Method and system for detecting vertical floating of segment in shield tunnel assembly stage

Abstract

This invention discloses a method and system for detecting the vertical floating of tunnel segments during the assembly stage of a shield tunnel, belonging to the field of tunnel engineering monitoring. The method acquires laser point clouds and surface images of the inner surface of the tunnel segments, constructing a set of ring-position data slices indexed by ring number; performs refined cross-sectional fitting on the slice set, taking into account local deformation of the segments, and extracts the measured center of each ring through physical rigid body separation or statistical joint optimization; performs adaptive parametric interpolation based on fitting sensitivity features to generate the measured tunnel centerline; constructs a self-calibrating optimization objective function containing multi-source constraints, and uses a sliding window to incrementally update sensor extrinsic parameters and ring-by-ring attitude; establishes a unified coordinate system with multi-directional components, decomposes spatial deviations, and determines the segment floating mode. This invention solves the coupling problem between rigid body displacement and elastic deformation, achieving high-precision, drift-resistant closed-loop detection of segment floating.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering monitoring technology, and in particular to a method and system for detecting the vertical floating of tunnel segments during the assembly stage of a shield tunnel. Background Technology

[0002] Segment assembly in shield tunneling is a crucial part of tunnel construction, and detecting the vertical float of the segments helps ensure the quality of the tunnel's axis and structural safety. As an important geometric parameter for assessing construction quality, the float directly reflects the effectiveness of synchronous grouting and the level of segment attitude control. In the underground construction environment, obtaining the three-dimensional spatial morphology of the assembled segments through high-precision measurement methods helps guide the adjustment of shield machine excavation parameters, prevent segment misalignment and cracking, and achieve accurate fitting of the tunnel's designed alignment.

[0003] Currently, segment inspection technology based on 3D laser scanning can be used to replace traditional total station single-point measurements. The mainstream method uses non-contact scanning to acquire point clouds of the tunnel inner wall, projects the 3D point cloud slices along the tunnel axis onto a 2D plane, and uses the least squares method to fit the cross-sectional point cloud into a standard elliptical model. The geometric center of the fitted ellipse is then extracted to calculate the spatial position deviation of the segment. This type of method utilizes the full coverage of point cloud data, and under ideal conditions, it can quickly achieve full-section morphological reconstruction and center positioning.

[0004] However, existing technologies suffer from the problem of coupled calculation of rigid body displacement and elastic deformation when dealing with complex deformations during the assembly stage. Therefore, further research and innovation are needed to solve the aforementioned problems in existing technologies. Summary of the Invention

[0005] Purpose of the invention: In view of the above-mentioned problems of the prior art, this application provides a method and system for detecting the vertical floating of tunnel segments during the assembly stage of a shield tunnel.

[0006] Technical solution: Firstly, a method for detecting the vertical floating of tunnel segments during the shield tunnel assembly stage, comprising:

[0007] The laser point cloud and appearance image of the inner surface of the tunnel segment are obtained, and combined with the shield tunneling information, a set of ring position data slices indexed by ring number is constructed.

[0008] A refined cross-sectional fitting that takes into account the local deformation of the pipe segments is performed on the annular data slice set to extract the set of measured center points of each annular cross section;

[0009] Based on the fitting sensitivity characteristics of the measured center point set, adaptive parametric interpolation is performed to generate the measured tunnel centerline;

[0010] The self-calibration optimization objective function containing multi-source constraints is invoked, and the extrinsic parameter offset and ring-by-ring attitude correction of the ring position data slice set are incrementally updated using a sliding window.

[0011] A multi-dimensional unified coordinate system is established, and the spatial deviation of the measured tunnel centerline relative to the design axis is decomposed into a set of structured displacement components. Based on this set of structured displacement components, the segment floating mode is determined.

[0012] Secondly, a segment vertical floating detection system during the shield tunnel assembly stage includes:

[0013] The data acquisition unit is used to acquire laser point cloud and appearance images of the inner surface of the tunnel segments, as well as shield tunneling status data;

[0014] Memory, used to store computer programs and pre-configured design parameters;

[0015] A processor, connected to a data acquisition unit and a memory, is used to execute a computer program to implement the method as described in any of the first aspects.

[0016] Beneficial effects: This invention solves the coupling problem between rigid body displacement and elastic deformation, achieving high-precision, drift-resistant closed-loop detection of segment floating. The related technical effects will be described in detail below with reference to specific embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for detecting the vertical floating of tunnel segments during the assembly stage of a shield tunnel, as provided in an embodiment of this application.

[0018] Figure 2 This is a flowchart illustrating an example of constructing a ring position data slice set indexed by ring number, provided in an embodiment of this application.

[0019] Figure 3 This is a flowchart illustrating a cross-sectional projection and arc segment division, provided as an embodiment of this application.

[0020] Figure 4 This is a flowchart illustrating an example of adaptive parametric interpolation provided in an embodiment of this application.

[0021] Figure 5 This is a flowchart illustrating an example of establishing a multi-component unified coordinate system and decomposing the spatial deviation of the measured tunnel centerline relative to the design axis into a set of structured displacement components, as provided in an embodiment of this application. Detailed Implementation

[0022] 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.

[0023] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] To address the aforementioned issues, the applicant conducted in-depth searches and analyses, and discovered:

[0025] Correspondingly, the segment ring is assembled from multiple prefabricated components and is not an ideal continuum. Its actual shape includes both the rigid body upward displacement of the entire ring and the local elastic deformation at the joints (such as misalignment and opening). Traditional overall fitting methods forcibly smooth the local joint deformations into elliptical geometric parameters, which cannot separate the actual rigid body center displacement, resulting in systematic biases in the estimation of upward displacement.

[0026] Furthermore, existing centerline fitting and calibration methods lack consideration for physical uncertainty and have not established a closed-loop calibration mechanism that integrates visual texture confidence and geometric fitting covariance. This results in sensor extrinsic drift accumulating with tunneling mileage, making it difficult to meet the requirements for long-distance, high-precision detection.

[0027] To solve these problems, combined with Figures 1 to 5 The present invention will be specifically described through the following embodiments.

[0028] On the one hand, an exemplary scheme for detecting the vertical floating of tunnel segments during the shield tunnel assembly stage is provided. This is an automatic monitoring system that integrates multi-source data and performs intelligent analysis, used to solve the problems of discontinuous and low-precision segment floating detection, and difficulty in distinguishing between true displacement and local deformation. Specifically, it includes:

[0029] Step 101: Obtain the laser point cloud and appearance image of the inner surface of the tunnel segment, and combine it with the shield tunneling information to construct a set of ring position data slices indexed by ring number.

[0030] In this step, the ring location data slice set is a structured data organization form, using the segment ring number as a unique index key to encapsulate heterogeneous multi-source data. Specifically, the system uses a laser scanner installed on the tunnel boring machine trolley to acquire a high-density 3D point cloud of the tunnel inner wall, which can characterize the geometry of the segment; and uses an industrial camera to acquire a high-resolution appearance image of the inner surface of the segment, which includes visual features such as joint texture and misaligned edges.

[0031] Simultaneously, the system receives real-time data on the tunnel boring machine's (TBM) tunneling status, including mileage, speed, and attitude angle. The process of constructing the slice set involves spatiotemporal alignment of data; that is, based on the physical spatial range of each tunnel segment ring, the point cloud, images, and corresponding tunneling parameters belonging to that ring are associated and stored. The ring-indexed structure provides a standardized data foundation for subsequent independent ring-by-ring calculations and sliding-window recursive optimization.

[0032] Step 102: Perform refined cross-sectional fitting on the ring data slice set to take into account the local deformation of the segments, and extract the set of measured center points of each ring cross section.

[0033] Furthermore, the refined cross-sectional fitting considering local segment deformation refers to an algorithm that overcomes the limitations of traditional least-squares elliptic fitting. In actual engineering, after segment assembly, not only will overall rigid body displacement occur (i.e., floating or sinking), but local elastic deformations such as misalignment and opening will also occur at the joints. Traditional overall fitting will incorrectly smooth local deformations to the center point coordinates, leading to errors in the estimation of floating. The method used in this step can distinguish and handle local deformations.

[0034] In practical implementation, a rigid body separation method based on a physical model can be used, treating the tunnel segments as rigid bodies and separating the rigid body displacements of each segment to determine the center; alternatively, a shared center optimization method based on a statistical model can be used, introducing radial offset variables to absorb local errors. Regardless of the specific path used, the goal is to eliminate the interference of joint deformation on center positioning and extract a set of measured center points that can reflect the overall spatial position of the tunnel segment ring. These center points constitute a discrete skeleton describing the shape of the tunnel axis.

[0035] Step 103: Based on the fitting sensitivity features of the measured center point set, perform adaptive parametric interpolation to generate the measured tunnel centerline.

[0036] In this step, the fitting sensitivity feature refers to a quantitative index that reflects the degree of drastic local geometric changes or physical uplift anomalies of the tunnel axis. Spline interpolation typically uses uniform parameterization or chord parameterization, which performs well in straight sections of the tunnel, but is prone to fitting bias in areas of abrupt uplift changes or small-radius curves due to sparse nodes. This step introduces an adaptive mechanism to calculate sensitivity indices including geometric curvature and uplift gradient. In areas of high sensitivity, i.e., where the axis curvature is large or the uplift changes rapidly, the interpolation parameter nodes are automatically densified; in areas of low sensitivity, they are kept sparse. Based on this non-uniformly distributed parameter sequence, a quadratic B-spline curve is solved to generate a smooth, continuous, and closely approximate three-dimensional spatial curve that closely approximates the measured centerline of the tunnel.

[0037] Step 104: Call the self-calibration optimization objective function containing multi-source constraints, and use a sliding window to incrementally update the extrinsic parameter offset and ring-by-ring attitude correction of the ring position data slice set.

[0038] Alternatively, the measured center point set and / or the measured tunnel centerline can be corrected using the updated extrinsic parameter offset and ring-by-ring attitude correction.

[0039] Specifically, due to the harsh environment of tunnel boring machine (TBM) construction, sensors develop installation errors or zero-point drift after long-term operation, and relying on initial calibration parameters leads to the accumulation of detection errors over time. Based on this, an online self-calibration mechanism was established. The self-calibration optimization objective function is a mathematical model that combines geometric residual terms with regularization constraint terms.

[0040] Optionally, the function also incorporates visual confidence and geometric uncertainty, using a neural network to score image quality and dynamically adjust the weights of each data point. The calculation process employs a sliding window strategy, selecting only a few recently constructed segments (e.g., the last 10 segments) at a time to construct the optimization problem. The objective function is minimized to solve for the sensor's extrinsic parameter offset relative to the tunnel boring machine and the attitude correction for each segment. The corrections are written back to the slice set in real time, achieving dynamic zero-point correction of the measurement system and suppressing systematic drift in long-distance detection.

[0041] Step 105: Establish a multi-component unified coordinate system, decompose the spatial deviation of the measured tunnel centerline relative to the design axis into a set of structured displacement components, and determine the segment floating mode based on this (structured displacement components).

[0042] Accordingly, this step establishes a local orthogonal coordinate system that varies with the tunnel axis. This coordinate system includes the axial direction tangential to the axis, the radial direction pointing upwards perpendicular to the axis, and the circumferential direction laterally. Projecting the spatial deviation vector between the measured centerline and the design axis at the same mileage onto these three axes yields a structured set of displacement components. Based on the magnitudes of the three components and their combinations, the segment uplift pattern can be intelligently determined. For example, if only the radial component is large, it is determined to be uniform uplift; if both the radial and axial components are large and correlated, it is determined to be uplift caused by longitudinal displacement; if the circumferential component is abnormal or accompanied by tilting, it is determined to be rotational warping. This structured analytical result provides construction personnel with a decision-making basis for fault diagnosis.

[0043] On the other hand, this paper describes alternative implementation methods for data preprocessing and the construction of ring-shaped data slice sets, particularly how to utilize tunneling pulse signals and software interpolation techniques to achieve high-precision spatiotemporal alignment of multi-source data, and how to clean and estimate the normal vectors of the original point clouds to provide basic input for subsequent fitting. Specifically, this approach includes:

[0044] Step 201: Respond to the pulse signal of the shield tunneling system, determine the start and end times of tunneling of a single ring segment, and align the timestamps of the laser point cloud and the appearance image accordingly (start and end times of tunneling) to divide the data units belonging to the corresponding ring number.

[0045] In this application, the tunnel boring machine's propulsion system triggers a pulse signal every fixed distance advanced (e.g., 1 mm or 10 mm). By counting and accumulating these pulses, the real-time mileage of the tunnel boring machine can be accurately determined. Since the standard segment width (usually 2 meters) is fixed, the system can determine whether one ring of tunneling has been completed based on the mileage increment. Specifically, the system records the system timestamp at the start and end of the k-th ring of tunneling, and these two times constitute a closed time window. All laser point cloud frames and apparent image frames acquired within this time window are logically attributed to the k-th ring.

[0046] Furthermore, considering the potential clock asynchrony or transmission delay between the laser scanner, camera, and tunneling control system, this step may employ software linear interpolation for timestamp correction. Assume the scanner's sampling time is t, and the tunnel boring machine's position P at this time... _raw The original position is recorded at time t. If there is a known system delay Δt, and the current advance speed of the tunnel boring machine is v... _scan Then the corrected accurate position P _aligned Calculated using the following formula:

[0047] P _aligned =P _raw +v _scan ×Δt;

[0048] Among them, P _aligned P represents the aligned position coordinates. _raw v represents the original record coordinates. _scan The vector represents the moving speed of the scanning platform or tunnel boring machine, and Δt represents the timestamp deviation. This method eliminates mileage matching errors caused by hardware latency, ensuring that each point cloud and each frame of imagery can be mapped to its physical location on the tunnel axis.

[0049] Step 202: Perform density-based clustering on the laser point cloud in the data unit to filter out outliers and noise, and use principal component analysis to calculate the local normal vector of each point cloud to generate a ring-shaped data slice set containing geometric normal information and texture information.

[0050] Optionally, the original point cloud often contains outliers caused by dust, water mist, or movement of construction workers. To ensure fitting accuracy, cleaning is necessary. This step employs a density-based clustering algorithm (DBSCAN): for any point in the point cloud, the number of points within its preset radius is counted. If the number of points is below a preset threshold, it is marked as noise and filtered out; if the number of points meets the requirement, it is retained and considered a valid surface point. This method preserves the main structure of the tunnel segment while eliminating sparse, suspended noise.

[0051] Based on this, the local geometric properties of each point need to be calculated to support subsequent feature extraction, such as seam recognition. Specifically, for each retained point p... _i The system searches for its k nearest neighbors, for example, k can be 10 to 20, and constructs a local covariance matrix. Principal component analysis (PCA) is then performed on this matrix, and the eigenvector corresponding to its smallest eigenvalue is the local normal vector of that point. At this point, each data point in the ring-shaped data slice set not only contains three-dimensional coordinates (x, y, z) and reflection intensity I, but also includes normal vector information (nx, ny, nz).

[0052] Furthermore, the specific data structure of the ring position data slice set is defined. Each slice contains the following fields:

[0053] ring_id, an integer, represents the ring index;

[0054] timestamp, a floating-point number, represents the time of data collection.

[0055] point _c loud, an N×6 floating-point array, stores coordinates and normals;

[0056] images, a List structure, stores multiple corresponding appearance images;

[0057] design _paramsThe structure stores the design center coordinates and the tangent of the design axis of the ring.

[0058] In addition, the slice set must also include the upward tilt angle φ, which is defined as the rotational component of the cross-section about the tunnel design axis (Z-axis), and its calculation formula is as follows:

[0059] φ=arctan2(R _21 R _11 );

[0060] Among them, R _21 and R _11 This represents the corresponding element in the rotation matrix R output by the tunnel boring machine's attitude measurement unit, where arctan2 is the four-quadrant arctangent function. A local coordinate system is assumed, with the Z-axis along the tunnel's extension direction and the Y-axis vertically upwards. This data structure definition allows subsequent steps to unambiguously retrieve the required information.

[0061] On the other hand, this paper describes an alternative technical solution for extracting the cross-sectional center based on rigid segment constraints. This solution utilizes the rigidity of the segments to separate deformation and displacement, solving the technical problem that overall fitting cannot distinguish between actual uplift and joint deformation. Accordingly, this method can be implemented using the following steps:

[0062] Step 301: Based on the design center angle range of the tunnel segment, divide the cross-sectional point cloud in the ring position data slice set into a segment point cloud subset corresponding to each tunnel segment.

[0063] In this application, the tunnel segment ring is typically assembled from several precast concrete segments, such as a capping block K, adjacent blocks B, and a standard block A, totaling M segments. During the design phase, each segment has a fixed coverage angle range on its circumference. Let φ be the design starting angle for the m-th segment. _start_m The termination angle is φ _end_m For any point p in the slice set _j Project it onto a polar coordinate system with the design center as the origin, and calculate its polar angle φ. _j If φ _j Falling in the interval [φ _start_m , φ _end_m If the point is within the range of [m], then it is assigned to the m-th segment point cloud subset P. _km In this way, the disordered point cloud of the entire ring is deconstructed into M independent subsets with physical correspondences.

[0064] Step 302: Perform rigid body registration for each segment point cloud subset and its design reference position, and estimate the independent rigid body transformation parameters of each segment.

[0065] Specifically, it utilizes the physical assumption that a single tunnel segment can be considered a rigid body during assembly and does not undergo deformation. For the m-th tunnel segment, let its design model point cloud be Q. _m ={q _1 , ..., q _Nm The measured subset of the point cloud is P. _km ={p _1 , ..., p _Nm}. Or, in other words, subscript. _Nm N is the number of points in this subset. _m To find a better rotation matrix R _km Translation vector t _km This makes the transformed model point cloud closer to the measured point cloud. Specifically, the singular value decomposition (SVD) algorithm is used for solving the problem.

[0066] Accordingly, the centroid p of the measured point cloud subset is calculated. - And the centroid q of the point cloud in the design model - Next, we construct the cross-covariance matrix H. _km The calculation formula is as follows:

[0067] H _km = (q _j -q - (p) _j -p - ) T ;

[0068] Above, H _km Let N be the cross-covariance matrix of the m-th segment in the k-th ring. _m The summation is performed on all corresponding pairs of points within the subset, with the superscript representing the number of points in the subset. T The corresponding matrix transpose operation, q _j p _j These correspond to the j-th point in the design model point cloud and the measured point cloud subset, respectively. Next, for matrix H... _km Perform singular value decomposition:

[0069] H _km =UΣV T ;

[0070] Where U and V are orthogonal matrices, and Σ is a singular value diagonal matrix.

[0071] Based on this, the optimal rotation matrix R _km Calculated as: R _km =VU T ;

[0072] Optimal translation vector t _km Calculated as: t _km =p- -R _km q - ;

[0073] Through this process, a set of parameters (R) describing the independent attitude of each segment in space can be obtained. _km , t _km ).

[0074] Step 303: Based on the minimum deformation energy criterion, the rigid body displacement component of the whole ring and the elastic deformation component of the joint are separated from the rigid body transformation parameters of each segment. The cross-sectional position is corrected by using the rigid body displacement component of the whole ring and the set of measured center points is extracted.

[0075] Furthermore, the independent orientation of each segment includes the following two components: the overall rigid displacement of the entire ring as the strata move, i.e., actual uplift; and the relative displacement between segments due to compression or misalignment, i.e., elastic deformation.

[0076] Based on this, this step introduces the minimum deformation energy criterion to separate the two, that is, assuming that the segment ring tends to maintain its designed circular topology in the absence of external force, the overall deformation energy should be minimized.

[0077] Step 304: Using the rigid body transformation parameters of the entire ring as optimization variables, call the optimization objective function that minimizes the weighted difference between the rigid body transformation parameters of each segment and the rigid body transformation parameters of the entire ring.

[0078] In this application, the transformation parameters of the complete rigid body to be solved are defined as (R... _k , t _k The constructed optimization objective function J is as follows:

[0079] J= w _m (λ _R ||R _km -R _k || _F 2 +λ _t ||t _km -t _k || 2 );

[0080] in, w represents the summation over all segments (m=1 to M). _m λ is the weight of the m-th segment, which is usually proportional to the number of point clouds contained in that segment; _R and λ _t It is the regularization coefficient of the dimensions of the rotation and translation terms in the equilibrium, and optionally, λ _R Let λ be 1. _t Set to 0.1; ||…|| _FLet ||Frobenius| represent the Frobenius norm, and ||…|| represent the Euclidean norm (L2 norm). This objective function is used to find the average attitude that minimizes the deviation of each individual segment from this average attitude.

[0081] Step 305: Solve the objective function to obtain the whole ring translation vector. Use this (whole ring translation vector) as the actual upward displacement. Calculate the sum of the difference norms between the rigid body transformation parameters of each segment and the rigid body transformation parameters of the whole ring to obtain the cross-sectional deformation index that reflects the degree of joint deformation.

[0082] Alternatively, the optimal integral ring translation vector t can be calculated by minimizing the objective function J, for example, by using the differentiation and zeroing method or the iterative method. _k . t _k The vertical component (the component along the Y-axis) in the figure represents the pure rigid body float after eliminating the interference of joint deformation, and its corresponding center position is the high-precision measured center point C. _k Simultaneously, the residuals of each segment's attitude relative to the entire ring are calculated, and the cross-sectional deformation index D is defined. _k for:

[0083] D _k =(1 / M) (||δR _km -I|| _F +||δt _km ||);

[0084] Wherein, δR _km =R _km R _k T Let δt be the rotation residual matrix. _km =t _km -t _k Let I be the translation residual vector, and let D be the identity matrix. _k The larger the value, the more severe the misalignment or opening of the joints of the ring segment. This index, as a structural health status parameter output, can be used as an input factor for sensitivity analysis in subsequent steps.

[0085] In some scenarios, a specific technical solution for joint optimization of cross-section extraction based on shared center and arc segment offset is provided. This solution utilizes a statistical optimization model to address local misalignment interference in tunnel segments, particularly in scenarios with large non-rigid deformation of tunnel segments or uneven point cloud distribution. This method can provide center estimation including uncertainty information. Specifically, this solution is as follows:

[0086] Step 401: Construct a rotation matrix using the tangential vector of the measured centerline and the collected tilt angle data, calculate the normal vector of the corrected normal plane that takes into account the segment attitude, and project the cross-section point cloud onto the plane determined by the normal vector of the corrected normal plane.

[0087] Alternatively, by using the tangential vector of the design axis at the current ring position and the collected tilt angle data to construct a rotation matrix, the normal vector of the corrected normal plane that takes into account the segment attitude is calculated, and the cross-sectional point cloud is projected onto the plane determined by the normal vector of the corrected normal plane.

[0088] Correspondingly, during the tunnel boring machine's excavation process, rolling and pitching occur, causing the physical normal of the segment cross-section to no longer be parallel to the tunnel's design axis. Directly projecting the point cloud onto the plane perpendicular to the design axis will result in torsional deformation errors. This step acquires the attitude information of the segment in real time. Let τ be the tangential vector of the measured centerline at the k-th ring. _k The tilt angle (Roll angle) is φ _k Construct the rotation matrix R(τ) based on Rodrigues' formula. _k , φ _k Alternatively, in a simplified model, first rotate the Z-axis to be aligned with τ. _k Overlap, then rotate φ around the new Z-axis _k The corrected normal vector n of the normal plane. _k The calculation is as follows:

[0089] n _k =R(φ _k )×τ _k ;

[0090] Wherein, R(φ) _k ) represents the rotation φ about the current tangential axis. _k The rotation transformation matrix; or, in other words, the matrix based on the tilt angle φ. _k The rotation transformation matrix. Next, for each point cloud p within the ring. _i Calculate its value in terms of the normal vector n _k The projection point p' on the defined plane _i =p _i -((p _i -C _approx )•n _k )•n _k Among them, C _approx The approximate center of the ring can be replaced by the design center, and • denotes the vector dot product. This attitude-corrected projection eliminates projection errors, ensuring that subsequent fitting is performed on the actual cross-section of the segment.

[0091] Step 402: Perform multi-source consistency screening on the normal abrupt change features of the cross-sectional point cloud and the seam texture features of the corresponding appearance image. The position that simultaneously satisfies the geometric abrupt change significance and the visual texture significance is determined as the stable arc segment boundary. Based on this (stable arc segment boundary), the projected cross-sectional point cloud is divided into arc segment point sets.

[0092] In other words, a local normal curvature greater than the geometric threshold indicates that the geometric abrupt change is significant, and the gradient magnitude of the corresponding pixel region greater than the visual threshold indicates that the visual texture is significant.

[0093] In this step, accurate identification of the segment joints is crucial. Relying solely on point cloud normal abrupt changes is susceptible to interference from pipes such as water pipes and cables, while relying solely on image texture is easily affected by water stains and oil contamination. Therefore, a multi-source consistency strategy is adopted: for a single point in the point cloud, its local normal curvature κ is calculated. _geom If κ _geom Geometric threshold T _geom If the value is positive, it is marked as a geometric candidate boundary; simultaneously, the point is mapped onto the appearance image, and the gradient magnitude G of the corresponding pixel region is calculated. _vis If G _vis Visual threshold T _vis If a location is marked as a geometric candidate boundary, it is then identified as a true seam boundary. Only when a location is simultaneously marked as both a geometric and visual candidate boundary is it determined to be a true seam boundary. Based on the true boundaries, the entire ring projection point cloud is divided into J independent arc segment point sets.

[0094] In some alternative implementations, if the apparent image is missing or of poor quality, the system automatically downgrades to using only geometric normal change features for boundary screening, but the geometric threshold will be increased accordingly to reduce the false detection rate.

[0095] Step 403: Divide the cross-sectional point cloud into several arc segment point sets, and call the joint optimization objective function containing shared center variables and arc segment local radial offset variables. The shared center variables constrain the geometric position of the entire ring cross-section, and the arc segment local radial offset variables are used to absorb the independent deformation of each arc segment relative to the ideal ellipse.

[0096] Specifically, traditional methods assume the entire ring is a perfect ellipse, but cannot explain the misalignment phenomenon. This step proposes a shared center + local offset model. The shared geometric center of the entire ring is defined as c. _k Simultaneously, a radial offset Δr is defined for each arc segment j. _j This is used to characterize the degree of expansion or contraction of the segment relative to a standard ellipse.

[0097] Furthermore, the joint optimization objective function also includes a regularization constraint term for the elliptical geometry;

[0098] The regularization constraint term is used to constrain the fitted major axis and fitted minor axis of each ring section, keeping them within the preset tolerance range of the designed ellipse axis length, and preventing the ellipse shape from degrading due to the lack of local point cloud.

[0099] To prevent distortion of the ellipse shape when point cloud is missing in certain areas (such as when the arch is obscured by ductwork), a regularization constraint on the ellipse geometry needs to be added to the objective function. The constructed joint optimization objective function F is as follows:

[0100] F(c _k , Δr _j a _k b _k )= ||d(p,c _k a _k b _k )-Δr _j || 2 +λ _shape [(a _k -a _design ) 2 +(b _k -b _design ) 2 ]+λ _smooth ∑ j (Δr _j -Δr _j-1}) 2 ;

[0101] Wherein, d(p, c) _k a _k b _k ) represents the distance from point p to point c. _k Centered on, with semi-major axis a _k The minor semi-axis is b _k The algebraic or geometric distance of the standard ellipse; a _design and b _design λ is the design axis length of the tube segment; _shape Set the shape constraint weights (e.g., 10). 3 ), used to force the fitted ellipse to not deviate too far from the designed shape; λ _smooth To smooth out the weights, the radial offsets of adjacent arc segments are constrained to prevent abrupt changes. In other words, c... _k Let Δr be the coordinates of the shared center of the entire ring to be solved. _j Let a be the radial offset of the j-th arc segment; _k b is the measured semi-major axis to be solved; _k Let Δr be the measured semi-minor axis to be solved. _j-1 Let J be the radial offset of the (j-1)th arc segment, j be the arc segment number of the k-th ring, J be the total number of arc segments in the k-th ring, and p be the identifier of the point cloud point, representing the point set S of the j-th arc segment traversed in the k-th ring. _k_j All point cloud points. While allowing each segment to move freely radially, find the global center that best interprets all data points.

[0102] Step 404: Solve the joint optimization objective function to obtain the measured center point of each ring section. Based on the residual statistical characteristics and Hessian matrix in the optimization process, calculate the measurement covariance matrix that characterizes the uncertainty of the center point position.

[0103] The objective function F is a nonlinear least squares problem, typically solved iteratively using the LM (Levonburg-Marquardt) algorithm. When the iteration converges, the measured center point is obtained. At this point, the uncertainty of this center point must be quantified to provide a confidence basis for subsequent self-calibration.

[0104] At the convergence point of the LM algorithm, the Hessian matrix H of the objective function with respect to all parameters is calculated. According to estimation theory, the covariance matrix Cov(x) of the parameters is approximately equal to the inverse of the Hessian matrix H. -1 With residual variance σ 2 The product of these two products is given by the following formula:

[0105] Cov(x)≈σ 2 ×H -1 ;

[0106] Where, σ 2 This is the unit weight variance estimate. Extracting the 2×2 submatrix corresponding to the center coordinates from Cov(x) yields the measurement covariance matrix Σ. _k Its eigenvalues ​​reflect the positioning uncertainty of the center point along the major and minor axes. For example, a large eigenvalue indicates insufficient point cloud constraint or high noise in that direction. This statistic will be passed as input to the self-calibration module.

[0107] On the other hand, the formula for estimating the central covariance can also be:

[0108] Σ _k ≈σ 2 (H -1 ) _c_k ;

[0109] Where, Σ _k Center point c _k The measurement covariance matrix; σ 2 H is the unit weight variance estimate; H is the Hessian matrix of the objective function at the optimal solution; H -1 The inverse of the Hessian matrix; (H -1 ) _c_k This represents the sub-block in the inverse of the Hessian matrix that corresponds to the center parameter.

[0110] In other scenarios, exemplary solutions are provided for adaptive parametric interpolation methods based on fitting sensitivity features, particularly how to adaptively adjust the interpolation density using physical and geometric indices to address the insufficient fitting accuracy of traditional spline interpolation in abrupt change regions. This solution includes:

[0111] Step 501: Calculate the rate of change of the direction angle of the line connecting adjacent points in the set of measured center points as a geometric curvature index; calculate the change of the vertical deviation of adjacent measured center points relative to the design axis as an upward gradient index; and calculate the spatial distance between each measured center point and the corresponding position of the design axis as a design deviation index.

[0112] Furthermore, three dimensions of sensitivity indicators were defined:

[0113] Geometric curvature index κ _geo This reflects the degree of curvature of the axis in space. For the i-th measured center point C... _i The formula for calculating its geometric curvature index is:

[0114] κ _geo_i =||v _i -v _i-1 || / ||C _i -C _i-1 ||;

[0115] Among them, v _i =(C _i+1 -C _i ) / ||C _i+1 -C _i || represents the unit tangent vector, i is the index of the measured center point, and C _i Let C be the i-th measured center point. _i-1 C _i+1 v represents the (i-1)th and (i+1)th measured center points, respectively. _i Let be the unit tangent vector at the i-th measured center point.

[0116] Float gradient index κ _float This reflects the abrupt changes in the buoyancy of the tunnel segments and is a physical characteristic. Let C... _i The vertical coordinate is z _i The vertical coordinate of the design axis is z. _design_i Then the upward buoyancy Δz _i =z _i -z _design_i The buoyancy gradient is calculated as: κ _float_i =|Δz _i -Δz _i-1 | / L _step Among them, L _step For the ring width, Δz _i-1Let |i| represent the amount of segment rise at position i-1, where |…| represents the absolute value. This index can capture the rise jump caused by sudden changes in grouting pressure.

[0117] Design deviation index d _dev This reflects the cumulative degree of construction errors. The design deviation index value d at the i-th location... _dev_i =||C _i -C _design_i ||;C _design_i Let be the design center point of the design axis at position i. Three indicators quantify the difficulty of fitting from the perspectives of geometric shape, physical change, and construction error, respectively.

[0118] Step 502: The geometric curvature index, the buoyancy gradient index, and the design deviation index are weighted and fused to obtain the comprehensive sensitivity of each measured center point. Based on the comprehensive sensitivity, the parameter increment of the spline interpolation is adjusted in reverse. In the region where the comprehensive sensitivity is higher than the preset threshold, the parameter increment is reduced and the parameter nodes are densified. The quadratic B-spline curve is solved using the adjusted parameter sequence.

[0119] Optionally, the three indicators are normalized to the interval [0, 1], denoted as κ'. _geo , κ' _float ,d' _dev These correspond to the normalized geometric curvature index, the upward gradient index, and the design deviation index, respectively. The overall sensitivity S is calculated. _i =w _g ×κ' _geo +w _f ×κ' _float +w _d ×d' _dev The recommended weight configuration is w. _g =0.2, w _f =0.5, w _d =0.3, emphasizing the influence of the upward gradient. In B-spline interpolation, the parameter increment assigned by traditional chord length parameterization is proportional to the chord length.

[0120] An adaptive parameterization strategy is adopted, and the corrected parameter increment Δu' _i The calculation is as follows:

[0121] Δu' _i =L _i / (1+η×S _i );

[0122] Where η is the sensitivity amplification factor, for example, taking a value from 3 to 5, L _iLet be the chord length between adjacent center points. The formula shows that when the overall sensitivity is high, the denominator increases, and the parameter increment decreases. In B-spline theory, the smaller the parameter increment, the higher the density of control points corresponding to that segment, and the stronger the curve's ability to approximate the data points—that is, the weaker the rigidity and the stronger the flexibility. Conversely, in straight segments where the overall sensitivity is close to 0, the parameter increment is larger, and the curve maintains a high degree of smoothness. By constructing a node vector using this non-uniform parameter sequence and solving for the control vertices of the quadratic B-spline, a measured tunnel centerline with a good fit can be generated.

[0123] Step 503: After obtaining the quadratic B-spline curve in the initial solution, calculate the fitting residuals from each measured center point to the curve, normalize the fitting residuals and add them to the comprehensive sensitivity, and use the updated comprehensive sensitivity to recalculate the parameter increments and solve the spline curve until the fitting residuals converge.

[0124] Specifically, a residual feedback mechanism is introduced. Based on the initial comprehensive sensitivity, the curve C(u) is fitted, and C is calculated for each measured point. _i The shortest Euclidean distance to curve C(u), i.e., the residual r _i Update the overall sensitivity S _new_i =S _old_i +γ×(r _i / r _max ), where γ is the feedback gain, for example 0.5, S _new_i S _old_i This represents the overall sensitivity before and after the update, r _max This represents the maximum residual value. Areas where the fit is not close enough (large residuals) will be considered high-sensitivity regions for further refinement in the next round. Using S... _new_i Recalculate the parameter increments and reconstruct the curve. Repeat the above steps until the rate of change of the root mean square error (RMSE) of the residuals at all points is less than a preset threshold, such as 1%, or the maximum number of iterations is reached. This mechanism helps to achieve a better fit to the curve in terms of local details.

[0125] In other scenarios, an exemplary scheme for multi-source joint sliding window self-calibration is described. Specifically, it utilizes the complementary characteristics of vision and geometry to construct a robust self-calibration mathematical model, addressing the problem of long-distance sensor drift during tunnel boring machine (TBM) construction. Accordingly, this scheme can be:

[0126] Step 601: The optimization variables of the self-calibration optimization objective function include at least the sensor's extrinsic offset vector relative to the tube segment, the ring-by-ring attitude correction amount, and the ring-by-ring float correction amount.

[0127] In this application, the joint state vector X to be optimized is defined as: X = [δe, δT] _k ,δT _k+1 , ..., δu _k ,δu_k+1 ...]; where δe is the extrinsic offset of the sensor (laser scanner) relative to the tunnel boring machine trolley, which usually contains 3 translational components and 3 rotational components, belongs to the se(3) Lie algebra space, is a global variable, and is shared within the sliding window; δT _k δT _k+1 δu represents the attitude correction for the k-th and k+1-th ring segments, used to compensate for errors in the tunnel boring machine's own attitude measurement (such as the guidance system), and is a local variable; _k δu _k+1 This is the upward correction amount for the kth and k+1th rings, which directly affects the final vertical coordinate of the segment.

[0128] Or, in other words, X = [δe, δT] _k ,δT _k+1 , ...,δT _m ,δu _k ,δu _k+1 , ..., δu _m ] T ; m is any terminating index.

[0129] Step 602, the self-calibration optimization objective function includes a Huber robust term to characterize the reprojection error, and a regularization constraint term to constrain the physical rationality of the constraint variables;

[0130] The regularization constraints include extrinsic energy constraints that limit the drift amplitude of extrinsic parameters, smoothness constraints that limit abrupt changes in the attitude of adjacent segments, and upward energy constraints that limit the magnitude of upward correction and its rate of change.

[0131] In other words, the self-calibration optimization objective function is solved to obtain the optimized extrinsic offset vector, the ring-by-ring attitude correction, and the ring-by-ring float correction. The ring-by-ring float correction is used to correct the measured center point set, and / or the extrinsic offset vector is used to perform coordinate transformation on the subsequently acquired laser point cloud.

[0132] Optionally, the overall objective function J is defined as the data item J. _data With regularization term J _reg The sum. Data item J _data Describe the degree of agreement between the observed data and the theoretical model:

[0133] J _data =∑ i∈Ω w _i ×ρ _Huber (||π(C _design ,δe,δT _k ,δu _k )-p _i ||);

[0134] Where Ω corresponds to the set of data points within the sliding window, w_i The dynamic weight corresponding to the i-th data point, p _i Represents the measured point cloud or feature points; π(...) represents the projection function that projects the design model points onto the measurement space; ρ _Huber For Huber robust kernel function, when residual r < δ, ρ _Huber (r) = 0.5 × r 2 When the residual r ≥ δ, ρ _Huber (r)=δ×(|r|-0.5×δ) This function is used to suppress the pull of gross errors (such as temporary occlusion) on optimization. The threshold δ can be set to 1.345 times the median absolute deviation (MAD) of the residuals.

[0135] Regular term J _reg To prevent overfitting and parameter drift, it can be expressed as:

[0136] J _reg =λ _e ||δe|| 2 +λ _T ∑||δT _k -δT _k-1 || 2 +λ _u ∑||δu _k || 2 ;

[0137] Where, λ _e The extrinsic parameters should not deviate too far from their initial calibration values; λ _T Constrain the smoothness of attitude changes between adjacent rings (the attitude of the segments cannot change abruptly); λ _u The energy of the upward correction is constrained, reflecting the principle of minimal intervention.

[0138] On the other hand, the overall objective function for self-calibration can also be described by the following formula:

[0139] J=∑ i∈Ω w _i ×ρ _Huber (||π(C _design ,δe,δT _k ,δu _k )-p _i ||)+λ _e ||δe|| 2 +λ _T ∑||δT _k -δT _k-1 || 2 +λ _u ∑||δu _k || 2 .

[0140] Step 603: The dynamic weights of each data item in the self-calibration optimization objective function are determined by visual confidence, measurement covariance, and the current fitting residual.

[0141] Among them, visual confidence is derived from feature analysis or network inference of the segment appearance image, and measurement covariance is derived from the center point position uncertainty statistics output by the refined section fitting process.

[0142] Dynamic weights are positively correlated with visual confidence and negatively correlated with the norm of the measured covariance and the magnitude of the current fitting residual, in order to automatically reduce the optimization contribution of the corresponding data in regions where visual texture is blurred or geometric fitting is unstable.

[0143] For the i-th data point (or feature point), its weight w in the objective function _i It is not fixed, but calculated using the following formula:

[0144] w _i =ρ _vis_k ×exp(-||r _i || 2 / (2×σ geo 2 ))×(1 / (trace(Σ _k )+ε));

[0145] Where ε is a small constant to prevent the denominator from being zero, and σ geo The scale parameter represents the geometric residual, exp(...) represents the exponential function, and trace(Σ) _k Let be the trace of the measurement covariance matrix of the k-th ring, representing the geometric uncertainty. This formula contains three factors, specifically:

[0146] Visual confidence ρ _vis_k The score is given to the ring image by a deep learning network. If the image is clear and has rich texture, ρ _vis_k Approximately 1; if the image is blurry or completely black, ρ _vis_k The value is close to 0, which allows the algorithm to automatically reduce the weight of visual data when visual failure occurs.

[0147] Current fitted residual term exp(-||r _i || 2 / (2×σ geo 2 )), residual r _i The larger the value, the more exponentially the weight decreases.

[0148] Geometric uncertainty term (1 / trace(Σ) _k )), where Σ _k It measures the covariance matrix. If the center fit of the loop is highly unstable (Σ... _kIf the trace is large, the voice weight of that loop data is automatically reduced during self-calibration to prevent it from skewing the extrinsic parameter estimation of the entire system. Through the coupling mechanism, the system can rely on vision when visual features are rich but geometric features are degraded (such as pure rotation of a circular tunnel), and rely on geometry when vision is obstructed (such as water mist) but geometric features are good, thus achieving robust detection under various working conditions.

[0149] In some embodiments, specific implementation methods for intelligent determination of buoyancy modes based on structured displacement components are described. This explains how to establish a dynamic local coordinate system adapted to the tunnel curve orientation, decouple spatial displacement vectors into structured components, and automatically classify and diagnose the specific buoyancy modes of the tunnel segments based on the quantization logic of these components. Specifically, this includes:

[0150] Step 701: Construct a local orthogonal coordinate system at the design center position of each segment ring. The local orthogonal coordinate system includes the axial direction along the tangent direction of the design axis, the radial direction perpendicular to the design axis and pointing towards the crown, and the circumferential direction perpendicular to both the axial and radial directions.

[0151] In this application, since shield tunnels typically include straight sections, transition curve sections, and circular curve sections, their spatial orientation changes continuously with mileage. Therefore, it is necessary to establish a dynamically changing local reference system that varies with mileage, rather than using a fixed global coordinate system. For the k-th ring segment, based on the tunnel design axis equation, which is typically a combination of horizontal and vertical curves, the unit tangent vector t at the design center point of that ring is calculated. _k The three basis vectors of the locally orthogonal coordinate system are defined as follows:

[0152] Axial basis vector e _z It is directly taken as the tangent vector of the design axis, with the positive direction pointing towards the tunneling direction;

[0153] Radial basis vector e _r is defined as a unit vector perpendicular to the design axis and pointing towards the tunnel arch (i.e., the projection direction opposite to the direction of gravity). Mathematically, if the global vertical upward vector is V... _up =(0, 0, 1), then e _r =(V _up -(V _up ×e _z )e _z ) / ||V _up -(V _up ×e _z )e _z ||;

[0154] Circular basis vector e _θ , is defined as the cross product of the axial and radial directions, i.e., e _θ =e _z ×e _rThe positive direction points to the right side of the tunnel (facing the excavation direction). This set of basis vectors constitutes the local Frenet frame of the k-th ring, which can orthogonally decompose any displacement vector in three-dimensional space into three independent components: longitudinal (axial), vertical (radial), and transverse (circumferential).

[0155] Step 702: Calculate the spatial deviation vector between the measured tunnel centerline and the design axis at the corresponding mileage, and project the spatial deviation vector onto the three axes of the local orthogonal coordinate system to obtain a structured displacement component set containing axial displacement components, radial displacement components and circumferential displacement components.

[0156] Accordingly, let P be the coordinate of the measured tunnel centerline obtained by fitting at the kth ring mileage. _measure_k The corresponding design axis coordinates are P. _design_k Calculate the original spatial deviation vector ΔP. _k =P _measure_k -P _design_k Next, using the vector dot product operation, ΔP _k Projecting onto the aforementioned local coordinate system, we obtain the structured displacement component set {Δz}. _axial , Δr _radial Δs _circumf The specific calculation formula is as follows:

[0157] Axial displacement component Δz _axial =ΔP _k •e _z ;

[0158] Radial displacement component Δr _radial =ΔP _k •e _r ;

[0159] Circumferential displacement component Δs _circumf =ΔP _k •e _θ ;

[0160] Where, Δr _radial It directly reflects the vertical rise (positive value) or sink (negative value) of the tunnel lining segment and is the core indicator for buoyancy detection; Δz _axial This reflects the longitudinal movement of the tunnel segments along the tunnel axis; Δs _circumf It reflects the left and right lateral offset of the tunnel segment, and • represents the vector dot product.

[0161] Step 703: When the radial displacement component exceeds the preset floating threshold, and both the axial displacement component and the circumferential displacement component are less than the preset static threshold, it is determined to be a uniform floating mode.

[0162] The uniform upward movement mode is a pure translational motion of the entire segment ring along the vertical direction, typically caused by the buoyancy of the slurry. The system sets two thresholds: a significant upward movement threshold T... _float and static tolerance threshold T _static For example, based on engineering experience, T _float It can be set to 10 mm, T _static It can be set to 2 millimeters. When the condition (Δr) is met... _radial ≥T _float And (|Δz) _axial | <T _static And (|Δs) _circumf | <T _static When this occurs, the system outputs a diagnostic conclusion indicating uniform upward movement. In this case, the construction recommendation is usually to adjust the grout mix ratio or grouting pressure during synchronous grouting.

[0163] Step 704: When the radial displacement component and the axial displacement component both exceed their respective significant thresholds, and the two show a correlation change between continuous loops, it is determined to be a longitudinal displacement upward mode.

[0164] The longitudinal displacement upward floating mode occurs when the jack thrust is uneven or when there is jamming between the segment and the shield tail. The segment floats upward while being pushed out, i.e., it floats while being pushed. The judgment condition considers not only the value of a single ring but also the trend of multiple consecutive rings (e.g., three consecutive rings). If consecutive rings k, k+1, and k+2 all satisfy (Δr) _radial ≥T _float And (|Δz) _axial |≥T _axial_significant ), where T _axial_significant For example, if the value is set to 5 mm, and the absolute value of the correlation coefficient ρ(Δr, Δz) between the two sets of sequences is greater than 0.5, then it is determined to be longitudinal displacement upward. This indicates a strong coupling between upward movement and longitudinal thrust, and construction recommendations typically involve checking the zonal hydraulic pressure settings of the jacks. Or, T _axial_significant The corresponding axial displacement significant threshold.

[0165] Step 705: When the radial displacement component has a distribution difference in the circumferential direction that exceeds a preset non-uniform threshold, or when it is accompanied by a large segment tilt angle, it is determined to be a rotational warping and floating mode.

[0166] In another possible approach, when the cross-sectional deformation index exceeds the deformation threshold, it is determined to be a rotational warping and floating mode.

[0167] Furthermore, the rotational warping and floating mode is characterized by the segment ring undergoing torsion around its axis or elliptic deformation simultaneously with displacement. The determination of this mode relies on the following auxiliary indicators: segment tilt angle φ, and the radial offset range Δr of the cross-section. _range .

[0168] Where, Δr _range Defined as the difference between the maximum and minimum radial offsets at different locations (such as the crown, crown, left and right sides) of the annular cross-section, i.e., Δr _range =max(Δr _j )-min(Δr _j ), where Δr _j For each arc segment's radial offset, max(...) and min(...) represent the functions for taking the maximum and minimum values, respectively. The specific judgment logic is: if (|φ| ≥ tilt threshold T) _tilt ) or (Δr _range ≥ Non-uniform threshold T _uneven ), or (D _k ≥ Deformation index threshold T _def If T is positive, it is determined to be a case of rotational warping and upward floating. _tilt The setting can be set to 0.1 degrees; T _uneven Optional setting is 20 mm; T _def The setting can be set to 0.02. Once this mode is triggered, the tunnel segment structure may face the risk of cracking, and the system will issue the highest level red alert.

[0169] In other embodiments, a hardware architecture and a self-circulating network structure for the detection system are provided, specifically a physical device structure for implementing the method of the present invention, and details of a deep learning network for extracting visual features.

[0170] One of them is a segment vertical floating detection system during the shield tunnel assembly stage, comprising:

[0171] The data acquisition unit is used to acquire laser point cloud and appearance images of the inner surface of the tunnel segments, as well as shield tunneling status data;

[0172] Memory, used to store computer programs and pre-configured design parameters;

[0173] A processor, connected to a data acquisition unit and a memory, is used to execute a computer program to implement the method as described in any one of the present invention.

[0174] Optionally, the segment vertical floating detection system provided by this invention during the shield tunnel assembly stage mainly includes a data acquisition unit, a computing and processing unit, and a storage unit at the hardware level. The data acquisition unit is installed on the shield machine trolley or a dedicated detection trolley. Specifically, it includes:

[0175] A 3D laser scanner, such as a phase-type laser scanner, is mounted on a robotic arm or elliptical track slider that can rotate tangentially along the tunnel to acquire point cloud data of the full-circle cross-section.

[0176] High-definition industrial cameras, rigidly connected to or synchronously triggered with laser scanners, are used to acquire visible light images of the inner surface of tube segments.

[0177] Attitude measurement modules, such as high-precision inertial measurement units (IMUs) or dual-axis tiltmeters, are used to monitor the pitch and roll angles of the acquisition equipment in real time and assist in normal plane correction.

[0178] A mileage encoder, connected to the tunnel boring machine's propulsion cylinders or traveling wheels, provides high-frequency pulse signals for data alignment. A processor (such as an industrial computer or embedded GPU module) is connected to a memory pre-stored with a computer program. When the processor executes the program, it sequentially calls the various algorithm modules described in this embodiment of the invention.

[0179] Step 801: The method uses a pre-constructed self-circulating multi-scale attention network to extract visual features or visual confidence of the appearance image.

[0180] Accordingly, this invention designs a feature extraction network for accurately identifying segment joints and textures under complex tunnel lighting conditions. The network takes a flattened image of the segment as input and outputs a texture gradient map for boundary screening and a visual confidence score for weight calculation.

[0181] Step 802: The self-circulating multi-scale attention network includes a cyclic dilated convolutional layer. The convolutional kernels in the cyclic dilated convolutional layer are divided into several groups. Each group of convolutional kernels is configured with different dilation rates according to a preset cyclic sequence, so that the same convolutional layer can capture multi-scale image receptive field features simultaneously in a single forward propagation.

[0182] In this application, the convolutional layers employ a grouped cyclic dilation strategy. The specific implementation is as follows: Assume the number of input channels in a certain convolutional layer is C. _in The number of output channels is C _out The convolution kernel size is 3×3. The output channel C... _out Divide the data evenly into G groups, for example, G=4. Preset the expansion rate cyclic sequence D=[d _1 d _2 , ..., d _G For example, D=[1, 2, 5, 1, 2, 5, ...], or D=[1, 2, 5], assign values ​​cyclically according to this sequence. For the g-th output channel (index range from g×C) _out / G to (g+1)×C _out / G-1), whose corresponding convolution kernel is forced to use the dilation rate d during convolution operations. _g .

[0183] For example, the first group of convolutional kernels uses dilation rate 1 (standard convolution) to capture local fine textures; the second group uses dilation rate 2 to capture medium-scale features; and the third group uses dilation rate 5 to capture large-scale contextual information. Based on this, the output feature maps of each group of convolutions are concatenated along the channel dimension. Through this parameter grouping and dilation rate loop within a single layer, the network can simultaneously perceive the micro-cracks (small scale) and the overall assembly outline (large scale) of the pipe segment in a single forward propagation.

[0184] Step 803: After the recurrent dilation rate convolutional layer, the self-circulating multi-scale attention network is further configured with a channel attention module and a spatial attention module.

[0185] Furthermore, an attention mechanism (CBAM structure, convolutional block attention module) is chained after the recurrent convolutional layers. This is a channel attention module. The input feature map F is processed by global average pooling and global max pooling to obtain two one-dimensional vectors. These two vectors are processed by a shared multilayer perceptron (MLP) and then added together. A sigmoid activation function is then used to generate channel weight vectors. These channel weight vectors are multiplied by the original feature map F to obtain the channel recalibration features F'. Feature channels more valuable for float detection, such as edge response channels, are selected.

[0186] Furthermore, it is also a spatial attention module. Taking F' as input, it performs max pooling and average pooling along the channel dimension to obtain two two-dimensional feature maps. After concatenating the two maps, it is passed through a 7x7 or larger scale convolutional layer and activated by Sigmoid to generate a spatial weight map. The spatial weight map is multiplied by F' to obtain a weighted feature map F''. The weights are used to suppress background regions of the image (such as textureless concrete surfaces) and highlight geometric feature regions such as seams and bolt holes, thereby improving visual confidence ρ. _vis Reliability.

[0187] In step 804, the channel attention module recalibrates the feature channels using global pooling and one-dimensional convolution, while the spatial attention module extracts spatially dependent features using large-scale dilated convolution. Together, they generate a weighted feature map to characterize the saliency of the pipe segment joints.

[0188] Specifically, for convolutional operations in the spatial attention module, to further capture long-distance spatial dependencies (such as long seams running through the entire image), large-scale dilated convolutions (e.g., kernel size of 7x7, dilation rate of 3) can be used instead of standard convolutions. The generated spatial weight map can maintain the continuity of the seams and prevent seam feature breaks caused by uneven lighting. Based on this, the generated weighted feature map is fed into the subsequent regression head or classification head, which outputs visual confidence and auxiliary judgment probability of the floating mode, respectively.

[0189] In other embodiments, the specific implementation process of offline model training and system initialization configuration is described, particularly the training strategy of the self-recurrent multi-scale attention network and the calibration process of the detection system before deployment. In this embodiment, the following steps can be performed:

[0190] Step 901: Construct a training dataset of tunnel segment images containing various working conditions, and define a multi-task loss function for jointly training a self-recurrent multi-scale attention network.

[0191] Accordingly, an offline image dataset of tunnel segments is required, encompassing various conditions such as clear / blurred, dry / wet, and intact / damaged. The dataset is then manually annotated, including pixel-level masks of the seams and image quality confidence scores (e.g., scalars from 0 to 1). The overall training loss function L for the network is defined. _total The weighted sum of the segmentation loss and the confidence regression loss:

[0192] L _total =L _seg +λ _conf ×L _reg ;

[0193] Among them, L _seg Binary cross-entropy loss or Dice loss (a loss function derived from the Dice coefficients based on the Sorenson-Dixson coefficients) is used to supervise the network's ability to extract features of seam texture; L _reg Mean squared error loss is used to supervise the network output confidence level to approximate human ratings. This loss function is minimized by stochastic gradient descent (SGD) or the Adam optimizer (adaptive moment estimator), enabling the pre-trained network to not only extract texture gradients but also perceive the observation quality of the current environment, providing visual confidence input for self-calibrating weights.

[0194] Or, L _total L represents the total training loss of the network. _seg For the segmentation loss; λ _conf The weights for the confidence regression task; L _reg For regression loss.

[0195] Step 902: Perform initial external parameter calibration and elliptical orbit parameter correction of the detection system to generate a pre-configured system parameter file.

[0196] The initial value of the extrinsic parameter offset vector needs to be calibrated during the system installation phase. A total station is used to measure high-precision control points (GCPs) within the tunnel, and a laser scanner is used to scan these control points. The correspondence between the control points in the world coordinate system and the scanner coordinate system is established. Singular value decomposition (SVD) is used to solve for the initial rotation matrix and translation vector, which are then used as the starting values ​​for sliding window optimization, i.e., the zeros of δe. Furthermore, for scanning systems using elliptical track sliders, the major and minor axis radii of the track must be pre-determined, and the geometric parameters are written into the system's configuration file. During real-time detection, the processor reads the pre-configured parameters to convert the scanner's local polar coordinate data into Cartesian coordinate data for the tunnel cross-section.

[0197] In some other embodiments, the method of the present invention may also be:

[0198] Accordingly, by fusing multi-source data such as laser point cloud and segment appearance images, and taking the real-time spatial pose of the tunnel boring machine during the tunneling process as a benchmark, the collected multi-source data is associated and integrated with the segment ring number to construct a ring position data slice set arranged according to the segment ring number. The original multi-source dataset contains laser point cloud data, segment appearance image data and pose parameter information.

[0199] In this embodiment, the acquisition of raw, multi-source heterogeneous data relies on a multi-sensor system integrated on the tunnel boring machine (TBM), which forms the data input foundation for the entire inspection process. This system typically includes a laser scanner and a linear array camera positioned approximately 3 to 4 rings behind the tail of the shield, as well as an attitude measurement unit mounted on the TBM. The laser scanner is used to acquire high-density three-dimensional point cloud data of the inner surface of the tunnel segments. This data is stored as a series of discrete points with three-dimensional spatial coordinates (XYZ), characterizing the geometry, spatial undulations, and microscopic deformations of the segment surface.

[0200] Simultaneously, a high-resolution area array camera acquires digital images of the cross-sectional area, obtaining visual texture information of the tunnel segment surface, including the appearance features of the concrete tunnel segments and imaging of the segment connection joints. The attitude measurement unit is responsible for real-time monitoring and outputting the three-dimensional attitude angles of the tunnel boring machine during the tunneling process, i.e., the tilt angle, providing a pose reference for subsequent data calculations.

[0201] Specifically, a data slice set arranged by ring number is constructed based on the tunneling progress information of the tunnel segments. This process relies on the displacement information provided by the tunnel boring machine's own propulsion to establish a data organization system strictly arranged according to the ring number of the tunnel segments. The propulsion system generates pulse signals proportional to the tunneling distance by monitoring the stroke of the tunnel boring machine's hydraulic cylinders. The system sets the standard width (2 meters) of each ring segment as the displacement cycle. By analyzing the pulse count in real time, the data processing unit can accurately determine the start and end times of tunneling for each ring segment, defining a unique data acquisition period for each ring number.

[0202] Based on this timeline, the system performs automated classification and aggregation of multi-source data. All 3D laser point cloud slices, optical image slices of the tube edge, and spatial attitude parameters recorded by the inertial measurement unit within the same ring number and corresponding time period are associated and packaged into independent ring position data slice sets identified by ring number.

[0203] Optionally, a density-based spatial clustering algorithm can be introduced to preprocess the original point cloud when constructing the ring-shaped data slice set. This algorithm identifies and filters out noise points by setting two parameters: the neighborhood search radius and the minimum number of neighborhood points. The system traverses each data point in the point cloud, counts the number of points in its neighborhood, and separates discretely distributed noise points from high-density point clusters representing the real structure based on a density threshold. Next, the algorithm aggregates points belonging to the same high-density region into independent clusters, calculates the geometric center of each cluster as a representative point, and reconstructs the denoised point cloud accordingly. This method can eliminate outliers caused by environmental noise such as shadows and illumination, as well as high-frequency vibrations of equipment, while preserving the true geometric features of the pipe segment surface, thus improving the robustness and accuracy of subsequent normal vector estimation, feature extraction, and geometric fitting processes.

[0204] Based on the ring-shaped data slice set, after cross-sectional feature extraction and center point calculation, and combined with optimized sensor-acquired data and tilt angle parameters, an accurate set of cross-sectional center points is generated. This set of center points is then used to obtain the measured tunnel centerline through curve fitting, and its displacement deviation from the design axis is used to construct a self-calibrating optimized dataset.

[0205] In this step, extracting the dense point cloud features of all cross-sections refers to systematically identifying, separating, and reconstructing geometric information that can characterize the joints and structural deformation of the tunnel segments from a massive collection of ring-shaped data slices. This process forms the data foundation for subsequent high-precision pose alignment and deformation analysis. The edges of the tunnel segments, as the mechanically weak points and structural discontinuities between precast concrete rings, are characteristic areas reflecting the overall deformation and relative displacement of the segments. In particular, the joints typically exhibit obvious abrupt changes in planarity and curvature discontinuities; extracting these local features is a prerequisite for achieving high-precision spatial registration and deformation parameter inversion between the tunnel segment rings.

[0206] Specifically, based on the local surface fitting method, the normal vectors of the 3D point cloud at each point of the cross-section are calculated to construct a point cloud normal vector field. By analyzing the abrupt changes in the normal vector direction, regions where the rate of change of the normal vector exceeds a preset threshold are identified. These regions correspond to the edge contour of the annular cross-section. The point cloud data within these regions are extracted into independent cross-sectional point cloud subsets, which serve as inputs for subsequent feature matching and alignment operations.

[0207] Furthermore, based on the ring number arrangement, the point cloud data corresponding to each ring is cross-sectionally sliced ​​and reassembled in three-dimensional space. Using the Iterative Closest Point (ICP) algorithm, the cross-sectional point clouds of adjacent rings are accurately spatially aligned. Through three-dimensional modeling technology, seamless stitching of point cloud data from multiple continuous rings is achieved, constructing a continuous three-dimensional solid model covering a long tunnel section, providing a complete geometric context for the overall fitting and deformation analysis of the tunnel axis.

[0208] Furthermore, by calculating the spatial collinearity between the measured tunnel centerline and the design centerline, an accurate assessment of the pose of the center point of each segment ring is achieved. For each segment ring, the projection deviation of the measured center point on the design axis normal plane is calculated; this deviation vector represents the overall pose offset of the segment ring in three-dimensional space. The deviation values ​​of all rings along the entire line are aggregated to generate an upward deviation result set arranged by ring number and recording the spatial deformation of the tunnel axis, providing a data foundation for subsequent construction feedback and safety assessment.

[0209] Based on this, tunneling mileage information can be introduced as an auxiliary alignment benchmark. The cumulative advance distance recorded in real time is compared with the designed ring width of the tunnel segments to perform preliminary time-domain matching. The theoretical number of ring segments is calculated, mapped to the timestamps of the data stream, and the approximate time range of data acquisition for each ring is estimated, determining an initial time offset. This offset can serve as the initial iterative value for subsequent high-precision self-calibration optimization, narrowing the parameter search space.

[0210] The centerline parameter set obtained from the preliminary calculation is transformed from the Cartesian coordinate system to the polar coordinate system. This mapping converts the spatial distribution of the centerline into a sequential representation based on radial distance and angle. This representation data is then input into a pre-trained self-recurrent multi-scale attention neural network to capture and output a self-recurrent multi-scale attention feature set.

[0211] Furthermore, transforming the preliminarily calculated centerline parameter set from the Cartesian coordinate system to the polar coordinate system means mapping the three-dimensional rectangular coordinates (X, Y, Z) of the center points of each section to a polar coordinate representation with the tangent point of the design axis as the origin, using the tunnel design axis as the reference. Core parameters include the arc length along the design axis (determining the longitudinal position), the radial offset relative to the design axis (characterizing the magnitude of uplift or subsidence), and the circumferential angle within the normal plane of that section. This not only transforms the complex problem of comparing spatial curves into a more intuitive radial distance sequence analysis, but also generates a polar coordinate sequence whose data structure better conforms to the geometric characteristics of the tunnel axis as a one-dimensional extension.

[0212] Specifically, the self-circulating multi-scale attention feature set refers to inputting the polar coordinate transformed image data into a pre-designed deep neural network. Unlike traditional global convolutional networks, this network employs a multi-scale parallel modular structure to extract aggregated features across different dimensions and spaces. For example, based on the edge response characteristics of the cross-sectional image, the network decomposes the feature extraction process into three core weighting strategies: channel domain compression and recalibration, spatial domain feature focusing, and context enhancement based on the cyclic dilation rate. For instance, one module extracts shallow feature spatial location information in the network width direction, while another module extracts high-level semantic information in the network depth direction. These features are then fused through attention weighting to form an enhanced feature map that is highly sensitive to the floating of the tunnel segments.

[0213] Furthermore, by integrating the floating geometry of the ring segment data slice set and the deep visual features of the self-circulating multi-scale attention feature set, the spatial pose of the center of each ring segment relative to the design axis is collaboratively analyzed and accurately quantified to generate a structured floating deviation result set.

[0214] In this embodiment, the geometric shape of the floating segment refers to a multi-dimensional vector data representation, which is deconstructed into displacement components in three orthogonal dimensions: axial, radial, and circumferential. The geometric displacement is calculated based on high-precision point cloud and pose data to quantify the physical scale of the floating phenomenon; while the visual confidence level originates from an attention feature map generated by a deep learning model, providing visual evidence for the existence and spatial distribution of the floating phenomenon.

[0215] Correspondingly, calculating the uplift geometric vector of multidimensional vector data refers to establishing a unified standard coordinate system at the spatial location of each point cloud data point, calculating the relative displacement between the measured centerline and the design centerline of the tunnel, and projecting this displacement onto the axial, radial, and circumferential axes. For example, the axial component reflects the extent of the uplift, and the radial component reflects the displacement difference. Generating the uplift deviation result set refers to classifying the uplift as uniform overall uplift, longitudinal displacement uplift, or rotational warping uplift based on the numerical proportions of the components and verification using visual features, outputting specific misalignment values ​​and classification labels. This helps provide construction personnel with more targeted rectification suggestions.

[0216] This application employs a deformation separation algorithm based on rigid segment constraints and a joint optimization model based on shared center and radial offset. This approach can eliminate the interference of local non-rigid deformation from both mathematical and physical perspectives, extract the rigid center representing the true spatial position of the segment, and eliminate the systematic errors of traditional ellipse fitting methods.

[0217] Furthermore, a sliding window self-calibration mechanism integrating visual confidence and geometric measurement covariance was constructed. By coupling visual texture quality with the stability (covariance) of geometric fitting and optimizing the weights, online dynamic correction of sensor parameters was achieved, ensuring that detection accuracy does not decrease with increasing mileage. Simultaneously, adaptive parametric interpolation based on the uplift sensitivity gradient was employed to automatically densify nodes in areas of severe deformation, resolving the fitting distortion problem under complex linear shapes.

[0218] In addition, a unified coordinate system with multiple components was established, which decouples the abstract spatial deviation into axial, radial and circumferential components with clear physical meaning. Combined with intelligent judgment logic, it solves the engineering problem that a single index is not enough to distinguish complex modes such as uniform floating, longitudinal movement and rotational warping.

[0219] The optional embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solution of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for detecting the vertical floating of tunnel segments during the assembly stage of a shield tunnel, characterized in that, include: The laser point cloud and appearance image of the inner surface of the tunnel segment are obtained, and combined with the shield tunneling information, a set of ring position data slices indexed by ring number is constructed. A refined cross-sectional fitting that takes into account the local deformation of the pipe segments is performed on the annular data slice set to extract the set of measured center points of each annular cross section; Based on the fitting sensitivity characteristics of the measured center point set, adaptive parametric interpolation is performed to generate the measured tunnel centerline; The self-calibration optimization objective function containing multi-source constraints is invoked, and the extrinsic parameter offset and ring-by-ring attitude correction of the ring position data slice set are incrementally updated using a sliding window. Establish a multi-dimensional component unified coordinate system, decompose the spatial deviation of the measured tunnel centerline relative to the design axis into a set of structured displacement components, and determine the segment floating mode accordingly. The optimization variables of the self-calibration optimization objective function include at least the extrinsic parameter offset vector of the sensor relative to the segment, the ring-by-ring attitude correction, and the ring-by-ring buoyancy correction. The self-calibration optimization objective function includes a Huber robust term to characterize the reprojection error, and a regularization constraint term to constrain the physical rationality of the variables. The regularization constraint term includes extrinsic parameter energy constraints that limit the extrinsic parameter drift amplitude, smoothness constraints that limit the abrupt changes in the attitude of adjacent segments, and buoyancy energy constraints that limit the magnitude and rate of change of the buoyancy correction.

2. The method according to claim 1, characterized in that, Constructing a ring-position data slice set indexed by ring number, specifically including: In response to the pulse signal of the shield tunneling system, the start and end times of tunneling of a single ring segment are determined. Based on this, the timestamps of the laser point cloud and the appearance image are aligned, and the data units belonging to the corresponding ring numbers are divided. Density-based clustering is performed on the laser point cloud in the data unit to filter out outliers and noise. Principal component analysis is used to calculate the local normal vector of each point cloud to generate a ring-shaped data slice set containing geometric normal information and texture information.

3. The method according to claim 1, characterized in that, To perform refined cross-sectional fitting on the annular segment data slice set, taking into account the local deformation of the segments, a deformation-displacement separation method based on rigid segment constraints is adopted, specifically including: Based on the design center angle range of the tunnel segment, the cross-sectional point cloud in the ring location data slice set is divided into a segment point cloud subset corresponding to each tunnel segment; For each segment point cloud subset, rigid body registration is performed with its design reference position to estimate the independent rigid body transformation parameters of each segment. Based on the minimum deformation energy criterion, the rigid body displacement component of the entire ring and the elastic deformation component of the joint are separated from the rigid body transformation parameters of each segment. The cross-sectional position is corrected by using the rigid body displacement component of the entire ring, and the set of measured center points is extracted.

4. The method according to claim 3, characterized in that, Estimate the independent rigid body transformation parameters of each segment by using the singular value decomposition algorithm to solve for the rotation matrix and translation vector of each segment relative to the design position. The displacement components of the entire ring rigid body and the elastic deformation components of the joints are separated, including: Using the rigid body transformation parameters of the entire ring as optimization variables, the optimization objective function is called to minimize the weighted difference between the rigid body transformation parameters of each segment and the rigid body transformation parameters of the entire ring. Solving the objective function yields the translation vector of the entire ring, which is used as the actual upward displacement. The sum of the difference norms between the rigid body transformation parameters of each segment and the rigid body transformation parameters of the entire ring is calculated to obtain the cross-sectional deformation index that reflects the degree of joint deformation.

5. The method according to claim 1, characterized in that, The refined cross-section fitting employs a joint optimization method based on shared center and arc segment offset, specifically including: The cross-sectional point cloud is divided into several arc segment point sets, and a joint optimization objective function containing shared center variables and arc segment local radial offset variables is called. The shared center variables constrain the geometric position of the entire ring cross-section, and the arc segment local radial offset variables are used to absorb the independent deformation of each arc segment relative to the ideal ellipse. The measured center points of each ring section are obtained by solving the joint optimization objective function. Based on the residual statistical characteristics and Hessian matrix in the optimization process, the measurement covariance matrix characterizing the uncertainty of the center point position is calculated.

6. The method according to claim 5, characterized in that, Before invoking the joint optimization objective function, the process also includes section projection and arc segment division, specifically: A rotation matrix is ​​constructed using the tangential vector of the measured centerline and the collected tilt angle data. The normal vector of the corrected normal plane, taking into account the segment attitude, is calculated. The cross-sectional point cloud is then projected onto the plane determined by the normal vector of the corrected normal plane. Multi-source consistency screening is performed on the normal abrupt change features of the cross-sectional point cloud and the seam texture features of the corresponding appearance image. The positions that simultaneously satisfy the geometric abrupt change significance and visual texture significance are determined as stable arc segment boundaries. Based on this, the projected cross-sectional point cloud is divided into arc segment point sets.

7. The method according to claim 1, characterized in that, Fit sensitivity features include geometric curvature indices, upward gradient indices, and design deviation indices; adaptive parametric interpolation is performed, specifically including: The rate of change of the direction angle of the line connecting adjacent points in the set of measured center points is calculated as the geometric curvature index; the change of the vertical deviation of adjacent measured center points relative to the design axis is calculated as the upward gradient index; and the spatial distance between each measured center point and the corresponding position of the design axis is calculated as the design deviation index. The geometric curvature index, the buoyancy gradient index, and the design deviation index are weighted and fused to obtain the comprehensive sensitivity of each measured center point. Based on the comprehensive sensitivity, the parameter increment of the spline interpolation is adjusted in reverse. In the region where the comprehensive sensitivity is higher than the preset threshold, the parameter increment is reduced and the parameter nodes are densified. The quadratic B-spline curve is solved using the adjusted parameter sequence.

8. The method according to claim 1, characterized in that, A multi-component unified coordinate system is established, and the spatial deviation of the measured tunnel centerline relative to the design axis is decomposed into a set of structured displacement components, specifically including: A local orthogonal coordinate system is constructed at the design center of each segment ring. The local orthogonal coordinate system includes the axial direction along the tangent of the design axis, the radial direction perpendicular to the design axis and pointing towards the crown, and the circumferential direction perpendicular to both the axial and radial directions. Calculate the spatial deviation vector between the measured tunnel centerline and the design axis at the corresponding mileage, and project the spatial deviation vector onto the three axes of the local orthogonal coordinate system to obtain a structured displacement component set containing axial displacement components, radial displacement components and circumferential displacement components.

9. A segment vertical floating detection system during the shield tunnel assembly stage, characterized in that, include: The data acquisition unit is used to acquire laser point cloud and appearance images of the inner surface of the tunnel segments, as well as shield tunneling status data; Memory, used to store computer programs and pre-configured design parameters; A processor, connected to a data acquisition unit and a memory, is used to execute a computer program to implement the method as described in any one of claims 1 to 8.

Images