Data acquisition and analysis method and system for tunnel structure apparent disease detection

Abstract

This invention discloses a data acquisition and analysis method and system for detecting apparent defects in tunnel structures. The method includes: simultaneously scanning the tunnel lining surface using visible light, infrared thermal imaging, and three-dimensional point cloud to obtain a spatiotemporally aligned multimodal raw dataset; performing point cloud registration and multi-source data fusion processing on the multimodal raw dataset to construct a three-dimensional model of the lining surface with temperature attributes; based on the three-dimensional model, using a dual-channel convolutional neural network to extract visible light crack features and infrared leakage features, respectively, and outputting candidate defect maps; performing spatial topological structure analysis on the candidate defect maps to generate a set of quantitative indicators; inputting the set of quantitative indicators into a spatiotemporally evolving hidden Markov model to determine the immediate risk level and evolution trend report of each defect area, and associating it with the spatial coordinates of the three-dimensional model. Using this invention, defect risks can be dynamically assessed and development trends predicted, improving detection efficiency and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically a data acquisition and analysis method and system for detecting apparent defects in tunnel structures. Background Technology

[0002] As tunnels age, the need for detecting surface defects in the lining structure (such as cracks, spalling, and water seepage) is becoming increasingly prominent. Traditional manual inspections are inefficient and highly subjective, while automated detection methods using single sensors (such as visible light cameras or infrared thermal imagers) are insufficient to comprehensively characterize the geometric features and physical mechanisms of defects. Existing technologies often face problems such as insufficient spatiotemporal alignment accuracy, weak cross-modal feature correlation, and a lack of quantitative prediction of defect evolution trends. Furthermore, the collaborative analysis of 3D point clouds and infrared temperature fields has not yet been effectively applied to dynamic risk assessment of defects. Summary of the Invention

[0003] The purpose of this invention is to provide a data acquisition and analysis method and system for detecting apparent defects in tunnel structures, so as to overcome the shortcomings of the prior art, dynamically assess the risk of defects and predict their development trend, and improve detection efficiency and accuracy.

[0004] One embodiment of this application provides a data acquisition and analysis method for detecting apparent defects in tunnel structures, the method comprising: According to the preset tunnel detection path, a mobile acquisition platform equipped with a multispectral imaging module and lidar is used to simultaneously scan the tunnel lining surface with visible light, infrared thermal imaging and three-dimensional point cloud to obtain a spatiotemporally aligned multimodal raw dataset. The multimodal raw dataset is subjected to point cloud registration and multi-source data fusion processing. Based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space, a three-dimensional model of the lining surface with temperature attributes is constructed. Based on the three-dimensional model, a dual-channel convolutional neural network is used to extract visible light crack features and infrared leakage features respectively. Cross-modal correlation analysis is performed through a feature cascade module to output a candidate disease map containing cracks, spalling, and water seepage areas. Spatial topology analysis is performed on the candidate disease map. Based on the geometric continuity, depth gradient and temperature distribution of the disease area, a set of quantitative indicators including crack length-weighted curvature, spalling area-to-volume ratio and seepage path thermodynamic parameters is generated. The set of quantitative indicators is input into the spatiotemporal evolution hidden Markov model. Based on the degradation patterns in the historical disease database, the instantaneous risk level and evolution trend report of each disease area are determined and associated with the spatial coordinates of the three-dimensional model.

[0005] Optionally, according to a preset tunnel detection path, a mobile acquisition platform equipped with a multispectral imaging module and lidar performs simultaneous scanning of the tunnel lining surface using visible light, infrared thermal imaging, and three-dimensional point clouds to obtain a spatiotemporally aligned multimodal raw dataset, including: The raw three-dimensional point cloud data is generated by scanning the tunnel lining surface with a lidar mounted on a mobile platform. The laser interferometric positioning base station is used to measure the platform's pose offset in real time, and the gyroscope is used to correct the inertial navigation drift, outputting a high-precision spatial pose matrix. The spatial pose matrix triggers the synchronous acquisition of visible light camera and infrared thermal imager to generate a spatiotemporally aligned set of visible light images and infrared temperature field sequence. By fusing the original 3D point cloud, visible light image set, and infrared temperature field sequence, a multimodal original dataset aligned with spatiotemporal reference is output.

[0006] Optionally, the step of performing point cloud registration and multi-source data fusion processing on the multimodal raw dataset, and constructing a three-dimensional model of the lining surface with temperature attributes based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space, includes: LiDAR point clouds are extracted from the multimodal raw dataset, and curvature adaptive filtering is performed to output the denoised geometric point cloud. The thermal expansion coefficient of the lining material is calculated based on the infrared temperature field, and a temperature-deformation compensation vector field is generated. The temperature-deformation compensation vector field is applied to the geometric point cloud, and the visible light texture coordinates are registered using a non-rigid ICP algorithm to output a geometrically corrected 3D mesh model. The infrared temperature field is mapped to the vertices of the 3D mesh model, and the initial 3D model with temperature attributes is output. Temperature field gradient optimization is performed on the initial 3D model. An MLP network is used to enhance the temperature continuity of the curved surface, and a gradient-optimized 3D model of the lining surface is output.

[0007] Optionally, based on the three-dimensional model, a dual-channel convolutional neural network is used to extract visible light crack features and infrared leakage features respectively. Cross-modal correlation analysis is then performed through a feature cascade module to output a candidate disease map containing cracks, spalling, and seepage areas, including: The gradient-optimized 3D model is projected to generate visible light texture unfolded map and infrared temperature unfolded map, and a bimodal 2D unfolded map is output. A dual-channel CNN is used to process the two-dimensional unfolded graph of the two modes respectively. The visible light channel outputs the crack feature tensor, and the infrared channel outputs the leakage feature tensor. Input the crack feature tensor and the leakage feature tensor into the cross-attention module to generate an association-enhanced feature map; The association-enhanced feature map is passed through a cascaded decoder to generate a disease probability map. The boundary is optimized by CRF, and the candidate disease map is output.

[0008] Optionally, the spatial topological structure analysis of the candidate disease map is performed, and a set of quantitative indicators including crack length-weighted curvature, spalling area-to-volume ratio, and seepage path thermodynamic parameters are generated based on the geometric continuity, depth gradient, and temperature distribution of the disease area. Extract the crack skeleton line from the candidate disease map, calculate the length-weighted integral of the curvature, and output the crack length-weighted curvature value. The surface area to volume ratio of the exfoliated region is calculated based on a gradient-optimized 3D model, and the exfoliation volume ratio parameter is output. Solve for the temperature gradient vector field along the seepage path and output the thermodynamic seepage direction matrix; Path integration is performed on the thermodynamic permeation direction matrix to generate the seepage thermodynamic attenuation coefficient; A set of structured quantitative indicators is constructed based on crack length-weighted curvature values, spalling volume ratio parameters, and seepage thermodynamic attenuation coefficients.

[0009] Optionally, the step of inputting the set of quantitative indicators into a spatiotemporally evolving hidden Markov model, determining the immediate risk level and evolution trend report of each disease area based on the degradation patterns in the historical disease database, and linking it to the spatial coordinates of the three-dimensional model includes: The quantitative index set is gridded according to the three-dimensional model space coordinates to generate a gridded state vector sequence; The gridded state vector sequence is input into a two-layer hidden Markov model, where the lower HMM outputs the unit state transition probability and the upper HMM outputs the neighborhood propagation probability. Input the unit state transition probability and neighborhood propagation probability into the graph neural network, aggregate the topological risk features, and output the grid risk level distribution map; Based on the grid risk level distribution map and historical database, Bayesian evolution prediction is performed to generate a risk report with three-dimensional spatial coordinate association.

[0010] Another embodiment of this application provides a data acquisition and analysis system for detecting apparent defects in tunnel structures, the system comprising: The scanning module is used to scan the surface of the tunnel lining simultaneously with visible light, infrared thermal imaging and three-dimensional point cloud by a mobile acquisition platform equipped with a multispectral imaging module and lidar, according to the preset tunnel detection path, so as to obtain a spatiotemporally aligned multimodal raw dataset. The construction module is used to perform point cloud registration and multi-source data fusion processing on the multimodal raw dataset, and to construct a three-dimensional model of the lining surface with temperature attributes based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space. The extraction module is used to extract visible light crack features and infrared leakage features based on the three-dimensional model using a dual-channel convolutional neural network. It performs cross-modal correlation analysis through the feature cascade module and outputs a candidate disease map containing cracks, spalling, and water seepage areas. The analysis module is used to perform spatial topological structure analysis on the candidate disease map, and generate a set of quantitative indicators including crack length weighted curvature, peeling area-to-volume ratio, and seepage path thermodynamic parameters based on the geometric continuity, depth gradient, and temperature distribution of the disease area. The determination module is used to input the set of quantitative indicators into the spatiotemporal evolution hidden Markov model, determine the real-time risk level and evolution trend report of each disease area based on the degradation patterns in the historical disease database, and associate it with the spatial coordinates of the three-dimensional model.

[0011] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0012] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0013] Compared with existing technologies, this invention provides a data acquisition and analysis method for detecting apparent defects in tunnel structures. It involves simultaneous scanning of the tunnel lining surface using visible light, infrared thermal imaging, and 3D point clouds to obtain a spatiotemporally aligned multimodal raw dataset. The multimodal raw dataset undergoes point cloud registration and multi-source data fusion processing to construct a 3D model of the lining surface with temperature attributes. Based on the 3D model, a dual-channel convolutional neural network is used to extract visible light crack features and infrared leakage features, outputting candidate defect maps. Spatial topological structure analysis is performed on the candidate defect maps to generate a set of quantitative indicators. This set of quantitative indicators is input into a spatiotemporally evolving hidden Markov model to determine the immediate risk level and evolution trend report for each defect area, and is associated with the spatial coordinates of the 3D model. This allows for dynamic assessment of defect risk and prediction of development trends, improving detection efficiency and accuracy. Attached Figure Description

[0014] Figure 1 Hardware structure block diagram of a computer terminal for a data acquisition and analysis method for detecting apparent defects in tunnel structures, provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a data acquisition and analysis method for detecting apparent defects in tunnel structures, provided in an embodiment of the present invention. Figure 3This is a schematic diagram of a data acquisition and analysis system for detecting apparent defects in tunnel structures, provided in an embodiment of the present invention. Detailed Implementation

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] The present invention first provides a data acquisition and analysis method for detecting apparent defects in tunnel structures. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0017] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a data acquisition and analysis method for detecting apparent defects in tunnel structures, provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0018] The non-volatile storage medium can store an operating system and a computer program. This computer program includes program instructions that, when executed, cause the processor to perform any data acquisition and analysis method for detecting apparent defects in tunnel structures.

[0019] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0020] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to perform any data acquisition and analysis method for detecting apparent defects in tunnel structures.

[0021] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0022] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0023] See Figure 2 The present invention provides a data acquisition and analysis method for detecting apparent defects in tunnel structures, which may include the following steps: S201, according to the preset tunnel detection path, through a mobile acquisition platform equipped with a multispectral imaging module and lidar, performs simultaneous scanning of visible light, infrared thermal imaging and three-dimensional point cloud on the tunnel lining surface to obtain a spatiotemporally aligned multimodal raw dataset. Specifically, the surface of the tunnel lining can be scanned using a lidar mounted on a mobile platform to generate raw three-dimensional point cloud data; The mobile platform uses a tracked or rail-mounted vehicle equipped with a Light Detection and Range (LiDAR) system. This LiDAR typically employs a line-scan or rotating design, emitting laser beams with wavelengths of 905 nm or 1550 nm, and a pulse repetition frequency (PRF) set between 100 Mbps and 2 million Mbps. As the platform moves at a constant speed along a pre-set tunnel detection path, the LiDAR emits laser pulses at a fixed elevation angle (e.g., 30 degrees) towards the tunnel lining surface. After the laser beam contacts the concrete surface, it reflects, and the receiver records the round-trip time (ToF) of each pulse. Combining the laser emission angle and the platform's real-time position, the three-dimensional spatial coordinates (X, Y, Z) of each reflection point are calculated. During the scanning process, the LiDAR's angular resolution (AR) is set to 0.05 degrees to ensure that the distance between adjacent points is within 5 mm. Simultaneously, the radar's built-in intensity sensor records the echo intensity (IV), reflecting the surface material's reflectivity. A single scan generates raw 3D point cloud data containing millions of points, in a structured array format that includes coordinates (X, Y, Z), intensity (IV), and timestamp (TS).

[0024] To address interference from dust and moisture within the tunnel, the lidar is equipped with a Dynamic Gain Control (DGC) module. When the detected echo signal-to-noise ratio (SNR) falls below a preset threshold (e.g., 20 dB), the DGC automatically increases the transmission power or extends the integration time to ensure point cloud integrity. Simultaneously, the platform velocity (PV) and laser scan rate (SR) are dynamically matched: the maximum permissible speed (e.g., 0.5 m / s) is calculated using the formula PV ≤ SR × minimum point spacing to prevent point cloud stretching and deformation. During scanning, the raw point cloud is transmitted in real-time to the onboard processing unit for preliminary filtering: a Statistical Outlier Removal (SOR) algorithm is used, taking 50 neighboring points around each point as samples, and removing noise points deviating from the mean by more than 3 standard deviations.

[0025] To address the curved surface characteristics of the tunnel, the LiDAR employs a multi-beam (e.g., 32- or 64-beam) layered scanning design. Each beam is distributed vertically (VFOV) to cover a range of -30° to +30°, ensuring no blind spots from the arch to the sidewalls. During real-time stitching of the scanned data, a sliding window (SW) mechanism is used: local point clouds are generated in 10-meter segments, and overlapping areas are registered using an Iterative Closest Point (ICP) algorithm, with the registration error (RE) controlled within 2 millimeters. The final output of the raw 3D point cloud data contains millimeter-level precision point sets in absolute coordinates, laying the geometric foundation for subsequent fusion processing.

[0026] The laser interferometric positioning base station is used to measure the platform's pose offset in real time, and the gyroscope is used to correct the inertial navigation drift, outputting a high-precision spatial pose matrix. Laser Interferometric Base Stations (LIBS) are deployed at the tunnel entrance and key nodes along the detection path. Each LIBS emits a 632.8 nm helium-neon laser, forming an interferometric measurement network covering the entire length of the tunnel. A retroreflector target (RT) is mounted on top of the mobile platform, its core consisting of three equilateral triangularly arranged pyramidal prisms. The laser beam is reflected by the prisms, generating interference fringes. A photoelectric detector (PED) within the base station analyzes the phase difference (PD) in real time, and calculates the displacement changes (ΔX, ΔY, ΔZ) of the platform relative to the base station using the laser wavelength, achieving a measurement accuracy of 0.01 mm. Simultaneously, an inclinometer sensor (IS) built into the target measures the pitch angle (PA) and roll angle (RA), with the data transmitted wirelessly to the platform's main control system.

[0027] The platform integrates an Inertial Navigation System (INS), comprising a three-axis fiber optic gyroscope (FOG) and a three-axis accelerometer (ACC). The gyroscope measures angular velocity (AV) in degrees per second; the accelerometer measures linear acceleration (LA) in meters per second. 2Due to the drift error (DE) inherent in the INS, correction is required using laser interferometry data: LIBS displacement data is acquired every 0.1 seconds and compared with the INS integral result to generate an error vector (EV). A Kalman filter (KF) algorithm is used to fuse the two types of data. The KF state variable (SV) includes position, velocity, attitude angle, and sensor bias. During the filtering process, LIBS data is used as observation values ​​(OV) to update the state estimate, suppressing the INS position drift to within 1 centimeter per 100 meters.

[0028] The final output High-Precision Spatial Pose Matrix (HPSPM) is a 4×4 homogeneous transformation matrix containing a 3×3 rotation matrix (RM) and a 3×1 translation vector (TV). The matrix is ​​updated at a frequency of 100 Hz, with a time synchronization accuracy (TSA) of 1 microsecond. This matrix fully describes the platform's six-DOF pose (X, Y, Z, pitch, yaw, roll) in the global coordinate system, providing a spatiotemporal reference for multi-sensor data fusion. The matrix data is stored using a combination of quaternions (Q) and displacement vectors to avoid gimbal lock issues.

[0029] The spatial pose matrix triggers the synchronous acquisition of visible light camera and infrared thermal imager to generate a spatiotemporally aligned set of visible light images and infrared temperature field sequence. When the HPSPM (Hyperspace Pose Matrix) is updated, its timestamp (TS) is simultaneously activated via a hardware trigger signal (HTS) to activate the Visible Light Camera (VLC) and the Infrared Thermal Imager (IRTI). The Visible Light Camera uses a Global Shutter (GS) CMOS sensor with a resolution of 2448×2048 pixels and a pixel size of 3.45 micrometers. After triggering, the camera captures images at a rate of 10 frames per second, with the exposure time (ET) automatically adjusted according to the tunnel lighting (20-100 ms), and the ISO sensitivity (IS) fixed at 800 to suppress noise. A fixed-focus industrial lens (Focal Length, FL=35mm) is used, with the aperture (AP) set to F4.0 to ensure a depth of field coverage from 0.5 meters to infinity.

[0030] The infrared thermal imager employs an uncooled microbolometer (UM) with a resolution of 640×512 pixels and a thermal sensitivity (TS) ≤0.03℃. It operates in the 8-14 micrometer wavelength range, with a temperature measurement range of -20℃ to 150℃. Upon triggering, it synchronously acquires data at 10 Hz, generating infrared temperature field (ITF) data for each frame, including the absolute temperature value (TV) of each pixel. To eliminate environmental reflection interference, the imager incorporates a blackbody reference source (BRS) for real-time non-uniformity correction (NUC). During acquisition, the emissivity (EM) is set to a typical value of 0.93 for concrete, and the distance-to-spot ratio (DSR) is set to 50:1.

[0031] Spatiotemporal alignment (STA) is achieved through a three-level mechanism: Time synchronization: Pose matrix, visible light image, infrared temperature field averaged GPS time (GPST) and internal oscillator time (IOT), time deviation ≤100 microseconds.

[0032] Spatial mapping: Using pre-calibrated camera-radar extrinsic parameters (EP), the pose matrix is ​​converted into pose in the camera coordinate system, and then the 3D point cloud is mapped to the 2D image plane through perspective projection (PP).

[0033] Data Association: Each visible light image frame (IF) and infrared frame (IRF) is associated with its corresponding pose matrix, forming a triplet data structure (timestamp, pose matrix, image / temperature data). The final output Visible Light Image Set (VLIS) and Infrared Temperature Field Sequence (ITFS) have a strict spatiotemporal correspondence.

[0034] By fusing the original 3D point cloud, visible light image set, and infrared temperature field sequence, a multimodal original dataset aligned with spatiotemporal reference is output.

[0035] Fusion Processing (FP) is performed in the vehicle's edge computing unit, and the core process consists of three steps: Step 1: Point cloud-image registration.

[0036] Align the raw 3D point cloud (RPC) with the visible light image (VLIS) using point cloud projection registration (PCPR): Based on camera intrinsic parameters (IP: focal length fx, fy, principal point cx, cy, distortion coefficients k1-k5) and pre-calibrated extrinsic parameters (camera rotation matrix RC and translation vector TC relative to the radar). The point cloud is projected onto the image plane to generate a depth map (DM). The offset vector between the projection point and the image feature point is calculated using the Dense Optical Flow (DOF) algorithm. The Levenberg-Marquardt (LM) optimization algorithm was used to iteratively correct the extrinsic parameters, reducing the reprojection error (RE) to below 0.5 pixels.

[0037] Step 2: Temperature field-point cloud mapping.

[0038] Spatial alignment of the infrared temperature field (ITFS) with the point cloud is achieved through thermal radiation coordinate transformation (TRCT): Based on the thermal imager's intrinsic parameters (IFOV = 0.5 milliradians) and extrinsic parameters (rotation matrix RT and translation vector TT relative to the radar); Transform the point cloud to the thermal imager coordinate system; Establish a bilinear interpolation (BI) mapping relationship between each point in the point cloud and the pixels in the infrared image; Distance Attenuation Correction (DAC) is applied to the temperature field: the temperature value is corrected according to the formula T_corrected = T_raw × (d / d0)^0.25, where d is the measurement distance and d0 is the calibration distance (1 meter).

[0039] Step 3: Constructing a multimodal dataset.

[0040] Output the spatiotemporally aligned multimodal raw dataset (SAMRD), with the following data structure: Geometry layer: Millimeter-precision point cloud, each point contains coordinates (X,Y,Z), echo intensity (IV), RGB color value (from visible light image) and temperature value (T); Image layer: Time-synchronized visible light image sequence (JPEG2000 compression) and infrared temperature field sequence (14-bit RAW format). Pose layer: A high-precision pose matrix (4×4 homogeneous matrix) associated with each frame of data. Metadata layer: sensor parameters (such as camera exposure time, thermal imager emissivity), environmental parameters (temperature and humidity), timestamps (nanosecond-level accuracy); The dataset uses hierarchical data format (HDF5) storage and achieves fast retrieval through spatial index (Octree) and time index (B+Tree), with the total data rate controlled within 200MB / s.

[0041] S202, perform point cloud registration and multi-source data fusion processing on the multimodal original dataset, and construct a three-dimensional model of the lining surface with temperature attributes based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space. Specifically, LiDAR point clouds can be extracted from the multimodal raw dataset, curvature adaptive filtering can be performed, and the denoised geometric point cloud can be output. The system first separates the raw 3D point cloud data acquired by the mobile platform's LiDAR from the spatiotemporally aligned multimodal raw dataset. This point cloud contains millions of spatial coordinate points (XYZ coordinates), but due to the influence of dust, water vapor, and equipment vibration in the tunnel environment, it contains a large number of outlier noise points (ONPs) and surface fluctuation noise (SFN). To eliminate noise, a curvature-adaptive filtering algorithm (CAFA) is adopted. The core of this algorithm is to dynamically adjust the filtering intensity: strong filtering is used in areas with gentle curvature on the lining surface (such as large flat walls), and a large-size filtering window (such as a radius of 30 cm) is used to merge neighboring points; in areas with high curvature (such as crack edges or segment joints), weak filtering is switched to use a small-size window (such as a radius of 5 cm) to preserve geometric details. In practice, the system calculates the Local Curvature Value (LCV) of the neighborhood surrounding each point (e.g., 50 neighboring points). The LCV is obtained by estimating the rate of change of the surface normal vector through Principal Component Analysis (PCA). When the LCV of a point is lower than a threshold (e.g., 0.05 mm -1), it is identified as a low curvature region, triggering the Statistical Outlier Removal (SOR) algorithm to remove points that deviate from the neighborhood average distance by more than 3 standard deviations. When the LCV is higher than the threshold (e.g., 0.2 mm -1), the Radius Filter (RF) is activated, retaining only points whose neighborhood density is higher than a set value. This adaptive process ensures that while eliminating floating noise, the geometric sharpness of the disease features is fully preserved.

[0042] The implementation of curvature adaptive filtering relies on a hierarchical processing architecture. The first layer performs voxel grid downsampling (VGD), dividing the original point cloud into uniform spatial cubic units (voxel side length 5 mm), retaining only one centroid point within each voxel, thus initially reducing the data volume. The second layer performs curvature zoning identification (CZI): based on the downsampled point cloud, it calculates the curvature weight factor (CWF) for each point. The CWF is dynamically generated by comparing the LCV with a preset curvature threshold (range 0~1). The third layer applies curvature-guided bilateral filtering (CGBF). This filtering not only considers spatial proximity (Gaussian distance weight) but also introduces curvature similarity weight: for high CWF regions (rich detail regions), the spatial smoothing intensity is reduced and the curvature constraint weight is increased to avoid edge blurring; for low CWF regions (flat regions), the spatial smoothing intensity is increased. The filtering parameters are adjusted in real time using gradient descent optimization (GDO), with the objective function being to minimize the Hausdorff distance (HD) between the filtered point cloud and the ideal surface. The final output is a denoised geometric point cloud (DGPC) with a pixel pitch error controlled within ±0.3 mm, providing an accurate geometric basis for subsequent modeling.

[0043] To verify the filtering effect, a Noise Suppression Evaluation Module (NSEM) was added to the system. This module imports the point clouds before and after filtering into the Local Surface Reconstruction (LSR) algorithm, generates triangular meshes, and calculates the Average Mesh Ruggedness (AMR). The rate of decrease in AMR value (e.g., from 15% before filtering to 3%) is used as the main evaluation indicator. Simultaneously, Feature Point Displacement Error (FPDE) is calculated at known geometric feature locations (e.g., pre-set calibration plates) to ensure that critical structures are not deformed. All processing is performed on the Edge Computing Unit (ECU), and the processing time for a single 100-meter tunnel point cloud (approximately 20 million points) is no more than 5 minutes, meeting real-time requirements. The output DGPC data format includes XYZ coordinates, normal vectors, and curvature labels for downstream processes.

[0044] The thermal expansion coefficient of the lining material is calculated based on the infrared temperature field, and a temperature-deformation compensation vector field is generated. Due to the thermal expansion and contraction effect of tunnel lining materials (concrete, masonry, etc.), the geometric point cloud acquired by lidar will deviate from the infrared temperature field in terms of physical deformation. To correct this error, the system calculates the dynamic thermal expansion coefficient (TEC) based on the synchronously acquired infrared temperature field sequence (ITFS). The specific process is as follows: First, select data frames from the ITFS during a period of stable temperature (temperature fluctuation less than 0.5 degrees Celsius for more than 10 seconds) and extract the reference temperature value (RTV) of the lining surface, typically using the tunnel's annual average temperature (e.g., 15 degrees Celsius) as the benchmark. Next, for the instantaneous temperature field (ITF) at the current scanning moment, calculate the temperature difference value (TDV) for each pixel region, i.e., ΔT = ITF - RTV. According to the principles of materials physics, the local deformation caused by temperature is proportional to ΔT and TEC.

[0045] The TEC calculation employs a multi-material zoning inversion strategy (MMZIS). Based on visible light image segmentation results, the system divides the lining surface into different material regions (e.g., concrete region, stone repair region, metal bolt region). For each material, its specific heat capacity (SHC) and thermal conductivity (TC) parameters are retrieved from a historical database. The mapping relationship between the temperature field and thermal stress is established using the transient heat conduction equation (THCE), and small displacements in the continuous temperature field are analyzed using digital image correlation (DIC) to deduce the equivalent thermal expansion coefficient (ETEC) for each material region. For example, the typical ETEC value for concrete is 10 × 10⁻⁶. -6 / °C, stone is 8×10 -6 / °C. Finally, a Temperature-Deformation Compensation Vector Field (TDCVF) is generated. This field is a three-dimensional vector field, and each vector represents the displacement compensation in the XYZ directions caused by temperature changes. The calculation formula is: Displacement vector = ETEC × ΔT × Feature length (taken as the local curvature radius of the point cloud).

[0046] Generating TDCVFs requires addressing the spatial interpolation problem. Since the resolution of infrared thermal imagers (e.g., 640×480 pixels) is lower than the point cloud density, the system employs the Kriging Interpolation Method (KIM) to interpolate pixel-level TDV data to each point cloud location. A Temperature Gradient Constraint (TGC) is introduced during interpolation to ensure that areas with a temperature difference greater than 2°C between adjacent points (such as seepage boundaries) do not produce smoothing distortion. The generated TDCVFs are overlaid on the point cloud using a Vector Field Visualization Engine (VFVE), allowing operators to manually correct anomalous vectors (such as abnormal high-temperature deformation in metal component areas). This vector field data is stored as a three-channel floating-point matrix, corresponding one-to-one with the geometric point cloud.

[0047] The temperature-deformation compensation vector field is applied to the geometric point cloud, and the visible light texture coordinates are registered using a non-rigid ICP algorithm to output a geometrically corrected 3D mesh model. The Temperature-Deformation Compensation Vector Field (TDCVF) obtained in the previous step is applied to the denoised Geometric Point Cloud (DGPC). Point-by-point displacement correction is performed: for each point P_i in the DGPC, its corresponding compensation vector V_i is found, and the corrected coordinates P'_i = P_i + V_i are calculated. This step eliminates the geometric deviation caused by temperature deformation, resulting in the Temperature-Compensated Point Cloud (TCPC). Next, the Texture Image Coordinates (TIC) acquired by the visible light camera need to be accurately mapped to the TCPC. Due to the difference in viewing angle between the camera and the LiDAR and lens distortion, non-rigid registration is required.

[0048] Registration employs the Non-Rigid Iterative Closest Point (NR-ICP) algorithm. This algorithm consists of four stages: Feature point extraction: SIFT (Scale-Invariant Feature Transform) feature points are extracted from the visible light image, and corresponding 3D feature points are generated in TCPC (through inverse projection).

[0049] Deformation field initialization: Based on feature point pairs, the initial deformation field is constructed using the Thin Plate Spline (TPS) function.

[0050] Iterative optimization: In each iteration, TCPC is projected onto the image plane to generate virtual 2D points, and the Euclidean distance (ED) between these virtual points and real image feature points is calculated. The deformation field parameters are updated by minimizing the weighted sum of ED and TPS bending energy (BE). The weights are dynamically adjusted by the annealing scheduler (AS), initially emphasizing feature matching and later emphasizing deformation smoothing.

[0051] Convergence criterion: Stop when the rate of change of ED is less than 0.1% for 10 consecutive iterations, and output the optimal deformation field.

[0052] The optimal deformation field is applied to TCPC to obtain a texture-registered point cloud (TRPC). Poisson surface reconstruction (PSR) is then performed based on TRPC to generate a continuous triangular mesh model. During reconstruction, the octree depth (OD=10) is set to control the level of detail, and the smoothing factor (SF=0.4) is used to eliminate staircase effects. The output geometry-corrected 3D mesh model (GC3M) includes vertex coordinates, triangular facets, and texture mapping relationships. The geometric error of the model has been calibrated and verified to be less than 1 mm.

[0053] The infrared temperature field is mapped to the vertices of the 3D mesh model, and the initial 3D model with temperature attributes is output. The temperature field data (TFD) acquired by the infrared thermal imager is mapped to the vertices of the GC3M model generated in the previous step. First, the spatial correspondence between the temperature data and the mesh vertices is established: based on the camera's intrinsic and extrinsic parameters (intrinsic parameters: focal length f=35mm, principal point coordinates; extrinsic parameters: given by the pose matrix), the projection coordinates (PC) of each mesh vertex on the infrared image are calculated. Due to the limited resolution of infrared images (e.g., 320×240 pixels), a single pixel may cover multiple mesh vertices.

[0054] To address the resolution mismatch issue, a bilinear interpolation-weighted allocation (BIWA) strategy is employed. For each vertex V_i, calculate its projection point P_i on the infrared image.

[0055] Take the four pixels surrounding P_i (Q) 11 Q 12 Q 21 Q 22 ), calculate the normalized distance (ND) from P_i to the center of each pixel.

[0056] Weights are calculated inversely proportional to distance: W 11 = (1-dx)(1-dy), where dx and dy are P_i and Q_i, respectively. 11 The difference between the horizontal and vertical coordinates.

[0057] Vertex temperature T_i = Σ(W_ij × T_j), where j is the four neighboring pixels.

[0058] At the same time, a confidence mask (CM) is introduced to mark vertices in the image edge region (confidence <85%) as low reliability.

[0059] After temperature mapping is completed, an initial 3D model with temperature attributes (I3MTA) is generated. This model's data format is an enhanced OBJ file: vertices record XYZ coordinates + temperature values ​​(range -20~50°C, accuracy 0.1°C), faces record vertex indices, and texture coordinates are stored independently. The model is rendered in real-time using a WebGL visualization engine, supporting overlay display of temperature heatmaps, and operators can interactively check the mapping quality.

[0060] Temperature field gradient optimization is performed on the initial 3D model. An MLP network is used to enhance the temperature continuity of the curved surface, and a gradient-optimized 3D model of the lining surface is output.

[0061] The initial model I3MTA's temperature field suffers from two types of problems: first, abrupt temperature changes caused by infrared image noise (e.g., a single vertex temperature being more than 5°C higher than its neighborhood); and second, missing data regions caused by viewpoint occlusion (e.g., the junction of the vault and sidewalls). To address these issues, temperature field gradient optimization is performed, the core of which involves constructing a Temperature Continuity Constraint (TCC) and a Physical Conduction Constraint (PCC).

[0062] The optimization employs a multilayer perceptron network (MLP). The network structure is as follows: Input layer: Vertex coordinates (XYZ) + initial temperature + curvature value + material label (4D); Hidden layers: 3 fully connected layers (128-64-32 neurons), activation function ReLU (Rectified LinearUnit, ReLU); Output layer: 1 neuron (predicts temperature value).

[0063] During training, the entire temperature measurement area is used as the sample set, and the loss function includes: Temperature reconstruction loss (L1 norm): the difference between the predicted temperature and the original temperature; Gradient smoothing loss: Gaussian weighted sum of temperature differences between adjacent vertices; Thermal conduction loss: based on the matching degree between the predicted temperature gradient and the thermal conductivity of the material according to Fourier's law.

[0064] After training, the network can predict the temperature of missing regions and smooth out outliers.

[0065] The optimized temperature values ​​are written to the model vertices to generate a gradient-optimized 3D model (GO3DM) of the lining surface. The optimization effect is evaluated using the root mean square gradient error (RMSGE), with a typical value decreasing from 8.7°C / m before optimization to 2.3°C / m. The final model is stored in a standardized format (such as glTF 2.0) and includes complete geometry, texture, and temperature attributes.

[0066] S203. Based on the three-dimensional model, a dual-channel convolutional neural network is used to extract visible light crack features and infrared leakage features respectively. Cross-modal correlation analysis is performed through the feature cascade module to output a candidate disease map containing cracks, spalling, and water seepage areas. Specifically, the gradient-optimized 3D model can be projected to generate visible light texture unfolded maps and infrared temperature unfolded maps, outputting a bimodal 2D unfolded map; The system receives a gradient-optimized 3D model of the lining surface, processed through point cloud registration, data fusion, temperature attribute mapping, and gradient optimization. This model is a high-precision 3D mesh structure, where each mesh vertex contains not only spatial coordinates (X, Y, Z) but also precisely correlated visible light texture color values ​​(RGB) and optimized infrared temperature values ​​(TV). To facilitate subsequent deep learning-based 2D image processing, this 3D surface model needs to be "unfolded" into a planar image. The core of this step is projection. The system employs a geodesic distance-based adaptive cylindrical projection algorithm. First, based on the tunnel centerline or the actual acquired trajectory of the moving platform, a virtual projection cylinder is determined. The radius of the projection cylinder is slightly larger than the maximum radius of the tunnel lining surface to ensure that all surface points can be projected. Then, for each vertex on the 3D mesh model, its radial distance (RD) to the central axis and its position along the central axis (AP) are calculated. Simultaneously, the angular position (AngP) of that point on the annular cross-section perpendicular to the central axis is calculated. By mapping the 3D spatial coordinates (X, Y, Z) of each vertex to the 2D planar coordinates (U, V) consisting of the axial position (AP) and the angular position (AngP), the 3D surface can be mapped to a 2D plane.

[0067] During projection, to generate the Visible Texture Unfolded Image (VTUI), the system fills the corresponding pixel positions with the visible texture color (RGB value) associated with each mesh vertex according to the mapped two-dimensional coordinates (U, V). Since the tunnel is a long, narrow structure, the axial position (AP) corresponds to the long side (Width, W) of the image, and the angular position (AngP) corresponds to the short side (Height, H). Because the tunnel surface is a closed curved surface, the angular direction (AngP) covers the entire circumference from 0 degrees to 360 degrees; therefore, the generated VTUI is connected end-to-end in the angular direction. To ensure image quality, the system uses bilinear interpolation to process the pixel values ​​within the mesh patches, avoiding jagged edges. For high-curvature areas (such as construction joints or near pipe segment joints), the algorithm adaptively adjusts the projection density based on the local curvature (LC) of the mesh, increasing sampling points in areas of high curvature to ensure that detailed textures (such as fine crack textures) are clearly visible in the unfolded image. Simultaneously, an Infrared Temperature Unfolded Image (ITUI) is generated. Its generation principle is the same as VTUI, but instead of color values, it fills in the gradient-optimized temperature value (TV) corresponding to each vertex. The system typically maps the temperature values ​​to a pseudocolor scale (e.g., from blue representing low temperature to red representing high temperature), or directly saves them as a grayscale image, where the pixel intensity (I) directly corresponds to the temperature value (higher temperature, brighter). The final output dual-modal 2D unfolded images are a pair of strictly spatially aligned images: VTUI and ITUI. Each pixel in their 2D plane coordinates (U, V) precisely corresponds to the same physical location point on the original 3D model, and respectively carries the visual appearance information and temperature information of that point. This pair of images provides a precise and unified input benchmark for subsequent dual-channel feature extraction.

[0068] A key challenge in the projection process lies in handling the irregularities and potential local deformations (such as convergence deformation) of the tunnel surface. Simple cylindrical projection can produce severe distortion in certain areas, particularly the arch and arch base. To address this, the system embeds an adaptive correction module based on the geometric properties of the 3D mesh model itself into the projection algorithm. This module utilizes the triangular facet information of the 3D mesh to calculate the area distortion ratio (ADR) or angular distortion ratio (AngDR) of each facet during the projection process. When the distortion ratio of a local area exceeds a preset threshold (e.g., ADR > 1.5 or AngDR > 15 degrees), the system dynamically introduces additional control points (CPs) or applies local non-rigid transformations, such as thin plate spline (TPS) interpolation, to fine-tune the projection mapping of that area, minimizing the impact of deformation on subsequent image analysis and spatial positioning accuracy. Furthermore, to ensure the integrity of the entire unfolded image, the system performs special processing on image boundaries (especially at the beginning and end points in the angular direction), such as copying edge pixels or applying smoothing transitions, to avoid generating discontinuity artifacts at the boundaries. After these processing steps, the final output bimodal 2D unfolded images (VTUI and ITUI) are not only complete in information and clear in detail, but also have precise spatial correspondence, laying a reliable foundation for the next step of feature extraction.

[0069] A dual-channel CNN is used to process the two-dimensional unfolded graph of the two modes respectively. The visible light channel outputs the crack feature tensor, and the infrared channel outputs the leakage feature tensor. After obtaining accurately registered bimodal 2D unfolded images (VTUI and ITUI), the system uses a specially designed dual-channel convolutional neural network (DC-CNN) to process these two images in parallel, extracting deep features related to the disease. The core structure of DC-CNN consists of two independent but structurally similar (usually sharing some shallow layers) feature extraction channels: one specifically processes the visible light texture unfolded image (VTUI), called the Visible Channel (VC); the other specifically processes the infrared temperature unfolded image (ITUI), called the Infrared Channel (IC). Each channel is a deep convolutional neural network architecture. Taking the visible light channel as an example, its input is the VTUI (e.g., dimensions of Width W=8000 pixels x Height H=2000 pixels x Channels C=3, corresponding to RGB three channels). The first layer of the network typically consists of multiple (e.g., 64) convolutional kernels (CK), each of which is a small 2D filter (e.g., 3 pixels x 3 pixels). These convolutional kernels slide across the input image, performing convolution operations to calculate the dot product of the receptive field (RF) and the kernel weights, generating a preliminary feature map (FM). Each convolutional layer is typically followed by a non-linear activation function (such as ReLU, Rectified LinearUnit) to introduce non-linearity, and a pooling layer (PL, such as Max Pooling) to downsampling, gradually increasing the receptive field (RF) and reducing the spatial dimension (SD) while preserving key feature information.

[0070] The core task of the visible light channel (VC) is to identify and extract visual features associated with cracks and spalling. During training, this channel uses a large number of labeled visible light images of tunnel linings (containing various crack types such as transverse cracks, longitudinal cracks, mesh cracks, and spalling areas) for supervised learning. Through learning, the network's deep convolutional layers can automatically extract increasingly abstract and discriminative features. For example, shallow convolutional kernels may learn filters to detect edges and corners; mid-layers may learn filters to detect textures and blobs in specific directions; and deep layers may learn filters to represent complex crack morphologies (such as linear and dendritic), irregular boundaries of spalling areas, and internal broken textures. The network finally outputs a high-dimensional feature tensor (FT) at a specific layer (usually after the last convolutional layer or a global pooling layer), called the crack feature tensor (CFT). This tensor (e.g., with dimensions W'=50 x H'=12.5 x D=512) is no longer the original pixel image, but a compact representation containing rich abstract feature information about the probability, morphology, orientation, and density of cracks and spalling at all spatial locations in the image (corresponding to the original tunnel surface locations). Similarly, the input to the infrared channel (IC) is an infrared temperature unfolded map (ITUI, which may be a single-channel grayscale image or a pseudo-color image). Its training objective is to identify and extract thermal anomaly features associated with water seepage. Water seepage areas typically exhibit localized temperature anomalies (low or high temperature areas, depending on the environment) due to the heat absorption of water evaporation or the difference in heat capacity with the surrounding dry lining. The IC network learns to detect such anomalous patterns of temperature distribution (e.g., localized low-temperature patches, low-temperature lines extending along cracks, or temperature gradient fields formed by humidity diffusion). The final output of the IC network is a seepage feature tensor (SFT), whose dimensions may be the same as or similar to those of the CFT (e.g., W'=50 x H'=12.5 x D=512), which contains deep feature information such as the probability of leakage at various locations in the image, the leakage range, and the direction of the temperature gradient.

[0071] The key advantage of the dual-channel (VC and IC) design lies in its optimization capabilities for different physical modalities (light vs. heat). The network structure and parameters of the visible light channel (VC) (such as kernel size, number of layers, and number of channels) are optimized for optical image characteristics (such as edges, texture, and color contrast). The infrared channel (IC) is optimized for thermal image characteristics (such as absolute temperature values, local temperature contrast, temperature gradient, and spatial continuity). The two channels may employ different loss functions (LF) or data augmentation (DA) strategies during training. For example, VC training may prioritize geometric accuracy (e.g., using pixel-level cross-entropy loss), while IC training may focus more on the sensitivity of temperature anomaly detection (e.g., using focal loss to address class imbalance issues where leakage areas are typically small). During the inference phase, the system feeds the prepared VTUI and ITUI into the trained VC and IC networks, respectively. The network performs forward propagation, undergoing layers of convolution, activation, and pooling operations, ultimately outputting two high-dimensional feature tensors in parallel: CFT (primarily encoding crack and spalling information) and SFT (primarily encoding seepage information). These two feature tensors are perfectly aligned in the spatial dimensions (W', H'), meaning that the feature vector at position (i, j) in the tensor corresponds to the deep feature representation of the same physical location point in the original two-dimensional unfolded graph (U, V coordinate system) (and thus to a point on the three-dimensional model). This spatial alignment is crucial, providing the foundation for subsequent cross-modal correlation analysis.

[0072] Input the crack feature tensor and the leakage feature tensor into the cross-attention module to generate an association-enhanced feature map; While the Crack Feature Tensor (CFT) and Leakage Feature Tensor (SFT) output by a dual-channel CNN each contain rich intra-modal features, the disease information they reflect is interrelated. For example, a crack (represented by a high response in the CFT) accompanied by seepage (represented by a high response in the SFT) significantly increases the disease risk in that area; conversely, an isolated thermal anomaly may not necessarily indicate leakage. To effectively utilize this complementary cross-modal information and correlation, the system inputs the CFT and SFT into a Cross-Attention Module (CAM) for fusion and interaction. The core idea of ​​the Cross-Attention Mechanism is to allow features from one modality (called Query, Q) to "attention" to features from another modality (called Key, K and Value, V), thereby generating a correlated and enhanced feature that integrates information from both modalities. In practical implementations, it is typically designed as a bidirectional interaction. First, the system projects the CFT and SFT separately through independent Linear Transformation Layers (LTLs) to generate their respective query (Q), key (K), and value (V) representations. For example, define: Queries from CFT: Q_c = Linear_QC(CFT); Key from SFT: K_s = Linear_KS(SFT); Values ​​derived from SFT: V_s = Linear_VS(SFT).

[0073] Similarly, a query (Q_s) from the SFT can be defined to attend to the key (K_c) and value (V_c) from the CFT.

[0074] Taking the crack feature as a query (Q_c) and focusing on the leakage feature (K_s, V_s) as an example, the module calculates the similarity matrix (SM) between Q_c and K_s. Each element SM[i, j] of the similarity matrix SM represents the correlation or matching degree between the feature vector at position i in the CFT and the feature vector at position j in the SFT. The similarity is usually calculated using the dot product or scaled dot product. Then, the Softmax function is applied to each row of the similarity matrix SM (corresponding to a position in the CFT) for normalization, resulting in the attention weight matrix (AWM). The value of AWM[i, j] indicates how much weight should be assigned to the feature value (V_s[j]) at position j in the SFT when generating the associated enhancement feature at position i in the CFT. The attention weight matrix AWM intuitively reflects the "degree of attention" of the crack feature position to the leakage feature position. For example, a point on a crack (CFT position i) will pay special attention to points (SFT positions j) that are spatially adjacent to or extend along the crack's direction and have significant leakage responses, assigning them high weights. Next, the value V_s is weighted and summed using the attention weight AWM: for each position i in the CFT, its output feature incorporating leakage information is: Output_i = ∑_j (AWM[i, j] * V_s[j]). This process essentially allows the crack features to "selectively" fuse with their associated leakage feature information, thereby generating a feature representation with enhanced correlation at the crack location. Similarly, the system performs the reverse operation: allowing the leakage features (Q_s) to focus on the crack features (K_c, V_c), generating a leakage feature representation that incorporates crack information.

[0075] After bidirectional cross-attention, the system obtains two enhanced feature tensors: one is an Enhanced Crack Tensor (ECT) dominated by crack features but incorporating relevant leakage information, and the other is an Enhanced Seepage Tensor (EST) dominated by leakage features but incorporating relevant crack information. To fully utilize the bidirectional information, the system typically concatenates the ECT and EST along the channel dimension (CD) to form the final Association-Enhanced Feature Map (AEFM). The size of the AEFM is consistent with the input CFT / SFT in the spatial dimensions (W', H'), but the channel dimension (D') is usually the sum of the two (e.g., D'=1024, if ECT and EST each have 512 channels). The AEFM is the core output of the cross-attention module. It significantly improves the model's depth of understanding of the disease: crack information is no longer isolated, and leakage information also finds possible source paths. More importantly, it helps identify complex diseases (such as active seepage cracks) that are easily overlooked or misjudged in a single modality. Furthermore, the cross-attention mechanism exhibits spatial adaptability, dynamically adjusting the fusion strategy based on the characteristics of local image regions, resulting in more accurate fusion. AEFM provides a comprehensive feature representation with strong spatial correlation, integrating visible light crack spalling information and infrared seepage information, for the next step of disease atlas generation.

[0076] The association-enhanced feature map is passed through a cascaded decoder to generate a disease probability map. The boundary is optimized by CRF, and the candidate disease map is output.

[0077] While Association Enhancement Feature Maps (AEFM) contain rich disease information, their spatial resolution is low (due to pooling operations in CNNs downsampling the size from the original image size W x H to W' x H'), and they are high-dimensional abstract features, not intuitive disease segmentation results. To transform AEFM into an image that labels disease location and category at the original resolution (or close to the original resolution), the system employs a cascaded decoder (CD). A cascaded decoder typically consists of a series of transposed convolutional layers (TCL, also known as deconvolution layers) or upsampling layers (UL, such as bilinear upsampling) alternately stacked with convolutional layers (CL). Its function is to gradually upsample the low-resolution, high-channel-count feature map (AEFM) to restore it to a high spatial resolution, while refining features through convolution operations, and finally outputting a multi-channel feature map with the same or similar size as the input unfolded map (VTUI / ITUI) (e.g., W x H).

[0078] The decoder operates in a step-by-step manner. It begins at the highest level of abstraction (AEFM). The first-level decoder typically includes a transposed convolution operation (e.g., using a 4x4 kernel with a stride of 2), doubling the spatial size of the feature map (e.g., from W'=50, H'=12.5 to ~100 x ~25), while potentially reducing the number of channels (e.g., from 1024 to 512). This is followed by one or more standard convolutional layers that refine the upsampled, coarse feature map, extracting finer spatial details. Then, to further recover these details, the system performs a skip connection between the current upsampled and refined feature map and an intermediate feature map (IFM) of the corresponding layer in the encoder (i.e., a two-channel CNN) that has the same spatial size (before pooling downsampling). This is typically achieved by concatenating or adding the two maps along the channel dimension. This step is crucial; it re-injects low-level spatial details (such as edges and textures) discarded by the encoder into the decoding process, significantly improving the boundary accuracy and detail restoration capability of the final segmentation. After multiple levels (e.g., 3-4 levels) of such upsampling-refining-skip connection operations, the final output is a feature map with the same spatial dimensions (W x H) as the original input image (VTUI / ITUI) and the number of channels equal to the number of disease categories. For tunnel disease detection, there are typically three categories: background, cracks / spare, and seepage. Therefore, the output feature map may have three channels. Then, the Softmax function is applied to this multi-channel feature map for pixel-wise normalization, converting the value of each channel into a probability value (between 0 and 1). The final result is the Disease Probability Map (DPM). In the DPM, each pixel location has three probability values, representing the probability that the pixel belongs to the background, crack / spare, or seepage category, respectively. For example, if a pixel has a crack / peeling probability of 0.92, a water seepage probability of 0.85, and a background probability of 0.03, it indicates that the pixel is very likely to have both cracks / peeling and water seepage.

[0079] While cascaded decoders combined with skip connections can generate high-quality disease probability maps (DPMs), their boundary segmentation may sometimes lack precision or contain small noise. To further improve the boundary accuracy and spatial consistency of the segmentation results, the system introduces a Conditional Random Field (CRF) as a post-processing optimization module after the decoder output. A CRF is a probabilistic graphical model that models the dependencies between pixel labels in an image. In the disease segmentation task, CRF considers two main factors: 1. Unary Potential: the independent class probability of each pixel from the decoder output (the probability value provided by the DPM). 2. Pairwise Potential: the energy defining the label consistency between adjacent pixels (usually spatially close pixels). The pairwise potential encourages adjacent pixels with similar color / texture / temperature to have the same label (e.g., both are cracks), while penalizing adjacent pixels with large feature differences to have the same label. The system inputs the DPM as the unary potential into the CRF model. Simultaneously, the original bimodal unfolded image (the RGB values ​​of VTUI and the temperature or grayscale values ​​of ITUI) is used as features to define the similarity between pixels for calculating the pairwise potential. Then, the CRF model minimizes the global energy function through iterative optimization algorithms (such as mean field approximation) to find the pixel label configuration that maximizes the overall probability. This process effectively smooths the segmentation results: eliminating isolated noise, filling small holes, and sharpening blurred boundaries, making the crack direction clearer and more coherent, and the boundaries of the seepage area more consistent with the continuity of temperature field diffusion. The result after CRF optimization is the final Candidate Disease Map (CDM). The CDM is an image of the same size as the original unfolded image, with each pixel assigned a specific disease category label (background, crack / stripping, seepage). This map clearly marks all detected crack (potentially including spalling) and seepage areas on the tunnel lining surface. It is the core output of the defect identification step, providing a spatial basis for subsequent quantitative index calculations and risk assessments. Each defect patch in the map is associated with its coordinates in a two-dimensional unfolded diagram (U, V), which can then be mapped back to the original three-dimensional model spatial coordinates through inverse projection, enabling precise location of the defect on the tunnel inner wall.

[0080] S204, Perform spatial topological structure analysis on the candidate disease map, and generate a set of quantitative indicators including crack length weighted curvature, peeling area-to-volume ratio, and seepage path thermodynamic parameters based on the geometric continuity, depth gradient, and temperature distribution of the disease area. Specifically, crack skeleton lines can be extracted from candidate disease maps, the curvature along the length can be calculated as a weighted integral, and the crack length-weighted curvature value can be output. Crack skeleton line extraction: The Candidate Disease Map (CDM) is a pixel-level disease probability map generated by a dual-channel convolutional neural network, where crack regions are represented by thin, continuous, dark lines. First, a skeletonization algorithm is used to refine the crack regions: iterative erosion operations remove boundary pixels, retaining connected center lines with a width of only one pixel, forming complete crack skeleton lines (CSLs). For example, the classic Zhang-Suen parallel thinning algorithm is used, which defines the 8-neighbor connection count (NCC) of pixels and progressively removes non-critical boundary points, ensuring the skeleton line remains unbroken and the original crack topology is preserved. After skeletonization, a branch pruning technique is used to remove short burrs smaller than a preset threshold (e.g., 5 cm) to avoid noise interference. The final output skeleton line network accurately reflects the main trunk and branch directions of the cracks, laying the foundation for subsequent geometric analysis.

[0081] Curvature calculation and length weighting: Along each crack skeleton line, key control points (KCPs) are selected at fixed sampling intervals (e.g., every 1 cm). At each KCP, the adjacent points before and after it (e.g., the third point before and the third point after) form a local curve segment. The local curvature (LC) of this point is calculated using the three-point circle fitting method. Specifically, the three-point circle fitting method obtains the curvature value (in radians per meter) by solving for the reciprocal of the circumcircle radius of the three non-collinear points. For example, a curvature value of 0.25 rad / m at a point indicates that the crack is highly curved at that location. Since the hazard of different sections of a crack is positively correlated with its curvature and length, a length-weighted integration (LWI) is required. This involves multiplying the curvature value of each key crack point (KCP) by the micro-arc length (MAL, which is half the distance between adjacent KCPs), and then summing the products over all segments of the crack to obtain the crack length-weighted curvature value (CLWCV). This value comprehensively reflects the overall curvature of the crack. For example, a crack 10 meters long with an average curvature of 0.15 rad / m has a CLWCV of approximately 1.5 rad·m, significantly higher than that of a short, straight crack.

[0082] Engineering significance and parameter output: CLWCV is a core indicator for assessing the structural risk of cracks: high-curvature cracks (such as circumferential cracks) are prone to stress concentration, accelerating lining deterioration. The system binds the CLWCV of each crack to its spatial coordinates and associates it with the 3D model. For example, a crack in the arch with a CLWCV of 2.3 rad·m is marked as high risk, while a crack in the sidewall with a CLWCV of 0.7 rad·m is low risk. The CLWCV of all cracks constitutes the first parameter of the quantitative indicator set, providing a basis for subsequent risk classification.

[0083] The surface area to volume ratio of the exfoliated region is calculated based on a gradient-optimized 3D model, and the exfoliation volume ratio parameter is output. 3D reconstruction of the exfoliated area: The spalling regions (SRs) in the candidate lesion maps appear as irregular, sheet-like patches. Based on a gradient-optimized 3D model (GO3DM), the geometry of the spalling regions is reconstructed using a triangular mesh tessellation algorithm: the boundary point cloud of the SR is extracted, and Delaunay triangulation is used to generate a triangular mesh (MeshResolution, MR, typically set to 5 mm) covering the region. Mesh quality is optimized using the Max-Min Angle Criterion to avoid elongated triangular patches. The reconstructed spalling body is a closed surface composed of thousands of triangular patches, accurately representing its convex-concave morphology and depth variations.

[0084] Surface area and volume calculation: Surface Area (SA): Calculated by summing the areas of all triangular facets of the exfoliated body (using the modulus of the cross product of vectors). For example, an exfoliated body containing 2000 facets has an SA of 0.85 square meters.

[0085] Volume (VOL): Calculated using the Ray Casting Method: Virtual rays are emitted from the surface of the exfoliated body into its interior. The number of voxels completely located within the exfoliated area (Voxel Size, VS, typically a 1 cm cube) is counted using a voxel flood fill algorithm. The total volume is the number of voxels multiplied by the volume of a single voxel. For example, if an exfoliated body occupies 5200 voxels, then VOL = 0.052 cubic meters.

[0086] Finally, calculate the Spalling Area-Volume Ratio (SAVR), which is the ratio of SA to VOL (unit: square meters per cubic meter). For example, SA = 0.85 m. 2 VOL=0.052 m 3 At that time, SAVR≈16.35 m -1 .

[0087] Physical meaning and output of parameters: SAVR reflects the severity of spalling: high ratios (e.g., >15 m) -1 This indicates shallow exfoliation (such as hollow areas) and low ratios (such as <5m). -1This indicates deep spalling (such as large-area concrete spalling). This parameter is recorded as the Spalling Volume Ratio Parameter (SVRP) and associated with the 3D model coordinates. For example, the SVRP of a spalled area in a sidewall is 18.2m. -1 Marked as "shallow hollow layer", it requires close attention to the risk of adhesion failure.

[0088] Solve for the temperature gradient vector field along the seepage path and output the thermodynamic seepage direction matrix; Seepage path extraction and temperature field mapping: Seepage regions (SERs) in candidate lesion maps typically exhibit a dendritic diffusion pattern. First, the main seepage path (SMP) is extracted through skeletonization and main path tracking. Based on a gradient-optimized 3D model (GO3DM), infrared temperature field data is mapped to each vertex of the SMP, forming a path temperature sequence (PTS). For example, an 8-meter-long SMP might contain 800 vertices, each corresponding to a temperature value (in degrees Celsius).

[0089] Temperature gradient vector field calculation: Along the SMP path, with a step size of 1 cm between adjacent vertices, calculate the Temperature Gradient Vector (TGV) for each point: Gradient direction: along the tangent of the path, from the high temperature point to the low temperature point (i.e., the direction of water flow).

[0090] Gradient magnitude: Using the Central Difference Method, for example, for points A (temperature T_A), B (T_B), and C (T_C), the gradient magnitude of point B = |T_A - T_C| / (2 × step size).

[0091] The TGV at all points constitutes a Temperature Gradient Vector Field (TGVF). This field reveals the direction of water flow and the intensity of heat diffusion; for example, the TGV at a certain point is 0.8℃ / cm, and its direction points towards the tunnel invert.

[0092] Permeation direction matrix generation: The tunnel lining surface was divided into a 1m × 1m spatial grid (SG). Within each grid cell, all TGV data falling into that cell were collected. Dominant Seepage Direction (DSD): The direction of TGV synthesis within the grid (angle 0-360 degrees) is calculated by vector averaging.

[0093] Direction Confidence (DC): The ratio of the magnitude of the composite vector to the magnitude of the total vector (0-1), reflecting directional consistency.

[0094] The output is a Thermodynamic Seepage Direction Matrix (TSDM), with each element being a (DSD, DC) pair. For example, if the element in the 3rd row and 5th column of the matrix is ​​(125°, 0.78), it indicates that the water flow in this area is mainly in the 125° direction, with a confidence level of 78%.

[0095] The thermodynamic permeation direction matrix is ​​integrated by path to generate the seepage thermodynamic attenuation coefficient; a structured quantitative index set is constructed based on the crack length weighted curvature value, the spalling volume ratio parameter and the seepage thermodynamic attenuation coefficient.

[0096] Thermodynamic modeling of seepage path: Based on the thermodynamic seepage direction matrix (TSDM), a continuous seepage path (CSP) is reconstructed in the 3D model. The path starts at the seepage development point (e.g., the highest temperature point) and ends at the point where the water flow disappears (temperature close to the ambient value). The path extends according to the main direction indicated by the TSDM: between adjacent grids, probabilistic path connections are made based on the DSD and DC values, with higher DC values ​​(e.g., >0.7) being preferred.

[0097] Temperature decay path integral: Along the reconstructed CSP, samples are taken at fixed step sizes (e.g., 0.1 meters), recording the temperature value (T) and path length (L) from the starting point at each point. The Normalized Temperature Attenuation Function (NTAF) is then calculated: ; The function value decreases from 1 (starting point) to 0 (ending point). The area under the NTAF curve is calculated using the trapezoidal numerical integration method to obtain the thermodynamic attenuation coefficient (STAC). For example, for a path length of 5 meters, STAC = 0.42, indicating that the temperature decreases gradually along the path.

[0098] Parameter meaning and indicator integration: STAC quantifies the concealment and persistence of seepage: high coefficients (e.g., >0.6) indicate slow flow (e.g., capillary seepage), while low coefficients (e.g., <0.3) indicate rapid dispersion (e.g., through-cracks). A Structured Quantitative Indicator Set (SQIS) is constructed by combining crack length-weighted curvature (CLWCV), slough volume ratio (SVRP), and STAC. For example: { "Crack_001": {CLWCV: 1.8 rad·m, Position: [x1,y1,z1]}, "Spalling_005": {SVRP: 12.4 m -1 Position: [x2,y2,z2]}, "Seepage_012": {STAC: 0.38, Position: [x3,y3,z3]} }

[0099] This set of indicators serves as a unified standard for measuring the severity of diseases and is input into subsequent risk prediction models.

[0100] S205, the set of quantitative indicators is input into the spatiotemporal evolution hidden Markov model, and the real-time risk level and evolution trend report of each disease area are determined according to the degradation pattern in the historical disease database, and associated with the spatial coordinates of the three-dimensional model.

[0101] Specifically, the set of quantitative indicators can be gridded according to the coordinates of the three-dimensional model space to generate a gridded state vector sequence; The system receives a structured set of quantitative indicators from previous steps, including key parameters such as crack length-weighted curvature (CLWC), spalling volume ratio (SVR), and seepage thermodynamic attenuation coefficient (STAC). First, a spatial mapping relationship is established with the 3D model (i.e., a 3D model of the lining surface with temperature attributes): using the vertex coordinate system (VCS) of the 3D model as a reference, the tunnel inner surface is divided into cubic grid cells (CGC) with sides of 10 cm. The center point coordinates (CPC) of each grid cell serve as the location identifier for that cell. The system traverses all grid cells, retrieving the quantitative indicators of defects falling within their spatial range. For example, a grid cell covers a 15 cm long crack with a CLWC value of 0.25 radians / meter (rad / m), and also contains a spalling area with a surface area of ​​0.3 square meters and a volume of 0.01 cubic meters (SVR=30 m). 2 / m 3 The STAC value is 0.85 (dimensionless). The state vector (SV) of this unit is defined as [CLWC=0.25, SVR=30, STAC=0.85], and the vector dimension strictly corresponds to the quantification index type.

[0102] To achieve time-series analysis, the system introduces a time dimension (TD). Each tunnel inspection task generates an independent set of quantitative indicators, corresponding to an inspection timestamp (IT). Assuming the historical database contains four inspection data points from 2020 to 2023, four time-ordered state vectors will be generated for the same grid cell: SV_2020, SV_2021, SV_2022, and SV_2023. This constitutes the state vector sequence (SVS) for that cell. The system performs this operation in parallel for all grid cells, ultimately generating a gridded state vector sequence (GSVS) covering the entire tunnel surface. During this process, the grid alignment problem needs to be addressed: when the 3D models from different inspection periods have slight offsets due to scanning errors, an affine transformation algorithm (ATA) is used for spatial calibration to ensure that grid cells at the same physical location have a consistent spatial coordinate index (SCI) across different data periods. The grid resolution (10 cm) needs to be determined based on the minimum feature size of the disease (such as a crack width threshold of 5 mm) and the balance of computational resources.

[0103] Special scenario handling is a crucial aspect of mesh generation. For defects spanning multiple mesh cells (such as a 2-meter-long crack), area-weighted interpolation (AWI) is used to allocate index values: if 70% of a crack segment's area is located in mesh cell A and 30% in cell B, its CLWC value is proportionally split across the SV values ​​of the two cells. For defect-free mesh cells, their SV is filled with a default zero vector ([CLWC=0, SVR=0, STAC=1]), where STAC=1 indicates no seepage thermodynamic attenuation (ideal state). The final output GSVS is a four-dimensional data structure (spatial mesh coordinates X / Y / Z + time + index vector), which can be directly input into subsequent spatiotemporal evolution models.

[0104] The gridded state vector sequence is input into a two-layer hidden Markov model, where the lower HMM outputs the unit state transition probability and the upper HMM outputs the neighborhood propagation probability. A Two-Layer Hidden Markov Model (TL-HMM) consists of a lower-layer HMM (LL-HMM) and an upper-layer neighborhood-level HMM (UL-HMM). First, the lower-layer HMM is processed: each grid cell is treated as an independent state machine (SM). Its hidden state (HS) is defined as the disease risk level, for example, divided into five levels: HS1 (no risk), HS2 (low risk), HS3 (medium risk), HS4 (high risk), and HS5 (critical risk). The observation state (OS) is the SV of that cell (e.g., [CLWC=0.25, SVR=30, STAC=0.85]). The core of LL-HMM is to calculate the Unit State Transition Probability (USTP), which is the probability P(HS_t | HS_t-1) of transitioning from the previous state HS_t-1 to the current state HS_t. For example, if a unit's state is HS3 in 2022 and HS4 in 2023, then a "HS3→HS4" transition event is recorded. The system uses a historical database to count the frequency of state transitions of all units between adjacent periods and constructs a global transition probability matrix (TPM). The matrix element TPM[i][j] represents the statistical probability of transitioning from state HS_i to HS_j (e.g., TPM[3][4]=0.15 means that 15% of medium-risk units will be upgraded to high-risk units in the next period).

[0105] The upper-level HMM is used to model the spatial propagation characteristics of the disease. The neighborhood (NH) of each grid cell is defined as its nine neighboring cells in a 3×3×1 (planar expansion, ignoring height) range. The upper-level hidden state is defined as the Neighborhood Risk Propagation Intensity (NRPI), divided into three levels: NRPI1 (no propagation), NRPI2 (weak propagation), and NRPI3 (strong propagation). The observed state is the pattern of state changes (PSC) of all cells within the neighborhood. The UL-HMM outputs the Neighborhood Propagation Probability (NPP), which is the probability P(NRPI | HS_center, {HS_neighbors}) of the risk spreading to the neighborhood given the current state of the central cell and the states of its neighbors. For example, if the central cell is HS4 (high risk) and there are three HS3 (medium risk) neighboring cells, the system might calculate NPP = 0.8 (strong propagation probability 80%). This probability is determined by the proportion of disease spread that actually occurred under similar neighborhood configurations in historical data.

[0106] The collaborative operation of the two-layer Hidden Markov Model (HMM) is achieved through the Viterbi-Bayesian Coupling Algorithm (VBCA). Taking the 2023 detection as an example: First, the LL-HMM calculates the USTP (e.g., P(HS4|HS3)=0.15) based on the SV change of the cell itself from 2022 to 2023; simultaneously, the UL-HMM calculates the NPP (e.g., P(NRPI3|HS4, {HS3,HS2,...})=0.8) based on the neighborhood state in 2023. Both are jointly input into the Posterior Probability Fusion (PPF), which outputs the final state confidence of the cell in 2023 (e.g., HS4 confidence 85%). During model training, the Baum-Welch Algorithm (BWA) is used to learn the optimal TPM and NPP parameters from historical GSVS data, ensuring that the probabilistic predictions conform to the actual evolution of tunnel defects.

[0107] Input the unit state transition probability and neighborhood propagation probability into the graph neural network, aggregate the topological risk features, and output the grid risk level distribution map; The Graph Neural Network (GNN) is constructed using a 3D mesh model as its topological framework. Each mesh cell is defined as a Graph Node (GN), and the Node Feature Vector (NFV) contains three core inputs: 1) the unit state transition probability USTP output from the lower-level HMM (e.g., P(HS4|HS3)=0.15); 2) the neighborhood propagation probability NPP output from the upper-level HMM (e.g., P(NRPI3)=0.8); and 3) the latest quantization index SV of the cell (e.g., [CLWC=0.25, SVR=30, STAC=0.85]). Connections between nodes (i.e., graph edges) are established based on spatial adjacency: if two cells share a face, edge, or corner (based on Delaunay triangulation rules), an undirected edge is added, and the edge weight (EW) is initialized to the reciprocal of the Euclidean distance between the center points of the two cells (e.g., if the distance between adjacent cells is 0.1 meters, then EW=10 m). -1 ).

[0108] GNN uses a message passing-aggregation mechanism (MPAM) for feature propagation. Each node iteratively performs three steps: Message Generation (MG): Node A sends a message to its neighbor B. The message content = A's NFV × the edge weight EW_AB from A to B × the propagation factor (PF). PF is taken as A's NPP value (e.g., 0.8), representing the intensity of risk diffusion.

[0109] Message Aggregation (MA): Node B collects messages from all its neighbors (e.g., 8) and fuses them into an aggregated feature (AF) through a gated recurrent unit (GRU).

[0110] Node Update (NU): Concatenate the original NFV of B with AF and input it into the Fully Connected Layer (FCL) to update the NFV.

[0111] After three rounds of iteration, the node features fully integrate local topological relationships (such as multiple units on a crack mutually reinforcing risk confidence). The final output NFV includes topological risk features (TRF), such as "the risk of the crack end unit is increased by 20% due to its proximity to the spalling area".

[0112] The generation of a gridded risk level distribution map (GRLDM) involves two steps: Risk classification: The updated NFV is input into the Softmax Classifier (SC), which outputs a five-dimensional probability vector [P(HS1), P(HS2), ..., P(HS5)]. The level corresponding to the highest probability is taken as the instantaneous risk level of the unit (e.g., P(HS4) = 0.92 → Level HS4).

[0113] Spatial rendering: Based on the unit's spatial coordinates (CPC) and risk level, color-coded rendering is performed on the surface of the 3D model (e.g., HS1=green, HS5=red). The output is an interactive 3D heatmap, and users can click on any mesh to view its TRF details (e.g., "This unit's risk has increased to HS4 due to crack curvature CLWC=0.28 and seepage in the neighborhood STAC=0.7").

[0114] Based on the grid risk level distribution map and historical database, Bayesian evolution prediction is performed to generate a risk report with three-dimensional spatial coordinate association.

[0115] Bayesian Evolution Prediction (BEP) starts with the previously generated GRLDM (Instant Risk Status) and combines it with degradation patterns from the Historical Defect Database (HDD) to perform multi-period predictions. The HDD stores three key types of information: 1) Statistical transitions of defects of specific risk levels in similar tunnels after 1 / 3 / 5 years (e.g., after three years, 60% of HS3 units remain HS3, 30% upgrade to HS4, and 10% downgrade to HS2); 2) Environmental Factors (EF), including groundwater pH and freeze-thaw cycles; 3) Maintenance Intervention Records (MIR). For each grid cell of the target tunnel, a Bayesian Prediction Network (BPN) is constructed, where nodes represent the risk status at different future time points (e.g., 2024, 2025, 2026), and the transition probabilities between nodes are initialized by statistical patterns from similar environments in the HDD.

[0116] The prediction process incorporates Dynamic Weight Adjustment (DWA): Time decay correction: The prediction weight W_t = e^(-λt) for the future year t (λ=0.1 is the decay coefficient), which means that the further away from the present, the greater the uncertainty of the prediction.

[0117] Environmental factor matching: The 10 historical records closest to the current tunnel EF are retrieved from the HDD, and the transfer probability is adjusted by a weighted average of their degradation trajectories (weight = similarity). For example, if the current groundwater pH is 6.8, the weight of historical data with pH values ​​of 6.5 to 7.0 is increased to 0.9.

[0118] Maintenance strategy simulation: If the user inputs a planned maintenance scheme (such as "grouting the HS5 area in 2024"), the system retrieves the effects of similar measures from the MIR (such as the probability of HS5→HS2 after grouting being 70%), and forcibly modifies the transfer path of the affected unit.

[0119] The final output is the probability distribution of risk level for each unit at various future time points. For example, for the unit with coordinates (X=120.5m, Y=3.2m, Z=0m): P(HS4)=0.75, P(HS5)=0.25 in 2024; P(HS4)=0.4, P(HS5)=0.6 in 2025.

[0120] The Spatial Coordinate-Linked Risk Report (SC-LRR) consists of four parts: Real-time Risk Map: This is a 3D visualization model of GRLDM that supports rotation / zoom operations.

[0121] Key lesion list: Automatically filter HS4 and above units, generate independent lesion areas by spatial clustering (e.g., "crack cluster C001: involving 35 grids, highest risk HS5"), and associate their coordinate boundaries.

[0122] Evolutionary trend chart: Outputs a line chart for each lesion area, showing the probability changes of each risk level over the next few months (assuming the HS5 probability of C001 increases from 25% in March to 80% in August).

[0123] Maintenance Recommendation Matrix: Based on the treatment effects of similar lesions in HDD, the optimal measures are recommended (e.g., "C001 recommends epoxy resin grouting within 6 months, with an expected success rate of 92%"). All data points in the report are bound to the coordinates of the 3D model; clicking on any lesion will link to its actual location in the tunnel.

[0124] As can be seen, simultaneous scanning of the tunnel lining surface using visible light, infrared thermal imaging, and 3D point cloud yields a spatiotemporally aligned multimodal raw dataset. Point cloud registration and multi-source data fusion processing are then performed on the multimodal raw dataset to construct a 3D model of the lining surface with temperature attributes. Based on the 3D model, a dual-channel convolutional neural network is used to extract visible light crack features and infrared leakage features, respectively, outputting candidate defect maps. Spatial topological structure analysis is performed on the candidate defect maps to generate a set of quantitative indicators. This set of quantitative indicators is then input into a spatiotemporally evolving hidden Markov model to determine the immediate risk level and evolution trend report for each defect area, and correlated with the spatial coordinates of the 3D model. This allows for dynamic assessment of defect risk and prediction of development trends, improving detection efficiency and accuracy.

[0125] Another embodiment of the present invention provides a data acquisition and analysis system for detecting apparent defects in tunnel structures, see [link to relevant documentation]. Figure 3 The system may include: The scanning module 301 is used to perform simultaneous scanning of visible light, infrared thermal imaging and three-dimensional point cloud on the surface of the tunnel lining by a mobile acquisition platform equipped with a multispectral imaging module and a lidar, according to a preset tunnel detection path, so as to obtain a spatiotemporally aligned multimodal raw dataset. The construction module 302 is used to perform point cloud registration and multi-source data fusion processing on the multimodal raw dataset, and to construct a three-dimensional model of the lining surface with temperature attributes based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space. Extraction module 303 is used to extract visible light crack features and infrared leakage features respectively based on the three-dimensional model using a dual-channel convolutional neural network, and perform cross-modal correlation analysis through feature cascade module to output candidate disease map containing crack, spalling and seepage areas. The analysis module 304 is used to perform spatial topological structure analysis on the candidate disease map and generate a set of quantitative indicators including crack length weighted curvature, peeling area-to-volume ratio, and seepage path thermodynamic parameters based on the geometric continuity, depth gradient, and temperature distribution of the disease area. The determination module 305 is used to input the set of quantitative indicators into the spatiotemporal evolution hidden Markov model, determine the real-time risk level and evolution trend report of each disease area according to the degradation pattern in the historical disease database, and associate it with the spatial coordinates of the three-dimensional model.

[0126] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0127] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps: S201, according to the preset tunnel detection path, through a mobile acquisition platform equipped with a multispectral imaging module and lidar, performs simultaneous scanning of visible light, infrared thermal imaging and three-dimensional point cloud on the tunnel lining surface to obtain a spatiotemporally aligned multimodal raw dataset. S202, perform point cloud registration and multi-source data fusion processing on the multimodal original dataset, and construct a three-dimensional model of the lining surface with temperature attributes based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space. S203. Based on the three-dimensional model, a dual-channel convolutional neural network is used to extract visible light crack features and infrared leakage features respectively. Cross-modal correlation analysis is performed through the feature cascade module to output a candidate disease map containing cracks, spalling, and water seepage areas. S204, Perform spatial topological structure analysis on the candidate disease map, and generate a set of quantitative indicators including crack length weighted curvature, peeling area-to-volume ratio, and seepage path thermodynamic parameters based on the geometric continuity, depth gradient, and temperature distribution of the disease area. S205, the set of quantitative indicators is input into the spatiotemporal evolution hidden Markov model, and the real-time risk level and evolution trend report of each disease area are determined according to the degradation pattern in the historical disease database, and associated with the spatial coordinates of the three-dimensional model.

[0128] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0129] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0130] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program: S201, according to the preset tunnel detection path, through a mobile acquisition platform equipped with a multispectral imaging module and lidar, performs simultaneous scanning of visible light, infrared thermal imaging and three-dimensional point cloud on the tunnel lining surface to obtain a spatiotemporally aligned multimodal raw dataset. S202, perform point cloud registration and multi-source data fusion processing on the multimodal original dataset, and construct a three-dimensional model of the lining surface with temperature attributes based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space. S203. Based on the three-dimensional model, a dual-channel convolutional neural network is used to extract visible light crack features and infrared leakage features respectively. Cross-modal correlation analysis is performed through the feature cascade module to output a candidate disease map containing cracks, spalling, and water seepage areas. S204, Perform spatial topological structure analysis on the candidate disease map, and generate a set of quantitative indicators including crack length weighted curvature, peeling area-to-volume ratio, and seepage path thermodynamic parameters based on the geometric continuity, depth gradient, and temperature distribution of the disease area. S205, the set of quantitative indicators is input into the spatiotemporal evolution hidden Markov model, and the real-time risk level and evolution trend report of each disease area are determined according to the degradation pattern in the historical disease database, and associated with the spatial coordinates of the three-dimensional model.

[0131] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A data acquisition and analysis method for detecting apparent defects in tunnel structures, characterized in that, The method includes: According to the preset tunnel detection path, a mobile acquisition platform equipped with a multispectral imaging module and lidar is used to simultaneously scan the tunnel lining surface with visible light, infrared thermal imaging and three-dimensional point cloud to obtain a spatiotemporally aligned multimodal raw dataset. The multimodal raw dataset is subjected to point cloud registration and multi-source data fusion processing. Based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space, a three-dimensional model of the lining surface with temperature attributes is constructed. Based on the three-dimensional model, a dual-channel convolutional neural network is used to extract visible light crack features and infrared leakage features respectively. Cross-modal correlation analysis is performed through a feature cascade module to output a candidate disease map containing cracks, spalling, and water seepage areas. Spatial topology analysis is performed on the candidate disease map. Based on the geometric continuity, depth gradient and temperature distribution of the disease area, a set of quantitative indicators including crack length-weighted curvature, spalling area-to-volume ratio and seepage path thermodynamic parameters is generated. The set of quantitative indicators is input into the spatiotemporal evolution hidden Markov model. Based on the degradation patterns in the historical disease database, the instantaneous risk level and evolution trend report of each disease area are determined and associated with the spatial coordinates of the three-dimensional model.

2. The method according to claim 1, characterized in that, The process involves simultaneously scanning the tunnel lining surface using visible light, infrared thermal imaging, and 3D point clouds via a mobile acquisition platform equipped with a multispectral imaging module and lidar, based on a preset tunnel detection path. This yields a spatiotemporally aligned multimodal raw dataset, including: The raw three-dimensional point cloud data is generated by scanning the tunnel lining surface with a lidar mounted on a mobile platform. The laser interferometric positioning base station is used to measure the platform's pose offset in real time, and the gyroscope is used to correct the inertial navigation drift, outputting a high-precision spatial pose matrix. The spatial pose matrix triggers the synchronous acquisition of visible light camera and infrared thermal imager to generate a spatiotemporally aligned set of visible light images and infrared temperature field sequence. By fusing the original 3D point cloud, visible light image set, and infrared temperature field sequence, a multimodal original dataset aligned with spatiotemporal reference is output.

3. The method according to claim 2, characterized in that, The step of performing point cloud registration and multi-source data fusion processing on the multimodal raw dataset, and constructing a three-dimensional model of the lining surface with temperature attributes based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space, includes: LiDAR point clouds are extracted from the multimodal raw dataset, and curvature adaptive filtering is performed to output the denoised geometric point cloud. The thermal expansion coefficient of the lining material is calculated based on the infrared temperature field, and a temperature-deformation compensation vector field is generated. The temperature-deformation compensation vector field is applied to the geometric point cloud, and the visible light texture coordinates are registered using a non-rigid ICP algorithm to output a geometrically corrected 3D mesh model. The infrared temperature field is mapped to the vertices of the 3D mesh model, and the initial 3D model with temperature attributes is output. Temperature field gradient optimization is performed on the initial 3D model. An MLP network is used to enhance the temperature continuity of the curved surface, and a gradient-optimized 3D model of the lining surface is output.

4. The method according to claim 3, characterized in that, Based on the aforementioned 3D model, a dual-channel convolutional neural network is used to extract visible light crack features and infrared leakage features, respectively. Cross-modal correlation analysis is then performed through a feature cascade module to output a candidate disease map containing cracks, spalling, and seepage areas, including: The gradient-optimized 3D model is projected to generate visible light texture unfolded map and infrared temperature unfolded map, and a bimodal 2D unfolded map is output. A dual-channel CNN is used to process the two-dimensional unfolded graph of the two modes respectively. The visible light channel outputs the crack feature tensor, and the infrared channel outputs the leakage feature tensor. Input the crack feature tensor and the leakage feature tensor into the cross-attention module to generate an association-enhanced feature map; The association-enhanced feature map is passed through a cascaded decoder to generate a disease probability map. The boundary is optimized by CRF, and the candidate disease map is output.

5. The method according to claim 4, characterized in that, The spatial topological structure analysis of the candidate disease maps, based on the geometric continuity, depth gradient, and temperature distribution of the diseased areas, generates a set of quantitative indicators including crack length-weighted curvature, spalling area-to-volume ratio, and thermodynamic parameters of seepage paths, including: Extract the crack skeleton line from the candidate disease map, calculate the length-weighted integral of the curvature, and output the crack length-weighted curvature value. The surface area to volume ratio of the exfoliated region is calculated based on a gradient-optimized 3D model, and the exfoliation volume ratio parameter is output. Solve for the temperature gradient vector field along the seepage path and output the thermodynamic seepage direction matrix; Path integration is performed on the thermodynamic permeation direction matrix to generate the seepage thermodynamic attenuation coefficient; A set of structured quantitative indicators is constructed based on crack length-weighted curvature values, spalling volume ratio parameters, and seepage thermodynamic attenuation coefficients.

6. The method according to claim 5, characterized in that, The process involves inputting the set of quantitative indicators into a spatiotemporally evolving hidden Markov model, determining the immediate risk level and evolution trend report for each diseased area based on degradation patterns in a historical disease database, and linking this report to the spatial coordinates of the three-dimensional model, including: The quantitative index set is gridded according to the three-dimensional model space coordinates to generate a gridded state vector sequence; The gridded state vector sequence is input into a two-layer hidden Markov model, where the lower HMM outputs the unit state transition probability and the upper HMM outputs the neighborhood propagation probability. Input the unit state transition probability and neighborhood propagation probability into the graph neural network, aggregate the topological risk features, and output the grid risk level distribution map; Based on the grid risk level distribution map and historical database, Bayesian evolution prediction is performed to generate a risk report with three-dimensional spatial coordinate association.

7. A data acquisition and analysis system for detecting apparent defects in tunnel structures, characterized in that, The system includes: The scanning module is used to scan the surface of the tunnel lining simultaneously with visible light, infrared thermal imaging and three-dimensional point cloud by a mobile acquisition platform equipped with a multispectral imaging module and lidar, according to the preset tunnel detection path, so as to obtain a spatiotemporally aligned multimodal raw dataset. The construction module is used to perform point cloud registration and multi-source data fusion processing on the multimodal raw dataset, and to construct a three-dimensional model of the lining surface with temperature attributes based on the mapping relationship between visible light texture and infrared temperature field in three-dimensional space. The extraction module is used to extract visible light crack features and infrared leakage features based on the three-dimensional model using a dual-channel convolutional neural network. It performs cross-modal correlation analysis through the feature cascade module and outputs a candidate disease map containing cracks, spalling, and water seepage areas. The analysis module is used to perform spatial topological structure analysis on the candidate disease map, and generate a set of quantitative indicators including crack length weighted curvature, peeling area-to-volume ratio, and seepage path thermodynamic parameters based on the geometric continuity, depth gradient, and temperature distribution of the disease area. The determination module is used to input the set of quantitative indicators into the spatiotemporal evolution hidden Markov model, determine the real-time risk level and evolution trend report of each disease area based on the degradation patterns in the historical disease database, and associate it with the spatial coordinates of the three-dimensional model.

8. The system according to claim 7, characterized in that, The scanning module is specifically used for: The raw three-dimensional point cloud data is generated by scanning the tunnel lining surface with a lidar mounted on a mobile platform. The laser interferometric positioning base station is used to measure the platform's pose offset in real time, and the gyroscope is used to correct the inertial navigation drift, outputting a high-precision spatial pose matrix. The spatial pose matrix triggers the synchronous acquisition of visible light camera and infrared thermal imager to generate a spatiotemporally aligned set of visible light images and infrared temperature field sequence. By fusing the original 3D point cloud, visible light image set, and infrared temperature field sequence, a multimodal original dataset aligned with spatiotemporal reference is output.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-6 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-6.

Images