Large part integrated forming method based on graph neural network and laser additive manufacturing
Through the graph neural network combined with laser additive manufacturing technology, the area division and integrated forming of large parts are achieved, which solves the problems of deformation and internal stress accumulation in laser additive manufacturing, and improves the processing quality and service life of parts.
Patent Information
- Application Number
- CN202510349126.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-24
- Publication Date
- 2025-07-11
AI Technical Summary
Large parts are prone to warping and deformation and accumulation of internal stress during laser additive manufacturing, resulting in cracking or failure of parts, and it is difficult for the prior art to effectively predict and reduce deformation.
The finite element simulation is used based on graph neural network to determine the minimum stiffness area, and the minimum stiffness of the processing area is identified through the graph neural network model. Double-sided alternating deposition manufacturing technology and thermal cycling processing are used to realize the area division and integrated forming of large parts.
It effectively reduces the deformation of parts, improves processing quality and service life, reduces processing difficulty, improves efficiency and flexibility, and ensures efficient and integrated forming of parts.
Smart Images
Figure CN120286729A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of additive manufacturing, and particularly relates to an integrated forming method for large parts based on graph neural network and laser additive manufacturing. Background Technique
[0002] Laser additive manufacturing technology is a new and rapidly developing advanced manufacturing technology. For the integrated forming of large parts, using laser direct energy deposition for manufacturing is a very reasonable processing method. Compared with traditional processing technologies, it can significantly reduce production costs and improve the material utilization rate, which is particularly suitable for the current social development needs. Therefore, the integrated forming of large parts by laser additive manufacturing has been widely used in the aerospace field. However, during the process of laser additive manufacturing of large parts, warping deformation on both sides will occur. Therefore, reducing and minimizing the deformation amount is the ultimate goal of laser additive manufacturing. Since laser additive manufacturing adopts a rapid prototyping processing method, during the repeated processing of the deposited specimen, it is affected by extremely cold and hot temperatures, and large internal stress accumulations will occur inside the deposited specimen. As the height of the specimen increases during the deposition process, under the action of thermal stress, the overall part will produce large deformations, resulting in cracking or failure of the part.
[0003] In order to better integrate the forming of large metal parts by laser additive manufacturing, it is necessary to predict the deformation and strain of the parts. Dividing large parts in a reasonable way can better solve the internal stress accumulation caused by the long-term thermal cycle of the laser beam during the deposition process. Reduce the deformation caused by non-equilibrium solid-state phase transformation (thermal stress, tissue stress, and binding force). Since high-quality large metal parts cannot be directly integrally formed by laser additive manufacturing, there is an urgent need for an advanced graph neural network combined with a laser additive manufacturing method to achieve the integrated forming of large parts. Summary of the Invention
[0004] The present invention proposes an integrated forming method for large parts based on graph neural network and laser additive manufacturing to solve the problems existing in the above-mentioned prior art.
[0005] To achieve the above object, the present invention provides an integrated forming method for large parts based on graph neural network and laser additive manufacturing, including the following steps:
[0006] Perform finite element simulation on a known part to obtain the maximum deformation amount generated by the part under the action of internal stress, and determine the minimum stiffness region according to the maximum deformation amount;
[0007] Obtain the sample features of the known part, where the sample features include the part attribute graph and other influencing factors;
[0008] Training the graph neural network model with the minimum stiffness region and sample features to obtain a trained graph neural network model;
[0009] Identifying the part to be machined through the trained graph neural network model to obtain the minimum stiffness of the machining area;
[0010] Dividing the part to be machined according to the minimum stiffness of the machining area to obtain several sub-components;
[0011] Depositing the sub-components by the method of double-sided alternating deposition, clamping and positioning and additive connecting each sub-component according to the positional relationship during division to obtain a large part.
[0012] Preferably, the maximum deformation amount generated by the part under the action of internal stress includes:
[0013] Performing shape simulation on the part through an elastic finite element model, and introducing the initial temperature of the substrate, the thermal accumulation of the specimen during deposition and the internal stress caused during the machining process; setting boundary conditions and applying loads and constraints including axial, tangential and bending moments for solution, calculating the relationship between the internal stress and the bending moment generated during the machining process, and obtaining the maximum vertical displacement amount generated by the specimen during deposition.
[0014] Preferably, obtaining the part attribute graph includes:
[0015] The part attribute graph is composed of vertices, edges and an attribute matrix, and the attribute matrix is composed of attribute vectors composed of part geometric attributes;
[0016] In the part attribute graph, the machining area is corresponded to the sub-graph in the part attribute graph.
[0017] Preferably, the graph neural network model includes a graph convolutional layer, a machining area representation layer and a fully connected prediction layer;
[0018] The graph convolutional layer is used to aggregate vertex attributes to obtain vertex representation vectors;
[0019] The machining area representation layer is used to combine the vertex representation vectors to obtain geometric information, and concatenate the representation of the geometric information with the vectorized non-geometric influencing factors to obtain a complete representation vector of the machining area;
[0020] The fully connected prediction layer is used to make a prediction according to the complete representation vector to obtain the prediction result of the minimum stiffness of the machining area.
[0021] Preferably, the graph neural network model includes 4 to 5 fully connected layers, and each layer except the last prediction layer consists of a linear combination and a non-linear activation.
[0022] Preferably, the method of depositing sub-components by double-sided alternating deposition includes:
[0023] Deposit all sub-components on both sides of the same substrate respectively. After completing one layer of deposition, turn the substrate over and deposit another sub-component. When the cumulative deposition height on each side reaches 10 mm, the maximum number of consecutive deposition layers at one time reaches 10 layers.
[0024] Preferably, after the deposition of the sub-components, use wire cutting to remove the excess material, and then perform secondary processing on the sub-components. The secondary processing includes: removing the sub-components from the substrate, removing the machining allowance on the sub-components and finishing them into standard-sized sub-components.
[0025] Preferably, the method further includes: performing a thermal cycle treatment on the integral part, with a heating rate of 10 °C / min, heating to 1070 °C and holding for 60 min, cooling the specimen to 770 °C and then heating it up to 1070 °C again, with a cooling rate of 2 °C / min, and air-cooling after 4 cycles.
[0026] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the method.
[0027] The present invention also discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the method.
[0028] Compared with the prior art, the present invention has the following advantages and technical effects:
[0029] The present invention discloses an integrated forming method for large parts based on graph neural network and laser additive manufacturing, including: performing finite element simulation on a known part to obtain the maximum deformation amount generated by the part under the action of internal stress, and determining the minimum stiffness region according to the maximum deformation amount; obtaining the sample features of the known part, where the sample features include the part attribute graph and other influencing factors; training the graph neural network model through the minimum stiffness region and the sample features to obtain a trained graph neural network model; identifying the to-be-machined part through the trained graph neural network model to obtain the minimum stiffness of the machining region; dividing the to-be-machined part according to the minimum stiffness of the machining region to obtain several sub-components; depositing the sub-components by the method of double-sided alternating deposition, and clamping, positioning and additively connecting each sub-component according to the positional relationship during division to obtain a large part. The present invention uses advanced graph neural network combined with laser additive manufacturing technology to produce integrated large parts with a wide range of applications. It not only reduces the deformation caused by the accumulation of internal stress in components, but also improves the machining quality and service life of components, ensures the efficient realization of integrated forming of parts, reduces the machining difficulty, improves the machining efficiency, and has high flexibility, simplicity and easy operation. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments and descriptions thereof of this application are used to explain this application and do not constitute an improper limitation to this application. In the drawings:
[0031] Figure 1 is the flowchart of integrated additive manufacturing of large parts according to an embodiment of the present invention;
[0032] Figure 2 is the system structure diagram of the graph neural network model according to an embodiment of the present invention;
[0033] Figure 3 is the calculation process diagram of the minimum stiffness of the machining region according to an embodiment of the present invention;
[0034] Figure 4 is the schematic diagram of minimum stiffness sub-component segmentation for deformation prediction according to an embodiment of the present invention;
[0035] Figure 5 is the schematic diagram of the double-sided alternating deposition method according to an embodiment of the present invention;
[0036] Figure 6 is the schematic diagram of thermal cycle treatment according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0037] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine with the embodiments to detail this application.
[0038] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.
[0039] Example 1
[0040] As Figure 1 shown, in this embodiment, an integrated forming method for large parts based on graph neural network and laser additive manufacturing is provided, including the following steps:
[0041] Perform finite element simulation on the known part to obtain the maximum deformation amount generated by the part under the action of internal stress, and determine the minimum stiffness region according to the maximum deformation amount;
[0042] Obtain the sample features of the known part, and the sample features include the part attribute graph and other influencing factors;
[0043] Train the graph neural network model with the minimum stiffness region and the sample features to obtain a trained graph neural network model;
[0044] Identify the part to be processed through the trained graph neural network model to obtain the minimum stiffness of the processing area;
[0045] Divide the part to be processed according to the minimum stiffness of the processing area to obtain several sub-components;
[0046] Deposit the sub-components by the method of double-sided alternating deposition, and clamp, position and additively connect each sub-component according to the positional relationship during division to obtain a large part.
[0047] Furthermore, obtaining the maximum deformation includes: using an elastic finite element model in Abaqus. The part is meshed using C3D8R elements, with a global node distance of 1 mm. The mesh of the surface layer affected during deposition is refined to add machining-induced internal stresses. The measurement of deformation depends on the boundary conditions. The deformation of simple geometries is mainly caused by bending. The displacements of the left and right edges are restricted (ux = uy = uz = 0). These boundary conditions allow predicting the curvature generated during bending deformation. In the first step of the simulation, the initial temperature of the entire substrate is introduced into the deposition process of the raw material. A user-defined subroutine SIGINI is introduced as a coordinate function in all active elements. It is assumed that the initial temperature distribution across the workpiece is uniform. The initial temperature is determined using the isocontour method. In the second step of the simulation, the increase in the specimen height during the deposition experiment is added. In this model, using model change and delete interaction, all elements contained in the layer from the bottom surface (z = 0 mm) to the final surface (z = 100 mm) are deleted. This interaction allows deactivating and reactivating elements to simulate the addition of a part of the model, either temporarily or for the rest of the analysis. The thickness of the increase in the specimen height during deposition used is 0.8 mm. Finally, in the third step, the internal stresses caused by thermal accumulation during deposition are added. The machining-induced internal stresses are generated in very thin layers. Therefore, it is necessary to use a refined mesh near the machining surface. The influence of the internal stresses on the deformation is calculated, and the maximum vertical displacement amount generated by the specimen during deposition is obtained.
[0048] As Figure 3 shown, this embodiment provides a process for establishing and analyzing a finite element model. First, start with establishing a partial model, then obtain material properties and boundary conditions, and then mesh and analyze the data. Next, calculate the minimum stiffness direction of all nodes, establish a Q coordinate system, and divide the nodes into Q groups according to the minimum stiffness direction. A counter I is set in the process to start from 1, and the stiffness matrix of each group of nodes is calculated, including coordinate system transformation, combination of element stiffness matrices, construction of the global stiffness matrix, and inversion of the global compliance matrix. Finally, the stiffness of the I-th group of nodes is calculated using the flexibility method. This process is repeated until all groups are processed (i.e., I is greater than Q), and finally, the stiffness of all nodes is found and the minimum stiffness is determined, and the process ends.
[0049] Furthermore, obtaining the part property diagram includes:
[0050] The part property diagram is composed of vertices, edges, and an attribute matrix, and the attribute matrix is composed of attribute vectors consisting of part geometric attributes;
[0051] In the part property diagram, the machining area is corresponded to a sub-diagram in the part property diagram.
[0052] Specifically: The three-dimensional model of the part is regarded as a closed entity surrounded by multiple faces. Among these entities, the topological faces are interconnected. Then, the part can be represented as a graph g=(v, e), where v is a set of vertices and e is a set of edges, corresponding to the faces and edges on the part model respectively. In addition, considering that each face has geometric properties (11 in number), including normal vector, area, perimeter, etc. These geometric properties constitute an attribute vector h, and all attribute vectors form an attribute matrix H. Therefore, the part attribute graph is represented as g=(v, e, H). After the part model is represented as an attribute graph, the machining area corresponds to a subgraph in the graph, and a relationship matrix is constructed to describe the correlation between each vertex and each subgraph, where m is the number of vertices and t is the number of machining areas. Let f j be a machining area, and its vertices are represented as a set F j . Then, for a vertex v j in F i and an element r ij in the matrix R, the relationship is as follows: Finally, the part model is represented as an attribute graph, and each machining area is associated with a subgraph. Then, in the representation learning process, vectorization of geometric information is obtained based on the graph neural network model. Non-geometric influencing factors include material properties (such as elastic modulus, Poisson's ratio, and density), the area of the constrained end, and the minimum thickness of the machining area, etc., and these factors can be directly vectorized.
[0053] Furthermore, as Figure 2 shown, the graph neural network model includes a graph convolutional layer, a machining area representation layer, and a fully connected prediction layer;
[0054] The graph convolutional layer is used to perform aggregation of vertex attributes to obtain vertex representation vectors;
[0055] The machining area representation layer is used to combine the vertex representation vectors to obtain geometric information, concatenate the representation of geometric information with the vectorized non-geometric influencing factors, and obtain the complete representation vector of the machining area;
[0056] The fully connected prediction layer is used to make predictions based on the complete representation vector to obtain the prediction result of the minimum stiffness of the machining area.
[0057] In this embodiment, the graph convolutional layer mainly follows the aggregation of vertex attributes, that is, each vertex absorbs its adjacent vertices. After multiple iterations of aggregation, the representation vector of each vertex contains both geometric information and topological information. The graph attention mechanism realizes the aggregation of vertex attributes as B(·) is an activation function, n i is the set of adjacent vertices of vertex v i , a ij is the gravity coefficient of the node attributes, The learned weight matrix is the representation vector of vertex v in the l-th layer, m is the number of vertices, and F is the dimension of the attribute vector. To make the learning result more stable, after performing p aggregations simultaneously, the results are combined as i The ∪ represents the summation operation. Considering the size of the subgraph, 2 convolutional layers are required to complete it. The processing area representation layer is a single layer used to describe the complete features of the processing area. First, since each processing area is associated with a subgraph, the representations of the vertices contained in the subgraph are combined using the summation method to represent the geometric information as f f d is the geometric information representation vector of the processing area; G sub corresponds to the subgraph, and h′ j is the representation vector of vertex v j which is the output of the graph convolutional layer. Second, the representation of the geometric information is concatenated with the vectorized non-geometric influencing factors to completely represent a processing area, denoted as f = f d || f non . Where f and f non are the complete representation vector of the processing area and the vectorized non-geometric influencing factors respectively, and || is the concatenation operation.
[0058] The fully connected prediction layer is used to further extract features from the representation of the processing area, and then generate the minimum stiffness prediction. In the constructed graph neural network model, there are 4 - 5 fully connected layers. Except for the last prediction layer, each layer consists of a linear combination and a non-linear activation. Considering that the output minimum stiffness is a continuous value, the mean square error is taken as the loss function as: Where and y k are the predicted value and the labeled value respectively, and z is the total number of processing areas. In the fully connected prediction layer, further feature extraction of the processing area is performed and processed to generate predictions, and the minimum stiffness of all processing areas can be generated. Finally, the generated stiffness data can be applied to the process planning of the entire large part.
[0059] More specifically, in the processing area, the generated prediction error value of the minimum stiffness area does not exceed 5%, meeting the design usage requirements. The recognition model based on the graph neural network only needs 30 minutes to obtain the minimum stiffness of all processing areas. When the accuracy meets the engineering requirements, the efficiency is greatly improved. Therefore, the proposed method can be applied to the actual process planning of parts.
[0060] Based on the graph neural network, the attribute graph of the processing area is mapped to its minimum stiffness. By controlling the deformation of the deposited part and combining with the minimum stiffness, the part is divided into multiple sub-components. The deformation prediction of the minimum stiffness sub-components is segmented asFigure 4 as shown
[0061] Furthermore, as Figure 5 shown, the method of depositing sub-components by double-sided alternating deposition includes:
[0062] Deposit all sub-components on both sides of the same substrate respectively. To avoid bending the substrate due to thermal stress and interfering with the deposition of specimens on the other side. Once a layer of deposition is completed, the substrate needs to be flipped over for the deposition of another sub-component. When the cumulative deposition height on each side reaches 10 mm, 10 layers can be continuously deposited at one time. With the double-sided alternating deposition method, neither the substrate nor the sub-components are bent or deformed.
[0063] After the deposition of the sub-components is completed, wire cutting is used to remove the excess material, and then secondary processing is carried out on the sub-components. The secondary processing includes: removing the sub-components from the substrate, removing the machining allowance on the sub-components and finishing them into standard-sized sub-components. Initially observe whether there are pores, collapses, lack of fusion or failure to reach the design size inside each sub-component.
[0064] Clamp and position each sub-component according to the disassembled positional relationship through a fixture to prevent deformation during the integral forming of large parts.
[0065] Anneal the integral part to optimize the microstructure, reduce internal stress and improve its mechanical properties.
[0066] As Figure 6 shown, this embodiment will also perform thermal cycle treatment on the integral part. The heating rate is 10 °C / min, heat it to 1070 °C and hold for 60 min. After the specimen cools to 770 °C, continue to heat it to 1070 °C, and the cooling rate is 2 °C / min. After cycling 4 times, air cool. Optimize the microstructure, reduce internal stress and improve mechanical properties.
[0067] This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the method.
[0068] This embodiment also discloses a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the method.
[0069] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. An integrated forming method for large parts based on graph neural network and laser additive manufacturing, characterized in that, It includes the following steps: Perform finite element simulation on known parts to obtain the maximum deformation of the parts under the action of internal stress, and determine the minimum stiffness region according to the maximum deformation; Obtain the sample features of the known parts, where the sample features include the part attribute map and other influencing factors; Train the graph neural network model through the minimum stiffness region and the sample features to obtain a trained graph neural network model; Identify the parts to be processed through the trained graph neural network model to obtain the minimum stiffness of the processing region; Divide the parts to be processed according to the minimum stiffness of the processing region to obtain several sub-components; Deposit the sub-components by the method of double-sided alternating deposition, clamp and position the sub-components according to the positional relationship during division and perform additive connection to obtain large parts.
2. The method according to claim 1, characterized in that, The obtaining of the maximum deformation of the parts under the action of internal stress includes: Perform shape simulation on the parts through an elastic finite element model, and introduce the initial temperature of the substrate, the thermal accumulation of the specimen during the deposition process, and the internal stress caused during the processing process; set boundary conditions and apply loads and constraints including axial, tangential, and bending moments to solve, calculate the relationship between the internal stress and the bending moment generated during the processing process, and obtain the maximum vertical displacement of the specimen during the deposition process.
3. The method according to claim 1, characterized in that, Obtaining the part attribute map includes: The part attribute map is composed of vertices, edges, and an attribute matrix, and the attribute matrix is composed of attribute vectors composed of part geometric attributes; In the part attribute map, the processing region is corresponded to the sub-graph in the part attribute map.
4. The method according to claim 1, characterized in that, The graph neural network model includes a graph convolution layer, a processing region representation layer, and a fully connected prediction layer; The graph convolution layer is used to aggregate vertex attributes to obtain vertex representation vectors; The processing region representation layer is used to combine the vertex representation vectors to obtain geometric information, and concatenate the representation of the geometric information with the vectorized non-geometric influencing factors to obtain a complete representation vector of the processing region; The fully connected prediction layer is used to make predictions according to the complete representation vector to obtain the prediction result of the minimum stiffness of the processing region.
5. The method according to claim 4, characterized in that, The graph neural network model includes 4 to 5 fully connected layers. Except for the last prediction layer, each layer consists of a linear combination and a non-linear activation.
6. The method according to claim 1, wherein The depositing the sub-components by the method of double-sided alternating deposition includes: Deposit all sub-components on both sides of the same substrate respectively. After completing one layer of deposition, turn the substrate over and deposit another sub-component. When the cumulative deposition height on each side reaches 10 mm, the maximum number of consecutive deposition layers reaches 10 layers at a time.
7. The method according to claim 6, characterized in that, After the deposition of the sub-components is completed, use wire cutting to remove the excess material, and then perform secondary processing on the sub-components. The secondary processing includes: removing the sub-components from the substrate, removing the machining allowance on the sub-components and finishing them into standard-sized sub-components.
8. The method according to claim 1, wherein The method further includes: Perform thermal cycling treatment on the overall part, with a heating rate of 10 °C / min, heat up to 1070 °C and hold for 60 min, cool the specimen to 770 °C and then continue to heat up to 1070 °C, with a cooling rate of 2 °C / min, and perform 4 cycles and then air-cool.
9. A computer device, comprising a memory, a processor, and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, the steps of the method according to any one of claims 1-8 are implemented.