Fan blade non-stop intelligent inspection method based on unmanned aerial vehicle autonomous cruise

Abstract

The invention belongs to the technical field of new energy equipment operation and maintenance, and relates to a fan blade non-stop intelligent inspection method based on unmanned aerial vehicle autonomous cruise. In order to solve the problems of low manual inspection efficiency, power generation loss caused by shutdown of a traditional unmanned aerial vehicle and low fusion degree of a preliminary intelligent technology, the method configures hardware and a support environment to generate a flight path, the unmanned aerial vehicle accurately cruises and collects data by means of a flight control and fusion algorithm, damage is detected by using a lightweight YOLOv11 model after preprocessing, and the unmanned aerial vehicle is subjected to intelligent inspection. Visualization and life prediction are realized in combination with digital twinning, modeling is carried out only by first shutdown, and follow-up inspection without shutdown is carried out. The method can be used for blade inspection of land and offshore wind power plants and can also be expanded to the field of inspection of infrastructures such as photovoltaic power stations and power transmission lines.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy equipment operation and maintenance, and relates to unmanned aerial vehicle autonomous control, multi-modal sensor fusion, deep learning detection and digital twin technology, in particular to a fan blade non-stop intelligent inspection system based on unmanned aerial vehicle autonomous cruise, which is suitable for land and offshore wind farm blade damage detection, life prediction and operation and maintenance management, and can be extended to photovoltaic power station, power transmission line and other infrastructure inspection fields. BACKGROUND

[0002] In the field of fan blade inspection, although the current mainstream application technical solution has replaced the traditional manual operation to a certain extent, it still has limitations that are difficult to overcome and cannot meet the needs of non-stop, high-precision and intelligent operation and maintenance of wind farms, which are as follows:

[0003] Although the manual assisted inspection technology has low technical threshold, low initial investment cost, does not require complex algorithm debugging or equipment adaptation, and has low technical ability requirement for the operator, it can be directly completed by manual visual inspection combined with simple tools, but the inspection efficiency of this technology is low, 2-3 personnel are required for a single fan, and the time consumption for a single inspection is 4-6 hours, which is difficult to meet the high-frequency inspection needs of large-scale wind farms; moreover, the detection accuracy is seriously dependent on the experience of personnel, the identification accuracy rate of micro-cracks with a length less than 5mm and early corrosion with a depth less than 0.5mm is low, the missed detection rate is high, and the key parameters such as the depth and volume of damage cannot be quantified, only the qualitative judgment of "whether there is damage" can be made; at the same time, the operation and maintenance personnel need to climb to high altitude, which has a falling risk and is seriously affected by the environment.

[0004] Although the traditional unmanned aerial vehicle non-stop inspection technology greatly improves the efficiency compared with manual inspection, the time consumption for a single fan inspection is shortened to dozens of minutes, and the time consumption for 30 fan inspections is only a few hours, and it can cover the tip and back areas of the blade that are difficult for manual inspection to reach, and avoid the safety risk of high-altitude operation of personnel, some solutions can also automatically plan the flight path through the ground station software or define the body coordinate system based on the fan modeling to optimize the flight path, and reduce the limitation on the blade angle, but the core limitation of this technology is that it depends on the fan shutdown, and each inspection needs to be stopped for 2-4 hours, the annual loss of power generation of a single unit due to shutdown is tens of thousands of degrees, and frequent start and stop will cause the fan parts to bear additional mechanical stress and current impact, which will increase the risk of failure and even cause the tower to collapse; in addition, this technology relies on a single visual sensor, which is greatly affected by environmental interference such as backlight, rain and fog, has a high false positive rate, can only identify surface damage, cannot detect dynamic hidden dangers such as size deformation and resonance of the blade during operation, and the detection result is mostly isolated images and qualitative judgment, which lacks quantitative data support for maintenance priority decision-making.

[0005] Although the preliminary intelligent inspection technology attempts to introduce multiple sensors (such as cameras and laser radars) or digital twin technology, compared with single visual detection, it improves the adaptability to complex environments, and the application of laser radar can reduce the misjudgment caused by light interference. Some schemes also build a static model of the fan to assist in inspection, but this technology has not achieved deep integration of multiple technologies. Sensor data has spatial and temporal dislocation, time synchronization error, and spatial registration error, which cannot form the dual-mode collaborative detection capability of "point cloud + image", and the damage quantification precision is insufficient. The digital twin function also only stays at the level of static geometric model, and cannot reflect the damage development trend in real time, which lags behind the operation decision. Moreover, most schemes still need the fan to reduce the speed or stop to ensure data acquisition quality, and have not broken through the bottleneck problem of "stop inspection", and cannot realize real sense of non-stop intelligent inspection. SUMMARY

[0006] In view of the low efficiency of artificial auxiliary inspection, the large loss of traditional unmanned aerial vehicle stop inspection, the low integration degree of preliminary intelligent inspection technology and the problem of not breaking through the stop bottleneck, the present application provides a fan blade non-stop intelligent inspection method based on unmanned aerial vehicle autonomous cruise, which integrates unmanned aerial vehicle autonomous flight control, dual-mode damage detection and digital twin operation and maintenance technology, realizes high-precision and full-process automatic inspection of fan blades under non-stop state, and the specific technical scheme is as follows:

[0007] An EKF attitude estimation algorithm with an extended 11-dimensional state vector is adopted, on the basis of traditional attitude angle and angular velocity, accelerometer bias and horizontal wind disturbance velocity parameters are added, the sensor zero drift problem caused by fan electromagnetic interference is solved; the blade rotation induced wind speed is calculated combined with the blade element momentum theory, and the gust disturbance is compensated in real time, so that the attitude angle fluctuation of the unmanned aerial vehicle in the gust environment is controlled within ±1.2°, and the trajectory tracking error is ≤0.5m; at the same time, an adaptive trajectory correction function is designed, when the unmanned aerial vehicle deviates from the preset route, a smooth correction route is automatically generated to ensure a stable and safe distance from the blade.

[0008] Based on the fan model, a "surrounding + following" composite route is generated by MissionPlanner, the unmanned aerial vehicle first surrounds the fan at a fixed height to obtain the overall profile, and then completes the full surface coverage scanning along the blade rotation trajectory; the flight parameters are preset according to the fan specifications, such as the laser radar point cloud density of 6.5MW fan ≥80 points / cm², the camera slicing interval is 3s, and when the abnormal area triggers the close-up mode, the slicing interval is shortened to 0.5s.

[0009] During the UAV flight, the laser radar and camera are time-synchronized and spatially registered through the ROS Noetic system. The laser radar collects real-time three-dimensional point cloud data of the blade, and the camera captures surface texture images. The data is stored in dual mode of "real-time transmission + local caching". The ground station receives real-time data through the data transmission module, and the local SD card is used for backup to avoid data loss due to network interruption.

[0010] In the point cloud preprocessing stage, first, the isolated points with a neighborhood point number less than μ-2σ (μ is the average number of neighborhood points, and σ is the standard deviation) are deleted through statistical filtering to remove dust interference. Then, the radius filtering is used to remove bird droppings and other impurities attached to the blade surface. Subsequently, the region growing segmentation algorithm is used to separate the blade from the tower and the sky background. Finally, the data density is unified through 1mm×1mm×1mm voxel resampling, and the pure blade point cloud is output.

[0011] In the image preprocessing stage, for the motion blur of the collected images, adaptive Wiener filtering is used in the frequency domain to restore clear images and improve the peak signal-to-noise ratio. For uneven lighting, the Retinex algorithm is used to separate the illumination and reflection components, and only the illumination component is subjected to histogram equalization to preserve the true texture of the blade. For rain, fog, and dust noise, guided filtering is used to remove noise while preserving cracks and corrosion boundaries, maintaining good edge clarity.

[0012] A lightweight YOLOv11 model is deployed on the Jetson Nano edge. The Neck layer of the model is embedded with a damage depth perception unit, and a "classification loss + depth regression loss" dual loss function is introduced. The model accurately identifies the damage type (cracks, corrosion, coating shedding, etc.) by inputting the local image region mapped by the laser radar, and quantifies the key parameters. The crack length is calculated by the pixel distance of the bounding box diagonal and the pixel equivalent, the corrosion depth is estimated by the gray difference-depth mapping model, and the coating shedding area is calculated by the number of mask pixels and the pixel equivalent square. Subsequently, a scoring model is constructed based on "damage type-size-depth-location", and the score is used to divide the level. Different levels correspond to different operation and maintenance recommendations, and automatic labeling and pushing reminders are provided.

[0013] During the first inspection, the fan is stopped for 1-2 hours, and the laser radar scans the fan in all dimensions to obtain high-density point cloud data of the tower, nacelle, and blades. Combined with parameterized modeling, a high-precision static model is generated to lay the foundation for subsequent dynamic modeling. The real-time operating parameters of the integrated fan SCADA system are used to control the static model to simulate blade rotation and nacelle yaw through WebSocket technology. The finite element analysis results of material mechanics are introduced to add blade deformation compensation at different wind speeds in the model, making the dynamic state of the twin body closer to the physical fan.

[0014] After the subsequent non-stop inspection, the ICP registration algorithm is used to compare the new and old point cloud data, only the damage area with large deviation is updated, the full model reconstruction is avoided, and the update time is shortened to 30 minutes; the platform develops three-dimensional display, data panel, life prediction function, supports damage visualization marking, historical data tracing and residual life assessment, automatically generates inspection report and operation and maintenance work order, realizes "detection-decision-maintenance" closed-loop management.

[0015] Overall, the above technical solutions conceived by the present application mainly have the following beneficial effects:

[0016] Only the first shutdown modeling is required, and subsequent non-stop inspection is realized, greatly reducing the power generation loss and equipment start-stop loss caused by shutdown; unmanned aerial vehicle autonomous cruise combined with dual-mode damage detection technology not only improves the inspection efficiency and reduces the safety risk of manual work, but also realizes high-precision identification and quantification of damage, avoids the problems of missed detection and misdiagnosis in traditional inspection, and only qualitative and not quantitative; combined with the digital twin dynamic operation and maintenance platform, a "inspection-detection-evaluation-maintenance" closed loop can be quickly formed, promoting the transformation of wind farm operation from passive response to active prediction, providing support for wind farm cost reduction and efficiency improvement and ensuring safe and stable operation of equipment. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a system configuration and flight control flowchart of a wind turbine blade non-stop intelligent inspection method based on unmanned aerial vehicle autonomous cruise provided by the present application;

[0018] Figure 2 is an intelligent detection and data analysis flowchart of the method provided by the present application;

[0019] Figure 3 is a digital twin and operation and maintenance decision flowchart of the method provided by the present application;

[0020] Figure 4 is an effect diagram of real-time detection of wind turbine blades using lightweight YOLOv11;

[0021] Figure 5 is a 3D point cloud diagram of the whole wind turbine constructed by the unmanned aerial vehicle according to the planned path;

[0022] Figure 6 is an effect diagram of a wind turbine operation and maintenance system interface; DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0024] Specifically, the inspection method mainly includes the following steps:

[0025] Step one, before implementing the inspection task, first need to complete the hardware deployment and software configuration, the invention uses Pixhawk 6c flight control mainboard, running Ardupilot open source firmware, responsible for the attitude control and autonomous navigation of unmanned aerial vehicle, carries NVIDIA Jetson Nano edge computing platform, used for real-time data processing and model inference on site, integrates MID-360 laser radar and 4K SJCAM high-definition camera, used for synchronous acquisition of three-dimensional point cloud and surface texture image of blade, configures data radio to ensure long-distance data backhaul, and uses 6S high-capacity lithium battery to ensure single endurance time, deploys a high-performance notebook computer as a ground station at the inspection site, installs MissionPlanner for flight path planning and flight monitoring, and configures a dedicated image transmission receiving device to obtain real-time video stream.

[0026] Step two, parameter configuration, adopt "surround + follow" composite route, the unmanned aerial vehicle first surrounds the tower drum at a certain height to obtain the overall structure information, then enters the blade follow mode, sets the safety distance between the unmanned aerial vehicle and the leading edge of the blade to 3-5 meters, the laser radar scanning frequency is set to 10Hz, and the point cloud density is not less than 80 points / cm²; the default slice shooting interval of the camera is 3 seconds, when a suspected abnormality is detected, it automatically switches to a close-up mode of 0.5 seconds, sets the curvature mutation threshold to 0.05mm⁻¹, the height standard deviation threshold to 5μm, and the reflection intensity threshold to 500, used for screening suspected damage point cloud area, loads the pre-trained YOLOv11 lightweight model, and sets the confidence threshold to 0.7 to balance the detection accuracy and speed.

[0027] Step three, implement the inspection process. The operator selects the target wind turbine in the ground station software, the system automatically loads the preset wind turbine inspection route and generates a unique task ID, after the unmanned aerial vehicle is powered on, the system automatically completes the self-check of GPS, IMU, battery capacity and other states, after the self-check is passed, the ground station sends the take-off instruction, the unmanned aerial vehicle takes off autonomously to the preset height, when the unmanned aerial vehicle deviates from the preset route by more than 0.5 meters due to gust, the system immediately starts the adaptive trajectory correction function, predicts the future changes and generates a smooth correction path to ensure stable and safe distance, at the same time, the laser radar point cloud and camera image are transmitted in real time to the ground station through the data radio, and are backed up on the local SD card of the unmanned aerial vehicle to prevent data loss caused by network interruption.

[0028] Step four, data processing and damage detection, the method sequentially carries out statistical filtering, radius filtering, region growing segmentation and voxel resampling on the returned point cloud data, removes noise and background, and extracts pure leaf point cloud; the image collected synchronously is subjected to adaptive wiener filtering to remove blur, Retinex algorithm light equalization and guided filtering to remove noise, so as to improve the image quality, especially under backlight and complex light conditions, based on the pretreated point cloud, the algorithm quickly identifies the area with abnormal curvature, height and reflection intensity, generates a three-dimensional bounding box, and maps the coordinates to the corresponding image. The YOLOv11 model detects the image area marked by the laser radar, accurately identifies the damage type, the confidence value, and quantifies the damage area and maximum depth. According to the quantification result, the system combines the damage position, calculates the comprehensive score of the damage according to the preset scoring model, judges the risk level, and automatically marks.

[0029] Step five, the detection result is connected to the digital twin operation and maintenance platform, the system carries out ICP registration on the new point cloud obtained in this inspection and the historical point cloud, only the area with deviation exceeding a certain threshold is subjected to model updating, and the damage area is accurately marked on the three-dimensional digital twin of the fan. The operation and maintenance personnel can rotate, scale and section the model through the platform, and view the damage details from any angle. Clicking the label, the depth change trend of the damage in the last three inspections can be viewed, the platform predicts the remaining safe life of the blade based on the damage expansion rate and the material fatigue model, and displays the curve on the interface. At the same time, the system automatically generates a maintenance work order and pushes it to the relevant operation and maintenance person in charge.

Claims

1. A wind turbine blade non-stop intelligent inspection method based on unmanned aerial vehicle autonomous cruise, characterized in that, The method comprises the following steps: (1) First, configure the hardware environment and software support environment required by the method, determine the number of wind turbines to be inspected in the target wind farm, automatically generate or manually adjust the unmanned aerial vehicle flight path based on the wind turbine model, including loop type, spiral type, fixed point scanning mode, set flight height, camera resolution, point cloud density, sampling frequency and other parameters; (2) Start the unmanned aerial vehicle scheduling module, issue motion instructions through the unmanned aerial vehicle flight control communication protocol, and based on the open source flight control and Kalman filter + particle filter fusion algorithm, maintain accurate pose estimation in strong wind and complex airflow environment, simultaneously track the motion trajectory of the wind turbine blade in operation, use the laser radar with non-repeating scanning technology and 4K visible light camera to collect point cloud and image data, and realize sensor time synchronization and space calibration through the ROS framework; (3) Preprocess the point cloud, and extract pure point cloud from the laser radar point cloud through statistical filtering, radius filtering, region growing segmentation and voxel resampling; deploy a lightweight YOLOv11 model on a high-performance edge computing unit to perform real-time damage detection, first screen abnormal deformation areas by laser radar, and then classify local areas by YOLOv11 model in fine granularity, damage types include cracks, corrosion, peeling, wear, foreign matter attachment, etc., label damage positions and quantify area, maximum depth, volume and confidence score; (4) Upload the original data and detection results to the cloud through the Internet of Things communication protocol based on the publish / subscribe mode, store them in the relational database management system and full-text search engine, and realize blade motion state mapping and damage visualization in the dynamic digital twin body built in PyCharm; complete wind turbine complete modeling when stopping for the first time, and only scan new or changed areas through incremental update in the future, calculate the blade damage development trend and remaining service life, and generate a PDF format inspection report containing defect list, health score, maintenance suggestion, and automatically push the warning work order to the responsible person.

2. The unmanned aerial vehicle autonomous cruise based wind turbine blade no shutdown intelligent inspection method of claim 1, wherein: The unmanned aerial vehicle flight control adopts a "dynamic trajectory matching + adaptive attitude compensation" architecture, combined with the RTK-GNSS module and the IMU inertial navigation system, it can still maintain stable flight near the tower with signal shielding, and when detecting sudden changes in wind speed or signal loss, it can automatically trigger the obstacle avoidance mechanism and alarm; The damage detection adopts a "laser radar screening + deep learning detection" dual-mode mechanism to ensure the comprehensive recognition accuracy. 3.The unmanned aerial vehicle (UAV) -based autonomous cruise fan blade non-stop intelligent inspection method of claim 1, wherein: The digital twin visualization panel supports multi-view mode, including point cloud mode, image fusion mode, damage superposition mode and interactive operation, including zooming, rotating, slicing models, cross-section cutting, and can display the damage evolution process through the timeline, supports cross-terminal access, and all operation records can be saved as scene snapshots and traced back to specific flight time and shooting angle. 4.The unmanned aerial vehicle (UAV) -based autonomous cruise fan blade non-stop intelligent inspection method of claim 1, wherein: The patrol task management module supports batch creation of tasks, the task state includes to be started, to be executed, to be completed, and to fail, a timing task can be set, task data is permanently stored, search can be performed according to a project, time and fan number, data is stored in a private cloud or a local server, permission grading, operation trace and audit tracing are realized, and the power industry information security compliance requirements are met.

Images