High-precision asphalt pavement repairing system and method based on roof dual-unmanned aerial vehicle heterogeneous cooperation

Abstract

The application discloses a high-precision asphalt pavement repairing system and method based on roof dual-unmanned aerial vehicle (UAV) heterogeneous cooperation, and belongs to the technical field of intelligent maintenance of asphalt pavement. The system takes an intelligent unmanned repairing vehicle as a carrier, integrates a dual-UAV heterogeneous carrying operation and maintenance unit and a vehicle-machine cooperative central control unit, a first UAV carries visible light and infrared equipment to complete wide-area inspection and suspected disease rapid identification, and a second UAV carries a laser radar and a long-focus camera to realize high-precision three-dimensional scanning and imaging of a disease area; the vehicle-machine cooperative central control unit fuses infrared and point cloud data through a multi-modal AI identification module, analyzes disease types, outlines, depths and internal void ratios, generates a smooth and continuous repairing track according to a sub-pixel level path planning module, and through a dynamic closed-loop control module of a three-ring nested structure, linkage automatic driving, asphalt heating, paving and compaction and the like are completed to realize automatic repairing operation. The application realizes unmanned closed loop of the whole process of disease detection, path planning, accurate repairing and quality review.

Description

Technical Field

[0001] This invention relates to the field of intelligent maintenance technology for asphalt pavements, specifically to a high-precision asphalt pavement repair system and method with heterogeneous collaboration of dual unmanned aerial vehicles on a vehicle roof. Background Technology

[0002] Asphalt pavement, as the mainstream pavement structure for highways and urban roads in my country, is subjected to heavy traffic loads, temperature fluctuations, and rainwater erosion over long periods, making it prone to defects such as cracks, potholes, loosening, and internal voids. These defects directly affect road traffic quality, structural safety, and service life. Traditional pavement maintenance relies mainly on manual inspections and semi-mechanical repairs, which suffer from low detection efficiency, inaccurate quantification of defects, poor operational precision, long traffic closure times, and high labor intensity. This makes it difficult to meet the rapid, refined, and unmanned maintenance needs of high-grade highways and high-density urban road networks. With the rapid penetration of drones, autonomous driving, multimodal perception, and AI recognition technologies, air-ground collaboration and vehicle-machine integration are becoming the core development trends in the field of intelligent asphalt pavement repair.

[0003] Existing academic literature and conventional technologies have explored UAV-based pavement defect detection and intelligent repair. The paper "UAV Detection of Road Defects Using Multi-Scale Fusion Strategy and Improved YOLOv5," published in the *Journal of Geoinformation Science* in 2025, proposes a defect identification method based on aerial imagery, enabling two-dimensional image classification and localization. The paper "Road Crack Localization and Quantification Method Based on UAV Monocular Video," published in *Engineering Mechanics* in 2025, attempts to estimate crack size, but neither can obtain three-dimensional structural parameters such as defect depth and internal porosity, and both suffer from large errors in calculating filler volume. Several other studies based on visible light / infrared single-modal detection generally suffer from significant shortcomings, including a disconnect between wide-area inspection and fine-scale detection, a lack of data-driven paving operations, and a deficiency in three-dimensional quantization capabilities.

[0004] Chinese patent CN117371807A discloses a road-air collaborative digital management and control system for pavement defects. This system uses drones to collect multimodal data and links it with ground terminals to achieve digital identification of defects and planning of work paths, thus improving the problems of slow and inefficient manual inspections to some extent. However, this technology still has significant limitations: it uses a single drone platform, making it unable to simultaneously handle wide-area rapid screening and local high-precision 3D scanning; it does not achieve feature-level fusion of infrared thermal imaging and laser point clouds, making it difficult to accurately invert internal porosity; and the work path is only planned at the pixel level, lacking dynamic adjustment of paving flow and online closed-loop control of compaction quality, thus failing to meet the requirements for high-precision, high-quality integrated repair of asphalt pavements.

[0005] Therefore, it is urgent to fundamentally solve the industry pain points of existing technologies, such as incomplete detection, inaccurate identification, imprecise positioning, suboptimal operation, and uncontrollable quality, and to achieve a closed-loop operation of rapid detection, accurate quantification, intelligent repair, and completion verification of asphalt pavement defects. Summary of the Invention

[0006] Based on the above-mentioned technical problems, this application discloses a high-precision asphalt pavement repair system and method with heterogeneous collaboration of dual unmanned aerial vehicles on the roof; the high-precision asphalt pavement repair system with heterogeneous collaboration of dual unmanned aerial vehicles on the roof includes an intelligent unmanned asphalt pavement repair vehicle body, the vehicle body is provided with a vehicle chassis, an autonomous driving unit, an asphalt storage and heating unit, an automatic paving and compaction repair operation unit, and also includes a dual unmanned aerial vehicle heterogeneous mounting operation and maintenance unit fixedly installed on the top of the vehicle chassis, as well as a vehicle-machine collaborative central control unit;

[0007] The dual-drone heterogeneous operation and maintenance unit includes a dual-mode inspection drone group, a vehicle-mounted automatic charging mechanism, and a drone control submodule. The dual-mode inspection drone group includes a first drone and a second drone. The drone control submodule is electrically connected to the vehicle-mounted automatic charging mechanism, the first drone, and the second drone, respectively.

[0008] The vehicle-machine collaborative central control unit is communicatively connected to the autonomous driving unit, the automatic paving and compaction repair operation unit, and the drone control submodule. The vehicle-machine collaborative central control unit has a built-in road surface defect multimodal AI recognition module, a sub-pixel level operation path planning module, and a repair operation dynamic closed-loop control module.

[0009] The first drone is equipped with a wide-angle visible light camera and an infrared thermal imager, and is used to perform wide-area road inspections and quickly screen suspected defect areas;

[0010] The second drone is equipped with a high-precision lidar and a telephoto camera, and is used to perform close-range, high-precision three-dimensional scanning and imaging of suspected disease areas screened by the first drone.

[0011] The road surface defect multimodal AI recognition module is used to fuse the infrared thermal imaging data of the first UAV and the point cloud data of the second UAV to identify the type, outline, depth and internal porosity of road surface defects.

[0012] The subpixel-level operation path planning module is used to generate a repair operation path with subpixel-level accuracy based on the identified defect location coordinates and the real-time vehicle pose, and output it to the repair operation dynamic closed-loop control module.

[0013] The repair operation dynamic closed-loop control module is used to drive the autonomous driving unit to park the vehicle at the work position, and to control the asphalt storage and heating unit and the automatic paving and compaction repair operation unit to complete the repair operation, and to adjust the paving parameters in real time during the operation.

[0014] Preferably, the road surface defect multimodal AI recognition module includes a feature fusion submodule, which is used to integrate the two-dimensional infrared feature map output by the first UAV. The three-dimensional point cloud feature map output by the second UAV Registration and fusion are performed; the registration process is achieved by solving the following nonlinear optimization problem, as shown in the formula: ,in, Let be a rotation matrix. It is a translation vector. For feature points in a 3D point cloud, These are the corresponding feature points in the two-dimensional infrared feature map. For camera projection model, The number of matching feature point pairs;

[0015] fused feature vector Used as input to a deep residual network, outputting the depth of the disease. and the internal porosity of the diseased area .

[0016] Preferably, the sub-pixel-level job path planning module is specifically used for:

[0017] Based on the identified disease outline and location coordinates, a local coordinate system is established with the center of the vehicle's rear axle as the origin.

[0018] Within the local coordinate system, a smooth paving path is generated based on B-spline curves. This ensures that the movement trajectory of the paving head can completely cover the diseased area and that the path curvature is continuous, wherein: ,in, To control the vertices, For p-th degree B-spline basis functions;

[0019] Meanwhile, based on the internal porosity Calculate the required asphalt mixture fill volume : ,in, The area affected by the disease. To design the paving thickness, is the pore filling coefficient.

[0020] Preferably, the repair operation dynamic closed-loop control module includes a material rheology control submodule; the material rheology control submodule is used to control the material rheology based on the real-time paving speed of the automatic paving and compaction repair operation unit. and paving width Dynamically adjust the conveying flow rate of asphalt mixture To ensure the uniformity of the paving thickness, the flow rate... The regulation follows the following rheological formula: ,in, For the target thickness, The density of the asphalt mixture, This is the rheological correction factor based on temperature compensation.

[0021] Preferably, the automatic paving, compaction and repair operation unit includes a six-degree-of-freedom robotic arm and a repair execution head disposed at the end of the robotic arm;

[0022] The repair execution head integrates an infrared thermometer, a laser displacement sensor, and a compaction wheel;

[0023] The vehicle-machine collaborative central control unit also has a built-in online compaction quality monitoring module, which is used to calculate the paving smoothness in real time based on the road surface elevation data returned by the laser displacement sensor. and in When the preset threshold is exceeded, dynamic compensation is performed by adjusting the posture of the six-degree-of-freedom robotic arm.

[0024] Preferably, the collaborative control logic between the first UAV and the second UAV is as follows:

[0025] The first UAV performs wide-area scanning along a zigzag route while the vehicle is moving. When a suspected disease is detected, it locks the approximate geographical coordinates of the disease. ;

[0026] The drone control submodule controls the second drone to take off from the storage mechanism and fly to the approximate geographical coordinates. Hover over;

[0027] The second UAV activates a high-precision lidar to scan the target defect from all angles, generates a three-dimensional model, and transmits the model back to the vehicle-machine collaborative central control unit for precise identification.

[0028] Preferably, the vehicle-machine collaborative central control unit further includes a vision-assisted positioning submodule; the vision-assisted positioning submodule is used to control the on-board high-definition camera to capture road surface texture features when the vehicle approaches the target operation position, and match them with feature points in the high-precision three-dimensional model transmitted back by the second UAV to calculate the sub-pixel level deviation of the vehicle relative to the center of the defect. The deviation is then fed back to the autonomous driving unit for fine-tuning.

[0029] Preferably, the dual-drone heterogeneous mounting and maintenance unit also includes a composite storage compartment mounted on the roof of the vehicle; the composite storage compartment is equipped with a shock absorption mechanism and an automatic locking mechanism. When the first drone or the second drone lands, the automatic locking mechanism fixes the drone landing gear by electromagnetic adsorption to prevent vibration during vehicle operation from damaging the drone.

[0030] Preferably, the asphalt storage and heating unit includes a phase change heat storage layer and a microwave-assisted heating layer; the microwave-assisted heating layer is used to rapidly heat the local asphalt mixture before repair work, so that it reaches the optimal paving rheological temperature. The vehicle-mounted central control unit adjusts according to the ambient temperature. and the current temperature of the asphalt mixture Dynamic control of microwave power : ,in, For specific heat capacity, For quality, The heat exchange coefficient, is the surface area.

[0031] The high-precision asphalt pavement repair method using a dual-UAV heterogeneous collaborative system on a vehicle roof, applied to the aforementioned system, includes the following steps:

[0032] The first drone was launched to conduct a wide-area inspection, using infrared thermal imaging to quickly detect areas of abnormal road surface temperature and mark them as suspected defects;

[0033] The second drone was controlled to perform a high-precision 3D scan of the suspected disease area to obtain data on the depth, outline, and internal porosity of the disease.

[0034] Based on the damage data, B-spline curves are used to plan sub-pixel level paving paths and calculate the required asphalt filling volume;

[0035] Drive the vehicle to the work position, use visual-assisted positioning to perform sub-pixel-level correction, and start the automatic paving, compaction and repair work unit after stopping;

[0036] During the paving process, the material flow rate is dynamically adjusted according to the rheological formula, and the paving smoothness is monitored in real time using a laser displacement sensor.

[0037] After the repair is completed, the second drone is controlled to scan the area again for verification. If the flatness or density is found to be substandard, the secondary repair process is triggered.

[0038] Compared with the prior art, the technical solution of this application has the following technical effects:

[0039] This invention enables efficient wide-area screening and high-precision three-dimensional detection of road surface defects through heterogeneous collaborative operation of two drones on the roof. The first drone quickly locates suspected defect areas, while the second drone performs detailed scanning and imaging of the target area, comprehensively acquiring information on the external morphology and internal structure of the defects. This provides complete and reliable raw data support for subsequent repair work, ensuring the stable and efficient operation of the defect detection process.

[0040] The multimodal AI recognition module of this invention can integrate infrared imaging and three-dimensional point cloud data to accurately identify the contour depth and internal porosity of disease types. Based on feature fusion and deep learning algorithms, it improves the completeness and accuracy of disease analysis, providing a scientific basis for work path planning and material usage calculation, and making disease quantitative analysis more in line with actual repair needs.

[0041] This invention employs sub-pixel-level operation path planning and dynamic closed-loop control technology, which can generate smooth and continuous operation paths, accurately match the scope of the damaged area, and adjust the asphalt delivery flow rate and paving and compaction parameters in real time according to the operation status. With the help of dynamic compensation from the robotic arm, it can improve the paving smoothness and compaction uniformity, ensure that the repair operation is stable and controllable throughout the process, and improve the quality of road repair.

[0042] This invention enables integrated intelligent operation of vehicles and drones, forming a complete closed loop from defect detection, path planning, positioning and docking to paving, compaction and quality verification. The vehicle-mounted storage and automatic charging mechanism ensures stable operation and maintenance of the drone, and the asphalt storage and heating unit can precisely control the material temperature. The overall system has a high degree of automation and can continuously and reliably complete high-precision repair of road defects, improving the efficiency of road maintenance.

[0043] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0044] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0046] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0047] Figure 1 Schematic diagram of the overall structure of the intelligent unmanned asphalt road repair vehicle and the dual-UAV heterogeneous operation and maintenance unit;

[0048] Figure 2 Schematic diagram of the integrated operation process of dual drone inspection and road surface repair under vehicle-machine collaboration;

[0049] Figure 3 : Schematic diagram of the dual-UAV heterogeneous inspection architecture and the principle of all-domain perception and high-precision positioning of road defects;

[0050] Figure 4 Schematic diagram of the three-loop nested control structure of the dynamic closed-loop control module for repair operations;

[0051] Figure 5 Schematic diagram of a network architecture for identifying road surface distress parameters by fusing multimodal infrared and point cloud data;

[0052] Figure 6 Flowchart of closed-loop control logic for fully unmanned operation of the vehicle-machine collaborative central control unit. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0054] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0055] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0056] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0057] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0058] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0059] Example 1

[0060] This embodiment mainly describes a high-precision asphalt pavement repair system with heterogeneous collaboration between two unmanned aerial vehicles on a vehicle roof, such as... Figure 1 As shown, it specifically includes the intelligent unmanned asphalt road repair vehicle body, which is equipped with a vehicle chassis, an autonomous driving unit, an asphalt storage and heating unit, an automatic paving, compaction and repair operation unit, and also includes a dual unmanned heterogeneous mounted operation and maintenance unit fixedly installed on the top of the vehicle chassis, as well as a vehicle-machine collaborative central control unit.

[0061] The dual-drone heterogeneous operation and maintenance unit includes a dual-mode inspection drone group, a vehicle-mounted automatic charging mechanism, and a drone control submodule. The dual-mode inspection drone group includes a first drone and a second drone. The drone control submodule is electrically connected to the vehicle-mounted automatic charging mechanism, the first drone, and the second drone, respectively.

[0062] The vehicle-machine collaborative central control unit is communicatively connected to the autonomous driving unit, the automatic paving and compaction repair operation unit, and the drone control submodule. The vehicle-machine collaborative central control unit has a built-in road surface defect multimodal AI recognition module, a sub-pixel level operation path planning module, and a repair operation dynamic closed-loop control module.

[0063] The first drone is equipped with a wide-angle visible light camera and an infrared thermal imager, and is used to perform wide-area road inspections and quickly screen suspected defect areas;

[0064] The second drone is equipped with a high-precision lidar and a telephoto camera, and is used to perform close-range, high-precision three-dimensional scanning and imaging of suspected disease areas screened by the first drone.

[0065] The road surface defect multimodal AI recognition module is used to fuse the infrared thermal imaging data of the first UAV and the point cloud data of the second UAV to identify the type, outline, depth and internal porosity of road surface defects.

[0066] The subpixel-level operation path planning module is used to generate a repair operation path with subpixel-level accuracy based on the identified defect location coordinates and the real-time vehicle pose, and output it to the repair operation dynamic closed-loop control module.

[0067] The repair operation dynamic closed-loop control module is used to drive the autonomous driving unit to park the vehicle at the work position, and to control the asphalt storage and heating unit and the automatic paving and compaction repair operation unit to complete the repair operation, and to adjust the paving parameters in real time during the operation.

[0068] It also includes the addition of an automatic placement and recycling warning cone module to the front of the vehicle chassis and above the autonomous driving unit of the intelligent unmanned asphalt road repair vehicle. This module is connected to the vehicle-machine collaborative central control unit and the autonomous driving unit, and is synchronously integrated into the control logic of the repair operation dynamic closed-loop control module.

[0069] The automatic deployment and recycling warning cone module can automatically deploy warning cones at safe intervals around the work area before repair work begins, based on the location of the damage and the vehicle's trajectory. This forms a closed warning zone, effectively reminding passing vehicles to give way and ensuring road safety during unmanned repair work. After the work is completed, the module can automatically recycle the warning cones and put them back in their proper places. The entire process requires no manual operation and forms a fully unmanned collaborative process with dual drone inspections and asphalt paving and compaction operations, improving the safety protection and automated closed-loop capabilities of asphalt pavement repair operations.

[0070] like Figure 2 As shown, this demonstrates the closed-loop operation process of dual-UAV heterogeneous inspection and road surface defect repair under vehicle-machine collaboration. Within the scanning area, the dual-mode inspection UAV group performs wide-area road surface inspection according to the optimal path planning. It acquires road surface temperature field and three-dimensional point cloud data through composite scanning, and completes the preliminary judgment and coordinate locking of suspected defects through the road surface defect identification module. After receiving the defect information, the vehicle-machine collaborative central control unit schedules the road surface defect repair operation unit in the repair area. Based on the three-ring nested control logic, it carries out precise repair operations, realizing unmanned collaborative operation closed loop of wide-area screening, defect identification, and precise repair. The entire process does not require human intervention and can efficiently complete the full-area perception and automated repair of road surface defects.

[0071] like Figure 3 As shown, the dual-UAV heterogeneous operation and maintenance unit is the core hardware carrier for realizing rapid perception and high-precision positioning of road surface defects across the entire area. Its overall structure and working logic are designed in an integrated manner around the roof mounting space and vehicle driving conditions. The dual-mode inspection UAV group adopts a heterogeneous division of labor architecture. The first UAV is designed with wide-area coverage and rapid screening as its core design goals. It adopts a large wheelbase multi-rotor aerodynamic layout. The fuselage payload bay integrates a global shutter wide-angle visible light camera and an uncooled infrared thermal imager. The visible light camera has a resolution of 4K or higher and global exposure characteristics, which can avoid motion blur when the vehicle is traveling at high speed. The infrared thermal imager has a temperature measurement accuracy of less than 0.5℃ and a frame output frequency of more than 25fps. It can capture the small temperature differences caused by defects such as structural damage, internal voids, and water content in asphalt pavement. Its field of view is not less than 60°, and it can achieve single-frame large-area road surface coverage at a flight altitude of 5m to 10m.

[0072] The second UAV is designed for close-range, high-precision measurement. It adopts a small wheelbase, high-precision multi-rotor configuration and is equipped with a 128-line automotive-grade LiDAR and a long-focus fixed-focus industrial camera. The LiDAR has a ranging accuracy of ±5mm, an angular resolution better than 0.1°, and a single-point ranging frequency of no less than 100kHz. It can generate high-density three-dimensional point clouds within a working distance of 0.5m to 5m. The long-focus camera has micron-level imaging pixels and can simultaneously acquire high-definition texture information of the diseased area.

[0073] The UAV control submodule uses an embedded real-time processor as its core, with a main frequency of no less than 1GHz. It integrates CAN bus, 4G / 5G, and Wi-Fi 6 tri-mode communication interfaces to establish low-latency point-to-point communication links with the first and second UAVs. The end-to-end latency is controlled within 10ms. This submodule has built-in independent trajectory calculation, attitude stabilization control, and task scheduling units. It can autonomously complete the entire process control of UAV takeoff, cruise, hovering, scanning, return, and landing, and only uploads status information and perception data to the vehicle-machine collaborative central control unit.

[0074] The vehicle-mounted automatic charging mechanism adopts a composite charging mode of wireless induction and contact probe. It integrates a wireless charging transmitting coil, a flexible contact charging probe, a charging status monitoring circuit, and an over-temperature and over-voltage protection module. The output voltage is continuously adjustable in the range of 12V to 24V, and the output current is smoothly adjusted from 0A to 10A. The charging conversion efficiency is not less than 90%. It can automatically trigger the charging process after the drone lands. When the battery power reaches 95%, it automatically cuts off the charging circuit and transmits parameters such as battery voltage, charging current, cell temperature, and remaining power back to the drone control submodule in real time, so as to achieve autonomous maintenance of the drone's endurance.

[0075] Furthermore, the multimodal AI recognition module for road surface defects is the core for achieving quantitative and accurate identification of defects. It consists of an infrared data preprocessing submodule, a point cloud preprocessing submodule, a feature fusion submodule, a deep residual network recognition submodule, and a defect parameter calculation submodule. The raw infrared thermal imaging data transmitted back by the first UAV undergoes a multi-level preprocessing process, including non-uniformity correction, defect removal, temperature and grayscale linear calibration, and Gaussian noise filtering, outputting a standardized two-dimensional infrared feature map. The pixel grayscale values ​​of this feature map have a strict linear mapping relationship with the actual road surface temperature, which can accurately reflect the road surface temperature field distribution characteristics.

[0076] The raw 3D point cloud data transmitted back by the second UAV needs to undergo noise reduction filtering, ground point segmentation, disease area clustering, and downsampling optimization processing to reduce data redundancy while preserving the core geometric features of the disease, and output a 3D point cloud feature map. The effective point cloud density is no less than 500 points / cm², providing a high-quality data foundation for subsequent feature registration and fusion. The feature fusion submodule performs rigid registration of 2D infrared features and 3D point cloud features, achieving precise alignment of spatial coordinates by solving a nonlinear optimization problem. The optimization objective function is: In the formula It is a 3×3 orthogonal rotation matrix that satisfies Furthermore, the determinant value is 1, describing the rotational attitude of the 3D point cloud in space. It is a three-dimensional translation vector containing displacement components in the X, Y, and Z directions, describing the spatial position offset of the point cloud. For the 3D point cloud feature map, the first The three-dimensional coordinates of each feature point are directly acquired by the lidar. In the two-dimensional infrared feature map and Matching the first The two-dimensional pixel coordinates of each feature point are used to establish the correspondence through the SIFT feature matching algorithm. The specific formula for the pinhole projection model of an infrared camera is as follows: ,in and These are the focal lengths of the infrared camera in the X and Y directions, respectively. and The coordinates of the camera's principal point were obtained in advance through camera intrinsic parameter calibration experiments. To effectively match the number of feature point pairs, the system constrains... The value should be no less than 100 to ensure registration stability and accuracy. After registration, a feature-level fusion operation is performed to fuse the feature vectors. It is obtained by weighted combination of infrared features and point cloud features, and the calculation formula is as follows: ,in This is a 128-dimensional depth feature vector extracted from a two-dimensional infrared feature map. This is a 256-dimensional geometric feature vector extracted from a 3D point cloud feature map. and For adaptive learning weight coefficients, satisfying It can automatically adjust according to the type of disease and the complexity of the scene. The feature vector after fusion is input into a 50-layer deep residual network for disease parameter inference.

[0077] The deep residual network consists of an input layer, a batch normalization layer, a convolutional layer, a residual module, a pooling layer, and a dual-branch fully connected output layer. The residual module adopts a bottleneck structure to reduce computation and alleviate the gradient vanishing problem. After training and convergence on a large scale of multi-type disease samples, the network has stable generalization ability. The dual-branch output layer realizes the regression output of disease depth and internal porosity, respectively. The depth output branch is in millimeters with an output accuracy of ±0.1mm, and the porosity output branch has a value range of 0% to 100% with an output accuracy of ±0.5%. At the same time, the network is equipped with a semantic segmentation branch, which obtains disease contour information through a sub-pixel edge extraction algorithm. The contour positioning error is controlled within 0.1 pixels, and the geometric shape and spatial distribution information of the disease can be output completely.

[0078] The sub-pixel-level operation path planning module is the core algorithm linking defect perception and precise operation. The module establishes a local right-handed coordinate system with the vehicle's rear axle center as the origin, the vehicle's longitudinal direction of travel as the X-axis, the lateral direction as the Y-axis, and the vertical upward direction as the Z-axis. This coordinate system is coupled in real-time with the vehicle's GPS / IMU combined positioning system. The positioning system outputs a position accuracy of ±1cm and a heading angle accuracy of ±0.01°. It can stably convert the global geographic coordinates of the defect into coordinate values ​​in the local coordinate system, achieving dynamic binding between the defect location and the vehicle's real-time pose. The core of the path generation uses a cubic B-spline curve to construct a continuous and smooth paving trajectory. The curve formula is as follows: In the formula This represents a continuous trajectory for paving the path. The normalized path parameters increase uniformly within the range of 0 to 1. To control the number of vertices, the system adaptively determines the number based on the complexity of the disease outline, and the number is not less than 5. The degree of the B-spline curve is fixed at 3 to ensure that the trajectory is second-order continuous and differentiable, thus meeting the requirements for smooth robot arm motion. For the first The three-dimensional coordinates of each control vertex are generated by the joint constraints of the key points of the defect contour, the operational safety boundary, and the motion space range of the robotic arm. The basis functions are p-order B-spline functions, calculated using the recursive formula. The zero-order basis function is defined as follows: Higher-order basis functions are obtained recursively from lower-order basis functions, and the recursive formula is: The node vectors are constructed uniformly to ensure a uniform distribution of trajectory curvature. The generated paving path must meet the following constraints: complete coverage of the defect contour by 5mm, maximum path curvature not exceeding 0.01mm⁻¹, path length error ±1mm, and sub-pixel positioning accuracy ≤0.01 pixels. Simultaneously with path generation, the module calculates the asphalt mixture filling volume using the following formula: In the formula To fill the theoretical volume, The area of ​​the lesion is a two-dimensional projected area, calculated using a contour integral algorithm with an accuracy of ±0.1 cm². The designed paving thickness ranges from 20mm to 50mm, determined based on the pre-set road surface grade. This is the pore filling coefficient, ranging from 1.0 to 1.5, determined by the asphalt mixture gradation type. The porosity of the defect is directly output by the multimodal AI recognition module. To compensate for material loss during operation, the system introduces a volume compensation term. ,in The loss compensation coefficient, ranging from 0.02 to 0.05, represents the final total filling volume. The total volume value is output to the repair operation dynamic closed-loop control module in real time, providing a precise quantitative basis for material transportation.

[0079] Furthermore, such as Figure 4 As shown, the dynamic closed-loop control module for repair operations adopts a nested control structure of position loop, speed loop, and flow loop, with a control cycle of no more than 10ms. It can achieve integrated control of precise vehicle parking, stable material delivery, and dynamic adjustment of paving and compaction. The material rheology control submodule is the core unit ensuring uniform paving thickness. It dynamically adjusts the asphalt mixture delivery flow rate based on real-time operating parameters. The flow rate calculation formula is... In the formula To output traffic in real time, The real-time movement speed of the paving head is collected by the robotic arm encoder, with an accuracy of ±0.1 mm / s. The paving width is determined by the structural parameters of the actuator head or adjusted in real time. The target paving thickness should be consistent with the design paving thickness. The density of the asphalt mixture is fixed at 2.3 g / cm³ to 2.5 g / cm³. The temperature-compensated rheological correction factor is derived from the exponential formula. The calculation yielded, where This is the reference correction factor at standard temperature. For temperature sensitivity coefficient, For optimal paving rheological temperature, For real-time asphalt temperature control, a PID closed-loop control algorithm is used for flow regulation, and the output control quantity is... In the formula To set the flow rate value, This is the actual traffic feedback value. , , These are the proportional, integral, and derivative coefficients, respectively. By tuning these parameters, the flow error can be controlled within ±1%.

[0080] The automated paving, compaction, and repair unit is equipped with a six-degree-of-freedom robotic arm and an integrated repair actuator. The actuator integrates an infrared thermometer, a laser displacement sensor, and a compaction wheel. The infrared thermometer has a sampling frequency of at least 10Hz and a temperature measurement accuracy of ±1℃, monitoring the asphalt temperature in real time after paving. The laser displacement sensor has a sampling interval of ≤0.5mm and a distance measurement accuracy of ±0.05mm, continuously collecting road surface elevation data. The compaction wheel has a diameter of 100mm to 150mm, a width of 50mm to 80mm, and an adjustable operating pressure of 0N to 500N. The online compaction quality monitoring module calculates the paving smoothness based on the elevation data. The calculation formula is: In the formula For the first Values ​​of each elevation sampling point, This represents the average elevation of the sampling points. The number of effective sampling points must be no less than 100, when the flatness... Exceeding the preset threshold At that time, the system automatically generates compensation instructions, including elevation compensation amounts. Robotic arm joint angle compensation ,in and To compensate for the coefficient, the flatness of the paved surface is dynamically corrected by adjusting the posture of the robotic arm in real time, ensuring that the quality of the operation meets the specifications.

[0081] Furthermore, the dual-UAV collaborative control logic follows a layered perception strategy of first wide-area screening and then fine measurement. The first UAV performs full-area inspection along a zigzag route during vehicle movement. The route spacing is determined by the infrared camera's field of view and flight altitude, with the value set at 0.8 times the ground width corresponding to the field of view. The flight altitude is controlled between 5m and 10m, and the flight speed is synchronized with the vehicle's speed in real time. The infrared thermal imager continuously collects road surface temperature field data. When the road surface temperature gradient exceeds the system's set threshold... Immediately identify the area as a suspected disease target, locate and upload the approximate geographic coordinates of the disease. Upon receiving the coordinates of a suspected defect, the UAV control submodule immediately sends a takeoff command to the second UAV.

[0082] The second UAV takes off vertically from the roof-mounted composite storage compartment, flies along the shortest path to the target coordinates, enters hovering operation mode with a hovering positioning accuracy of ±5cm, and an operation height of 1m to 3m. Then, it activates a high-precision LiDAR to perform a full-angle 3D scan, completing a 360° rotation scan in the horizontal direction and a reciprocating scan in the pitch direction within the range of -15° to +15°, generating a high-precision 3D mesh model of the disease. The model has no less than 10,000 triangular facets, which can completely restore the geometric shape and internal structure information of the disease. The 3D model data is transmitted back to the vehicle-machine collaborative central control unit in real time to complete the precise identification of the disease. After the scanning operation is completed, the second UAV automatically returns and lands in the storage compartment. The UAV control submodule switches its status to standby charging mode, waiting for the next task scheduling command. The entire process can complete the heterogeneous collaborative perception closed loop without human intervention.

[0083] Furthermore, the visual-assisted positioning submodule is crucial for achieving sub-pixel-level precise vehicle parking. It comprises an onboard high-definition camera, a feature extraction unit, a pose calculation unit, and a deviation feedback unit. When the vehicle autonomously approaches the vicinity of the work area, the visual-assisted positioning submodule is automatically activated. The onboard high-definition camera acquires real-time raw road images, and the feature extraction unit simultaneously extracts SIFT and SURF feature points from the images. The real-time acquired road feature points are robustly matched with the feature points of the high-precision 3D model of the defect previously transmitted by a second UAV. The feature matching similarity threshold is set to no less than 0.8 to ensure matching accuracy and robustness. The pose calculation unit uses the PnP algorithm to solve the 3D pose relationship between the vehicle and the defect center, outputting the sub-pixel-level lateral deviation of the vehicle relative to the defect center. longitudinal deviation The positioning accuracy reaches ±0.01 pixels. The calculated position deviation is transformed into a vehicle micro-movement command, which is fed back to the autonomous driving unit in real time to control the vehicle to perform high-precision fine-tuning in the lateral and longitudinal directions. Ultimately, the vehicle parking accuracy is controlled within ±2mm, providing a precise position reference for subsequent paving and compaction operations.

[0084] Furthermore, the roof-mounted composite storage compartment serves as the vehicle-mounted support and protection mechanism for the dual drones. It features an integrated structural design fixed to the top of the vehicle chassis. The compartment is independently divided into a first drone storage area and a second drone storage area. Internally, it integrates a multi-stage shock absorption mechanism and an electromagnetic automatic locking mechanism. The shock absorption mechanism uses a composite buffer structure of rubber damping pads and metal springs, with adjustable stiffness in the vertical direction. This effectively attenuates vibrations and impacts generated during vehicle operation, preventing the drone's onboard sensors from becoming inaccurate or damaged due to long-term vibration. The automatic locking mechanism operates using electromagnetic adsorption, with a rated adsorption force of no less than 50N. When the drone accurately lands in the designated storage position, the locking mechanism is immediately energized to generate electromagnetic attraction, firmly adhering to the drone's landing gear. When the drone needs to take off for operation, the locking mechanism is de-energized and released, with a response time of no more than 50ms. Temperature and humidity sensors are also installed inside the storage compartment to monitor environmental parameters in real time. The overall protection level reaches IP67, making it suitable for complex outdoor working conditions.

[0085] Furthermore, the asphalt storage and heating unit adopts a combined temperature control mode of phase change thermal storage and microwave-assisted heating to provide a stable operating temperature for the asphalt mixture. The phase change thermal storage layer uses paraffin-based composite phase change thermal storage material with a phase change temperature range of 120℃ to 140℃ and a thermal storage density of not less than 500kJ / kg, enabling long-term heat storage and slow release, thus reducing heating energy consumption. The microwave-assisted heating layer operates at a frequency of 2450MHz with a continuously adjustable output power from 0kW to 5kW, enabling rapid heating of the asphalt mixture. The vehicle-machine collaborative central control unit dynamically adjusts the microwave output power based on a thermodynamic model. The power control formula is: In the formula This represents the real-time output power of the microwave. The specific heat capacity of asphalt mixture is taken as 1.2 kJ / (kg·℃) to 1.5 kJ / (kg·℃). The current quality of the asphalt material to be heated. To set the heating rate, values ​​ranging from 5℃ / min to 10℃ / min were used. The heat exchange coefficient, The effective heating surface area of ​​the material. This refers to the real-time temperature of the asphalt mixture. To maintain the optimal paving rheological temperature, the system collects real-time data on both the material temperature and the ambient temperature, and then uses a closed-loop adjustment mechanism to control the microwave output power. The temperature fluctuation range in the vicinity is no higher than ±2℃, ensuring that the asphalt mixture has good rheological properties and construction performance.

[0086] This detailed implementation utilizes a dual-UAV heterogeneous collaborative operation, employing a fully automated collaborative mode that first performs wide-area screening and then performs fine measurement. Combining the collaborative control logic of heterogeneous UAV deployment and unmanned intervention, it achieves comprehensive, rapid, and accurate perception and location of pavement defects, significantly improving the efficiency and accuracy of pavement defect detection, avoiding errors and safety hazards caused by manual intervention, effectively reducing operation and maintenance costs, significantly improving the intelligence and efficiency of pavement defect detection, and ensuring detection accuracy and operational safety.

[0087] Example 2 describes in detail the multimodal AI recognition of pavement defects based on ResNet-50, such as... Figure 5 As shown, specifically:

[0088] The ResNet-50 deep residual network used in this embodiment is a core defect identification network designed specifically for a high-precision asphalt pavement repair system with heterogeneous collaboration between dual unmanned aerial vehicles on a vehicle roof. It uniquely receives the feature vector obtained by registration and fusion of infrared thermal imaging features and lidar point cloud features. As network input, it is used to complete the deep regression, internal porosity regression and disease contour segmentation of road surface defects. The overall network depth is 50 layers, and it is constructed using a bottleneck residual structure. It solves the gradient vanishing and model degradation problems in the deep network training process through cross-layer identity mapping. It can achieve real-time inference above 50Hz on the vehicle embedded platform. It forms strict data interaction and logical coupling with the road surface defect multimodal AI recognition module and the sub-pixel level operation path planning module, and is the core algorithm support for the disease quantitative perception link of the entire system.

[0089] ResNet-50 deep residual network fuses feature vectors As input, this feature vector is derived from a two-dimensional infrared feature map. Extracted 128-dimensional features and 3D point cloud feature map The extracted 256-dimensional features are weighted and fused to obtain the result, which satisfies the following conditions. and After the input features are fed into the network, dimensionality reshaping and batch normalization operations are performed. The normalization operation follows the formula. ,in The characteristic mean, The characteristic standard deviation, To prevent tiny constants with zero denominators, normalization can unify the numerical distribution of multimodal features, improve network training stability and convergence speed. The preprocessed features are mapped to feature maps of a specified size and fed into the initial convolutional layer of the network for basic feature extraction.

[0090] The initial segment of the ResNet-50 deep residual network uses convolutional layers with 7×7 kernels. The convolution operation formula is as follows: ,in For convolution kernel weights, For the input feature map, For bias terms, The ReLU activation function is used, the stride of this convolutional layer is set to 2, and the number of output channels is 64. This allows for the initial spatial abstraction and size downsampling of multimodal fusion features. A 3×3 max-pooling layer is then connected, with pooling operations following... The feature map size is further compressed while retaining the strong response features of the disease area. The initial convolution and pooling layers work together to extract low-level features such as disease edges, temperature changes, and geometric changes, providing high signal-to-noise ratio basic features for subsequent deep residual modules.

[0091] The ResNet-50 deep residual network consists of four stacked residual modules. All residual modules employ a bottleneck structure, with each bottleneck residual block containing three convolutional layers: a 1×1 convolution, a 3×3 convolution, and a 1×1 convolution. The 1×1 convolution is used for dimensionality reduction and expansion in the channel dimension, reducing overall computational cost. The 3×3 convolution is used for core spatial feature extraction. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The core mathematical mapping relationship of the residual block is as follows: ,in The input features for the residual block are... This is the residual function obtained after three layers of convolution operations. As the final output feature of the residual block, this identity mapping structure allows the gradient to be directly transmitted along the direct connection channel, avoiding gradient vanishing in deep networks and ensuring that a 50-layer network can be trained stably and converge.

[0092] In the ResNet-50 deep residual network, the number of stacked residual modules and the number of output channels are strictly configured according to a customized approach. The first group of residual modules has 3 stacked layers and 256 output channels, used to extract shallow geometric and temperature distribution features of the disease. The second group of residual modules has 4 stacked layers and 512 output channels, used to extract mid-level structural features of the disease. The third group of residual modules has 6 stacked layers and 1024 output channels, used to extract deep abstract features of the disease. The fourth group of residual modules has 3 stacked layers and 2048 output channels, used to fuse all multimodal information and output high-dimensional disease features. The four groups of residual modules sequentially complete the feature extraction from shallow to deep, gradually deeply coupling infrared temperature information with point cloud geometric information, and outputting a 2048-dimensional high-dimensional disease feature vector.

[0093] After completing the computation of all residual modules, the ResNet-50 deep residual network is connected to a global average pooling layer. The pooling operation follows the formula... ,in and These represent the width and height of the feature map, respectively. Global average pooling compresses spatial features into a fixed-length vector, eliminating the influence of disease location and size on the parameter output. This ensures the network maintains stable recognition accuracy for diseases of different sizes and locations. The pooled feature vector is then fed into a two-branch fully connected regression layer, where it is used to represent the depth of the disease. With internal porosity Continuous numerical regression.

[0094] The ResNet-50 deep residual network's bi-branch fully connected regression layers employ an independent parallel structure, and the fully connected layer computation follows the formula... ,in For fully connected weights, For the input vector, For bias terms, As a linear activation function, the disease depth regression branch outputs continuous values ​​in millimeters with an output accuracy of ±0.1mm, covering the full depth range of pavement diseases; the internal porosity regression branch outputs continuous values ​​from 0 to 100% with an output accuracy of ±0.5%, which can accurately characterize the degree of porosity development inside the disease. The output results of both branches are directly transmitted to the sub-pixel level operation path planning module for paving path generation and asphalt filling volume calculation.

[0095] This embodiment describes in detail the ResNet-50 deep residual network with a semantic segmentation auxiliary branch. This branch shares residual features with the backbone network and outputs a disease contour mask through upsampling, convolution operations, and sub-pixel level edge extraction algorithms. The contour positioning error is ≤0.1 pixels, providing accurate geometric boundaries for operation path planning. The network as a whole adopts a multi-task joint loss function for end-to-end training. The loss function includes depth regression loss, porosity regression loss, and contour segmentation loss, which can simultaneously optimize multiple output indicators, making the network fully adaptable to the operation requirements of heterogeneous collaborative perception of dual UAVs on the roof and high-precision repair of asphalt pavement.

[0096] Example 3 describes in detail an intelligent unmanned asphalt pavement repair method using dual drones inspecting vehicle rooftops. Figure 6 As shown, specifically:

[0097] The system performs power-on initialization and full module self-test; the vehicle-machine collaborative central control unit completes self-test of algorithm module and communication link; the drone control submodule wakes up the first drone and completes pre-takeoff preparations.

[0098] After the first drone takes off, it conducts a wide-area road inspection along a preset zigzag route. The infrared thermal imager continuously collects road surface temperature data and transmits it back in real time. After preprocessing, the data is input into the road surface defect multimodal AI recognition module. The module performs anomaly detection on the temperature field data, marks areas with abnormal temperatures as suspected defects, and records the rough geographical coordinates.

[0099] The UAV control submodule schedules the second UAV to take off based on the coordinates of the suspected disease. After the second UAV arrives at the target area, it enters a stable hovering state and starts a high-precision lidar and telephoto camera to perform all-round three-dimensional scanning and high-definition imaging of the disease area. After preprocessing, feature registration and feature fusion, the raw perception data is input into the depth residual network to complete the accurate calculation of the disease depth, outline and internal porosity.

[0100] The calculation results are synchronously transmitted to the sub-pixel level operation path planning module. The module establishes a local coordinate system with the center of the vehicle's rear axle as the origin, converts the global coordinates of the defect into local coordinates, and generates a sub-pixel level paving path that meets the coverage requirements and smoothness of movement based on cubic B-spline curves. At the same time, it calculates the asphalt filling volume and loss compensation volume based on the defect area, design paving thickness, and internal void ratio, generates complete path and material parameters, and sends them to the repair operation dynamic closed-loop control module.

[0101] The repair operation dynamic closed-loop control module sends the operation position command to the autonomous driving unit. The vehicle autonomously drives to the vicinity of the damaged area. The visual auxiliary positioning submodule starts road feature matching and pose calculation, outputs sub-pixel level position deviation and controls the vehicle to complete precise fine-tuning and parking. After parking, the asphalt storage and heating unit heats the asphalt mixture to the optimal paving temperature. The automatic paving and compaction repair operation unit enters the operation ready state.

[0102] After the paving operation starts, the material rheology control submodule collects parameters such as paving speed, paving width, and material temperature in real time, calculates and adjusts the asphalt delivery flow rate in a closed loop according to the rheological formula to ensure that the paving thickness is uniform. The laser displacement sensor continuously collects road surface elevation data along the paving path, and the online compaction quality monitoring module calculates the paving smoothness in real time. When the smoothness exceeds the threshold, the six-degree-of-freedom robotic arm is automatically driven to perform attitude compensation, and the compaction wheel completes the compaction operation synchronously.

[0103] After the repair work is completed, the second drone takes off again to scan the repaired area for verification. The multimodal AI recognition module analyzes the verification data and detects key indicators such as flatness, density, and contour matching. If the indicators do not meet the set requirements, the secondary repair process is automatically triggered, the work path is replanned and supplementary repairs are performed. If all indicators meet the standards, the repair work is marked as completed. The second drone returns to the roof storage compartment and automatically charges. The system waits for the next repair instruction. The entire process realizes unmanned, high-precision, and closed-loop asphalt pavement repair.

[0104] This embodiment describes in detail how dual-UAV heterogeneous collaboration and fully unmanned control enable wide-area rapid screening, accurate parameter calculation, and closed-loop repair of pavement defects, significantly improving the efficiency and accuracy of defect repair, reducing manual intervention, ensuring stable repair quality, and achieving efficient, accurate, and automated treatment of asphalt pavement defects.

[0105] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A high-precision asphalt pavement repair system with heterogeneous collaboration between two unmanned aerial vehicles on a vehicle roof, comprising an intelligent unmanned asphalt pavement repair vehicle body, wherein the vehicle body is equipped with a vehicle chassis, an autonomous driving unit, an asphalt storage and heating unit, and an automatic paving and compaction repair operation unit, characterized in that, It also includes a dual-UAV heterogeneous mounted operation and maintenance unit fixedly installed on the top of the vehicle chassis, and a vehicle-machine collaborative central control unit; The dual-drone heterogeneous operation and maintenance unit includes a dual-mode inspection drone group, a vehicle-mounted automatic charging mechanism, and a drone control submodule. The dual-mode inspection drone group includes a first drone and a second drone. The drone control submodule is electrically connected to the vehicle-mounted automatic charging mechanism, the first drone, and the second drone, respectively. The vehicle-machine collaborative central control unit is communicatively connected to the autonomous driving unit, the automatic paving and compaction repair operation unit, and the drone control submodule. The vehicle-machine collaborative central control unit has a built-in road surface defect multimodal AI recognition module, a sub-pixel level operation path planning module, and a repair operation dynamic closed-loop control module. The first drone is equipped with a wide-angle visible light camera and an infrared thermal imager, and is used to perform wide-area road inspections and quickly screen suspected defect areas; The second drone is equipped with a high-precision lidar and a telephoto camera, and is used to perform close-range, high-precision three-dimensional scanning and imaging of suspected disease areas screened by the first drone. The road surface defect multimodal AI recognition module is used to fuse the infrared thermal imaging data of the first UAV and the point cloud data of the second UAV to identify the type, outline, depth and internal porosity of road surface defects. The subpixel-level operation path planning module is used to generate a repair operation path with subpixel-level accuracy based on the identified defect location coordinates and the real-time vehicle pose, and output it to the repair operation dynamic closed-loop control module. The repair operation dynamic closed-loop control module is used to drive the autonomous driving unit to park the vehicle at the work position, and to control the asphalt storage and heating unit and the automatic paving and compaction repair operation unit to complete the repair operation, and to adjust the paving parameters in real time during the operation.

2. The system according to claim 1, characterized in that, The road surface defect multimodal AI recognition module includes a feature fusion submodule, which is used to convert the two-dimensional infrared feature map output by the first UAV into a feature fusion model. The three-dimensional point cloud feature map output by the second UAV Registration and fusion are performed; the registration process is achieved by solving the following nonlinear optimization problem, as shown in the formula: ,in, Let be a rotation matrix. It is a translation vector. For feature points in a 3D point cloud, These are the corresponding feature points in the two-dimensional infrared feature map. For camera projection model, The number of matching feature point pairs; fused feature vector Used as input to a deep residual network, outputting the depth of the disease. and the internal porosity of the diseased area .

3. The system according to claim 2, characterized in that, The sub-pixel-level job path planning module is specifically used for: Based on the identified disease outline and location coordinates, a local coordinate system is established with the center of the vehicle's rear axle as the origin. Within the local coordinate system, a smooth paving path is generated based on B-spline curves. This ensures that the movement trajectory of the paving head can completely cover the diseased area and that the path curvature is continuous, wherein: ,in, To control the vertices, For p-th degree B-spline basis functions; Meanwhile, based on the internal porosity Calculate the required asphalt mixture fill volume : ,in, The area affected by the disease. To design the paving thickness, is the pore filling coefficient.

4. The system according to claim 1, characterized in that, The repair operation dynamic closed-loop control module includes a material rheology control submodule; the material rheology control submodule is used to control the material rheology based on the real-time paving speed of the automatic paving and compaction repair operation unit. and paving width Dynamically adjust the conveying flow rate of asphalt mixture To ensure the uniformity of the paving thickness, the flow rate... The regulation follows the following rheological formula: ,in, For the target thickness, The density of the asphalt mixture, This is the rheological correction factor based on temperature compensation.

5. The system according to claim 1, characterized in that, The automated paving, compaction, and repair unit includes a six-degree-of-freedom robotic arm and a repair execution head located at the end of the robotic arm; The repair execution head integrates an infrared thermometer, a laser displacement sensor, and a compaction wheel; The vehicle-machine collaborative central control unit also has a built-in online compaction quality monitoring module, which is used to calculate the paving smoothness in real time based on the road surface elevation data returned by the laser displacement sensor. and in When the preset threshold is exceeded, dynamic compensation is performed by adjusting the posture of the six-degree-of-freedom robotic arm.

6. The system according to claim 1, characterized in that, The collaborative control logic between the first and second UAVs is as follows: The first UAV performs wide-area scanning along a zigzag route while the vehicle is moving. When a suspected disease is detected, it locks the approximate geographical coordinates of the disease. ; The drone control submodule controls the second drone to take off from the storage mechanism and fly to the approximate geographical coordinates. Hover over; The second UAV activates a high-precision lidar to scan the target defect from all angles, generates a three-dimensional model, and transmits the model back to the vehicle-machine collaborative central control unit for precise identification.

7. The system according to claim 1, characterized in that, The vehicle-machine collaborative central control unit also includes a vision-assisted positioning submodule; the vision-assisted positioning submodule is used to control the on-board high-definition camera to capture road surface texture features when the vehicle approaches the target operation position, and match them with feature points in the high-precision three-dimensional model transmitted back by the second UAV to calculate the sub-pixel level deviation of the vehicle relative to the center of the defect. The deviation is then fed back to the autonomous driving unit for fine-tuning.

8. The system according to claim 1, characterized in that, The dual-drone heterogeneous operation and maintenance unit also includes a composite storage compartment mounted on the roof of the vehicle. The composite storage compartment is equipped with a shock absorption mechanism and an automatic locking mechanism. When the first drone or the second drone lands, the automatic locking mechanism fixes the drone landing gear by electromagnetic adsorption to prevent vibration during vehicle operation from damaging the drone.

9. The system according to claim 1, characterized in that, The asphalt storage and heating unit includes a phase change heat storage layer and a microwave-assisted heating layer; the microwave-assisted heating layer is used to rapidly heat the local asphalt mixture before repair work, so that it reaches the optimal paving rheological temperature. The vehicle-mounted central control unit adjusts according to the ambient temperature. and the current temperature of the asphalt mixture Dynamic control of microwave power : ,in, For specific heat capacity, For quality, The heat exchange coefficient, is the surface area.

10. A high-precision asphalt pavement repair method using a dual-UAV heterogeneous collaborative system on a vehicle roof, applied to the system as described in any one of claims 1 to 9, characterized in that... Includes the following steps: The first drone was launched to conduct a wide-area inspection, using infrared thermal imaging to quickly detect areas of abnormal road surface temperature and mark them as suspected defects; The second drone was controlled to perform a high-precision 3D scan of the suspected disease area to obtain data on the depth, outline, and internal porosity of the disease. Based on the damage data, B-spline curves are used to plan sub-pixel level paving paths and calculate the required asphalt filling volume; Drive the vehicle to the work position, use visual-assisted positioning to perform sub-pixel-level correction, and start the automatic paving, compaction and repair work unit after stopping; During the paving process, the material flow rate is dynamically adjusted according to the rheological formula, and the paving smoothness is monitored in real time using a laser displacement sensor. After the repair is completed, the second drone is controlled to scan the area again for verification. If the flatness or density is found to be substandard, the secondary repair process is triggered.

Images