A Deep Learning-Based Method for Configuring Automotive Stamping Process Parameters

Abstract

This invention discloses a deep learning-based method for configuring automotive stamping process parameters, relating to the field of computer-aided design technology. The method includes the following steps: S1, parsing stamping design data and dividing the material domain; S2, associating the material domains in the stamping process and configuring allowable deformation allowances to form a material domain transfer chain; S3, constructing a stamping process graph sequence; S4, improving the StemGNN model to construct a graph Laplacian matrix and allocating invariant subspaces; S5, aligning the invariant subspaces of adjacent processes and extracting cumulative deformation components and newly added deformation components; S6, determining the deformation increment of the material domain processes, locating the first overdraft process, allowance overdraft amount, and the first overdraft invariant subspace; S7, tracing back the responsible process parameters and narrowing the allowable interval; S8, converging the material domain parameter intervals, determining the execution value, and writing it back to the stamping process, with the configuration result formed after allowance verification. This invention achieves precise cross-process configuration of automotive stamping process parameters.

Description

Technical Field

[0001] This invention relates to the field of computer-aided design technology, and in particular to a method for configuring automotive stamping process parameters based on deep learning. Background Technology

[0002] Existing automotive stamping process parameter configurations typically rely on engineering experience, trial die adjustments, or finite element simulations, involving repeated corrections of parameters such as blank holder force, stamping speed, friction coefficient, and die clearance. These methods require multiple rounds of simulation or trial production, resulting in long configuration cycles, high adjustment costs, and parameter configuration accuracy heavily influenced by personnel experience.

[0003] With the development of deep learning technology, methods for predicting forming defects and recommending process parameters based on historical stamping data have been gradually applied. However, existing methods mostly analyze single stamping processes or final forming results, making it difficult to correlate the deformation transfer process of the same sheet metal area in continuous stamping processes, and also unable to accurately distinguish between cumulative deformation from previous processes and new deformation in the current process.

[0004] Furthermore, existing methods typically cannot pinpoint the first stamping process where insufficient allowance occurs based on the allowable deformation margin of the material region. This makes it difficult to trace local deformation anomalies back to the corresponding process parameters, resulting in excessively large parameter adjustment ranges or inaccurate adjustment directions, which affects the reliability of automotive stamping process parameter configuration.

[0005] Therefore, how to provide a method for configuring automotive stamping process parameters based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based method for configuring automotive stamping process parameters. This invention combines cross-process material domain tracking with an improved StemGNN model. By constructing a material domain transfer chain and a stamping process diagram sequence, it uses complementary spectrum projection to separate cumulative deformation components from newly added deformation components. Furthermore, it assigns a fixed subspace to the first instance of deformation allowance overdraft, thus achieving process location for the first overdraft and backtracking of responsible process parameters. Based on this, it determines the execution value by considering the allowable range of the allowance overdraft shrinkage parameter and their intersection. This method offers advantages such as accurate parameter configuration, clear responsibility location, stable cross-process deformation tracking, and reduced repeated trial molding.

[0007] A method for configuring automotive stamping process parameters based on deep learning according to an embodiment of the present invention includes the following steps: S1. Analyze automotive stamping design data, divide material domains according to the initial sheet metal forming continuity, and form a set of material domains; S2. Associat material domains along the automotive stamping process and configure deformation allowances to form a material domain transfer chain; S3. Convert the material domain transfer chain into an automotive stamping process diagram, using the material domain as nodes and the deformation transfer relationship as edges, and associate process parameters to form a stamping process diagram sequence. S4. Improve the latent correlation layer of the StemGNN model to process the stamping process graph sequence, construct the graph Laplacian matrix, and allocate the invariant subspace to two StemGNN blocks through complementary spectrum projection. S5. Align the invariant subspaces of adjacent automotive stamping processes and inherit block ownership. Extract the cumulative deformation components from the first StemGNN block and extract the newly added deformation components from the second StemGNN block. S6. Superimpose the cumulative deformation components and the newly added deformation components, determine the material domain process deformation increment and deduct the allowable deformation allowance, locate the first overdraft process, the allowance overdraft amount and the first overdraft invariant subspace, and fix the first overdraft invariant subspace to the second StemGNN block. S7. Determine the responsible process parameters based on the deformation transfer relationship and process parameter correlation relationship of the material domain corresponding to the first overdraft invariant subspace, and shrink the allowable range of the responsible process parameters according to the margin overdraft amount to form the material domain parameter range; S8. Intersect the material domain parameter ranges of the same responsible process parameters, determine the execution value and write it back to the stamping process, check the remaining deformation allowance, form the configuration result when not overdrawn, and re-shrink the corresponding parameter range when overdrawn.

[0008] Optionally, S1 specifically includes: S11. Read the initial sheet metal mesh and stamping part forming mesh from the automotive stamping design data, and establish the mesh node mapping relationship according to the node inheritance number in the forming process; S12. Check the adjacency status of the mapped nodes in the stamping forming grid along the adjacent edges of the initial sheet metal grid. Group the adjacent grid cells into the same forming continuous region and separate the grid cells with interrupted adjacency. S13. Merge the initial sheet metal mesh units according to the continuous forming region, divide the material domains and form a set of material domains.

[0009] Optionally, S2 specifically includes: S21. Compare the inheritance relationship of sheet metal grid units in adjacent processes according to the automotive stamping process sequence, establish a correspondence between material domains with common inheritance units, and form a material domain process correspondence relationship. S22. Based on material properties, initial sheet metal condition and stamping quality requirements, determine the available deformation range of each material domain from the initial state to the forming limit, and configure the allowable deformation allowance. S23. Connect the material domain process correspondences along the automotive stamping process sequence. Continue the connection for material domains that retain common inheritance units, and terminate the connection for material domains that no longer have inheritance units. Write the deformation allowance into the continuously connected material domains to form a material domain transfer chain.

[0010] Optionally, S3 specifically includes: S31. According to the automotive stamping process, cut the material domain transfer chain, establish the material domains retained in each process as nodes, and maintain the corresponding nodes of the same material domain in adjacent processes. S32. Check the sheet metal mesh connection between material domains within the same process, and connect the material domain nodes that maintain the common mesh boundary as deformation transfer edges to form an automotive stamping process diagram. S33. According to the mold contact area and range of action corresponding to the process parameters, associate the process parameters with the covered material domain nodes, and arrange the automotive stamping process diagrams along the automotive stamping process to form a stamping process diagram sequence.

[0011] Optionally, the improved StemGNN model includes a latent correlation layer, a StemGNN layer, and an output layer. The StemGNN layer includes a first StemGNN block, a second StemGNN block, and residual connections. The graph Fourier transforms of the first StemGNN block and the second StemGNN block are respectively connected to complementary spectrum projection operators.

[0012] Optionally, S4 specifically includes: S41. The potential correlation layer encodes the material domain node sequence along the automotive stamping process, and the self-attention mechanism calculates the correlation weights between the material domain nodes to form a correlation weight matrix. S42. Perform symmetry and degree normalization on the relevant weight matrix to construct the normalized graph Laplace matrix, and obtain the graph Fourier basis by eigenvalue decomposition. S43. The first StemGNN block generates a backtracking output through the graph Fourier basis, calculates the residual between the material domain node sequence and the backtracking output through the residual connection, and projects the backtracking output and the residual onto the graph Laplace invariant subspace. S44. The graph Laplacian invariant subspace with backtracking output spectral energy not lower than residual spectral energy is assigned to the first StemGNN block, and the rest are assigned to the second StemGNN block. The first spectral projection operator and the second spectral projection operator are constructed respectively.

[0013] Optionally, S5 specifically includes: S51. Using the Fourier basis of the previous automotive stamping process as a reference, perform orthogonal Procrustes alignment on the Laplace invariant subspace of the current automotive stamping process to form an aligned invariant subspace. S52. Calculate the projection amount of the aligned invariant subspace within the range of the first and second spectral projection operators in the previous automotive stamping process. S53. The invariant subspace with a first projection amount not less than the second projection amount is assigned to the first StemGNN block, and the rest are assigned to the second StemGNN block to form the first invariant subspace and the second invariant subspace. S54. The first spectral projection operator projects the material domain node sequence of the current automotive stamping process onto the first invariant subspace. The first StemGNN block sequentially performs graph Fourier transform, spectral sequence unit processing, graph convolution, and inverse graph Fourier transform to form the cumulative deformation component. S55. The residual connection is formed by subtracting the backtracking output of the first StemGNN block from the material domain node sequence. S56. The second spectral projection operator projects the residual onto the second invariant subspace. The second StemGNN block sequentially performs graph Fourier transform, spectral sequence unit processing, graph convolution, and inverse graph Fourier transform to form the newly added deformed components.

[0014] Optionally, S6 specifically includes: S61. Add the newly added deformation components of the current automotive stamping process to the cumulative deformation components according to the corresponding material domain nodes, update the cumulative deformation components, and determine the newly added deformation components as the material domain process deformation increment. S62. Deduct the allowable deformation allowance sequentially along the material domain transfer chain, and determine the material domain and corresponding automotive stamping process whose remaining deformation allowance is less than zero for the first time as the first overdraft material domain and the first overdraft process. S63. The newly added deformation components of the material domain overdrafted for the first time are orthogonally decomposed according to the graph Laplace invariant subspace occupied by the second StemGNN block, and restored to the material domain node components by graph inverse Fourier transform. S64. Determine the first overdraft invariant subspace according to the residual deduction amount corresponding to the material domain node components; S65. Lock the feature vector corresponding to the first overdraft invariant subspace, and use Wielandt elimination to exclude the corresponding feature vector from the graph Fourier basis of the first StemGNN block. S66. The second spectral projection operator preserves the corresponding feature vectors, forming a graph Laplacian invariant subspace that is fixed and belongs to the second StemGNN block.

[0015] Optionally, S7 specifically includes: S71. The material domain node corresponding to the first overdraft invariant subspace that is fixedly assigned to the second StemGNN block is determined as the backtracking starting point. S72. Backtrack along the deformation transfer edge to the material domain nodes, extract the associated process parameters, and form a set of candidate responsible process parameters; S73. Take equidistant parameter values ​​around the current value of the candidate responsible process parameter in the increasing and decreasing directions, keep other process parameters unchanged, and calculate the process deformation increment of the first overdraft material domain respectively; S74. The process parameters that can reduce the deformation increment of the material domain process are identified as the responsible process parameters, and the adjustment direction for reducing the deformation increment of the material domain process is determined. S75. Divide the margin overdraft amount by the material domain process deformation increment reduction amount corresponding to the unit change of the responsible process parameter to determine the allowable interval boundary movement amount. Move the allowable interval boundary along the adjustment direction to form the material domain parameter interval.

[0016] Optionally, S8 specifically includes: S81. Collect material domain parameter intervals according to the responsible process parameters, select the maximum lower boundary and the minimum upper boundary, and form a common interval for the responsible process parameters. S82. When the current value of the responsible process parameter is within the common range, the current value is retained. When it is outside the common range, the boundary of the adjacent range is selected to form the execution value and write it back to the corresponding automotive stamping process. S83. Based on the executed value after writing back, redetermine the deformation increment of the material domain process, deduct the allowable deformation allowance along the material domain transfer chain, and form the remaining deformation allowance of the material domain. S84. When the remaining deformation allowance of all material domains is not overdrawn, the configuration result is formed. When there is overdrawn, the corresponding responsible process parameter range is shrunk, the material domain parameter range is re-converged, the execution value is determined, and the remaining deformation allowance is reviewed.

[0017] The beneficial effects of this invention are: This invention establishes a material domain transfer chain and a stamping process sequence by dividing the material into continuously traceable domains across stamping processes. This allows for continuous recording of the deformation transfer process of the same sheet metal area in different stamping processes, and by gradually deducting local deformation in each process based on allowance for deformation. Compared to methods that rely solely on single processes or final forming results, this invention can more accurately pinpoint the process where the material domain first experiences allowance overrun, improving the targeted nature of stamping forming risk identification.

[0018] Furthermore, this invention improves the StemGNN model by allocating the Graph Laplacian invariant subspace to two StemGNN blocks through complementary spectrum projection. This allows for the extraction of cumulative deformation components and newly added deformation components, respectively. The invariant subspaces are aligned and their assigned blocks are inherited between adjacent stamping processes. The invariant subspace that initially triggers excess allowance is fixedly assigned to the second StemGNN block, thus enabling continuous tracking of newly added deformation components causing local deformation anomalies and reducing judgment bias caused by the mixing of cumulative deformation and newly added deformation in the current process.

[0019] Based on this, the present invention traces back the responsible process parameters along the deformation transfer relationship of the material domain corresponding to the initial overdraft invariant subspace, shrinks the allowable range of parameters according to the margin overdraft amount, and determines the execution value through the intersection of parameter ranges of different material domains. This method can narrow the parameter adjustment range, reduce overall trial and error and repeated trial molding, and improve the accuracy, stability and efficiency of automotive stamping process parameter configuration. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a deep learning-based method for configuring automotive stamping process parameters proposed in this invention. Figure 2 This is a schematic diagram of the improved StemGNN model for a deep learning-based method for configuring automotive stamping process parameters proposed in this invention. Figure 3 This is a schematic diagram of cross-process alignment and fixed affiliation of the graph Laplace invariant subspace for a deep learning-based method for configuring automotive stamping process parameters proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A method for configuring automotive stamping process parameters based on deep learning includes the following steps: S1. Analyze automotive stamping design data, divide material domains according to the initial sheet metal forming continuity, and form a set of material domains; S2. Associat material domains along the automotive stamping process and configure deformation allowances to form a material domain transfer chain; S3. Convert the material domain transfer chain into an automotive stamping process diagram, using the material domain as nodes and the deformation transfer relationship as edges, and associate process parameters to form a stamping process diagram sequence. S4. Improve the latent correlation layer of the StemGNN model to process the stamping process graph sequence, construct the graph Laplacian matrix, and allocate the invariant subspace to two StemGNN blocks through complementary spectrum projection. S5. Align the invariant subspaces of adjacent automotive stamping processes and inherit block ownership. Extract the cumulative deformation components from the first StemGNN block and extract the newly added deformation components from the second StemGNN block. S6. Superimpose the cumulative deformation components and the newly added deformation components, determine the material domain process deformation increment and deduct the allowable deformation allowance, locate the first overdraft process, the allowance overdraft amount and the first overdraft invariant subspace, and fix the first overdraft invariant subspace to the second StemGNN block. S7. Determine the responsible process parameters based on the deformation transfer relationship and process parameter correlation relationship of the material domain corresponding to the first overdraft invariant subspace, and shrink the allowable range of the responsible process parameters according to the margin overdraft amount to form the material domain parameter range; S8. Intersect the material domain parameter ranges of the same responsible process parameters, determine the execution value and write it back to the stamping process, check the remaining deformation allowance, form the configuration result when not overdrawn, and re-shrink the corresponding parameter range when overdrawn.

[0023] In this embodiment, S1 specifically includes: The automotive stamping design data is derived from the sheet metal unfolding model, the die forming analysis model, and the finite element calculation results. The initial sheet metal mesh is taken from the discretized results of the sheet metal before stamping, and the forming mesh of the stamped part is taken from the mesh results after the forming calculation is completed. The two types of meshes adopt a unified coordinate system and a unified element type.

[0024] The node inheritance number is assigned a unique integer according to the node storage order when the initial sheet metal mesh is generated; the original number is retained when the mesh node updates its coordinates with stamping deformation, and the new node generated by local remeshing records the source element number, local interpolation position and the original node number that participated in the interpolation.

[0025] The mesh node mapping relationship is established by retrieving nodes with the same number in the formed mesh using the initial node number; when a node with the same number is missing, the source cell and interpolation weight of the new node record are read, and the initial node is mapped to the source node with the largest weight. When the source record is missing, the corresponding mesh cell is marked as a missing mapping cell.

[0026] The adjacency status check reads adjacent mesh cells one by one along the common edge of the initial sheet metal mesh; if the nodes at both ends of the common edge are still adjacent cells in the formed mesh after mapping, the adjacency is maintained; if any adjacent cell is deleted or the nodes at both ends of the common edge cannot simultaneously form the common edge of the formed mesh, the adjacency is determined to be interrupted.

[0027] Adjacent mesh cells are merged into formed continuous regions through connected components; each formed continuous region records the mesh cells, nodes, and common boundary edges it contains, and the adjacent break points serve as the boundary points between different formed continuous regions.

[0028] When some node mappings are missing in a boundary mesh cell, the length of the common edge shared by the mesh cell and the adjacent formed continuous region is accumulated, and the cell is assigned to the formed continuous region with the largest total length of shared edges; when the total length of shared edges is the same, the cell is assigned to the formed continuous region with the closest centroid distance.

[0029] After the boundary assignment is completed, the initial sheet metal mesh cells in each continuous forming area are merged into a material domain, and a unique number and corresponding mesh cell set are written for the material domain; all material domains are gathered to form a material domain set, which is used to establish the material domain correspondence across stamping processes.

[0030] In this embodiment, S2 specifically includes: The inheritance relationship of sheet metal mesh elements in adjacent automotive stamping processes is established through the initial sheet metal element number; during the finite element solution process, the original element number is retained, and the new elements generated by local remeshing are written with the source element number and area ratio to restore the inheritance correspondence between the old and new elements.

[0031] The matching of common inheritance units uses the initial sheet metal unit number contained in the material domain as the retrieval object; when the material domain of the previous process and the material domain of the current process have the same number, or when the source unit number of the current unit falls into the material domain of the previous process, the material domain process correspondence is established.

[0032] When a material domain is separated in the current process, the number of previous process units inherited by each separated material domain is counted. Each separated material domain establishes a branch connection with the raw material domain, and the inheritance ratio is obtained by dividing the area of ​​the shared inherited units by the total area of ​​the raw material domain units.

[0033] Material properties are determined by the results of sheet tensile tests and finite element material cards; yield strength, hardening parameters, initial thickness and forming limit curves are obtained from the calibration values ​​in the material cards, initial strain and rolling direction are read from the initial sheet model, and thinning rate limit and forming safety margin are read from the stamping quality specifications.

[0034] For mesh element e, the principal strain utilization rate is equal to the maximum principal strain divided by the ultimate principal strain of the forming limit curve at the current strain. When the current strain is between two curve sampling points, the ultimate principal strains of the two sampling points are used for linear interpolation. When it exceeds the curve endpoint, the adjacent endpoint value is used.

[0035] The thinning utilization rate of mesh element e is equal to the difference between the initial thickness and the current thickness divided by the initial thickness, and then divided by the allowable thinning rate given by the quality specification; the plastic strain utilization rate is equal to the equivalent plastic strain divided by the allowable equivalent plastic strain given by the material card.

[0036] The deformation utilization rate of the mesh element is taken as the maximum value among the principal strain utilization rate, thinning utilization rate and plastic strain utilization rate; the initial deformation utilization rate of the material domain is taken as the maximum value of the deformation utilization rate of all effective mesh elements in the domain.

[0037] The forming safety margin is written as a dimensionless proportion; when the quality specification uses a percentage, divide by 100. The allowable allowance for material domain deformation is equal to 1 minus the initial deformation utilization rate of the material domain minus the forming safety margin. If the result is less than 0, write 0 and mark that the initial allowance is insufficient.

[0038] When a material domain maintains a common inherited unit in the next automotive stamping process, the current remaining allowable deformation is transferred to the corresponding material domain. When a material domain separates, each separated material domain inherits the remaining allowable deformation before separation. The inheritance ratio is only used as the area weight of the branch connection and does not perform proportional reduction on the dimensionless allowance.

[0039] When a material domain no longer has a common inherited unit in the next process, the connection is terminated at the current process node; a material domain that is removed due to trimming or punching takes the process in which the unit is deleted as the connection endpoint and no longer participates in the subsequent allowance transfer.

[0040] All continuous connections, branch connections, and termination positions together form the material domain transfer chain; the material domain transfer chain retains the material domain number, process sequence, inheritance ratio, and allowable deformation, which is used to generate the node correspondence when constructing the stamping process diagram.

[0041] In this embodiment, S3 specifically includes: The material domain transfer chain is cut off at the start and end positions of the automotive stamping process; each cut-off result retains the material domain number, mesh cell set, deformation allowance, and correspondence with adjacent processes that still exist in the current process.

[0042] Material domain nodes are established from the material domains retained in the current process; when adjacent processes have a common inherited unit, the two material domain nodes are written with the same material domain number to maintain the correspondence between cross-process nodes.

[0043] When a material domain exits the current process due to trimming or punching, the material domain nodes in the previous process are retained and the cross-process connection is terminated; the exited material domain does not enter the node set of the next process, nor does it participate in deformation transfer calculation.

[0044] Starting from the identification of common edges of the common mesh boundary through the material domain mesh cells, and ending with the re-accumulation of the remaining common edge lengths and the updating of edge properties, the overall replacement is as follows: The common grid boundary is identified using the cell edge connection table in the current automotive stamping process grid; each cell edge is indexed by arranging the two endpoint node numbers in ascending order. When the same edge index exists in two different material domains, the cell edge is determined as the common grid boundary.

[0045] When there are multiple common element edges between the same pair of material domains, the lengths of each common element edge are added together to obtain the total length of the common edge, and the two material domain nodes are connected as a deformation transfer edge; the deformation transfer edge records the material domain numbers at both ends, the common edge index set, the total length of the common edge, and the connection status.

[0046] When the common edge index set is empty due to edge trimming, punching, or element deletion, the deformation transfer edge is written as disconnected; if some common edge indices are still retained, the invalid indexes are deleted and the total length of the common edge is recalculated, and the updated deformation transfer edge continues to participate in the deformation transfer between material domains.

[0047] The mold contact area is read from the finite element contact results; when a mesh element comes into contact with a punch, die, blank holder, or draw bead, the contact element and its corresponding material domain are written into the corresponding mold contact area.

[0048] The range of application of process parameters is determined by the mold components and contact areas associated with the parameters; the blank holder force is associated with the blank holder contact material domain, the drawbead resistance is associated with the drawbead contact material domain, the friction coefficient is associated with the corresponding contact surface covering material domain, and the mold clearance is associated with the relative mold surface covering material domain in the stamping direction.

[0049] When a process parameter covers a material domain node, the process parameter name, current value, and applicable process are written into the node parameter association item; when the same parameter covers multiple material domains, associations are established separately, and the parameter is not written into material domain nodes that are not within the scope of application.

[0050] After completing the association of nodes, deformation transfer edges, and process parameters, the current automotive stamping process diagram is formed; each automotive stamping process diagram is arranged according to the process execution order to form a stamping process diagram sequence for potential related layers to read.

[0051] In this embodiment, the improved StemGNN model includes a latent correlation layer, a StemGNN layer, and an output layer. The StemGNN layer includes a first StemGNN block, a second StemGNN block, and residual connections. The graph Fourier transforms of the first StemGNN block and the second StemGNN block are respectively connected to complementary spectrum projection operators. A typical StemGNN model consists of a latent correlation layer, a StemGNN layer, and an output layer connected in sequence. The StemGNN layer processes multivariate sequences through two StemGNN blocks with residual connections. Each StemGNN block sequentially performs graph Fourier transform, spectral sequence unit processing, graph convolution, and inverse graph Fourier transform.

[0052] In a typical StemGNN model, two StemGNN blocks continuously read the same Fourier space without setting exclusive processing ranges for different spectral components. When changes occur in the automotive stamping process, the same deformation components can repeatedly enter or change their affiliation between the two blocks, making it difficult to continuously distinguish between cumulative deformation components and newly added deformation components in the current process.

[0053] The improved StemGNN model connects the first spectral projection operator and the second spectral projection operator to the graph Fourier transform inputs of the first StemGNN block and the second StemGNN block, respectively; the first spectral projection operator receives the output of the latent correlation layer, and the second spectral projection operator receives the output of the residual connection.

[0054] The ranges of the first and second spectral projection operators do not overlap. The two ranges are merged to cover the graph Fourier space of the current automotive stamping process. The first StemGNN block processes the cumulative deformation components, and the second StemGNN block processes the newly added deformation components in the residuals.

[0055] The complementary spectral projection operator does not set gradient update parameters, and its matrix elements are calculated by the block assignment of the graph Fourier basis and the graph Laplace invariant subspace; the weights and biases of the latent correlation layer, spectral sequence unit, graph convolution and output layer are obtained through model training.

[0056] The model input includes a B×T×N×F node feature tensor, a B×T×N effective node mask, and a B×T×N×4 process parameter coverage mask; B represents the number of samples in a single batch, T represents the number of automotive stamping processes, N represents the number of material domain nodes, and F represents the number of node feature channels.

[0057] In one embodiment, F can be 8; the 8 channels are sequentially written with equivalent plastic strain, maximum principal strain, thickness reduction rate, contact pressure, blank holder force, drawbead resistance, friction coefficient and die clearance.

[0058] Each feature channel is first standardized using the mean and standard deviation of the training set. When the process parameters do not cover the material domain nodes, the standardized parameter channels are multiplied by the corresponding coverage mask and then written as 0. When the material domain exits the stamping part, all feature channels are multiplied by the effective node mask and then written as 0.

[0059] The latent correlation layer outputs a B×N×N correlation weight matrix in each automotive stamping process; the first and second spectral projection operators are both N×N matrices, acting on the B×N×F node feature matrix along the material domain node dimension.

[0060] The first StemGNN block outputs B×N×1 of the accumulated deformation components before the start of the current process and B×N×F of the backtracking output; the residual connection outputs B×N×F of the residual; the second StemGNN block outputs B×N×1 of the newly added deformation components in the current process.

[0061] The output layer adds the cumulative deformation component before the start of the current process to the newly added deformation component of the current process along the material domain node position to obtain the total deformation component B×N×1 at the end of the current process; the output layer does not perform mixing between different material domain nodes.

[0062] The model training data consists of automotive stamping finite element analysis records that have completed the solution convergence check and process records that have been verified by trial molds; the finite element records must meet the following requirements: the solver convergence flag is valid, there are no negative volume elements, and the energy error is within the convergence range set by the solution task.

[0063] Each training sample contains a complete stamping process sequence for the same stamped part, process parameters for each process, effective node mask for the material domain, process parameter overlay mask, and material domain deformation utilization rate; the thickness of the test die is interpolated to the material domain node through the shape function of the grid cell where the measurement point is located.

[0064] All samples are divided into training set, validation set and test set based on the combination of process parameters; in one implementation, the ratio of the three can be 70%, 15% and 15%, and all process records generated by the same parameter combination are only included in one dataset.

[0065] The cumulative deformation component label before the start of the current process is taken as the material domain deformation utilization rate at the end of the previous automotive stamping process; the newly added deformation component label in the current process is equal to the deformation utilization rate at the end of the current process minus the deformation utilization rate before the start of the current process, and the initial sheet deformation utilization rate is used as the subtraction for the first process.

[0066] Each forward computation first maps the effective node space unit as a temporary first spectral projection operator and writes the temporary second spectral projection operator as a zero matrix; the first StemGNN block performs a temporary backtracking computation and uses the temporary backtracking output and residual to determine the ownership of the invariant subspace.

[0067] After the invariant subspace assignment is completed, the first and second spectral projection operators are generated, and then the formal forward computation is performed. The cumulative deformation component loss, the newly added deformation component loss, and the backtracking loss are all calculated from the formal forward results, and the temporary forward results are not included in the loss function.

[0068] Eigenvalue decomposition, spectral energy comparison, and block attribution determination do not participate in gradient backpropagation; the weights of the latent correlation layer, spectral sequence unit, graph convolution, predicted output branch, backtracking output branch, and output layer are updated through formal forward loss.

[0069] The cumulative deformation component loss and the newly added deformation component loss are both the sum of the squares of the differences between the predicted values ​​of effective nodes and their corresponding labels, divided by the number of effective nodes; the backtracking loss is the mean square error between the backtracking output at the effective node and the input node features.

[0070] The total loss is equal to the weighted sum of the cumulative deformation component loss, the newly added deformation component loss, and the backtracking loss; in one embodiment, the three weights can be 0.4, 0.4, and 0.2, and the sum of the three weights is 1.

[0071] The trainable weights are initialized uniformly using Xavier, and all biases are initialized to 0. The model parameters are updated using the AdamW optimizer. In one implementation, the initial learning rate can be 0.001, the first moment decay coefficient can be 0.9, the second moment decay coefficient can be 0.999, the weight decay coefficient can be 0.0001, and the number of samples per batch can be 16.

[0072] If the aggregate loss does not decrease for 8 consecutive rounds, the learning rate is multiplied by 0.5, and the lower limit of the learning rate can be 0.00001. If the L2 norm of the parameter gradient exceeds 5, it is scaled by the same proportion so that the L2 norm of the scaled gradient is equal to 5.

[0073] Training ends when the relative decrease of the total loss is less than 0.0001 for 15 consecutive rounds; the relative decrease is equal to the total loss of the previous round minus the total loss of the current round, and then divided by the total loss of the previous round. Training ends synchronously when 300 rounds are reached.

[0074] After training, all trainable parameters of the validation set with the lowest total loss are saved. During model inference, the saved parameters are loaded, and temporary forward computation, complementary spectrum projection operator construction, and formal forward computation are performed for each automotive stamping process.

[0075] In this embodiment, S4 specifically includes: The potential relevance layer is composed of gated loop units and self-attention components connected together; the gated loop units read the T×N×F material domain node sequence sequentially along the automotive stamping process and retain the N×d node hidden matrix in each automotive stamping process.

[0076] For the current automotive stamping process t, the update gate is equal to the product of the current node features and the update gate input weights, the product of the previous process hidden matrix and the update gate loop weights, and the sum of the update gate biases, compressed by the Sigmoid function.

[0077] The reset gate uses the same computation path and uses independent reset gate weights and biases; the candidate hidden matrix is ​​equal to the sum of the linear mapping of the current node's features and the linear mapping after the reset gate is multiplied by the previous hidden matrix, compressed by the hyperbolic tangent function.

[0078] The current hidden matrix is ​​equal to 1 minus the result of the element-wise multiplication of the previous hidden matrix after the update gate, plus the result of the element-wise multiplication of the candidate hidden matrix after the update gate; the previous hidden matrix of the first automotive stamping process is written into an all-zero matrix.

[0079] The current hidden matrix is ​​multiplied by the query weight matrix and the key weight matrix respectively, and the corresponding biases are added to form an N×d query matrix and an N×d key matrix; the query matrix is ​​multiplied by the transpose of the key matrix and divided by the square root of d to form an N×N attention score matrix.

[0080] Before the Softmax calculation, the key positions corresponding to invalid material domain nodes are written as -109. After the Softmax calculation is completed, the entire row and column corresponding to the invalid node are multiplied by the outer product of the mask of the valid node and written as 0 to avoid the generation of uniform correlation weights in the fully masked rows.

[0081] The sum of the relevant weights of the effective node rows is normalized to 1; rows with a normalization denominator of 0 remain all zero, ultimately forming an N×N relevant weight matrix for the current automotive stamping process.

[0082] The rows and columns corresponding to invalid material domain nodes are written with masking values ​​before Softmax calculation; in one embodiment, the masking value can be -109, so that material domain nodes that exit the current process do not participate in the relevant weight allocation.

[0083] The correlation weight matrix is ​​added position by position to its own transpose matrix and then multiplied by 0.5 to form a symmetric correlation matrix; the correlation weights between the current node and itself are written in the diagonal positions, and the off-diagonal positions represent the bidirectional correlation strength between two material domain nodes.

[0084] The degree matrix is ​​an N×N diagonal matrix; the i-th diagonal element is equal to the sum of all elements in the i-th row of the symmetric correlation matrix. Invalid nodes with a degree value of 0 do not participate in the matrix inversion, and their corresponding normalization coefficients are written to 0.

[0085] The normalized graph Laplacian matrix is ​​obtained by subtracting the product of the degree matrix (power of -1 / 2), the symmetric correlation matrix, and the degree matrix (power of -1 / 2) from the identity matrix; the i-th row and i-th column of the matrix correspond to the i-th material domain node.

[0086] The normalized graph Laplace matrix is ​​a real symmetric matrix, which can be written as the product of the graph Fourier basis matrix, the eigenvalue diagonal matrix, and the graph Fourier basis transpose matrix using eigenvalue decomposition; each column of the graph Fourier basis is a unit orthogonal eigenvector.

[0087] Eigenvalues ​​with equal values ​​correspond to the same Laplace invariant subspace of the graph; when the difference between two eigenvalues ​​does not exceed 10⁻⁶ times the larger of the absolute value of the largest eigenvalue and 1, the corresponding eigenvectors are assigned to the same invariant subspace.

[0088] The effective node mask of the current automotive stamping process is written as an N×N diagonal matrix M; the temporary first spectrum projection operator takes M, and the temporary second spectrum projection operator takes the zero matrix, so that the first StemGNN block reads all effective graph Fourier spectrum directions in the temporary forward computation.

[0089] The first StemGNN block uses the temporary first spectral projection operator to generate an N×F temporary backtracking output; the current node features are subtracted from the temporary backtracking output to form an N×F temporary residual.

[0090] The k-th graph Laplace invariant subspace is represented by an N×rk basis matrix Uk; the backtracking spectral energy is equal to the sum of the squares of all elements of the matrix obtained by left multiplying the transpose of Uk by the temporary backtracking output, and the residual spectral energy is equal to the sum of the squares of all elements of the matrix obtained by left multiplying the transpose of Uk by the temporary residual.

[0091] If the backtracking spectrum energy is not lower than the residual spectrum energy, Uk is assigned to the first StemGNN block; if the backtracking spectrum energy is lower than the residual spectrum energy, Uk is assigned to the second StemGNN block.

[0092] The first spectral projection operator is equal to the sum of the products of all the basis matrices of the first invariant subspace and their respective transposes; the second spectral projection operator is equal to M minus the first spectral projection operator, ensuring that the ranges of the two operators are orthogonal and that their intersection contains only the zero vector.

[0093] The sum of the first and second spectral projection operators equals M, and the product of the two operators is zero. Each invariant subspace can only enter one StemGNN block, and the formal forward computation uses the two spectral projection operators that have been verified.

[0094] The improved internal structure changes by rewriting the inter-block division of labor in ordinary StemGNN, which relies on the implicit formation of serial residuals, into a mutually exclusive spectral domain responsibility boundary defined by the invariant subspace of the graph Laplacian. The two StemGNN blocks, instead of jointly reading the entire spectrum, occupy complementary spectral spaces respectively. This allows the first block to stably carry the cumulative deformation that can be explained by the backtracking path, while the second block concentrates on carrying the new deformation within the residual path. This eliminates the problem of repeated extraction of the same spectral component and cross-block drift from within the model, establishing a fixed spectral domain foundation for continuously tracking the source of new deformation across stamping processes.

[0095] In this embodiment, S5 specifically includes: The invariant subspace alignment of adjacent automotive stamping processes is calculated using the N-dimensional node space where the material domain node number remains unchanged; the corresponding row of the material domain exiting the current process is written with 0, and the corresponding row of the valid material domain participates in the inner product and norm calculation.

[0096] The k-th invariant subspace of the current automotive stamping process is represented by an N×r basis matrix Ut,k. The absolute value of the inner product between each column of Ut,k and all Fourier basis columns of the previous automotive stamping process is calculated, and r non-repeating preceding basis columns are selected to maximize the sum of the absolute values ​​of the matching inner products.

[0097] The selected r preorder basis columns form an N×r reference basis matrix Vt−1,k; the transpose of Ut,k is multiplied by Vt−1,k to obtain an r×r cross-correlation matrix, and the cross-correlation matrix is ​​decomposed by singular value to obtain the left singular matrix A, the singular value matrix and the right singular matrix B.

[0098] The orthogonal Procrustes rotation matrix is ​​equal to the product of the transposes of A and B; Ut,k is right-multiplied by the rotation matrix to form the initial alignment basis matrix.

[0099] The initial alignment base matrix performs a one-to-one column matching with the reference base matrix to maximize the sum of the absolute values ​​of the inner products of the corresponding columns; when the inner product of the corresponding columns after matching is less than 0, the current column is multiplied by -1 to complete the column order exchange and sign reversal correction.

[0100] The orthogonal projection matrix Qk of the alignment-invariant subspace is equal to the product of the alignment basis matrix and its transpose; the first projection quantity is equal to the square of the Frobenius norm after left multiplying the first spectral projection operator of the previous step by the alignment basis matrix, and the second projection quantity is calculated in the same way using the second spectral projection operator of the previous step.

[0101] The first projection difference is equal to Qk minus the Frobenius norm of the product of the first spectral projection operator and Qk in the previous process; the second projection difference is calculated in the same way using the second spectral projection operator in the previous process, and the Frobenius norm is equal to the square root of the sum of the squares of all elements of the matrix.

[0102] When the first projection amount is not less than the second projection amount, the alignment-invariant subspace is allocated to the first StemGNN block; when the first projection amount is less than the second projection amount, it is allocated to the second StemGNN block; when the absolute value of the difference in projection amounts does not exceed 10−8, it is treated as equal and allocated to the first StemGNN block.

[0103] After all invariant subspaces have been allocated, the products of the corresponding aligned basis matrix and its own transpose are accumulated to form the first and second spectral projection operators for the current automotive stamping process.

[0104] The spectrum sequence unit reads a continuous automotive stamping process window of length L; in one embodiment, L can be the number of complete automotive stamping processes. If there are less than L processes before the current process, all-zero node features are added to the front of the window and an invalid time mask is written.

[0105] For the q-th automotive stamping process within the window, the first spectral projection operator is left-multiplied by the N×F material domain node feature matrix, and then the projection result is left-multiplied by the aligned graph Fourier basis transpose to form the N×F first graph spectrum matrix.

[0106] The L first graph spectrum matrices within the window are arranged in the order of automobile stamping processes to form an L×N×F graph spectrum sequence; the discrete Fourier transform is performed along the L dimensions, so that each graph spectrum position and feature channel forms L complex frequency coefficients.

[0107] The real and imaginary parts of the complex frequency coefficients are respectively entered into a parameter-independent one-dimensional convolution; each frequency coefficient in the convolution kernel is multiplied by the convolution weights and the bias is accumulated to form the content channel and the gate channel.

[0108] After being compressed by the Sigmoid function, the gated channel is multiplied element-wise with the content channel; the real part and the imaginary part are combined to form the complex frequency coefficients, and the inverse discrete Fourier transform restores the complex frequency coefficients to the L×N×C spectral sequence characteristics.

[0109] Graph convolution configures a trainable spectral kernel for each input and output channel; each graph spectral position is multiplied by the corresponding spectral kernel and accumulated along the input channel, and the graph inverse Fourier transform restores the accumulated result to the hidden features of L×N×C nodes.

[0110] The first StemGNN block prediction output branch multiplies the N×C hidden features of the current process with the C×1 prediction weights and adds the bias, then multiplies it by the first spectral projection operator to form the N×1 cumulative deformation components before the start of the current process.

[0111] The first StemGNN block backtracking output branch multiplies the N×C hidden features with the C×F backtracking weights and adds the bias, then multiplies them by the first spectral projection operator to form the N×F backtracking output; the current node features are subtracted from the backtracking output to form the N×F residual.

[0112] After left-multiplying the residual by the second spectral projection operator, N×C hidden features are generated using the same window arrangement, graph Fourier transform, spectral sequence unit, graph convolution, and graph Fourier inverse transform path.

[0113] The second StemGNN block prediction output branch maps the N×C hidden features to an N×1 initial new output, which is then multiplied by the second spectral projection operator to form a new deformed component that is only located within the value range of the second spectral projection operator.

[0114] This internal improvement changes the temporary division of labor between StemGNN blocks, which relies on the magnitude of a single residual, to a spectral space identity division of labor that is continuously inherited across automotive stamping processes. The cumulative deformation channel retains its original computational responsibilities after graph Fourier basis rotation, column order exchange, and sign flipping. New deformations are generated only in the orthogonal complement space that the cumulative channel cannot interpret, enabling the model to obtain stable and traceable inter-block responsibility boundaries and preventing historical cumulative deformations from being reinterpreted as new deformations in the current process due to spectral basis drift.

[0115] In this embodiment, S6 specifically includes: Both the cumulative deformation components and the newly added deformation components adopt N×1 node vectors; the i-th row of the two vectors corresponds to the same material domain number, and the material domain that exits the current automotive stamping process is written with 0 in the corresponding row.

[0116] The newly added deformation components of the current automotive stamping process are added row by row to the cumulative deformation components of the previous automotive stamping process; the sum of these components forms the updated cumulative deformation components, and the newly added deformation components are directly used as the process deformation increment of the material domain in this process.

[0117] When the deformation increment of a material domain process enters the material domain transfer chain, it is deducted sequentially according to the execution order of the stamping process; the initial allowable deformation of each material domain is used as the chain head balance, and the remaining deformation allowance of the material domain minus the deformation increment of the process in this material domain is written to the corresponding material domain node.

[0118] When a material domain branches, each branch reads the allocated deformation allowance and deducts it independently; when a material domain exits the stamping part, the deduction stops, and the last material domain node before exiting retains the corresponding remaining deformation allowance.

[0119] If the remaining deformation allowance in the material domain is not less than 0, continue reading the next automotive stamping process; if the remaining deformation allowance in the material domain is less than 0 for the first time, the corresponding material domain will be determined as the first overdraft material domain, and the automotive stamping process in which this deduction occurs will be determined as the first overdraft process.

[0120] The allowance overdraft amount is taken as the absolute value of the remaining deformation allowance of the material domain; when multiple material domains are overdrafted for the first time in the same automotive stamping process, the allowance overdraft amount of each material domain is retained separately, and the overdraft amounts of different material domains are not averaged or merged.

[0121] The k-th graph Laplacian invariant subspace occupied by the second StemGNN block is represented by an N×rk basis matrix Uk; the transpose of Uk is multiplied by the N×1 newly added deformed components to obtain rk×1 spectral coefficients, and then multiplied by the spectral coefficients to obtain N×1 subspace node components.

[0122] The newly added deformed components are recovered by adding all the node components of the second invariant subspace row by row. When the difference between the recovered result and the Frobenius norm of the newly added deformed components exceeds 10⁻⁶, the second block of all basis matrices is re-orthogonally normalized and decomposed again.

[0123] When the node component of the row corresponding to the first overdraft of the material domain in the k-th invariant subspace is greater than 0, this value is recorded as the corresponding surplus deduction; when the node component is not greater than 0, it is recorded as 0.

[0124] After removing the node component of the k-th invariant subspace, if the process deformation increment of the first overdraft material domain no longer exceeds the remaining allowable deformation margin before deduction, the k-th invariant subspace is marked as a necessary overdraft subspace; if there are multiple necessary overdraft subspaces, the subspace with the largest margin deduction is selected.

[0125] If there is no necessary overdraft subspace that satisfies the removal criterion alone, the subspace with the largest remaining amount deduction is selected as the first overdraft invariant subspace; if the maximum values ​​are the same, the subspace with the smaller graph Laplace eigenvalue is selected.

[0126] The first overdraft of all eigenvectors of the invariant subspace forms the basis matrix U* of the locked invariant subspace; after orthogonal normalization, U* satisfies that the product of the transpose of U* and U* is equal to the r*-order identity matrix.

[0127] The invariant subspace ownership record table is written with the first overdraft process, the first overdraft material domain, the locked invariant subspace basis matrix, the corresponding eigenvalue, the second StemGNN block ownership flag, and the locking status; this record serves as the fixed state input for the subspace alignment of the next automotive stamping process.

[0128] After the next automotive stamping process completes Procrustes alignment, the locked invariant subspace is mapped to the current locked basis matrix U*′; when the current locked basis matrix is ​​ready to enter the first StemGNN block, Wielandt elimination is performed on the first block's temporary graph Laplacian spectrum operator.

[0129] The first temporary graph Laplacian spectral operator is subtracted from the product of U*′, the corresponding eigenvalue diagonal matrix, and the transpose of U*′; the first spectral projection operator is reconstructed in the orthogonal complement space of the current locked basis matrix; the second spectral projection operator is equal to the effective node space unit map minus the first spectral projection operator.

[0130] The product of the first spectral projection operator and U*′ must be a zero matrix, and the product of the second spectral projection operator and U*′ must be equal to U*′; the fixed assignment is completed when the Frobenius norm of both errors does not exceed 10−6.

[0131] This improvement changes the lifecycle of the spectral domain responsibilities of the two residual blocks in StemGNN; the second StemGNN block no longer receives only the instantaneous residual generated by a single forward computation, but continuously occupies the complete invariant subspace that has already caused the deformation margin to be overdrawn. The model output can reversely change the reachable spectral space of the next automotive stamping process, so that the newly added deformation that causes the risk still maintains the same computational identity after the graph Fourier basis rotation and process diagram changes.

[0132] This part changes the original StemGNN residual branch from a one-time supplementary prediction channel to a responsibility state channel that can occupy, lock, and continuously retain the hazard spectrum direction; the first overdraft no longer stays at the output end to form an abnormal value, but reverses the spectral space that the two StemGNN blocks can reach, so that the first StemGNN block loses the hazard spectrum direction and the second StemGNN block exclusively occupies the hazard spectrum direction, establishing a closed loop of the prediction result reverse constraint feature propagation path from within the model, and ensuring that the new deformation component that causes the surplus overdraft maintains a single affiliation and traceable identity in the cross-process recursion.

[0133] In this embodiment, S7 specifically includes: Multiply the locked invariant subspace basis matrix by its transpose and then multiply it by the newly added deformation component to obtain N×1 locked subspace node components; material domain nodes with node components greater than 0 are written into the backtracking queue, and the node components are used as the backtracking contribution values ​​of the corresponding nodes.

[0134] During the initial overdraft process, the backtracking queue only expands along the deformation transfer edge with a valid connection state; when the locked subspace node component of an adjacent material domain node is greater than 0, it is added to the queue, and when the node component is not greater than 0, it is no longer expanded along that node.

[0135] After completing the queue expansion of the current automotive stamping process, the inherited nodes of each queue node in the previous process are read through the material domain transfer chain; the inherited nodes enter the backtracking queue of the previous process and repeat the deformation transfer edge expansion until the starting point of the material domain transfer chain, the material domain exit position, or the queue is empty.

[0136] Read the parameter association items of the material domain node within the traceback range; when the parameter's action process is no later than the first overdraft process and the parameter's action range covers the current traceback material domain, write the parameter name, action process, associated mold component, current value, and original allowed range of the responsible process parameter into the candidate set.

[0137] The original allowable range of the responsible process parameters is read from the stamping process card, the allowable adjustment range of the mold, and the executable range of the process equipment; the lower boundary of the range is the maximum value among the three types of lower limits, and the upper boundary of the range is the minimum value among the three types of upper limits.

[0138] The parameter detection step size is 5% of the original allowable range width of the responsible process parameter; when the detection step size is less than the minimum adjustment resolution of the process equipment, the minimum adjustment resolution is used, and both increasing and decreasing the detection value are limited to the original allowable range of the responsible process parameter.

[0139] Each detection changes only one candidate responsible process parameter; the detection value is first standardized using the mean and standard deviation of the corresponding channel in the training set, and then written into the material domain node parameter channel with parameter overlay mask 1, while other parameter channels retain their current values.

[0140] The model calculates the material domain process deformation increment corresponding to increasing the detection value, decreasing the detection value, and the current value, respectively; the root mean square error of the newly added deformation component in the validation set is recorded as the model judgment tolerance, and it is determined to be an effective reduction only when the reduction after detection is greater than the model judgment tolerance.

[0141] If neither direction meets the model's judgment tolerance, the corresponding candidate responsible process parameter is removed; if both directions are effective, the direction with the larger reduction in material domain process deformation increment is selected; if the reduction is the same, the direction closer to the center of the original allowable interval of the responsible process parameter is selected.

[0142] The reduction in unit parameter change is equal to the current material domain process deformation increment minus the material domain process deformation increment after detection, and then divided by the absolute value of the difference between the detected value and the current value; when the reduction in unit parameter change is not greater than 0, the corresponding parameter is removed.

[0143] The allowable range boundary movement is equal to the margin overdraft divided by the unit parameter change reduction; in the increasing direction, the lower boundary of the material domain parameter range is moved to the sum of the current value and the boundary movement, and in the decreasing direction, the upper boundary is moved to the difference between the current value and the boundary movement.

[0144] After the boundary is moved, it is first restricted to the original allowable range of the responsible process parameters, and then corrected to the executable scale of the process equipment. When the lower boundary after correction is greater than the upper boundary, the corresponding parameter is removed, and the remaining parameters form the material domain parameter range.

[0145] In this embodiment, S8 specifically includes: The material domain parameter range is grouped by the name of the responsible process parameter, the action process, and the associated mold component; for each grouped item, the maximum value among all lower boundaries is taken as the common lower boundary, and the minimum value among all upper boundaries is taken as the common upper boundary.

[0146] The common lower boundary is corrected upwards to the nearest executable scale of the process equipment, and the common upper boundary is corrected downwards to the nearest executable scale of the process equipment; when the corrected common lower boundary is not greater than the common upper boundary, a common interval of responsible process parameters is formed.

[0147] When the common lower boundary is greater than the common upper boundary, a conflict is formed; the material domains that generate the maximum lower boundary and the minimum upper boundary are written into the conflict set, and the conflict interval is obtained by subtracting the common upper boundary from the common lower boundary.

[0148] The conflict material domain reads other responsible process parameters that have been tested but have not yet assumed the responsibility for interval contraction; the margin overdraft amount is divided by the unit parameter change reduction to obtain the expected movement amount, and the expected movement amount is then divided by the original allowable interval width of the responsible process parameter to obtain the replacement adjustment ratio.

[0149] Each time, the responsible process parameter with the smallest substitution adjustment percentage is selected to handle the conflicting material domain; when the substitution adjustment percentages are the same, the parameter with the larger reduction in unit parameter change is selected; if they are still the same, the parameter with the highest parameter name is selected.

[0150] The conflicting material domain is removed from the original collection item, and the parameter interval of the new material domain is recalculated using the adjustment direction and the reduction in unit parameter change that have already been detected; the original collection item and the new collection item are recalculated for their common interval respectively.

[0151] If no other valid responsible process parameters exist in the conflicting material domain, mark the current parameter combination as unconfigurable and record the conflicting material domain, conflict interval, and removed parameters.

[0152] When the current value of the responsible process parameter is within the common range, the current value is retained; when the current value is less than the common lower boundary, the common lower boundary is taken; when the current value is greater than the common upper boundary, the common upper boundary is taken. The resulting value is corrected to the executable scale of the process equipment to form the executed value.

[0153] After all execution values ​​are written back, the earliest automotive stamping process with parameter changes is read; the gated loop hidden matrix, cumulative deformation components, invariant subspace attribution record table, and material domain remaining deformation allowance at the end of the previous process are retained.

[0154] Starting from the earliest automotive stamping process where parameters change, the gated cyclic hidden matrix, relevant weight matrix, graph Laplacian matrix, complementary spectrum projection operator, cumulative deformation component, and newly added deformation component are recalculated; the locked invariant subspace continues to maintain the affiliation of the second StemGNN block through Procrustes alignment.

[0155] The deformation increment of the re-obtained material domain process is deducted step by step along the material domain transfer chain; the loop ends and the configuration result is output when the remaining deformation margin of all material domains is not less than 0.

[0156] When there is still a margin of overdraft, the additional boundary movement is equal to the remaining overdraft divided by the current unit parameter change reduction; the additional boundary movement is corrected upward to an integer multiple of the minimum adjustment resolution of the process equipment, and the corresponding interval boundary is moved along the already determined effective adjustment direction.

[0157] All responsible process parameter execution values ​​generated in each round are arranged in a fixed parameter order as an execution value vector; the execution value vector, the remaining deformation margin of the material domain, and the boundaries of each interval are written into the iteration record table.

[0158] When the new execution value vector already exists in the iteration record table, other responsible process parameters with shrinkage space are selected in descending order of the reduction amount of unit parameter change, and the corresponding boundary is moved to the minimum adjustment resolution of one process device.

[0159] The number of executable values ​​of the responsible process parameter j within the original allowable range is equal to the range width divided by the minimum adjustment resolution of the process equipment, rounded down, and then added by 1; the total number of executable combinations of all responsible process parameters is equal to the product of the number of executable values ​​of each parameter.

[0160] The loop ends when all remaining deformation margins in all material domains are not less than 0, all responsible process parameters have no remaining shrinkage space, the execution value vector cannot generate untested combinations, or the number of iteration records reaches the total number of executable combinations.

[0161] If there is still a margin of overdraft at the end of the cycle, the output parameter range is infeasible, and the first overdraft process, the overdraft material domain, the remaining overdraft amount, the executed value vector that has been verified, and the responsible process parameters that have reached the original allowable boundary are recorded.

[0162] Example 1: To verify the feasibility of this invention in practice, it was applied to the stamping process parameter configuration of a side panel of an automotive bodywork. The side panel uses DC06 steel sheet with a thickness of 0.8mm, which undergoes drawing, trimming and punching, flanging, and forming processes in sequence. After flanging, thinning concentration and microcracks are prone to occur in the corner area of ​​the door frame. After reading the initial sheet metal mesh and the forming mesh of each process, the mesh node mapping relationship of adjacent processes is established according to the node inheritance number. Then, the mapping nodes are checked along the adjacent edges of the initial sheet metal mesh to see if they remain adjacent. The continuously inherited and adjacent mesh units are grouped into the same material domain, forming a total of 326 material domains. The target material domain at the corner of the door frame consists of 136 shell units, and its equivalent plastic strain corresponding to the material forming limit is 0.312. The initial equivalent plastic strain is 0.018. Considering a mass safety margin of 0.020, the allowable deformation margin of this material domain is configured as 0.274. Based on the node inheritance relationship, the material domain maintains a continuous connection between the drawing, trimming and punching, flanging and shaping processes, thus forming a material domain transfer chain that can transfer the allowable deformation allowance across processes.

[0163] Material domains in each process are established as nodes. Nodes with common mesh boundaries are connected as deformation transfer edges. Blank holder force, drawbead resistance, friction coefficient, and flanging clearance are associated with the material domain nodes within the corresponding die contact area. A single process diagram includes 326 material domain nodes and 714 deformation transfer edges. The node sequence records equivalent plastic strain, thickness reduction rate, contact pressure, and curvature changes. A latent correlation layer calculates the correlation weights between nodes and constructs a normalized graph Laplacian matrix. After eigenvalue decomposition, a graph Fourier basis is obtained. A residual is formed between the backtracking output of the first StemGNN block and the material domain node sequence. Based on the spectral energy of the backtracking output and the residual in each invariant subspace, the invariant subspaces are assigned to the first StemGNN block and the second StemGNN block, respectively. The invariant subspaces of adjacent processes inherit their original block affiliation after orthogonal Procrustes alignment. The first StemGNN block outputs the cumulative deformation component, and the second StemGNN block outputs the newly added deformation component. The deformation increments of the target material domain in the drawing, trimming, punching, and flanging processes are 0.151, 0.007, and 0.127, respectively. The remaining deformation allowances after deduction are 0.123, 0.116, and -0.011, respectively. Therefore, the flanging process is identified as the first overdraft process, with an allowance overdraft of 0.011. After orthogonal decomposition and inverse graph Fourier transform of the invariant subspace occupied by the second StemGNN block, the allowance deduction corresponding to the node of the target material domain in an invariant subspace is 0.014, which exceeds the allowance overdraft. Therefore, it is identified as the first overdraft invariant subspace and is fixedly assigned to the second StemGNN block.

[0164] Starting from the target material domain node corresponding to the first overdraft of the invariant subspace, and following the deformation transfer edge and the correlation with process parameters, three candidate parameters are obtained: flanging gap, friction coefficient, and drawing blank holder force. Keeping other parameters constant, changing the flanging gap from 0.82 mm to 0.80 mm and 0.84 mm respectively results in flanging deformation increments of 0.134 and 0.116 for the target material domain; changing the friction coefficient from 0.12 to 0.11 and 0.13 respectively results in flanging deformation increments of 0.115 and 0.136; and changing the drawing blank holder force from 1180 kN to 1130 kN and 1230 kN respectively results in flanging deformation increments of 0.126 and 0.128. Based on the reduction in deformation increment corresponding to the unit parameter change, the allowable range for the flange gap in the target material domain is 0.844mm to 0.880mm, and the allowable range for the friction coefficient is 0.100 to 0.111. For the other two material domains with overdraft risk, flange gap ranges of 0.838mm to 0.868mm and 0.842mm to 0.865mm, and friction coefficient ranges of 0.106 to 0.116 and 0.108 to 0.114, respectively, are given. After convergence, a common flange gap range of 0.844mm to 0.865mm and a common friction coefficient range of 0.108 to 0.111 are obtained. The values ​​of 0.844mm and 0.111, which are adjacent to the original parameters, are selected as the execution values ​​and written back to the flange process. After recalculation, the flange deformation increment in the target material domain is reduced to 0.108, and the remaining deformation allowance is 0.008. The remaining deformation allowances in the other material domains are not less than zero. Further, 120 sets of process configuration records that were verified by finite element analysis and mold testing were selected. The manual method combined with the finite element iteration method, the traditional StemGNN method and the method of this invention were compared. The verification results are shown in Table 1.

[0165] Table 1. Comparison of Cross-Process Deformation Tracking and Parameter Configuration Effects for Automotive Body Panels

[0166] As can be seen from the data in Table 1 above, the method of the present invention outperforms the manual combined with finite element iteration method and the traditional StemGNN method in terms of the accuracy rate of the first overdraft process location, the accuracy rate of the identification of responsible process parameters, the first configuration pass rate, the average number of finite element review rounds, and the average parameter configuration time. The accuracy rate of the first overdraft process location of the manual combined with finite element iteration method is 73.3%, the accuracy rate of the identification of responsible process parameters is 69.2%, and the first configuration pass rate is 76.7%. This indicates that when relying on the final defect location and engineering experience for reverse adjustment, it is difficult to accurately determine in which process the local deformation allowance is insufficient for the first time, and it is also easy to include irrelevant process parameters in the adjustment range. The traditional StemGNN method improves the three accuracy indicators to 84.2%, 80.0%, and 85.0% respectively through graph structure and temporal feature processing. However, due to the lack of stable and exclusive spectral domain division between the two StemGNN blocks, the cumulative deformation of the previous process and the new deformation of the current process may still be mixed, and the responsibility location capability is still limited. The method of this invention achieves a 94.2% accuracy rate in initial overdraft process positioning, a 92.5% accuracy rate in identifying responsible process parameters, and a 93.3% initial configuration pass rate. The average number of finite element verification rounds is reduced from 4.2 and 2.8 to 1.6, and the average parameter configuration time is reduced from 84 min and 58 min to 35 min. The initial overdraft process positioning accuracy rate measures the consistency between the system-determined initial insufficient margin process and the manual verification results; the responsible process parameter identification accuracy rate measures the consistency between the backtracked parameters and the parameters confirmed by trial molding; and the initial configuration pass rate represents the proportion of samples whose executed values ​​meet quality requirements immediately after the first write-back. These results demonstrate that this invention can simultaneously improve cross-process anomaly positioning, responsible parameter identification, and one-time parameter configuration success capabilities, while reducing repeated simulations and trial molding.

[0167] This embodiment divides the material domain by establishing the node inheritance relationship between the initial sheet metal mesh and the forming meshes of each stamping process. This ensures that the deformation of the same sheet metal region during drawing, trimming and punching, flanging, and shaping processes continues to correspond, and the allowable deformation allowance is configured based on material properties, initial state, and quality requirements. Based on this, the material domain and deformation transfer relationship are converted into a stamping process graph sequence. The improved StemGNN model uses complementary spectrum projection to allocate the graph Laplacian invariant subspace to the first StemGNN block and the second StemGNN block, respectively. The first StemGNN block extracts the cumulative deformation components that continue across processes, while the second StemGNN block extracts the newly generated deformation components from the current process. The invariant subspaces of adjacent processes inherit the original block affiliation after orthogonal Procrustes alignment. The invariant subspace that initially causes the material domain allowance to be overdrawn is fixedly assigned to the second StemGNN block, thus preventing the newly generated deformation components that cause abnormalities from re-mixing into the cumulative deformation components in subsequent processes. The system then uses the material domain node corresponding to the initial overdraft invariant subspace as the starting point for backtracking, extracts associated process parameters along the deformation transfer edge, determines the responsible process parameter and adjustment direction through bidirectional adjustment of a single parameter, and shrinks the parameter allowable range based on the allowance overdraft amount. The parameter ranges corresponding to different material domains intersect to form a common range, which is used to determine the execution value and write it back to the stamping process, then deducts the deformation allowance again to complete the verification. This process achieves a complete closed loop from the discovery of local deformation anomalies, the location of the initial overdraft process, the backtracking of responsible parameters, to the determination of the execution value, improving the accuracy, stability, and engineering implementation efficiency of automotive stamping process parameter configuration.

[0168] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for configuring automotive stamping process parameters based on deep learning, characterized in that, Includes the following steps: S1. Analyze automotive stamping design data, divide material domains according to the initial sheet metal forming continuity, and form a set of material domains; S2. Associat material domains along the automotive stamping process and configure deformation allowances to form a material domain transfer chain; S3. Convert the material domain transfer chain into an automotive stamping process diagram, using the material domain as nodes and the deformation transfer relationship as edges, and associate process parameters to form a stamping process diagram sequence. S4. Improve the latent correlation layer of the StemGNN model to process the stamping process graph sequence, construct the graph Laplacian matrix, and allocate the invariant subspace to two StemGNN blocks through complementary spectrum projection. S5. Align the invariant subspaces of adjacent automotive stamping processes and inherit block ownership. Extract the cumulative deformation components from the first StemGNN block and extract the newly added deformation components from the second StemGNN block. S6. Superimpose the cumulative deformation components and the newly added deformation components, determine the material domain process deformation increment and deduct the allowable deformation allowance, locate the first overdraft process, the allowance overdraft amount and the first overdraft invariant subspace, and fix the first overdraft invariant subspace to the second StemGNN block. S7. Determine the responsible process parameters based on the deformation transfer relationship and process parameter correlation relationship of the material domain corresponding to the first overdraft invariant subspace, and shrink the allowable range of the responsible process parameters according to the margin overdraft amount to form the material domain parameter range; S8. Intersect the material domain parameter ranges of the same responsible process parameters, determine the execution value and write it back to the stamping process, check the remaining deformation allowance, form the configuration result when not overdrawn, and re-shrink the corresponding parameter range when overdrawn.

2. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Read the initial sheet metal mesh and stamping part forming mesh from the automotive stamping design data, and establish the mesh node mapping relationship according to the node inheritance number in the forming process; S12. Check the adjacency status of the mapped nodes in the stamping forming grid along the adjacent edges of the initial sheet metal grid. Group the adjacent grid cells into the same forming continuous region and separate the grid cells with interrupted adjacency. S13. Merge the initial sheet metal mesh units according to the continuous forming region, divide the material domains and form a set of material domains.

3. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Compare the inheritance relationship of sheet metal grid units in adjacent processes according to the automotive stamping process sequence, establish a correspondence between material domains with common inheritance units, and form a material domain process correspondence relationship. S22. Based on material properties, initial sheet metal condition and stamping quality requirements, determine the available deformation range of each material domain from the initial state to the forming limit, and configure the allowable deformation allowance. S23. Connect the material domain process correspondences along the automotive stamping process sequence. Continue the connection for material domains that retain common inheritance units, and terminate the connection for material domains that no longer have inheritance units. Write the deformation allowance into the continuously connected material domains to form a material domain transfer chain.

4. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. According to the automotive stamping process, cut the material domain transfer chain, establish the material domains retained in each process as nodes, and maintain the corresponding nodes of the same material domain in adjacent processes. S32. Check the sheet metal mesh connection between material domains within the same process, and connect the material domain nodes that maintain the common mesh boundary as deformation transfer edges to form an automotive stamping process diagram. S33. According to the mold contact area and range of action corresponding to the process parameters, associate the process parameters with the covered material domain nodes, and arrange the automotive stamping process diagrams along the automotive stamping process to form a stamping process diagram sequence.

5. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, The improved StemGNN model includes a latent correlation layer, a StemGNN layer, and an output layer. The StemGNN layer includes a first StemGNN block, a second StemGNN block, and residual connections. The graph Fourier transforms of the first StemGNN block and the second StemGNN block are respectively connected to complementary spectrum projection operators.

6. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. The potential correlation layer encodes the material domain node sequence along the automotive stamping process, and the self-attention mechanism calculates the correlation weights between the material domain nodes to form a correlation weight matrix. S42. Perform symmetry and degree normalization on the relevant weight matrix to construct the normalized graph Laplace matrix, and obtain the graph Fourier basis by eigenvalue decomposition. S43. The first StemGNN block generates a backtracking output through the graph Fourier basis, calculates the residual between the material domain node sequence and the backtracking output through the residual connection, and projects the backtracking output and the residual onto the graph Laplace invariant subspace. S44. The graph Laplacian invariant subspace with backtracking output spectral energy not lower than residual spectral energy is assigned to the first StemGNN block, and the rest are assigned to the second StemGNN block. The first spectral projection operator and the second spectral projection operator are constructed respectively.

7. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Using the Fourier basis of the previous automotive stamping process as a reference, perform orthogonal Procrustes alignment on the Laplace invariant subspace of the current automotive stamping process to form an aligned invariant subspace. S52. Calculate the projection amount of the aligned invariant subspace within the range of the first and second spectral projection operators in the previous automotive stamping process. S53. The invariant subspace with a first projection amount not less than the second projection amount is assigned to the first StemGNN block, and the rest are assigned to the second StemGNN block to form the first invariant subspace and the second invariant subspace. S54. The first spectral projection operator projects the material domain node sequence of the current automotive stamping process onto the first invariant subspace. The first StemGNN block sequentially performs graph Fourier transform, spectral sequence unit processing, graph convolution, and inverse graph Fourier transform to form the cumulative deformation component. S55. The residual connection is formed by subtracting the backtracking output of the first StemGNN block from the material domain node sequence. S56. The second spectral projection operator projects the residual onto the second invariant subspace. The second StemGNN block sequentially performs graph Fourier transform, spectral sequence unit processing, graph convolution, and inverse graph Fourier transform to form the newly added deformed components.

8. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Add the newly added deformation components of the current automotive stamping process to the cumulative deformation components according to the corresponding material domain nodes, update the cumulative deformation components, and determine the newly added deformation components as the material domain process deformation increment. S62. Deduct the allowable deformation allowance sequentially along the material domain transfer chain, and determine the material domain and corresponding automotive stamping process whose remaining deformation allowance is less than zero for the first time as the first overdraft material domain and the first overdraft process. S63. The newly added deformation components of the material domain overdrafted for the first time are orthogonally decomposed according to the graph Laplace invariant subspace occupied by the second StemGNN block, and restored to the material domain node components by graph inverse Fourier transform. S64. Determine the first overdraft invariant subspace according to the residual deduction amount corresponding to the material domain node components; S65. Lock the feature vector corresponding to the first overdraft invariant subspace, and use Wielandt elimination to exclude the corresponding feature vector from the graph Fourier basis of the first StemGNN block. S66. The second spectral projection operator preserves the corresponding feature vectors, forming a graph Laplacian invariant subspace that is fixed and belongs to the second StemGNN block.

9. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. The material domain node corresponding to the first overdraft invariant subspace that is fixedly assigned to the second StemGNN block is determined as the backtracking starting point. S72. Backtrack along the deformation transfer edge to the material domain nodes, extract the associated process parameters, and form a set of candidate responsible process parameters; S73. Take equidistant parameter values ​​around the current value of the candidate responsible process parameter in the increasing and decreasing directions, keep other process parameters unchanged, and calculate the process deformation increment of the first overdraft material domain respectively; S74. Identify the process parameters that can reduce the incremental deformation of the process as the responsible process parameters, and determine the adjustment direction to reduce the incremental deformation of the process. S75. Divide the margin overdraft amount by the reduction in process deformation increment corresponding to the unit change of the responsible process parameter to determine the allowable range boundary movement amount. Move the allowable range boundary along the adjustment direction to form the material domain parameter range.

10. The method for configuring automotive stamping process parameters based on deep learning according to claim 1, characterized in that, S8 specifically includes: S81. Collect material domain parameter intervals according to the responsible process parameters, select the maximum lower boundary and the minimum upper boundary, and form a common interval for the responsible process parameters. S82. When the current value of the responsible process parameter is within the common range, the current value is retained. When it is outside the common range, the boundary of the adjacent range is selected to form the execution value and write it back to the corresponding automotive stamping process. S83. Based on the executed value after writing back, redetermine the deformation increment of the material domain process, deduct the allowable deformation allowance along the material domain transfer chain, and form the remaining deformation allowance of the material domain. S84. When the remaining deformation allowance of all material domains is not overdrawn, the configuration result is formed. When there is overdrawn, the corresponding responsible process parameter range is shrunk, the material domain parameter range is re-converged, the execution value is determined, and the remaining deformation allowance is reviewed.

Images