Turbine blade surface image acquisition method and system based on 3D visual path planning

Abstract

The invention discloses a steam turbine blade surface image acquisition method and system based on 3D visual path planning, the system comprises an industrial robot, a product tooling platform, a surface data acquisition unit and a geometrical optical unit, the surface data acquisition unit comprises a 3D sensor and a 2D industrial camera, the geometrical optical unit is formed by combining a plurality of optical light sources, and the 3D sensor is connected with the 2D industrial camera. The blade is illuminated at different angles and under different illumination conditions, details and features of the surface of the blade can be highlighted, and the contrast ratio and definition of the image are improved. The method can be widely applied to industrial detection and maintenance, the position and angle of the acquisition equipment can be dynamically adjusted by constructing the 3D model of the blade in real time and intelligently planning an acquisition path algorithm, and a data acquisition system is guided to realize comprehensive and accurate image acquisition on the surface of the blade through accurate path planning and multi-view fusion.

Description

Technical Field

[0001] The present invention relates to the technical field of image acquisition and processing, and in particular to a method and system for acquiring images of a turbine blade surface based on 3D visual path planning. Background Art

[0002] The existing method of acquiring images of the surface of turbine blades mainly relies on a predetermined path and a fixed angle for acquisition, which has many defects. First, the acquisition method of the predetermined path and fixed angle is inefficient and cannot flexibly cope with the complex geometric structure of the blades, resulting in some areas being unable to be effectively covered, and the comprehensiveness and accuracy of image acquisition are difficult to guarantee. Secondly, the complex structure of the blade surface and the diversity of products further increase the difficulty and time cost of image acquisition, especially when faced with blades of different types and shapes. Traditional methods are difficult to adapt to rapidly changing needs. This method is not only time-consuming and labor-intensive, but also prone to missing key details, affecting subsequent quality inspection and analysis.

[0003] With the development of 3D vision technology, combined with path planning algorithms, a new solution is provided for blade surface image acquisition. By building a 3D model of the blade in real time and intelligently planning the acquisition path algorithm, the position and angle of the acquisition device can be dynamically adjusted to guide the data acquisition system to achieve comprehensive and accurate image acquisition of the blade surface. This method can not only significantly improve the acquisition efficiency, reduce unnecessary repetitions and pauses, but also relatively improve the image quality. Through path planning, adaptive adjustments can be made according to the different characteristics of the blade, improve the flexibility and adaptability of the system, and meet the image acquisition requirements of complex surface structures. Therefore, in order to better acquire the surface image of the turbine blade, the design uses 3D vision combined with intelligent planning acquisition path technology to ensure that every detail of the image is clearly captured and acquired, and the surface image of the turbine blade can be acquired efficiently, comprehensively and accurately. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for collecting surface images of turbine blades based on 3D visual path planning, so as to solve the problems of low efficiency, incomplete collection, low image quality, etc. in the prior art; the present invention uses 3D sensors and intelligent algorithms to plan the surface image collection path, and guides the data collection system to realize automatic and high-precision blade surface image collection.

[0005] The technical solution of the present invention is as follows:

[0006] The method for collecting surface images of steam turbine blades based on 3D visual path planning includes the following steps:

[0007] S1. Install the 3D sensor at the end of the robotic arm and perform the hand-eye calibration process to achieve accurate fusion of the sensor camera coordinate system and the world coordinate system in which the robotic arm works.

[0008] S2.3D model construction: use 3D sensors to obtain 3D point cloud data of the turbine blade surface. This step includes data preprocessing, such as denoising and filtering to ensure the accuracy of the model, and constructing a high-precision 3D surface model to provide a basis for subsequent path planning.

[0009] S3. Based on the constructed 3D blade surface model, the present invention implements a grid clustering algorithm to divide the point cloud into several blocks. Specifically, firstly, the minimum and maximum values ​​of the coordinates of the point cloud to be divided are obtained, and the width and height of each point cloud to be divided are set, such as width width and height height. Then, each point is traversed, and a grid relationship is established according to each point and the minimum point cloud coordinate, and the width and height relationship corresponding to each point is calculated, and the relationship between the point and the grid is judged, as shown in the following formula:

[0010] Width grid X = (current point X value - point cloud X minimum value) / width width

[0011] Width grid Y = (current point Y value - point cloud Y minimum value) / height height

[0012] Then the point set is grid clustered to obtain a target point cloud area set with set width and height;

[0013] S4. Traverse the point cloud set in S3 above, and find the centroid and normal vector of each point cloud. Usually, the centroid coordinates of the object are P C The calculation formula is as follows:

[0014]

[0015] Among them, r i =(x i ,y i ,z i ), i=1,2...n are the coordinates of each particle, m i is the mass corresponding to the particle. Let the above formula m i =1, the calculation formula of the point cloud centroid coordinates can be obtained as follows:

[0016]

[0017] S5. Calculate the normal vector of the region block particles in step S4 based on the local surface fitting method;

[0018] First, the surface of the point cloud is smoothed and filtered. For this reason, the local neighborhood of any point can be well fitted with a plane. For each scan point p in the point cloud, the K nearest neighboring points are searched, and then the local plane P in the least squares sense of the K neighboring points is calculated. This plane can be expressed as

[0019]

[0020] In the formula, is the normal vector of plane P, d is the distance from plane P to the origin of the coordinate system, p i For each scan point in the point cloud, the normal vector of the plane fitted by the K nearest points is the normal vector of the current scan point;

[0021] The normal vector of plane P can be obtained by principal component analysis PCA. It is known from the operation that P passes through the centroid p0 of its K adjacent points, and the normal vector satisfy First, perform eigenvalue decomposition on the covariance matrix M in equation (4) to obtain the eigenvalues ​​of M. The eigenvector corresponding to the minimum eigenvalue of M is the normal vector of P.

[0022] S6. Normal vector orientation. The normal vector calculated above is ambiguous, that is, only the straight line where the normal vector is located is obtained, but the direction of the straight line is not determined as the final direction of the normal vector. Next, the normal vector will be oriented.

[0023] Assuming that the point cloud is dense enough and the sampling plane is smooth everywhere, the normal vectors of two adjacent points will be close to parallel. If the inner product is negative, it means that the normal vector of one of the points needs to be flipped. Therefore, first set a normal vector direction for a point in the point cloud, and then traverse all other points. If the normal vector of the current point is set to is the next point to be traversed. If Then Flip, otherwise remain unchanged.

[0024] S7. Calculate the surface curvature of the regional point cloud block, perform eigenvalue decomposition on the covariance matrix M in the above formula (4), and obtain the eigenvalues ​​of M. If the eigenvalues ​​satisfy λ0≤λ1≤λ2, the surface curvature of point P is:

[0025]

[0026] The smaller the δ, the flatter the area, and the larger the δ, the greater the fluctuation of the area. The curvature of the surface can be used to determine the curvature of the turbine blade surface, that is, to divide the area into blocks more reasonably, so that the data collected on the blade surface is clear, stable and reliable.

[0027] S8. Based on the above content, the centroid of each point cloud after segmentation is offset along the normal vector to the working distance of the surface image acquisition system. Combined with the results of the 3D sensor and robot hand-eye calibration, the set of acquisition points is converted to the robot's coordinate system, which can guide the robot to achieve comprehensive and accurate image acquisition of the blade surface.

[0028] A turbine blade surface image acquisition system suitable for the above method specifically includes an industrial robot, a product tooling platform, a surface data acquisition unit and a geometric optical unit. The product tooling platform is located in the working direction of the industrial robot, the surface data acquisition unit is installed at the working end of the robot, and the geometric optical unit is installed below and in front of the surface data acquisition unit. The light source illumination range of the geometric optical unit covers the acquisition range of the surface data acquisition unit.

[0029] The product tooling platform is provided with several groups of fixtures for simultaneously clamping blades of different specifications, and a distance is left between each two groups of fixtures to ensure that the clamped blades are within the collection range of the surface data collection unit.

[0030] The moving range of the industrial robot covers the entire product tooling platform; the industrial robot determines the moving path according to the position transformation relationship between the robot end joint coordinate system and the camera coordinate system.

[0031] The surface data acquisition unit includes a 3D sensor and a 2D industrial camera. Both the 3D sensor and the 2D industrial camera are installed at the end of the industrial robot, and the 3D sensor in the working state is located at the upper end of the 2D industrial camera.

[0032] The surface data acquisition unit is provided with a camera light source connecting piece installed along the shooting direction of the 2D industrial camera.

[0033] The geometric optical unit includes at least a ring light source and a strip light source; the ring light source and the strip light source are both installed at the end of the industrial robot through a camera light source connector, and both can adjust the brightness, color temperature, irradiation direction and angle according to the environment and requirements. The ring light source is installed at the front end of the 2D industrial camera, and the strip light source is installed at the front end of the ring light source. In the working state, the height of the strip light source is between the 3D sensor and the ring light source.

[0034] The technical solution of the present invention has the following beneficial effects:

[0035] (1) Improve detection accuracy. The present invention uses 3D vision technology to obtain high-resolution three-dimensional point cloud data of the blade surface, and intelligently plans the path to guide the high-precision 2D vision system to collect surface images, which can capture subtle defects on the blade surface, such as tiny cracks, inclusions, fisheyes, etc. This high-precision detection greatly improves the reliability and accuracy of the detection results.

[0036] (2) Enhanced detection efficiency. Through the intelligent path planning algorithm, the present invention can automatically generate the best image acquisition path to cover the entire blade surface. Compared with traditional manual detection and two-dimensional imaging systems, automatic path planning significantly reduces the time of image acquisition and improves detection efficiency. At the same time, the system can continuously and seamlessly acquire images, avoiding pauses and repetitions in manual operations.

[0037] (3) Reduce human errors. The present invention uses automated 3D visual modeling to intelligently plan paths to guide high-precision 2D acquisition of surface images, effectively reducing human errors, achieving standardization and consistency of the detection process, and ensuring the objectivity and stability of the detection results.

[0038] (4) Improve data management and analysis capabilities. All relevant data of the present invention can be directly stored in the database, which is convenient for subsequent management and analysis. Combined with data analysis tools, the historical data of the blades can be tracked and compared to help predict the service life and maintenance cycle of the blades, thereby optimizing the maintenance strategy.

[0039] (5) Support for multiple blade types. The present invention has high flexibility and adaptability, and can adapt to turbine blades of different types and sizes, achieve accurate detection of different blades, and expand the application scope and practicality of the system.

[0040] (6) Strong integration and easy maintenance. The present invention can be seamlessly integrated with the existing steam turbine production line and maintenance system, which is easy to implement and promote. The system design focuses on modularization and maintainability, which is convenient for subsequent upgrades and maintenance, and reduces the operating cost of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is an implementation flow chart of the present invention.

[0042] Figure 2 This is a flow chart for implementing the grid clustering algorithm of the present invention.

[0043] Figure 3 Schematic diagram of the photo taking points of the point cloud segmentation and block-segmentation system with image acquisition system of the present invention.

[0044] Figure 4 This is an example diagram of the hand-eye calibration principle in an embodiment of the present invention.

[0045] Figure 5 It is a schematic diagram of the structure of the steam turbine blade surface image acquisition system of the present invention. DETAILED DESCRIPTION

[0046] Example 1

[0047] The image acquisition method of turbine blade surface based on 3D visual path planning, such as Figure 1As shown, the following steps are included:

[0048] S1. Install the 3D sensor of the image acquisition system used in this method at the end of the robotic arm, perform the hand-eye calibration process, and realize the fusion of the sensor camera coordinate system and the world coordinate system of the robotic arm;

[0049] The position transformation relationship between the robot end joint coordinate system and the camera coordinate system can be solved by the equation AX = XB. A represents the transformation relationship between the robot end joints during two adjacent movements, B represents the relative movement of the camera coordinates during two adjacent movements, and X represents the hand-eye relationship.

[0050] Hand-eye calibration is actually solving the matrix equation: AX=XB; A is the homogeneous matrix of the two spatial transformations of the camera (monocular or binocular); B is the homogeneous matrix of the two transformations of the coordinate system of the end of the robot; X is the hand-eye matrix to be solved; by solving this equation multiple times, X can be solved (the method for solving the A and B matrices is as follows).

[0051] Method for finding A:

[0052] A is the transformation matrix between the two camera coordinate systems. Assuming that there are three calibration images in the above camera calibration, the external parameter calibration results are Hc1, Hc2, and Hc3, respectively, then two A matrices can be obtained:

[0053] A1=H c2 *H c1 -1

[0054] A2=H c3 *H c2 -1

[0055] Method for finding B:

[0056] B is the transformation matrix between the two robot coordinate systems. Assuming that the coordinate system description matrices of the robot coordinate systems corresponding to the three calibration images in the above camera calibration are Hg1, Hg2, and Hg3 in the base coordinate system (need to be read or converted from the robot controller or teaching pendant), two B matrices can be obtained:

[0057] B1=Hc2-1*Hc1

[0058] B2=Hc3-1*Hc2

[0059] From the above two groups A and B, substituting into AX=XB, we can get the unique solution X.

[0060] (1) Calculation method

[0061] ① Known camera extrinsic matrix Hc1 , H c2 , H c3

[0062] ② Read the corresponding manipulator coordinate end posture description data from the robot controller. The coordinate system description data of the robot system may have two forms:

[0063] a) Translation vector + Euler angle: x, y, z, r z , r y , r x

[0064] b) Translation vector + four-element mode: x, y, z, q w ,q x ,q y ,q z

[0065] (4) Solve the transformation matrix X from the camera coordinate system to the robot coordinate system

[0066] This example uses a conventional honeycomb calibration plate to collect 16 images for camera calibration, and obtains 25 external parameter matrices. At the same time, 25 corresponding attitude matrices can be read or converted from the controller interface. Combined with (1) and (2) in the above embodiment 1, we can obtain Group A and B, and then use AX=XB to calculate X.

[0067] (5) Hand-eye calibration principle

[0068] This example uses the eyes-on-hands approach, such as Figure 4 As shown, the principle of this method is explained as follows.

[0069] b T g : The homogeneous matrix of the mechanical gripper relative to the base coordinate system

[0070] c T t : The homogeneous matrix of the calibration plate relative to the camera

[0071] b T t : The homogeneous matrix of the calibration plate relative to the robot base coordinate system

[0072] g T c : The homogeneous matrix of the camera relative to the mechanical gripper (i.e. the solved hand-eye matrix)

[0073] For each pose, the homogeneous transformation between the gripper and the robot base coordinate system is recorded using the robot kinematics:

[0074]

[0075] For each pose, the pose estimation method based on the point correspondence of the camera coordinate system is used to record the homogeneous transformation between the calibration plate coordinate system and the camera coordinate system:

[0076]

[0077] The homogeneous matrix is ​​obtained through hand-eye calibration, that is, the target hand-eye matrix is ​​obtained:

[0078]

[0079] S2.3D model construction: use 3D sensors to obtain 3D point cloud data of the turbine blade surface. This step includes data preprocessing, such as denoising and filtering to ensure the accuracy of the model, and constructing a high-precision 3D surface model to provide a basis for subsequent path planning.

[0080] S3. Based on the constructed 3D blade surface model, Figure 2-3 As shown in the figure, the grid clustering algorithm is used to divide the point cloud into several blocks. Specifically, the minimum and maximum values ​​of the coordinates of the point cloud to be divided are first obtained, and the width and height of each point cloud to be divided are set, such as width width and height height; then each point is traversed, and a grid relationship is established according to each point and the minimum point cloud coordinate, and the width and height relationship corresponding to each point is calculated, and the relationship between the point and the grid is judged, as shown in the following formula:

[0081] Width grid X = (current point X value - point cloud X minimum value) / width width

[0082] Width grid Y = (current point Y value - point cloud Y minimum value) / height height

[0083] Then the point set is grid clustered to obtain a target point cloud area set with set width and height;

[0084] S4. Traverse the point cloud set in S3 above, and find the centroid and normal vector of each point cloud. Usually, the centroid coordinates of the object are P C The calculation formula is as follows:

[0085]

[0086] Among them, r i =(x i ,y i ,z i ), i=1,2...n are the coordinates of each particle, m i is the mass corresponding to the particle. Let the above formula m i =1, the calculation formula of the point cloud centroid coordinates can be obtained as follows:

[0087]

[0088] S5. Calculate the normal vector of the region block particles in step S4 based on the local surface fitting method;

[0089] First, the surface of the point cloud is smoothed and filtered. For this reason, the local neighborhood of any point can be well fitted with a plane. For each scan point p in the point cloud, the K nearest neighboring points are searched, and then the local plane P in the least squares sense of the K neighboring points is calculated. This plane can be expressed as

[0090]

[0091] In the formula, is the normal vector of plane P, d is the distance from plane P to the origin of the coordinate system; the normal vector to the plane fitted by the K nearest points is the normal vector of the current scanning point;

[0092] The normal vector of plane P can be obtained by principal component analysis PCA. It is known from the operation that P passes through the centroid p0 of its K adjacent points, and the normal vector satisfy First, perform eigenvalue decomposition on the covariance matrix M in equation (4) to obtain the eigenvalues ​​of M. The eigenvector corresponding to the minimum eigenvalue of M is the normal vector of P.

[0093] S6. Normal vector orientation. The normal vector calculated above is ambiguous, that is, only the straight line where the normal vector is located is obtained, but the direction of the straight line is not determined as the final direction of the normal vector. Next, the normal vector will be oriented.

[0094] Assuming that the point cloud is dense enough and the sampling plane is smooth everywhere, the normal vectors of two adjacent points will be close to parallel. If the inner product is negative, it means that the normal vector of one of the points needs to be flipped. Therefore, first set a normal vector direction for a point in the point cloud, and then traverse all other points. If the normal vector of the current point is set to is the next point to be traversed. If Then Flip, otherwise remain unchanged.

[0095] S7. Calculate the surface curvature of the regional point cloud block, perform eigenvalue decomposition on the covariance matrix M in the above formula (4), and obtain the eigenvalues ​​of M. If the eigenvalues ​​satisfy λ0≤λ1≤λ2, the surface curvature of point P is:

[0096]

[0097] The smaller the δ, the flatter the area, and the larger the δ, the greater the fluctuation of the area. The curvature of the surface can be used to determine the curvature of the turbine blade surface, that is, to divide the area into blocks more reasonably, so that the data collected on the blade surface is clear, stable and reliable.

[0098] S8. Based on the above content, the centroid of each point cloud after segmentation is offset along the normal vector to the working distance of the surface image acquisition system. Combined with the results of the 3D sensor and robot hand-eye calibration, the set of acquisition points is converted to the robot's coordinate system, which can guide the robot to achieve comprehensive and accurate image acquisition of the blade surface.

[0099] Corresponding to the above method, the turbine blade surface image acquisition system used is as follows: Figure 5 As shown, it includes an industrial robot, a product tooling platform, a surface data acquisition unit and a geometric optical unit. The product tooling platform is located in the working direction of the industrial robot, the surface data acquisition unit is installed at the working end of the robot, and the geometric optical unit is installed below and in front of the surface data acquisition unit. The light source illumination range of the geometric optical unit covers the collection range of the surface data acquisition unit.

[0100] Several groups of clamps are arranged on the product tooling platform, which can clamp blades of different specifications at the same time, and there is a certain distance between each two groups of clamps to ensure that there is enough distance between the clamped blades, the blades do not block each other, and do not affect the surface data acquisition unit to collect the surface image of each group of blades.

[0101] The industrial robot has a moving range covering the entire product tooling platform. When the industrial robot receives an instruction, it drives the surface data acquisition unit and the geometric optical unit to move to a specified position, and the surface data acquisition unit and the geometric optical unit start working.

[0102] The industrial robot determines a moving path according to a position transformation relationship between a robot end joint coordinate system and a camera coordinate system.

[0103] The surface data acquisition unit includes a 3D sensor and a 2D industrial camera. Both the 3D sensor and the 2D industrial camera are installed at the end of the industrial robot, and the 3D sensor in the working state is located at the upper end of the 2D industrial camera.

[0104] The surface data acquisition unit is provided with a camera light source connecting piece installed along the shooting direction of the 2D industrial camera.

[0105] The geometric optical unit includes at least a ring light source and a strip light source, which illuminate the blades on the product tooling platform under different angles and lighting conditions, highlight the details and features of the blade surface, and improve the contrast and clarity of the image; the ring light source and the strip light source are both installed at the end of the industrial robot through a camera light source connector, the ring light source is installed at the front end of the 2D industrial camera, and the strip light source is installed at the front end of the ring light source.

[0106] Furthermore, in the operating state steam turbine blade surface image acquisition system, the height of the strip light source is located between the 3D sensor and the annular light source.

[0107] Furthermore, in order to adapt to the acquisition environment and requirements of the turbine blade surface image acquisition system, the annular light source can adjust the brightness, color temperature, irradiation direction and angle, and the strip light source can also adjust the brightness, irradiation direction and angle.

[0108] During operation, the system:

[0109] The industrial robot used is a 6-axis industrial robot with a working range radius of 900mm and a load of 0.9kg.

[0110] The 3D sensor used is a binocular face scanning grating 3D camera, which has the characteristics of small and medium field of view, high precision, and high beat. Its working indicators include: working distance 700mm~1300mm, effective field of view 400mm*400mm~1100mm*1100mm, so it can be compatible with about

[0111] Product specifications: 400mm*400~1000mm*1000mm.

[0112] The 2D industrial camera used has a resolution of 5120*5120 and a pixel size of 2.5um.

[0113] The focal length of the industrial lens used is 35mm.

[0114] When the system is in operation, the 3D sensor acquires the three-dimensional point cloud data of the turbine blade surface for building a high-precision 3D model. This step includes data preprocessing, such as denoising and filtering to ensure the accuracy of the model. Building a high-precision 3D surface model provides a basis for subsequent path planning. Then, based on the high-precision 3D model, automatic path planning is implemented to guide the 2D industrial camera in the data acquisition unit to acquire blade surface images; finally, feature analysis is performed through the acquired blade surface image data to obtain analysis conclusions.

Claims

1. A method for acquiring a surface image of a steam turbine blade based on 3D visual path planning, comprising the following steps: S1. Install the 3D sensor at the end of the robotic arm and perform the hand-eye calibration process to achieve the fusion of the sensor camera coordinate system and the world coordinate system of the robotic arm; S2.3D model construction: use 3D sensors to obtain 3D point cloud data of the turbine blade surface. This step includes data preprocessing, such as denoising and filtering to ensure the accuracy of the model, and constructing a high-precision 3D surface model to provide a basis for subsequent path planning. S3. Based on the constructed 3D blade surface model, the grid clustering algorithm is used to segment the point cloud into several blocks. Specifically, the minimum and maximum values ​​of the point cloud coordinates to be segmented are first obtained, and the width and height of each segmented point cloud are set, such as width width and height height; then each point is traversed, and a grid relationship is established according to each point and the minimum point cloud coordinate, and the width and height relationship corresponding to each point is calculated, and the relationship between the point and the grid is determined, as shown in the following formula: Width grid X = (current point X value - point cloud X minimum value) / width width Width grid Y = (current point Y value - point cloud Y minimum value) / height height Then the point set is grid clustered to obtain a target point cloud area set with set width and height; S4. Traverse the point cloud set in S3 above, and find the centroid and normal vector of each point cloud. Usually, the centroid coordinates of the object are P C The calculation formula is as follows: Among them, r i =(x i ,y i ,z i ), i=1,2...n are the coordinates of each particle, m i is the mass corresponding to the particle. Let the above formula m i =1, the calculation formula of the point cloud centroid coordinates can be obtained as follows: S5. Calculate the normal vector of the region block particles in step S4 based on the local surface fitting method; First, the surface of the point cloud is smoothed and filtered. For this reason, the local neighborhood of any point can be well fitted with a plane. For each scan point p in the point cloud, the K nearest neighboring points are searched, and then the local plane P in the least squares sense of the k neighboring points is calculated. This plane can be expressed as In the formula, is the normal vector of plane P, d is the distance from plane P to the origin of the coordinate system, p i For each scan point in the point cloud, the normal vector of the plane fitted by the K nearest points is the normal vector of the current scan point; The normal vector of plane P can be obtained by principal component analysis PCA. It is known from the operation that P passes through the centroid p0 of its K adjacent points, and the normal vector satisfy First, perform eigenvalue decomposition on the covariance matrix M in equation (4) to obtain the eigenvalues ​​of M. The eigenvector corresponding to the minimum eigenvalue of M is the normal vector of P. S6. Normal vector orientation. The normal vector calculated above is ambiguous, that is, only the straight line where the normal vector is located is obtained, but the direction of the straight line is not determined as the final direction of the normal vector. Next, the normal vector will be oriented. Assuming that the point cloud is dense enough and the sampling plane is smooth everywhere, the normal vectors of two adjacent points will be close to parallel. If the inner product is negative, it means that the normal vector of one of the points needs to be flipped. Therefore, first set a normal vector direction for a point in the point cloud, and then traverse all other points. If the normal vector of the current point is set to is the next point to be traversed. If Then Flip, otherwise remain unchanged. Is the i here consistent with the previous i? If not, please replace it with another character. S7. Calculate the surface curvature of the regional point cloud block, perform eigenvalue decomposition on the covariance matrix M in the above formula (4), and obtain the eigenvalues ​​of M. If the eigenvalues ​​satisfy λ0≤λ1≤λ2, the surface curvature of point P is: The smaller the δ, the flatter the area, and the larger the δ, the greater the fluctuation of the area. The curvature of the surface can be used to determine the curvature of the turbine blade surface, that is, to divide the area into blocks more reasonably, so that the data collected on the blade surface is clear, stable and reliable. S8. Based on the above content, the centroid of each point cloud after segmentation is offset along the normal vector to the working distance of the surface image acquisition system. Combined with the results of the 3D sensor and robot hand-eye calibration, the set of acquisition points is converted to the robot's coordinate system, which can guide the robot to achieve comprehensive and accurate image acquisition of the blade surface.

2. The method for collecting surface images of steam turbine blades based on 3D visual path planning according to claim 1, characterized in that: The position transformation relationship between the robot end joint coordinate system and the camera coordinate system is specifically solved by the equation AX=XB; where A represents the transformation relationship between the robot end joints during two adjacent movements, B represents the relative movement of the camera coordinates during two adjacent movements, and X is the hand-eye relationship.

3. A steam turbine blade surface image acquisition system applicable to the method described in claim 1 or 2, characterized in that: It includes an industrial robot, a product tooling platform, a surface data acquisition unit and a geometric optical unit. The product tooling platform is located in the working direction of the industrial robot, the surface data acquisition unit is installed at the working end of the robot, and the geometric optical unit is installed below and in front of the surface data acquisition unit. The light source illumination range of the geometric optical unit covers the collection range of the surface data acquisition unit.

4. The steam turbine blade surface image acquisition system based on 3D visual path planning according to claim 3 is characterized in that: The product tooling platform is provided with several groups of fixtures for simultaneously clamping blades of different specifications, and a distance is left between each two groups of fixtures to ensure that the clamped blades are within the collection range of the surface data collection unit.

5. The steam turbine blade surface image acquisition system based on 3D visual path planning according to claim 3 is characterized in that: The moving range of the industrial robot covers the entire product tooling platform; the industrial robot determines the moving path according to the position transformation relationship between the robot end joint coordinate system and the camera coordinate system.

6. The steam turbine blade surface image acquisition system based on 3D visual path planning according to claim 3 is characterized in that: The surface data acquisition unit includes a 3D sensor and a 2D industrial camera. Both the 3D sensor and the 2D industrial camera are installed at the end of the industrial robot, and the 3D sensor in the working state is located at the upper end of the 2D industrial camera.

7. The steam turbine blade surface image acquisition system based on 3D visual path planning according to claim 6, characterized in that: The surface data acquisition unit is provided with a camera light source connecting piece installed along the shooting direction of the 2D industrial camera.

8. The steam turbine blade surface image acquisition system based on 3D visual path planning according to claim 7 is characterized in that: The geometric optical unit includes at least a ring light source and a strip light source; the ring light source and the strip light source are both installed at the end of the industrial robot through a camera light source connector, the ring light source is installed at the front end of the 2D industrial camera, and the strip light source is installed at the front end of the ring light source.

9. The steam turbine blade surface image acquisition system based on 3D visual path planning according to claim 8, characterized in that: In the steam turbine blade surface image acquisition system in an operating state, the height of the strip light source is located between the 3D sensor and the annular light source.

10. The steam turbine blade surface image acquisition system based on 3D visual path planning according to claim 6, characterized in that: The brightness, color temperature, irradiation direction and angle of the annular light source are adjusted according to the environment and requirements, and the brightness, irradiation direction and angle of the strip light source are adjusted according to the environment and requirements.

Images