Tunnel lining disease automatic identification method and system based on multi-source data fusion

Abstract

This invention discloses an automatic identification method and system for tunnel lining defects based on multi-source data fusion, belonging to the field of facility inspection technology. The method includes: acquiring multimodal time-series data and performing spatiotemporal alignment to obtain a time-series multi-source dataset; reconstructing the tunnel centerline based on vehicle pose and constructing a lining structure consistency coordinate framework; performing structured projection and distortion correction on the aligned data to obtain a multimodal fusion dataset; dividing a two-dimensional structural mesh within this coordinate framework, extracting and fusing geometric, texture, depth, and energy features, calculating the structural consistency failure index, and extracting suspected defect areas; combining multimodal physical evidence to calculate a defect confidence index and determine the defect type; mapping the suspected defect areas back to three-dimensional space; completing defect boundary extraction and geometric quantization; and outputting structured defect information. This invention, by constructing a lining structure consistency coordinate framework, achieves the organized expression and structured alignment of cross-modal data under a unified geometric reference.

Description

Technical Field

[0001] This invention relates to the field of facility inspection technology, and in particular to an automatic identification method and system for tunnel lining defects based on multi-source data fusion. Background Technology

[0002] With the continuous expansion of the construction scale of rail transit, urban tunnels, and highway tunnels, the long-term safe operation of tunnel lining structures has become an important part of infrastructure operation and maintenance. In recent years, structural inspection technologies based on machine vision, depth imaging, laser scanning, and ground-penetrating radar have been gradually applied to tunnel defect identification. By observing surface cracks, bulges, voids, and internal defects in the lining, the technology has evolved from manual inspection to equipment-assisted detection. At the same time, the development of multi-source sensor platforms has enabled the simultaneous acquisition of color images, depth maps, point clouds, and radar signals on mobile carriers, providing conditions for multimodal perception of lining structures.

[0003] While existing tunnel lining defect detection technologies can utilize single modalities such as vision, point clouds, or ground-penetrating radar for identification, they still suffer from two key shortcomings. First, existing technologies lack a unified structural reference system, making it impossible to organize multimodal data within the same geometric framework. This results in inconsistent representation of defect location, structural scale, and anomaly amplitude across different modalities, limiting the reliability of multimodal fusion. Second, existing technologies lack a cross-validation mechanism based on multimodal physical evidence, making it impossible to simultaneously and consistently assess texture damage, geometric shifts, and internal energy changes. This makes it difficult to distinguish suspected anomalies from actual defects, and also prevents reliable quantification of defect type, depth, and volume. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an automatic identification method for tunnel lining defects based on multi-source data fusion to solve the problems of existing technologies, such as the lack of a unified structural reference system leading to inaccurate alignment of multimodal data and the lack of multimodal physical evidence resulting in insufficient reliability of defects.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an automatic identification method for tunnel lining defects based on multi-source data fusion, which includes: collecting time-series data, performing spatiotemporal alignment, and obtaining a time-series multi-source dataset;

[0008] Based on a time-series multi-source dataset, the trajectory of the tunnel centerline is reconstructed using vehicle pose, and a consistent coordinate framework for the lining structure is obtained.

[0009] By using a lining structure consistency coordinate framework to perform structured projection and distortion correction on time-series data, a multimodal fusion dataset is obtained.

[0010] The tunnel structure is divided based on the lining structure consistency coordinate framework to obtain a two-dimensional structural mesh. The structural consistency multimodal feature set is then extracted from the two-dimensional structural mesh based on the multimodal fusion dataset.

[0011] Based on the unified fusion of structural consistency multimodal features, the structural consistency failure index is calculated, and a structural failure distribution map is generated according to the structural consistency failure index. Suspected disease areas are then extracted from the structural failure distribution map.

[0012] Multimodal physical evidence is extracted from suspected disease areas. Self-verifying multimodal cross-validation of suspected disease areas is performed using multimodal physical evidence to calculate the disease credibility index and obtain the disease type.

[0013] Based on the disease credibility index and disease type, suspected disease areas are mapped to the lining structure consistency coordinate frame, and the three-dimensional boundary is extracted and geometrically quantified to obtain structured disease information.

[0014] As a preferred embodiment of the automatic identification method for tunnel lining defects based on multi-source data fusion described in this invention, the specific steps for collecting time-series data, performing spatiotemporal alignment, and obtaining a time-series multi-source dataset are as follows.

[0015] Color images, depth maps, laser point clouds, echo timing signals, and high-frequency pose sequences are acquired. Time alignment is performed using a unified trigger timestamp. The time-aligned color images, depth maps, laser point clouds, echo timing signals, and high-frequency pose sequences are then bound to coordinates to obtain a time-series multi-source dataset.

[0016] As a preferred embodiment of the automatic identification method for tunnel lining defects based on multi-source data fusion described in this invention, the specific steps for obtaining a consistent coordinate framework for the lining structure by reconstructing the tunnel centerline trajectory using vehicle pose based on a time-series multi-source dataset are as follows.

[0017] The high-frequency pose sequences in the time-series multi-source dataset are smoothed to obtain smooth position trajectories. The smooth position trajectories are then connected in time order to obtain the preliminary tunnel centerline trajectory.

[0018] Based on the preliminary tunnel centerline trajectory, the slight deviation of the vehicle in a local position is corrected by distance constraint to obtain the tunnel centerline;

[0019] Distance filtering, angle selection, and statistical filtering are performed on laser point clouds in a time-series multi-source dataset to obtain an effective set of point clouds.

[0020] A local cross-section reference plane is constructed based on the tunnel centerline, and the effective point cloud set is projected onto the local cross-section reference plane to obtain the cross-section point cloud set;

[0021] By constructing the principal surface function through the radial distribution of cross-sectional point clusters in the circumferential direction, and determining the unit normal vector and curvature field based on the local geometric changes of the principal surface of the tunnel lining, a consistent coordinate framework for the lining structure is constructed.

[0022] As a preferred embodiment of the automatic identification method for tunnel lining defects based on multi-source data fusion described in this invention, the step of obtaining a multimodal fusion dataset by performing structured projection and distortion correction on time-series data through a lining structure consistency coordinate framework is as follows:

[0023] Spatial back projection is performed on the time-series multi-source dataset to obtain three-dimensional observation points. The longitudinal parameters, circumferential parameters and radial offset of the three-dimensional observation points are determined by the principal surface function to obtain the structural parameter triplet.

[0024] Based on the normal direction of the master surface function, point cloud error correction, depth tilt compensation and energy attenuation compensation are performed on the structural parameter triplet to obtain the corrected three-dimensional observation point, the corrected radial offset and the corrected echo energy value.

[0025] The corrected 3D observation points, corrected radial offset, and corrected echo energy values ​​are uniformly expressed within the lining structure consistency coordinate framework to obtain a multimodal fusion dataset.

[0026] As a preferred embodiment of the automatic identification method for tunnel lining defects based on multi-source data fusion described in this invention, the method involves: dividing the tunnel structure based on a lining structure consistency coordinate frame to obtain a two-dimensional structural mesh; and extracting a structural consistency multi-modal feature set from the two-dimensional structural mesh based on a multi-modal fusion dataset. The specific steps are as follows:

[0027] Based on the lining structure consistency coordinate frame, the tunnel is divided at fixed mileage intervals on the longitudinal parameter axis, and the lining perimeter is divided at equal angular intervals on the circumferential parameter axis to obtain a two-dimensional structural mesh.

[0028] A multimodal fusion dataset is collected at each structural location of a two-dimensional mesh, and geometric features, texture features, depth features, and energy features are extracted.

[0029] Geometric features, texture features, depth features, and energy features are combined according to the lining structure consistency coordinate framework to obtain a multimodal feature set of structure consistency.

[0030] As a preferred embodiment of the automatic identification method for tunnel lining defects based on multi-source data fusion described in this invention, the steps of unifying and fusing multi-modal features of structural consistency, calculating the structural consistency failure index, and generating a structural failure distribution map based on the structural consistency failure index are as follows.

[0031] Based on the statistical feature benchmark value of the structural consistency multimodal feature set, the deviation of each structural position from the benchmark value is calculated to obtain the geometric anomaly, texture anomaly, depth anomaly and energy anomaly.

[0032] The geometric anomaly, texture anomaly, depth anomaly and energy anomaly are integrated to obtain the structural consistency violation index;

[0033] The structural integrity failure index is mapped onto a two-dimensional unfolded plane according to the longitudinal and circumferential parameters of the structural elements, and a structural failure distribution map is drawn.

[0034] As a preferred embodiment of the automatic identification method for tunnel lining defects based on multi-source data fusion described in this invention, the specific steps for extracting suspected defect areas from the structural damage distribution map are as follows:

[0035] Adaptive threshold segmentation is performed on the structural damage distribution map by constructing a damage index histogram and traversing candidate thresholds to select the optimal threshold;

[0036] Structural locations above the optimization threshold are marked as suspected defects, while structural locations below the optimization threshold are marked as normal locations.

[0037] Connectivity analysis and morphological screening were performed on the locations marked as suspected diseases to obtain suspected disease areas.

[0038] As a preferred embodiment of the automatic identification method for tunnel lining defects based on multi-source data fusion described in this invention, the steps of extracting multimodal physical evidence in suspected defect areas, performing self-verifying multimodal cross-validation on the suspected defect areas using the multimodal physical evidence, calculating the defect confidence index, and obtaining the defect type are as follows.

[0039] Evidence of texture inconsistency disruption, geometric morphology shift, and energy reflection changes was extracted from suspected diseased areas.

[0040] The evidence of texture consistency disruption, geometric shape shift, and energy reflection change is subjected to self-verifying multimodal cross-validation to calculate the disease credibility index.

[0041] The disease type is identified by using the disease credibility index.

[0042] As a preferred embodiment of the automatic identification method for tunnel lining defects based on multi-source data fusion described in this invention, the steps of mapping suspected defect areas to a lining structure consistency coordinate frame based on defect confidence index and defect type, extracting three-dimensional boundaries and performing geometric quantization of defects to obtain structured defect information are as follows.

[0043] The suspected defect area is rasterized within the lining structure consistency coordinate frame to obtain the three-dimensional surface distribution of the suspected defect.

[0044] Perform connected region analysis and contour extraction on the three-dimensional surface distribution of suspected defects, obtain the three-dimensional boundary line of the defects by boundary tracking, and calculate the width index by the tunnel centerline trajectory.

[0045] The diseased area is regularly discretized within the lining structure consistency coordinate frame to obtain the diseased surface area. Radial damage statistics are performed using the disease confidence index and the corrected radial offset to calculate the equivalent depth of the disease.

[0046] By combining disease equivalence depth, width indicators, disease type, and disease credibility index, structured disease information is obtained.

[0047] Secondly, the present invention provides an automatic identification system for tunnel lining defects based on multi-source data fusion, including the acquisition and alignment module, which acquires time-series data, performs spatiotemporal alignment, and obtains a time-series multi-source dataset.

[0048] The centerline reconstruction module, based on a time-series multi-source dataset, uses vehicle pose to reconstruct the tunnel centerline trajectory and obtain a consistent coordinate framework for the lining structure.

[0049] The structural correction module performs structured projection and distortion correction on time-series data through the lining structure consistency coordinate framework to obtain a multimodal fusion dataset;

[0050] The mesh feature module divides the tunnel structure based on the lining structure consistency coordinate framework, obtains a two-dimensional structural mesh, and extracts a structural consistency multimodal feature set on the two-dimensional structural mesh based on the multimodal fusion dataset;

[0051] The damage assessment module integrates multimodal features of structural consistency to calculate the structural consistency damage index, generates a structural damage distribution map based on the structural consistency damage index, and extracts suspected disease areas from the structural damage distribution map.

[0052] The trusted verification module extracts multimodal physical evidence in suspected disease areas, performs self-verifying multimodal cross-verification of suspected disease areas using multimodal physical evidence, calculates the disease trust index, and obtains the disease type.

[0053] The three-dimensional quantization module maps suspected disease areas to the lining structure consistency coordinate frame based on the disease confidence index and disease type, performs three-dimensional boundary extraction and geometric quantization of the disease, and obtains structured disease information.

[0054] The beneficial effects of this invention are as follows: by constructing a lining structure consistency coordinate framework, the organization, expression, and structured alignment of cross-modal data under a unified geometric reference are realized; by proposing a multimodal structural consistency violation index, four types of physical features—geometric offset, texture violation, depth anomaly, and energy attenuation—are integrated into structural consistency, achieving a comprehensive quantification of the nature of the disease; by introducing a self-proving multimodal cross-validation mechanism, a disease credibility index is constructed by combining evidence of texture consistency violation, evidence of geometric shape offset, and evidence of energy reflection changes, enabling automatic judgment of the authenticity, type, and level of the disease. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of an automatic identification method for tunnel lining defects based on multi-source data fusion.

[0057] Figure 2 This is a schematic diagram of an automatic identification system for tunnel lining defects based on multi-source data fusion.

[0058] Figure 3 This is a flowchart for multimodal cross-validation and disease type determination in suspected disease areas.

[0059] Figure 4 This is a flowchart for extracting the three-dimensional boundary of the disease and outputting structured disease information. Detailed Implementation

[0060] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0061] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0062] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0063] Reference Figures 1-4 This is one embodiment of the present invention, which provides an automatic identification method for tunnel lining defects based on multi-source data fusion, comprising the following steps:

[0064] S1. Collect time-series data, perform spatiotemporal alignment, and obtain a multi-source time-series dataset.

[0065] A color camera is mounted at the front of the vehicle, facing the tunnel lining wall, using a fixed-focus industrial lens to capture color images. A depth camera is mounted parallel to the color camera, sharing a portion of the imaging area, and uses infrared structured light to acquire depth information, capturing a depth map synchronized with the color image. A laser scanner is installed at the center of the vehicle roof, using a 3D rotating lidar to scan the entire tunnel cross-section and capture point clouds. A ground-penetrating radar is installed below the front side of the vehicle, emitting electromagnetic pulses into the lining at a fixed frequency and receiving reflected echoes to acquire echo timing signals. An inertial measurement unit (IMU) is installed near the vehicle's center of gravity to acquire three-axis acceleration and three-axis angular velocity in real time, obtaining high-frequency pose sequences.

[0066] By using a unified trigger timestamp, a one-to-one correspondence is established between color images, depth maps, point clouds, echo timing signals, and high-frequency pose sequences on the same time axis for time alignment. The imaging coordinate system with the camera center as the origin is used as the color camera coordinate system, the depth imaging coordinate system related to the depth measurement optical axis is used as the depth camera coordinate system, the local coordinate system of the laser point cloud with the laser emitter as the origin is used as the laser scanner coordinate system, and the echo measurement coordinate system with the radar antenna position as the origin is used as the ground penetrating radar coordinate system. The color camera coordinate system, depth camera coordinate system, laser scanner coordinate system, and ground penetrating radar coordinate system are uniformly transformed to the vehicle coordinate system through sensor extrinsic parameter calibration and rigid coordinate transformation methods, realizing the spatial position alignment of all sensors under the same spatial reference. The spatiotemporally aligned color images, depth maps, point clouds, echo timing signals, and high-frequency pose sequences are combined to obtain a time-series multi-source dataset.

[0067] S2. Based on a time-series multi-source dataset, the trajectory of the tunnel centerline is reconstructed using vehicle pose to obtain a consistent coordinate framework for the lining structure.

[0068] The high-frequency pose sequences in the temporal multi-source dataset are affected by vehicle vibration and sensor noise. Therefore, within a preset time window, the average value of the three-dimensional position coordinates of the continuous sampling points within the window is calculated, and the average value is used as the smooth position coordinates at the center of the window, thereby eliminating the influence of single abrupt sampling on the trajectory. The spatial position points of the vehicle are connected in chronological order to obtain the preliminary tunnel centerline trajectory. For vehicle positions with slight deviations, a distance-constrained centerline correction method is used to ensure that the centerline maintains geometric consistency with the actual direction of the tunnel.

[0069] The point cloud in the time-series multi-source dataset is processed. Specifically, background and interference points are filtered out based on distance, tunnel wall points are selected based on vertical angle and reflection intensity, and residual isolated points are statistically filtered to obtain a point cloud set containing only the tunnel lining outline. Based on the preliminary tunnel centerline trajectory, a local cross-sectional reference plane perpendicular to the tangential direction of the centerline is constructed at a specified mileage position. The point cloud corresponding to the specified mileage position is uniformly projected onto the local cross-sectional reference plane through cross-sectional projection operation to obtain the cross-sectional point cloud set.

[0070] The cross-sectional point cloud is divided according to the circumferential angle, and the circumferential direction of the cross section is divided into several equal-angle sectors. The radial median of the point cloud in each sector to the center line is calculated to obtain the radius distribution of the cross section in each angular direction. The radial distribution of all mileage positions is continuously connected along the longitudinal direction to obtain the principal surface function of the tunnel lining surface. At each parameter position of the principal surface function, the corresponding unit normal vector is determined by the local surface morphology of the neighborhood of the principal surface function. Based on the local geometric changes of the principal surface function, the curvature field is calculated. The curvature field includes the principal curvature, the secondary curvature and their corresponding directions. Based on the principal surface function, the unit normal vector and the curvature field, a consistent coordinate framework for the lining structure is constructed.

[0071] S3. The time series data is structured and distortion corrected by using the lining structure consistency coordinate framework to obtain a multimodal fusion dataset.

[0072] Spatial back projection is performed on color images, depth images, point clouds, and echo time-series signals to obtain a set of three-dimensional observation points in the vehicle coordinate system. Using the principal surface function as a geometric reference, the three-dimensional observation points are projected onto the tunnel centerline trajectory, and the corresponding nearest mileage position is taken as the longitudinal parameter. In the local cross-section reference plane constructed at the nearest mileage position, the direction angle between the centerline direction and the three-dimensional observation point is calculated to determine the circumferential parameter. The radial offset is determined based on the distance between the principal surface function and the three-dimensional observation point in the normal direction. The longitudinal parameter, circumferential parameter, and radial offset are combined to obtain the structural parameter triplet.

[0073] Based on the normal direction of the principal surface function, dynamic fusion correction is performed on the local offset of the point cloud. Depth camera measurements at a certain viewing angle will produce tilt deviation, so depth compensation is performed to address this tilt deviation. Ground penetrating radar echo intensity decreases with distance and medium, so energy compensation is performed based on the structural distance. The expression is:

[0074] ;

[0075] ;

[0076] ;

[0077] in, Represents a three-dimensional observation point. This represents the corrected three-dimensional observation points. This represents the point cloud error correction coefficient. Indicates the meridional offset. Represents the unit normal vector. This represents the depth offset correction factor. Represents a curvature field. This represents the corrected radial offset. Indicates the echo energy value. This represents the energy attenuation compensation coefficient. This represents the corrected echo energy value.

[0078] It should be noted that the echo energy value It is obtained by performing envelope calculations on the echo timing signal and squaring and integrating the envelope amplitude within the target time window; The method involves collecting ground-penetrating radar echo signals within a calibration area with known lining materials, thickness, and dielectric properties, recording the attenuation trend of echo energy as the incident distance changes, and obtaining the energy attenuation compensation coefficient by fitting the attenuation curve with an exponential function. In front of a standard planar target with known geometry, depth maps are acquired by a depth camera from different angles and postures. The offset error between the depth measurement value and the true depth is calculated, and the law of error variation with incident angle and surface curvature is statistically analyzed. Then, the offset error distribution is curve-fitted to obtain the depth offset correction coefficient. The process involves acquiring laser point cloud data in front of a standard cylindrical or planar target with known geometric dimensions and surface normals, at different postures and distances. The acquired point cloud is then projected onto the actual surface of the target, the offset error of the point cloud points in the normal direction is calculated, and the distribution of the offset error as a function changes with the ranging conditions is fitted to obtain the point cloud error correction coefficient.

[0079] The corrected 3D observation points, corrected radial offset, and corrected echo energy values ​​are combined to obtain a multimodal fusion dataset.

[0080] S4. Based on the lining structure consistency coordinate frame, the tunnel structure is divided into two-dimensional structural meshes. Based on the multimodal fusion dataset, the structural consistency multimodal feature set is extracted on the two-dimensional structural mesh.

[0081] The structural mesh is generated based on the lining structure consistency coordinate frame. Specifically, the tunnel is divided into a series of continuous longitudinal strip regions along the longitudinal parameter axis at fixed mileage intervals. Then, the lining perimeter is divided into multiple adjacent fan-shaped strip regions along the circumferential parameter axis at equal angular intervals. Since the main surface is at a position with zero radial offset in the lining structure consistency coordinate frame, the longitudinal and circumferential strip regions intersect on the main surface, forming a two-dimensional structural mesh composed of regular rectangular grids.

[0082] A multimodal fusion dataset is collected within each structural unit of the two-dimensional mesh. To characterize the structural geometry of the local lining region, a geometric discretization feature based on the deviation between the corrected point cloud and the main surface is performed, expressed as:

[0083] ;

[0084] in, Indicates the first The vertical grid and the first Geometric features of structural elements within the intersection region of a circumferential grid This indicates the number of observation points within a structural unit. Represents the first structural unit The corrected radial offset in each three-dimensional observation point.

[0085] By jointly back-projecting color and depth images onto the vehicle coordinate system using camera intrinsic and extrinsic parameters and depth information, three-dimensional observation points with color attributes are obtained. These three-dimensional observation points with color attributes are mapped to the corresponding structural parameter positions according to the lining structure consistency coordinate framework. Within each structural grid cell, the color information is resampled according to its corresponding longitudinal and circumferential parameters to obtain texture color values. The grayscale mean, grayscale variance, and local texture contrast are extracted from the texture color values ​​within each structural grid cell to obtain texture features. The mean, extreme values, and normal offset gradient of the corrected radial offset are extracted as depth features. Based on the corrected echo energy values, the local energy mean, the energy change rate through differential neighboring cells, and the local energy anomaly index are extracted as energy features. The geometric features, texture features, depth features, and energy features are combined to obtain structural consistency multimodal features.

[0086] It should be noted that the local energy anomaly index is obtained by comparing the difference between the corrected echo energy within a structural unit and the energy baseline of adjacent structural units, and then normalizing the difference.

[0087] S5. Based on the unified fusion of structural consistency multimodal features, calculate the structural consistency failure index, generate a structural failure distribution map based on the structural consistency failure index, and extract suspected disease areas from the structural failure distribution map.

[0088] A structurally intact tunnel section without significant defects is selected as a reference section. Within the reference section, the average levels of geometric, texture, depth, and energy features are statistically analyzed, and the natural fluctuation ranges of these features within the reference section are calculated to obtain the feature baseline values ​​and fluctuation ranges. The deviations of the geometric, texture, depth, and energy features from the baseline values ​​are calculated to obtain geometric anomalies, texture anomalies, depth anomalies, and energy anomalies. Specifically, the structural consistency multimodal features are compared with the baseline values, and the differences are calculated. The differences are normalized according to the fluctuation range so that the degree of deviation reflects the abnormal amplitude of the structural consistency multimodal features relative to the normal level. The larger the difference, the higher the degree of deviation. The normalized data are used as the geometric anomalies, texture anomalies, depth anomalies, and energy anomalies.

[0089] The geometric anomaly, texture anomaly, depth anomaly, and energy anomaly are unified and fused to calculate the structural consistency violation index, expressed as:

[0090] ;

[0091] in, Indicates the structural consistency violation index. Indicates geometric anomaly degree. Indicates the depth anomaly degree. Indicates the degree of energy anomaly. Indicates texture anomaly degree. Indicates the modal amplification factor. Indicators representing structural stability This represents the stability modulation coefficient.

[0092] It should be noted that the modal amplification factor This involves selecting disease-free sections and typical sections confirmed as having true diseases from existing historical testing data. The joint distribution of texture anomalies and energy anomalies in both types of sections is statistically analyzed. By comparing the differences in the simultaneous occurrence of texture and energy anomalies between the two types of sections, the coefficient value that significantly amplifies the joint anomaly in diseased sections while maintaining a low response in normal sections is selected as the modal amplification coefficient; structural stability index Within the coordinate framework of the lining structure, a certain range of neighborhood locations are selected around a specific spatial location. The distributions of geometric anomalies, texture anomalies, depth anomalies, and energy anomalies in the longitudinal and circumferential directions are measured, and the variance of the anomalies within the neighborhood is statistically analyzed. When the anomalies within the neighborhood change relatively smoothly with space and the values ​​are similar, Take a larger value when the four types of anomalies in the neighborhood vary drastically and show significant differences with spatial variation. Take the smaller value; stability modulation coefficient By analyzing the typical value range of structural stability index in disease-free areas, a modulation coefficient that can effectively suppress isolated noise without weakening the response of continuous abnormal areas is selected as the stability modulation coefficient.

[0093] Using the longitudinal and circumferential parameters of each structural unit as planar coordinates, and mapping the structural consistency failure index corresponding to the structural unit to color or grayscale values, a two-dimensional distribution map is drawn on the unfolded planes of the longitudinal and circumferential positions to obtain a structural failure distribution map reflecting the degree of structural failure throughout the entire tunnel.

[0094] An adaptive threshold segmentation method is used to determine the segmentation threshold of the structural consistency failure index. Specifically, statistical analysis is performed on the structural consistency failure index across the entire tunnel, and a histogram distribution of the failure index is constructed. Threshold sequences are obtained by traversing possible segmentation positions within the range of the structural consistency failure index at a fixed step size. The intra-class variance and inter-class variance under the threshold sequences are calculated. Based on the optimization criterion of minimizing the intra-class variance, a global adaptive threshold is selected from the threshold sequences. Structural units with a structural consistency failure index higher than the global adaptive threshold are marked as suspected defective units, while structural units with a structural consistency failure index lower than the global adaptive threshold are marked as normal units. Connectivity analysis and morphological screening are performed on suspected defective units to remove isolated noise units and retain spatially continuous suspected defective regions.

[0095] S6. Extract multimodal physical evidence from suspected disease areas, perform self-verifying multimodal cross-validation on suspected disease areas using multimodal physical evidence, calculate disease credibility index, and obtain disease type.

[0096] Evidence of texture consistency disruption is extracted by: Specifically, local statistical analysis of texture color values ​​around suspected disease areas, comparing grayscale change rates, local contrast, and texture gradients with those of normal reference areas to obtain the amount of texture consistency disruption; evidence of geometric morphological offset is extracted by: statistical analysis of the corrected radial offset of suspected disease areas and their surroundings, comparing the offset trend with the local structural baseline, calculating the positional difference between the corrected radial offset value at the current location and the local structural baseline, and observing the upward or downward trend of the offset with space by combining the change amplitude of radial offset within the neighborhood to obtain geometric morphological offset as evidence; and the extraction of internal reflection changes as evidence of energy reflection changes by comparing the difference between the corrected echo energy value at the suspected disease area and the neighborhood energy baseline.

[0097] It should be noted that the local structural reference shape is determined by selecting a neighborhood range in the longitudinal and circumferential directions with the current position as the center within the lining structure consistency coordinate frame, extracting the positions of the main surface parameters and the corrected three-dimensional observation points within the neighborhood, and constructing a smooth surface that is closest to the overall shape of the neighborhood through local polynomial fitting as the local structural reference shape.

[0098] Self-verifying multimodal cross-validation was conducted using evidence of texture consistency disruption, geometric shift, and energy reflection changes to construct a disease reliability index. This index is used to comprehensively evaluate the multimodal consistency of texture disruption, geometric shift, and energy changes. The expression is as follows:

[0099] ;

[0100] in, This indicates the reliability index of the disease. This indicates evidence of texture inconsistency. This indicates evidence of geometric shape deviation. This indicates evidence of changes in energy reflection. This represents the uniformity modulation coefficient.

[0101] It should be noted that the uniformity modulation coefficient This involves calculating evidence of texture consistency disruption, geometric morphology shift, and energy reflection changes on existing normal areas and actual diseased areas, then substituting these into the disease confidence index for calculation, and applying different values... Multiple comparisons were conducted to select the disease locations that significantly increased the reliability of actual lesion locations compared to normal locations, and that also suppressed single-modal anomalies. As a uniform modulation coefficient.

[0102] To determine the type of disease, specifically, when the texture is significantly damaged and the geometric shift and energy change are both weak, it is determined to be surface cracks or peeling; when the geometric shift is prominent, the texture is generally damaged, and the energy change is not strong, it is determined to be bulging, deformation, or flaking; when the energy change is significant and the texture and geometric features are relatively normal, it is determined to be internal voids or internal cracks; when the texture is significantly damaged, the geometric shift is prominent, and the energy change is significant, it is determined to be a complex disease.

[0103] S7. Based on the disease credibility index and disease type, the suspected disease area is mapped to the lining structure consistency coordinate frame, and the three-dimensional boundary is extracted and geometrically quantified to obtain structured disease information.

[0104] Within the coordinate framework of the lining structure, the suspected defect areas are rasterized, so that each raster position corresponds to a clear parameter coordinate and defect mark. Connectivity analysis is performed on the raster mark matrix. Using the four-neighbor connectivity or eight-neighbor connectivity rule, the raster positions that are adjacent to each other in the longitudinal and circumferential directions and whose defect marks are all suspected defects are grouped into the same connected region, forming a continuous set of raster blocks. The longitudinal and circumferential parameter ranges of each raster block are recorded. The longitudinal and circumferential parameters of each raster block are converted into a set of three-dimensional coordinate points on the main surface through the main surface function, forming a three-dimensional surface distribution of suspected defects on the main surface of the lining.

[0105] On the lining structure consistency coordinate frame, the three-dimensional surface distribution of suspected defects is analyzed by connected region analysis and contour extraction. The boundary polyline of the outer contour of the defect in the two-dimensional unfolded plane is obtained by boundary tracking. Each parameter point on the boundary polyline is mapped to the three-dimensional space of the lining main surface to obtain the three-dimensional boundary line of the defect.

[0106] The spatial location and length of the defects are quantified by tracing the tunnel centerline. Specifically, the starting and ending mileages of the defects along the tunnel direction are obtained by calculating the projection of the closest point between the three-dimensional boundary line of the defects and the tunnel centerline. These mileages serve as the longitudinal location and length of the defects. At multiple mileage locations, lining sections along the circumferential direction are intercepted, and the maximum chord length of the defect boundary within each section is calculated. The chord lengths are then statistically analyzed to obtain the width index representing the transverse scale of the defects.

[0107] On the lining structure consistency coordinate frame, the diseased area is divided into regular grids. Each grid block or triangle is mapped to the three-dimensional space of the lining main surface. The area of ​​the corresponding local face is calculated. The areas of all face are summed to obtain the total surface area covered by the disease.

[0108] The radial damage extent of the affected area was statistically analyzed, and the equivalent depth of the disease was calculated using the following expression:

[0109] ;

[0110] in, Indicates the equivalent depth of the disease. Indicates the number of grid cells. Indicates the first The reliability index of disease at each grid location. Indicates the first Corrected radial offset for each grid position Indicates the first The corresponding area of ​​each grid position on the lining main curved surface.

[0111] The product of the total surface area covered by the disease and the equivalent depth of the disease is used as the disease volume. The disease volume, the width index representing the lateral scale of the disease, the disease type, and the disease reliability index are combined to obtain structured disease information.

[0112] This embodiment also provides an automatic identification system for tunnel lining defects based on multi-source data fusion, including: a data acquisition and alignment module, which acquires time-series data, performs spatiotemporal alignment, and obtains a time-series multi-source dataset;

[0113] The centerline reconstruction module, based on a time-series multi-source dataset, uses vehicle pose to reconstruct the tunnel centerline trajectory and obtain a consistent coordinate framework for the lining structure.

[0114] The structural correction module performs structured projection and distortion correction on time-series data through the lining structure consistency coordinate framework to obtain a multimodal fusion dataset;

[0115] The mesh feature module divides the tunnel structure based on the lining structure consistency coordinate framework, obtains a two-dimensional structural mesh, and extracts a structural consistency multimodal feature set on the two-dimensional structural mesh based on the multimodal fusion dataset;

[0116] The damage assessment module integrates multimodal features of structural consistency to calculate the structural consistency damage index, generates a structural damage distribution map based on the structural consistency damage index, and extracts suspected disease areas from the structural damage distribution map.

[0117] The trusted verification module extracts multimodal physical evidence in suspected disease areas, performs self-verifying multimodal cross-verification of suspected disease areas using multimodal physical evidence, calculates the disease trust index, and obtains the disease type.

[0118] The three-dimensional quantization module maps suspected disease areas to the lining structure consistency coordinate frame based on the disease confidence index and disease type, performs three-dimensional boundary extraction and geometric quantization of the disease, and obtains structured disease information.

[0119] In summary, this invention achieves the organized expression and structured alignment of cross-modal data under a unified geometric reference by constructing a lining structure consistency coordinate framework; by proposing a multimodal structural consistency violation index, it integrates four types of physical features—geometric offset, texture violation, depth anomaly, and energy attenuation—to achieve comprehensive quantification of the nature of the disease; and by introducing a self-verifying multimodal cross-validation mechanism, it constructs a disease credibility index by combining evidence of texture consistency violation, geometric shape offset, and energy reflection change, thereby achieving automatic judgment of the authenticity, type, and level of the disease.

[0120] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automatic identification method for tunnel lining defects based on multi-source data fusion, characterized in that: include, Collect time-series data, perform spatiotemporal alignment, and obtain a multi-source time-series dataset; Based on a time-series multi-source dataset, the trajectory of the tunnel centerline is reconstructed using vehicle pose, and a consistent coordinate framework for the lining structure is obtained. By using a lining structure consistency coordinate framework to perform structured projection and distortion correction on time-series data, a multimodal fusion dataset is obtained. The tunnel structure is divided based on the lining structure consistency coordinate framework to obtain a two-dimensional structural mesh. The structural consistency multimodal feature set is then extracted from the two-dimensional structural mesh based on the multimodal fusion dataset. Based on the unified fusion of structural consistency multimodal features, the structural consistency failure index is calculated, and a structural failure distribution map is generated according to the structural consistency failure index. Suspected disease areas are then extracted from the structural failure distribution map. Multimodal physical evidence is extracted from suspected disease areas. Self-verifying multimodal cross-validation of suspected disease areas is performed using multimodal physical evidence to calculate the disease credibility index and obtain the disease type. Based on the disease confidence index and disease type, suspected disease areas are mapped to the lining structure consistency coordinate frame, and the three-dimensional boundary of the disease is extracted and geometrically quantified to obtain structured disease information. The specific steps for reconstructing the tunnel centerline trajectory using vehicle pose based on a time-series multi-source dataset and obtaining a consistent coordinate framework for the lining structure are as follows: The high-frequency pose sequences in the time-series multi-source dataset are smoothed to obtain smooth position trajectories. The smooth position trajectories are then connected in time order to obtain the preliminary tunnel centerline trajectory. Based on the preliminary tunnel centerline trajectory, the slight deviation of the vehicle in a local position is corrected by distance constraint to obtain the tunnel centerline; Distance filtering, angle selection, and statistical filtering are performed on laser point clouds in a time-series multi-source dataset to obtain an effective set of point clouds. A local cross-section reference plane is constructed based on the tunnel centerline, and the effective point cloud set is projected onto the local cross-section reference plane to obtain the cross-section point cloud set; By constructing the principal surface function through the radial distribution of cross-sectional point clusters in the circumferential direction, and determining the unit normal vector and curvature field based on the local geometric changes of the tunnel lining principal surface, a consistent coordinate framework for the lining structure is constructed. The process involves mapping suspected defect areas to a lining structure consistency coordinate frame based on defect confidence index and defect type, extracting three-dimensional boundaries and performing geometric quantization of the defects to obtain structured defect information. The specific steps are as follows: The suspected defect area is rasterized within the lining structure consistency coordinate frame to obtain the three-dimensional surface distribution of the suspected defect. Perform connected region analysis and contour extraction on the three-dimensional surface distribution of suspected defects, obtain the three-dimensional boundary line of the defects by boundary tracking, and calculate the width index by the tunnel centerline trajectory. The diseased area is regularly discretized within the lining structure consistency coordinate frame to obtain the diseased surface area. Radial damage statistics are performed using the disease confidence index and the corrected radial offset to calculate the equivalent depth of the disease. By combining disease equivalence depth, width indicators, disease type, and disease credibility index, structured disease information is obtained.

2. The method for automatic identification of tunnel lining defects based on multi-source data fusion as described in claim 1, characterized in that: The specific steps for collecting time-series data, performing spatiotemporal alignment, and obtaining a multi-source time-series dataset are as follows: Color images, depth maps, laser point clouds, echo timing signals, and high-frequency pose sequences are acquired. Time alignment is performed using a unified trigger timestamp. The time-aligned color images, depth maps, laser point clouds, echo timing signals, and high-frequency pose sequences are then bound to coordinates to obtain a time-series multi-source dataset.

3. The method for automatic identification of tunnel lining defects based on multi-source data fusion as described in claim 2, characterized in that: The step of performing structured projection and distortion correction on time-series data using a lining structure consistency coordinate framework to obtain a multimodal fusion dataset is as follows: Spatial back projection is performed on the time-series multi-source dataset to obtain three-dimensional observation points. The longitudinal parameters, circumferential parameters and radial offset of the three-dimensional observation points are determined by the principal surface function to obtain the structural parameter triplet. Based on the normal direction of the master surface function, point cloud error correction, depth tilt compensation and energy attenuation compensation are performed on the structural parameter triplet to obtain the corrected three-dimensional observation point, the corrected radial offset and the corrected echo energy value. The corrected 3D observation points, corrected radial offset, and corrected echo energy values ​​are uniformly expressed within the lining structure consistency coordinate framework to obtain a multimodal fusion dataset.

4. The method for automatic identification of tunnel lining defects based on multi-source data fusion as described in claim 3, characterized in that: The tunnel structure is divided into two-dimensional structural meshes based on the lining structure consistency coordinate framework. Then, a structural consistency multimodal feature set is extracted from the two-dimensional structural mesh using a multimodal fusion dataset. The specific steps are as follows: Based on the lining structure consistency coordinate frame, the tunnel is divided at fixed mileage intervals on the longitudinal parameter axis, and the lining perimeter is divided at equal angular intervals on the circumferential parameter axis to obtain a two-dimensional structural mesh. A multimodal fusion dataset is collected at each structural location of a two-dimensional mesh, and geometric features, texture features, depth features, and energy features are extracted. Geometric features, texture features, depth features, and energy features are combined according to the lining structure consistency coordinate framework to obtain a multimodal feature set of structure consistency.

5. The method for automatic identification of tunnel lining defects based on multi-source data fusion as described in claim 4, characterized in that: The process involves unified fusion of multimodal features related to structural consistency, calculation of a structural consistency failure index, and generation of a structural failure distribution map based on this index. The specific steps are as follows: Based on the statistical feature benchmark value of the structural consistency multimodal feature set, the deviation of each structural position from the benchmark value is calculated to obtain the geometric anomaly, texture anomaly, depth anomaly and energy anomaly. The geometric anomaly, texture anomaly, depth anomaly and energy anomaly are integrated to obtain the structural consistency violation index; The structural integrity failure index is mapped onto a two-dimensional unfolded plane according to the longitudinal and circumferential parameters of the structural elements, and a structural failure distribution map is drawn.

6. The method for automatic identification of tunnel lining defects based on multi-source data fusion as described in claim 5, characterized in that: The specific steps for extracting suspected disease areas from the structural damage distribution map are as follows: Adaptive threshold segmentation is performed on the structural damage distribution map by constructing a damage index histogram and traversing candidate thresholds to select the optimal threshold; Structural locations above the optimization threshold are marked as suspected defects, while structural locations below the optimization threshold are marked as normal locations. Connectivity analysis and morphological screening were performed on the locations marked as suspected diseases to obtain suspected disease areas.

7. The method for automatic identification of tunnel lining defects based on multi-source data fusion as described in claim 6, characterized in that: The steps involve extracting multimodal physical evidence from suspected disease areas, performing self-verifying multimodal cross-validation on the suspected disease areas using this evidence, calculating a disease confidence index, and determining the disease type. Evidence of texture inconsistency disruption, geometric morphology shift, and energy reflection changes was extracted from suspected diseased areas. The evidence of texture consistency disruption, geometric shape shift, and energy reflection change is subjected to self-verifying multimodal cross-validation to calculate the disease credibility index. The disease type is identified by using the disease credibility index.

8. An automatic identification system for tunnel lining defects based on multi-source data fusion, based on the automatic identification method for tunnel lining defects based on multi-source data fusion as described in any one of claims 1 to 7, characterized in that: This includes a data acquisition and alignment module, which collects time-series data, performs spatiotemporal alignment, and obtains a multi-source time-series dataset. The centerline reconstruction module, based on a time-series multi-source dataset, uses vehicle pose to reconstruct the tunnel centerline trajectory and obtain a consistent coordinate framework for the lining structure. The structural correction module performs structured projection and distortion correction on time-series data through the lining structure consistency coordinate framework to obtain a multimodal fusion dataset; The mesh feature module divides the tunnel structure based on the lining structure consistency coordinate framework, obtains a two-dimensional structural mesh, and extracts a structural consistency multimodal feature set on the two-dimensional structural mesh based on the multimodal fusion dataset; The damage assessment module integrates multimodal features of structural consistency to calculate the structural consistency damage index, generates a structural damage distribution map based on the structural consistency damage index, and extracts suspected disease areas from the structural damage distribution map. The trusted verification module extracts multimodal physical evidence in suspected disease areas, performs self-verifying multimodal cross-verification of suspected disease areas using multimodal physical evidence, calculates the disease trust index, and obtains the disease type. The three-dimensional quantization module maps suspected disease areas to the lining structure consistency coordinate frame based on the disease confidence index and disease type, performs three-dimensional boundary extraction and geometric quantization of the disease, and obtains structured disease information.

Images