CN116912715A - A UAV visual servo control method and system for wind turbine blade inspection - Google Patents

Abstract

The invention discloses an unmanned aerial vehicle vision servo control method and system for fan blade inspection, wherein through the shooting function of an unmanned aerial vehicle, panoramic information of fan blades is shot at different angles, camera pose information during shooting is recorded, coordinates of key points of a fan are solved according to multiple groups of information, and path planning of unmanned aerial vehicle inspection starting points and directions for the fan blades is completed; and finally, performing visual servo by using a reinforcement learning method, enabling the unmanned aerial vehicle to be close to the blade, attaching the curved surface of the blade to fly and photographing. According to the method, under the condition that the length and direction information of the fan blade is not required to be obtained, the positioning of key points of the fan is automatically completed, high-precision visual tracking is performed, and the problem of routing inspection route deviation caused by irregular shapes of the blade and short-time wind gusts is solved.

Description

A UAV visual servo control method and system for wind turbine blade inspection

Technical field

The invention belongs to the technical field of UAV inspection, and mainly relates to a UAV visual servo control method and system for inspection of wind turbine blades.

Background technique

As a renewable energy source with mature technology, wind power has been widely used around the world. However, the surface layer of wind turbine blades is prone to wear and tear, leading to damage such as trachoma and scratches. Through automatic inspection by drones, the details of the blade surface can be captured and analyzed at close range. The detection of early blade defects can effectively reduce later maintenance costs. It is of great significance to design a drone inspection method for wind turbine blades that is autonomous, stable, adaptable and economical.

Research based on the autonomous completion of tasks by UAVs has also achieved many results. Many autonomous inspection solutions have been proposed. Some use positioning systems to complete position control based on the appearance and prior posture data of the wind turbine. Some are based on a variety of three-dimensional sensors. For example: millimeter-wave radar, lidar, ultrasonic sensors, multi-camera, etc. perform inspections after completing local three-dimensional reconstruction, and direct inspections after linear fitting of the attitude of the wind turbine blades. Most of these methods rely on priori data and additional sensors, however, wind turbines are usually installed in areas with harsh environments, which poses a great challenge to flight stability. At the same time, these additional sensors are often expensive and difficult to maintain.

In order to improve the adaptability of UAVs in complex environments, reduce maintenance costs and improve reliability, it is of great significance to use machine vision methods to design UAV inspection methods for wind turbine blades that are autonomous, stable, adaptable and economical. significance.

Contents of the invention

The present invention is aimed at the existing technology of automatic inspection of wind turbine blades by drones, which relies on a priori information and has a low degree of automation. The drone lacks adjustment during inspection of wind turbine blades, resulting in low inspection stability and poor inspection images. To solve the problem of low quality, a drone visual servo control method and system for wind turbine blade inspection is provided. Through the drone's shooting function, the panoramic information of the wind turbine blades is captured at different angles and the camera pose information is recorded during shooting. , solve the coordinates of key points of the wind turbine based on multiple sets of information, and complete the path planning of the starting and ending points and directions of the UAV inspection facing the wind turbine blades; finally, use the reinforcement learning method to perform visual servoing to make the UAV close to the blades and fit the blades Fly over curved surfaces and take photos. The method of the present invention can independently complete the positioning of key points of the wind turbine without obtaining information on the length and direction of the wind turbine blades, perform high-precision visual tracking, and correct the inspection route deviation problem caused by irregular blade shapes and short-term gusts.

In order to achieve the above objectives, the technical solution adopted by the present invention is: a UAV visual servo control method for wind turbine blade inspection, including the following steps:

S1: Obtain location information: Use the shooting function of the drone to shoot the panoramic information of the wind turbine blades at different angles and record the camera pose information during shooting. The panoramic information and camera pose information include at least three groups, which were shot in Different positions of the wind turbine blades, and each position includes the entire blade of the wind turbine;

S2: Solve the coordinates of the key points of the wind turbine: The coordinates of the key points of the wind turbine are obtained according to the multiple sets of panoramic information and camera pose information obtained in step S1. The key points of the wind turbine are the tip points of the wind turbine blades; the straight line Li of the key point is The corresponding straight line equation in space is:

Among them, t is the parameter, x _i is the x value of the three-dimensional direction vector of the straight line Li, Xi is the X value of the camera coordinates when the camera acquires the image, yi is the y value of the three-dimensional direction vector of the straight line Li, and Yi is the camera The Y value of the coordinate, z_i is the z value of the three-dimensional direction vector of the straight line Li, Zi is the Z value of the camera coordinate when the camera acquires the image;

S3: Path planning: According to the coordinates of the key points of the wind turbine obtained in step S2, complete the path planning of the starting and ending points and directions of the drone inspection facing the wind turbine blades; the requirements of the path planning are to ensure that the largest local image of the wind turbine blades is in the camera The area occupied in the field of view is between 50% and 75%;

S4: Trajectory implementation: According to the path obtained in step S3, use the reinforcement learning method to perform visual servoing, so that the drone is close to the blade, flies along the blade curved surface and takes pictures.

As an improvement of the present invention, step S2 specifically includes:

S21: Use the semantic segmentation model to separate the pixels belonging to the wind turbine blades in the image from the panoramic information, and obtain the panoramic segmentation map of the wind turbine blades;

S22: According to the panoramic segmentation image obtained in step S21, extract the pixel coordinates of the points on the three blade tips of the wind turbine as the pixel coordinates of the key points;

S23: According to the pixel coordinate values extracted in step S22, the camera pose when taking the image, and the camera parameters, solve the coordinates of the key points of the wind turbine blades in space, specifically: plane S ₁ is point C when the camera shoots at point A. The image plane where it is located, point B on the plane is the center of the image captured by the camera, that is, the intersection point of the optical axis of the camera at A and the plane S ₁ , the pixel coordinate of point B in the image is OK,/> Column, the direction of the camera is the direction of vector AB;

Plane S ₂ is the image plane where point C is located when the camera shoots at point D. Point C is on the intersection of plane S ₁ and plane S _2. The plane pixel coordinates of point C on plane S1 are row a1 and b1. column, then the pixel distance from the camera center point to the image plane can be obtained in the pixel coordinate system as shown in the following formula,

Among them, θ is the angle of the camera's FOV angle; with camera A as the origin, the vector is the direction vector of straight line L1, vector/> The value of is specifically:

Among them, a ₁ is the row number of point C on plane S1, b ₁ is the column number of point C on plane S1, and W is the heading angle of the camera in the local space coordinate system o-xyz;

Since it is known that the straight line passes through point A and the coordinates of point A are (X ₀ , Y ₀ , Z ₀ ), the parametric equation of the straight line Li can be obtained as follows:

Among them, t is a parameter, _Li is the straight line equation corresponding to a key point of the wind turbine in space, and the intersection of multiple space straight lines is the spatial position of the key point.

As an improvement of the present invention, step S3 specifically includes:

S31: Based on the three-dimensional coordinate information of the key points, obtain the direction of the wind turbine and the coordinates of the center point of the wind turbine blade;

S32: According to the fan orientation and fan blade center point coordinates obtained in step S31, within a fixed distance from the front and back of the fan blade, determine the starting point and end point of the visual servo and provide the inspection direction; the fixed distance is determined by the fan blade size The size is determined by a combination of the camera's FOV angle.

As another improvement of the present invention, the method for solving the direction of the wind turbine in step S31 is specifically: take the three-dimensional coordinates of two of the three blades of the wind turbine and make a difference to obtain a vector, and then calculate the three-dimensional coordinates of the other two blades. Difference, get another vector, take the outer product of the two vectors, and get the direction vector of the plane, that is, the direction of the fan;

The solution of the center point coordinates of the wind turbine blades is specifically: averaging the three-dimensional coordinates of the three blades of the wind turbine.

As another improvement of the present invention, step S4 specifically includes:

S41: Input the captured image into the image segmentation algorithm to separate the leaf image and background image;

S42: After data compression of the segmented image, it is used as a state input to the reinforcement learning model for training. The output action of the reinforcement learning model is a fixed positive vector value perpendicular to the inspection direction, performs correction actions, and calculates the reward function. ;

S43: Update the reinforcement learning parameters and optimize the reinforcement learning model according to the reward value, status and action; when the reward value reaches the threshold, the training is completed, and the trained modified model is used for visual servoing to complete the trajectory implementation.

As a further improvement of the present invention, the reward function in step S42 is specifically:

w＝-(d ² +k·s)

In the formula, w is the reward function value, d is the offset value in visual servoing, s is the number of servo steps, and k is the coefficient.

As a further improvement of the present invention, in step S1, the edge of the image captured by the drone camera is cropped by 5% before use, and the camera is calibrated using an image calibration card. The FOV value of the camera is after cropping and calibration after edited.

In order to achieve the above object, the present invention also adopts a technical solution: a UAV visual servo control system for wind turbine blade inspection, including a computer program. When the computer program is executed by a processor, any of the above-mentioned methods are implemented. Describe the steps of the method.

Compared with the existing technology, the present invention has beneficial effects: The present invention proposes a UAV visual servo control method and system for wind turbine blade inspection, which has a high degree of autonomy and is robust during the inspection process. The performance is good and the pictures obtained are of high quality. This invention has the ability to independently calculate the key points of the wind turbine blades when the length of the wind turbine blades, the blade orientation and the blade stopping angle are unknown a priori, and the expected inspection path of each blade is calculated through the key points, and at the same time based on the images captured by the camera , using the method of visual servoing to continuously correct the position of the drone during the inspection process. When the shape of the wind turbine blades changes greatly and encounters short-term gusts, the position of the drone can be corrected and high-quality wind turbine blades can be photographed. Picture, compared with the existing wind turbine drone automatic inspection method, it has a higher level of autonomy, applicability and robustness, and has high engineering value.

Description of the drawings

Figure 1 is a step flow chart of the method of the present invention;

Figure 2 is a schematic diagram of spatial key point positioning based on a monocular camera in step S2 of the present invention;

Figure 3 is a schematic diagram of visual servoing during the inspection process of the method of the present invention.

Detailed ways

The present invention will be further clarified below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

Example 1

A UAV visual servo control method for wind turbine blade inspection, as shown in Figure 1, includes the following steps:

S1: Acquisition of position information: The drone flies to no less than three different positions in front of the wind turbine blades, captures panoramic information of the wind turbine blades at different angles, and records the camera pose information during shooting.

S2: Calculating the coordinates of key points of the wind turbine: Calculate the coordinates of the three tip points of the wind turbine blades as key points based on the solution algorithm of taking no less than three pictures and the camera pose information during the shooting; the steps specifically include:

S21: Use the semantic segmentation model to separate the pixels belonging to the wind turbine blades in the image from the background, and obtain a panoramic segmentation image of the wind turbine blades;

S22: Based on the panoramic segmentation map and combined with the obvious external features of the wind turbine, extract the pixel coordinates of the points on the three blade tips of the wind turbine as the pixel coordinates of the key points;

S23: Based on the extracted pixel coordinate values, the camera pose when taking the image, and the camera parameters, solve the coordinates of the key points of the wind turbine blades in space. In the local space coordinate system o-xyz, point A (X ₀ , Y ₀ , Z ₀ ) and point D are the camera shooting positions in space. When the camera resolution is row A and column B (unit is pixel, the same below), the angle of the camera's FOV angle is θ degrees, and the camera position is at point A (X ₀ , Y ₀ , Z ₀ ), the camera shoots horizontally, with The positive direction of the X-axis is 0 degrees, and the heading angle is W. Plane S ₁ is the image plane where point C is located when the camera shoots at point A. Point B on the plane is the center of the image captured by the camera, that is, the intersection of the optical axis of the camera at A and plane S _1. The pixel of point B in the image The coordinates are OK,/> List. Therefore, the orientation of the camera is the orientation of vector AB. Plane S ₂ is the image plane where point C is located when the camera shoots at point D. Point C is the target point that needs to be solved. Point C is on the intersection of plane S ₁ and plane S _2. Point C is on the plane on plane S1. The pixel coordinates are row a1 and column b1. In the pixel coordinate system, the pixel distance from the camera center point to the image plane can be obtained as follows:

Among them, with camera A as the origin, the vector is the direction vector of straight line L1, vector/> The value of is shown in the following formula, where W is the heading angle of the camera in the local spatial coordinate system o-xyz, and the remaining parameters are obtained according to the linear imaging model of the camera.

Since it is known that the straight line passes through point A and the coordinates of point A are (X ₀ , Y ₀ , Z ₀ ), the parametric equation of straight line L can be obtained as follows:

(t is a parameter)

L ₁ is the straight line equation corresponding to a key point of the wind turbine in space in one picture. Using the same method to solve the same key point in several other pictures, we can finally obtain no less than three non-parallel lines. The equation of the space straight line, the spatial position of the key point is the intersection point of these space straight lines.

S3: Path planning: Based on the coordinates of the three key points of the wind turbine, perform path planning of the starting and ending points and directions of the drone inspection of the blades; the steps include:

S31: Based on the three-dimensional coordinate information of the three key points, the direction of the wind turbine and the coordinates of the center point of the wind turbine blade can be obtained.

S32: Based on the coordinates of the fan center point and the fan tip coordinates, as well as the direction of the fan, determine the starting point and end point of the visual servo at a certain distance from the front and back sides of the three blades of the fan, and give the inspection direction. The distance is determined by the size of the wind blade and the FOV angle of the camera, ensuring that the largest partial image of the wind blade occupies between 50% and 75% of the area in the camera's field of view.

S4: Trajectory implementation: According to the path obtained in step S3, use the reinforcement learning method to perform visual servoing, so that the drone is close to the blade, flies along the blade curved surface and takes pictures. This step specifically includes:

S41: Input the captured image into the image segmentation algorithm to separate the leaf image from the background;

S42: After data compression of the segmented image, it is used as a state input to the reinforcement learning model for training. The output action of the reinforcement learning is a fixed positive vector value perpendicular to the inspection direction. The correction action is performed and the reward function is calculated. The reward function The design is as follows:

w＝-(d ² +k·s)

In the formula, w is the reward function value, d is the offset value in visual servoing, s is the number of servo steps, k is the coefficient, k·s is used to penalize the direction calculation incorrectly, the servo algorithm leads to the problem of too long flight time, the servo When the camera loses its target, the reinforcement learning round ends. In order to punish the decision-making of this kind of agent in this situation, the reward function is given an additional negative number with a large absolute value. The specific value is determined according to the specific parameters of the system.

S43: Update the reinforcement learning parameters according to the reward value, status and action to obtain a better control strategy; when the reward value reaches the threshold, the training is considered complete, and the trained modified model is used for visual servoing to reduce errors during the inspection process. , to achieve the purpose of flying according to the curved surface of the blade.

Example 2

A UAV visual servo control method for wind turbine blade inspection. The UAV flies to three different positions in front of the wind turbine blades, takes panoramic information of the wind turbine blades at different angles, and records the camera pose information during the shooting. The panoramic view It means that the entire blade of the wind turbine needs to be included in order to subsequently solve the three key points of the blade tip. Based on the three pictures taken and the camera pose information during the shooting, the solution algorithm calculates the coordinates of the three key points of the wind turbine blades; based on the coordinates of the three key points of the wind turbine, the start and end points and direction of the visual servo are planned; using reinforcement learning The method uses visual servoing to make the drone get close to the blades, fly according to the shape of the blades and take pictures. The method specifically includes the following steps:

S1: The drone flies to three different positions in front of the wind turbine blades, captures the panoramic information of the wind turbine blades at different angles, and records the camera pose information during shooting.

In this embodiment, the drone is equipped with RTK to obtain latitude and longitude information, and the orientation information is obtained through a 9-axis IMU. Through the conversion algorithm between the longitude and latitude high coordinate system and the northeast sky coordinate system, with the take-off point as the origin and the aircraft position at take-off as the positive direction, a northeast sky coordinate system is established. This coordinate system is the spatial local coordinate system in this example. The camera is connected to the drone through a two-axis gimbal, and the pose information of the camera is the pose information in this coordinate system.

In this embodiment, the height of the wind turbine is 80 meters, the blade length is 40 meters, and the FOV angle of the camera carried by the drone is 90 degrees. Based on the above information, the drone flew 45 meters in front of the wind turbine blades to capture images of the wind turbine blade options.

S2: Find the coordinates of the three key points of the wind turbine based on the solving algorithm of the three pictures taken and the camera pose information at the time of shooting.

The semantic segmentation algorithm is used to perform image segmentation on the wind turbine blade panorama. The entire image is classified into two categories. The pixel value of the wind blade part is set to 255, which is white, and the remaining pixel values are set to 0, which is black. In the above binary image, select the top white pixels and find their average coordinates as the first key point coordinates. Select the leftmost white pixels and find their average coordinates as the second key point coordinates. , select the rightmost white pixel point, find the mean of their coordinates, and use it as the third key point coordinate. According to the outline of the wind turbine, noise points are eliminated, and special circumstances such as judging whether it is at the edge of the image are eliminated to eliminate interference data of non-key points. ;

Based on the extracted pixel coordinate values, the camera pose when taking the image, and the camera parameters, the coordinates of the key points of the wind turbine blades in space are solved. As shown in Figure 2, in the local space coordinate system o-xyz, point A (X ₀ , Y ₀ , Z ₀ ) and point D are the two points in the space where the camera shooting position is located. When the camera resolution is row A and column B (unit is pixel, the same below), the angle of the camera's FOV angle is θ degrees, and the camera position is at point A (X ₀ , Y ₀ , Z ₀ ), the camera shoots horizontally, with The positive direction of the X-axis is 0 degrees, and the heading angle is W. Plane S ₁ is the image plane where point C is located when the camera shoots at point A. Point B on the plane is the center of the image captured by the camera, that is, the intersection of the optical axis of the camera at A and plane S _1. The pixel of point B in the image The coordinates are OK,/> List. Therefore, the orientation of the camera is the orientation of vector AB. Plane S ₂ is the image plane where point C is located when the camera shoots at point D. Point C is the target point that needs to be solved. Point C is on the intersection of plane S ₁ and plane S _2. Point C is on the plane on plane S1. The pixel coordinates are row a1 and column b1. In the pixel coordinate system, the pixel distance from the camera center point to the image plane can be obtained as shown in the following formula.

In this embodiment, when the camera resolution is 1024×1024 pixels, the camera’s FOV angle is 90 degrees. pixels.

With camera A as the origin, the vector is the direction vector of straight line L1, vector/> The value of is shown in the following formula, where W is the heading angle of the camera in the local spatial coordinate system o-xyz, and the remaining parameters are obtained according to the linear imaging model of the camera.

Since it is known that the straight line passes through point A, and the coordinates of point A are (X ₀ , Y ₀ , Z ₀ ), the parametric equation of straight line L can be obtained as follows. When i is 1, it is the equation of straight line L ₁ .

Among them, t is a parameter, and L ₁ is the straight line equation corresponding to a key point of the wind turbine in one picture in space. Use the same method to solve the same key point in the other two pictures, and finally obtain The equations of three non-parallel space straight lines, the spatial positions of the three key points of the wind turbine blade are the intersection points of the three space straight lines.

S3: Based on the coordinates of the key points of the wind turbine obtained from the solution, complete the path planning of the starting and ending points and direction of the drone inspection facing the wind turbine blades.

As shown in Figure 3, based on the three-dimensional coordinate information of the three key points, the orientation of the wind turbine and the coordinates of the center point of the wind turbine blade can be obtained.

In this example, the step of obtaining the orientation can number 3 points. For example, the key point of the upper blade of the wind turbine is point 1, the key point of the left blade is point 2, and the key point of the right blade is point 3. Take point 1. Difference the coordinates of point 2 to get a vector, take the difference of the coordinates of point 1 and point 3 to get another vector, take the outer product of the two vectors, and you can get the direction vector perpendicular to the plane formed by the three points, the center point To obtain it, just average the three points.

According to the coordinates of the fan center point and the fan tip coordinates, as well as the direction of the fan, the starting point and end point of the visual servo are determined at a certain distance from the front and back of the three blades of the fan, and the inspection direction is given. The distance from the fan blade is given by The size of the wind blade and the FOV angle of the camera are jointly determined to ensure that the largest partial image of the wind blade occupies between 50% and 75% of the area in the camera's field of view.

In this example, the camera resolution is 1024×1024 pixels, the camera’s FOV angle is 90 degrees, the blade length is 40 meters, and the distance between the drone and the wind turbine blades is set to 2.5 meters.

S4: Input the captured image into the image segmentation algorithm to separate the close-up image of the blade from the background;

The semantic segmentation algorithm is used to perform image segmentation on the wind turbine blade panorama. The entire image is classified into two categories. The pixel value of the wind blade part is set to 255, which is white, and the remaining pixel values are set to 0, which is black.

After data compression of the segmented image, it is used as a state input to the reinforcement learning model for training. The output action of the reinforcement learning is a one-dimensional positive vector value perpendicular to the inspection direction. The correction action is performed and the reward function is calculated. The reward function is designed as follows :

w＝-(d ² +k·s)

In the formula, w is the reward function value, d is the offset value in visual servoing, s is the number of servo steps, k is the coefficient, k·s is used to penalize the direction calculation incorrectly, the servo algorithm leads to the problem of too long flight time, the servo When the camera loses its target, the round of reinforcement learning ends. In order to punish the agent's decision-making in this situation, the reward function is given an additional negative number with a large absolute value. The specific value is determined by the specific parameters of the system.

In this example, a 1024×1024 pixel camera is used for inspection, and the penalty coefficient is set to -90000.

The reinforcement learning algorithm used in this example can be the DDPG algorithm. The input state variable is the compressed image information of several consecutive frames. The output correction variable is a vector perpendicular to the inspection path. The input image is set to four consecutive frames of images.

Update reinforcement learning parameters based on reward values, states, and actions to obtain better control strategies;

When the reward value reaches the threshold, the training is considered completed, and the trained correction model is used for visual servoing to reduce errors during the inspection process and achieve the purpose of fitting the blade shape.

Example 3

The difference between this embodiment and Embodiments 1 and 3 is that the UAV in this method can also be equipped with a UWB-based positioning device. Therefore, the establishment of the spatial local coordinate system can be based on the spatial local coordinate system of the UWV equipment positioning, without the need to convert the longitude and latitude coordinate system into the northeast sky coordinate system.

In addition, because the edge of the inspection camera mounted on the drone has large distortion, direct use will lead to positioning errors. Therefore, during the flight, the edge of the image captured by the camera was clipped with 5% distortion, and an image calibration card was used to calibrate the camera. In this example, the coordinates of the pixels in the image used for positioning are modified by the calibration model, and the FOV value of the camera is modified by clipping and calibration.

In summary, this method uses the sensors carried by drones and inspection cameras to determine the coordinate positions of key points of wind turbine blades without relying on a priori data such as the scale and shape of the wind turbine blades. Based on the determined key point coordinates, this method uses visual servo technology to correct the inspection trajectory during the inspection process along the blades, solving problems such as unstable imaging accuracy and lack of details of turbine blades in dynamic environments. In the feedback control link of visual servoing, this method uses reinforcement learning technology, which achieves better control performance and wider adaptability than traditional PID manual parameter adjustment. Compared with existing technical solutions, the present invention has higher autonomy and robustness.

It should be noted that the above content only illustrates the technical idea of the present invention and cannot limit the protection scope of the present invention. For those of ordinary skill in the technical field, without departing from the principle of the present invention, they can also make Several improvements and modifications are made, and these improvements and modifications fall within the protection scope of the claims of the present invention.

Claims (8)

1. A UAV visual servo control method for wind turbine blade inspection, which is characterized by including the following steps: S1: Obtain location information: Use the shooting function of the drone to shoot the panoramic information of the wind turbine blades at different angles and record the camera pose information during shooting. The panoramic information and camera pose information include at least three groups, which were shot in Different positions of the wind turbine blades, and each position includes the entire blade of the wind turbine; S2: Solve the coordinates of the key points of the wind turbine: Solve to obtain the coordinates of the key points of the wind turbine based on the multiple sets of panoramic information and camera pose information obtained in step S1. The key points of the wind turbine are the tip points of the wind turbine blades; the straight line L passing through the key points The straight line equation corresponding to _i in space is: Among them _, t is a parameter _{, x i} is the x value of the three _- dimensional direction vector of straight line _Li , When acquiring an image, the Y value of the camera coordinates, z _i is the z value of the three _- dimensional direction vector of the straight line Li, and Z _i is the Z value of the camera coordinates when the camera acquires an image; The intersection of multiple spatial straight lines is the spatial position of the key point; S3: Path planning: According to the coordinates of the key points of the wind turbine obtained in step S2, complete the path planning of the starting and ending points and directions of the drone inspection facing the wind turbine blades; the requirements of the path planning are to ensure that the largest local image of the wind turbine blades is in the camera The area occupied in the field of view is between 50% and 75%; S4: Trajectory implementation: According to the path obtained in step S3, use the reinforcement learning method to perform visual servoing, so that the drone is close to the blade, flies along the blade curved surface and takes pictures. 2. A UAV visual servo control method for wind turbine blade inspection as claimed in claim 1, characterized in that step S2 specifically includes: S21: Use the semantic segmentation model to separate the pixels belonging to the wind turbine blades in the image from the panoramic information, and obtain the panoramic segmentation map of the wind turbine blades; S22: According to the panoramic segmentation image obtained in step S21, extract the pixel coordinates of the points on the three blade tips of the wind turbine as the pixel coordinates of the key points; S23: According to the pixel coordinate values extracted in step S22, the camera pose when taking the image, and the camera parameters, solve the coordinates of the key points of the wind turbine blades in space, specifically: plane S ₁ is point C when the camera shoots at point A. The image plane where it is located, point B on the plane is the center of the image captured by the camera, that is, the intersection point of the optical axis of the camera at A and the plane S ₁ , the pixel coordinate of point B in the image is OK,/> Column, the direction of the camera is the direction of vector AB; Plane S ₂ is the image plane where point C is located when the camera shoots at point D. Point C is on the intersection of plane S ₁ and plane S _2. The plane pixel coordinate of point C on plane S1 is row a ₁ , line 1. b ₁ column, then the pixel distance from the camera center point to the image plane can be obtained in the pixel coordinate system as follows: Among them, θ is the angle of the camera's FOV angle; with camera A as the origin, the vector is the direction vector of straight line L1, vector/> The value of is specifically: Among them, a ₁ is the row number of point C on plane S1, b ₁ is the column number of point C on plane S1; W is the heading angle of the camera in the local space coordinate system o-xyz; Since it is known that the straight line passes through point A and the coordinates of point A are (X ₀ , Y ₀ , Z ₀ ), the parametric equation of the straight line Li can be obtained as follows: Among them, t is a parameter, _Li is the straight line equation corresponding to a key point of the wind turbine in space, and the intersection of multiple space straight lines is the spatial position of the key point. 3. A UAV visual servo control method for wind turbine blade inspection as claimed in claim 2, characterized in that step S3 specifically includes: S31: Based on the three-dimensional coordinate information of the key points, obtain the direction of the wind turbine and the coordinates of the center point of the wind turbine blade; S32: According to the fan orientation and fan blade center point coordinates obtained in step S31, within a fixed distance from the front and back of the fan blade, determine the starting point and end point of the visual servo and provide the inspection direction; the fixed distance is determined by the fan blade size The size is determined by a combination of the camera's FOV angle. 4. A UAV visual servo control method for wind turbine blade inspection as claimed in claim 3, characterized in that: the solution method for the direction of the wind turbine in step S31 is specifically: any one of the three blades of the wind turbine is selected. Difference the three-dimensional coordinates of the two blades to get a vector. Difference the three-dimensional coordinates of the other two blades again to get another vector. Take the outer product of the two vectors to obtain the direction vector of the plane, that is, the direction of the wind turbine; The solution of the center point coordinates of the wind turbine blades is specifically: averaging the three-dimensional coordinates of the three blades of the wind turbine. 5. A UAV visual servo control method for wind turbine blade inspection as claimed in claim 4, characterized in that step S4 specifically includes: S41: Input the captured image into the image segmentation algorithm to separate the leaf image and background image; S42: After data compression of the segmented image, it is used as a state input to the reinforcement learning model for training. The output action of the reinforcement learning model is a fixed positive vector value perpendicular to the inspection direction, performs correction actions, and calculates the reward function. ; S43: Update the reinforcement learning parameters and optimize the reinforcement learning model according to the reward value, status and action; when the reward value reaches the threshold, the training is completed, and the trained modified model is used for visual servoing to complete the trajectory implementation. 6. A UAV visual servo control method for wind turbine blade inspection as claimed in claim 5, characterized in that: the reward function in step S42 is specifically: w＝-(d2+k·s) In the formula, w is the reward function value, d is the offset value in visual servoing, s is the number of servo steps, and k is the coefficient. 7. A UAV visual servo control method for wind turbine blade inspection as claimed in claim 1, characterized in that: in step S1, the edge of the image captured by the UAV camera is cropped by 5% before use. An image calibration card is used to calibrate the camera, and the FOV value of the camera is modified after clipping and calibration. 8. A UAV visual servo control system for wind turbine blade inspection, including a computer program, characterized in that when the computer program is executed by a processor, the steps of any of the above methods are implemented.