Aviation blade tip track generation method based on structured light vision

Abstract

The invention discloses an aviation blade tip track generation method based on structured light vision. The aviation blade tip track generation method comprises the following steps: capturing a tip point cloud of an aviation blade through a 3D surface structured light camera, and filtering and denoising to obtain a homogenized point cloud; determining a local normal direction and a tangent plane by adopting a principal component analysis method, and determining an edge point cloud in the homogenized point cloud according to a projection vector angle on the tangent plane; a principal component analysis method is adopted to determine a blade tip first principal direction represented by edge points in the homogenized point cloud, a new space coordinate system is constructed, a fitting intermediate line is solved in XY and XZ dimensions by a polynomial with endpoint constraint through a least square method, and then the fitting intermediate line is converted to an original coordinate. The method can adapt to aviation blade tips of different sizes, carries out middle line extraction on the aviation blade tips, and is improved in the aspects of adaptability, calculation efficiency and anti-noise capability.

Description

Technical Field

[0001] The invention relates to a blade tip trajectory generation method, belonging to the technical field of irregular profile measurement. Background Art

[0002] Aircraft blades are the core components of aircraft engines. As a typical representative of complex curved surface workpieces, their surfaces have non-uniform curvature and complex three-dimensional geometric structures, and must meet strict fluid mechanics requirements. This characteristic leads to drastic changes in the tip surface and frequent small radius bending, which places high demands on the accuracy of midline extraction, and at the same time requires strict control of the smoothness and stability of the midline.

[0003] The existing technical solutions mainly have the problem of poor adaptability to asymmetric regular blades or areas with high curvature changes on blades, resulting in poor extraction accuracy and high computational complexity. Summary of the invention

[0004] In view of the above-mentioned defects of the prior art, the task of the present invention is to provide a method for generating aviation blade tip trajectories based on structured light vision, so as to solve the problems of poor adaptability to the curvature change area of ​​the blade, low accuracy and high computational complexity.

[0005] The technical solution of the present invention is as follows: A method for generating aviation blade tip trajectories based on structured light vision, comprising the steps of: The tip point cloud of the aviation blade is captured by a 3D surface structured light camera and filtered to remove noise to obtain a homogenized point cloud; For each point in the homogenized point cloud, the principal component analysis method is used to determine the local normal and construct the tangent plane. The projection of the neighboring points of the point on the tangent plane is calculated, and the points whose angle difference of the projection polar angle exceeds the set angle threshold are selected as edge points. The principal component analysis method is used to determine the first principal direction of the blade tip represented by the edge points in the homogenized point cloud and construct a new coordinate system make Align the first principal direction, transform the edge point cloud from the original coordinate system to the new coordinate system, solve the fitted middle line in the XY and XZ dimensions using the least squares method with a polynomial with endpoint constraints, and then transform the fitted middle line back to the original coordinate system.

[0006] Furthermore, the middle line is smoothed and evenly segmented after being converted back to the original coordinates.

[0007] Furthermore, the step of uniform segmentation includes first calculating the cumulative instantaneous height of the entire middle line, then determining the position of each sampling point according to the total length and the target number of segments, and obtaining the coordinates of the corresponding points by linear interpolation.

[0008] Furthermore, the filtering and denoising includes first performing height filtering by setting a Z-axis value interval, and then performing voxel filtering.

[0009] Furthermore, the method of using the principal component analysis method to determine the local normal direction and construct the tangent plane for each point in the homogenized point cloud includes selecting a point from the point cloud of k The nearest neighbor points are ,in ,in , Yes The horizontal axis of Yes The vertical coordinate of Yes The height coordinate of Zero-mean the point cloud data ,in: , , , in: ; Calculate the eigenvalues ​​and eigenvectors of the covariance matrix G and get three eigenvalues and the corresponding eigenvector , and the minimum eigenvalue The corresponding eigenvector Yes The normal vector , and the normal vector The orthogonal plane is the tangent plane. ,in yes The transpose of .

[0010] Further, k The value is not less than 50.

[0011] Furthermore, the difference between the first and last angles when filtering the angle difference of the projection polar angle for , represents the first projection polar angle, Indicates the last projected polar angle.

[0012] Further, using the principal component analysis method to determine the first principal direction of the edge point in the homogenized point cloud includes: Assume that the 3D edge point cloud data in the original coordinate system ,in: , , , Zero-mean the point cloud data ,in: , , , Among them, N is the total number of edge point clouds, , Calculate the eigenvalues ​​and eigenvectors of the covariance matrix C. The eigenvector corresponding to the largest eigenvalue is represents the first main direction of the data, ,in for The transpose of .

[0013] Furthermore, when constructing a new coordinate system, select a coordinate system in space that is consistent with Reference vectors for axes that are not parallel , Axis , for Axis unit vector, then determined by the right-hand screw rule axis, Axis .

[0014] Further, solving the fitting middle line by least square method with polynomials with endpoint constraints in XY and XZ dimensions includes: Taking the extreme point on the first main direction of the blade tip as the endpoint, a linear interpolation function passing through the endpoint is constructed. , , The coordinates of the extreme points are ; The residual error between the point cloud data and the linear basis is defined by polynomial correction: , , in, , is the weighted term, k is the polynomial order, The problem is converted into a system of linear equations and the polynomial coefficients a and b are solved using the minimum square method.

[0015] The advantages of the present invention compared with the prior art are: The present invention can adapt to aviation blade tips of different sizes and stably extract the middle line under their different postures and positions. The method has improvements in adaptability, computational efficiency and noise resistance, and optimizes trajectory consistency through a uniform segmentation strategy, making it have a wider application potential in aviation blade tip feature extraction tasks under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a graph of the extraction results after smoothing of the aviation blade tip trajectory generation method based on structured light vision in the embodiment.

[0017] Figure 2 This is a diagram of the extraction results after uniform segmentation of the aviation blade tip trajectory generation method based on structured light vision in the embodiment.

[0018] Figure 3 This is a graph of the extraction results of the aviation blade tip trajectory generation method based on structured light vision on five types of data in the embodiment. DETAILED DESCRIPTION

[0019] The present invention will be further described below in conjunction with the embodiments, but are not intended to limit the present invention.

[0020] The method for generating the aviation blade tip trajectory based on structured light vision in this embodiment includes the following steps: The leaf tip point cloud is captured by a 3D structured light camera.

[0021] The captured point cloud data contains a large amount of outlier noise, which may come from environmental interference, sensor errors, etc. These noises will have an adverse effect on the subsequent feature extraction process. Therefore, the point cloud data needs to be further cleaned and filtered to reduce noise and unnecessary information.

[0022] The process of denoising point cloud data includes height filtering and voxel filtering. This embodiment line uses a height-based PassThrough filter to filter the target point cloud by manually setting a specific Z-axis value range according to the tip height of blades of different models. Height filtering helps data dimensionality reduction and noise suppression, and can extract effective information in a specific height area, thereby improving the accuracy and efficiency of subsequent analysis and processing algorithms.

[0023] After high filtering, there are still a lot of point clouds, which need to be downsampled and homogenized through voxel filtering. The voxel grid filter divides the space into cubic voxels, averages the points in each voxel, and generates a new point cloud. Voxel filtering reduces the resolution of the point cloud, eliminates redundant points in dense areas, reduces the number of data points, makes the point cloud more uniform, and retains the key feature information of the original point cloud, thereby improving processing speed and reducing the impact of noise.

[0024] Extract edge point cloud from the homogenized point cloud. Specifically, Define the point cloud coordinate system as , assuming that the 3D point cloud data in the original coordinate system , the total number of point clouds after voxel filtering and downsampling is ,in: , , , Select points from the point cloud of k (This embodiment selects k=60 ) nearest neighbor points, denoted as ,in .in , Yes The horizontal axis of Yes The vertical coordinate of Yes The height coordinate of .

[0025] Zero-mean the point cloud data ,in: , , , in: .

[0026] Compute the covariance matrix: ,in yes The transpose of .

[0027] By formula Compute the eigenvalues ​​and eigenvectors of the covariance matrix G, where is the eigenvalue, indicating the data variance in this direction. is the eigenvector, indicating the main distribution trend of the data in this direction, and three eigenvalues ​​are obtained and the corresponding eigenvector , agreed , then with the minimum eigenvalue The corresponding eigenvector Yes The normal vector .

[0028] Definition Point The tangent plane (i.e. For a point Each field point , the vector Projection onto the tangent plane to get the projection vector ,in" " indicates the dot product operation. All the projected vectors Convert to polar coordinates (on the tangent plane), calculate its polar angles, and let these angles be After sorting all the angles, calculate the difference between two adjacent angles and consider connecting the end to the end.

[0029] .

[0030] If there is an angle difference greater than a preset angle threshold (the threshold set in this embodiment is ), we can consider that the point is on the edge, thus finding all edge points in the homogenized point cloud.

[0031] Aviation blade tip point cloud data usually have problems such as surface mutation areas, local data missing or uneven sampling. Normal estimation alone cannot fully distinguish these boundary features. The above method can sensitively identify and locate the boundary areas on the surface where the point cloud density changes significantly, the geometric features are discontinuous or there are data gaps by analyzing the projection distribution and angle interval of the neighborhood points on the local tangent plane, thereby effectively determining the true edge point position in the point cloud.

[0032] The tip of an aviation blade has an obvious streamlined structure. No matter what position or posture it is in, PCA can always adaptively find its first principal direction, that is, the first principal direction of the blade tip represented by the edge points in the homogenized point cloud is calculated by the principal component analysis method, which is also the direction with the largest data variance in three-dimensional space. Specifically, Assume that the total number of edge point clouds is N, and the three-dimensional edge point cloud data in the original coordinate system is ,in: , , , Zero-mean the point cloud data ,in: , , , in .

[0033] Compute the covariance matrix: ,in for The transpose of .

[0034] By formula Calculate the eigenvalues ​​and eigenvectors of the covariance matrix C. The eigenvector corresponding to the largest eigenvalue is Represents the first principal direction of the data, that is, the direction in which the data variance is the largest.

[0035] Establish a new local space coordinate system so that The axis is aligned with the first principal direction calculated above to facilitate the subsequent extraction of the middle line.

[0036] Specifically, the first main direction is normalized. , Will As the new coordinate system axis.

[0037] Then select a Non-parallel reference vectors ,calculate ,in" " represents the cross product, which is further normalized: , Will As the new coordinate system axis.

[0038] Determined by the right-hand screw rule , .right Normalize ,Will As the new coordinate system axis.

[0039] Finally, the new coordinate system is, .

[0040] Construct a rotation matrix R to transform the edge point cloud from the original coordinate system to the new coordinate system.

[0041] .

[0042] The matrix R is used to transform the point cloud from the original coordinate system to the new coordinate system. The column vectors of the rotation matrix R are the three normalized unit direction vectors, that is, .

[0043] The points in the new coordinate system are transformed by the formula It is obtained, which expands to , in is the 3D point cloud data under the new coordinates, is the centroid of the point cloud: .

[0044] The middle line of the aviation blade tip is a curve, and the subsequent blade tip repair work requires the fitted middle line to pass through the front and rear endpoints of the blade tip. Therefore, the middle line is fitted by polynomial through endpoint constraints.

[0045] The point cloud in the new coordinate system is The endpoint constraint is based on the first main direction ( The starting point and the end point of the middle line correspond to the two points at the front and the rear of the first main direction. The two endpoints are: .

[0046] First, a linear interpolation function passing through the endpoints is constructed as a reference curve, which is used to ensure that the fitting curve passes through the endpoints.

[0047] , , Define the residual between point cloud data and linear basis as , in, .

[0048] The residual is corrected by polynomials, while ensuring that the residual at the endpoint is 0. , , in, , is the weighted term, k is the polynomial order.

[0049] The weighting term is used to: when or hour, , that is, it is 0 at the endpoint, ensuring that the fitting curve strictly passes through the endpoint.

[0050] When hour, , that is, the polynomial is allowed to fit the residuals freely in the intermediate region.

[0051] Then solve it by the least squares method.

[0052] Convert the problem into a system of linear equations: , in ,Right now: , in: .

[0053] , .

[0054] Expanding A gives: , Where a and b are the coefficients to be solved: .

[0055] Minimize the residual sum of squares using the least squares method and , calculate the optimal solution of polynomial coefficients a and b respectively, , , thus obtaining a higher precision fitting effect of the middle line of the point cloud under the endpoint constraints.

[0056] The point on the middle line Back to the original coordinate system, the conversion formula is: . Expanding it yields the following result: .

[0057] Since the middle line extracted by polynomial fitting is usually accompanied by certain noise and local fluctuations, it is necessary to smooth it. This process helps to effectively suppress noise and eliminate irregular fluctuations, so that the middle line is more in line with the requirements of the actual trajectory and improves the stability and availability of the trajectory. The above structure is smoothed. In this embodiment, a sliding window averaging method is specifically used, that is, for each point, the average value of the points within a certain range before and after it is taken as the coordinates of the new point.

[0058] Indicates the half-width of the window: ,parameter Specifies the size of the sliding window.

[0059] Assume that the total number of point clouds on the fitted middle line is H. For the i-th point in the middle line, , we need to extract the domain points within the window. If the i-th point , then its domain point set is: , in, The domain set representing the i-th point contains all points within the legal index range.

[0060] Calculate new smoothing points ,in , , , n is the number of valid points in the domain: After this processing, the coordinates of each point are smoothed by the average value of its surrounding area. The extraction result of the smoothed middle line is as follows Figure 1 shown.

[0061] Figure 1 (a) is the middle line in the original coordinate system. Figure 1 (b) is the position diagram of the extracted middle line and edge point cloud, indicating that the extracted middle line is located at the center of the edge point cloud in the original coordinate system. Figure 1 (c) shows the position diagram of the extracted middle line and the original point cloud, indicating that the final extracted middle line is successfully located at the center of the original point cloud. The above comparison shows that the extraction process of the middle line is highly consistent with the structure of the point cloud.

[0062] After the smoothing process is completed, the middle line contains a large number of trajectory points. If all the trajectory points are directly used for the subsequent trajectory guidance, the system's computational burden will be significantly increased, affecting the operating efficiency. In order to effectively control the computational complexity while ensuring the trajectory accuracy, the middle line needs to be downsampled using a length-based uniform segmentation strategy to keep the spacing between adjacent trajectory points consistent, thereby generating uniformly distributed and representative sampling points as the final guidance trajectory points. In terms of the selection of the number of segments, a reasonable balance must be achieved between accuracy and computational load: if the number of segments is too large, although the trajectory fitting accuracy can be improved, it will also significantly increase the amount of system calculations; conversely, if the number of segments is too small, the trajectory points may be too sparsely distributed, thus affecting the smoothness and stability of the guidance process. Therefore, a reasonable setting of the number of segments is crucial to achieving efficient and accurate trajectory guidance.

[0063] The steps of downsampling in uniform segmentation strategy include: For every two points on the smoothed middle line and , calculate the Euclidean distance between them: .

[0064] Then construct a cumulative length array ,in The calculation formula is , among which hour, ,when hour, The total length of the middle line I .

[0065] Divide the middle line evenly segments, the length of each segment is , for each segment sampling point j ( j From 0 to ), the target cumulative length on its entire line is If the cumulative length of the target sampling point is equal to the cumulative length of the existing points on the middle line, then the existing point on the middle line is the first target sampling points.

[0066] For each target cumulative length , we need to find the interval in the cumulative length array. However, the cumulative length of the target sampling point is not necessarily equal to the cumulative length of the existing points on the middle line. , is the cumulative length of the point before the target sampling point, It is the cumulative length of the point after the target sampling point. idx is k The scale factor for this segment is: , t The value range is (0,1), use t Perform linear interpolation to calculate new target sampling points.

[0067] , , , in As the middle line New target sampling points, Indicates that there is a point on the middle line i The horizontal coordinate of the point, Indicates that there is a point on the middle line i The vertical coordinate of the point, Indicates that there is a point on the middle line i The height coordinates of a point. Indicates that there is a point on the middle line The horizontal coordinate of a point, and so on.

[0068] When the target length is exactly equal to 0, the first point of the middle line is directly taken; when the target length is longer than the total length, the last point is directly taken. Through the above process, a new set of points can be resampled at equal intervals of equal length on the middle line after smoothing to achieve uniform segmentation. is 4, and finally 5 key feature points are obtained. The final result after uniform segmentation is as follows Figure 2 shown.

[0069] exist Figure 2 The five points in (a) are new target sampling points, which evenly divide the middle line into four segments. Figure 2 The five points in (b) are Figure 1 The five new target sampling points in (a) are used as trajectory points to represent the generated trajectory. Figure 2 As can be seen in (c), the generated trajectory is located at the center of the original point cloud, indicating that the generation of the trajectory is highly consistent with the geometric distribution of the original point cloud. The above comparison shows that the extracted trajectory points have high accuracy and consistency, thus ensuring the smoothness of the trajectory.

[0070] The effect of the method of the present invention was verified by experiments. The experimental data collection was divided into five categories. The first category of data was collected by using a 3D surface structured light camera to collect a one-to-one model of a real aviation blade tip, and a total of 14 groups of data were collected. The 14 groups of data were divided into two types: one was to collect a group every 45 degrees at the original position unchanged, with different postures of the aviation blade tip model, and a total of 8 groups (A1-A8) were collected. The second was to collect at any position and in any posture, and a total of 6 groups (A9-A14) were collected.

[0071] The second type of data is real aerial large leaf tips collected by 3D surface structured light camera, the third type of data is real aerial small leaf tips collected by 3D surface structured light camera, and the fourth type of data comes from real aerial large leaf tip and small leaf tip data collected by the second sensor. In order to verify the anti-noise performance of the algorithm, the fifth type of data is also real aerial large leaf tip and small leaf tip data collected by the second sensor, but it is obtained under overexposure conditions, and its exposure rate is 1.37 times higher than that of the fourth type of data.

[0072] Figure 3 The a, b, c, d, and e rows correspond to the first, second, third, fourth, and fifth categories of data, respectively. Figure 3The first column is the original point cloud, the five trajectory points in the second column evenly segment the middle line and are highly consistent with the structure of the original point cloud, and the five trajectory points in the third column are highly consistent with the structure of the original point cloud. Although the c-row data is missing part of the point cloud, the middle line can still be extracted. Due to the missing point cloud data, the data distribution may be uneven when performing polynomial fitting, resulting in insufficient fitting constraints in local areas, thereby reducing the smoothness of the middle line. The e-row data was collected under overexposure conditions, but the method of the present invention can still extract the middle line well, further illustrating that the method has good robustness.

[0073] The accuracy of the extraction results is further verified. For the same type of data, the consistency of the middle line extraction may be affected by the different acquisition positions and postures. To this end, a point cloud is selected from the same type of data, and the extracted middle line is used as the reference point cloud, while the middle lines extracted from the remaining point clouds are used as the registration point clouds. The ICP algorithm is used to rigidly register the registration point cloud with the reference point cloud, and the registration error is calculated. For the same type of data, multiple repeatability tests are performed using the ICP technology to analyze the changes in the registration errors of different point clouds to evaluate the stability and consistency of the middle line extraction. Finally, the repeatability and reliability of the algorithm are quantified by calculating the Final RMS error. The results of the first type of data are shown in Table 1.

[0074] Table 1 Final RMS error

[0075] The same operation is performed on the remaining four types of data, and the Final RMS error of each type of data is calculated respectively. Finally, the Final RMS errors of the five types of data are summarized and their Final RMS mean is calculated to evaluate the stability and consistency of the algorithm. The results are shown in Table 2.

[0076] Table 2 Final RMS error mean

[0077] There are 37 sets of data in the experiment. Excluding the case of self-registration, among the remaining 30 sets of data, the Final RMS error is greater than 0.1, accounting for 6.7%. The reason is that when extracting the middle line, the endpoint is not found accurately, causing the entire extracted middle line to deviate slightly from the center trajectory. The proportion is less than 0.1 is 93.3%. Among them, the error is between 0.05 and 0.1, accounting for 73.3%, and less than 0.05 accounts for 26.7%. It can be seen from Table 2 that even under the influence of overexposure, the mean Final RMS error of the data is still below 0.1, which shows that the method of the present invention has good robustness. Multiple repeated experimental verifications with the help of ICP technology also illustrate the stability and feasibility of the method of the present invention.

Claims

1. A method for generating aviation blade tip trajectories based on structured light vision, characterized in that: The following steps are involved: The tip point cloud of the aviation blade is captured by a 3D surface structured light camera and filtered to remove noise to obtain a homogenized point cloud; For each point in the homogenized point cloud, the principal component analysis method is used to determine the local normal and construct the tangent plane. The projection of the neighboring points of the point on the tangent plane is calculated, and the points whose angle difference of the projection polar angle exceeds the set angle threshold are selected as edge points. The principal component analysis method is used to determine the first principal direction of the blade tip represented by the edge points in the homogenized point cloud and construct a new coordinate system make Align the first principal direction, transform the edge point cloud from the original coordinate system to the new coordinate system, solve the fitted middle line in the XY and XZ dimensions using the least squares method with a polynomial with endpoint constraints, and then transform the fitted middle line back to the original coordinate system.

2. The method for generating aviation blade tip trajectories based on structured light vision according to claim 1, characterized in that: The midline is smoothed and evenly segmented after it is converted back to the original coordinates.

3. The method for generating aviation blade tip trajectories based on structured light vision according to claim 2, characterized in that: The uniform segmentation step includes first calculating the cumulative length of the entire middle line, then determining the position of each sampling point according to the total length and the target number of segments, and obtaining the coordinates of the corresponding points by linear interpolation.

4. The method for generating aviation blade tip trajectories based on structured light vision according to claim 1, characterized in that: The filtering and denoising includes first performing height filtering by setting a Z-axis value interval, and then performing voxel filtering.

5. The method for generating aviation blade tip trajectory based on structured light vision according to claim 1, characterized in that: The method of using the principal component analysis method to determine the local normal direction of each point in the homogenized point cloud and constructing the tangent plane includes selecting a point from the point cloud of k The nearest neighbor points are ,in ,in , Yes The horizontal axis of Yes The vertical coordinate of Yes The height coordinate of Zero-mean the point cloud data ,in: ， ， ， in: ; Calculate the eigenvalues ​​and eigenvectors of the covariance matrix G and get three eigenvalues and the corresponding eigenvector , and the minimum eigenvalue The corresponding eigenvector Yes The normal vector , and the normal vector The orthogonal plane is the tangent plane. ,in yes The transpose of .

6. The method for generating aviation blade tip trajectory based on structured light vision according to claim 5, characterized in that: k The value is not less than 50.

7. The method for generating aviation blade tip trajectory based on structured light vision according to claim 1, characterized in that: The difference between the first and last angles when filtering the angle difference of the projection polar angle for , represents the first projection polar angle, Indicates the last projected polar angle.

8. The method for generating aviation blade tip trajectory based on structured light vision according to claim 1, characterized in that: The principal component analysis method is used to determine the first principal direction of edge points in the homogenized point cloud, including: Assume that the 3D edge point cloud data in the original coordinate system ,in: ， ， ， Zero-mean the point cloud data ,in: ， ， ， Among them, N is the total number of edge point clouds, , Calculate the eigenvalues ​​and eigenvectors of the covariance matrix C. The eigenvector corresponding to the largest eigenvalue is represents the first main direction of the data, ,in for The transpose of .

9. The method for generating aviation blade tip trajectory based on structured light vision according to claim 1, characterized in that: When constructing a new coordinate system, select a coordinate system in space that matches Reference vectors for axes that are not parallel , Axis , for Axis unit vector, then determined by the right-hand screw rule axis, Axis .

10. The method for generating aviation blade tip trajectory based on structured light vision according to claim 1, characterized in that: The method of fitting the middle line in the XY and XZ dimensions using a polynomial with endpoint constraints by least squares method includes: Taking the extreme point on the first main direction of the blade tip as the endpoint, a linear interpolation function passing through the endpoint is constructed. ， ， The coordinates of the extreme points are ; The residual error between the point cloud data and the linear basis is defined by polynomial correction: ， ， in, , is the weighted term, k is the polynomial order, The problem is converted into a system of linear equations and the polynomial coefficients a and b are solved using the minimum square method.

Images