Unmanned aerial vehicle self-adaptive inspection method based on multi-layer characterization learning

Abstract

The invention discloses an unmanned aerial vehicle self-adaptive inspection method based on multi-layer characterization learning, which comprises two stages of automatic coarse inspection and self-adaptive fine inspection, and learns the space structure, component characteristics and disease characteristics of a bridge layer by layer. In the automatic rough detection stage, in combination with a spatial clustering post-processing method and a point cloud semantic segmentation network, instance attributes of bridge components are quickly learned. Then, the point cloud space is compressed and converted by using a simulation view field model and a dimension reduction technology, and the unmanned aerial vehicle is guided to carry out efficient and comprehensive space patrol in plane geometry. In the adaptive fine detection stage, a space-channel two-dimensional optimized hybrid convolution structure Light-PVIT is utilized to learn and extract disease features in the automatic coarse detection stage as prior information. The information guides the unmanned aerial vehicle to carry out fine detection on a disease area in a small field of view. According to the unmanned aerial vehicle self-adaptive inspection method, the bridge inspection efficiency and precision are remarkably improved, and reliable technical support is provided for bridge maintenance work.

Description

Technical Field

[0001] The present invention belongs to the field of structural health detection in civil engineering, and in particular relates to an unmanned aerial vehicle adaptive inspection method based on multi-layer representation learning. Background Art

[0002] As a key component of the transportation network, bridges play an important role in ensuring public safety and smooth economic operation. However, affected by various factors such as the natural environment, vehicle loads and engineering construction, bridge structures are prone to various diseases, such as cracks, spalling, and rust. These diseases may not only damage the structural integrity and bearing capacity of the bridge, but also cause serious safety accidents. Therefore, regular bridge inspection is essential to ensure the stability and safety of the bridge structure. Bridge inspection covers multiple aspects such as the bridge deck system, superstructure and substructure. Especially for the superstructure of the bridge, most of its components are located in a high-altitude environment. Traditional inspection methods mainly rely on technicians to carry out inspections on vehicles or other high-altitude work platforms. However, these traditional methods have many limitations: first, they usually require expensive equipment and professionals, resulting in high costs and long time; second, due to the limitations of the inspection platform, it is difficult to achieve a thorough inspection of the entire range of the bridge, which may lead to the omission of important defects or hidden dangers; in addition, the inspection results are often affected by the experience and subjective judgment of the engineering personnel, lacking objectivity and accuracy. Therefore, the development of efficient and objective new technologies for regular bridge inspections, especially in high-altitude environments, has become an urgent problem to be solved in the engineering community.

[0003] In order to meet this challenge, researchers are committed to exploring and developing advanced intelligent detection equipment, using advanced machine vision and sensor technology to improve the accuracy and efficiency of bridge field inspection operations. For the inspection of bridge bottom structures, mobile robot inspection systems, cameras of zipline structures and spherical drones are used to quickly detect the defects at the bottom of the bridge. In response to the inspection needs of high-altitude suspension structures, innovative climbing robot structures and tank-type robots with multiple types of sensors are used to collect detailed data on the surface defects of steel cables. In the face of the inspection difficulties of high piers and high tower facade structures of bridges, drone systems and ring-shaped climbing robots using new sensors and stereoscopic vision-inertial fusion are used to evaluate the cracks of high-rise bridge piers. The challenge of bridge data analysis lies in the massiveness and complexity of data, which cannot provide effective information for daily inspection and maintenance tasks. Manual processing methods are not only inefficient, but also often have differences in analysis results due to individual subjectivity. Most of the data is saved in the form of images, and traditional image analysis technology still has difficulty in getting rid of manual dependence. In the process of meeting this challenge, the academic community has gradually turned its attention to deep learning technology. The deep learning network structure can automatically learn and extract useful feature information from large-scale image data, realize automatic recognition of bridge data, and significantly improve data processing efficiency and accuracy. At present, the research and practice of defect detection mainly revolve around three core levels of tasks: classification, detection, and segmentation. The goal of the classification task is to assign the input image to predefined categories, each category representing a specific disease state. Compared with classification learning, the detection task is not only to determine the overall category of the image, but also to calibrate the location of the defect in the image by predicting the bounding box. The segmentation task involves classifying each pixel in the image into different categories, so as to accurately determine the outline and location of the defect.

[0004] In the current field of mobile detection platform development, the mainstream method still relies on manual operation. This not only limits the comprehensiveness and efficiency of detection, but also makes it difficult to focus on the diseased area. In addition, although high-precision visual analysis technology in data processing tasks can improve detection accuracy, its high demand for computing resources and strict requirements for hardware equipment also limit its widespread application.

[0005] Traditional methods of manually flying drones to inspect bridges have significant limitations and risk factors, mainly due to operational dependencies and limited field of view. These challenges highlight the need for more efficient and safe drone path planning methods.

[0006] In addition, traditional point cloud segmentation methods mainly include: attribute method, region segmentation, model fitting, etc. These methods usually rely on predefined rules or models and are difficult to adapt to complex or irregular point cloud data. For large-scale point cloud data, it takes a long time to process. In recent years, in order to improve the recognition effect of the network, many studies on deep learning technology have focused on designing complex local feature extraction structures and building larger models. This method often consumes a lot of resources when processing large scene models. Summary of the invention

[0007] In view of the above technical problems, the present invention proposes a new strategy for UAV adaptive inspection based on multi-layer representation learning. This strategy innovatively designs a two-stage inspection process: the first stage is an automatic rough inspection based on bridge space understanding. In this stage, the large field of view camera carried by the UAV is used to quickly obtain the geometric information of the bridge, and the three-dimensional model of the bridge is quickly constructed in combination with the motion recovery structure algorithm. Further in-depth analysis of the attribute information at the bridge component level, a point cloud instance segmentation framework based on an optimized clustering algorithm is proposed, which realizes the end-to-end comprehensive extraction of the attribute information of each bridge component. By simulating the field of view model and geometric space dimensionality reduction projection technology to compress and transform the point cloud space, an efficient spatial inspection path is automatically planned for each independent component of the bridge in plane geometry, ensuring that the UAV can conduct a comprehensive coverage inspection of the bridge components. The second stage is an adaptive detailed inspection based on disease information feedback. In order to improve the recognition speed and accuracy of diseases, the self-attention mechanism is reorganized from the two key dimensions of space and channel, and then a lightweight model named Light-PVIT is constructed. By combining the identified damaged areas with the established spatial transformation criteria, the inspection path is adaptively adjusted to guide the UAV to conduct detailed inspections of the damaged areas with a smaller field of view.

[0008] To achieve the above object, the technical solution of the present invention is as follows:

[0009] The UAV adaptive inspection method based on multi-layer representation learning is applied to the inspection of the bottom structure of the bridge, including the following steps:

[0010] S1, automatic rough inspection based on bridge spatial understanding;

[0011] S11. Use the large field of view camera carried by the drone to quickly obtain the geometric information of the bridge, and combine it with the motion recovery structure algorithm to quickly construct a three-dimensional model of the bridge;

[0012] S12. Analyze the attribute information at the bridge component level and propose a point cloud instance segmentation framework based on an optimized clustering algorithm to achieve end-to-end comprehensive extraction of attribute information of each bridge component;

[0013] S13. By simulating the field of view model and the geometric space dimensionality reduction projection method, the point cloud space is compressed and transformed, and efficient spatial inspection paths are automatically planned for each independent component of the bridge in the plane geometry, ensuring that the drone can conduct a comprehensive coverage inspection of the bridge components;

[0014] S2: Adaptive detailed inspection based on disease information feedback;

[0015] S21. Reorganize the self-attention mechanism from two key dimensions: space and channel, and build a lightweight model called Light-PVIT to identify the damaged areas of bridge components.

[0016] S22. Combining the damaged area of ​​the bridge component identified in step S21 with the geometric space dimensionality reduction projection method, the inspection path planned in step S13 is adaptively adjusted to guide the drone to perform detailed inspection of the damaged area.

[0017] Step S11 specifically includes the following sub-steps:

[0018] S111, using satellite maps and prior information of bridges to plan the path of the UAV;

[0019] S112. Collect multi-view 3D reconstruction data for the bridge, and then construct a 3D model of the entire scene.

[0020] Step S12 specifically includes the following sub-steps:

[0021] S121. Using PointNeXt network as a tool for learning information representation of bridge components, using the design concept of encoder and decoder to process point cloud data and finely segment the structure; in this process, the encoder captures high-level semantic features in the point cloud, and the decoder remaps these high-level semantic features to construct a finely segmented structure of the point cloud;

[0022] S122. Further intra-class division is performed based on semantic segmentation, and a strategy combining density-based spatial clustering post-processing method with semantic segmentation network is adopted until all points in the point cloud model are visited.

[0023] Step S13 specifically includes the following sub-steps:

[0024] S131. According to the simulated field of view model, a rough inspection route is planned for each component in the bridge point cloud space. The bridge component shape belongs to a regular spatial geometric body. Combining these characteristics, the point cloud space is compressed and transformed using the dimensionality reduction idea to describe the spatial relationship in the plane geometry.

[0025] S132, randomly extract three non-collinear points P from the point cloud data set 1,,P2,P3, constitute the initial point set as the basis for subsequent model fitting; calculate the normal vector of the initial point set Based on this, an initial spatial plane reference model is constructed; each remaining point in the component point cloud data is traversed to perform distance measurement operations, and any point whose distance to the plane model is lower than the tolerance threshold is identified as an inner point t;

[0026] S133, repeat the above iterative process, select the model with the largest consensus set as the representative space plane, compress and project the remaining point cloud (X, Y, Z) in the bridge component space onto the representative space plane (X pj ,Y pj ,Z pj ); construct an inspection plane parallel to the representative spatial plane according to the distance coefficient D, reducing the spatial dimension that needs to be considered when designing the UAV flight path; further, transform and project the component point cloud on the representative spatial plane onto the inspection plane, lock the camera azimuth, that is, the camera acquisition plane is parallel to the plane of the bridge to be inspected, and the two key factors of the inspection planning are unified on the inspection plane, converting the inspection plane in the three-dimensional space into a two-dimensional plane model;

[0027] S134. Based on the regularity of bridge components, the center line of the bridge component is used as the reference of the UAV inspection path on the two-dimensional inspection plane, and the center line of the bridge component is quickly extracted by the boundary fitting method; the boundary fitting direction follows the aspect ratio criterion, that is, the width of the bridge component and the height of the bridge component are compared numerically. When the width of the bridge component is greater than the height of the bridge component, the width direction of the bridge component is taken; when the height of the bridge component is greater than the width of the bridge component, the height direction of the bridge component is taken; a polynomial is used to fit the boundary;

[0028] S135. Plan camera acquisition points on the center line of the bridge component in combination with the field of view model, and consider different combinations of distance coefficients D and focal lengths f to ensure that the coverage range is greater than the boundary spacing; in the rough inspection stage, increase the distance coefficient D, reduce the focal length f, increase the acquisition field of view, and then reduce the overlap rate δ to cover the bridge component in a manner with an optimal number of acquisition points;

[0029] S136. Restore the collection point coordinates on the two-dimensional plane to three-dimensional coordinates through reverse mapping to obtain a complete inspection path.

[0030] Step S21 specifically includes the following sub-steps:

[0031] S211, the Light-PVIT lightweight model consists of six stages of learning tasks. The first five stages of learning tasks consist of a convolution module CVB, a deep separation convolution module (RDB2, RDB1) with an inverse residual structure, and a partial visual transformer module PVIT. The deep separation convolution module with an inverse residual structure is used as an efficient replacement for the convolution layer.

[0032] The first step in each learning stage is to downsample the features using an inverse deep separation convolution module without residual structure.

[0033] The partial visual transformer module PVIT consists of three submodules, namely the local representation submodule, the global representation submodule and the feature fusion submodule. A depth separation convolution with a convolution kernel size of 3×3 and a mapping convolution with a convolution kernel size of 1×1 are used to model the local information of the features and output the features. in,

[0034] H, W, C s Respectively represent the height, width and number of channels of the output feature map;

[0035] The global representation submodule introduces the transformer's self-attention mechanism to capture global context information on the feature map, so as to obtain the global information of the H×W feature map;

[0036] S213, introduce a partial split operation in the channel dimension. Specifically, the spatial compression feature X is split by the split coefficient n. s Divided into global modeling features and local modeling features To manage the redundancy of high-resolution features in the early stages of the model, the split factor n in stage 3 is stage-3 Set to 4; the split coefficient n of stage 4 stage-4 and the splitting factor n in stage five stage-5 Set them to 2 to avoid information loss caused by excessive compression;

[0037] S214, the global modeling feature X st Feeding the transformer module captures the long-range dependencies in the input features, and the local modeling features X sc The output of the transformer module will be combined with the global modeling feature X st Merge in the channel direction, restore the feature shape, and output the hybrid modeling feature

[0038] The feature fusion submodule is used for the mixed modeling feature Integrate global and local information;

[0039] S215, the final stage of the model uses a convolution submodule, a pooling flattening submodule, and a fully connected submodule to convert the features of stage six into the final prediction results.

[0040] Step S22 specifically includes the following sub-steps:

[0041] S221, using a disease detection network to automatically screen the data collected in the rough inspection stage for diseases, and at the same time independently summarize the image data containing diseases to form a disease set that is different from the image data of normal areas;

[0042] Before the disease screening stage, the order of collection points based on the rough inspection route Assign a label x to each collected image i ,After the screening stage, the real spatial location of the disease collection point can be ,quickly indexed based on the disease image label;

[0043] An adaptive inspection method based on rough inspection information feedback is proposed to enable the UAV to inspect efficiently while keeping attention on the diseased area. First, the collection points of the UAV are divided into normal and abnormal categories based on the feedback information from the disease screening stage.

[0044] At normal collection points, i.e., areas where bridge components are not damaged, the drone uses the rough inspection path as the inspection benchmark;

[0045] At abnormal collection points, i.e., areas where bridge components are damaged, the drone adaptively adjusts the detection path based on the disease feedback information to obtain more accurate detailed information;

[0046] Reduce the size of the drone's field of view and record the local details of the disease; the field of view can be reduced by reducing the distance coefficient D or increasing the focal length f. Keep the camera focal length unchanged and construct a distance target plane D in the inspection space. s The solution of fine detection plane realizes the fine acquisition of field of view (W s ,H s ) can be adjusted flexibly, W s ,H s Indicates the height and width of the acquisition field of view;

[0047] S222, further, projecting the diseased area as the optimization target onto the newly constructed fine detection plane, transforming the detection plane in the three-dimensional space into a two-dimensional plane, and performing grid-cutting on the diseased area corresponding to a single collection point, wherein the width of each grid is high S w ,S h Respectively represent the division coefficients in width and height. The center coordinates of these sub-grids represent the acquisition points on the detailed inspection path;

[0048] S223, traverse all diseased areas using the same operation as step S222, and formulate a detailed collection plan;

[0049] The overlapping area measurement criterion is used to filter and delete the collection points that exceed the limit threshold δ;

[0050] A heuristic search algorithm is used to connect discrete acquisition points and construct a detailed inspection path by minimizing the distance cost between neighboring points;

[0051] S224. Finally, the spatial conversion relationship between the collection point and the inspection target point is used, that is, the collection point is projected onto the bridge surface through the distance coefficient D to locate the defective area in real space.

[0052] The beneficial effects of the present invention are:

[0053] 1. The present invention uses satellite maps and prior information of bridges to plan the path of the drone. This method not only significantly improves the efficiency of drone flight, ensuring that all necessary inspection areas are covered in the shortest time, but also greatly improves the safety of the entire flight process.

[0054] 2. The present invention uses the PointNeXt network as a tool for learning information representation of bridge components, which achieves an excellent balance between model accuracy and size. Under a certain category attribute, a bridge may contain multiple components, and the differences between these components cannot be identified and detected by drones. In this process, a strategy combining density-based spatial clustering post-processing methods with semantic segmentation networks is adopted. Through this method, instance-level division can be achieved while maintaining processing simplicity.

[0055] 3. Convolutional neural networks (CNNs) have long been the mainstream choice for image processing and computer vision tasks. They are good at capturing local features of processed images, but the lack of understanding of global context restricts the performance of CNN network structures. The Transformer architecture that has emerged in recent years has shown excellent performance in processing global information. However, they usually require more computing resources, which becomes impractical on mobile devices with limited resources. Therefore, the present invention combines the efficiency of CNN with the excellent performance of Transformer, and proposes Light-PVIT as an efficient disease detection tool with the Mobile-ViT structure as the network baseline.

[0056] 4. The present invention has constructed an efficient interactive closed-loop system by deeply integrating the inspection activities of drones with data analysis. The system uses a multi-layer representation learning method to comprehensively learn and analyze the spatial structure, component properties and disease characteristics of the bridge, ensuring the optimization of the entire chain from data collection to processing. Through this strategy, the large amount of data collected during the inspection process is not only used for real-time analysis, but the conclusions drawn from the analysis can be quickly fed back to the inspection task, thereby achieving the continuous accumulation and self-improvement of bridge inspection experience. In addition, the present invention further explores the lightweight problem of internal analysis methods. Through comprehensive lightweight processing from three-dimensional space to two-dimensional planes, the computational burden and hardware requirements of the algorithm are significantly reduced, making the inspection system more efficient and practical.

[0057] In summary, the drone inspection strategy proposed in this invention is not only innovative in theory, but also can effectively improve the efficiency and accuracy of bridge inspection in practical applications. Through the combination of multi-layer representation learning and lightweight analysis methods, the method of this invention provides a new technical path for the field of bridge inspection, which helps to promote the further development and application of related technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the overall framework of the present invention;

[0059] Figure 2 The bridge component semantic learning model diagram proposed by the present invention;

[0060] Figure 3 The bridge component instance segmentation diagram based on spatial distribution characteristics proposed by the present invention;

[0061] Figure 4 The camera projection geometry diagram mentioned in the present invention;

[0062] Figure 5 Schematic diagram of the path planning method based on spatial dimensionality reduction proposed in the present invention;

[0063] Figure 6 The structure diagram of the lightweight disease identification model proposed in the present invention;

[0064] Figure 7 Detailed diagram of the spatial channel dual-dimensional optimization module proposed in the present invention;

[0065] Figure 8 Figure 1 shows the adaptive inspection method based on disease information feedback proposed by the present invention;

[0066] Fig. 9 The structural diagram of the bridge mentioned in the example;

[0067] Fig.10 The implementation process of the rough inspection of drones mentioned in the example;

[0068] Fig.11 The detection result diagram of the rough inspection stage mentioned in the example;

[0069] Fig.12 Schematic diagram of the two-stage results of the adaptive coarse-fine inspection of the bridge arch member proposed in the example.

[0070] Fig.13 Schematic diagram of the two-stage results of the adaptive coarse-fine inspection of the bridge beam component proposed in the example.

[0071] Fig.14 Schematic diagram of the two-stage results of the adaptive coarse-fine inspection of the bridge cross brace member proposed in the example. DETAILED DESCRIPTION

[0072] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, the modifications of various equivalent forms of the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0073] A UAV adaptive inspection method based on multi-layer representation learning. The overall flow chart of the solution is as follows: Figure 1 As shown, it includes the following components:

[0074] S1: Automatic rough inspection based on spatial understanding of bridges. In this stage, the large field of view camera carried by the drone is used to quickly obtain the geometric information of the bridge and quickly build a three-dimensional model of the bridge. Further in-depth analysis of the attribute information at the bridge component level automatically plans an efficient spatial inspection path for each independent component of the bridge in plane geometry, ensuring that the drone can conduct a comprehensive coverage inspection of the bridge components.

[0075] S2: Adaptive detailed inspection based on disease information feedback. Combining the identified damaged area with the established spatial transformation criteria, the inspection path is adaptively adjusted to guide the drone to conduct detailed inspection of the damaged area with a smaller field of view.

[0076] Furthermore, the automatic rough inspection based on bridge spatial understanding described in step S1 includes the following steps:

[0077] Step S11: Reconstructing the three-dimensional model of the bridge: quickly obtaining the overall geometric information of the bridge;

[0078] Step S12: Point cloud instance segmentation framework based on optimized clustering algorithm: deeply analyzes the attribute information at the bridge component level and realizes end-to-end comprehensive extraction of attribute information of each bridge component.

[0079] Step S13: Geometric space dimensionality reduction projection technology based on the simulated field of view model: realizing automatic planning of efficient spatial inspection paths for each independent component of the bridge in plane geometry.

[0080] Furthermore, S11: the specific method of reconstructing the three-dimensional model of the bridge is as follows:

[0081] Satellite maps and prior information of bridges are used to plan the UAV path. First, satellite maps provide the UAV with macroscopic geographic information, which enables the UAV to identify key features of the bridge and its surrounding environment before flight, including the precise spatial location of the bridge, adjacent buildings, and other possible flight interference factors. Based on this information, a preliminary flight path can be designed to avoid potential danger areas (such as obstacles such as high-voltage towers). This method not only significantly improves the efficiency of UAV flight and ensures that all necessary inspection areas are covered in the shortest time, but also greatly improves the safety of the entire flight process. Comprehensive multi-view 3D reconstruction data collection of the bridge is performed according to the planned UAV route. Feature extraction is performed on each collected image, usually using feature point algorithms such as SIFT and SURF. Then, the correspondence between images is established by matching these feature points between each pair of images. Finally, these correspondences are used to calculate the position of each feature point in the 3D coordinate system, and then a 3D model of the entire scene is constructed.

[0082] Furthermore, S12: The specific method of the point cloud instance segmentation framework based on the optimized clustering algorithm is as follows:

[0083] PointNeXt network is used as a tool to learn the information representation of bridge components, which achieves an excellent balance between model accuracy and size. Inspired by the U-Net architecture, the design concept of encoder and decoder is adopted to process point cloud data and achieve fine structure segmentation. In this process, the encoder is responsible for capturing high-level semantic features in the point cloud, while the decoder is responsible for remapping these features to construct a detailed segmentation structure of the point cloud, such as Figure 2As shown in the figure. N in [N,x] represents the number of point clouds output by each layer, and x represents the feature dimension of the point cloud. The encoder consists of a multi-layer perceptron layer (MLP), a point set abstraction (SA) layer, and an inverse residual multi-layer perceptron layer (IRM). The point set abstraction module is the core of the encoder and contains three key components: the farthest point sampling layer, the spherical grouping layer, and the feature extraction layer. The sampling layer uses the farthest point algorithm to select the center points of 1 / 4 of the number of point clouds, and optimizes the topological structure of the center points to ensure that the sampling covers the entire point cloud. The spherical grouping layer creates a spherical neighborhood with a radius of R around each center point, thereby sampling K feature points in each local area. The feature extraction layer extracts the features of the local area through the maximum pooling operation. Although this operation may cause partial loss of feature information, by using MLP to expand the feature dimension before pooling, these features can be enriched and compensated to reduce the impact of information loss. The inverse residual multi-layer perceptron layer is a feature enhancement structure built on the point set abstraction layer. It effectively alleviates the problem of gradient disappearance by introducing residual connections between input and output. Compared with the original structure, this design cancels the sampling layer and adds two MLP layers to enhance the feature extraction capability. At the same time, the grouping layer now uses a spherical neighborhood with a larger radius R, which expands the receptive field of the model, so that a more robust local feature relationship can be established. Stacking MLP layers can effectively enhance the feature extraction capability of the model, but it will also significantly increase the amount of calculation. In order to optimize this, the model uses a separable MLP structure to gradually perform feature extraction tasks. The MLP layer between the grouping layer and the pooling layer focuses on extracting features between each neighborhood, while the subsequent MLP layer is dedicated to extracting single-point features. This layered and refined processing strategy not only enhances the feature extraction capability of the model, but also maintains computational efficiency.

[0084] The decoder adopts a strategy based on distance interpolation and skip connections, using several feature propagation modules (FP) to gradually restore and map the compressed features. Each FP module mainly consists of a distance interpolation layer and a multi-layer perceptron layer. In order to compensate for the loss of detail information during the downsampling process, the decoder fuses the downsampled features of each stage during the feature transfer process. These features are introduced into the corresponding FP modules and integrated through the MLP layer to produce a more refined feature representation.

[0085] Further intra-class division is performed on the basis of semantic segmentation. The strategy of combining density-based spatial clustering post-processing method with semantic segmentation network is adopted, such as Figure 3As shown in the figure. First, the point cloud region condensation processing is introduced. The specific method is to construct a three-dimensional grid in the point cloud space according to the set three-dimensional pixel size S_p, and each pixel appears in the form of a cube. The point cloud data is mapped to the created grid space, and the geometric centroid is selected as the representative point for the sub-point cloud space in each cube. Then a statistical filter is used as a supplementary step to filter out abnormal points by using the distance statistical characteristics of each point in the point cloud data and its neighboring point set. For each point in the point cloud, the nearest N_k points are selected as the neighboring point set. Then, for each point, the distance between it and its neighboring points is calculated and the global data, including the mean and standard deviation, are counted. Based on these statistical data, the global average distance plus the standard deviation of α times is used as the screening threshold T_k. If the average distance between a point and its neighbors is greater than this threshold, it is regarded as an abnormal point and removed from the point cloud. Finally, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm based on the local point density distribution characteristics is adopted. Determine the neighborhood space search range ε ​​of each point cloud element and the minimum number of point clouds n that each dense area needs to contain. Perform a neighborhood search on each point cloud based on the spatial search range ε. If its neighborhood contains at least the number of points n required for the dense area, the point is defined as the core point P_ci. All points in the point cloud space that meet the definition of core points are the key to cluster formation. All core points whose distances between adjacent core points are within the neighborhood range are connected and grouped into a cluster in an iterative manner. When the current clustering is completed, retrieve the next new unvisited point cloud and repeat the above steps until all points in the point cloud model are visited.

[0086] Furthermore, step S13: the specific method of the geometric space dimensionality reduction projection technology based on the simulated field of view model is as follows:

[0087] For different inspection tasks, it is crucial to choose the optimal camera field of view. The camera field of view, that is, the spatial range that the camera can cover at a specific distance and focal length, is determined by multiple factors, including the camera's sensor size, focal length, and acquisition distance. The camera's sensor size H s ,W s and focal length f, these parameters determine the width of the camera's field of view. The acquisition distance D directly affects the coverage area of ​​the field of view. Under the premise that these parameters are known, the actual acquisition field of view of the camera in real space can be simulated, such as Figure 4 The formula for calculating the field of view is as follows:

[0088]

[0089] In order to ensure that the UAV inspection path can completely cover the target area, it is necessary to ensure that the images between adjacent collection points have sufficient overlapping area W. o ×H o Appropriate overlap provides a buffer for environmental changes or minor errors during flight, thereby ensuring the quality and integrity of drone inspections. Therefore, it is necessary to reasonably set the overlap rate between adjacent collection points. It is the key to achieve efficient and comprehensive inspection.

[0090] According to the simulated field of view model, a rough inspection route is planned for each component in the bridge point cloud space. The shape of bridge components is basically a regular spatial geometric body. Combining these characteristics, the dimensionality reduction idea is used to compress and transform the point cloud space to accurately describe the spatial relationship in the plane geometry, such as Figure 5 shown.

[0091] For different bridge components, it is necessary to find a representative spatial plane with morphological characteristics. Here, the idea based on random sample consistency is adopted to accurately depict the main geometric morphological characteristics of the component. The initial step involves randomly extracting three non-collinear points P1, P2, and P3 from the point cloud data set to form an initial point set as the basis for subsequent model fitting. The normal vector of the initial point set is calculated using mathematical methods. Based on this, an initial spatial plane benchmark model is constructed. Each remaining point in the component point cloud data is traversed to perform distance measurement operations. Any point whose distance to the plane model is lower than the tolerance threshold is identified as an inlier and included in the Consensus Set. Repeat the above iterative process and select the model with the largest consensus set as the representative spatial plane. The remaining point cloud (X, Y, Z) in the component space is compressed and projected onto the representative spatial plane (X pj ,Y pj ,Z pj ), realize the transformation from point cloud space to spatial plane. The projection formula is as follows:

[0092]

[0093] Here, a, b, c, and d are the model coefficients of the representative space plane. This plane is also regarded as the inspection object. According to the distance coefficient D, an inspection plane parallel to the representative plane is constructed to reduce the spatial dimension that needs to be considered when designing the UAV flight path. Furthermore, according to the above formula, the component point cloud on the representative space plane is transformed and projected onto the inspection plane, and the camera azimuth is locked, that is, the camera acquisition plane is parallel to the inspection object plane. The two key factors of inspection planning are unified on the inspection plane, and the inspection plane in three-dimensional space is converted into a two-dimensional plane model.

[0094] Based on the regularity of bridge components, the center line of the bridge component is used as the reference of the UAV inspection path on the two-dimensional inspection plane. The center line of the bridge component is quickly extracted by the boundary fitting method. The boundary fitting direction needs to follow the aspect ratio criterion, that is, a polynomial is used to fit the boundary in the direction with a larger proportional coefficient. Then, the camera acquisition point planning is performed on the center line of the bridge component in combination with the field of view model, and the appropriate rough inspection acquisition field of view (W) is calculated by considering the combination of different distance coefficients D and focal lengths f. p ,H p ) Ensure that the coverage is greater than the boundary spacing. At the same time, an appropriate field of view overlap rate is an indispensable condition to ensure a comprehensive inspection of the outer surface of the component. The rough inspection stage focuses on efficiency, and the acquisition field of view is increased by combining a large distance coefficient D and a small focal length f. Then reduce the overlap rate δ to cover the component with the optimal number of acquisition points. Finally, the acquisition point coordinates on the two-dimensional plane are restored to three-dimensional coordinates through inverse mapping to obtain a complete inspection path.

[0095] Furthermore, the adaptive detailed inspection based on disease information feedback described in step S2 includes the following steps: step S21: Light-PVIT lightweight model: effectively capture and integrate local features and global context information; step S22: Fine inspection method based on disease information feedback: adaptively adjust the inspection path to guide the drone to perform detailed inspection of the damaged area with a smaller field of view.

[0096] Furthermore, the specific method of the lightweight model of step S21: Light-PVIT is as follows:

[0097] Combining the efficiency of CNN with the excellent performance of Transformer, Light-PVIT is proposed as an efficient disease detection tool with Mobile-ViT structure as the network baseline. Light-PVIT consists of six stages of learning tasks, using a multi-scale architecture to perform deep feature extraction on images, such as Figure 6 As shown in the figure. The learning tasks in the first five stages are mainly composed of convolutional modules (CVB), deep separation convolutional modules with reverse residual structures (RDB2, RDB1) and partial visual transformer modules (PVIT). In order to optimize subsequent learning tasks, deep separation convolutional modules with reverse residual structures are used as efficient replacements for convolutional layers. The addition of residual structures can effectively alleviate the gradient vanishing problem, thereby accelerating the convergence of the network.

[0098] In the deep separation convolution module of the residual structure, 1×1 mapping convolution is first used to reduce the feature dimension, then 3×3 deep separation convolution is used for efficient feature extraction, and finally 1×1 mapping convolution is used for feature dimension increase. The feature dimension adjustment operation of the reverse residual structure is the opposite. First, 1×1 mapping convolution is used to increase the feature dimension, and finally 1×1 mapping convolution is used to reduce the feature dimension. The first step in each learning stage is to downsample the features. The reverse deep separation convolution module without residual structure is used to downsample the features.

[0099] A more efficient visual transformer module PVIT was invented, which can capture and fuse local and global information of feature maps without generating too many parameters, such as Figure 7 As shown in Figure 2. This module mainly consists of three submodules: local representation submodule, global representation submodule, and feature fusion submodule. The local representation submodule is used for input A depth separation convolution with a convolution kernel size of 3×3 and a mapping convolution with a convolution kernel size of 1×1 are used to model the local information of the features and output the features of the effective receptive field of 3×3. The deep separation convolution with a convolution kernel size of 3×3 can lightly encode the local spatial information of the feature map through its characteristics, and then learn the linear combination information between the input feature channels through the mapping convolution with a convolution kernel size of 1×1. In order to obtain the global information of the effective receptive field of H×W feature map, the global representation submodule introduces the transformer's self-attention mechanism to capture the global context information of the feature map. The calculation process of the self-attention mechanism can be expressed by the following formula:

[0100]

[0101] Where ⊕ represents the Concat operation, denote the weight matrices of the Q, K, and V branches in the i-th linear layer, respectively. σ denotes the softmax function, and <·,·> denotes the dot product operation.

[0102] In order to solve the challenges of the traditional self-attention mechanism, a fast self-attention mechanism is reorganized from the two dimensions of space and channel. In the global representation submodule, the input feature map has established information dependencies between pixels through local operations, that is, each pixel has established connections with adjacent pixels. The idea of ​​hole convolution is introduced into the transformer module to hollow out the calculation mode between token vectors. Specifically, each token only performs self-attention calculations with the token at the corresponding hole distance position in the feature map. First, according to the pre-set window patch size and X f Perform a flattening operation to form Here Indicates the number of blocks. This means that a block of size Instead of treating the patch as a single feature unit, it is divided into N independent tokens, each of which corresponds to a pixel sub-region in the patch. Usually, the patch used in the network is square, that is, Furthermore, each token will only be spatially distant from it. The adjacent tokens of are used to construct global information, which effectively improves the computational efficiency of self-attention.

[0103] A similar processing strategy is adopted in the spatial dimension, and a partial split operation is introduced in the channel dimension to optimize the computing resource usage of the model while maintaining the model performance. Specifically, the feature X after spatial compression is split by the split coefficient n. s Divided into global modeling features and local modeling features In order to effectively manage the redundancy of high-resolution features in the early stages of the model and improve the computational efficiency of the model, the splitting factor n in stage 3 is stage-3 Set it to 4. Appropriately reduce the value of the split coefficient and set n stage-4 ,n stage-5 is set to 2 to avoid excessive compression and loss of information. st Feed it into the transformer module and use the above Att formula to capture the long-distance dependencies in the input features and enhance the global representation ability of feature representation. Local modeling feature X sc Will be combined with X at the output stage of the transformer module st Merge in the channel direction and restore the characteristic shape at the same time, output

[0104] The fusion submodule targets the input A mapping convolution with a convolution kernel size of 1×1 and a deep separation convolution with a convolution kernel size of 3×3 are used to fuse the global and local information of the features. The output features The main function of the mapping convolution with a kernel size of 1×1 is to recombine and map the feature channels while keeping the spatial dimension unchanged. Next, a deep separation convolution with a kernel size of 3×3 is used to efficiently fuse spatial information. Through the collaborative work of these two steps, the fusion submodule can effectively integrate global and local information. The output feature representation not only contains rich global context information, but also captures key details of the local area.

[0105] The final stage of the model is the key link to convert the internal learned features into the final prediction results. This stage consists of three components: convolution submodule, pooling flattening submodule and fully connected submodule. The three submodules work together to ensure that the model can effectively extract useful information from the learned high-dimensional features.

[0106] Furthermore, step S22: the specific method of the fine inspection method based on disease information feedback is as follows:

[0107] The disease detection network is used to automatically perform high-precision disease screening on the data collected in the rough inspection stage, and the image data containing diseases are independently summarized to form a disease set that is different from the image data of normal areas, such as Figure 8 As shown. Analyzing the images at this stage can achieve the effect of disease recognition, that is, to understand the existence of bridge diseases. In order to build a connection between the image and the real space, before the disease screening stage, the sequence of acquisition points based on the rough inspection route is Assign a label x to each collected image i After the screening stage, the real spatial location of the disease collection point can be quickly indexed based on the disease image label.

[0108] An adaptive inspection method based on rough inspection information feedback is proposed. While the UAV is performing efficient inspections, it can still maintain high-precision attention on the diseased area, such as Figure 8 As shown. First, combined with the information fed back during the disease screening phase, the drone's collection points are divided into two categories: normal and abnormal. At normal collection points, the drone uses the rough inspection path as the inspection benchmark. However, at abnormal collection points, the drone will adaptively adjust the inspection path based on the disease feedback information to obtain more accurate detail information. Reduce the size of the drone's collection field of view so that the image focuses on the local details of the disease. The field of view of the drone is mainly determined by two factors: the distance coefficient D and the focal length f. The collection field of view can be reduced by reducing D or increasing f. Keep the camera focal length unchanged and construct a distance target plane D in the inspection space. sThe solution of fine detection plane realizes the fine acquisition of field of view (W s ,H s ). Furthermore, the diseased area is projected onto the newly constructed fine detection plane as the optimization target. Here, the concept of dimensionality reduction analysis is adhered to, and the detection plane in the three-dimensional space is transformed into a two-dimensional plane. The diseased area corresponding to a single collection point is cut into a grid, where the size of each grid is S w ,S h Represent the division coefficients in width and height respectively. The center coordinates of these subgrids represent the collection points on the detailed inspection path. The same operation as above is used to traverse all diseased areas to develop a more detailed collection plan. The overlapping area measurement criterion is used to automatically filter and delete the collection points that exceed the limit threshold δ, so as to reduce resource waste and improve inspection efficiency. An efficient heuristic search algorithm is used to solve the connection problem between discrete collection points, and the detailed inspection path is constructed by minimizing the distance cost between neighboring points.

[0109] The inspection method based on fine inspection path can obtain a series of fine local image data, screen the collected data for diseases through the disease detection network, and quickly locate the real collection position according to the label index. Finally, the spatial transformation relationship between the collection point and the inspection target point is used, that is, the collection point is reversely mapped to the target representative space plane through the distance coefficient D, so as to realize the fine positioning of the disease area in the real space.

[0110] Example 1

[0111] Fangshan Bridge was used as a test object. Fig. 9 As shown. Fangshan Bridge is a bottom-supported prestressed concrete tie-arch bridge with a main span of 96.76m and a deck width of 7m. The main arch rib adopts a reinforced concrete structure, and the arch axis uses a quadratic parabola with a calculated span of 95m and a rise of 14m. In the process of the drone performing adaptive inspection tasks, it gets rid of the dependence on direct human control and realizes autonomous flight. In order to take into account both safety and convenience, the compact DJI M3T industrial drone was specially selected as the equipment carrier. The sensor size of the camera it carries is 1 / 2" (6.4×4.8mm), and the equivalent focal length is 24mm. The actual focal length obtained by conversion is 4.47mm. When the drone is located at a distance of 14m from the inspection surface of the bridge, 5 times the original focal length is used for data collection, and the actual field of view size is 4×3m (with an accuracy of 1mm / pixel). The drone in this state is used as the flight criterion for the rough inspection stage of the inspection method, that is, the flight distance D p =14m, camera focal length f p =22.35mm, collecting field width and height W p ×Hp =4009×3007mm≈4×3m. This stage focuses on the overall information of the bridge. When the UAV is 5m away from the target, using the same focal length, the actual field of view is 1.4×1m (accuracy is 0.35mm / pixel). The UAV in this state is used as the flight criterion for the detailed inspection stage of the inspection method, that is, the distance from the target plane D s =5m,f s =22.35mm,W s ×H s =1432×1074mm≈1.4×1m.

[0112] A drone was used to collect large scene views of the experimental bridge from multiple angles at a long distance, and then the 3D reconstruction task of multi-view geometry was performed using the motion recovery structure algorithm. The initial exploration of the properties of bridge components was carried out using the PointNeXt network. The instance segmentation results are shown in Figure 2. Fig.10 The inspection targets include cross braces, arch structures, cantilevers and beams. According to the distance coefficient D in the rough inspection stage p The inspection plane of each component is constructed. According to the detailed component attribute information, a customized inspection plan is made for each component, such as Fig.10 shown.

[0113] The data collected during the rough inspection phase will be input into the Light-PVIT model for rapid disease screening, such as Fig.11 shown.

[0114] In order to provide effective bridge operation advice, the real spatial location of the disease collection point can be quickly inferred based on the association between the disease collection image and the actual space, such as Fig.12 , 13 , 14 The rough inspection plane of the bridge arch, beam and cross brace is shown. The dark collection points in the figure indicate that the area covered by the location has defects, while the light collection points indicate intact areas. In the defective area, the drone will adaptively adjust the inspection path according to the defect information fed back in the rough inspection stage to obtain more accurate detailed information.

[0115] Construct the distance target plane D in the inspection space s The distance from the coarse inspection plane to the fine inspection plane is D ps =D p -D s The drone’s field of view is W p ×H p Adjust to W s ×H s. Divide all diseased areas into grids, where the size of each grid is w = 1mm, h = 0.75mm. At the same time, calculate the center points of these sub-grids as detailed inspection collection points. Automatically filter and delete the collection points with an overlap rate threshold exceeding 30%.

[0116] In summary, the specific embodiments verify the effectiveness of the solution proposed in the present invention and its applicability to complex projects.

[0117] What is disclosed above is only a typical embodiment of the present invention, but the embodiment of the present invention is not limited thereto. Any homogeneous modification made to this patent by any technician in this field after reading the patent should fall within the protection scope of the present invention.

Claims

1. An adaptive inspection method based on multi-layer representation learning using drones is applied to the inspection of bridge bottom structures, which is characterized by: The steps include: S1, automatic rough inspection based on bridge spatial understanding; S11. Use the large field of view camera carried by the drone to quickly obtain the geometric information of the bridge, and combine it with the motion recovery structure algorithm to quickly construct a three-dimensional model of the bridge; S12. Analyze the attribute information at the bridge component level and propose a point cloud instance segmentation framework based on an optimized clustering algorithm to achieve end-to-end comprehensive extraction of attribute information of each bridge component; S13. By simulating the field of view model and the geometric space dimensionality reduction projection method, the point cloud space is compressed and transformed, and efficient spatial inspection paths are automatically planned for each independent component of the bridge in the plane geometry, ensuring that the drone can conduct a comprehensive coverage inspection of the bridge components; S2: Adaptive detailed inspection based on disease information feedback; S21. Reorganize the self-attention mechanism from two key dimensions: space and channel, and build a lightweight model called Light-PVIT to identify the damaged areas of bridge components. S22. Combining the damaged area of ​​the bridge component identified in step S21 with the geometric space dimensionality reduction projection method, the inspection path planned in step S13 is adaptively adjusted to guide the drone to perform detailed inspection of the damaged area.

2. The UAV adaptive inspection method based on multi-layer representation learning according to claim 1 is characterized in that: Step S11 specifically includes the following sub-steps: S111, using satellite maps and prior information of bridges to plan the path of the UAV; S112. Collect multi-view 3D reconstruction data for the bridge, and then construct a 3D model of the entire scene.

3. The UAV adaptive inspection method based on multi-layer representation learning according to claim 1 is characterized in that: Step S12 specifically includes the following sub-steps: S121, using PointNeXt network as a tool for learning bridge component information representation, using encoder and decoder design concepts to process point cloud data and finely segment the structure; In this process, the encoder captures high-level semantic features in the point cloud, and the decoder remaps these high-level semantic features to construct a detailed segmentation structure of the point cloud; S122. Further intra-class division is performed based on semantic segmentation, and a strategy combining density-based spatial clustering post-processing method with semantic segmentation network is adopted until all points in the point cloud model are visited.

4. The UAV adaptive inspection method based on multi-layer representation learning according to claim 1 is characterized in that: Step S13 specifically includes the following sub-steps: S131. According to the simulated field of view model, a rough inspection route is planned for each component in the bridge point cloud space. The bridge component shape belongs to a regular spatial geometric body. Combining these characteristics, the point cloud space is compressed and transformed using the dimensionality reduction idea to describe the spatial relationship in the plane geometry. S132, randomly extract three non-collinear points P from the point cloud data set 1, ,P2,P3, constitute the initial point set as the basis for subsequent model fitting; calculate the normal vector of the initial point set Based on this, an initial spatial plane reference model is constructed; each remaining point in the component point cloud data is traversed to perform distance measurement operations, and any point whose distance to the plane model is lower than the tolerance threshold is identified as an inner point t; S133, repeat the above iterative process, select the model with the largest consensus set as the representative space plane, compress and project the remaining point cloud (X, Y, Z) in the bridge component space onto the representative space plane (X pj ,Y pj ,Z pj ); construct an inspection plane parallel to the representative spatial plane according to the distance coefficient D, reducing the spatial dimension that needs to be considered when designing the UAV flight path; further, transform and project the component point cloud on the representative spatial plane onto the inspection plane, lock the camera azimuth, that is, the camera acquisition plane is parallel to the plane of the bridge to be inspected, and the two key factors of the inspection planning are unified on the inspection plane, converting the inspection plane in the three-dimensional space into a two-dimensional plane model; S134. Based on the regularity of bridge components, the center line of the bridge component is used as the reference of the UAV inspection path on the two-dimensional inspection plane, and the center line of the bridge component is quickly extracted by the boundary fitting method; the boundary fitting direction follows the aspect ratio criterion, that is, the width of the bridge component and the height of the bridge component are compared numerically. When the width of the bridge component is greater than the height of the bridge component, the width direction of the bridge component is taken; when the height of the bridge component is greater than the width of the bridge component, the height direction of the bridge component is taken; a polynomial is used to fit the boundary; S135. Plan camera acquisition points on the center line of the bridge component in combination with the field of view model, and consider different combinations of distance coefficients d and focal lengths f to ensure that the coverage range is greater than the boundary spacing; in the rough inspection stage, increase the distance coefficient d, reduce the focal length f, increase the acquisition field of view, and then reduce the overlap rate δ to cover the bridge component in a manner with an optimal number of acquisition points; S136. Restore the collection point coordinates on the two-dimensional plane to three-dimensional coordinates through reverse mapping to obtain a complete inspection path.

5. The UAV adaptive inspection method based on multi-layer representation learning according to claim 1 is characterized in that: Step S21 specifically includes the following sub-steps: S211, the Light-PVIT lightweight model consists of six stages of learning tasks. The first five stages of learning tasks consist of a convolution module CVB, a deep separation convolution module (RDB2, RDB1) with an inverse residual structure, and a partial visual transformer module PVIT. The deep separation convolution module with an inverse residual structure is used as an efficient replacement for the convolution layer. The first step in each learning stage is to downsample the features using an inverse deep separation convolution module without residual structure. The partial visual transformer module PVIT consists of three submodules, namely the local representation submodule, the global representation submodule and the feature fusion submodule. A depth separation convolution with a convolution kernel size of 3×3 and a mapping convolution with a convolution kernel size of 1×1 are used to model the local information of the features and output the features. in, H, W, C s Respectively represent the height, width and number of channels of the output feature map; The global representation submodule introduces the transformer's self-attention mechanism to capture global context information on the feature map, so as to obtain the global information of the H×W feature map; S213, introduce a partial split operation in the channel dimension. Specifically, the spatial compression feature X is split by the split coefficient n. s Divided into global modeling features and local modeling features To manage the redundancy of high-resolution features in the early stages of the model, the split factor n in stage 3 is stage-3 Set to 4; the split coefficient n of stage 4 stage-4 and the splitting factor n in stage 5 stage-5 Set them to 2 to avoid information loss caused by excessive compression; S214, the global modeling feature X st Feeding the transformer module captures the long-range dependencies in the input features, and the local modeling features X sc The output of the transformer module will be combined with the global modeling feature X st Merge in the channel direction, restore the feature shape, and output the hybrid modeling feature The feature fusion submodule is used for the mixed modeling feature Integrate global and local information; S215, the final stage of the model uses a convolution submodule, a pooling flattening submodule, and a fully connected submodule to convert the features of stage six into the final prediction results.

6. The UAV adaptive inspection method based on multi-layer representation learning according to claim 1 is characterized in that: Step S22 specifically includes the following sub-steps: S221, using a disease detection network to automatically screen the data collected in the rough inspection stage for diseases, and at the same time independently summarize the image data containing diseases to form a disease set that is different from the image data of normal areas; Before the disease screening stage, the order of collection points based on the rough inspection route Assign a label x to each collected image i ,After the screening stage, the real spatial location of the disease collection point can be ,quickly indexed based on the disease image label; An adaptive inspection method based on rough inspection information feedback is proposed to enable the UAV to inspect efficiently while keeping attention on the diseased area. First, the collection points of the UAV are divided into normal and abnormal categories based on the feedback information from the disease screening stage. At normal collection points, i.e., areas where bridge components are not damaged, the drone uses the rough inspection path as the inspection benchmark; At abnormal collection points, i.e., areas where bridge components are damaged, the drone adaptively adjusts the detection path based on the disease feedback information to obtain more accurate detailed information; Reduce the size of the drone's field of view and record the local details of the disease; the field of view can be reduced by reducing the distance coefficient d or increasing the focal length f. Keep the camera focal length unchanged and construct a distance target plane D in the inspection space. s The solution of fine detection plane realizes the fine acquisition of field of view (W s ,H s ) can be adjusted flexibly, W s ,H s Indicates the height and width of the acquisition field of view; S222, further, projecting the diseased area as the optimization target onto the newly constructed fine detection plane, transforming the detection plane in the three-dimensional space into a two-dimensional plane, and performing grid-cutting on the diseased area corresponding to a single collection point, wherein the width of each grid is high S w ,S h Respectively represent the division coefficients in width and height. The center coordinates of these sub-grids represent the acquisition points on the detailed inspection path; S223, traverse all diseased areas using the same operation as step S222, and formulate a detailed collection plan; The overlapping area measurement criterion is used to filter and delete the collection points that exceed the limit threshold δ; A heuristic search algorithm is used to connect discrete acquisition points and construct a detailed inspection path by minimizing the distance cost between neighboring points; S224. Finally, the spatial conversion relationship between the collection point and the inspection target point is used, that is, the collection point is projected onto the bridge surface through the distance coefficient D to locate the defective area in real space.

Images