A high-fidelity finite element model automatic modeling method for blade disk structure

Abstract

The present invention relates to the field of aeroengine blade disk modeling, and specifically discloses a high-fidelity finite element model automatic modeling method for blade disk structure, including obtaining the point cloud data of the whole blade disk measured by an optical scanner; extracting the point cloud data of the blade, and distinguishing the point cloud data set of each blade by using a blade recognition algorithm; step 3, using a multi-feature clustering analysis algorithm to perform feature classification on each blade point cloud data, and respectively obtain the subdivided data sets of the blade top, pressure surface, and suction surface; step 4, using a mesh deformation algorithm to move the corresponding nodes in the standard finite element model to fit the classified subdivided data set, and using a radial basis kernel function to maintain the unit shape and mapping, and obtain a geometric detuned finite element model matching the measured whole blade disk. The method can quickly convert the subdivided scanning data into a high-fidelity finite element model, while ensuring high robustness and accuracy.

Description

Technical Field

[0001] The invention relates to the field of aeroengine blade disk modeling, and in particular to a high-fidelity finite element model automatic modeling method for a blade disk structure. Background Art

[0002] Due to factors such as processing errors, the dispersion of the material itself, and wear during use, the physical and geometric parameters of each blade in the blade disk cannot actually be completely consistent, but there are slight differences. This slight difference in the blade disk structure is called "detuning". Detuning will cause the vibration energy to be concentrated in a few sectors, which is the so-called "vibration localization" phenomenon. The actual geometric detuning in the blade disk also changes the stiffness and mass distribution of the actual model, and in some cases it will have a significant impact on the vibration characteristics. Therefore, the standard finite element model directly established according to the physical and geometric parameters is not conducive to studying the dynamic characteristics of the blade disk, and a high-fidelity model of the blade disk structure needs to be established.

[0003] In the prior art, non-contact optical measurement can measure the geometric coordinates of the entire blade disk with high precision, and use the measurement data to reconstruct the tessellated scan data (TSD) of the geometric detuned blade disk. These tessellated models can be reverse engineered into computer-aided design (CAD) models and input into mesh generation software to divide the finite element mesh for finite element analysis.

[0004] However, the process of generating a CAD model from machine scan data and then dividing the mesh cannot guarantee the accuracy of the mesh model. It is difficult to maintain consistent parameters for each manually divided mesh, and it lacks robustness. In addition, this operation is very cumbersome and time-consuming. For a single blade model, from TSD→CAD→FEM, the user needs to operate for a long time. If finite element modeling is performed on detuned blade scan data in batches, it is difficult to achieve. Therefore, it is crucial to quickly and accurately convert the segmented scan data into a high-fidelity finite element model while ensuring high robustness and accuracy for blade geometry detuning analysis. Summary of the invention

[0005] The present invention aims to provide a method for automatically modeling a finite element model of a detuned blade disk, which can automatically generate a high-fidelity finite element model through the point cloud data of the blade disk, so as to solve the technical problems in the prior art that the process of generating a CAD model from machine scanning data and then dividing the grid cannot guarantee the accuracy of the grid model, and the operation is very cumbersome and time-consuming and lacks robustness.

[0006] The high-fidelity finite element model automatic modeling method for the blade disk structure of the present invention includes:

[0007] Step 1, obtaining the point cloud data of the whole leaf disk measured by using an optical scanner;

[0008] Step 2, extracting the point cloud data of the blades from the overall blade disk point cloud data, and using the blade recognition algorithm to distinguish the point cloud data sets of each blade;

[0009] Step 3: Use a multi-feature clustering analysis algorithm to perform feature classification on each blade point cloud data to obtain segmented data sets of the blade top, pressure surface, and suction surface respectively;

[0010] Step 4: Use the mesh deformation algorithm to move the corresponding nodes in the standard finite element model to fit the classified segmented data set, and use the radial basis kernel function to maintain the unit shape and mapping to obtain a geometrically detuned finite element model that matches the measured integral blade disk.

[0011] Furthermore, in step 2, the blade shape function f disc (x) Perform Boolean operation on the point cloud data of the entire leaf disk to divide the point cloud data of the entire leaf disk into wheel disk P D With blade P B Two parts of the dataset.

[0012] Furthermore, for the blade shape function f disc (x) Set the redundancy ε0.

[0013] Furthermore, the process of distinguishing each leaf using the leaf recognition algorithm in step 2 includes: B Randomly select a point P from Bij , and find its connection point, and then use the found connection point as the starting point to find the connection point; iterate this operation until there are no more connection points. All the connection points and point P found in the above process Bij The collection is the point cloud dataset P of a single leaf Bi ;

[0014] Then from P B Remove the point cloud dataset P of the leaf Bi , and repeat the above operation in the remaining point cloud data until all leaves are distinguished.

[0015] Furthermore, step 2 also includes sorting the leaves.

[0016] Furthermore, the sorting process includes: Bi Calculate the angle between the centroid vector of each blade and the central axis of the blade disk, and sort the blades according to the size of the angle.

[0017] Furthermore, in step 3, the process of feature classification of each blade point cloud data includes:

[0018] First clustering: Using the leaf point cloud dataset P Bi The coordinates of the point P in Bi-XYZ As a feature for cluster analysis, P Bi Divided into the blade upper segment point cloud dataset P Bi-UP and the blade lower segment point cloud dataset P Bi-Bottom ;

[0019] Second clustering: Using the leaf point cloud dataset P Bi The normal vector P in Bi-UVW P Bi-UP Perform cluster analysis and extract the blade top surface point cloud dataset P Bi-TS ;

[0020] The third clustering: using the leaf point cloud dataset P Bi The normal vector P in Bi-UVW As a feature, remove P Bi-TS The leaf point cloud dataset P Bi The remaining part is clustered and the blade pressure surface point cloud dataset P is extracted. Bi-PS and the blade suction surface point cloud dataset P Bi-SS .

[0021] Furthermore, in step 4, the process of moving the edge nodes in the standard finite element model includes:

[0022] In the surface point cloud dataset P Bi-TS , P Bi-SS and P Bi-PS Find edge points in the image, extract these edge points and fit a smooth edge curve;

[0023] For the edge nodes in the standard finite element model, calculate the position parameters of each node on the edge curve where it is located, where the distance between each segment of nodes is taken as the arc length;

[0024] Taking isoparametric points in the edge curve according to the position parameters of the edge nodes, and subtracting the obtained isoparametric points from the corresponding edge nodes to obtain the displacement vectors of the edge nodes in the standard finite element model;

[0025] These displacement vectors are assigned to the neighboring nodes of the edge nodes using radial basis functions;

[0026] The corresponding nodes in the standard finite element model are moved according to the displacement vector and the assigned displacement vector.

[0027] Furthermore, the edge curve is fitted by a NURBS method.

[0028] Furthermore, in step 4, the process of moving the surface and internal nodes in the standard finite element model includes:

[0029] Obtain the normal vector of the surface node in the standard finite element model, project the node along the normal vector direction, find the subdivision surface triangle that intersects with the normal vector in the surface point cloud dataset, and calculate the distance from the node to the subdivision surface triangle;

[0030] Move each corresponding surface node along the normal vector direction according to the distance;

[0031] These distances are assigned to each internal node of the standard finite element model using radial basis functions;

[0032] Move each corresponding internal node along the normal vector direction according to this distance.

[0033] The present invention uses the point cloud data of the entire blade disk measured by an optical scanner to modify the standard blade disk finite element mesh model and directly generate a high-fidelity finite element mesh. There is no need to generate a CAD model for each measured scan data. The generated mesh model has consistent parameters to ensure model accuracy. The point cloud classification algorithm improves the accuracy and stability of the deformation algorithm, and the mesh deformation algorithm directly converts the standard finite element model into a finite element model of the actual geometric shape. While the present invention quickly and accurately converts the detuned blade scanning data into a high-fidelity finite element model, it has high robustness and accuracy, can realize batch conversion, and greatly improves efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the measurement results of geometric deviation and damage type of the engineering blade disk used in the embodiment of the present invention;

[0035] Figure 2 It is a flow chart of a method for automatically building a high-fidelity finite element model for a blade disk structure in an embodiment of the present invention;

[0036] Figure 3 It is a flow chart of the leaf disk point cloud data processing and classification algorithm in the embodiment of the present invention;

[0037] Figure 4 is a flow chart of a finite element mesh deformation algorithm in an embodiment of the present invention;

[0038] Figure 5 It is a comparison chart of the finite element calculation results and experimental data of the high-fidelity model of the engineering blade disk established in the embodiment of the present invention. DETAILED DESCRIPTION

[0039] This embodiment takes the compressor blade disk as an example to illustrate the method of the present invention. First, it is necessary to obtain the point cloud data of the entire blade disk. The specific form of the geometric deviation of each blade of the engineering blade disk to be measured is as follows: Figure 1As shown, the top of blade 3 is greatly deformed, and blades 11, 12, 14, and 15 are damaged blades with missing corners at the leading edge of the blade tip. The average deviation of the undamaged blades is less than 0.05mm, and the average deviation of the damaged blades is 0.1mm-0.25mm.

[0040] The high-fidelity finite element model automatic modeling method for the blade disk structure used in this embodiment is basically as follows: Figure 1 As shown, first, the measured whole blade disk point cloud data (TSD model) is processed, and the wheel disk point cloud data is deleted by Boolean operation. The blade recognition algorithm is used to distinguish and sort the blades, and then the point cloud data set of each blade is classified to obtain the point cloud data of the blade top, pressure surface, and suction surface.

[0041] like Figure 2 As shown, the above leaf disk point cloud processing and classification algorithm is specifically operated as follows:

[0042] (1.1) First, extract the point cloud data of a single leaf from the leaf disk point cloud data, such as Figure 2 As shown on the left, according to the blade shape function f disc (x) Perform Boolean operation on the point cloud data of the entire blade disk to divide it into two parts: the wheel disk and the blade. In order to avoid the connection between the blade and the disk due to manufacturing errors, in this embodiment, disc (x) Redundancy ε0 is also set.

[0043] Next, using the rule that there is no triangle mesh connection between different leaves, a single leaf is extracted from all the leaf point cloud data. B Randomly select a point P from Bij , and find its connection point, and then use the found connection point as the starting point to find the connection point; iterate this operation until there are no more connection points. All the connection points and point P found in the above process Bij The collection is the point cloud dataset P of a single leaf Bi ; Then from P B Remove the point cloud dataset P of the leaf Bi , and repeat the above operation in the remaining point cloud data until all leaves are distinguished.

[0044] In addition, due to P Bi is randomly selected, and it is impossible to determine P Bi The position of the extracted P Bi Needs to be sorted correctly.

[0045] Specifically, according to the point cloud dataset P of each blade Bi Calculate the angle between the centroid vector of each blade and the central axis of the blade disk, and finally calculate P BiThe angle between the centroid vector and the central axis is used to sort the blades according to the angle between 0-360° and the blades at each position.

[0046] (1.2) The surface features of the blade point cloud data are detected by clustering analysis of multiple geometric characteristics to adapt to the large deformation of the finite element mesh. The leading edge, trailing edge, pressure surface, suction surface and top surface feature points of the blade need to be identified from the blade point cloud data. Since the top of the blade is an uncertain curve, and the pressure surface and suction surface are both complex surfaces, only P is used. Bi-XYZ It is difficult to extract each face using point coordinates as features, so the normal vector feature P of the point cloud is used. Bi-UVW As the second identified geometric property.

[0047] like Figure 2 As shown on the right, three cluster analyses were performed:

[0048] The first cluster analysis takes into account the excessive bending of the blade root of the integral blade disk. The normal vector of the blade root will also affect the recognition of the top surface. Therefore, the blade point cloud dataset P is used. Bi The coordinates of the point P in Bi-XYZ As a feature for cluster analysis, P Bi Divided into the blade upper segment point cloud dataset P Bi-UP and the blade lower segment point cloud dataset P Bi-Bottom ;

[0049] Second clustering: Using the leaf point cloud dataset P Bi The normal vector P in Bi-UVW P Bi-UP Perform cluster analysis and extract the blade top surface point cloud dataset P Bi-TS ;

[0050] The third clustering: using the leaf point cloud dataset P Bi The normal vector P in Bi-UVW As a feature, remove P Bi-TS The leaf point cloud dataset P Bi The remaining part is clustered and the blade pressure surface point cloud dataset P is extracted. Bi-PS and the blade suction surface point cloud dataset P Bi-SS (1.3) After identifying the feature surface in the blade point cloud data, find its edge points in the surface point cloud data.

[0051] like Figure 2 As shown in , the three clustering analyses in this embodiment all adopt the gradient descent-K nearest neighbor algorithm (GD-KNN).

[0052] Since the points between the pressure surface and the suction surface belong to the leading edge line and the trailing edge line, these points are extracted and fitted to obtain a smooth edge curve. After fitting the edge points, the curve can be used to infer the function value without point cloud data and generate more edge parameter points, which helps to align the edge nodes of the standard finite element model. The fitting curve is obtained by the NURBS method, which is defined as a vector-valued function of one or more parameters, which transforms a multidimensional space into at least one single dimension. The Cox-de Boor recursive method is used to evaluate the basic functions of NURBS. The specific form of NURBS is:

[0053]

[0054] Where ξ is the parameter value, N i,p is the i-th order basis function, and p is the control point. Since C(0) and C(1) are the extreme points of the edge NURBS curve, it can be found by searching for the vertices of the blade edge farthest from the central axis (including the leading edge vertex and the trailing edge vertex).

[0055] After the classification is completed, the mesh deformation algorithm is used to move the nodes of the standard finite element model to fit the classified subdivided data set, and the radial basis kernel function is used to maintain the unit shape and mapping to obtain a geometrically detuned finite element model that matches the measured model. The specific process of this part is as follows Figure 3 As shown, it includes two parts: the movement of blade edge nodes and the movement of blade internal point cloud data, as follows:

[0056] (2.1) Obtain the displacement vector of the finite element edge node. To this end, it is necessary to take isoparametric points on the NURBS edge curve, and the isoparametric points correspond to the finite element nodes one by one. For the edge nodes in the standard finite element model, calculate the position parameters of each edge node on the NURBS edge curve (the starting point is 0 and the end point is 1), where the distance between each node is taken as the arc length. Then, based on the position parameters of the edge node, take isoparametric points on the edge curve, and subtract the obtained isoparametric points from the corresponding edge nodes to obtain the displacement vector of the edge node in the standard finite element model;

[0057] (2.2) When the displacement of the tip node is too large when moving the edge curve, it will cause unit deformation, surface depression and unevenness. In order to solve the above problems, after obtaining the edge node displacement vector, the radial basis function (RBF) is used to distribute these displacements to adjacent nodes, which can reduce the possibility of unit deformation. The influence between node displacements depends on the Euclidean distance between FEM nodes, so the specific form of the Gaussian kernel is:

[0058]

[0059] Where s is the hyperparameter of the Gaussian kernel, R is the Euclidean distance between two vectors (u and v in the formula), V is the variance vector, and V[x i ] is the variance calculated over the i-th component at all points.

[0060] Each edge node has its own displacement vector and will affect each other. The local displacement vector is solved by the local Gaussian kernel matrix so that the final displacement conforms to the node displacement vector. The movement of all nodes in the standard finite element model is shown in the following formula:

[0061]

[0062] Where N i-1 is the node coordinate before moving, N i is the node coordinate after movement, and K is the Gaussian kernel radial basis matrix.

[0063] (2.3) After the first move, obtain the normal vectors of all surface nodes, project the surface nodes along the normal vector direction, and find the subdivision surface triangles that intersect with the normal vector. The general idea of ​​the algorithm for determining the intersection of a ray and a triangle is to first calculate the intersection point of the ray and the plane where the triangle is located, and then determine whether the intersection point is inside the triangle, as follows:

[0064] For a ray in space, the starting point is O (finite element node), and the ray direction is D (finite element node normal vector). According to the parameter formula of the ray, any point on it (that is, the intersection point with the triangle plane) is N(t):

[0065] N(t)=O+tD (5)

[0066] When calculating the intersection of a ray and a triangular plane, we actually evaluate the parametric equation of the ray and the parametric equation of the plane together. Assuming that we know the parametric equation of the spatial triangle, we can directly calculate the intersection point by combining it with the parametric equation of the ray. For a spatial triangle composed of vertices V0, V1, and V2, assuming that any point N(t) in the triangle has the following parametric equation:

[0067] N(t)=(1-uv)V1+uV2+uV3 (6)

[0068] Here, u and v are the weights of V2 and V3, 1-uv is the weight of V1, and u>=0, v>=0, u+v<=1. Combining the ray formula (5) with the triangle parameter equation (6), we can get the following equation.

[0069] (1-uv)V1+u V2+v V3=O+tD (7)

[0070] Obviously, u, v, t are all unknowns. After sorting, we can get the following linear equations:

[0071]

[0072] Let E1=V2-V1, E2=V3-V1, T=O-V1, and according to Cramer's law, derive the following formula:

[0073]

[0074] The above algorithm needs to traverse all surface nodes and the corresponding normal vectors for calculation, and find the triangles that intersect with them. However, the number of subdivided surface triangles is huge. In order to reduce the calculation time of the intersection of rays and triangles in each step, it is necessary to find several triangles closest to the finite element node to test whether they intersect. In this embodiment, these similar triangles are selected using the nearest neighbor algorithm based on Euclidean distance.

[0075] Finally, the distances between the surface nodes and the subdivision surface triangles intersecting with them are calculated, and the corresponding surface nodes are moved along the normal vector direction according to the distances; and the radial basis function is used to assign these distances to the internal nodes of the standard finite element model, and the corresponding internal nodes are moved along the normal vector direction according to the distances.

[0076] Through the above two steps, the updated finite element model is finally obtained.

[0077] Table 1 shows the comparison between the natural frequencies of each order of the blade obtained by finite element calculation and the test data. The comparison results of the natural frequencies of the engineering blade with one bending and one twist are as follows: Figure 5 As shown in the figure, it can be seen that the relative error range of the frequency calculated by the finite element model obtained in this embodiment for the test data is 0.0133% to 0.289%, and the relative errors of 10 blade frequencies are below 0.1%. The average relative error of the natural frequency of all blade bends is 0.087632%. The relative error range of the frequency of undamaged blades in the natural frequency of one torsion of the engineering blade disk is 0.01415% to 0.328%, and the relative errors of 8 blade frequencies are below 0.1%. The average relative error of the natural frequency of one torsion of the undamaged blade is 0.09791%. Therefore, for high-fidelity structural modeling, the error between the numerical results and the measured results of the blade bend and torsion of the engineering blade disk is less than 0.1%.

[0078] Table 1 Experimental data of natural frequency of each blade of engineering blade disk and statistical data of detuning model

[0079]

[0080] Specific examples are used in this application to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is to help understand the core idea of ​​the present invention. It should be pointed out that it is obvious that those familiar with the art can easily make various modifications to these embodiments, or replace some or all of the technical features therein, and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A high-fidelity finite element model automatic modeling method for a blade disk structure, characterized in that: include Step 1, obtaining the point cloud data of the whole leaf disk measured by using an optical scanner; Step 2: extract the point cloud data of the blades in the overall blade disk point cloud data, and use the blade recognition algorithm to distinguish the point cloud data sets of each blade, including: Using the blade shape function f disc (x) Perform Boolean operation on the point cloud data of the entire leaf disk to divide the point cloud data of the entire leaf disk into wheel disk P D With blade P B Two parts of the data set; in P B Randomly select a point P from Bij , and find its connection point, and then use the found connection point as the starting point to find the connection point; iterate this operation until there are no more connection points. All the connection points and point P found in the above process Bij The collection is the point cloud dataset P of a single leaf Bi ; Then from P B Remove the point cloud dataset P of the leaf Bi , and repeat the above operation in the remaining point cloud data until all leaves are distinguished; Step 3: Use a multi-feature clustering analysis algorithm to perform feature classification on each blade point cloud data to obtain segmented data sets of the blade top, pressure surface, and suction surface, including: First clustering: Using the leaf point cloud dataset P Bi The coordinates of the point P in Bi-XYZ As a feature for cluster analysis, P Bi Divided into the blade upper segment point cloud dataset P Bi-UP and the blade lower segment point cloud dataset P Bi-Bottom ; Second clustering: Using the leaf point cloud dataset P Bi The normal vector P in Bi-UVW P Bi-UP Perform cluster analysis to extract the blade top surface point cloud dataset P Bi-TS ; The third clustering: using the leaf point cloud dataset P Bi The normal vector P in Bi-UVW As a feature, remove P Bi-TS The leaf point cloud dataset P Bi The remaining part is clustered and the blade pressure surface point cloud dataset P is extracted. Bi-PS and the blade suction surface point cloud dataset P Bi-SS ; Step 4: Use the mesh deformation algorithm to move the corresponding nodes in the standard finite element model to fit the classified segmented data set, and use the radial basis kernel function to maintain the unit shape and mapping to obtain a geometrically detuned finite element model that matches the measured integral blade disk.

2. The method according to claim 1, characterized in that The blade shape function f disc (x) Set the redundancy ε0.

3. The method according to claim 1, characterized in that Step 2 also includes sorting the leaves.

4. The method according to claim 3, characterized in that The sorting process includes: Bi Calculate the angle between the centroid vector of each blade and the central axis of the blade disk, and sort the blades according to the size of the angle.

5. The method according to claim 1, characterized in that In step 4, the process of moving the edge nodes in the standard finite element model includes: In the surface point cloud dataset P Bi-TS , P Bi-SS and P Bi-PS Find edge points in the image, extract these edge points and fit a smooth edge curve; For the edge nodes in the standard finite element model, calculate the position parameters of each node on the edge curve where it is located, where the distance between each segment of nodes is taken as the arc length; Taking isoparametric points in the edge curve according to the position parameters of the edge nodes, and subtracting the obtained isoparametric points from the corresponding edge nodes to obtain the displacement vectors of the edge nodes in the standard finite element model; These displacement vectors are assigned to the neighboring nodes of the edge nodes using radial basis functions; The corresponding nodes in the standard finite element model are moved according to the displacement vector and the assigned displacement vector.

6. The method according to claim 5, characterized in that The edge curve is fitted by the NURBS method.

7. The method according to claim 5, characterized in that In step 4, the process of moving the surface and internal nodes in the standard finite element model includes: Obtain the normal vector of the surface node in the standard finite element model, project the node along the normal vector direction, find the subdivision surface triangle that intersects with the normal vector in the surface point cloud dataset, and calculate the distance from the node to the subdivision surface triangle; Move each corresponding surface node along the normal vector direction according to the distance; These distances are assigned to each internal node of the standard finite element model using radial basis functions; Move each corresponding internal node along the normal vector direction according to this distance.

Images