equipment coordinate system to the engineering coordinate system are calculated using a rigid body transformation solution method; based on the shield attitude information recorded in the attitude data, the mapping relationship between the shield machine body coordinate system and the engineering coordinate system is derived. By combining the extrinsic parameters of the laser scanning device with the shield attitude mapping relationship, the initial coordinate mapping parameters for projecting point cloud data and circumferential seam image data onto the engineering coordinate system are obtained, and these initial coordinate mapping parameters are written into the aligned initial dataset.

[0121] The system reads the multi-source time series data index information, encoder displacement pulse data from the ring-level data packets, and the corresponding ring number from the initial alignment dataset. Based on the shield tunneling distance changes recorded in the encoder displacement pulse data and combined with the designed ring length of each ring, a reference acquisition time window is determined for each ring-level data packet. A representative timestamp is selected within this window as the reference alignment timestamp for that ring. The acquisition timestamps of the 3D point cloud data and ring gap image data are compared with the reference alignment timestamps. Point cloud frames and image frames whose acquisition times fall within the reference time window are assigned to the corresponding ring-level data packets, forming a coarse alignment time index for each ring. This coarse alignment time index is written into the initial alignment dataset. The initial coordinate mapping parameters and coarse alignment time index stored in the initial alignment dataset are read, and the corresponding 3D point cloud data and ring gap image data are read from the ring-level data packet set. According to the coarse alignment time index, point cloud frames and image frames adjacent to the reference alignment timestamp of each ring are selected from the multi-source time series, and these point cloud frames and image frames are used as the representative point cloud subset and representative image subset for that ring, respectively. Using initial coordinate mapping parameters, each point in the representative point cloud subset and each pixel in the representative image subset are projected into the engineering coordinate system, resulting in multi-source coarse alignment data represented in a unified engineering coordinate system. This multi-source coarse alignment data includes the approximate ring center position, approximate ring seam position, and the corresponding point cloud distribution and image coverage for each ring. Simultaneously with generating the multi-source coarse alignment data, the projection residual statistics and alignment quality indicators for each ring are updated and written into the initial alignment dataset. This information is used for screening and weighting abnormal rings during subsequent self-calibration optimization.

[0122] The approximate ring center position, approximate ring seam position, and spatial extent of the design ring seam are obtained from the multi-source coarse alignment data and the design ring seam position data in the initial alignment dataset. Based on the approximate ring seam position and the spatial extent of the design ring seam, a candidate ring seam region envelope containing a preset radial width and circumferential length is constructed for each ring seam in the engineering coordinate system. This envelope is then mapped back to the corresponding point cloud index and image pixel region index, resulting in the candidate ring seam point cloud region and candidate ring seam image region for each ring seam. The candidate ring seam point cloud region and candidate ring seam image region serve as inputs for subsequent local fine extraction, ensuring that the local extraction of the ring seam is spatially limited to the vicinity of the design ring seam, thus improving the constraint targeting of subsequent self-calibration optimization. Based on the candidate ring seam point cloud region, the 3D point cloud data within the corresponding region is read from the multi-source coarse alignment data. A neighborhood search is performed on the candidate point cloud to construct a local geometric neighborhood for each point, and the local normal vector of the point is calculated using the spatial distribution of points within the local neighborhood. Abrupt changes in normal vectors are identified based on the changes in the angle between the normal vectors of adjacent points. Points with abrupt changes in normal vector angles greater than a preset threshold are marked as suspected ring seam edge points. Based on this, connectivity analysis and small cluster removal are performed on the point cloud set of abrupt changes in normal vectors, retaining continuous point cloud bands that are close to the designed ring seam location as a subset of the ring seam's local point cloud. Based on the candidate ring seam image regions, ring seam image data within the corresponding regions are read from ring-level data sets or multi-source coarse-aligned data. Gray-level normalization and contrast enhancement are performed on the candidate ring seam image regions to improve the gray-level difference between the ring seam texture and the background. Edge detection operators or line structure detection operators are used to extract obvious linear texture responses within the candidate regions, and texture bands consistent with the designed ring seam direction are selected based on the connectivity and directional characteristics of the linear responses. For the selected texture bands, a seam centerline with an approximate single-pixel width is obtained through thinning and skeleton extraction algorithms. The pixel position of this centerline is mapped to the engineering coordinate system, forming a subset of the ring seam's local image containing the image domain seam position and the spatial domain projection position.

[0123] The point cloud positions and projected joint positions on the same circumferential seam are read from the local point cloud subset and the local image subset of the circumferential seam. In the engineering coordinate system, the points in the local point cloud subset of the circumferential seam are sorted in circumferential order, and a local circumferential seam plane or local circumferential seam curve is fitted based on these points to obtain preliminary actual circumferential seam geometric parameters. The spatial projection position of the local image subset of the circumferential seam is compared with the preliminary fitting result, and the preliminary geometric parameters are further corrected using the spatial distribution of the joint centerline in the image to reduce local deviations caused by point cloud noise. Through the above joint fitting process, an actual circumferential seam geometric parameter set that reflects both point cloud geometry and image texture is obtained for each circumferential seam, including the actual circumferential seam plane or surface position, the actual circumferential seam centerline position, and the positional deviation relative to the design circumferential seam. The actual circumferential seam geometric parameter sets of all circumferential seams are stored together with the corresponding ring number and circumferential seam number in the local circumferential seam dataset. The design circumferential seam geometric parameters and initial coordinate mapping parameters in the aligned initial dataset are read, and the local circumferential seam dataset is read simultaneously. For each annular gap, the designed annular gap geometry parameters and the actual annular gap geometry parameters are paired, and the spatial deviation between them is calculated. The deviation vector, the initial attitude parameters of the corresponding ring, the initial extrinsic parameters of the laser scanning equipment, and the quality indicators of the local point cloud and image of the annular gap are organized into a self-calibration optimization record. By traversing all rings and all annular gaps, these self-calibration optimization records are combined to form self-calibration optimization data containing multi-ring, multi-gap, and multi-source deviation information. This self-calibration optimization data serves as the direct input for constructing the alignment optimization objective function and performing multi-source spatiotemporal alignment solution.

[0124] The design ring gap geometry parameters, actual ring gap geometry parameters, initial attitude parameters, and initial extrinsic parameters are read from each self-calibration optimization record in the self-calibration optimization data. Using the sensor extrinsic parameter correction and the attitude correction of each ring as variables to be solved, an alignment optimization objective function is constructed with the primary objective of minimizing the spatial deviation between the design ring gap and the actual ring gap. Regularization terms that limit the variation range of sensor extrinsic parameters and the attitude correction range of a single ring are added to the objective function, ensuring that the optimization process corrects long-term drift while avoiding large, non-physical attitude jumps. Through this process, the self-calibration optimization data is converted into an alignment optimization objective description that can be used for numerical solution. Based on the ring number information and shield tunneling sequence in the self-calibration optimization data, a sliding window containing the most recent rings is selected for the tunneling position where the alignment parameters need to be updated. The self-calibration optimization records within this sliding window are read to form a subset of sliding window optimization data. This subset contains the design ring gap geometry parameters, actual ring gap geometry parameters, and corresponding initial attitude and extrinsic parameters for multiple consecutive rings. By using a sliding window to optimize a subset of data instead of the entire data segment, the computational scale can be controlled while ensuring sufficient constraints, and the self-calibration results are more sensitive to the latest construction status. The sliding window optimization subset and alignment optimization objective description are read. Based on the objective and regularization terms given in the alignment optimization objective description, the sensor extrinsic parameter corrections and the attitude corrections of each ring are jointly optimized using an iterative numerical solution method. In each iteration, the spatial deviation of all ring gaps within the window is estimated based on the current parameters, and the parameters are updated until the deviation converges to a preset threshold or the maximum number of iterations is reached. When the optimization iteration converges, the sensor extrinsic parameter correction results and the attitude correction results of each ring within the current sliding window are output, and these are organized into a multi-source alignment parameter set containing the latest extrinsic parameters and ring-level attitude parameters.

[0125] The latest sensor extrinsic parameters and ring-level attitude parameters are read from the multi-source alignment parameter set, while the corresponding 3D point cloud data and ring gap image data for each ring are read from the ring-level data package set. Based on the updated attitude and extrinsic parameters, the point cloud data of each ring is reprojected into the engineering coordinate system. Using the actual ring gap geometry parameters from the local ring gap dataset, the projection results are locally registered and corrected to obtain the high-precision ring center position and high-precision ring gap position for each ring. Based on the high-precision ring gap position, the corresponding ring gap image data is cropped to a spatial range consistent with the precise ring gap position, and the positional relationship of each image pixel or image block in the engineering coordinate system is recorded, forming a ring gap alignment result set containing both point cloud and image information. The multi-source alignment parameter set and alignment parameter history are read. For each ring, the currently obtained attitude correction and sensor extrinsic parameters are merged with the history records to form an alignment parameter history sequence containing the tunneling ring number, timestamp, and corresponding alignment parameters.

[0126] According to one aspect of this application, extracting the attention enhancement feature set can further involve: reading the ring center position, ring joint spatial position, and projection relationship of the ring joint image data in the engineering coordinate system for each ring from the ring joint alignment result set. Based on this information, a mapping rule from image pixel coordinates to engineering coordinates is constructed for each ring joint image. That is, based on the row and column position of the pixel in the image, and the corresponding camera intrinsic parameters and projection parameters, the spatial position of the intersection point of the pixel ray and the area near the ring joint in the engineering coordinate system is calculated. Through this mapping rule, each pixel in the ring joint image data is mapped to an approximate point in the engineering coordinate system, generating an image spatial mapping dataset containing pixel position, engineering coordinate position, and grayscale information. The spatial position of each pixel in the engineering coordinate system is read from the image spatial mapping dataset and combined with the ring center position and tunnel axis direction information of the corresponding ring in the ring joint alignment result set. Using the center of each ring as the origin of polar coordinates, the circumferential direction of the tunnel as the angle axis, and the radial direction as the radius axis, the engineering coordinates of each pixel are converted into polar coordinates with the ring center as the reference. Then, at preset angle and radial sampling intervals, the pixel grayscale values ​​are interpolated and resampled to construct a polar coordinate ring seam image set in a regular grid format. Each image in this polar coordinate ring seam image set uses angle and radius as coordinate axes, and grayscale values ​​as pixel values, providing a structured representation for subsequent ring seam narrowband mask generation and patch division.

[0127] Each polar coordinate image is read from the polar coordinate seam image set, and combined with the spatial location and estimated actual width of the seams recorded in the seam alignment result set. Based on the radial position of the seam in polar coordinates, a radial interval centered at this radial position and with a preset bandwidth is defined as the narrow band region of the seam, while simultaneously limiting the angular range covering the entire circumference, forming a ring-shaped narrow band region mask on the polar coordinate image. For cases with multiple seams, a corresponding narrow band region mask can be generated for each seam, and all masks are stored in the seam narrow band mask data according to the seam number, used to guide the differentiation between seam and non-seam regions during subsequent patch division. Each polar coordinate image and its corresponding narrow band mask are read from the polar coordinate seam image set and the narrow band mask data. Based on preset angular and radial step sizes, the polar coordinate image is divided into several regular small blocks, each of which is a polar coordinate patch unit. Based on the overlap between the pixel position within each patch unit and the narrow band mask of the annular seam, patches mainly located within the narrow band region are marked as annular seam patches, patches located at both ends of the narrow band region and near the corners of the segment are marked as corner patches, and patches far from the narrow band region and without corners are marked as background patches. Through this partitioning process, a set of annular seam patches, a set of corner patches, and a set of background patches are obtained for each ring, and the index relationship between each patch and the original feature map is recorded.

[0128] The feature regions corresponding to the circumferential seam patch set are read from the mid-to-high-level feature map output by the visual backbone network, and the index relationship between the circumferential seam patch set and the feature map is also read. For each circumferential seam patch, the corresponding feature sub-block is extracted from the feature map, and a one-dimensional convolution or difference operator is applied in the radial direction to enhance the gray-level and height abrupt response along the radial direction, obtaining the misalignment response feature reflecting the local misalignment height change. The misalignment response features of all circumferential seam patches are summarized by patch number to form the misalignment response feature branch output for subsequent weight calculation and feature fusion. Continuing to utilize the mid-to-high-level feature map and the circumferential seam patch set, one-dimensional convolution, deformable convolution, or strip convolution kernels are applied in the circumferential direction to the feature sub-block corresponding to each circumferential seam patch to extract the seam texture features and bolt arrangement features extending along the circumferential direction. By performing local statistical aggregation on the convolution output, seam texture features reflecting seam connectivity, seam integrity, and seam texture anomalies are obtained. The seam texture features of all circumferential seam patches are summarized by patch number to form the seam texture feature branch output. Feature regions corresponding to the corner patch set are read from the mid-to-high-rise feature map. For each feature sub-block corresponding to a corner patch, multi-scale convolution kernels and high-frequency enhancement operators are used to extract high-frequency features such as local step edges, sharp corners, and missing concrete edges and corners. By splicing and weighting the outputs of convolution at different scales, high-frequency corner features sensitive to corner misalignment and local damage are obtained. The high-frequency corner features of all corner patches are summarized according to patch number to form the output of the corner high-frequency feature branch.

[0129] The multi-branch feature representation for each patch is extracted from the misalignment response feature, seam texture feature, and corner high-frequency feature. For each patch, the statistics of each branch feature are calculated through operations such as global average pooling and max pooling, forming a feature statistical vector reflecting the importance of the patch in terms of misalignment height, seam texture, and corner high frequency. This feature statistical vector is input into a lightweight weight mapping network or normalization function to generate patch-level weight coefficients for the three branches, and the weight coefficients are normalized to ensure that their summation within each patch is a constant. Through the above process, adaptive patch-level weights are obtained for each patch. These weights reflect the relative importance of the three feature branches of misalignment response, seam texture, and corner high frequency under different spatial locations and structural features. From the patch-level weights and multi-branch features, the feature representation and corresponding weights of each patch on the three branches are extracted. For each patch, the misalignment response feature, seam texture feature, and corner high frequency feature are weighted and summed according to the patch-level weights to obtain the fused feature representation of the patch. The fused features of all patches are written back to their corresponding positions in the mid-to-high-level feature maps according to their spatial indices in the original feature maps. For overlapping areas, the features are normalized and superimposed based on the patch coverage count, thus forming a globally consistent attention-enhanced feature set. This attention-enhanced feature set has a higher response in the narrow band region of the annular seam and the corner region of the segment, and a higher significance for small target regions related to misalignment.

[0130] The attention-enhanced feature set is input into the object detection head network to detect and locate potential misalignment regions in and around the ring seam region. The output is a misalignment candidate result set containing image patches of candidate misalignment regions, their positions in the engineering coordinate system, and corresponding visual saliency scores. This misalignment candidate result set retains both the attention-enhanced visual feature information and maintains a one-to-one correspondence with the geometric positions in the ring seam alignment result set, providing joint visual and spatial input for subsequent geometric decomposition-based classification of intra-ring and inter-ring misalignments.

[0131] In one embodiment of this application, generating a misalignment classification result set may also involve: reading the ring center position, ring-level attitude information, and ring seam spatial position from the ring seam alignment result set, while simultaneously reading the misalignment candidate region position and corresponding visual saliency score from the misalignment candidate result set, associating each misalignment candidate region with its corresponding ring seam segment, and constructing a misalignment candidate association dataset containing ring number, ring seam number, candidate region spatial coordinates, and saliency score.

[0132] The ring center position and ring-level attitude information of each ring are read from the ring joint alignment result set. The spatial position and corresponding ring joint number of each misalignment candidate region in the engineering coordinate system are read from the misalignment candidate association dataset. Taking the ring center position of each ring as the origin, the axial basis vector along the tunnel axis, the radial basis vector pointing to the outside of the lining, and the circumferential basis vector along the circumferential direction are determined based on the ring-level attitude information, and a local ring-level coordinate system is established for each ring. The spatial position of each misalignment candidate region is represented in the local ring-level coordinate system of its ring, and the coordinates of the candidate position in the axial, radial, and circumferential directions are obtained. The definition information of this local coordinate system is recorded as local ring-level coordinate system description data. The ring joint spatial position and local coordinate system parameters of each misalignment candidate position are read from the ring joint alignment result set and the local ring-level coordinate system description data. At the same time, the aligned point cloud data of the corresponding ring is read from the ring-level data set or the aligned point cloud data. Centered on the candidate location, a sampling line is established along the normal direction of the seam in the local ring-level coordinate system, passing through the seam. Point cloud data are selected within a preset range on both sides of the sampling line to form a left local point cloud subset on one side of the seam and a right local point cloud subset on the other side. In this way, local aligned point cloud data corresponding to the local structures on both sides of the seam is obtained at each misalignment candidate location. Local point cloud coordinates are read from the left and right local point cloud subsets and represented in the corresponding local ring-level coordinate system. For the left local point cloud subset, a least-squares fitting method is used to fit a local plane or local surface, so that the fitted surface can approximate the spatial distribution of the left point cloud as closely as possible. The same fitting process is performed on the right local point cloud subset to obtain the right local fitted surface. After fitting, local geometric parameter sets describing the left and right geometric surfaces are obtained, including the normal direction and position offset information of the surface in the local coordinate system.

[0133] The geometric descriptions of the left and right locally fitted surfaces in the local ring-level coordinate system are read from the local geometric parameter set, and combined with the candidate position coordinates recorded in the local ring-level coordinate system description data. In the local coordinate system, the projection point of the candidate misalignment position near the joint center is used as the reference point. The coordinates of the corresponding point on the left and right locally fitted surfaces are calculated respectively. The coordinates of the corresponding point on the right are subtracted from the coordinates of the corresponding point on the left to obtain the misalignment geometric difference vector describing the height and position difference on both sides of the joint. This misalignment geometric difference vector fully reflects the relative displacement of the structures on both sides of the joint in the local coordinate system at the candidate position. The local coordinate axis direction information is read from the misalignment geometric difference vector and the local ring-level coordinate system description data. The misalignment geometric difference vector is decomposed on the axial basis vector, radial basis vector, and circumferential basis vector of the local ring-level coordinate system to obtain the axial component of the misalignment along the tunnel axis, the radial component of the misalignment along the lining radius, and the circumferential component of the misalignment along the circumferential direction. For each candidate misalignment location, the three-dimensional components are combined with information such as the ring number, ring seam number, and spatial location of the candidate location to form a misalignment geometric vector record for that location. By traversing all candidate misalignment locations, all misalignment geometric vector records are summarized to form a misalignment geometric vector set.

[0134] The axial, radial, and circumferential components of each candidate misalignment location are read from the geometric vector set. Simultaneously, the saliency score and local attention response intensity of the corresponding candidate location are read from the misalignment candidate association dataset. The three-dimensional components of the misalignment geometric vector are paired with the saliency score and attention response intensity according to the candidate location, constructing a misalignment feature combination data containing geometric components and visual features. This provides an information foundation that balances geometric information and visual saliency for subsequent classification logic. The axial, radial, and circumferential components of the misalignment, along with the corresponding saliency features, are read from the misalignment feature combination data for each candidate location. Based on engineering experience and statistical analysis results, multiple sets of judgment thresholds and proportional relationships are set for the axial, radial, and circumferential components of the misalignment. When the axial component is greater than the radial and circumferential components, the candidate location is classified as primarily an inter-ring misalignment; when the radial or circumferential component is greater than the axial component, the candidate location is classified as primarily an intra-ring misalignment; when the three-dimensional components are similar or near the threshold, the candidate location is classified as a composite misalignment. In the above rule, by weighting the saliency score and attention response intensity, candidate positions where the geometric components are consistent with the visual features are given priority, while anomalies where the geometric components and visual features clearly contradict each other are weakened, thereby improving the stability and robustness of the classification results. By applying this classification rule to all misalignment candidate positions, the misalignment type of each candidate position is determined as intra-ring misalignment, inter-ring misalignment, or composite misalignment.

[0135] Based on the classification results, the misalignment type, corresponding dominant direction (e.g., axial or radial component as the dominant component), and the three-dimensional components and comprehensive misalignment amount in the misalignment geometric vector of each candidate misalignment location are recorded together to form a misalignment classification record for that location. By traversing all candidate locations, all misalignment classification records are summarized to generate a misalignment classification result set containing candidate location identifiers, ring numbers, ring seam numbers, misalignment types, dominant directions, and misalignment amount values. The misalignment classification records of each ring are read from the misalignment classification result set, while the geometric information of the corresponding ring is read from the ring seam alignment result set, and the local response intensity related to the misalignment location is read from the attention enhancement feature set. For each ring, the number of intra-ring and inter-ring misalignments, the maximum misalignment amount, the average misalignment amount, and the circumferential distribution of misalignments at all candidate locations within it are statistically analyzed. These statistical indicators are correlated with the ring number, ring seam number, and local attention response intensity to form ring-level misalignment analysis entries used to describe the quality status and misalignment distribution characteristics of a single ring. By summarizing the ring-level misalignment analysis entries for all rings, a structured result of the overall tunnel segment misalignment analysis is generated. This result can be directly used to generate a quality assessment report, or it can be used as training data and input features for a dynamic threshold prediction model.

[0136] In summary, a method for detecting segment misalignment in shield tunnels includes: constructing a multi-source ring-level data set indexed by ring number; building a self-calibration optimization model using the spatial deviation between the actual ring joint and the designed ring joint, and achieving high-precision spatiotemporal alignment by jointly solving the sensor extrinsic parameters and ring-level attitude correction through a sliding window; mapping the alignment data to the polar coordinate domain, and extracting attention enhancement features using a parallel patch perception network (HJ-PPA) with three branches including misalignment response, joint texture, and high-frequency corners; and combining geometric information and visual features to generate structured results including intra-ring and inter-ring misalignment classifications through geometric vector decomposition in the local ring-level coordinate system.

[0137] This invention addresses the problems of accumulated error and alignment drift by abandoning traditional static calibration and global ICP methods and introducing a multi-source calibration spatiotemporal alignment technique based on annular joint constraints. By using the inherent strong geometric feature of the annular joint in the tunnel as a constraint, a joint optimization objective function incorporating sensor extrinsic parameters and annular-level attitude corrections is constructed and solved iteratively online using a sliding window mechanism. This enables the system to dynamically sense and compensate for physical deformation caused by vibration, forcibly pulling the measurement data back to the precise design baseline and ensuring millimeter-level geometric measurement accuracy. To address the problem of micro-feature overload, a parallel patch-aware attention mechanism in the polar coordinate domain is employed. By mapping the image to a polar coordinate system, curved joints are straightened into regular strips, and dedicated radial (elevation abrupt changes), circumferential (texture continuity), and corner (high-frequency edges) parallel convolution branches are designed. Utilizing structural priors, computational resources are precisely focused on the narrow joint bands, enhancing the signal-to-noise ratio of micro-misalignment features and making them stand out in complex concrete backgrounds. To address the difficulty of decoupling misalignment types, a geometric vector decomposition model in a local annular coordinate system is established. Instead of outputting a single scalar value, the misalignment is decomposed into three orthogonal components: axial, radial, and circumferential. Classification rules are then established based on visual saliency scores. This allows for a precise physical distinction between inter-ring misalignment (axially dominant) and intra-ring misalignment (radial / circumferential dominant), providing clear quantitative data for construction personnel to adjust the tunnel boring machine's attitude or optimize assembly processes.

[0138] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A segment misalignment detection method for a shield tunnel, characterized by, The method comprises the following steps: acquiring an original multi-source data set containing three-dimensional point cloud data and ring seam image data, and constructing a ring level data packet set indexed by ring number according to pre-stored propulsion encoder information; extracting ring seam local features from the ring level data packet set, constructing self-calibration optimization data using the spatial deviation between the actual ring seam and the design ring seam, jointly solving the sensor extrinsic parameter and the ring level attitude correction amount, and generating a ring seam alignment result set containing accurate ring seam spatial positions; mapping the ring seam alignment result set to the polar coordinate domain, and extracting an attention-enhanced feature set through a pre-configured parallel patch perception network; combining the geometric information in the ring seam alignment result set with the visual features in the attention-enhanced feature set to calculate misalignment geometric vectors decomposed into multi-directional components, and generating a misalignment classification result set; extracting the attention-enhanced feature set, comprising: using the accurate ring center position in the ring seam alignment result set as the polar coordinate origin, mapping the image data in the ring seam alignment result set into a polar coordinate ring seam image set represented by ring coordinate and radial coordinate; generating a ring seam narrowband area mask covering the joint area according to the radial distribution of the accurate ring seam spatial position in the ring seam alignment result set in the polar coordinate; dividing the polar coordinate ring seam image set into a predetermined number of regular image blocks, classifying and constructing ring seam patch sets, corner point patch sets located at the corner of the pipe piece, and background patch sets according to the spatial overlap relationship between each regular image block and the ring seam narrowband area mask; taking them as the input data of the parallel patch perception network to obtain the attention-enhanced feature set.

2. The method of claim 1, wherein, jointly solving the sensor extrinsic parameter and the ring level attitude correction amount, comprising: reading the design ring seam plane and the actual ring seam point cloud subset contained in the self-calibration optimization data; taking the minimization of the sum of squares of the spatial distances of each point in the actual ring seam point cloud subset to the corresponding design ring seam plane after the pose transformation as the main objective, combining the regularization term for limiting the change amplitude of the sensor extrinsic parameter and the numerical amplitude of the ring level attitude correction amount, and constructing an alignment optimization objective function; based on the alignment optimization objective function, simultaneously solving the sensor extrinsic parameter and the ring level attitude correction amount of each ring through a numerical iteration algorithm.

3. The method of claim 2, wherein, simultaneously solving the sensor extrinsic parameter and the ring level attitude correction amount of each ring through a numerical iteration algorithm, comprising: selecting continuous predetermined ring data from the ring level data packet set according to the pre-stored shield tunneling sequence to construct a sliding window optimization data subset; substituting the sliding window optimization data subset into the alignment optimization objective function to perform iterative optimization solving until the alignment optimization objective function value converges or the preset iteration number is reached, outputting the multi-source alignment parameter set corresponding to the current window, wherein the multi-source alignment parameter set comprises the sensor extrinsic parameter and the ring level attitude correction amount of each ring; updating the pose information of each ring in the ring level data packet set using the multi-source alignment parameter set, and moving the sliding window forward with the shield tunneling, and performing incremental alignment solving on the subsequently added ring level data packet.

4. The method of claim 1, wherein, constructing self-calibration optimization data, comprising: performing normal mutation detection on the three-dimensional point cloud data in the ring level data packet set to extract a ring seam local point cloud subset containing ring seam edge features; Performing joint texture recognition on the ring seam image data in the ring level data packet set, and extracting a ring seam local image subset containing the joint center line feature; Fitting a local geometric surface based on the ring seam local point cloud subset, and performing center line constraint correction on the local geometric surface by using the texture position in the spatial projection of the ring seam local image subset, to generate an actual ring seam geometric parameter set; Spatially pairing the actual ring seam geometric parameter set with the pre-set designed ring seam geometric parameter, calculating the geometric deviation therebetween, and constructing a self-calibration optimization data containing the geometric deviation.

5. The method of claim 1, wherein, Extracting the attention-enhanced feature set also includes performing parallel feature extraction on each patch set, specifically: For the ring seam patch set, a one-dimensional convolution kernel along the radial direction is used to extract a misalignment response feature reflecting height mutation, and a convolution kernel extending along the ring direction is used to extract a joint texture feature reflecting joint connectivity; For the corner point patch set, a multi-scale convolution kernel and a high-frequency enhancement operator are used to extract a corner high-frequency feature reflecting the corner step edge; The misalignment response feature, the joint texture feature and the corner high-frequency feature are used as basic feature components for generating the attention-enhanced feature set.

6. The method of claim 5, wherein, Extracting the attention-enhanced feature set also includes calculating patch-level weights, specifically: Performing a global pooling operation on the misalignment response feature, the joint texture feature and the corner high-frequency feature respectively, to obtain a feature statistical vector reflecting the intensity of each branch feature; The feature statistical vector is input into a pre-set weight mapping network to calculate the attention response value corresponding to each branch, and normalization processing is performed on the attention response value to generate a patch-level weight coefficient set satisfying the local constraint condition.

7. The method of claim 1, wherein, Calculating a misalignment geometric vector decomposed into multiple directional components, including: Using the ring center position and the ring level attitude information in the ring seam alignment result set, a local ring level coordinate system containing axial, radial and ring direction basis vectors is constructed for each misalignment detection position; In the local ring level coordinate system, the spatial displacement between the local fitting surfaces on both sides of the joint is calculated to obtain a misalignment geometric difference vector; The misalignment geometric difference vector is projected onto the basis vector direction of the local ring level coordinate system to generate a misalignment geometric vector set containing a misalignment axial component, a misalignment radial component and a misalignment ring component.

8. The method of claim 7, wherein, Generating a misalignment classification result set, including: When the misalignment axial component in the misalignment geometric vector set is greater than the misalignment radial component and the misalignment ring component, the corresponding detection position is determined as an inter-ring misalignment type; When the misalignment radial component or the misalignment ring component in the misalignment geometric vector set is greater than the misalignment axial component, the corresponding detection position is determined as an intra-ring misalignment type; The confidence of the determination result is weighted by combining the local response intensity in the attention-enhanced feature set, and the misalignment classification result set containing the misalignment type, the dominant direction and the misalignment quantity value is output.

9. The method of claim 4, wherein, Extracting a ring seam local point cloud subset containing ring seam edge features, including: A local geometric neighborhood is constructed for each three-dimensional point in the ring level data packet set, and the local normal vector of the three-dimensional point is calculated by using the spatial distribution in the local geometric neighborhood; The angle change value between the local normal vectors of adjacent three-dimensional points is calculated, and the region with an angle change value greater than a pre-set normal mutation threshold is marked as a ring seam local point cloud subset; Wherein, the normal mutation threshold is dynamically generated based on the statistical mean and standard deviation of the local point cloud normal angle distribution, or is set to a fixed empirical value between 15 degrees and 30 degrees.

Images