Method for predicting machining deformation of thin-walled workpiece

Abstract

The invention discloses a thin-walled workpiece machining deformation prediction method. The thin-walled workpiece machining deformation prediction method comprises the following steps that S1, a time change model of the machining process of a thin-walled workpiece is established; s2, matching the cutting load point cloud with the geometric point cloud of the part; s3, a graph reconstruction process of the overall stress point cloud of the part; s4, processing deformation prediction based on depth map convolution; part geometry and actual cutting load data are modeled into a point cloud model in a three-dimensional space, the point cloud model and the point cloud model are effectively combined through point cloud matching, and accurate characterization of overall stress of the part in the cutting process is achieved. Thirdly, reconstructing the point cloud of the overall stress of the part into a graph model, taking points in the point cloud as nodes of the graph model, and taking an influence relationship between cutting loads as edges of the graph model, so as to realize simulation of geometric and stress changes of the part in the cutting process; and finally, designing a depth map convolution network, and realizing prediction of graph node attributes, namely processing deformation, by combining graph convolution operation with a multi-layer perceptron.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning-driven engineering simulation and deformation prediction, and specifically to a method for predicting the machining deformation of thin-walled parts. Background Technique

[0002] As the core carrier of lightweight design and high-performance manufacturing in the aviation industry, thin-walled structural parts are widely used in key load-bearing components such as aviation integral panels, integral frames, and beams. However, due to their low stiffness characteristics, poor machining performance of materials such as titanium alloys and composite materials, and complex geometric features, thin-walled parts are easily affected by the coupling of multiple physical fields such as cutting force, clamping force, and residual stress during cutting, resulting in difficult-to-control machining deformation. Such deformation directly affects the assembly accuracy and service performance of parts and has become a technical bottleneck restricting the high-precision manufacturing of aviation equipment.

[0003] Existing methods of using data-driven approaches to establish the relationship between workpiece deformation and its influencing factors are considered a feasible way. However, when facing the problem of thin-walled part machining deformation, data-driven machining deformation prediction methods are always troubled by issues such as the complex coupling of non-structural geometry, cutting force, clamping force, and other cutting loads of the workpiece in three-dimensional space and time sequence, and the lack of suitable data-driven methods. Secondly, existing data-driven methods are affected by the monitoring process data model, mainly establishing a time sequence model between monitoring data and machining deformation, unable to capture the spatial correlation between part geometry and cutting load during the formation process of the above machining deformation, and it is difficult to predict the machining deformation of workpieces with complex geometric shapes. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for predicting the machining deformation of thin-walled parts to solve the problems raised in the above background technique.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A method for predicting the machining deformation of thin-walled parts, the machining deformation prediction method includes the following steps: S1: Establish a geometric time-varying model of the thin-walled part machining process, model the thin-walled part and the tool as point cloud models, and simulate the material removal process through the interference detection of the tool swept volume point cloud and the workpiece point cloud to generate the geometric time-varying point cloud of the workpiece for different processes; S2: Model the external loads of cutting force and clamping force as spatio-temporal point clouds, and perform spatio-temporal registration with the part geometric point cloud to construct an overall force model of the part; S3: Reconstruct the overall force point cloud of the thin-walled part into a graph model, where the nodes represent the points in the geometric point cloud and are associated with cutting load attributes, and the edges represent the spatial interaction relationships between the nodes; S4: Extract the spatial coupling features of the graph model based on the deep graph convolutional network and predict the machining deformation through a regression function.

[0006] Preferably, the method for constructing the tool swept volume point cloud is as follows: generating the tool swept volume surface point cloud according to the tool path parameters, and determining the material removal area by the direction consistency between the workpiece point cloud normal vector and the tool swept point cloud gradient. The specific formula for the tool swept volume surface point cloud is: where, is the tool swept volume point cloud generation function, D and H are the tool diameter and the cutting edge length respectively, is the tool cutting trajectory.

[0007] Preferably, the cutting force point cloud modeling includes the spatio-temporal point cloud of spatial position, time stamp and force vector. The clamping force point cloud modeling is the point cloud with a fixed spatial position and the force vector changing with time, and is matched with the workpiece geometric point cloud through the sampling consistency registration algorithm. The spatio-temporal point cloud is specifically represented as: where, is the position at time t and

[0008] is the cutting force vector at where, U, V, W are the local coordinate system basis vectors, and d is the Euclidean distance between two points.

[0009] Preferably, the edge attributes of the graph model include the Euclidean distance between nodes and the force influence factor , and the force influence factor is determined by the spatial transmission relationship of the cutting load, where: In the formula, and are the coordinates of adjacent nodes.

[0010] Preferably, the depth graph convolutional network is an equivariant graph neural network. The convolutional layer updates the node coordinate embedding to predict the deformation displacement by aggregating the node attributes, edge attributes and global embedding information. The update rule of the convolutional layer is: In the formula, is the amount of information transfer between nodes, C is the normalization factor, and It is a multi-layer perceptron.

[0011] Preferably, the initial node embedding of the equivariant graph neural network is composed of the splicing of cutting force, clamping force attributes and global embedding information, and the global embedding includes the average force influence factor of the graph and the clustering coefficient CC, and the calculations are respectively: In the formula, is the number of triangles in the node neighborhood, is the number of neighborhood nodes.

[0012] Preferably, the workpiece point cloud in the geometric time-varying model is generated by uniform sampling of triangular meshes, and the octree-based downsampling method is used to retain key geometric features.

[0013] Preferably, the point cloud normal vector is calculated by local quadratic surface fitting, and the neighborhood radius is adaptively adjusted according to the point cloud density and local curvature. The surface equation is: In the formula, the coefficient matrix is solved by the least squares method, and the normal vector is calculated from the surface partial derivatives; The neighborhood radius of the normal vector has the calculation formula: In the formula, A is the area of the point cloud bounding box, N is the number of point clouds, is the local Gaussian curvature, and α, β are constants.

[0014] Preferably, the regression function is a multi-layer perceptron. The input is the spatial coupling features extracted by the graph convolutional network, and the output is the displacement of the graph node coordinates. The calculation process is: Among them, a is the activation function, and are network parameters, represents matrix multiplication.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The present invention is different from the machining deformation prediction that focuses on the time flow, or the simplified part geometry deformation prediction. The present invention fully considers that the machining deformation formation process of thin-walled parts is a three-dimensional space. As the material is removed during the cutting process, the part geometry and cutting load evolve in correlation and eventually cause machining deformation. This spatial characteristic is that the part geometry, cutting force and clamping force are modeled as a point cloud model to simulate the part geometry changes and the changes in the overall force of the part during the machining process. Then, a deep graph convolution model is designed that can mine the evolution law of the part machining deformation in three-dimensional space. The aggregation and transmission process of information between graph nodes and edges is used to simulate the complex effects of part geometry and cutting load in three-dimensional space during the machining process, so as to better analyze the variation law of machining deformation in three-dimensional space, and overcome the problem that the existing data-driven machining deformation method focuses on two-dimensional geometry and time series relationship mining, resulting in a large limitation on the applicability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The figure is a flow chart of the method for predicting deformation of thin-walled parts during machining according to the present invention.

[0017] Figure 2 It is a schematic diagram of the tool sweeping point cloud of the present invention.

[0018] Figure 3 It is a schematic diagram of the part point cloud model of the present invention.

[0019] Figure 4 It is a schematic diagram of the matching of the cutting load and the part geometric point cloud according to the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] See also Figures 1 to 4 The present invention provides a technical solution: a method for predicting machining deformation of thin-walled parts, the method comprising the following steps: S1: Establish a geometric time-varying model of thin-walled parts machining process; First, the thin-walled parts and the corresponding tools are modeled as point cloud models respectively. For the parts, their three-dimensional models are converted into triangular mesh models. Then, uniform sampling based on the area of ​​the triangular mesh is adopted, and the point cloud is further down-sampled to obtain the part geometry point cloud.

[0022] Tool 3D point cloud model , the three-dimensional point cloud model of the tool represents the three-dimensional geometric model of the tool by a set of discrete three-dimensional point clouds, and each point represents a sampling point on the tool surface. The three-dimensional point cloud of the tool can be expressed as: Among them, is the tool point cloud model containing P points, represents the point in the tool point cloud under the coordinate system the coordinates of the sampling point on the workpiece surface.

[0023] Furthermore, during the cutting process of the tool, a series of tool time point clouds will be formed. Combining the effective cutting positions of the tool, a tool swept volume point cloud is formed. As Figure 2 shown, that is, there are multiple sets of three-dimensional point clouds of the tool: Among them, represents the point cloud formed by the tool sweeping from the processing time t = 0 to T.

[0024] The three-dimensional point cloud model of the workpiece represents the three-dimensional geometric model of the workpiece by a set of discrete three-dimensional point clouds, and each point represents a sampling point on the workpiece surface. The three-dimensional point cloud of the workpiece can be expressed as: Among them, is the workpiece point cloud model containing m points, represents the point in the workpiece point cloud under the workpiece coordinate system the coordinates of the sampling point on the workpiece surface, corresponding to a series of sequential point clouds formed by the tool during the cutting process. The workpiece will also form a series of workpiece time point clouds during the cutting process; at the same time, considering that the same workpiece may be cut by different tools, the workpiece spatio-temporal point cloud model is constructed as: It can be seen that contains a series of three-dimensional point cloud data of the workpiece formed by cutting with n different tools. Due to the removal of materials, different tools obtain different three-dimensional point cloud models of the workpiece at different processing times t are all different.

[0025] As Figure 3 shown, through part point cloud extraction, downsampling, and normal vector calculation, the part geometric model is finally transformed into a point cloud and the corresponding normal vector model.

[0026] After that, according to the tool point cloud , let P be the tool path and S be the volume point cloud swept by the tool, represented by the set of points on the surface of the swept volume. Then, the tool swept volume point cloud can be calculated according to the following formula: Where, is the tool swept volume point cloud generation function, and H are the tool diameter and cutting edge length respectively, is the tool cutting path, and g is the tool point cloud generated by the tool parameters and the initial cutting position,

[0027] For any point in the workpiece point cloud Pw, if the normal vector corresponding to this point and the gradient vector of the tool swept volume point cloud Ss at this point point in the same direction, it means that this point is close to or within the volume swept by the tool. On the contrary, if the normal vector n of any point in the workpiece point cloud Pw points in the opposite direction to the gradient vector of Ss at this point, it means that this point is far from or outside the volume swept by the tool.

[0028] Therefore, the dot product of the normal vector n corresponding to any point in the part point cloud Pw and the gradient vector of the tool swept volume point cloud Ss at this point is used to determine whether this point is within the tool swept volume, so as to determine whether to retain or delete this point; that is, according to the part point cloud model Pw and the tool swept volume point cloud Ss, the interference between the tool and the workpiece point cloud can be detected by checking the distance between each pair of points in the two, and the geometric time-varying description of the part machining process can be realized.

[0029] S2: Matching the cutting load point cloud and the part geometric point cloud; The matching of the cutting load point cloud and the part geometric point cloud is mainly based on the workpiece geometric point cloud, and the point cloud of cutting loads such as cutting force and clamping force is matched with the part geometric model to realize the description of the overall load state of the part affecting the machining deformation.

[0030] First, extract the fast point feature histogram of the point cloud model as the point cloud model feature. Specifically, for any point cloud, assume that there is a query point in the point cloud model, at the center of a sphere with a radius of r, and at the same time all k-nearest neighbor elements, that is, the points whose distance from the point is less than the radius r are connected to each other. Then calculate the positional relationship between any two points and in the k-neighborhood, as well as the corresponding normal vectors and

[0031] Taking the point as an example, define the local coordinate system UVW, and the calculation is as follows:​​ Next, in the local coordinate system UVW, the points and and the corresponding normal vectors and The features between them adopt the fast point feature histogram calculation method, and only calculate the tuples (α, φ, θ) between the query point and its neighborhood points, and the calculation is as follows: where U, V, and W are the basis vectors of the local coordinate system, and d is the Euclidean distance between two points.

[0032] Next, the sampling consistency registration algorithm is used to match different point cloud models. Global matching is to find the corresponding relationship between the features extracted from two point cloud models. Here, still taking the workpiece geometric point cloud Pw and the cutting force point cloud The matching process is used as an example to illustrate the implementation process of the algorithm. The specific implementation steps are as follows: S11: Select s sample points from the cutting force point cloud , and at the same time ensure that the distance between the sampled points is greater than the query point neighborhood distance to ensure that the feature histograms of the sampled points are different; S12: For each sample point in the cutting force point cloud , find the points with similar features to the sampled points in the part geometric point cloud Pw, and use them as the corresponding candidate points of the sampled point in the workpiece geometric point cloud Pw. Suppose m points with similar features are obtained after searching, then the sample point and the m candidate points can be expressed as C = {ci|ci = < , pw1, pw2,…, ,…, pwm>, 1 ≤ i ≤ s, 1 ≤ j ≤ s, ∈ , ∈Pw}, and randomly select a point from the candidate point set as the corresponding point pair < , >; S13: Calculate the rigid body transformation matrix T between the sampled point and its corresponding point , and use this transformation matrix and the cutting force point cloud after this transformation to judge the current matching performance with the distance error sum function between the part geometric point cloud Pw, as Figure 4 .

[0033] S3: Graph reconstruction process of the overall force application points of the parts; If the cutting force coordinate system is consistent with the machine tool coordinate system, the resultant force is calculated based on the angle between the cutting normal and the cutting force coordinate system.

[0034] The nodes in the graph are represented by the points in the point cloud, that is, the entire workpiece geometry is abstracted into a graph. The node attributes are represented by the coordinates of the node, as well as the cutting force, clamping force, or cutting residual stress applied to it, and all node types are consistent. Here, the node coordinates, cutting force, clamping force, and cutting residual stress values are concatenated into a vector to form the graph node matrix of the spatio-temporal point cloud graph G at time t. The graph node matrix consists of nodes and the node attribute matrix: In the formula, the graph node matrix Each row in represents a set of point cloud spatial coordinates, and each row in the graph node attribute matrix represents the information such as the cutting force, clamping force, and cutting residual stress applied to the corresponding point cloud; the symbol represents the one-to-one correspondence between the rows of the matrix and the matrix . The in the matrix represents the coordinate value of the point in the workpiece point cloud, while the in the matrix , and are the cutting force, clamping force, and cutting residual stress corresponding to this point after point cloud registration.

[0035] For any edge ek ∈ ε (k = 1, 2, …, M) in the spatio-temporal point cloud graph G, according to the above analysis, it can be expressed as: and represent the distance and force influence factor between two nodes respectively.

[0036] Furthermore, can be obtained by calculating the Euclidean distance between node coordinates: In the formula, and are the coordinates of adjacent nodes; The global embedding of the graph can be expressed as: In the formula, Ge represents the global embedding of the spatio-temporal point cloud graph G, N and M are the number of nodes and the number of edges of the graph G respectively: Iavg represents the average force influence factor, and CC represents the clustering coefficient of the graph.

[0037] The average force influence factor Iavg, as one of the global embedding properties of graph G, is calculated based on the force influence factor by the following formula: The clustering coefficient CC, one of the global embedding properties of the spatio-temporal point cloud graph G, can be calculated by the following formula: In the formula, represents the clustering coefficient of node .

[0038] The clustering coefficient of a node is calculated by the ratio of the number of existing triangles in the neighborhood of the node to the total number of triangles formed in the neighborhood, as follows: In the formula, represents the number of existing triangles in the neighborhood of node , Nb(vi) represents the neighborhood of node , which is the set of all nodes adjacent to the node, represents the number of elements in the neighborhood.

[0039] In the above formula, the number of elements in the neighborhood Nb(vi) of node can be obtained through the degree d(vi) of the node. The degree d(vi) represents the number of edges associated with node , that is: From the above formula, it can be obtained that .

[0040] S4: Prediction of machining deformation based on deep graph convolution.

[0041] An equivariant graph neural network is used as the graph convolution model for machining deformation prediction. Its convolution operation is EGCL. The input of the entire network structure is the initial graph model G, and the initial node embedding is the corresponding node attribute vi∈V, together with the edge information ε = ( ). The spatial coupling feature extraction based on EGCL is as follows. Let the set of node embeddings of the first layer of EGCL be: In the formula, the initial node embedding The node attribute matrix V' formed by the vector obtained by splicing the attribute information of each node of the input spatio-temporal point cloud graph G with the global embedding information Ge of the graph G. That is, the embedding information of the entire graph can be regarded as the information of a virtual key graph node. To enable the embedding information of the entire graph to participate in the convolution operation, Ge representing the graph embedding is used as the auxiliary information of each node and spliced with the vector vi (vi ∈ V) representing the node attribute information. The coordinate embedding set is as follows: In the formula, the initial coordinate embedding is the coordinate information of the node vi in the graph G. The node embedding and coordinate embedding of the (l + 1)-th layer after the EGCL convolution can be calculated by the following formula: Among them, the convolution process of EGCL is as follows: In the formula, , , are the multi-layer perceptrons MLP for edge, node, and coordinate operations respectively; is the information transfer between node and node ; C = 1 / (N - 1) is the normalization factor; is the attribute of the edge between node and node ; Finally, an operation is performed on node , taking the aggregated information and the node embedding as inputs, and outputting the updated node embedding .

[0042] Then, these features are input into the newly built MLP, and the coordinate attributes of the graph nodes are output. The displacement of the coordinates of the same node in space is used as the processing deformation of the node. The main calculation process of the MLP is as follows: In the formula, is the initial input of the network, that is, the spatially coupled features obtained by the EGNN graph convolution; represents the output of the l-th layer; a is the activation function; is the convolution kernel of the l-th layer convolution; represents the convolution operation; is the bias of the l-th layer convolution. By training this network to predict the coordinate attributes of the graph nodes, the processing deformation is obtained.

[0043] In summary, the present invention models the part geometry and actual cutting load data as a point cloud model in three-dimensional space, and effectively combines the two through the matching of point clouds to accurately characterize the overall force on the part during the cutting process. Then, the point cloud of the overall force on the part is reconstructed into a graph model, with the points in the point cloud as the nodes of the graph model and the influence relationship between cutting loads as the edges of the graph model, thereby realizing the simulation of the geometric and force changes of the part during the cutting process. Finally, a deep graph convolutional network is designed, taking the overall force graph model of the part as the input, regarding the machining deformation prediction as the prediction of the graph node attributes, and combining graph convolutional operations with a multi-layer perceptron to realize the prediction of the graph node attributes, that is, the machining deformation.

[0044] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting deformation of thin-walled parts during machining, characterized in that: The machining deformation prediction method comprises the following steps: S1: Establish a geometric time-varying model of the thin-walled part machining process, model the thin-walled part and the tool as a point cloud model, simulate the material removal process through interference detection between the tool swept volume point cloud and the workpiece point cloud, and generate the workpiece geometric time-varying point cloud of different processes; S2: Cutting force and clamping force external loads are modeled as spatiotemporal point clouds, and spatiotemporally aligned with the part geometry point clouds to construct the overall force model of the part; S3: reconstructing the overall force point cloud of the thin-walled part into a graphical model, wherein nodes represent points in the geometric point cloud and are associated with cutting load attributes, and edges represent spatial action relationships between nodes; S4: Extract the spatial coupling features of the graph model based on a deep graph convolutional network and predict the machining deformation through a regression function.

2. A method for predicting deformation of thin-walled parts according to claim 1, characterized in that: The method for constructing the tool swept volume point cloud is as follows: generating the tool swept volume surface point cloud according to the tool trajectory parameters, and determining the material removal area by the direction consistency between the workpiece point cloud normal vector and the tool swept point cloud gradient. The specific formula of the tool swept volume surface point cloud is: in, Generate function for point cloud of tool swept volume, and H are tool diameter and cutting edge length respectively, is the tool cutting trajectory.

3. A method for predicting deformation of thin-walled parts according to claim 1, characterized in that: The cutting force point cloud modeling includes the spatiotemporal point cloud of spatial position, timestamp and force vector. The clamping force point cloud modeling is a point cloud with a fixed spatial position and a force vector that changes over time. It is matched with the workpiece geometry point cloud through a sampling consistency registration algorithm. The spatiotemporal point cloud is specifically expressed as: in, is the position at time t Cutting force vector at .

4. A method for predicting deformation of thin-walled parts according to claim 3, characterized in that: The sampling consistency registration algorithm includes: extracting the fast feature histogram of the point cloud, calculating the rigid body transformation matrix between the cutting force point cloud and the workpiece point cloud, and optimizing the matching accuracy through the distance error function, specifically including the following tuple calculations: Among them, U, V, W are the basis vectors of the local coordinate system, and d is the Euclidean distance between two points.

5. The method for predicting deformation of thin-walled parts according to claim 1, characterized in that: The edge attributes of the graph model include the Euclidean distance between nodes Impact Factor , the force influence factor is determined by the spatial transfer relationship of the cutting load, where: In the formula, and are the coordinates of the adjacent nodes.

6. A method for predicting deformation of thin-walled parts according to claim 1, characterized in that: The deep graph convolutional network is an equivariant graph neural network. The convolutional layer aggregates node attributes, edge attributes and global embedding information to update the node coordinate embedding to predict the deformation displacement. The convolutional layer update rule is: In the formula, is the amount of information transmitted between nodes, C is the normalization factor, and is a multi-layer perceptron.

7. A method for predicting machining deformation of thin-walled parts according to claim 6, characterized in that: The initial node embedding of the equivariant graph neural network is composed of the cutting force and clamping force attributes and the global embedding information, and the global embedding includes the average force influence factor of the graph And the clustering coefficient CC are calculated as: In the formula, is the number of triangles in the node neighborhood, is the number of neighboring nodes.

8. The method for predicting deformation of thin-walled parts according to claim 1, characterized in that: The workpiece point cloud in the geometric time-varying model is generated by uniform sampling of triangular meshes, and an octree-based downsampling method is used to retain key geometric features.

9. A method for predicting machining deformation of thin-walled parts according to claim 1, characterized in that: The point cloud normal vector is calculated by fitting a local quadratic surface, and the neighborhood radius is adaptively adjusted according to the point cloud density and the local curvature. The surface equation is: In the formula, the coefficient matrix Solved by the least squares method, the normal vector is calculated by the partial derivative of the surface; Normal vector neighborhood radius The calculation formula is: Where A is the area of ​​the point cloud bounding box, N is the number of point clouds, is the local Gaussian curvature, α and β are constants.

10. The method for predicting deformation of thin-walled parts according to claim 1, characterized in that: The regression function is a multi-layer perceptron, the input is the spatial coupling features extracted by the graph convolutional network, and the output is the displacement of the graph node coordinates. The calculation process is: Among them, a is the activation function, and are network parameters, Represents matrix multiplication.

Images