Feature Extraction Method for Flash of Large Wind Turbine Blades Based on Point Cloud Normal Vector Difference

Abstract

The present invention discloses a method for extracting the flash feature of a large wind turbine blade based on the difference of point cloud normal vectors, which includes the following steps: S1. Point cloud preprocessing: Obtain the blade point cloud data of the wind turbine blade through an RGBD camera, and perform primary filtering on the blade point cloud data to obtain the preprocessed blade point cloud data; S2. Feature segmentation based on the difference of normal vectors: Calculate the point cloud resolution of the blade, determine two support radii of different sizes according to the point cloud resolution, calculate the normal vectors under the two support radii, perform conditional filtering according to the difference of the normal vectors of the two support radii, remove the points in the blade point cloud that are less than the set threshold, and separate the flash feature from the blade point cloud; S3. Clustering segmentation: Perform secondary filtering on the blade point cloud data, and completely separate the flash from the blade through clustering segmentation to obtain the feature point cloud of the blade flash. It can realize the extraction of the relatively small feature of the flash of the huge wind turbine blade, and the extraction method can be quickly deployed; the extraction effect of the flash feature is good and it has strong robustness.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional point cloud, and relates to a method for extracting features of point cloud, in particular to a method for extracting flash edges of large wind turbine blades based on the difference of normal vectors. Background Art

[0002] The southeast coastal areas of our country are rich in offshore wind energy resources, which are very conducive to the construction of offshore wind farms. The construction of an offshore wind farm requires a certain scale of wind turbines and a power transmission system, among which the wind turbine is the key equipment for converting wind energy into electrical energy. The blade is the most important component of the wind turbine, and the manufacturing cost of the wind turbine blade accounts for about 15-20% of the production cost of the wind turbine.

[0003] The blade production process is divided into a molding stage and a post-treatment stage. In the molding stage, materials are poured into the mold and the mold is closed. The post-treatment stage mainly processes the blade after demolding. After demolding, it is first necessary to cut off the flash edges and polish the flash edges. This process is currently mainly operated manually, with a harsh working environment and high operation difficulty, which is one of the problems that need to be solved in the automated processing of wind turbine blades.

[0004] Three-dimensional point cloud technology has been widely used in fields such as surveying and mapping, intelligent driving, and reverse engineering. In the field of surface processing of large wind turbine blades, three-dimensional point cloud technology has also been introduced. The flash edge feature is located on the surface of the wind turbine blade. After the flash edge of the blade after demolding is cut, its feature only protrudes about 1-2 mm from the blade, and the width is about 15 mm. The width and height of the common blade cross-section are both above 1 m. At the same time, the blade has a complex three-dimensional surface, and large deformation and torsion often occur when it is placed. Therefore, after obtaining the three-dimensional point cloud of the blade after this process, the extraction of the point cloud of the blade flash edge feature has become a major problem, and there is little research on the extraction of the flash edge feature of wind turbine blades at present. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem of difficult extraction of the point cloud of the flash edge feature of the existing wind turbine blade, and provide a method for extracting the flash edge feature of a large wind turbine blade based on the differential of the point cloud normal vector, which is convenient for the extraction of the point cloud of the flash edge feature of the wind turbine blade, has a good extraction effect of the flash edge feature, and has strong robustness.

[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A method for extracting the flash edge feature of a large wind turbine blade based on the difference of the point cloud normal vector, comprising the following steps:

[0008] S1. Point cloud preprocessing: Obtain the blade point cloud data of the wind turbine blade through an RGBD camera, and perform primary filtering on the blade point cloud data to obtain the preprocessed blade point cloud data;

[0009] S2. Feature segmentation based on the difference of normal vectors: Calculate the point cloud resolution of the blade, determine two support radii of different sizes according to the point cloud resolution, calculate the normal vectors under the two support radii, perform conditional filtering based on the difference of the normal vectors of the two support radii, remove the points in the blade point cloud that are less than the set threshold, and separate the flash feature from the blade point cloud;

[0010] S3. Clustering segmentation: Filter the blade point cloud data again, and completely separate the flash from the blade through clustering segmentation to obtain the feature point cloud of the blade flash.

[0011] Furthermore, the primary filtering includes direct filtering and statistical filtering. The direct filtering is used to remove the points in the irrelevant area according to the distance between the RGBD camera and the wind power blade, and the statistical filtering is used to remove the irrelevant points and noise points in the blade point cloud.

[0012] Furthermore, the process of statistical filtering is as follows:

[0013] By calculating the distance l from each point in the point cloud to its n neighboring points i , where i = 1, 2, 3... n,, and then calculate the average distance i according to the distance l and the variance σ of the average distance value. The calculation result conforms to the Gaussian distribution Statistical filtering removes the points where the distance l i is outside the range of , where m is the set value. By changing the size of m, the filtering effect can be changed.

[0014] Furthermore, the calculation method of the point cloud resolution is as follows: Traverse each point in the point cloud data, find the point closest to each point, and calculate the distance between the two points; Accumulate all the distances and divide by the number of points in the point cloud data.

[0015] Furthermore, the principal component analysis method is used for normal vector calculation, and a viewing point is introduced to determine the normal vector direction. The viewing point is default to the origin of the coordinate system, and the viewing point can also be set to other points in the coordinate system separately.

[0016] Furthermore, for the point p in the point cloud, first calculate the centroid of the area under the support radius of the point p. The formula for the centroid of the area is:

[0017]

[0018] In the formula, k represents the number of points within the support radius of the point p, is the three-dimensional centroid of this area.

[0019] Furthermore, according to the centroid of the area of the support radius of the point p, establish a covariance matrix. The formula for the covariance matrix is:

[0020]

[0021] In the formula, C represents the covariance matrix.

[0022] Furthermore, according to the calculated covariance matrix C, calculate the eigenvalues and eigenvectors of the covariance matrix C. The formulas for the eigenvalues and eigenvectors are:

[0023]

[0024] In the formula, λ i represents the i-th eigenvalue of the covariance matrix, represents the j-th eigenvector of the covariance matrix.

[0025] Furthermore, according to the formulas for the eigenvalues and eigenvectors, the eigenvector corresponding to the minimum eigenvalue is the normal vector at point p, and the normal vector needs to satisfy the formula:

[0026]

[0027] In the formula, is the normal vector, and v p is the viewing point.

[0028] Furthermore, the normal vector difference of any point p in the point cloud is:

[0029]

[0030] In the formula, is the normal vector of point p under the large support radius, is the normal vector of point p under the small support radius.

[0031] Compared with the prior art, the present invention can achieve the extraction of relatively small features such as flash on a huge wind turbine blade, and at the same time, the method for extracting flash features of a large wind turbine blade based on the normal vector difference of the point cloud of the present invention can be quickly deployed without relying on a large amount of data for model training; the extraction effect of the flash features of the present invention is good, and the extraction effect of the flash features is not affected by scanning angles, sampling noise, etc., and has strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a schematic diagram of the flash feature of the wind turbine blade.

[0033] Figure 2 is a schematic flow diagram of the method for extracting flash features of a large wind turbine blade based on the normal vector difference of the point cloud of the present invention.

[0034] Figure 3 is a schematic diagram of the normal vector difference at the flash feature.

[0035] Figure 4 Schematic diagram of segmenting the flash features of wind turbine blades, where a is the original point cloud image, b is the effect image after straight-through filtering, c is the effect image after statistical filtering, d is the schematic diagram of the normal vector difference feature calculation results, e is the effect image after conditional filtering, and f is the clustering effect image.

[0036] Figure 5 Schematic diagram of the segmented pure wind turbine blade flash feature point cloud, where a is the schematic diagram of the flash feature point cloud from a side perspective, and b is the schematic diagram of the flash feature point cloud from a front perspective. DETAILED DESCRIPTION

[0037] The flash feature extraction method of large wind turbine blades based on point cloud normal vector difference of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0038] The width and height of the wind turbine blade section are both over 1 meter. After demoulding, the blade needs to be flash cut first. After cutting, the remaining flash features are located at the front and rear ends of the section. Figure 1 The residual flash feature protrudes about 2mm from the blade surface, and the flash width is about 15mm. Compared with the whole blade, the residual flash feature is not obvious enough, and visual recognition is very difficult. Therefore, an effective flash feature extraction method is urgently needed to extract the point cloud of the blade flash feature.

[0039] See also Figure 2 In view of the difficulty in extracting blade flash features, the present invention discloses a method for extracting flash features of large wind turbine blades based on point cloud normal vector difference, comprising the following steps:

[0040] S1. Point cloud preprocessing: The blade point cloud data of the wind turbine blade is obtained through the RGBD camera, and the blade point cloud data is initially filtered to obtain the preprocessed blade point cloud data.

[0041] S2. Feature segmentation based on normal vector difference: calculate the point cloud resolution of the blade, determine the large and small support radii based on the point cloud resolution, calculate the normal vector under the two support radii, perform conditional filtering based on the normal vector difference of the two support radii, remove points in the blade point cloud that are smaller than the set threshold, and separate the flash features from the blade point cloud.

[0042] S3. Clustering segmentation: The blade point cloud data is filtered again, and the flash edge and the blade are completely separated through clustering segmentation to obtain the characteristic point cloud of the blade flash edge.

[0043] In step S1, the initial filtering includes direct filtering and statistical filtering. The direct filtering is used to remove the points in the irrelevant regions according to the distance between the RGBD camera and the wind turbine blade, and the statistical filtering is used to remove the irrelevant points and noise points in the blade point cloud. When performing the initial filtering on the blade point cloud data, first perform direct filtering according to the distance between the RGBD camera and the wind turbine blade to be measured, and remove the points beyond the distance. Subsequently, use statistical filtering to remove the invalid points and noise points in the blade point cloud.

[0044] The statistical filtering calculates the distance l from each point in the point cloud to its n neighboring points i , where i = 1, 2, 3... n,, and then calculates the average distance i and the variance σ of the average distance value. The calculation results conform to the Gaussian distribution The statistical filtering removes the points outside the range of the distance l where m is a set value. By changing the size of m, the filtering effect can be changed. m = 1, 2, 3. To obtain a better filtering effect, m can be taken as 1. i within the range, where m is a set value. By changing the size of m, the filtering effect can be changed. m = 1, 2, 3. To obtain a better filtering effect, m can be taken as 1.

[0045] In step S2, calculate the point cloud resolution of the preprocessed blade point cloud data, and determine two support radii for calculating the normal vectors, one large and one small. Calculate the normal vectors based on the large and small support radii, calculate the normal vector difference feature of the large and small support radii, and perform conditional filtering based on the calculation results.

[0046] The calculation method of the point cloud resolution is as follows: traverse each point in the point cloud data, find the point closest to each point, and calculate the distance between the two points; accumulate all the distances and divide by the number of points in the point cloud data. The large and small support radii can be determined according to the point cloud resolution. The large support radius is generally 2 times the small support radius.

[0047] The calculation of the normal vector adopts the principal component analysis method (PCA), and a viewing point is introduced to determine the direction of the normal vector. The viewing point is defaulted to the origin of the coordinate system, and the viewing point can also be set separately to other points in the coordinate system.

[0048] For the point p in the point cloud, first calculate the centroid of the region under the support radius of the point p. The formula for the centroid of the region is:

[0049]

[0050] In the formula, k represents the number of points within the support radius of the point p, is the three-dimensional centroid of this region.

[0051] According to the centroid of the region under the support radius of the point p establish a covariance matrix. The formula for the covariance matrix is:

[0052]

[0053] In the formula, C represents the covariance matrix.

[0054] According to the calculated covariance matrix C, calculate the eigenvalues and eigenvectors of the covariance matrix C. The formulas for the eigenvalues and eigenvectors are:

[0055]

[0056] In the formula, λ i represents the i-th eigenvalue of the covariance matrix, represents the j-th eigenvector of the covariance matrix.

[0057] According to the formulas for the eigenvalues and eigenvectors, the eigenvector corresponding to the minimum eigenvalue is the normal vector at point p. Since the calculated normal vector is not unique, a viewing point is introduced to solve this problem. The normal vector needs to satisfy the formula:

[0058]

[0059] In the formula, is the normal vector, and v p is the viewing point.

[0060] The difference of the normal vectors of any point p in the point cloud is:

[0061]

[0062] In the formula, is the normal vector of point p under the large support radius, is the normal vector of point p under the small support radius.

[0063] Please refer to Figure 3 , it can be seen from Figure 3 that the calculation of the normal vector under the large support radius is less affected by the small-scale structure, while the calculation of the normal vector under the small support radius is easily affected by the small-scale structure. If the change in the normal vectors calculated under the two support radius values is small, the surface structure will not change significantly. If the change in the normal vectors is large, the surface structure will change significantly.

[0064] According to the calculated difference of the normal vectors under the large and small support radii, set a threshold for conditional filtering, and the flash feature can be completely separated from the point cloud of the blade surface. The flash segmentation threshold of the wind turbine blade is generally less than 0.1. Screen the raised positions or the positions where the curvature changes suddenly on the surface according to the normal vector difference feature, and remove the points in this part to completely separate the flash feature from the blade surface.

[0065] In step S3, clustering segmentation is adopted. By setting the search radius, the minimum and maximum number of clustering points, the flash point cloud is completely segmented from the blade surface, and a pure flash feature point cloud is obtained.

[0066] Please refer to Figure 4 and Figure 5 , for the method for extracting the flash feature of large wind turbine blades based on the point cloud normal vector difference in the present invention, taking the wind turbine blade model scanned by the RGBD camera as the experimental object, the experimental steps are as follows:

[0067] 1. Preprocess the point cloud scanned by the RGBD camera, and adopt direct filtering and statistical filtering to remove noise points and irrelevant points to obtain a pure wind turbine blade surface point cloud.

[0068] 1.1. The camera is 600 mm away from the model, and the original point cloud data collected is as Figure 4 shown in a. Since the RGBD camera used in this experiment has a visual blind area of 300 mm, the direct filtering is set to remove the data more than 300 mm to 600 mm away from the camera optical center in the Z direction. At the same time, since the acquisition bracket is against the wall and the wall surface is 160 mm away from the optical center, it is necessary to remove the points greater than 160 mm in the X direction. The effect of direct filtering is as Figure 4 shown in b.

[0069] 1.2. The number of nearest neighbor points considered in the statistical filtering setting is 50, and σ = 1. The filtering effect is as Figure 4 shown in c.

[0070] 1.3. After preprocessing, the number of points in the point cloud data is 278378.

[0071] 2. Calculate the point cloud resolution, determine the size support radius, calculate the normal vectors of the point cloud under the size support radius, calculate the normal vector difference feature, and set a threshold for conditional filtering.

[0072] 2.1. The resolution of the experimental point cloud is 0.471509. The small support radius is 10 times the resolution, which is 4.71509. The large support radius is twice the small support radius, which is 9.43018.

[0073] 2.2. After the calculation of the normal vector difference feature is completed, the calculation result is as Figure 4 shown in d. Subsequently, conditional filtering is used to remove the points less than the threshold. In this experiment, 0.056 is taken. The effect is as Figure 4 shown in e.

[0074] 3. According to the previous calculation results, perform clustering segmentation to obtain a pure flash feature point cloud. The clustering effect is as Figure 4 shown in f.

[0075] 3.1. Perform statistical filtering again before clustering segmentation to remove irrelevant points.

[0076] 3.2. The search radius for experimental clustering is 1, the minimum number of clusters is 2000, and the maximum number of clusters is 60000.

[0077] 3.3. The flash point cloud obtained by segmentation is as Figure 5 shown Figure 5 Figure a is a flash feature point cloud diagram from the side view, Figure 5 Figure b is a flash feature point cloud diagram from the front view.

[0078] In summary, the present invention can achieve the extraction of relatively small features such as flashes on large wind turbine blades. At the same time, the method for extracting flash features of large wind turbine blades based on the point cloud normal vector difference of the present invention can be quickly deployed without relying on a large amount of data for model training. The extraction effect of the flash features of the present invention is good, and the extraction effect of the flash features is not affected by scanning angles, sampling noise, etc., and has strong robustness.

[0079] The above description is a detailed description of the preferred feasible embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. Any equivalent changes or modifications made under the technical spirit disclosed by the present invention shall fall within the scope of the patent covered by the present invention.

Claims

1. A method for extracting the flash characteristics of large wind turbine blades based on the difference of point cloud normal vectors, characterized in that, It includes the following steps: S1. Point cloud preprocessing: Obtain the blade point cloud data of the wind turbine blade through an RGBD camera, perform primary filtering on the blade point cloud data, and obtain the preprocessed blade point cloud data; S2. Feature segmentation based on normal vector difference: Calculate the point cloud resolution of the blade, determine two support radii of different sizes according to the point cloud resolution, calculate the normal vectors under the two support radii, perform conditional filtering according to the normal vector difference of the two support radii, remove the points in the blade point cloud that are less than the set threshold, and separate the flash feature from the blade point cloud; S3. Clustering segmentation: Perform secondary filtering on the blade point cloud data, and completely separate the flash from the blade through clustering segmentation to obtain the feature point cloud of the blade flash; The primary filtering includes passing filtering and statistical filtering. The passing filtering is used to remove the points in the irrelevant area according to the distance between the RGBD camera and the wind turbine blade, and the statistical filtering is used to remove the irrelevant points and noise points in the blade point cloud; The normal vector calculation uses the principal component analysis method, and a viewing point is introduced to determine the normal vector direction. The viewing point is defaulted to the origin of the coordinate system, and the viewing point can also be separately set to other points in the coordinate system; For a point p in the point cloud, first calculate the centroid of the area under the support radius of point p. The formula for the centroid of the area is: where k represents the number of points within the support radius of point p, is the three-dimensional centroid of this area; According to the centroid of the area under the support radius of point p Establish a covariance matrix, and the formula for the covariance matrix is: In the formula, C represents the covariance matrix; According to the calculated covariance matrix C, calculate the eigenvalues and eigenvectors of the covariance matrix C. The formulas for the eigenvalues and eigenvectors are: where λ i represents the i-th eigenvalue of the covariance matrix, represents the j-th eigenvector of the covariance matrix; According to the formulas for the eigenvalues and eigenvectors, the eigenvector corresponding to the minimum eigenvalue is the normal vector at point p, and the normal vector needs to satisfy the formula: In the formula, is the normal vector, and v p is the viewpoint; Normal vector difference of any point p in the point cloud is as follows: Wherein, is the normal vector of point p under the large support radius, is the normal vector of point p under the small support radius.

2. The method for extracting the flash characteristics of large wind turbine blades based on the difference of point cloud normal vectors according to claim 1, characterized in that, The process of statistical filtering is: By calculating the distance \(l\) from each point in the point cloud to its \(n\) neighboring points i , where \(i = 1, 2, 3,\cdots,n\), and then calculating the average distance according to the distance \(l\) i Calculate the average distance and the variance \(\sigma\) of the average distance value. The calculation result conforms to the Gaussian distribution Statistical filtering removes the points where the distance \(l\) i is outside the range of , where \(m\) is a set value. By changing the size of \(m\), the filtering effect can be changed.

3. The method for extracting the flash characteristics of large wind turbine blades based on the difference of point cloud normal vectors according to claim 1, characterized in that, The calculation method of the point cloud resolution is to traverse each point in the point cloud data, find the point closest to each point, and calculate the distance between the two points; accumulate all the distances and divide by the number of points in the point cloud data.

Images