Engineering facility detection system and method based on laser SLAM and multi-rotor unmanned aerial vehicle

Abstract

The invention discloses an engineering facility detection system and method based on laser SLAM and a multi-rotor unmanned aerial vehicle. A four-axis eight-rotor coaxial propeller reversing structure is adopted to improve the loading capacity and the wind resistance, base-station-free centimeter-level positioning is achieved through an FAST-LIO2 tight coupling algorithm, and disease recognition and three-dimensional space calibration are achieved based on the improved lightweight YOLOv5 network fusion key point detection technology. The problems that a traditional unmanned aerial vehicle loses efficacy in positioning in a closed space, the collision rate is high, and existing contact type detection efficiency is low are innovatively solved. The system comprises anti-collision unmanned aerial vehicle hardware, a dynamic point cloud filtering module and an embedded recognition module, the detection efficiency reaches 30 m / min in a GPS-free environment, and the collision damage rate is lt; 8%.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) application technology, and in particular to an engineering facility inspection system and method based on laser SLAM and multi-rotor UAVs for bridges, tunnels, high-rise buildings, and other engineering facilities. Specifically addressing the need for inspection in enclosed spaces without GPS, it combines collision-resistant UAV hardware design, laser SLAM positioning methods, and lightweight defect identification algorithms to achieve fully automated inspection with centimeter-level positioning accuracy. Background Technology

[0002] Regarding the current state of engineering facilities, a large number of engineering facilities worldwide are facing severe aging challenges, including over 1 million bridges and 500,000 high-rise buildings with aging risks and declining structural safety year by year. In China, this problem is equally prominent, with 30% of bridges exhibiting varying degrees of damage, including 100,000 deemed unsafe bridges, posing a serious threat to traffic safety. Simultaneously, infrastructure maintenance costs are rapidly increasing, with an average annual growth rate of 12%, while annual economic losses caused by engineering facility safety accidents exceed 5 billion yuan, making the need for engineering facility inspection and maintenance extremely urgent. Traditional inspection techniques have significant shortcomings. Manual inspection, a common method in the past, is extremely inefficient, requiring 3-5 days to inspect just one bridge, severely impacting the normal use of engineering facilities. More importantly, inspection personnel often need to work at heights, with a risk factor as high as 0.28, making accidents highly likely and posing a significant threat to their lives.

[0003] Ordinary drones also have many shortcomings in engineering facility inspection. Their payload capacity is limited, typically ≤2kg, making it impossible to carry specialized inspection equipment and meet the demands of high-precision testing. They also have weak wind resistance, potentially experiencing stability issues in winds of ≤5 levels, especially in turbulent environments such as under bridges, where the instability rate exceeds 40%, severely hindering the smooth progress of inspection operations. Furthermore, in environments with dense steel structures, the collision damage rate of ordinary drones exceeds 25%, not only causing equipment damage but also potentially affecting the completion of the inspection task.

[0004] Positioning technology also has limitations in engineering facility inspection. Visual positioning is easily affected by lighting conditions; in poor lighting environments such as cloudy days, the false detection rate increases by 35%, and the detection accuracy drops significantly. GPS / base station positioning methods are completely ineffective in enclosed environments, such as tunnels and under bridges, where signals are blocked, making accurate positioning impossible and hindering inspection work. Summary of the Invention

[0005] This invention addresses the pain points of existing engineering facility inspection technologies by constructing an inspection system that integrates collision-resistant UAV hardware, a high-precision positioning module, and a lightweight recognition algorithm. This system enables efficient and accurate inspection of engineering facilities in complex environments. The specific details are as follows:

[0006] An engineering facility testing system, comprising:

[0007] The multi-rotor drone has an energy-absorbing buffer structure at the end of its rotor arm, which is designed to absorb the impact energy of a collision through deformation.

[0008] The lidar-IMU tightly coupled positioning module, integrated into the UAV, includes lidar and IMU sensors, and generates point cloud maps and real-time pose data with centimeter-level accuracy through a tightly coupled SLAM algorithm;

[0009] A lightweight recognition module is mounted on the UAV and receives the real-time pose data. It runs an improved YOLOv5 network, removes the 9th and 12th convolutional layers in the backbone network, and adds an SE channel attention module after the 7th layer. The network parameter count is 2.1M. It outputs the disease type and the three-dimensional coordinates associated with the point cloud map space.

[0010] The real-time pose data drives the UAV flight path planning, and the three-dimensional coordinates are mapped to the point cloud map to form a closed-loop detection.

[0011] Preferably, the multi-rotor UAV is a quadcopter with an octocopter design, employing a coaxial counter-rotating rotor, and has a lift coefficient C. T and power coefficient C P :

[0012]

[0013] Therefore, the formula for calculating power load (PL) is:

[0014]

[0015] In the formula: T is the rotor lift, kg; P is the rotor power, W; A is the area of ​​the rotor disk, m². 2 Ω is the rotor speed, rad / s; R is the rotor radius, m; Q is the torque, N·m; ρ is the gas density, kg / m³ 3 C T C is the rotor thrust coefficient; P This represents the rotor's power coefficient.

[0016] Preferably, the tightly coupled SLAM algorithm is FAST-LIO2, the lidar is a 16-line scanning frequency of 20Hz, and the IMU sampling rate is 200Hz; the point cloud processing uses VoxelGrid filtering with a grid size of 0.05m.

[0017] Preferably, the energy-absorbing buffer structure is a double-layer frame, made of TPU and carbon fiber composite materials, with an angle of 18±2 degrees between the frame and the rotating arm, and a collision deformation threshold of ≥10mm.

[0018] This invention also provides a laser SLAM-based method for inspecting engineering facilities, comprising:

[0019] (1) The UAV performs manual flight to generate an initial three-dimensional point cloud model;

[0020] (2) Based on point cloud density coverage and flight efficiency, waypoint planning is optimized using point cloud data from the positioning module. The objective function is:

[0021]

[0022] Where C i E is a waypoint coverage metric. i As a flight efficiency indicator, the weights ω1 and ω2 are determined according to actual mission requirements or optimization algorithms. In this invention, ω1 = 0.7 and ω2 = 0.3 are used.

[0023] (3) The radar point cloud is downsampled by the VoxelGrid filtering algorithm of the positioning module;

[0024] (4) The positioning module fuses IMU data to execute the FAST-LIO2 algorithm and outputs a 6DOF pose.

[0025] (5) Image preprocessing: First, gamma correction is performed on the acquired images with a gamma value of γ = 1.2 to enhance the disease characteristics; then, according to the network requirements, the input dimension is adapted and normalized to obtain the network input data.

[0026] Feature extraction and detection: An improved YOLOv5 network is used, which removes the 9th and 12th convolutional layers of the backbone network and adds an SE channel attention module. Multi-scale features are extracted by combining the enhanced backbone network, and the features are fused by the Neck layer to enhance the expression. Based on the fused features, the detection head predicts the disease bounding box for accurate localization, disease category identification, and confidence level.

[0027] Post-processing output: Based on the crack identification confidence threshold, low-confidence predictions are filtered out, high-confidence detection boxes are retained, and information such as the location and type of the defects are output to complete the structural defect identification and detection.

[0028] (6) Map the disease identification results to the panoramic image. Coordinate transformation formula:

[0029] (X,Y)=(x×s x +Δx,y×s y +Δy)

[0030] Spatial calibration error ≤ 0.1m.

[0031] Preferred method: The hovering time at the waypoint in step (2) is calculated using the following formula: Among them, the number of shots per second S = 4, the number of shots per waypoint H = 8, the image processing time t1 = 1.2s, and the anti-disturbance buffer time t2 = 0.8s.

[0032] Preferred: The confidence threshold in step (5) is dynamically set according to the type of disease: confidence threshold for crack identification = 0.7; confidence threshold for corrosion identification = 0.65; confidence threshold for concrete spalling identification = 0.6.

[0033] Beneficial effects:

[0034] 1. The quadcopter octocopter UAV of the present invention adopts a coaxial counter-rotating rotor design. The rotor arm end is provided with a double-layer buffer frame, which is made of TPU and carbon fiber composite material. The frame and rotor arm are at a specific angle, which has strong anti-collision ability. When a collision occurs, it can generate a certain deformation to buffer the impact force and effectively reduce the risk of collision damage.

[0035] 2. The lidar-IMU tightly coupled positioning module of this invention integrates lidar and IMU (inertial measurement unit), and achieves real-time mapping and positioning through the FAST-LIO algorithm. It can output accurate pose information in the absence of GPS, providing centimeter-level positioning accuracy for UAVs. Simultaneously, point cloud filtering technology is used to process radar data and optimize point cloud quality.

[0036] 3. The lightweight recognition module of the present invention is equipped with an improved YOLOv5 network. By optimizing the network structure (such as removing some convolutional layers and adding attention modules), a lightweight design is achieved, which can efficiently perform disease recognition tasks and reduce the consumption of computing resources while ensuring recognition accuracy.

[0037] 4. This invention first generates an initial three-dimensional point cloud model through manual flight, and then optimizes waypoint planning based on point cloud density coverage and flight efficiency to balance detection coverage and flight efficiency, ensuring comprehensive and efficient detection.

[0038] 5. This invention performs downsampling processing on radar point clouds, fuses IMU data, and executes the FAST-LIO2 algorithm to output the UAV's pose information in real time, providing accurate spatial position reference for subsequent detection.

[0039] 6. Before disease identification, the present invention performs gamma correction on the image to improve image quality; the disease is identified by an improved YOLOv5 network, and the identification results are mapped to a panoramic image to achieve three-dimensional spatial calibration of the disease, ensuring that the positioning error is within a small range.

[0040] This invention is applicable to the inspection of engineering facilities such as bridges, tunnels, and high-rise buildings. It performs particularly well in enclosed environments or environments with weak GPS signals. It can efficiently identify defects such as cracks and corrosion, providing accurate data support for the maintenance of engineering facilities. Attached Figure Description

[0041] Figure 1 System architecture diagram;

[0042] Figure 2 Schematic diagram of a quadcopter with eight rotors;

[0043] Figure 3 Diagram of disease coordinate mapping;

[0044] Figure 4 : Schematic diagram of the improved YOLOv5 network architecture of this invention. Detailed Implementation

[0045] The technical solution of the present invention will be described in detail below with reference to specific embodiments and accompanying drawings:

[0046] Figure 1 This is a schematic diagram of the system architecture of the present invention, showing the collaborative relationship between the quadcopter octocopter UAV, the lidar-IMU positioning module, the lightweight identification module, and the ground station.

[0047] Figure 2 This is a schematic diagram of a quadcopter with eight rotors, highlighting the double-layer buffer frame (angle 18°±2°) at the end of the rotor arm and the coaxial counter-rotating rotor design.

[0048] Figure 3 This diagram illustrates the mapping of disease coordinates, revealing the transformation relationship from pixel coordinates (x, y) to panoramic coordinates (X, Y) (including scale factor s). x / s y (with offset Δx / Δy).

[0049] like Figure 4 As shown, the improved YOLOv5 network removes the convolutions in layers 9 and 12 of the backbone network and adds an SE channel attention module after layer 7, reducing the number of parameters to 2.1M.

[0050] This embodiment includes the following steps:

[0051] I. Collision-resistant UAV design

[0052] It adopts a quadcopter-eight-rotor structure with a rotor diameter of 35cm and an airfoil of NACA0012. Lift and stability are enhanced through a coaxial counter-rotating rotor design. The lift coefficient formula is... It has a maximum payload of 5kg and wind resistance of ≥7 levels. The rotor arm end is equipped with a double-layer buffer frame made of TPU and carbon fiber composite material. The frame and rotor arm have an angle of 18±2 degrees, and the collision deformation threshold is ≥10mm, which can effectively absorb collision energy and reduce the risk of damage. The shaft system is made of carbon fiber with a yield strength of ≥800MPa and is 40% lighter than aluminum shaft systems, balancing structural strength and lightweight requirements. It is also equipped with an anti-torque self-balancing mechanism with an angular velocity response time of <0.1s, ensuring flight stability.

[0053] II. LiDAR-IMU tightly coupled positioning module

[0054] Integrating a 16-line LiDAR and IMU (Inertial Measurement Unit), the LiDAR scans at a frequency of 20Hz, and the IMU samples at 200Hz. Real-time mapping and positioning are achieved through the FAST-LIO2 tightly coupled algorithm, with a computation latency of <20ms. It can output 6DOF (six degrees of freedom) pose and maintain centimeter-level positioning accuracy (±3cm) even in the absence of GPS. Point cloud processing employs the VoxelGrid filtering algorithm of PCL (Point Cloud Library), with a grid size of 0.05m. Voxel downsampling reduces data volume and improves processing efficiency while preserving key structural features, providing a precise 3D point cloud foundation for subsequent waypoint planning and defect localization.

[0055] III. Lightweight Recognition Module

[0056] The algorithm utilizes an improved YOLOv5 network as its core recognition algorithm, achieving a balance between lightweight design and high accuracy through network structure optimization: convolutional layers 9 and 12 in the backbone network are removed to reduce redundant computation; an SE channel attention module is added after layer 7 to enhance the extraction of disease features. The network has only 2.1M parameters, allowing for efficient operation on embedded platforms. Before disease identification, gamma correction (gamma value γ = 1.2) is applied to the image to improve contrast and facilitate feature extraction of diseases such as cracks and corrosion; the confidence threshold for crack identification is set to 0.7 to ensure accuracy. Combining a panoramic stitching algorithm (feature point matching error < 2 pixels) and a spatial calibration system (scale transformation matrix accuracy 0.01m), the disease identification results are transformed using the coordinate transformation formula (X,Y) = (x×s). x +Δx,y×s y +Δy) is mapped to the panoramic image to achieve three-dimensional spatial calibration of the disease, with an error ≤0.05m.

[0057] IV. System Autonomy Capability

[0058] It possesses full-process autonomous task execution capabilities, supports waypoint planning optimization based on an initial 3D point cloud model, and uses the point cloud / visual coverage integrity of a single waypoint i for the target detection area to represent the waypoint coverage index C. iThe operational efficiency E is reflected in the flight time and energy consumption of a single waypoint i. i The objective function is It can improve flight efficiency while ensuring coverage density. The system supports autonomous flight, autonomous defect detection and location, and autonomous data interaction and feedback. In the absence of GPS, the detection efficiency reaches 30m / min, and the collision damage rate is <8%, which greatly reduces the need for manual intervention and improves the practicality and intelligence of detection operations.

[0059] Applying the above embodiments to the inspection of steel box girder bridges specifically includes:

[0060] Initialization phase:

[0061] Operators manually control the drone to generate an initial 3D point cloud model of the bridge, achieving a point cloud density of 2000 points per square meter. The system then optimizes waypoint planning based on point cloud density coverage and flight efficiency, ensuring comprehensive coverage while maintaining high flight efficiency.

[0062] At the bridge pier, a drone was manually operated to fly at a height of 3 meters and a speed of 1 meter per second. A 3D point cloud model was generated using laser SLAM technology, achieving a point cloud density of 2000 points per square meter.

[0063] Automatic detection phase:

[0064] The drone flew along a preset route, maintaining a distance of 1.5 ± 0.1 meters from the bridge surface. During flight, the system dynamically adjusted the rotor speed to maintain stability. At each waypoint, the drone hovered for 2.5 seconds, taking 8 photos during this time. Simultaneously, the positioning module achieved real-time positioning using the FAST-LIO2 algorithm, outputting a frequency of 20Hz to provide the drone with accurate location information.

[0065] Disease identification:

[0066] First, gamma correction is applied to the captured images, with a gamma value set to 1.2. Then, the processed images are input into an improved YOLOv5 network. When identifying cracks, the confidence threshold is set to 0.7; when identifying corrosion, the confidence threshold is set to 0.65. Finally, the identified damage locations are mapped onto the panoramic image; for example, coordinates (8.75, -1.2) are mapped to (12.3, -0.7) on the panoramic image. Specifically:

[0067] Image preprocessing: The acquired images are first subjected to gamma correction with a gamma value of γ = 1.2 to enhance the disease features; then, according to the network requirements, the input dimensions are adapted and normalized to obtain the network input data.

[0068] Feature extraction and detection: An improved YOLOv5 network is used, which removes the 9th and 12th convolutional layers of the backbone network and adds an SE channel attention module. Multi-scale features are extracted by combining the enhanced backbone network, and the features are fused by the Neck layer to enhance the expression. Based on the fused features, the detection head predicts the disease bounding box for accurate localization, disease category identification, and confidence level.

[0069] Post-processing output: Based on the crack identification confidence threshold, low-confidence predictions are filtered out, high-confidence detection boxes are retained, and information such as the location and type of the defects are output to complete the structural defect identification and detection.

[0070] Experimental data:

[0071] Experimental data shows that the system can detect a 60m span bridge in just 25 minutes, with a crack identification rate of 98.3% and a positioning error of 0.04m, fully demonstrating its high efficiency and accuracy in bridge inspection.

[0072] Applying the above embodiments to tunnel detection specifically includes:

[0073] First, a 3D point cloud model of the tunnel is generated, with a vertical accuracy of 0.1m. Based on the tunnel's structural characteristics, the system sets a spiral flight path, with the UAV flying at a distance of 2.0m from the tunnel wall.

[0074] Considering the potential for high humidity inside the tunnel, when the humidity exceeds 85%, the anti-humidity lens on the drone can ensure the clarity of the captured images and avoid the impact of environmental factors on the detection results.

[0075] During the inspection, the drone autonomously identifies defects and transmits the results back to the ground station in real time, enabling staff to promptly grasp the defect situation inside the tunnel and providing timely and accurate data support for subsequent maintenance work.

Claims

1. An engineering facility inspection system based on laser SLAM and multi-rotor UAV, characterized in that, Include: The multi-rotor drone has an energy-absorbing buffer structure at the end of its rotor arm, which is designed to absorb the impact energy of a collision through deformation. The lidar-IMU tightly coupled positioning module, integrated into the UAV, includes lidar and IMU sensors, and generates point cloud maps and real-time pose data with centimeter-level accuracy through a tightly coupled SLAM algorithm; A lightweight recognition module is mounted on the UAV and receives the real-time pose data. It runs an improved YOLOv5 network, removes the 9th and 12th convolutional layers in the backbone network, and adds an SE channel attention module after the 7th layer. The network parameter count is 2.1M. It outputs the disease type and the three-dimensional coordinates associated with the point cloud map space. The real-time pose data drives the UAV flight path planning, and the three-dimensional coordinates are mapped to the point cloud map to form a closed-loop detection.

2. The system as described in claim 1, characterized in that: The multi-rotor UAV is a quadcopter with eight rotors, employing a coaxial counter-rotating rotor design, and the rotor lift coefficient C... T and power coefficient C P : The formula for calculating power load PL is: In the formula: T is the rotor lift; P is the rotor power; A is the area of ​​the rotor disk; Ω is the rotor speed; R is the rotor radius; Q is the torque; ρ is the gas density.

3. The system as described in claim 1, characterized in that: The tightly coupled SLAM algorithm is FAST-LIO2, the lidar is a 16-line scanning frequency of 20Hz, and the IMU sampling rate is 200Hz; the point cloud processing uses VoxelGrid filtering with a grid size of 0.05m.

4. The system as described in claim 1, characterized in that: The energy-absorbing buffer structure is a double-layer frame made of TPU and carbon fiber composite materials. The angle between the frame and the rotating arm is 18±2 degrees, and the collision deformation threshold is ≥10mm.

5. A method for inspecting engineering facilities based on laser SLAM, executed using the system described in any one of claims 1-4, characterized in that... include: (1) The initial three-dimensional point cloud model is generated by performing manual flight through the multi-rotor UAV; (2) Based on point cloud density coverage and flight efficiency, waypoint planning is optimized using point cloud data from the positioning module. The objective function is: Where C i E is a waypoint coverage metric. i ξ is the flight efficiency index, ω1 and ω2 are the weights, N is the total number of waypoints, and ξ is the set of waypoints. (3) The radar point cloud is downsampled by the VoxelGrid filtering algorithm of the positioning module; (4) The positioning module fuses IMU data to execute the FAST-LIO2 algorithm and outputs a 6DOF pose. (5) Perform gamma correction on the acquired images with a gamma value of γ = 1.2; run the improved YOLOv5 network through the lightweight recognition module to identify diseases, and filter the results according to the confidence threshold; (6) Map the disease identification results to the panoramic image. Coordinate transformation formula: (X,Y)=(x×s x +Δx,y×s y +Δy) Where (x,y) are the pixel coordinates of the disease on the original image, s x s is the horizontal scaling factor. y Δx is the vertical scale factor, Δy is the horizontal offset, and Δy is the vertical offset. The spatial calibration error is ≤0.1m.

6. The method as described in claim 5, characterized in that: The hovering time at the waypoint in step (2) is calculated using the following formula: Among them, the number of shots per second S = 4, the number of shots per waypoint H = 8, the image processing time t1 = 1.2s, and the anti-disturbance buffer time t2 = 0.8s.

7. The method as described in claim 5, characterized in that: The weights ω1 and ω2 are determined based on the actual task requirements or optimization algorithm, with ω1 = 0.7 and ω2 = 0.

3.

8. The method as described in claim 5, characterized in that: The confidence threshold mentioned in step (5) is dynamically set according to the type of disease: confidence threshold for crack identification = 0.7; confidence threshold for corrosion identification = 0.65; The confidence threshold for identifying concrete spalling is 0.6.

Images