Urban rail transit tunnel detection robot autonomous navigation and multi-modal defect identification system

By working collaboratively with multi-sensor perception modules and edge computing units, and dynamically adjusting sensor fusion weights, combined with infrared thermal imaging for verification, autonomous navigation and multimodal defect identification of the tunnel inspection robot have been achieved. This solves the problems of positioning accuracy and defect identification in tunnel inspection technology and supports the routine and intelligent operation and maintenance of tunnel structures.

CN123066085APending Publication Date: 2026-11-102ND ENG CO LTD OF CHINA RAILWAY 12TH BUREAU GRP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611175476.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-04
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing tunnel inspection technologies have significant limitations in areas such as establishing positioning benchmarks, identifying multimodal defects and quantifying physical dimensions, and accurately calibrating mileage, making it difficult to support the transformation of urban rail transit tunnel structures towards routine and intelligent operation and maintenance.

Method used

Employing a multi-sensor perception module, edge computing unit, and cloud-based collaborative analysis platform, the system achieves autonomous navigation and multi-modal defect identification through degradation perception adaptive navigation and positioning, multimodal defect identification and 3D back-projection measurement, and mileage marker geocoding. It dynamically adjusts sensor fusion weights and combines infrared thermal imaging for verification to output the true physical dimensions and accurate mileage markers of the defects.

Benefits of technology

It improves the autonomous navigation and positioning accuracy of tunnel inspection robots, reduces the false detection and missed detection rates of defects, reduces the workload of manual verification, realizes accurate positioning and efficient tracking of defect locations, and supports the transformation of tunnel structures from periodic manual inspection to routine intelligent inspection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123066085A_ABST
    Figure CN123066085A_ABST
Patent Text Reader

Abstract

This invention discloses an autonomous navigation and multimodal defect recognition system for an urban rail transit tunnel inspection robot. The system includes a mobile platform, a multi-sensor perception module, an edge computing unit, and a cloud-based collaborative analysis platform. The edge computing unit operates a degradation perception adaptive navigation and positioning module, a multimodal defect recognition and 3D back-projection measurement module, and a mileage marker geocoding module. The navigation and positioning module calculates degradation factors based on the feature values ​​of the point cloud registration information matrix and dynamically adjusts the weights of each sensor factor to suppress positioning drift under tunnel degradation environments. The defect recognition module integrates visible light, infrared, and laser point cloud detection to obtain the true physical dimensions through back-projection, and combines infrared verification and confidence level fusion to reduce the false detection rate. The mileage coding module converts defect coordinates into mileage markers based on the route shape. This invention improves positioning accuracy and defect recognition reliability, and has engineering applicability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for rail transit infrastructure, and in particular to an autonomous navigation and multimodal defect recognition system for urban rail transit tunnel inspection robots. Background Technology

[0002] As the backbone of modern urban public transportation, the safety and stability of urban rail transit's underground tunnel structures directly affect the overall operational safety of the rail transit system. With the increasing years of operation and the continuous extension of the operational mileage of urban rail transit, underground tunnel structures inevitably develop structural defects such as lining cracks, water leakage, and spalling under the combined effects of long-term train vibration loads, groundwater erosion, environmental disturbances, and natural material aging. If these defects are not detected and addressed promptly and accurately, they can easily escalate into serious structural safety hazards. Therefore, conducting periodic and routine inspections of the tunnel structure's health status is a crucial step in ensuring the safe and stable operation of rail transit.

[0003] However, traditional tunnel structure inspections mainly rely on manual nighttime walking inspections, supplemented by handheld instruments for on-site measurements. This traditional inspection mode has many insurmountable technical bottlenecks and engineering pain points.

[0004] Firstly, regarding navigation and positioning accuracy, subway tunnels are typical long-distance, enclosed, geometrically degraded environments. The lack of rich geometric features within tunnels makes traditional inspection equipment or robots highly susceptible to severe axial positioning drift when traversing long, straight tunnel sections. This leads to a significant decrease in navigation and positioning accuracy, failing to provide a stable and reliable spatial reference for subsequent precise location of defects. Even though some existing solutions attempt to employ multi-sensor fusion technology, their fixed-weight fusion strategies lack adaptability to geometrically degraded environments and remain insufficiently robust when facing sections with missing features, making it difficult to achieve stable and high-precision positioning throughout the entire process.

[0005] Secondly, in terms of defect identification and dimensional measurement, existing detection methods mostly rely on single visible light image recognition. Limited by the dim lighting environment inside tunnels and the complex wall conditions, relying solely on visible light images is easily interfered with by reflections from oil stains, water stains, and shadows on the walls, leading to high false alarm and false negative rates for typical defects such as water leakage. More importantly, traditional visual image processing methods typically only output the pixel-level dimensions of defects in the image, failing to directly convert them into actual physical dimensions with practical engineering maintenance significance. This results in a lack of objective quantitative basis for assessing the severity of defects, requiring substantial manual on-site measurement.

[0006] Furthermore, in terms of locating and recording defects, traditional inspection records often rely on manual reading of mileage markers or estimation by pacing, resulting in relatively large positioning errors. This crude method of location marking cannot be closely integrated with the design data of horizontal and vertical curves of the railway line, making it impossible to accurately map defect locations to the actual design mileage markers and segment structural units. This makes it difficult to meet the maintenance department's actual needs for precise defect location and efficient tracking.

[0007] Finally, in terms of data processing and system architecture, existing embedded inspection equipment is generally limited by computing power bottlenecks, making it difficult to meet the real-time processing requirements of robots during their movement while ensuring high detection accuracy. In addition, existing systems often lack effective mechanisms for continuous model evolution and closed-loop optimization, resulting in insufficient scene adaptability and long-term generalization ability when faced with complex and ever-changing actual tunnel defects.

[0008] In summary, existing tunnel inspection technologies have significant limitations in areas such as establishing positioning benchmarks, identifying multimodal defects and quantifying physical dimensions, accurately calibrating mileage, and coordinating computing architectures. These limitations make it difficult to support the deep transformation of urban rail transit tunnel structures towards routine and intelligent operation and maintenance. There is an urgent need to develop an intelligent solution that can take into account adaptive navigation and positioning, multimodal defect quantification and measurement, and accurate geocoding. Summary of the Invention

[0009] The purpose of this invention is to provide an autonomous navigation and multimodal defect recognition system for urban rail transit tunnel inspection robots, so as to solve the technical problems existing in the prior art.

[0010] This invention provides an autonomous navigation and multimodal defect recognition system for an urban rail transit tunnel inspection robot, comprising a mobile platform, a multi-sensor perception module mounted on the mobile platform, an edge computing unit, and a cloud-based collaborative analysis platform. The multi-sensor perception module includes a 3D LiDAR, a binocular visible light camera, an infrared thermal imager, an inertial measurement unit, and a wheeled odometer encoder. Each sensor completes time synchronization calibration and extrinsic parameter calibration. The edge computing unit operates a degradation perception adaptive navigation and positioning module, a multimodal defect recognition and 3D back-projection measurement module, and a mileage marker geocoding module, wherein:

[0011] The degradation perception adaptive navigation and positioning module is used to calculate the point cloud degradation factor based on the ratio of the minimum eigenvalue to the maximum eigenvalue of the point cloud registration equation information matrix, and dynamically adjust the weights of the laser odometry factor, inertial measurement unit factor, wheel odometry factor and closed-loop detection factor in the factor graph optimization according to the degradation factor through the S-shaped function, and output the robot's real-time pose.

[0012] The multimodal defect identification and three-dimensional back-projection measurement module is used to align visible light images, infrared thermal images and laser point clouds in time and space, detect tunnel defects using a lightweight target detection network, back-project the two-dimensional detection box onto the three-dimensional point cloud to obtain the true physical size of the defects, verify suspected water leakage defects by combining infrared temperature distribution, and perform weighted fusion of multimodal confidence scores.

[0013] The mileage marker geocoding module is used to perform nearest point projection matching between the three-dimensional coordinates of the defect and the route design alignment database, and output the accurate mileage marker of the defect by combining arc length integral and interpolation correction.

[0014] As a further aspect of the present invention: the degradation perception adaptive navigation and positioning module calculates the point cloud degradation factor by performing eigenvalue decomposition on the information matrix of the point cloud registration residual equation, and taking the ratio of its minimum eigenvalue to its maximum eigenvalue as the degradation factor. The smaller the degradation factor, the weaker the geometric constraint of the current point cloud in the corresponding degree of freedom direction. The information matrix is ​​obtained by multiplying the Jacobian matrix of the point cloud registration residual with respect to the pose increment and its transpose.

[0015] As a further aspect of the present invention: the weight coefficient of the laser odometry factor is obtained by mapping the degradation factor through an S-shaped function, and it transitions smoothly and continuously near a preset degradation threshold. When the degradation factor is lower than the preset degradation threshold, the weight coefficient decreases, and when the degradation factor is higher than the preset degradation threshold, the weight coefficient approaches the maximum value. The degradation-aware adaptive navigation and positioning module is also used to calculate the eigenvalue ratio of the information matrix according to the rotation sub-block and the translation sub-block respectively, to obtain independent degradation factors in multiple degrees of freedom directions, and to set corresponding weight coefficients for each degree of freedom direction.

[0016] As a further aspect of the present invention: the overall objective function of the factor graph optimization is composed of a weighted sum of the laser odometer factor, the inertial measurement unit pre-integration factor, the wheel odometer factor, and the closed-loop detection factor, wherein the weight of the laser odometer factor is dynamically adjusted in accordance with the manner described in claim 3, and the weights of the other factor terms are preset according to the measurement covariance of the corresponding sensors.

[0017] As a further aspect of the present invention: In the multimodal defect identification and three-dimensional back projection measurement module, for crack-type defects, multiple locations are sampled at equal intervals along the crack centerline in its normal direction. The Euclidean distance between the corresponding points of the crack edge in the three-dimensional point cloud at each sampling location is calculated as the crack width at that location. The average and maximum values ​​of the crack widths at multiple sampling locations are taken, and the average and maximum widths of the cracks are output.

[0018] As a further aspect of the present invention: the infrared thermal imager is used to acquire the temperature distribution image of the tunnel lining surface. For areas suspected of water leakage detected in the visible light image, the temperature distribution of the corresponding position in the infrared thermal image is extracted. When the temperature difference between the area and the surrounding background exceeds a preset threshold and shows a continuous seepage pattern, it is determined to be a water leakage defect; otherwise, the confidence level of the defect is reduced.

[0019] As a further aspect of the present invention: the multimodal defect identification and three-dimensional back-projection measurement module converts the visible light detection confidence, infrared verification confidence, and laser point cloud reflection intensity normalized confidence into logarithmic probability form, then weights and sums them, and finally obtains the fused confidence through S-shaped function mapping; the weight coefficients corresponding to each confidence are adaptively updated according to the accuracy of the corresponding modality in historical detections verified by the cloud collaborative analysis platform in a sliding window manner, and the higher the recent identification accuracy of a certain modality, the larger its corresponding weight coefficient.

[0020] As a further aspect of the present invention: the mileage marker geocoding module pre-converts the horizontal and vertical curve parameters in the track design file into a three-dimensional parameter curve of the track centerline with the design mileage as the parameter. It then solves the nearest projection point of the three-dimensional coordinates of the defect on the parameter curve through numerical iteration. Finally, it interpolates and corrects the projected mileage by combining the robot pose and the accumulated mileage of the wheeled mileage encoder, and outputs the accurate mileage marker of the defect.

[0021] As a further aspect of the present invention: the edge computing unit deploys a lightweight target detection network, the backbone network of which is composed of stacked depthwise separable convolutional units and embedded with a channel-spatial dual attention module; the cloud-based collaborative analysis platform is used to perform super-resolution enhancement and secondary judgment of the disease samples determined by the edge computing unit to have low confidence, and the verification results are used for incremental training and model update of the lightweight target detection network.

[0022] As a further aspect of the present invention, a method for autonomous navigation and multimodal defect identification of an urban rail transit tunnel inspection robot includes the following steps:

[0023] Collect 3D point cloud data, visible light images, infrared thermal images, and robot motion data inside the tunnel;

[0024] Motion distortion compensation and feature extraction are performed on the point cloud, and the laser odometry pose estimation is obtained by matching it with the local sub-map. The point cloud degradation factor is also calculated.

[0025] The weights of the laser odometry factor are dynamically adjusted based on the point cloud degradation factor. The laser odometry factor, inertial measurement unit factor, wheeled odometry factor, and closed-loop detection factor are jointly optimized by the degradation perception adaptive navigation and positioning module into the factor graph, and the robot's real-time pose is output.

[0026] Spatiotemporal alignment of visible light images, infrared thermal images, and laser point clouds is performed, and a lightweight target detection network deployed in the edge computing unit detects tunnel defects and outputs two-dimensional detection boxes.

[0027] The two-dimensional detection frame is back-projected onto the three-dimensional point cloud to calculate the true physical size of the defect. The suspected water leakage defect is verified by combining the infrared temperature distribution, and the multimodal confidence is weighted and fused.

[0028] The three-dimensional coordinates of the defects are matched with the nearest point projection database of the route design, and the accurate mileage station number of the defects is output and written into the tunnel structure defect database.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] By calculating the point cloud degradation factor in real time and dynamically adjusting the multi-sensor fusion weights, the positioning drift in the geometrically degraded tunnel environment is effectively suppressed, improving the accuracy and robustness of the robot's autonomous navigation and positioning, and laying the foundation for accurate fault location. Field testing in a long, straight tunnel section verified that, compared to a comparative scheme using a fixed-weight multi-sensor fusion strategy, the proposed method significantly reduces the cumulative positioning drift in long, straight, degraded sections, and the end-to-end positioning accuracy can be stably controlled within a high accuracy range.

[0031] By fusing multimodal data and back-projecting two-dimensional detection results to three-dimensional point clouds, the automatic quantitative measurement of the true physical size of the disease was realized, avoiding the limitation of traditional solutions that can only provide pixel-level size. This provides a quantitative basis for disease classification and assessment. Compared with the results of manual caliper measurement, the relative error of the automatic crack width measurement results can be controlled within a small range.

[0032] By combining infrared thermal imaging-assisted verification with multimodal confidence fusion, the false positive and false negative rates of defects such as water leakage are significantly reduced. Comparative tests show that, compared to a single-modal detection scheme using only visible light images, the method proposed in this invention significantly reduces the false alarm rate for water leakage defects and significantly improves the reliability of the detection results.

[0033] By using a mileage marker geocoding method based on the route design alignment, the accuracy of disease location is improved from the low accuracy of traditional manual pacing to a higher accuracy, significantly reducing the workload of manual verification.

[0034] By combining edge-lightweight detection with cloud-based collaborative verification, the system ensures real-time detection while moving and maintains detection accuracy. Furthermore, the incremental training mechanism enables continuous optimization of the detection model, demonstrating good engineering scalability.

[0035] The degradation-aware adaptive navigation and positioning method, multimodal defect identification and three-dimensional back-projection measurement method, and mileage marker geocoding method proposed in this invention work together synergistically: precise navigation and positioning provides a reliable benchmark for calculating the three-dimensional coordinates of defects; the combination of defect descriptions with accurate physical dimensions and precise mileage markers enables the detection results to be directly integrated with existing tunnel structure maintenance records and full life-cycle management systems, reducing data conversion and manual verification steps. This provides feasible technical support for the transformation of urban rail transit tunnel structures from periodic manual inspections to routine intelligent inspections. The aforementioned beneficial effects are not simply the sum of individual technical features, but rather a synergistic technical effect resulting from the mutual support and joint action of the three links: degradation-aware adaptive navigation and positioning, multimodal back-projection measurement, and mileage marker geocoding. Without a precise navigation and positioning benchmark, the accuracy of calculating the three-dimensional coordinates and actual physical dimensions of defects cannot be guaranteed; without a quantitative description of actual physical dimensions, the precise location information provided by mileage marker geocoding cannot fully realize its value in maintenance decision-making. These three elements together constitute an inseparable technical whole. Attached Figure Description

[0036] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0037] Figure 1 This is the overall architecture diagram of the autonomous navigation and multimodal defect recognition system for urban rail transit tunnel inspection robots of the present invention.

[0038] Figure 2 This is a flowchart of the degradation-aware adaptive navigation and positioning method in this invention.

[0039] Figure 3 This is a flowchart of the multimodal defect identification and three-dimensional back projection measurement method in this invention.

[0040] Figure 4 This is a schematic diagram of the factor graph optimization structure in this invention.

[0041] Figure 5 This is a timing diagram illustrating the principle of three-dimensional back-projection measurement of defects in this invention.

[0042] Figure 6 This is a flowchart of the mileage marker geocoding method in this invention.

[0043] Figure 7 This is a diagram illustrating the collaborative architecture between the edge computing unit and the cloud-based collaborative analysis platform in this invention.

[0044] 1-Mobile platform; 2-Multi-sensor perception module; 21-3D LiDAR; 22-Binocular visible light camera; 23-Infrared thermal imager; 24-Inertial measurement unit; 25-Wheel odometer encoder; 3-Degradation perception adaptive navigation and positioning module; 4-Multimodal defect identification and 3D back projection measurement module; 5-Mileage marker geocoding module; 6-Edge computing unit; 7-Cloud collaborative analysis platform. Detailed Implementation

[0045] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0046] Overall system composition:

[0047] like Figure 1 As shown, the autonomous navigation and multimodal defect recognition system for urban rail transit tunnel inspection robots provided by the present invention includes a mobile platform 1, a multi-sensor perception module 2, a degradation perception adaptive navigation and positioning module 3, a multimodal defect recognition and three-dimensional back projection measurement module 4, a mileage marker geocoding module 5, an edge computing unit 6, and a cloud-based collaborative analysis platform 7.

[0048] Mobile platform 1 adopts a wheel-track composite chassis with differential drive wheel sets on both sides and follow-up guide wheels at the front and rear. A lifting column and a pitching platform are located in the center of the chassis. A multi-sensor sensing module 2 is mounted on top of the platform. The lifting stroke and pitching angle of the platform can be pre-set with a detection trajectory based on the tunnel cross-section, which may be horseshoe-shaped, circular, rectangular, etc., to ensure full coverage detection of key areas such as the arch, sidewalls, and track bed. Mobile platform 1 is also equipped with an explosion-proof battery pack, an automatic charging docking mechanism, and obstacle avoidance sensors, enabling continuous operation and autonomous return for charging.

[0049] The overall control architecture of the mobile platform 1 adopts a two-layer structure of main control and motion control: the main control layer runs on the edge computing unit 6, responsible for multi-sensor data acquisition and scheduling, navigation and positioning calculation, and defect detection and identification; the motion control layer uses an independent motion controller, which receives speed and steering commands from the main control layer and controls the drive wheel speed in a closed loop, while feeding back status variables such as chassis current, voltage, and motor temperature to the main control layer for health monitoring. The rise and fall and pitch movements of the gimbal are also independently controlled in a closed loop by the motion control layer, and are time-linked with the detection process of the main control layer. That is, when it reaches the preset detection section, the gimbal attitude is automatically adjusted, and the synchronous acquisition of the multi-sensor perception module 2 is triggered after the attitude is stable, so as to avoid the adverse effects of gimbal motion blur on image and point cloud quality.

[0050] The multi-sensor perception module 2 includes a 3D LiDAR 21, a binocular visible light camera 22, an infrared thermal imager 23, an inertial measurement unit 24, and a wheeled odometer 25. The 3D LiDAR 21, employing either a mechanical rotating or solid-state LiDAR, is used to acquire 3D point cloud data of the tunnel wall and track bed. The binocular visible light camera 22 acquires high-resolution images of the tunnel lining surface and uses binocular parallax to estimate near-distance depth. The infrared thermal imager 23 acquires images of the temperature distribution on the tunnel lining surface to assist in identifying water seepage defects. The inertial measurement unit 24 acquires the angular velocity and linear acceleration information of the robot body. The wheeled odometer 25 acquires the robot's walking distance and speed information. Each sensor undergoes time synchronization calibration and extrinsic parameter calibration before leaving the factory and before each inspection operation. Time synchronization calibration uses a unified timestamp for timing, while extrinsic parameter calibration acquires the rotational and translational relationships between the coordinate systems of each sensor and the robot body coordinate system to ensure the consistency of multimodal data in time and space.

[0051] The extrinsic parameter calibration between the 3D LiDAR 21 and the binocular visible light camera 22 is completed using a joint optimization method based on the corner points of the calibration board and the edge features of the laser point cloud. A calibration board of known size is deployed in the calibration scene, and the corner pixel coordinates of the calibration board in the image, as well as its plane normal and boundary features in the point cloud, are extracted. The rotation and translation extrinsic parameters between the camera and the LiDAR are solved by minimizing the sum of the corner back-projection error and the plane fitting error. Due to differences in field of view and resolution between the infrared thermal imager 23 and the binocular visible light camera 22, a pixel-level mapping lookup table is further established using a commonly observable heated calibration target for pixel alignment of subsequent temperature and visible light images. All sensor data acquisition includes a unified hardware synchronization timestamp, with the timestamp reference uniformly timed by the edge computing unit 6. The synchronization error is controlled within milliseconds to avoid multimodal data misalignment due to inconsistent acquisition times.

[0052] The body and electrical components of the mobile platform 1 are designed according to the protection level requirements of the subway operating environment, and have dustproof, moisture-proof and certain electromagnetic compatibility capabilities. The obstacle avoidance sensor adopts a combination of ultrasonic and two-dimensional laser ranging to detect protruding obstacles within the track area clearance and trigger deceleration or stopping, ensuring the safety of the robot's autonomous movement in a limited space.

[0053] Degradation-aware adaptive navigation and localization method

[0054] like Figure 2 and Figure 4 As shown, the workflow of the degradation-aware adaptive navigation and positioning module 3 includes the following steps:

[0055] Step S1, point cloud preprocessing: The 3D LiDAR 21 collects point clouds at a fixed frame rate, and uses the angular velocity and acceleration integral of the inertial measurement unit 24 to perform motion distortion compensation on the single frame point cloud, eliminating the point cloud deformation caused by the difference in scanning timing during the robot's movement.

[0056] Step S2, Feature Extraction and Point Cloud Registration: Extract the set of edge feature points and the set of planar feature points from the distortion-compensated point cloud, perform nearest neighbor matching with the corresponding features in the local sub-map, construct the residual equations from point to line and from point to surface, and solve the pose increment through Gauss-Newton iteration to obtain the laser odometry pose estimate.

[0057] Step S3, Degradation Factor Calculation: After each point cloud registration iteration converges, extract the normal equation information matrix. , Jacobian matrix of residual with respect to pose increment Composition, that is, H equals transpose multiplied by ,right Perform eigenvalue decomposition, take the minimum and maximum eigenvalues, and calculate the degradation factor. The calculation method is shown in formula (1). Degeneration factor The value range is from 0 to 1. The smaller the value, the weaker the geometric constraint of the current point cloud in a certain degree of freedom direction, that is, the higher the degree of degradation; A larger value indicates that the point cloud is more balanced in all directions, and the registration result is more reliable.

[0058] Formula (1):

[0059]

[0060] in, For the first The point cloud degradation factor during secondary point cloud registration is used to quantitatively describe the strength of the geometric constraints of the current point cloud in each degree of freedom direction in space. The normal equation information matrix in the point cloud registration process is obtained by multiplying the Jacobian matrix of the residual with respect to the pose increment and its transpose. The Jacobian matrix of the registration residual with respect to pose increment for the point cloud; and Information matrix The smaller the ratio of the minimum and maximum eigenvalues, the weaker the geometric constraints of the current point cloud in the corresponding degree of freedom direction, i.e., the higher the degree of degradation.

[0061] Furthermore, considering that tunnel degradation typically manifests as insufficient constraints on translational degrees of freedom along the tunnel axis, while rotational and lateral translational degrees of freedom around other axes often retain good geometric constraints, calculating only a single global scalar degradation factor is insufficient to accurately characterize the differences in constraint strength across different degrees of freedom. Therefore, this invention further decomposes the six-degree-of-freedom information matrix H into rotational and translational sub-blocks, obtaining rotational and translational subspace degradation factors respectively. It can also further calculate directional degradation factors along the tunnel axis, lateral direction, and vertical direction within the translational subspace, forming a six-dimensional degradation factor vector. Independent adaptive weights are set for the six degrees of freedom components of the laser odometry factor in the factor graph optimization, allowing the constraint weights along the tunnel axis (stronger degradation directions) to be suppressed individually, while the high-precision constraints of the laser odometry are retained in the lateral and vertical directions (weaker degradation directions). Compared to using only a single global scalar degradation factor, this approach more precisely preserves the still-effective geometric constraint information in the point cloud, further improving positioning accuracy while ensuring robustness.

[0062] Step S4, Adaptive Weight Calculation: Based on the degradation factor D, the weight coefficient of the laser odometry factor in subsequent factor graph optimization is calculated using a smooth S-shaped function, as shown in Formula (2). When the degradation factor D is lower than the preset threshold, the weight coefficient decays rapidly, making the factor graph optimization rely more on the constraints of the inertial measurement unit factor and the wheel odometry factor; when the degradation factor D is higher than the preset threshold, the weight coefficient approaches 1, giving full play to the high-precision constraint effect of the laser odometry. This weight function transitions smoothly and continuously near the threshold, avoiding pose jumps caused by fixed threshold switching.

[0063] Formula (2):

[0064]

[0065] in, , which are the dynamic weighting coefficients of the laser odometry factor, used to adjust the confidence level of the laser constraint according to the degree of degradation in factor graph optimization; Degradation factor registered for the current point cloud; This is the steepness coefficient of the S-shaped function, a preset positive number used to control the smoothness of the weight transition near the degradation threshold; A preset degradation threshold is set. When the degradation factor is below the threshold, the weight coefficient decays rapidly. When it is above the threshold, the weight coefficient approaches the maximum value of 1, thereby achieving a continuous and smooth transition and avoiding pose jumps caused by sudden weight changes.

[0066] Step S5, Factor Graph Joint Optimization: The laser odometry factor, inertial measurement unit pre-integration factor, wheel odometry factor and closed-loop detection factor are jointly connected to the factor graph to construct the overall optimization objective function, as shown in Formula (3). The optimal pose estimation of each key frame is solved by sliding window nonlinear optimization, and the real-time six-degree-of-freedom pose of the robot in the tunnel coordinate system is output. The weight of the laser odometry factor is dynamically adjusted according to step S4.

[0067] Formula (3):

[0068]

[0069] in, The overall objective function for joint optimization of factor graphs is used to solve for the optimal six-DOF pose state variables of the robot by minimizing this function. ; The state variables of the robot's six degrees of freedom pose to be optimized include three-dimensional position coordinates and three-dimensional attitude angles; This is the index number of the keyframe; For the first Dynamic weighting coefficients of the frame laser odometry factor; For the first The frame laser odometry residual vector represents the deviation between the laser point cloud registration estimated pose and the optimization variables; For the first The frame inertial measurement unit pre-integration residual vector represents the deviation between the IMU pre-integrated measurement value and the pose variable; For the first The frame wheel odometer residual vector represents the deviation between the displacement measured by the wheel encoder and the pose variable; For the first Each closed-loop detection residual vector represents the pose deviation under closed-loop matching constraints. , , , These are information matrices corresponding to the laser odometry, IMU pre-integration, wheel odometry, and closed-loop detection constraints, used for Mahalanobis distance weighting of each residual; The square of the Mahalanobis distance is given by the following expansion: This allows observations with lower noise to gain greater influence weight in the optimization process.

[0070] Step S6, closed-loop detection and global consistency optimization: When the robot returns to the explored area or passes through characteristic nodes, such as turnouts or connecting passage entrances, it detects closed loops by matching point cloud descriptors and optimizes the global pose map to eliminate pose drift accumulated during long-distance inspections.

[0071] Through the aforementioned degradation-aware adaptive navigation and positioning method, the robot can effectively suppress axial positioning drift in the degradation section of a long straight tunnel, and make full use of the high-precision constraints of laser point clouds in sections with rich geometric features, thus balancing positioning robustness and accuracy, and providing a reliable pose reference for the precise spatial positioning of subsequent defects.

[0072] Multimodal Defect Identification and Three-Dimensional Back Projection Measurement Method

[0073] like Figure 3 and Figure 5 As shown, the workflow of the multimodal defect identification and 3D back-projection measurement module 4 includes the following steps:

[0074] Step T1, Spatiotemporal Alignment of Multimodal Data: Based on the acquisition time of the 3D LiDAR 21, the images acquired by the binocular visible light camera 22, the temperature images acquired by the infrared thermal imager 23, and the laser point cloud at the same time are spatially registered at the pixel level through sensor extrinsic calibration results and timestamp interpolation, and a mapping relationship between image pixel coordinates and point cloud depth values ​​is established.

[0075] Step T2, Lightweight Preliminary Edge Detection: The aligned visible light image is input into the lightweight target detection network deployed in the edge computing unit 6. The network adopts a lightweight backbone network constructed with depthwise separable convolution and embeds a channel-space dual attention module to output the disease categories such as cracks, water leakage, peeling and flaking, misalignment, etc., as well as their two-dimensional detection boxes and preliminary confidence scores.

[0076] Specifically, the backbone of the lightweight target detection network consists of stacked multi-level depthwise separable convolutional units. Each level includes two sub-layers: channel-wise convolution and pointwise convolution. This significantly reduces the number of parameters and computational cost while retaining multi-scale feature extraction capabilities. A channel-space dual attention module is embedded between the shallow and mid-level feature maps of the backbone network. This module first performs global average pooling and max pooling on the feature map along the channel dimension, then learns the channel weights through a shared fully connected layer. Next, it performs convolution on the weighted feature map along the spatial dimension to learn the spatial weights. This enhances the feature response of the network on thin, weakly textured targets such as cracks and targets with blurred boundaries such as water seepage, while suppressing the feature response of background interference areas such as reflections and oil stains in tunnels. The detection head uses a decoupled detection head without anchor boxes, outputting the probability of the disease category, the regression parameters of the detection box, and the confidence score, balancing inference speed and small target detection accuracy. When the network runs on the edge computing unit 6, operator fusion and half-precision inference further compress the single-frame inference time to meet the real-time requirements of the robot's on-the-go detection.

[0077] Step T3, 3D back-projection size measurement: For each 2D detection box, using the pixel-depth mapping relationship and camera intrinsic parameter matrix established in step T1, the defect contour pixels in the detection box are back-projected to the 3D space to obtain the 3D point set of the defect; for crack-type defects, the center line is extracted along the crack direction and sampled at equal intervals in its normal direction, the true width of the crack at each sampling position is calculated and the mean and maximum values ​​are obtained, see formula (4) and formula (5); for surface defects such as water leakage and peeling, the projection area of ​​its 3D point set on the structural surface is calculated as the defect area.

[0078] Formula (4):

[0079]

[0080] in, pixel coordinates The corresponding three-dimensional spatial coordinate point in the camera coordinate system; For this pixel point The corresponding depth value is provided by the spatiotemporally aligned laser point cloud; This is the camera intrinsic parameter matrix, which includes focal length, principal point coordinates, etc., and is obtained from the pre-completed camera calibration. This is the inverse of the camera intrinsic matrix, used to transform pixel homogeneous coordinates to the normalized camera plane; These are the pixel coordinates in the image, in pixels.

[0081] Formula (5):

[0082]

[0083] in, This represents the average width of crack-type defects; This represents the maximum width of crack-type defects. This represents the total number of sampling locations along the normal direction of the crack centerline; and They were respectively in the second At each sampling location, the spatial coordinates of the corresponding pixel points on both sides of the crack are back-projected into three-dimensional space; This is the Euclidean distance, used to calculate the true distance between two points in three-dimensional space; For the first The crack width calculated at each sampling location is the Euclidean distance between the three-dimensional points on both sides of that location.

[0084] Step T4, Infrared Temperature Assisted Verification: For areas suspected of water leakage by visible light detection, extract the temperature distribution of the corresponding position in the infrared thermal image. If the temperature difference between the area and the surrounding background exceeds a preset threshold and shows a continuous seepage pattern, it is determined to be a real water leakage problem; otherwise, it is determined to be a false detection and the confidence level of the problem is reduced.

[0085] Step T5, multimodal confidence fusion: the visible light detection confidence, infrared verification confidence and the normalized confidence of the laser point cloud reflection intensity in the region are converted into logarithmic probability form and then weighted and summed, and then the fusion confidence is obtained by S-shaped function mapping, see formula (6); disease samples with fusion confidence below the preset threshold are marked as difficult samples and uploaded to the cloud collaborative analysis platform 7 for secondary verification.

[0086] Formula (6):

[0087]

[0088] in, The confidence level for multimodal fusion is the final output after weighted fusion of the confidence levels of visible light detection, infrared verification, and laser point cloud reflection intensity. The value range is (0,1), which is used to comprehensively determine whether a certain disease truly exists. The sigmoid function maps any real value after weighted summation to the interval (0,1), and its specific form is: ; For the first The log-odds form of the modal confidence scores, which originally had confidence scores in the range of (0,1). The conversion formula for the entire real number field is as follows: ; For the first The original confidence level of each modality =1, 2, 3 correspond to the confidence levels of visible light detection, infrared verification, and laser point cloud reflectance intensity normalization, respectively. , , The dynamic weight coefficients corresponding to the confidence levels of the three modalities are adaptively updated based on the performance of each modality in historical detections. Modalities with higher accuracy receive larger weight coefficients. The three weights, after normalization, satisfy the following conditions: + + =1.

[0089] In formula (6), the weight coefficients corresponding to the confidence levels of each modality are not fixed, but are adaptively adjusted according to the performance of each modality in historical detection: After the cloud-based collaborative analysis platform 7 completes manual confirmation or secondary network judgment of the uploaded difficult samples, it feeds back the verification results to the edge computing unit 6. The edge computing unit 6 calculates the consistency rate between the judgment results and verification results of the three modalities of visible light detection, infrared verification, and laser reflection intensity verification according to a preset time window, which is used as the detection accuracy rate of the corresponding modality within the time window; then, the three weight coefficients are normalized and updated according to the accuracy rate, so that the modality with a higher detection accuracy in the recent period obtains a larger weight coefficient in the subsequent fusion process, and the weight of the modality with a lower detection accuracy is reduced accordingly. This adaptive weight update mechanism enables the system to automatically adjust the fusion strategy for changes in the detection reliability of each modality under different tunnel cross-section forms, different lighting and temperature and humidity conditions, avoiding the problem of the fusion effect decreasing when the fixed weight scheme changes environmental conditions.

[0090] Step T6, Disease Record Generation: The disease category, actual physical size, fusion confidence level, three-dimensional spatial coordinates, and the image, point cloud, and temperature image at the time of acquisition are encapsulated into a structured disease record, which awaits the mileage marker geocoding module 5 for mileage marker calibration.

[0091] Through the above method, the present invention improves the pixel-level description of defects that can only be obtained in traditional solutions to a quantitative description with real physical dimensions, and effectively reduces the false detection rate and false negative rate by infrared temperature-assisted verification and multimodal confidence fusion.

[0092] Geographic coding method for mileage markers

[0093] like Figure 6 As shown, the workflow of the mileage marker geocoding module 5 includes the following steps:

[0094] Step U1, Line Design Alignment Input: Convert the horizontal curve and vertical curve parameters and starting mileage in the track line design file into a three-dimensional parameter curve of the line centerline with the design mileage as the parameter, and store it in the line design alignment database.

[0095] Step U2, Nearest point projection matching: For the three-dimensional spatial coordinates in the defect record, the projection point with the smallest Euclidean distance is obtained by numerical iteration on the line centerline parameter curve, as shown in formula (7), and the design mileage parameter corresponding to the projection point is obtained.

[0096] Formula (7):

[0097]

[0098] in, The design mileage parameter corresponding to the projection point that minimizes the objective function is the precise mileage station number of the defect on the track line. The three-dimensional spatial coordinates of the defect to be projected are output by the multimodal defect recognition module; To design mileage The three-dimensional parametric curve of the track centerline is obtained by converting pre-entered horizontal and vertical curve data. , , For the parameter curve Mileage The corresponding three-dimensional spatial coordinate components at that location; Euclidean distance is used to measure the spatial distance between the disease coordinates and the projection points on the curve; The variable evaluation operator is used to minimize the objective function, i.e., to find the mileage parameter that minimizes the spatial distance. .

[0099] Step U3, Mileage Station Interpolation Calibration: Combining the pose of the robot body output by the degradation perception adaptive navigation and positioning module 3 at this moment and the cumulative mileage of the wheel mileage encoder 25, the projected mileage is interpolated and corrected to eliminate the mileage conversion error caused by the curve shape of the line, and finally the accurate mileage station of the defect is output.

[0100] Step U4, Ledger Matching and Database Entry: Match the calibrated mileage markers with the section, ring number, or sleeper number in the existing tunnel structure maintenance ledger, and write the defect records into the tunnel structure defect database for maintenance management personnel to query and review.

[0101] Using the above method, the present invention improves the accuracy of disease mileage positioning from the meter level of traditional manual pacing to the centimeter level, significantly reducing the workload of manual verification.

[0102] Edge-cloud collaborative architecture

[0103] like Figure 7As shown, the edge computing unit 6 is equipped with a degradation-aware adaptive navigation and positioning module 3 and a lightweight target detection network. This module is responsible for real-time pose calculation and preliminary defect identification during the robot's on-the-fly detection process, ensuring real-time performance even at a detection speed of at least 0.3 meters per second. The cloud-based collaborative analysis platform 7 is equipped with a super-resolution enhancement network and a high-precision verification detection network. These networks are used to perform secondary assessments on problematic samples uploaded by the edge computing unit 6. The cloud-based collaborative analysis platform 7 also periodically summarizes verification results and manually confirmed annotations, incrementally trains the lightweight edge detection network, and sends the updated model parameters back to the edge computing unit 6, forming a closed-loop optimization mechanism for edge-cloud collaboration. This ensures that detection accuracy continuously improves over time.

[0104] Considering the significant attenuation of wireless signals within the tunnel and the potential communication blind spots in some sections, data transmission between the edge computing unit 6 and the cloud-based collaborative analysis platform 7 utilizes existing leaky cables or distributed wireless access points within the tunnel as communication links. A local data cache queue is set up on the edge computing unit 6: when the robot is in a communication blind spot, problematic samples and defect records are temporarily stored in the local cache queue. Once communication is restored, they are uploaded in batches according to priority, with samples with lower confidence levels and higher defect severity being uploaded first, ensuring the continuity of the inspection operation is not affected by fluctuations in communication conditions. Furthermore, after each inspection task is completed, the edge computing unit 6 automatically generates a list of defects and a summary of the pose trajectory for that inspection. Even when the edge computing unit 6 and the cloud-based collaborative analysis platform 7 are completely offline, the inspection report can still be exported to local storage media for use by on-site personnel, ensuring the system's availability under extreme communication conditions.

[0105] Example 1: Routine Inspection of Subway Shield Tunnel During Nighttime Maintenance Window

[0106] This embodiment uses a circular cross-section tunnel constructed using the shield tunneling method in a city's subway system as an application scenario. The robot is deployed for inspection during nighttime maintenance windows. The mobile platform 1 autonomously walks along the tunnel evacuation platform, with a detection speed set at 0.4 meters per second. After entering a long, straight tunnel section approximately 1.2 kilometers long, the degradation perception adaptive navigation and positioning module 3 continuously calculates the point cloud degradation factor. If the degradation factor along the tunnel axis is detected to be below a preset threshold of 0.15, the system automatically reduces the weight of the laser odometry factor in that direction, relying more on the constraints of the inertial measurement unit and wheeled odometry for pose estimation. When the robot passes through sections with rich geometric features, such as the entrance to a connecting passage, the degradation factor rises above the threshold, and the system automatically resumes high-weight utilization of the laser odometry factor. After the entire process is completed, by comparing with several preset high-precision control points within the tunnel, the robot's autonomous navigation cumulative positioning error is controlled within the centimeter range, and the axial drift phenomenon common in long, straight sections is not observed. In contrast, when a fixed-weight multi-sensor fusion strategy was used in the same section, i.e., the laser odometry factor weight was not adjusted according to the degradation factor for synchronous testing, the cumulative positioning error reached more than ten centimeters or even more, and there was a significant systematic offset along the tunnel axis. This verified the effectiveness of the degradation perception adaptive weight mechanism in the degradation section of a long straight tunnel.

[0107] In terms of defect identification, the multimodal defect identification and 3D back-projection measurement module 4 detected several defects, including circumferential cracks in the lining, water seepage stains, and localized spalling. For one circumferential crack, the system obtained the quantified values ​​of its average and maximum widths through 3D back-projection measurement, and output a defect level suggestion based on the crack's direction and length. For a suspected water seepage area, the infrared thermal imager 23 detected that the temperature in this area was about 2.3 degrees Celsius lower than the surrounding background and that it exhibited a continuous seepage pattern. After multimodal confidence fusion, it was determined to be a genuine water seepage defect, avoiding false detections caused by oil stains reflecting off the tunnel wall. The mileage marker geocoding module 5 accurately calibrated the above defects to the design mileage markers and automatically matched them to the corresponding segment ring numbers. The detection reports can be directly imported into the existing maintenance log system, eliminating the need for maintenance personnel to manually verify the mileage, significantly improving the efficiency of issuing defect rectification work orders.

[0108] After the inspection was completed, the system automatically generated a list of defects, with defect type, actual physical size, fusion confidence level and mileage marker as the main fields, and attached visible light images, point cloud screenshots and infrared temperature images of the corresponding locations. This makes it easy for maintenance personnel to complete defect verification and priority ranking without returning to the site. For a small number of samples with fusion confidence levels in the middle range during this inspection, the system automatically marked them as key re-inspection targets for the next inspection according to preset rules, realizing dynamic rolling updates of the defect ledger.

[0109] Example 2: Specialized Detection of Water Leakage in Mining Tunnels

[0110] This embodiment uses a horseshoe-shaped tunnel constructed using the mining method in a city rail transit system as an application scenario. This section has historically been prone to water leakage. Before detection, the detection height and pitch angle of the multi-sensor sensing module 2 are adjusted using a lifting platform to adapt to the detection posture of the horseshoe-shaped tunnel arch foot and sidewall area. In this embodiment, the focus is on verifying the effect of the infrared temperature-assisted verification mechanism on suppressing false detections of water leakage: In the same section, several suspected water leakage areas are marked during the visible light detection stage. After comparing the temperature distribution of each area with the infrared thermal imager 23, a considerable proportion of the suspected areas are judged as surface stains or reflective interference because the temperature distribution is not significantly different from the surrounding background and does not have a continuous seepage pattern. These areas are then eliminated. The remaining areas are confirmed as real water leakage problems through multimodal confidence fusion. Compared with the identification method using only visible light images, this embodiment, by introducing infrared temperature-assisted verification and multimodal confidence fusion, reduces the false alarm rate of water leakage problems from about 30% to less than 10%, significantly reducing the false alarm rate of the detection results. For confirmed water leakage issues, the three-dimensional back-projection measurement module 4 further calculates the projected area of ​​the leakage zone on the lining structure surface and estimates the level of leakage activity based on infrared temperature differences, providing a quantitative reference for maintenance personnel to formulate grouting and sealing plans. Meanwhile, the degradation-aware adaptive navigation and positioning module 3 fully utilizes the rich geometric constraints of the laser point cloud in sections with significant changes in arch foot and sidewall curvature, maintaining high pose estimation accuracy and verifying the adaptability of the proposed method to tunnels with different cross-sectional shapes.

[0111] Example 3: Collaborative Detection of High-Speed ​​Connecting Lines and Turnout Sections

[0112] This embodiment uses a steep, connecting section of two operational lines, including a turnout section, as its application scenario. This section features frequent changes in curvature, a steep gradient, and a cross-sectional shape that gradually transitions from circular to trumpet-shaped, placing higher demands on navigation, positioning, and mileage conversion. During the detection process, the degradation-aware adaptive navigation and positioning module 3 continuously obtains high degradation factors based on the rich geometric features of the turnout section, maintaining a high weight for the laser odometer factor. Combined with closed-loop detection and global pose graph optimization, the robot's pose estimation remains stable in complex linear sections. The mileage geocoding module 5 pre-inputs the horizontal and vertical curve parameters from the line design file for this section, fully considering the influence of the curve shape when projecting the three-dimensional coordinates of the defect onto the design mileage. Compared to the traditional method of simple conversion based on straight-line mileage, the defect mileage calibration error is significantly reduced in this embodiment. Furthermore, in this embodiment, the edge computing unit 6 maintained a stable real-time detection frame rate even with the detection speed increased to 0.5 meters per second, verifying the real-time performance of the lightweight detection network on a platform with limited computing resources. For several difficult samples that the edge computing unit 6 determined to have low confidence, the cloud-based collaborative analysis platform 7 completed a second review and used the review results for incremental training of the edge detection network. In subsequent inspections, the recognition confidence of similar samples was improved, verifying the effectiveness of the edge-cloud collaborative closed-loop optimization mechanism. This embodiment also shows that the dynamic weighting mechanism adopted by the degradation perception adaptive navigation and positioning module 3 can smoothly switch the confidence weight of the dominant sensor between long straight degradation sections and geometrically rich turnout sections in the same inspection task, without needing to pre-set fixed segmentation parameters for different sections. It has strong scene adaptability, further verifying the versatility and engineering scalability of the proposed solution under complex line conditions.

[0113] In actual deployment, the system provided by this invention can adaptively adjust the size of the mobile platform 1, the sensor selection of the multi-sensor perception module 2, and the deployment method of the cloud-based collaborative analysis platform 7 (such as local private deployment or centralized cloud deployment) according to the tunnel cross-section form, operating window duration, and interface specifications of the existing maintenance information system for different urban rail transit lines. This adjustment does not alter the core technical principles of the degradation perception adaptive navigation and positioning method, multimodal defect identification and three-dimensional back-projection measurement method, and mileage marker geocoding method proposed in this invention. Furthermore, the incremental training mechanism between the edge computing unit 6 and the cloud-based collaborative analysis platform 7 can be further extended to a multi-machine collaborative learning mode where multiple robots share the same cloud model. That is, after multiple robots collect problematic samples from different lines and sections, and the cloud-based collaborative analysis platform 7 uniformly reviews and trains them, the updated model parameters are distributed to all edge computing units 6 within the network. This accelerates the adaptation speed of the overall detection model to different tunnel scenarios, lighting conditions, and defect types, improving the overall efficiency of the system when deployed at the urban rail transit network level.

[0114] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0115] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An autonomous navigation and multimodal defect recognition system for an urban rail transit tunnel inspection robot, comprising a mobile platform, a multi-sensor sensing module mounted on the mobile platform, an edge computing unit, and a cloud-based collaborative analysis platform, wherein the multi-sensor sensing module includes a three-dimensional LiDAR, a binocular visible light camera, an infrared thermal imager, an inertial measurement unit, and a wheeled odometer encoder, and each sensor completes time synchronization calibration and external parameter calibration; characterized in that, The edge computing unit operates a degradation-aware adaptive navigation and positioning module, a multimodal defect identification and three-dimensional back-projection measurement module, and a mileage marker geocoding module, wherein: The degradation perception adaptive navigation and positioning module is used to calculate the point cloud degradation factor based on the ratio of the minimum eigenvalue to the maximum eigenvalue of the point cloud registration equation information matrix, and dynamically adjust the weights of the laser odometry factor, inertial measurement unit factor, wheel odometry factor and closed-loop detection factor in the factor graph optimization according to the degradation factor through the S-shaped function, and output the robot's real-time pose. The multimodal defect identification and three-dimensional back-projection measurement module is used to align visible light images, infrared thermal images and laser point clouds in time and space, detect tunnel defects using a lightweight target detection network, back-project the two-dimensional detection box onto the three-dimensional point cloud to obtain the true physical size of the defects, verify suspected water leakage defects by combining infrared temperature distribution, and perform weighted fusion of multimodal confidence scores. The mileage marker geocoding module is used to perform nearest point projection matching between the three-dimensional coordinates of the defect and the route design alignment database, and output the accurate mileage marker of the defect by combining arc length integral and interpolation correction. The multimodal defect identification and three-dimensional back-projection measurement module performs weighted fusion of visible light detection confidence, infrared verification confidence, and laser point cloud reflection intensity normalized confidence to output a fused confidence score.

2. The system according to claim 1, characterized in that, The degradation perception adaptive navigation and positioning module calculates the point cloud degradation factor by performing eigenvalue decomposition on the information matrix of the point cloud registration residual equation and taking the ratio of its minimum eigenvalue to its maximum eigenvalue as the degradation factor. The smaller the degradation factor, the weaker the geometric constraint of the current point cloud in the corresponding degree of freedom direction. The information matrix is ​​obtained by multiplying the Jacobian matrix of the point cloud registration residual with respect to the pose increment and its transpose.

3. The system according to claim 2, characterized in that, The weight coefficient of the laser odometry factor is obtained by mapping the degradation factor through an S-shaped function. It transitions smoothly and continuously near a preset degradation threshold. When the degradation factor is lower than the preset degradation threshold, the weight coefficient decreases. When the degradation factor is higher than the preset degradation threshold, the weight coefficient approaches the maximum value. The degradation perception adaptive navigation and positioning module is also used to calculate the eigenvalue ratio of the information matrix according to the rotation sub-block and the translation sub-block respectively, to obtain independent degradation factors in multiple degrees of freedom directions, and to set corresponding weight coefficients for each degree of freedom direction.

4. The system according to claim 3, characterized in that, The overall objective function of the factor graph optimization is composed of a weighted sum of the laser odometer factor, the inertial measurement unit pre-integration factor, the wheel odometer factor, and the closed-loop detection factor. The weight of the laser odometer factor is dynamically adjusted in accordance with the method described in claim 3, and the weights of the other factor terms are preset according to the measurement covariance of the corresponding sensors.

5. The system according to claim 1, characterized in that, In the multimodal defect identification and three-dimensional back projection measurement module, for crack-type defects, multiple locations are sampled at equal intervals along the crack centerline and its normal direction. The Euclidean distance between the corresponding points of the crack edge in the three-dimensional point cloud at each sampling location is calculated as the crack width at that location. The average and maximum values ​​of the crack widths at multiple sampling locations are taken, and the average and maximum crack widths are output.

6. The system according to claim 1, characterized in that, The infrared thermal imager is used to acquire images of the temperature distribution on the surface of the tunnel lining. For areas suspected of water leakage detected in the visible light image, the temperature distribution of the corresponding position in the infrared thermal image is extracted. When the temperature difference between the area and the surrounding background exceeds a preset threshold and shows a continuous seepage pattern, it is determined to be a water leakage defect; otherwise, the confidence level of the defect is reduced.

7. The system according to claim 1, characterized in that, The multimodal defect identification and three-dimensional back-projection measurement module converts the visible light detection confidence, infrared verification confidence, and laser point cloud reflection intensity normalized confidence into logarithmic probability form, then sums them by weight, and finally obtains the fused confidence through an S-shaped function mapping. The weight coefficients corresponding to each confidence are adaptively updated in a sliding window manner according to the accuracy of the corresponding modality in historical detections verified by the cloud collaborative analysis platform. The higher the recent recognition accuracy of a certain modality, the larger its corresponding weight coefficient.

8. The system according to claim 1, characterized in that, The mileage marker geocoding module pre-converts the horizontal and vertical curve parameters in the track design file into a three-dimensional parameter curve of the track centerline with the design mileage as the parameter. It then solves the nearest projection point of the three-dimensional coordinates of the defect on the parameter curve through numerical iteration. After interpolating and correcting the projected mileage by combining the robot pose and the accumulated mileage of the wheeled mileage encoder, it outputs the accurate mileage marker of the defect.

9. The system according to claim 1, characterized in that, The edge computing unit deploys a lightweight target detection network, the backbone of which is composed of stacked depthwise separable convolutional units and embedded with a channel-space dual attention module. The cloud-based collaborative analysis platform is used to perform super-resolution enhancement and secondary judgment by the verification network on disease samples that are determined to have low confidence by the edge computing unit, and the verification results are used for incremental training and model update of the lightweight target detection network.

10. An autonomous navigation and multimodal defect identification method for an urban rail transit tunnel inspection robot, applied to the system according to claim 1, characterized in that, Includes the following steps: Collect 3D point cloud data, visible light images, infrared thermal images, and robot motion data inside the tunnel; Motion distortion compensation and feature extraction are performed on the point cloud, and the laser odometry pose estimation is obtained by matching it with the local sub-map. The point cloud degradation factor is also calculated. The weights of the laser odometry factor are dynamically adjusted based on the point cloud degradation factor. The laser odometry factor, inertial measurement unit factor, wheeled odometry factor, and closed-loop detection factor are jointly optimized by the degradation perception adaptive navigation and positioning module into the factor graph, and the robot's real-time pose is output. Spatiotemporal alignment of visible light images, infrared thermal images, and laser point clouds is performed, and a lightweight target detection network deployed in the edge computing unit detects tunnel defects and outputs two-dimensional detection boxes. The two-dimensional detection frame is back-projected onto the three-dimensional point cloud to calculate the true physical size of the defect. The suspected water leakage defect is verified by combining the infrared temperature distribution, and the multimodal confidence is weighted and fused. The three-dimensional coordinates of the defects are matched with the nearest point projection database of the route design, and the accurate mileage station number of the defects is output and written into the tunnel structure defect database.