Bridge disease unmanned aerial vehicle airborne active inspection method and system

Abstract

The application belongs to the technical field of bridge structure health monitoring, and provides a bridge disease unmanned aerial vehicle (UAV) airborne active inspection method and system. Bridge surface images obtained by the UAV flying along an initial global inspection route are acquired. Disease detection confidence and image comprehensive quality scores are obtained by combining the bridge surface images and a lightweight detection model of a multi-scale adaptive Mamba architecture fused with bridge physical priori. When the image comprehensive quality score is lower than a preset qualified threshold, flight and shooting parameters are adaptively adjusted according to image degradation causes, target area images are re-acquired, and re-checking is repeated until the image quality meets the standard. When the image quality meets the standard, detection area levels are divided based on disease detection confidence double thresholds, and corresponding resource allocation schemes are adopted for different confidence area levels. After global inspection is completed, a bridge three-dimensional disease distribution map is constructed by fusing multi-source inspection data in the whole process. High-precision, high-efficiency and intelligent autonomous inspection of bridge diseases is realized.

Description

Technical Field

[0001] This invention belongs to the field of bridge structural health monitoring technology, and in particular relates to a method and system for unmanned aerial vehicle (UAV)-borne active inspection of bridge defects. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Currently, bridge defect detection mainly relies on manual inspection and drone-assisted inspection. Manual inspection is inefficient, dangerous, and subjective, making it difficult to adapt to the complex structures of large bridges. Existing drone inspection uses a pre-planned flight path for static data acquisition and ground station post-processing, which improves detection efficiency and safety, but still has core shortcomings and cannot meet the needs of intelligent and refined inspection.

[0004] Static flight cannot adjust attitude and angle based on real-time image quality. Hidden areas such as the interior of the box girder, the back of the arch ribs, and the supports are susceptible to problems like blurring, overexposure, and poor viewing angles due to lighting, occlusion, and structural curvature, directly reducing the accuracy of subsequent inspections. Furthermore, the separation of data acquisition and detection means that inspection results cannot be fed back into flight control in real time. Suspected defect areas cannot be promptly re-inspected, and the confidence level lacks quantitative basis, leading to missed or false positives and potentially requiring secondary flights, thus increasing costs. In addition, the current system uses a uniform sampling density across the entire area, wasting flight resources in defect-free areas and failing to guarantee detection accuracy in suspected defect areas, making it difficult to balance accuracy and efficiency. Summary of the Invention

[0005] In order to solve at least one of the technical problems existing in the background art, the present invention provides a method and system for unmanned aerial vehicle (UAV) active inspection of bridge defects, so as to realize high-precision, high-efficiency, intelligent autonomous inspection of bridge defects.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides an unmanned aerial vehicle (UAV)-based active inspection method for bridge defects, comprising the following steps: An initial global inspection route is generated based on the constructed 3D model of the bridge; Acquire bridge surface images obtained by the UAV flying along the initial global inspection route; A lightweight detection model based on a multi-scale adaptive Mamba architecture, incorporating bridge surface images and bridge physical priors, is used to obtain the defect detection confidence score and overall image quality score. The detection model employs a shared backbone dual-branch parallel inference architecture to perform end-to-end inference on the bridge surface images. During the inference process, multi-scale defect features of the bridge are extracted, and after three-level multi-constraint confidence calibration, the defect detection confidence score is output. The image features extracted by the shared backbone are reused, and the overall image quality score is obtained through multi-dimensional index weighted calculation. When the overall image quality score is lower than the preset qualified threshold, the flight and shooting parameters are adaptively adjusted according to the cause of image degradation, the target area image is re-acquired and the verification is repeated until the image quality meets the standard. When the image quality meets the standard, the detection area level is divided based on the dual threshold of the disease detection confidence. For different confidence area levels, the corresponding resource allocation scheme is adopted to determine whether there are low confidence areas. If so, the active revisit navigation is performed and Bayesian iterative confidence update is performed. After the overall inspection is completed, the multi-source inspection data from the entire process is integrated to construct a three-dimensional map of bridge defects and generate a standardized inspection report.

[0007] Furthermore, the generation of the initial global inspection route based on the constructed 3D bridge model specifically includes: A 3D model of the bridge is constructed, and structural semantic segmentation is performed on the 3D model to obtain different structural regions; For the disease-prone areas in different regions, a sampling strategy combining structural priors and curvature quantification is used to generate initial inspection waypoints; An improved A* algorithm with UAV dynamic constraints is used to connect all inspection waypoints in the differentiated planning to generate an initial global inspection route that is collision-free, has continuous curvature, and has smooth velocity.

[0008] Furthermore, the initial inspection waypoints are generated using a sampling strategy combining structural priors and curvature quantization for disease-prone areas in different regions, including: For disease-prone areas in different regions, the severity level of disease-prone areas is determined based on established criteria. For any local surface patch of a bridge 3D point cloud model, the curvature calculation result of the target point is obtained by fitting the coefficients of the first and second basic forms of the surface. By combining the curvature calculation results and the level of disease-prone areas, the actual waypoint spacing of the target area is determined by the waypoint spacing quantification formula, and the initial inspection waypoints are obtained.

[0009] Furthermore, the UAV dynamics constraints include hard constraints and soft constraints. Hard constraints are mandatory verification thresholds, and soft constraints are flight quality optimization targets. Hard constraints include attitude hard constraints, maneuverability hard constraints, flight speed and acceleration / deceleration hard constraints, and safety distance hard constraints.

[0010] Furthermore, during the inference process, multi-scale bridge defects features are extracted through a scenario-based adaptive state space scanning mechanism. The multi-scale bridge defect feature vector at the k-th position is represented as: , Where k is the position index of the image feature sequence. The step size parameter for discretizing the feature sequence. This is the reference factor for step scaling. It is a sigmoid activation function. For linear projection layers with step size parameters, Let k be the input feature vector at the k-th position. For step size bias term, For multi-scale adaptive scan step size coefficients, To optimize the state transition matrix, , , , The learnable parameter matrix of the state-space model. The Hadamard product operator represents element-wise matrix multiplication. It is a natural exponential function. This is the spatial scan weight matrix. To optimize the input bias vector, The hidden state vector of the feature sequence at position k. For the kth The hidden state vector of the feature sequence at one location. This is the output feature vector at the k-th position.

[0011] Furthermore, the output of the disease detection confidence score after three-level multi-constraint confidence calibration includes: By combining multi-scale disease characteristics and the bridge physical prior knowledge set, the original probability of disease presence in the target area and the original classification probability of disease category are obtained. Based on the original probability of disease presence in the target area and the original classification probability of disease category, the basic confidence level is calculated. The intermediate confidence level, after bridge physical prior calibration, is generated by combining the base confidence level and the prior matching coefficient. The final defect detection confidence level of the image to be detected is obtained by combining the intermediate confidence level after bridge physical prior calibration and the image quality qualification threshold.

[0012] Furthermore, the multi-dimensional indicators include image sharpness, illumination uniformity, occlusion degree, and viewing angle effectiveness.

[0013] Furthermore, the method of classifying detection areas based on dual thresholds of disease detection confidence levels, and adopting corresponding resource allocation schemes for different confidence level areas, includes: When the confidence level of disease detection is greater than or equal to the high threshold, it is considered a high-confidence area and is determined to be a confirmed disease area. Disease information, such as category, location, and severity, is marked, and the sampling density of this area and adjacent areas is reduced, while the sampling interval is expanded. When the confidence level of disease detection is greater than or equal to the low threshold but less than the high threshold, it is considered a medium confidence area and is identified as a suspected disease area. The normal sampling density is maintained, the number of shots is increased, and images are collected from different angles. When the confidence level of disease detection is less than the low threshold, it is considered a low-confidence area and is determined to be an area to be re-inspected. The current flight path is immediately interrupted, and the active revisit mechanism is activated to conduct a close-up, multi-angle re-inspection.

[0014] Furthermore, the active revisit navigation execution and Bayesian iterative confidence update include: For low-confidence areas, the detection results of multiple re-inspection images are obtained. For the detection results of multiple re-inspection images, an iterative fusion method based on Bayes' theorem is adopted. The disease confidence of the first inspection is used as the initial prior probability. The detection results of each re-inspection image are used as independent observation evidence. Combined with the true positive rate and false positive rate of the pre-calibrated detection model, the posterior probability of the disease existence is calibrated round by round to complete the iterative update of the disease detection confidence.

[0015] A second aspect of the present invention provides an unmanned aerial vehicle (UAV)-based active inspection system for bridge defects, comprising: The 3D modeling and route planning module is used to generate an initial global inspection route based on the constructed 3D model of the bridge. The airborne image acquisition module is used to acquire images of the bridge surface obtained by the UAV flying along the initial global inspection route; The lightweight bridge defect detection module combines bridge surface images with a lightweight detection model based on a multi-scale adaptive Mamba architecture that integrates bridge physical priors to obtain defect detection confidence and overall image quality score. The detection model employs a shared backbone dual-branch parallel inference architecture to perform end-to-end inference on the bridge surface images. During inference, multi-scale defect features are extracted, and after three levels of multi-constraint confidence calibration, the defect detection confidence is output. The image features extracted from the shared backbone are reused, and the overall image quality score is calculated through multi-dimensional weighted indexes. The image quality assessment module is used to adaptively adjust flight and shooting parameters according to the causes of image degradation when the overall image quality score is lower than the preset qualified threshold, re-acquire images of the target area and repeat the verification until the image quality meets the standard. The active revisit planning module is used to classify the detection area level based on the dual threshold of the disease detection confidence when the image quality meets the standard. It adopts the corresponding resource configuration scheme for different confidence area levels, determines whether there are low confidence areas, and if so, performs active revisit navigation and Bayesian iterative confidence update. The report generation module is used to integrate multi-source inspection data from the entire process after the global inspection is completed, construct a three-dimensional map of bridge defects, and generate a standardized inspection report.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves an integrated, two-way closed-loop process for the entire inspection workflow. It improves the deployment of the Mamba detection module on an airborne platform, breaking through the traditional separation of data acquisition and detection, and enabling real-time synchronous processing of image acquisition and disease detection. Relying on bidirectional control of image quality and detection confidence, detection results are fed back to flight control in real time, forming a complete inspection closed loop. This shortens the inspection cycle, reduces repetitive flights, and comprehensively improves the intelligence and efficiency of inspection operations.

[0017] This invention improves imaging quality and detection reliability in complex areas. Based on feature sharing, it constructs a multi-index image quality evaluation system without requiring additional computing power. This system can determine image defect types and adaptively adjust flight attitude and shooting parameters, effectively solving problems such as abnormal lighting in concealed locations on bridges, blurred images, and poor shooting angles. It optimizes input image quality from the source, reducing missed and false detections of defects.

[0018] This invention achieves a refined allocation of inspection resources, balancing inspection accuracy and operational efficiency. It develops differentiated flight path planning schemes based on bridge structural characteristics and curvature differences, while simultaneously implementing tiered regional control based on inspection confidence levels. Sampling density is reduced in high-reliability areas, and re-inspection routes are proactively planned for suspected defect areas. Bayesian methods are used to update and correct inspection results, ensuring inspection accuracy while saving flight energy and reducing overall inspection costs.

[0019] Adapted to low-computing-power airborne environments, this invention balances detection accuracy and real-time performance. It employs a scenario-based improved Mamba architecture, relying on adaptive feature extraction and multi-level confidence calibration to enhance the identification of subtle defects, overcoming the shortcomings of traditional networks in terms of insufficient detection accuracy and inaccurate confidence levels. Coupled with a dedicated lightweight knowledge distillation solution, it compresses the model size while maintaining detection performance, meeting the real-time inference requirements of UAVs and stably identifying various typical bridge defects.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of an unmanned aerial vehicle (UAV) active inspection method for bridge defects provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the improved Mamba architecture disease detection module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the detection confidence grading and active revisit mechanism provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the active revisit mechanism and spiral re-inspection route provided in the embodiments of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] Example 1 like Figure 1 As shown in the figure, this embodiment provides a method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects, including the following steps: Step 1: Generate an initial global inspection route based on the constructed 3D model of the bridge; Specifically, the steps include the following: Step 101: Construct a 3D model of the bridge and perform structural semantic segmentation on the 3D model to obtain different structural regions; In this embodiment, a 3D point cloud model of the bridge scanned by an airborne lidar is imported to construct a high-precision 3D bridge model; the 3D model is then subjected to structural semantic segmentation, dividing it into different structural regions such as the bridge deck, the interior of the box girder, the arch ribs, the supports, and the piers. Step 102: For disease-prone areas in different regions, an initial inspection waypoint is generated using a sampling strategy that combines structural priors and curvature quantification. Specifically, the steps include the following: Step 1021: For disease-prone areas in different regions, determine the severity level of the disease-prone areas based on the established criteria; In this embodiment, disease-prone areas are determined using two levels: First, based on the semantic segmentation results of the bridge's 3D model, and combined with the high-risk structures such as bearings, arch feet, box girder end joints, and pier top anchorage zones as defined in the "Highway Bridge Technical Condition Assessment Standard," these are directly classified as Level 1 disease-prone areas. Second, for other structural areas, disease susceptibility is quantified through calculation of the principal curvature of the 3D surface. The maximum principal curvature of the bridge structural surface is positively correlated with the local stress concentration coefficient, and their mechanical quantitative relationship is as follows: , in, This represents the local stress concentration factor of the structure. The maximum principal curvature of the surface, The radius of the local chamfer of the structure, flat area When approaching infinity Approaching level 1, with no obvious stress concentration, areas with greater curvature are identified as secondary locations prone to disease.

[0027] Step 1022: For any local surface patch of the bridge's 3D point cloud model, solve for the orthogonal principal curvature by fitting the coefficients of the first and second basic forms of the surface; For arbitrary local surface patches of a bridge's 3D point cloud model, in order to quantify the local geometric bending characteristics of the bridge structure, match the requirements for graded judgment of defect-prone areas and differentiated route planning, this implementation method first solves for the orthogonal principal curvature by fitting the coefficients of the first and second basic forms of the surface based on the principle of surface differential geometry. The specific implementation steps are as follows: The first fundamental form of a surface is defined as the basic form describing the intrinsic metric of the tangent plane of the surface, and its expression is: , in, , In the local coordinate system axis, The square of the differential increment in the axial direction for axis, The product of the axial differential increments, all three are core differential units describing the intrinsic metric of the tangent plane of a surface; The coefficients of the first fundamental form of the surface are calculated using the first-order partial derivatives of the locally parameterized surface. The specific calculation formula is as follows: , in, For locally parameterized surfaces along First-order partial derivatives in the axial direction, For locally parameterized surfaces along The first-order partial derivative in the axial direction.

[0028] The second fundamental form of a surface is defined as the core form describing the local bending properties of a surface, and its expression is: , in: , , The meaning is consistent with that in the first basic form; The coefficients of the second fundamental form of the surface are calculated using the second-order partial derivatives and the unit normal vector of the locally parameterized surface. The specific implementation process is as follows: First, calculate the second-order partial derivatives of the locally parameterized surface: , Then calculate the target point. Unit normal vector of the surface at the location : , Substitute into the formula to calculate the coefficients of the second fundamental form: , Based on the first and second basic form coefficients obtained above, the target point is calculated. Gaussian curvature at the point With mean curvature The specific calculation formula is as follows: , in, For Gaussian curvature, it corresponds to the product of two orthogonal principal curvatures; The mean curvature is the average of the two orthogonal principal curvatures.

[0029] Based on Gaussian curvature With mean curvature Construct the principal curvature characteristic equation: , Solve the quadratic characteristic equation to obtain the target point. The principal curvatures corresponding to the two orthogonal directions are calculated using the following formula: , in, , The principal curvatures are two mutually orthogonal directions, and are the target points respectively. The maximum and minimum values ​​of the curvature at the point of application; The maximum absolute value of the two principal curvatures is taken as the maximum principal curvature of the region, that is: .

[0030] Step 1023: Combining the curvature calculation results and the level of disease-prone areas, determine the actual waypoint spacing of the target area through the waypoint spacing quantification formula, and further plan to obtain the initial waypoints; In this embodiment, based on the curvature calculation results and the disease susceptibility level, the actual waypoint spacing of the target area is determined using the waypoint spacing quantification formula: , in, This represents the actual waypoint spacing in the target area. The baseline waypoint spacing for the flat area of ​​the bridge deck. This is the curvature weighting coefficient. This is the weighting coefficient for the susceptibility level. This refers to the susceptibility level coefficient, with Level 1 susceptibility areas. Secondary common sites Flat, risk-free area .

[0031] This enables differentiated initial waypoint planning, which reduces waypoint spacing and increases sampling density in areas with high curvature and high susceptibility, and increases waypoint spacing and reduces sampling density in flat areas with low curvature.

[0032] Step 103: Using the improved A* algorithm with UAV dynamic constraints, connect all the inspection waypoints of the differentiated planning to generate an initial global inspection route that is collision-free, has continuous curvature, and has smooth speed, ensuring coverage of the entire area to be inspected.

[0033] To address the shortcomings of the original A* algorithm, which only pursues the shortest path and does not match the dynamic characteristics of the UAV, UAV dynamic constraints are embedded into the core components of the algorithm.

[0034] First, define the hard and soft constraints of UAV dynamics. The hard constraints are mandatory verification thresholds, and the soft constraints are flight quality optimization objectives.

[0035] The hard constraint formula for attitude is: , in, For pitch angle, For roll angle, For the maximum permissible posture, The yaw rate, This is the maximum permissible yaw rate.

[0036] The hard constraints on maneuverability are: , in, The baseline cruising speed for inspection. This is the minimum turning radius.

[0037] The formulas for flight speed and acceleration / deceleration hard constraints are as follows: , in, These are the minimum stable speed and the maximum safe flight speed, respectively. This represents the maximum permissible acceleration / deceleration.

[0038] The formula for the hard constraint of safety distance is: , in, The shortest distance between the route node and the bridge structure. The minimum safe distance is preset.

[0039] The core requirements of soft constraints are continuous curvature of the flight path, stable flight speed, and reasonable deviation between the heading and the camera's principal optical axis. To ensure that the flight path conforms to the dynamic characteristics of the UAV, the node expansion rules are reconstructed first, and child nodes are expanded only within the reach of the hard dynamic constraints, filtering out unreachable nodes and nodes with collision risks, thus ensuring path feasibility from the source. Based on this, the algorithm cost function is reconstructed to incorporate the soft-constraint optimization objective, and the specific formula is as follows: , in, The total cost of the current node. This represents the actual path length from the starting point to the current node. This is a soft constraint penalty item. , , To predefine non-negative weight coefficients and satisfy normalization constraints + + =1, The heuristic function uses the Dubins shortest path distance: , in, The target waypoint of the path, This represents the minimum turning radius of the drone. This is a Dubins path calculation function that outputs the minimum turning radius. Under constraints, the distance from the current node n to the target waypoint The shortest path length.

[0040] After path optimization is completed, cubic B-spline curves are used to smoothly fit the path nodes, ensuring that the curvature, speed, and acceleration of the flight path are continuously controllable. The fitting formula is as follows: , in, To smooth the flight path curve, For path control points, The basis functions are cubic B-spline functions. During the fitting process, the dynamic hard constraints must be enforced simultaneously: , in, For the curvature of the curve, The first and second derivatives of the curves are used to generate a global inspection route that satisfies both hard constraint safety requirements and soft constraint stability requirements, and is adapted to the dynamic characteristics of the UAV.

[0041] Step 2: Acquire bridge surface images obtained by the UAV flying along the initial global inspection route, and perform preprocessing; In this embodiment, the UAV is equipped with a high-definition industrial camera, a three-axis stabilized gimbal and an onboard computing unit, and flies along the initial global inspection route generated in step 1. During the flight, the three-axis stabilized gimbal adjusts the shooting posture in real time according to a preset angle, synchronously collects high-definition images of the bridge surface, and transmits the collected images to the onboard computing unit in real time.

[0042] The airborne computing unit performs grayscale conversion, noise reduction, and normalization preprocessing on the input image, providing standardized input for subsequent disease detection and image quality assessment. This enables simultaneous execution of image acquisition and airborne preprocessing, breaking the traditional separation mode of first acquiring and then processing during inspection.

[0043] Step 3: Combine the bridge surface image with the constructed lightweight detection model based on a multi-scale adaptive Mamba architecture that integrates bridge physical priors to obtain the defect detection confidence and overall image quality score; The detection model adopts a shared backbone dual-branch parallel inference architecture to complete end-to-end inference on the bridge surface image. During the inference process, multi-scale bridge defect features are extracted, and after three-level multi-constraint confidence calibration, the defect detection confidence is output. The image features extracted by the shared backbone are reused, and the comprehensive image quality score is obtained by multi-dimensional index weighted calculation. To address the core shortcomings of existing general-purpose UAV models for bridge defect inspection, such as susceptibility to falsely high or low detection confidence levels, high computational overhead, and inability to adapt to real-time airborne processing, this embodiment proposes a multi-scale adaptive Mamba architecture that integrates prior knowledge of bridge structure and defect physics. Figure 2 As shown, relying on three core designs—shared backbone dual-branch parallel inference, three-level multi-constraint confidence calibration, and scenario-based adaptive state space scanning—the accuracy of defect detection confidence calculation and the reliability of image comprehensive quality scoring are effectively improved, ultimately realizing airborne integrated end-to-end processing of bridge image acquisition, defect detection, and quality assessment.

[0044] To address the issues of high onboard computing power overhead and asynchronous inference caused by independent branches and repetitive feature extraction in traditional detection models, this architecture adopts end-to-end transmission and parallel inference with a shared backbone dual branches, standardizes module logic and data flow, enables feature reuse, and adapts to the requirements of low-computing-power and low-latency onboard deployment.

[0045] The overall end-to-end architecture is defined as follows: , in, This is a function compounding operator that represents the serial cascading relationship of modules and defines the end-to-end processing flow of data from right to left. The proposed end-to-end global mapping function for the multi-scale adaptive Mamba architecture; This is a mapping function for the detection head, used to output the bridge defect detection results and calibration confidence level; This is a multi-scale feature fusion mapping function used to output features that fuse global and local information; An adaptive selective state-space scan mapping function is used to optimize the feature extraction of minor bridge defects. A mapping function for extracting shared backbone features is used to output multi-scale shared depth features of bridge images; This is an image preprocessing mapping function used to perform standardization processing on the input image.

[0046] The shared backbone dual-branch parallel inference expansion is as follows: , in, The input is the original image of the bridge surface; To share multi-scale feature vectors, which are the basic features used by the detection branch and the quality assessment branch; This is a mapping function for the Mamba shared backbone network, used to extract multi-scale depth features from images; The original probability of the presence of disease in the target area. These are the original classification probabilities of the disease category, and both serve as the basic inputs for subsequent confidence level calculations. This is a branch mapping function for disease detection, used to output disease detection results and confidence levels; This is a set of prior knowledge about bridge physics, used to assist in optimizing the calibration of test confidence levels. is the quality feature vector output by the image quality assessment branch; is the image quality assessment branch mapping function used to evaluate the overall quality of the input image.

[0047] To address the shortcomings of the original Mamba architecture, such as fixed scan step size and easy loss of bridge image spatial structure and subtle disease features, a multi-scale adaptive scan step size coefficient is introduced to improve the original selective state space model. This not only solves the core defects of the original Mamba architecture but also improves the extraction accuracy of subtle disease features and the inference efficiency of disease-free areas.

[0048] The expression is: , Where k is the position index of the image feature sequence. The step size parameter for discretizing the feature sequence. This is the reference factor for step scaling. It is a sigmoid activation function. For linear projection layers with step size parameters, Let k be the input feature vector at the k-th position. For step size bias term, For multi-scale adaptive scan step size coefficients, To optimize the state transition matrix, , , , The learnable parameter matrix of the state-space model. The Hadamard product operator represents element-wise matrix multiplication. It is a natural exponential function. This is the spatial scan weight matrix. To optimize the input bias vector, The hidden state vector of the feature sequence at position k. For the kth The hidden state vector of the feature sequence at one location. This is the final output feature vector for the k-th feature sequence position.

[0049] Furthermore, the formula for calculating the adaptive scan step size coefficient is as follows: , , In the formula, The scanning step size reference coefficient is a positive real number, constrained within a preset positive value range adapted to the bridge inspection scenario. It serves as a fixed reference value for step size adjustment and does not change dynamically with the input image features. It is only used to unify the step size adjustment scale. is the texture weight coefficient, a real number in the interval [0,1], obtained by adaptive learning during model training. It is used to balance the adjustment weight of image texture features and bridge structure curvature on the scanning step size, and adapt to bridge images with different texture complexities. Let be the local image texture entropy, a non-negative real number whose value is constrained to [0, ..., ... The grayscale distribution of local regions in an image is statistically analyzed and used to quantify the texture complexity of the image region corresponding to the input features. is the maximum value of global image texture entropy, is a positive real number, and is the statistical upper limit of the texture entropy of the entire region of the input bridge image, which is used to normalize the local texture entropy and ensure the consistency of texture feature quantization. Let be the maximum principal curvature of the local structure, and be a non-negative real number whose value is constrained to [0, ]. The curve is calculated based on the geometric features of the bridge structure and is used to quantitatively characterize the curvature of the bridge structure region corresponding to the input features. is the maximum principal curvature of the global structure, is a positive real number, and is the statistical upper limit of the principal curvature of the entire target bridge structure, used to normalize the curvature of the local structure to ensure the consistency of curvature feature quantification; This is the lower limit coefficient for the scan step size. It is a positive real number and is constrained to a preset positive value range. It is used to limit the minimum value of the adaptive scan step size to avoid the model inference efficiency being reduced due to an excessively small step size. Let be the upper limit coefficient of the scan step size, be a positive real number, and satisfy . Constrained within a preset positive value range, this limit is used to set the maximum value of the adaptive scanning step size, preventing the loss of subtle bridge defects due to excessively large step sizes.

[0050] The quantization formula for the spatial scan weight matrix is ​​expressed as follows: , , In the formula, The probability of image region occlusion is a real number in the interval [0,1], which is adaptively calculated by the image occlusion detection module. The closer the value is to 1, the lower the degree of occlusion of the corresponding image region and the higher the proportion of effective features. The closer the value is to 0, the higher the degree of occlusion and the lower the proportion of effective features. The disease risk weight coefficient is a real number in the interval [0,1]. It is preset and adaptively fine-tuned by the model according to the bridge disease inspection requirements. It is used to adjust the weight distribution between the bridge disease-prone area and the ordinary structure area, and to strengthen the feature extraction of the disease-prone area. , where is the structural disease susceptibility coefficient, a non-negative real number, and its value ranges from [0, ]. The method is derived by combining bridge structural design parameters and historical defect data, and is used to characterize the defect susceptibility of the bridge structural area corresponding to the input features. The maximum value of the disease susceptibility level is a positive real number. It is the statistical upper limit of the disease susceptibility level of the entire target bridge structure. It is used to normalize the local disease susceptibility level coefficient to ensure the consistency of disease risk characteristic quantification.

[0051] Based on feature extraction, the three-level multi-constraint confidence calibration mechanism reconstructs the confidence generation process: design basic generation - physical prior calibration - image quality constraints. This solves the problem that the confidence of existing models only rely on visual features and the detection data quality is low, and significantly improves the calculation accuracy of detection confidence.

[0052] The basic confidence score is calculated based on the original probability of the presence of the disease in the target area and the original classification probability of the disease category. The expression for generating the basic confidence score is as follows: , , in, , where is the basic confidence level for disease detection, is the initial input value for the three-level calibration, and is a real number in the interval [0,1]. The original probability of the presence of disease in the target area. This represents the original classification probability of the disease category.

[0053] The intermediate confidence level after physical prior calibration of the bridge is generated by combining the base confidence level and the prior matching coefficient. The physical prior calibration expression is as follows: , in, is the intermediate confidence level after bridge physical prior calibration, and is a real number in the interval [0,1]. Here, is the prior matching coefficient, a real number in the interval [0,1], used to calibrate the confidence level of the foundation inspection by combining prior knowledge of bridge physics. Its calculation formula is: , , in, The weighting coefficient for the structure type is a real number in the interval [0,1], which is preset by the bridge inspection scenario. The fault susceptibility coefficient for the bridge structure in the detection area is a non-negative real number, with a value range of [0, 1]. ]; Let be the maximum value of the overall structural disease susceptibility level of the target bridge, be a positive real number, and be the normalized baseline value of the disease susceptibility level coefficient. The historical disease matching coefficient is a real number in the interval [0,1]. The higher the historical disease detection rate, the higher the recurrence frequency. The closer the value is to 1, the better for regular structural areas with no historical disease records. Use the preset baseline value.

[0054] The final defect detection confidence level of the image to be detected is obtained by combining the intermediate confidence level after bridge physical prior calibration and the image quality qualification threshold, and is expressed as: , in, The final disease detection confidence level is a real number in the interval [0,1]. The image quality score for the image to be detected is a non-negative real number in the interval [0,1]. Let be the image quality pass threshold, be a positive real number in the interval [0,1], and be the criterion for determining whether the image quality meets the requirements for disease detection. The calculation formula is: , , In the formula, The threshold reference value is a positive real number in the interval [0,1], which is pre-determined by the bridge inspection scenario calibration experiment. The structural type weight coefficient is a real number in the interval [0,1], preset by the bridge inspection scenario. This is the disease type weighting coefficient, a real number in the interval [0,1]. Higher values ​​are used for minor diseases, and lower values ​​are used for major diseases. , These are the upper and lower threshold coefficients, respectively, both being positive real numbers in the interval [0,1], and satisfying the following conditions: .

[0055] The image quality assessment module is based on the image quality features output above. The final score is calculated synchronously, eliminating the need for repeated image feature extraction. Simultaneously, a multi-dimensional weighted comprehensive quality scoring function is constructed, taking into account image sharpness, illumination uniformity, occlusion level, and viewpoint effectiveness. Its expression is as follows: , in, The overall image quality score ranges from [0,1]. For clarity indicators, As an indicator of illumination uniformity, For occlusion rate, As a measure of perspective effectiveness, , , , Let each indicator have a weight, and let the sum of the weights equal to 1.

[0056] The formula for calculating the sharpness index is as follows: , , Where I is the single-channel grayscale image of the RGB bridge image to be detected after conversion, with dimensions H×W and pixel values ​​in the range [0, 255] integers. The edge response map is obtained by convolving the grayscale image I with the 3×3 Laplacian operator. Let L(I) be the pixel value variance of the edge response map. To calibrate the maximum value of the Laplacian variance of clear images in bridge inspection scenarios, a preset positive real number is used, and Clip() is a numerical truncation function used for constraint. In the interval [0,1].

[0057] Formula for calculating illumination uniformity index:

[0058] Where I is a single-channel grayscale image of the image to be detected, with pixel values ​​in the range [0, 255] (integer). Let be the global pixel mean of the grayscale image I. Let I be the global pixel standard deviation of the grayscale image I. The default value is a very small positive real number, used to avoid denominators of 0. `Clip()` is a numeric truncation function used to constrain the value. In the interval [0,1].

[0059] The formula for calculating the occlusion rate index is as follows: , in, This represents the total number of pixels in high-frequency occlusion areas of the image. The total number of pixels in the image to be detected. This represents the percentage of areas with high-frequency occlusion.

[0060] Formula for calculating the effectiveness index of perspective: , , Where V is the unit vector of the UAV camera's shooting direction, and N is the unit normal vector of the target bridge structure surface. The angle between V and N is expressed in radians and ranges from [0, 0]. Clip() is a numeric truncation function used for constraint. In the interval [0,1].

[0061] To adapt to the computing power limitations of UAV-borne edge computing units, this architecture designs a two-stage bridge knowledge distillation model, setting up a teacher model and a student model to complete knowledge transfer. The teacher model is a large-scale pre-trained bridge defect detection model, which, after pre-training on massive amounts of bridge defect image data, possesses high-precision bridge defect feature extraction and confidence calibration capabilities, serving as the knowledge transfer provider. The student model is the multi-scale adaptive Mamba model of this architecture, specifically designed for the low computing power requirements of the airborne end, serving as the knowledge transfer receiver. Through this distillation model, the defect feature extraction and confidence calibration capabilities of the teacher model can be completely transferred.

[0062] The total loss function of the distillation process is: , in, The total loss during model training. To detect the basic loss, , , Let be the weighting coefficient, satisfying , , , The knowledge distillation loss is specifically: Characteristic distillation loss Calculation formula: , In the formula, This is an intermediate feature map of the teacher model. For the feature map corresponding to the student model , , These represent the feature map height, width, and number of channels, respectively. The loss is aligned with the feature distribution of the teacher-student model using mean squared error.

[0063] State-space model distillation loss Calculation formula: , In the formula, The hidden state vector is used for the Mamba module of the teacher model. The hidden state vector corresponding to the student model; , These represent the vector length and dimension, respectively. This loss is used to align the temporal modeling capabilities of the teacher-student model.

[0064] Confidence level distillation loss Calculation formula , In the formula, To determine the confidence level for detecting defects in the teacher model. The confidence level corresponds to the student model; This represents the number of detection boxes. This loss is used to align the confidence calibration logic of the teacher-student model.

[0065] Step 4: When the overall image quality score is lower than the preset qualified threshold, the flight and shooting parameters are adaptively adjusted according to the causes of image degradation, the target area image is re-acquired and the verification is repeated until the image quality meets the standard. Compare the output image overall quality score (IQE) with the preset image quality pass rate. By comparing the images, when the IQE is below the threshold, the system automatically identifies the causes of image degradation and executes the corresponding adaptive control strategy: if the image clarity is insufficient, the system controls the drone to reduce its flight speed and extend the camera exposure time; if the image lighting is abnormal, the system adaptively adjusts the camera exposure parameters and ISO sensitivity; if there is a deviation in the shooting angle, the system corrects the drone's pose and the gimbal's shooting angle in real time.

[0066] After the adjustment is completed, the target area is immediately re-acquired, the comprehensive quality score of the newly acquired image is recalculated and verified again until the score meets the standard, forming a closed-loop control of "evaluation-adjustment-reacquirement-verification", which ensures the quality of input images from the data source and reduces the risk of missed or false detection of diseases.

[0067] Step 5: When the image quality meets the standard, the detection area is divided into levels based on the dual threshold of the disease detection confidence, and corresponding resource allocation schemes are adopted for different confidence level areas. like Figure 3 As shown, in this embodiment, a dual confidence threshold is set: a high threshold and a low threshold. 0.85 is preferred; low threshold The optimal value is 0.50. The detection area is divided into three levels to achieve graded response: ① High confidence region, Score ≥ Once a diseased area is identified, disease information, including type, location, and severity, is marked. The sampling density in this area and adjacent areas is reduced, and the sampling interval is increased to four times the original, saving flight resources and time. ② Medium confidence region, ≤Score< For areas identified as suspected disease areas, maintain the usual sampling density, increase the number of shots, and collect images from different angles to assist in subsequent disease confirmation and improve detection reliability; ③ Low confidence region, Score < If an area is identified as requiring re-inspection, the current flight path will be immediately interrupted, and an active revisit mechanism will be initiated to conduct a close-range, multi-angle re-inspection to improve the confidence level of the detection.

[0068] Step 6: Determine if there is a low-confidence region. If so, actively revisit the voyage and perform Bayesian iterative confidence updates. To address the issues of missing features from single-view observations and inaccurate fusion of multi-frame results, a local coordinate system adapted to the bridge structure is established with the center point of the area as the origin for the low-confidence region to be re-inspected. An improved A* algorithm that incorporates UAV dynamic constraints is used to generate a spiral re-inspection route that enables multi-angle, unobstructed observation. The UAV flies close to the target area along the re-inspection route to acquire images from multiple perspectives and at multiple frequencies.

[0069] For the detection results of multiple re-inspection images, an iterative fusion method based on Bayes' theorem is adopted. The disease confidence score of the first inspection is used as the initial prior probability. The detection results of each re-inspection image are used as independent observation evidence. Combined with the true positive rate and false positive rate of the pre-calibrated detection model, the posterior probability of disease existence is calibrated round by round to complete the iterative update of the disease detection confidence score.

[0070] Meanwhile, based on the updated confidence level, a hierarchical and progressive closed-loop re-inspection mechanism is constructed: if the updated confidence level reaches a high threshold, it is determined to be a real disease, the information is recorded, and the system returns to the initial global inspection route; if the confidence level is still below a low threshold, the detection range is expanded, the sampling interval is reduced, and the re-inspection process is repeated; if the confidence level still does not meet the standard after reaching the preset maximum number of re-inspections, the area is marked as an uncertain area, and a manual review prompt is output.

[0071] like Figure 4 As shown, the confidence calibration accuracy is improved specifically through structurally coupled flight paths and Bayesian iteration, with the center point of the low-confidence region as the reference point. Establish a local coordinate system for the origin. The axes are along the bridge's length, width, and vertical height, respectively. An improved A* algorithm incorporating UAV dynamic constraints is used to generate a spiral re-inspection route. This spiral route enables unobstructed observation from multiple angles, providing sufficient feature support for confidence calibration. The constraints are as follows: , The objective function is: , in, This is the lower limit of the camera's pitch angle, typically set to 15°. This is the upper limit of the camera's pitch angle, typically set to 60°. This is the lower limit of the relative flight altitude of the drone, and is generally taken as 1 to 2 meters. This represents the current relative flight altitude of the drone. This is the upper limit for the relative flight altitude of drones, typically set at 5 to 8 meters. , , The local three-dimensional coordinates of the i-th waypoint form the basic nodes of the spiral route; , , Let be the local three-dimensional coordinates of the (i+1)th waypoint.

[0072] To address the shortcomings of existing methods that use simple averaging to fuse multi-frame detection results, which fail to fully consider the inherent performance characteristics of the detection model and are difficult to effectively eliminate systematic errors, this paper proposes an iterative fusion method for multi-frame detection results based on Bayes' theorem. This method uses the disease confidence score obtained from the initial inspection as the prior probability, and the detection results of each re-inspected image as independent observation evidence, combined with the pre-calibrated true positive rate of the detection model. With false positive rate The posterior probability of disease existence is calibrated round by round, and its formal update process is as follows: ① Prior initialization: using the confidence level of the defects obtained from the first inspection. As the initial prior probability of the existence of the disease, that is: , in, This indicates that "there are real diseases in the area." This refers to the event that "the area is free of diseases".

[0073] ② Single-frame evidence likelihood calculation: For the first frame Detection results of frame re-examination images Define the likelihood parameter, which represents the probability of the detection result occurring under conditions of disease presence and disease absence: , , ③ Bayesian Iterative Update: Before Fusion The formula for calculating the posterior probability of disease existence after frame observation evidence is as follows: , ④ Confidence output: Complete no less than After inference and analysis of the frame-by-frame re-examination images, the final disease detection confidence score is output, i.e.: , To achieve a response to the disease confidence level after Bayesian iterative updates and to address the issues of rigid response logic and difficulty in balancing inspection accuracy and efficiency in traditional re-inspection mechanisms, a high confidence threshold is preset. Low confidence threshold Maximum number of re-inspections Detection range expansion coefficient Sampling interval scaling factor Construct a hierarchical, progressive, closed-loop re-inspection mechanism: ①When When the system determines that there are real diseases in the area, it records the relevant disease information and then controls the drone to return to the initial global inspection route. ②When At that time, the testing scope was expanded to the original scope. Double, reduce the sampling interval to the original interval Double the number of tests, and repeat the re-inspection process; ③When passing through After the second re-examination, the confidence level of the disease Still below If the condition is not met, the area will be marked as an uncertain area, and a prompt for manual review will be output.

[0074] This mechanism achieves dynamic control of the re-inspection process through confidence-level response, ultimately realizing the synergistic optimization of bridge defect inspection accuracy and efficiency.

[0075] Step 6: After completing the overall inspection, integrate the multi-source inspection data from the entire process to construct a three-dimensional defect distribution map of the bridge, and generate a standardized inspection report containing high-confidence defect information, a three-dimensional distribution map, and maintenance recommendations.

[0076] The bridge area that has been inspected is divided into grids, and the sampling density of subsequent inspections is dynamically adjusted based on the final inspection confidence level of each grid cell. Specifically, the area that has been tested is divided into grids, preferably 1m×1m, with each grid cell corresponding to a test confidence value; for high confidence areas, the sampling interval between adjacent grid cells is increased to 4 times the original; for medium confidence areas, the normal sampling density is maintained and multi-angle shooting is added; for low confidence areas, an active revisit mechanism is used, and the sampling interval is reduced to within 0.5m.

[0077] It should be noted that the above embodiment is only one embodiment, and the specific data can be set according to the actual situation; After the overall inspection is completed, the multi-source data from the overall inspection and multiple rounds of re-inspection are integrated, the calibrated defect information is combined, a three-dimensional defect distribution map of the bridge is constructed, the defect level is classified according to the "Highway Bridge Technical Condition Assessment Standard", and finally a standardized inspection report containing defect statistics, a three-dimensional defect distribution map and maintenance treatment suggestions is generated. The marked uncertain areas are specially marked to provide data support for bridge maintenance decisions.

[0078] Example 2 This embodiment provides a bridge defect unmanned aerial vehicle (UAV) onboard active inspection system, including: The 3D modeling and route planning module is used to generate an initial global inspection route based on the constructed 3D model of the bridge. The airborne image acquisition module is used to acquire images of the bridge surface obtained by the UAV flying along the initial global inspection route; The lightweight bridge defect detection module combines bridge surface images with a lightweight detection model based on a multi-scale adaptive Mamba architecture that integrates bridge physical priors to obtain defect detection confidence and overall image quality score. The detection model employs a shared backbone dual-branch parallel inference architecture to perform end-to-end inference on the bridge surface images. During inference, multi-scale defect features are extracted, and after three levels of multi-constraint confidence calibration, the defect detection confidence is output. The image features extracted from the shared backbone are reused, and the overall image quality score is calculated through multi-dimensional weighted indexes. The image quality assessment module is used to adaptively adjust flight and shooting parameters according to the causes of image degradation when the overall image quality score is lower than the preset qualified threshold, re-acquire images of the target area and repeat the verification until the image quality meets the standard. The active revisit planning module is used to classify the detection area level based on the dual threshold of the disease detection confidence when the image quality meets the standard. It adopts the corresponding resource configuration scheme for different confidence area levels, determines whether there are low confidence areas, and if so, performs active revisit navigation and Bayesian iterative confidence update. The report generation module is used to integrate multi-source inspection data from the entire process after the global inspection is completed, construct a three-dimensional map of bridge defects, and generate a standardized inspection report.

[0079] It should be noted that the specific implementation of the bridge defect UAV-borne active inspection system in this embodiment of the invention is similar to the specific implementation of the bridge defect UAV-borne active inspection method in this embodiment of the invention. Please refer to the description in the method section for details. To reduce redundancy, it will not be repeated here.

[0080] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects.

[0081] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the above-described method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects.

[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0083] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0086] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects, characterized in that, Includes the following steps: An initial global inspection route is generated based on the constructed 3D model of the bridge; Acquire bridge surface images obtained by the UAV flying along the initial global inspection route; A lightweight detection model based on a multi-scale adaptive Mamba architecture, incorporating bridge surface images and bridge physical priors, is used to obtain the defect detection confidence score and overall image quality score. The detection model employs a shared backbone dual-branch parallel inference architecture to perform end-to-end inference on the bridge surface images. During the inference process, multi-scale defect features of the bridge are extracted, and after three-level multi-constraint confidence calibration, the defect detection confidence score is output. The image features extracted by the shared backbone are reused, and the overall image quality score is obtained through multi-dimensional index weighted calculation. When the overall image quality score is lower than the preset qualified threshold, the flight and shooting parameters are adaptively adjusted according to the cause of image degradation, the target area image is re-acquired and the verification is repeated until the image quality meets the standard. When the image quality meets the standard, the detection area level is divided based on the dual threshold of the disease detection confidence. For different confidence area levels, the corresponding resource allocation scheme is adopted to determine whether there are low confidence areas. If so, the active revisit navigation is performed and Bayesian iterative confidence update is performed. After the overall inspection is completed, the multi-source inspection data from the entire process is integrated to construct a three-dimensional map of bridge defects and generate a standardized inspection report.

2. The method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects as described in claim 1, characterized in that, The generation of the initial global inspection route based on the constructed 3D model of the bridge specifically includes: A 3D model of the bridge is constructed, and structural semantic segmentation is performed on the 3D model to obtain different structural regions; For the disease-prone areas in different regions, a sampling strategy combining structural priors and curvature quantification is used to generate initial inspection waypoints; An improved A* algorithm with UAV dynamic constraints is used to connect all inspection waypoints in the differentiated planning to generate an initial global inspection route that is collision-free, has continuous curvature, and has smooth velocity.

3. The method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects as described in claim 2, characterized in that, The method for generating initial inspection waypoints, using a sampling strategy combining structural priors and curvature quantification, targets disease-prone areas in different regions. For disease-prone areas in different regions, the severity level of disease-prone areas is determined based on established criteria. For any local surface patch of a bridge 3D point cloud model, the curvature calculation result of the target point is obtained by fitting the coefficients of the first and second basic forms of the surface. By combining the curvature calculation results and the level of disease-prone areas, the actual waypoint spacing of the target area is determined by the waypoint spacing quantification formula, and the initial inspection waypoints are obtained.

4. The method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects as described in claim 2, characterized in that, The UAV dynamics constraints include hard constraints and soft constraints. Hard constraints are mandatory verification thresholds, and soft constraints are flight quality optimization targets. Hard constraints include attitude hard constraints, maneuverability hard constraints, flight speed and acceleration / deceleration hard constraints, and safety distance hard constraints.

5. The method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects as described in claim 1, characterized in that, During the inference process, multi-scale bridge defects are extracted using a scenario-based adaptive state-space scanning mechanism. The multi-scale bridge defect feature vector at the k-th position is represented as follows: ， Where k is the position index of the image feature sequence. The step size parameter for discretizing the feature sequence. This is the reference factor for step scaling. It is a sigmoid activation function. For linear projection layers with step size parameters, Let k be the input feature vector at the k-th position. For step size bias term, For multi-scale adaptive scan step size coefficients, To optimize the state transition matrix, , , , The learnable parameter matrix of the state-space model. The Hadamard product operator represents element-wise matrix multiplication. It is a natural exponential function. This is the spatial scan weight matrix. To optimize the input bias vector, The hidden state vector of the feature sequence at position k. For the kth The hidden state vector of the feature sequence at one location. This is the output feature vector at the k-th position.

6. The method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects as described in claim 1, characterized in that, The three-level multi-constraint confidence level calibration outputs the disease detection confidence level, including: By combining multi-scale disease characteristics and the bridge physical prior knowledge set, the original probability of disease presence in the target area and the original classification probability of disease category are obtained. Based on the original probability of disease presence in the target area and the original classification probability of disease category, the basic confidence level is calculated. The intermediate confidence level, after bridge physical prior calibration, is generated by combining the base confidence level and the prior matching coefficient. The final defect detection confidence level of the image to be detected is obtained by combining the intermediate confidence level after bridge physical prior calibration and the image quality qualification threshold.

7. The method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects as described in claim 1, characterized in that, The multi-dimensional indicators include image sharpness, illumination uniformity, occlusion level, and viewpoint effectiveness.

8. The method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects as described in claim 1, characterized in that, The method of classifying detection areas based on dual thresholds of disease detection confidence levels, and adopting corresponding resource allocation schemes for different confidence levels, includes: When the confidence level of disease detection is greater than or equal to the high threshold, it is considered a high-confidence area and is determined to be a confirmed disease area. Disease information, such as category, location, and severity, is marked, and the sampling density of this area and adjacent areas is reduced, while the sampling interval is expanded. When the confidence level of disease detection is greater than or equal to the low threshold but less than the high threshold, it is considered a medium confidence area and is identified as a suspected disease area. The normal sampling density is maintained, the number of shots is increased, and images are collected from different angles. When the confidence level of disease detection is less than the low threshold, it is considered a low-confidence area and is determined to be an area to be re-inspected. The current flight path is immediately interrupted, and the active revisit mechanism is activated to conduct a close-up, multi-angle re-inspection.

9. A method for unmanned aerial vehicle (UAV)-based active inspection of bridge defects as described in claim 8, characterized in that, Active revisit navigation execution and Bayesian iterative confidence update include: For low-confidence areas, the detection results of multiple re-inspection images are obtained. For the detection results of multiple re-inspection images, an iterative fusion method based on Bayes' theorem is adopted. The disease confidence of the first inspection is used as the initial prior probability. The detection results of each re-inspection image are used as independent observation evidence. Combined with the true positive rate and false positive rate of the pre-calibrated detection model, the posterior probability of the disease existence is calibrated round by round to complete the iterative update of the disease detection confidence.

10. A bridge defect unmanned aerial vehicle (UAV) onboard active inspection system, characterized in that, include: The 3D modeling and route planning module is used to generate an initial global inspection route based on the constructed 3D model of the bridge. The airborne image acquisition module is used to acquire images of the bridge surface obtained by the UAV flying along the initial global inspection route; The lightweight bridge defect detection module combines bridge surface images with a lightweight detection model based on a multi-scale adaptive Mamba architecture that integrates bridge physical priors to obtain defect detection confidence and overall image quality score. The detection model employs a shared backbone dual-branch parallel inference architecture to perform end-to-end inference on the bridge surface images. During inference, multi-scale defect features are extracted, and after three levels of multi-constraint confidence calibration, the defect detection confidence is output. The image features extracted from the shared backbone are reused, and the overall image quality score is calculated through multi-dimensional weighted indexes. The image quality assessment module is used to adaptively adjust flight and shooting parameters according to the causes of image degradation when the overall image quality score is lower than the preset qualified threshold, re-acquire images of the target area and repeat the verification until the image quality meets the standard. The active revisit planning module is used to classify the detection area level based on the dual threshold of the disease detection confidence when the image quality meets the standard. It adopts the corresponding resource configuration scheme for different confidence area levels, determines whether there are low confidence areas, and if so, performs active revisit navigation and Bayesian iterative confidence update. The report generation module is used to integrate multi-source inspection data from the entire process after the global inspection is completed, construct a three-dimensional map of bridge defects, and generate a standardized inspection report.

Images