Deep learning-based engine key structure parameter optimization and automatic iteration method

Abstract

The application discloses a deep learning-based engine key structure parameter optimization and automatic iteration method, and relates to the technical field of computer-aided engineering, and comprises the following steps: collecting topology optimization data and constructing a region constraint graph; inputting an improved Swin-Unet model to generate a topology configuration and structure parameters; evaluating performance and process, and screening an initial optimization configuration; establishing a finite element model to generate a region real performance result; comparing the real and predicted results to generate a region error and a correction table; updating a change area grid, and cyclically executing generation, evaluation and finite element recalculation; terminating iteration according to a multi-round region error, and outputting an optimization result. Through the improved Swin-Unet model and the region error-driven finite element iteration, the application realizes the collaborative optimization of the engine topology configuration and key structure parameters, and solves the problem that the existing method is difficult to consider performance constraints, process constraints and iteration efficiency.

Description

Technical Field

[0001] This invention relates to the field of computer-aided engineering technology, and in particular to a method for optimizing and automatically iterating key structural parameters of engines based on deep learning. Background Technology

[0002] As engines evolve towards higher power density, lighter weight, and higher reliability, the wall thickness, stiffener spacing, bore diameter, transition fillets, and cavity dimensions of key engine structures directly impact load-bearing capacity, stiffness, structural quality, and fabrication feasibility. Existing technologies typically employ a combination of parametric modeling, finite element analysis, and topology optimization to design key engine structures. By setting load boundaries, performance constraints, and process constraints, structural parameters are repeatedly adjusted and finite element recalculations are performed to obtain a topological configuration that meets stress, stiffness, and structural quality requirements. Some methods also incorporate deep learning models to generate optimization results based on historical topological configurations and structural parameters, thereby reducing the computational load in traditional topology optimization processes.

[0003] Existing topology optimization methods often rely on fixed design domains, unified constraints, and multiple rounds of global finite element iterations. This makes it difficult to effectively divide and constrain the engine structure based on the differences in the transmission of burst pressure, bolt preload, crankshaft and connecting rod loads, and valve train loads at different locations and directions. This results in an unclear correspondence between local load characteristics and the overall optimization objective. Existing deep learning methods typically focus on topology generation, neglecting the coordinated utilization of performance constraints and process constraints such as minimum casting wall thickness, draft angle, machining limits, and structural continuity. Topology and structural parameters are often generated separately, easily leading to mismatches between the topology and structural parameters, or difficulties in simultaneously meeting performance and manufacturing requirements.

[0004] Furthermore, existing methods often employ a one-time finite element verification after generating the model's topology. The deviation between predicted performance results and actual finite element performance results fails to generate feedback-based correction information according to the load-bearing region, making it difficult to determine the correction object, direction, and amount for local performance deviations. When the topology changes, it is usually necessary to re-mesh the entire finite element mesh and perform a global recalculation, resulting in significant computational overhead. Additionally, there is a lack of iteration termination criteria based on multi-round regional performance errors, meaning the optimization process still relies on manual adjustments, making it difficult to achieve automatic closed-loop iteration between engine topology, key structural parameters, performance evaluation, and process discrimination.

[0005] Therefore, how to provide a method for optimizing key structural parameters of engines based on deep learning and for automatic iteration 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 optimizing key structural parameters of engines and for automatic iteration. This invention utilizes an improved Swin-Unet model and finite element analysis technology to achieve automatic optimization of engine topology and structural parameters, which has the advantages of high optimization accuracy, high computational efficiency and high degree of automation.

[0007] The deep learning-based method for optimizing and automatically iterating key structural parameters of an engine according to embodiments of the present invention includes:

[0008] Collect structural parameters, load boundaries, performance constraints, process constraints, and historical topology configurations to generate topology optimization data. Divide the load-bearing areas according to load location and direction and associate constraints to generate a region constraint map.

[0009] The topology optimization data and region constraint graph are input into the improved Swin-Unet model to generate topology configuration and structural parameters. The improved Swin-Unet model includes an input mapping layer, a shift window encoder, a double-constraint attention jump layer, a shift window decoder, and a parameter configuration collaborative output layer.

[0010] Input the topology and structural parameters into the performance evaluation subnetwork and the process discrimination subnetwork to generate predicted performance results and process discrimination results, and select the initial optimized configuration and corresponding structural parameters.

[0011] A finite element model is established based on the initial optimized configuration and corresponding structural parameters. Load boundaries are applied, and the actual performance results of the region are generated.

[0012] Compare the actual performance results and predicted performance results of the region, and generate a regional performance error and regional correction table based on the process discrimination results;

[0013] Input the region correction table into the improved Swin-Unet model, update the finite element mesh according to the configuration change region, and iteratively execute the topology configuration and structural parameter generation, performance evaluation and finite element recalculation.

[0014] The process terminates after determining the regional performance error in multiple rounds, and outputs the optimization results of the topology configuration and structural parameters.

[0015] Optionally, the structural parameters include wall thickness, stiffener thickness, stiffener spacing, hole diameter, transition fillet, and cavity size; the load boundaries include the load values, load locations, load directions, and support constraints corresponding to burst pressure, bolt preload, crankshaft and connecting rod load, valve train load, and external load; the performance constraints include upper limit of volume fraction, stress threshold, stress fluctuation threshold, and lower limit of stiffness; the process constraints include minimum casting wall thickness, draft angle, machining limits, and structural continuity; and the historical topology includes material distribution, stiffener layout, hole system location, and cavity boundaries.

[0016] Optionally, the generated region constraint map includes:

[0017] By associating structural parameters, load boundaries, performance constraints, process constraints, and historical topology configurations, topology optimization data is generated, and the load location and load direction are determined based on the load boundaries.

[0018] Based on the load location and load direction, combined with material distribution, stiffener layout, hole system location and cavity boundary, the load-bearing area is divided, and the load-bearing area nodes and area connection relationships are generated.

[0019] Structural parameters, performance constraints, and process constraints are associated with nodes in the load-bearing region. Connection edges are established based on the region connection relationships and load directions to generate a region constraint diagram.

[0020] Optionally, the generated topology and structural parameters include:

[0021] The topology optimization data and region constraint map are input into the input mapping layer of the improved Swin-Unet model. Mesh mapping and feature stitching are performed on structural parameters, load boundaries, performance constraints, process constraints, historical topology configurations and load-bearing region nodes. An initial mapping feature map is generated through linear transformation.

[0022] The initial mapped feature map is input into the shift window encoder, and window attention, shift window attention and hierarchical downsampling are performed to extract the associated features of material distribution, structural boundary and load-bearing area, and generate multi-scale encoded feature groups;

[0023] The multi-scale coded feature group, regional constraint map, performance constraint and process constraint are input into the dual-constraint attention jump layer. The coded features are aggregated according to the nodes of the carrying area. The matching weights of performance constraint and process constraint are calculated respectively. The matching weights of adjacent carrying areas are fused according to the connecting edges. The corresponding scale coded features are weighted to generate the multi-scale dual-constraint jump feature group.

[0024] The deepest scale encoded features are input into the shift window decoder, upsampling is performed scale by scale, and concatenated with the double-constrained jump features of the corresponding scale. Window attention and shift window attention are then performed to generate a topological decoding feature map.

[0025] The topology decoding feature map is input to the parameter configuration and output layer to generate the mesh material retention probability. The topology configuration is generated based on the configuration generation threshold and mesh connectivity. The topology configuration is used to perform region aggregation on the topology decoding feature map to generate configuration parameter feature vectors. The structural parameters are then generated from the configuration parameter feature vectors.

[0026] Using topology optimization data and region constraint graphs as training inputs, and historical topology configurations and corresponding structural parameters as supervision labels, the topology configuration loss, structural parameter loss, and region constraint consistency loss are calculated. These are then weighted to generate training optimization objectives and update the parameters of the improved Swin-Unet model until the training optimization objectives converge, resulting in the trained improved Swin-Unet model.

[0027] Optionally, the generation of prediction performance results includes:

[0028] The performance evaluation subnetwork includes a configuration coding branch, a parameter coding branch, a region fusion layer, and a performance regression layer. It performs finite element analysis on historical topological configurations and corresponding structural parameters, and generates stress monitoring labels, stiffness monitoring labels, and structural quality monitoring labels according to the load-bearing region.

[0029] The historical topology and corresponding structural parameters are input into the performance evaluation subnetwork. The performance training loss is calculated based on the stress supervision label, stiffness supervision label and structural quality supervision label. The parameters of the configuration encoding branch, parameter encoding branch, region fusion layer and performance regression layer are updated to obtain the trained performance evaluation subnetwork.

[0030] The topology and structural parameters are input into the trained performance evaluation subnetwork. Configuration coding features and parameter coding features are generated through configuration coding branch and parameter coding branch, respectively. After being fused by the region fusion layer, they are input into the performance regression layer to generate the stress prediction value, stiffness prediction value and structural mass prediction value of each load-bearing region. The volume fraction is calculated based on the topology and the stress fluctuation value is calculated based on the stress prediction value of each load-bearing region to form the predicted performance result.

[0031] Optionally, the generation of process discrimination results and the screening of initial optimized configurations and corresponding structural parameters include:

[0032] The process discrimination subnetwork includes a configuration coding branch, a parameter constraint coding branch, a feature fusion layer, and a process discrimination layer. It compares historical topological configurations and corresponding structural parameters with process constraints to generate process supervision labels.

[0033] The historical topology, corresponding structural parameters, and process constraints are input into the process discrimination subnetwork to generate process training discrimination values. The process training loss is calculated based on the process training discrimination values ​​and process supervision labels. The parameters of the configuration encoding branch, parameter constraint encoding branch, feature fusion layer, and process discrimination layer are updated to obtain the trained process discrimination subnetwork.

[0034] The topology configuration, structural parameters, and process constraints are input into the trained process discrimination subnetwork. Configuration process features and parameter constraint features are generated through configuration encoding branch and parameter constraint encoding branch, respectively. After being fused by the feature fusion layer, they are input into the process discrimination layer to generate the process discrimination result.

[0035] The predicted performance results are compared with the performance constraints. Combined with the process discrimination results, topological configurations and structural parameters that do not meet the performance constraints or process constraints are screened out. The remaining topological configurations are arranged in ascending order of the predicted structural quality values. The first topological configuration and its corresponding structural parameters are determined as the initial optimized configuration and its corresponding structural parameters.

[0036] Optionally, the step of establishing a finite element model based on the initial optimized configuration and corresponding structural parameters, and applying load boundaries, includes:

[0037] Based on the initial optimized configuration, the material retention area is determined, and the wall thickness, stiffeners, hole system, transition fillets and cavities are reconstructed according to the corresponding structural parameters to generate the engine structural geometric model.

[0038] The engine structure geometric model is meshed using the region constraint diagram, the correspondence between finite element elements and load-bearing regions is established, and the load boundary is applied to the corresponding finite element mesh to generate the finite element model.

[0039] The finite element model is solved by performing finite element analysis, and the stress and stiffness are collected according to the load-bearing region. The structural mass, volume fraction and stress fluctuation value are calculated to generate the actual performance results of the region.

[0040] Optionally, the generation of the regional performance error and regional correction table based on the process discrimination results includes:

[0041] The actual performance results and predicted performance results of the region are compared according to the bearing area and performance type, and the performance value difference is calculated to generate the regional performance error.

[0042] The actual performance results of the region are compared with the performance constraints, and the process discrimination results are compared with the process constraints. The correction objects and correction directions are determined by combining the regional performance error.

[0043] The correction amount is determined based on the regional performance error range and the deviation range between the process judgment result and the process constraint. The correction object, correction direction and correction amount are recorded according to the bearing area to generate a regional correction table.

[0044] Optionally, updating the finite element mesh according to the configuration change region includes:

[0045] Input the region correction table into the improved Swin-Unet model, adjust the characteristics of the corresponding bearing region according to the correction object, correction direction and correction amount, and generate the updated topology and structural parameters.

[0046] The updated topology configuration is compared with the original topology configuration by mesh correspondence, and the meshes where the material retention state changes and their adjacent meshes are identified as configuration change regions.

[0047] The finite element mesh of the configuration change region is redefined based on the updated topology and structural parameters, and the boundary is connected with the finite element mesh outside the configuration change region to generate an updated finite element model.

[0048] The updated topology and structural parameters are input into the performance evaluation subnetwork, and the updated topology, structural parameters, and process constraints are input into the process discrimination subnetwork to generate updated predicted performance results and process discrimination results.

[0049] The updated finite element model is recalculated using finite element analysis to generate the actual performance results of the region. The actual performance results of the region are compared with the predicted performance results. The region performance error and region correction table are updated in combination with the process discrimination results. The process of generating topology and structural parameters, performance evaluation, process discrimination and finite element analysis is repeated cyclically.

[0050] Optionally, the output topology configuration and structural parameter optimization results include:

[0051] Set the regional error threshold, error stability threshold, continuous judgment rounds and maximum iteration rounds. Record the regional performance error of each carrying area according to the iteration rounds. Calculate the maximum value of the absolute value of the regional performance error in each round and calculate the absolute value of the difference between the regional performance errors of adjacent rounds.

[0052] Automatic iteration terminates when the maximum absolute value of the regional performance error within consecutive judgment rounds is not greater than the regional error threshold, and the absolute value of the difference between regional performance errors in adjacent rounds is not greater than the error stability threshold; automatic iteration terminates when the number of iteration rounds reaches the maximum number of iteration rounds; otherwise, topology configuration and structural parameter generation, performance evaluation, and finite element recalculation continue.

[0053] From the topology configurations and structural parameters corresponding to consecutive decision rounds, select the round with the smallest absolute value of regional performance error, and determine the corresponding topology configuration and structural parameters as the optimization result of topology configuration and structural parameters.

[0054] The beneficial effects of this invention are:

[0055] This invention associates engine structural parameters, load boundaries, performance constraints, manufacturing constraints, and historical topology configurations. Based on load location and direction, it divides the load-bearing regions and constructs a region constraint diagram, enabling a unified expression of structural characteristics, load transfer relationships, and constraint conditions across different load-bearing regions. By inputting topology optimization data and region constraint diagrams into an improved Swin-Unet model, it achieves the collaborative generation of topology configurations and structural parameters, reducing mismatches between topology configurations and wall thickness, stiffener thickness, aperture, and cavity dimensions, and improving the accuracy of optimizing key engine structural parameters.

[0056] This invention generates predicted performance results and process discrimination results through a performance evaluation subnetwork and a process discrimination subnetwork, respectively. It then combines performance constraints and process constraints to screen initial optimized configurations, ensuring that the optimization results simultaneously consider stress, stiffness, structural mass, volume fraction, and casting and machining requirements. By comparing the actual performance results with the predicted performance results for a specific region, a regional performance error and regional correction table is generated according to the load-bearing region. This allows the model prediction bias to be transformed into correction objects, directions, and amounts for specific load-bearing regions, improving the targeting and reliability of the optimization process.

[0057] This invention determines the configuration change region based on topological configuration changes, re-meshes the finite element mesh only in the configuration change region, and retains the finite element mesh in the remaining regions, reducing the amount of mesh reconstruction in each round of finite element recalculation. By cyclically executing topological configuration and structural parameter generation, performance evaluation, process discrimination, and finite element recalculation, and terminating the iteration based on the performance error of multiple regions, automatic closed-loop optimization of key engine structural parameters is achieved, improving the efficiency of finite element recalculation and the degree of automation in optimization. Attached Figure Description

[0058] 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:

[0059] Figure 1 This is a flowchart of the deep learning-based engine key structural parameter optimization and automatic iteration method proposed in this invention;

[0060] Figure 2 This is a schematic diagram of the improved Swin-Unet model, which is based on deep learning for optimizing key structural parameters of engines and performing automatic iteration.

[0061] Figure 3 This is a flowchart illustrating the automatic iterative optimization process of the deep learning-based engine key structural parameter optimization and automatic iterative method proposed in this invention. Detailed Implementation

[0062] 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.

[0063] refer to Figure 1 , Figure 2 and Figure 3 Deep learning-based optimization and automatic iteration methods for key structural parameters of engines include:

[0064] Collect structural parameters, load boundaries, performance constraints, process constraints, and historical topology configurations to generate topology optimization data. Divide the load-bearing areas according to load location and direction and associate constraints to generate a region constraint map.

[0065] The topology optimization data and region constraint graph are input into the improved Swin-Unet model to generate topology configuration and structural parameters. The improved Swin-Unet model includes an input mapping layer, a shift window encoder, a double-constraint attention jump layer, a shift window decoder, and a parameter configuration collaborative output layer.

[0066] Input the topology and structural parameters into the performance evaluation subnetwork and the process discrimination subnetwork to generate predicted performance results and process discrimination results, and select the initial optimized configuration and corresponding structural parameters.

[0067] A finite element model is established based on the initial optimized configuration and corresponding structural parameters. Load boundaries are applied, and the actual performance results of the region are generated.

[0068] Compare the actual performance results and predicted performance results of the region, and generate a regional performance error and regional correction table based on the process discrimination results;

[0069] Input the region correction table into the improved Swin-Unet model, update the finite element mesh according to the configuration change region, and iteratively execute the topology configuration and structural parameter generation, performance evaluation and finite element recalculation.

[0070] The process terminates after determining the regional performance error in multiple rounds, and outputs the optimization results of the topology configuration and structural parameters.

[0071] In this embodiment, the structural parameters include wall thickness, stiffener thickness, stiffener spacing, hole diameter, transition fillet, and cavity size; the load boundaries include the load values, load positions, load directions, and support constraints corresponding to burst pressure, bolt preload, crank connecting rod load, valve train load, and external load; the performance constraints include upper limit of volume fraction, stress threshold, stress fluctuation threshold, and lower limit of stiffness; the process constraints include minimum casting wall thickness, draft angle, machining limits, and structural continuity; and the historical topology includes material distribution, stiffener layout, hole system position, and cavity boundaries.

[0072] In this embodiment, the generation of the region constraint map includes:

[0073] The structural parameters, load boundaries, performance constraints, manufacturing constraints, and historical topology configurations are correlated to generate topology optimization data. The load locations and directions are then determined based on the load boundaries. Specifically, the generation of topology optimization data involves:

[0074] A sample index is established using engine model and historical topology configuration number. Wall thickness, stiffener thickness, stiffener spacing, hole diameter, transition fillet, and cavity size are associated with the corresponding structural positions according to the geometric entity number. Using the engine's 3D geometric model's reference coordinate system—load position and support position—a load boundary table is created, comprising load values, load application surfaces, load directions, and support constraints. Volume fraction upper limit, stress threshold, stress fluctuation threshold, stiffness lower limit, minimum casting wall thickness, draft angle, machining limits, and structural continuity are associated with the corresponding structural positions according to the constraint objects. The structural design domain is divided using half the minimum casting wall thickness as the mesh edge length. The material retention state is recorded based on whether the center of each mesh is located inside the material entity, and stiffener layout, hole system position, and cavity boundary are mapped to the corresponding mesh. The sample index, structural parameters, load boundary table, performance constraints, process constraints, and meshed historical topology configuration are combined according to geometric entity number to generate topology optimization data.

[0075] The load location and load direction are determined based on the load boundary, specifically as follows:

[0076] The coordinates of the concentrated load application node in the reference coordinate system of the engine's three-dimensional geometric model are used to determine the concentrated load location. The area of ​​each grid surface of the distributed load application surface is multiplied by the corresponding center coordinate of the grid surface, the resulting product is summed, and then divided by the total area of ​​the grid surfaces to obtain the distributed load location. For loads with three coordinate axis components, the square root of the sum of the squares of the three components is used to obtain the load resultant, and each coordinate axis component is divided by the load resultant to obtain the load direction. For burst pressure, the normal vector of each grid surface of the pressure application surface is multiplied by the corresponding grid surface area, the summation of the resulting vectors is normalized according to the vector length, and the direction pointing inwards towards the material is determined as the load direction. The load location coordinates and load direction components are recorded according to the load type.

[0077] Based on the load location and direction, and combined with material distribution, rib layout, hole location, and cavity boundary, the load-bearing area is divided, and load-bearing area nodes and area connection relationships are generated. Specifically, the generation of load-bearing area nodes and area connection relationships is as follows:

[0078] Mesh with a material retention state of 1 in the historical topology is identified as the material mesh, and meshes covered by hole locations and cavity boundaries are identified as blocking meshes. Material meshes containing load locations or intersecting with the load application surface are identified to determine the load initiation mesh. Starting from the load initiation mesh, adjacent coplanar material meshes are traversed, and the angle between the direction vector from the center of the current material mesh to the center of the adjacent material mesh and the load direction is calculated. Adjacent material meshes with an angle not exceeding 90° are included in the load transfer range, and traversal in the corresponding direction stops when a blocking mesh is encountered. The load transfer range is continuously expanded along the stiffener layout, and material meshes connected to the same load initiation mesh and coplanar with each other are merged into the load-bearing region. For material meshes belonging to multiple load transfer ranges simultaneously, the load transfer range along the coplanar material mesh is calculated separately. The minimum number of meshes required to reach each load starting mesh is used to assign the material mesh to the load-bearing region corresponding to the minimum number of meshes. If the minimum number of meshes is the same, the material mesh is assigned to the load-bearing region with the larger load value. The load-bearing region nodes are generated by recording the region number, material mesh set, load type, load location, load direction, stiffener layout, hole system boundary, and cavity boundary according to the load-bearing region. The material meshes of different load-bearing regions are compared one by one. When two load-bearing regions have coplanar meshes, the region numbers, coplanar mesh numbers, and coplanar areas are recorded to generate a direct connection relationship. When two load-bearing regions are connected by continuous stiffeners and there are no blocking meshes in between, the region numbers and stiffener numbers are recorded to generate a stiffener connection relationship. The direct connection relationship and the stiffener connection relationship are combined to form a region connection relationship.

[0079] Structural parameters, performance constraints, and process constraints are associated with nodes in the load-bearing region. Connection edges are established based on the region connectivity and load direction to generate a region constraint diagram. Specifically, this process involves establishing connection edges based on region connectivity and load direction to generate the region constraint diagram.

[0080] The arithmetic mean of the center coordinates of the material mesh in each load-bearing region is taken to obtain the region center coordinates. The load value is multiplied by each component of the load direction and summed separately to obtain the region load resultant vector, which is then normalized to obtain the region load direction. The difference between the center coordinates of adjacent regions is used to form the region connection vector and normalized to obtain the region connection direction. The region load direction is multiplied by each component of the region connection direction and summed to obtain the direction determination value. When the direction determination value is greater than 0, a connection edge is established from the current region to the adjacent region; when it is less than 0, a connection edge is established in the opposite direction; when it is equal to 0, a bidirectional connection edge is established. The coplanar area or stiffener number, stiffener thickness, and stiffener cross-sectional area are written into the connection edge attributes, and the load-bearing region nodes, node association constraints, and connection edges are combined to generate a region constraint diagram.

[0081] In this embodiment, the generated topology and structural parameters include:

[0082] The topology optimization data and region constraint map are input into the input mapping layer of the improved Swin-Unet model. Mesh mapping and feature stitching are performed on structural parameters, load boundaries, performance constraints, process constraints, historical topology configurations, and load-bearing region nodes. An initial mapping feature map is generated through linear transformation. Specifically, the generation of the initial mapping feature map is as follows:

[0083] A unified grid coordinate system is established based on the historical topological configuration. Material retention status, stiffener layout, hole system location, and cavity boundary are mapped to grid values. Structural parameters are associated with the grids covered by the corresponding structural locations. Load values ​​and load direction components are written into the grids corresponding to the load application locations. Performance constraints, process constraints, and load-bearing area node numbers are written into the grids contained in the corresponding load-bearing areas. Continuous features are scaled to 0 to 1 using the minimum and maximum values ​​of each continuous feature in the training samples. One-hot encoding is used to convert the load type and load-bearing area node numbers. In the order of material features, structural parameter features, load features, performance constraint features, process constraint features, and area node features, grid features are spliced ​​in the channel direction. The spliced ​​features are multiplied by a linear transformation weight matrix and a linear transformation bias is added to generate an initial mapped feature map. The linear transformation weight matrix is ​​initialized using the Xavier method, and the initial value of the linear transformation bias is set to 0 and updated as the model is trained.

[0084] The initial mapped feature map is input into the shifted window encoder, and window attention, shifted window attention, and hierarchical downsampling are performed to extract the associated features of material distribution, structural boundaries, and load-bearing areas, generating a multi-scale encoded feature set. Specifically, the generation of the multi-scale encoded feature set includes:

[0085] The initial mapped feature map is divided into a non-overlapping window with four grids on each side. Layer normalization is performed on the grid features within the window, and then multiplied by the query, key, and value transformation matrices to generate query vectors, key vectors, and value vectors. The query vector is multiplied by the corresponding components of each key vector and summed, then divided by the square root of the key vector dimension. Positional offsets obtained according to the grid's relative coordinates are superimposed, and softmax normalization is performed to obtain attention weights. The value vectors are then weighted and summed, and added to the input features to generate the window attention features. The window is then cyclically moved two grids along each of the three spatial dimensions. For grids not belonging to the same window, a mask value of -100 is added. The grids are then calculated in the same way and moved in the reverse direction to restore their positions, generating shifted window attention features. These shifted window attention features are then subjected to layer normalization, a fully connected transformation that increases the channel dimension by 4 times, GELU activation, and channel restoration transformation, and added to the input features to generate the current scale coding features. Adjacent 2×2×2 grid features are then concatenated and linearly compressed to halve the spatial size and double the number of channels. This process is repeated for four levels of feature extraction and downsampling. The coding features at each level are then arranged by scale to generate a multi-scale coding feature group.

[0086] The multi-scale coded feature set, region constraint map, performance constraints, and process constraints are input into the dual-constraint attention skip layer. The coded features are aggregated according to the nodes of the carrying area, and the matching weights for performance constraints and process constraints are calculated separately. The matching weights of adjacent carrying areas are fused based on the connecting edges, and the coded features at the corresponding scales are weighted to generate a multi-scale dual-constraint skip feature set. The dual-constraint attention skip layer includes:

[0087] Regional feature aggregation unit: According to the grid index contained in the carrying area node, extract the corresponding grid features from the coding features of each scale, and perform average aggregation on the grid features in the same carrying area to generate multi-scale regional coding features;

[0088] Performance-constrained attention branch: Perform feature mapping and matching calculation on the multi-scale region coding features with the upper limit of volume fraction, stress threshold, stress fluctuation threshold and lower limit of stiffness to generate performance constraint matching weights for each load-bearing region;

[0089] Process constraint attention branch: Perform feature mapping and matching calculation on the coding features of multi-scale regions with minimum casting wall thickness, draft angle, machining limit and structural continuity to generate process constraint matching weights for each load-bearing region;

[0090] Neighborhood weight fusion unit: Extracts adjacent bearing areas based on the connecting edges in the region constraint graph, and fuses the performance constraint matching weights and process constraint matching weights of the adjacent bearing areas according to the direction and attributes of the connecting edges to generate region fusion matching weights;

[0091] Multi-scale jump weighted unit: Maps the region fusion matching weights to the grid positions contained in the corresponding carrying region, weights the encoded features of each scale grid by grid, and generates a multi-scale double-constraint jump feature group;

[0092] In the dual-constraint attention skip layer, the region feature aggregation unit reads the corresponding grid features from the multi-scale encoded feature group based on the grid index recorded by the node of the bearing region. It then adds the values ​​of the corresponding dimensions of each grid feature in the same bearing region and divides them by the number of grids to obtain the multi-scale region encoded features. The performance constraint attention branch and the process constraint attention branch perform linear mapping on the multi-scale region encoded features and the corresponding constraints, respectively. The corresponding dimensions of the mapping results are multiplied and summed, and then normalized by Softmax to generate performance constraint matching weights and process constraint matching weights. The neighborhood weight fusion unit determines the weight transmission direction based on the direction of the connecting edge. It averages the two types of matching weights of the adjacent bearing regions with the two types of matching weights of the current bearing region. Direct connecting edges participate in the fusion according to the ratio of the coplanar area to the contact area of ​​the two bearing regions, and stiffener connecting edges participate in the fusion according to the ratio of the stiffener cross-sectional area to the contact area of ​​the connected regions to generate region fusion matching weights. The multi-scale skip weighting unit writes the region fusion matching weights into the corresponding grid positions and multiplies them element-wise with the corresponding scale encoded features to form a multi-scale dual-constraint skip feature group for the shift window decoder to call.

[0093] The deepest-scale encoded features are input into the shift-window decoder, upsampling is performed scale by scale, and concatenated with the corresponding scale's double-constrained jump features. Window attention and shift-window attention are then performed to generate a topological decoding feature map. Specifically, the generation of the topological decoding feature map is as follows:

[0094] The channel dimension of the deepest-scale encoded feature is linearly increased to 4 times and rearranged according to grid coordinates, so that the three spatial dimensions are increased to 2 times and the channel dimension is reduced to half, generating upsampled features. The upsampled features and double-constrained skip features with the same spatial size are concatenated along the channel direction, and the channel dimension is compressed to the dimension corresponding to the current decoding scale through linear transformation. The concatenated features are divided into non-overlapping windows with a side length of 4 grids, and the window attention features are calculated after layer normalization and query, key, and value transformation. The window is cyclically moved 2 grids along each of the three spatial dimensions. A -100 mask value is added to the grids that do not belong to the same window before the move, and the shifted window attention features are calculated and the grid positions are restored. The attention features are added to the input features after a fully connected transformation that increases the channel dimension by 4 times, GELU activation, and channel restoration transformation. The upsampling, skip feature concatenation, and attention calculation are repeated scale by scale until the spatial size of the initial mapped feature map is restored, generating the topological decoding feature map.

[0095] The topology decoding feature map is input into the parameter configuration co-output layer to generate mesh material retention probabilities. A topology configuration is generated based on the configuration generation threshold and mesh connectivity. Region aggregation is performed on the topology decoding feature map using the topology configuration to generate configuration parameter feature vectors. Structural parameters are then generated from these configuration parameter feature vectors. The parameter configuration co-output layer includes:

[0096] Material probability output branch: Performs channel mapping on the topology decoding feature map to generate the material retention probability of each grid;

[0097] Topology generation unit: Based on the topology generation threshold and mesh connectivity, materials are selected to retain the mesh and generate the topology.

[0098] Regional feature aggregation unit: Aggregates topology decoding features according to topology configuration and carrying region nodes to generate regional configuration features;

[0099] Structural parameter regression branch: Integrates the configuration features of each region and maps them to the range of structural parameter values ​​to generate configuration parameter feature vectors and structural parameters;

[0100] In the parameter configuration collaborative output layer, the material probability output branch uses a 1×1×1 convolution to compress the topology decoding feature map into a single-channel mesh feature, which is then converted into a material retention probability between 0 and 1 using a Sigmoid function. The topology configuration generation unit sets the configuration generation threshold to 0.5, marking meshes with a material retention probability of not less than 0.5 as retained meshes and meshes with a material retention probability less than 0.5 as removed meshes. Starting from the retained meshes corresponding to the load application position and the support position, the retained meshes are traversed according to the six-adjacency relationship of the shared mesh surface. Meshes connected to the starting point are retained and isolated meshes are deleted, generating the configuration. Topology configuration; the regional feature aggregation unit extracts the corresponding retained grid features according to the grid index recorded by the node of the bearing area, sums the retained grid features of each channel in the same area and divides them by the number of retained grids to generate regional configuration features; the structural parameter regression branch splices the regional configuration features according to the node number of the bearing area, generates configuration parameter feature vectors through two layers of fully connected transformation and GELU activation, generates structural parameter ratio values ​​through the Sigmoid function, and performs interval mapping according to the lower limit and upper limit of parameters determined by process constraints and historical samples to generate wall thickness, stiffener thickness, stiffener spacing, hole diameter, transition fillet and cavity size;

[0101] Using topology optimization data and region constraint graphs as training inputs, and historical topology configurations and corresponding structural parameters as supervision labels, the topology configuration loss, structural parameter loss, and region constraint consistency loss are calculated. These are then weighted to generate a training optimization objective and update the parameters of the improved Swin-Unet model until the training optimization objective converges, resulting in a trained improved Swin-Unet model. Specifically, the trained improved Swin-Unet model is obtained as follows:

[0102] The training samples were divided into training, validation, and test sets in an 8:1:1 ratio, with 8 sets of topology optimization data and region constraint maps input in each batch. The binary cross-entropy between the mesh material retention probability and the material retention state of the historical topology configuration was calculated mesh-by-mesh, and the average value was taken to obtain the topology configuration loss. The predicted structural parameters and supervision labels were normalized to 0 to 1 using the lower limit of the process constraints on the structural parameters and the maximum value of the historical samples. The average value of the absolute differences of the corresponding parameters was calculated to obtain the structural parameter loss. The absolute difference between the mean of the predicted material retention probability and the mean of the historical material retention state was calculated for each node in the bearing area, and the absolute difference between the material retention probabilities of coplanar meshes in adjacent regions was calculated for each connecting edge. For the difference, the average of the two types of differences is taken to obtain the region constraint consistency loss. The weights of the three types of losses are 0.5, 0.3 and 0.2 respectively, and they are summed to generate the training optimization objective. The weights are determined by the combination with the smallest training optimization objective in the grid search of the validation set. The AdamW optimizer is used to update the model parameters. The initial learning rate is set to 0.0001 and the weight decay coefficient is set to 0.01. The learning rate is halved every 10 rounds of training. Training is stopped when the decrease of the training optimization objective in the validation set is less than 0.0001 for 10 consecutive rounds. The model parameters of the minimum training optimization objective in the validation set are saved to obtain the improved Swin-Unet model after training.

[0103] In this embodiment, generating the predicted performance results includes:

[0104] The performance evaluation subnetwork includes a configuration coding branch, a parameter coding branch, a region fusion layer, and a performance regression layer. It performs finite element analysis on historical topological configurations and corresponding structural parameters, generating stress monitoring labels, stiffness monitoring labels, and structural quality monitoring labels according to the load-bearing region. Specifically, the finite element analysis on historical topological configurations and corresponding structural parameters, and the generation of stress monitoring labels, stiffness monitoring labels, and structural quality monitoring labels according to the load-bearing region, are as follows:

[0105] The material retention area is determined according to the historical topology. Wall thickness, stiffeners, hole systems, transition fillets, and cavities are reconstructed based on the corresponding structural parameters. A finite element mesh is generated, and the correspondence between finite element elements and the load-bearing area is established according to the mesh positions contained in the nodes of the load-bearing area. The material's elastic modulus, Poisson's ratio, and density are read. Load boundaries and support constraints are applied to the corresponding finite element nodes, and nodal displacement, element strain, and element equivalent stress are solved. For each load-bearing area, the maximum equivalent stress within the area is recorded as a stress monitoring label. The total load in the area is divided by the maximum nodal displacement in the load direction to obtain the stiffness monitoring label. When the maximum nodal displacement is 0, the corresponding training sample is deleted. The volume of each finite element element within the area is multiplied by the material density, and the resulting products are summed to obtain the structural quality monitoring label. The stress monitoring label, stiffness monitoring label, and structural quality monitoring label are recorded according to the load-bearing area number.

[0106] The performance evaluation subnetwork is constructed as follows:

[0107] The system establishes a configuration coding branch, a parameter coding branch, a region fusion layer, and a performance regression layer. The configuration coding branch uses three sets of convolutional layers to encode the material distribution and structural boundaries of the topological configuration. Each convolutional layer sequentially performs 3×3×3 convolution, batch normalization, GELU activation, and 2x downsampling to generate configuration coding features. The parameter coding branch normalizes wall thickness, stiffener thickness, stiffener spacing, aperture, transition fillet, and cavity size to 0 to 1, and generates parameter coding features through two fully connected transformations. The region fusion layer averages and aggregates the configuration coding features according to the grid index contained in the load-bearing region nodes, concatenates the region configuration features with the parameter coding features, and performs a weighted summation of adjacent load-bearing region features based on the connection edges. The weight of direct connection edges is the proportion of the coplanar area to the region contact area, and the weight of stiffener connection edges is the proportion of the stiffener cross-sectional area to the region contact area. The performance regression layer inputs the fused features into three parallel fully connected output branches to generate predicted stress, stiffness, and structural quality values ​​for each load-bearing region, forming a performance evaluation sub-network.

[0108] The historical topological configuration and corresponding structural parameters are input into the performance evaluation subnetwork. The performance training loss is calculated based on the stress monitoring label, stiffness monitoring label, and structural quality monitoring label. The parameters of the configuration encoding branch, parameter encoding branch, region fusion layer, and performance regression layer are then updated to obtain the trained performance evaluation subnetwork. Specifically, the trained performance evaluation subnetwork is as follows:

[0109] The historical topology and corresponding structural parameters are divided into training, validation, and test sets in an 8:1:1 ratio. Using the minimum and maximum values ​​of various supervision labels in the training set, the stress, stiffness, and structural quality supervision labels are normalized to 0 to 1. The absolute differences between the output of the performance evaluation sub-network and the corresponding supervision labels are calculated and averaged to generate stress, stiffness, and structural quality training losses. These losses are then weighted and summed with weights of 0.4, 0.35, and 0.25 to generate the performance training loss, with weights taken from the combination with the smallest performance training loss in the validation set grid search. The Adam optimizer is used to back-update the parameters of the configuration encoding branch, parameter encoding branch, region fusion layer, and performance regression layer. The initial learning rate is set to 0.0001, and it is halved every 10 training rounds. Training stops when the performance training loss in the validation set decreases by less than 0.0001 for 10 consecutive rounds, and the network parameters of the round with the smallest performance training loss in the validation set are saved, resulting in the trained performance evaluation sub-network.

[0110] The topology and structural parameters are input into the trained performance evaluation subnetwork. Configuration-encoded features and parameter-encoded features are generated through configuration-encoded branches and parameter-encoded branches, respectively. After fusion by the region fusion layer, these features are input into the performance regression layer to generate predicted stress, stiffness, and structural mass values ​​for each load-bearing region. The volume fraction is calculated based on the topology, and stress fluctuation values ​​are calculated based on the predicted stress values ​​for each load-bearing region, forming the predicted performance results. Specifically, the generation of predicted stress, stiffness, and structural mass values ​​for each load-bearing region is as follows:

[0111] The configuration encoding branch performs convolutional encoding on the material-preserving mesh and structural boundaries of the topological configuration. The parameter encoding branch normalizes each structural parameter according to the minimum and maximum values ​​of the training set and performs fully connected mapping. The region fusion layer averages the configuration encoding features channel by channel according to the mesh index contained in the nodes of the bearing region, concatenates the obtained region configuration features with the parameter encoding features, and performs weighted summation of the features of adjacent bearing regions according to the connection edges. The performance regression layer inputs the fused features of each bearing region into the stress regression branch, stiffness regression branch and structural quality regression branch respectively, generates a prediction ratio value between 0 and 1 through fully connected transformation, and then multiplies it by the difference between the maximum and minimum values ​​of the corresponding supervision label and adds the minimum value to obtain the stress prediction value, stiffness prediction value and structural quality prediction value of each bearing region.

[0112] The volume fraction is calculated based on the topological configuration, and the stress fluctuation value is calculated based on the predicted stress values ​​of each load-bearing region, specifically as follows:

[0113] The total material retention volume is obtained by summing the mesh volumes of all material-retaining meshes in the topology. The total design domain volume is obtained by summing the mesh volumes of all meshes within the design domain. The volume fraction is obtained by dividing the total material retention volume by the total design domain volume. The average stress of each load-bearing region is obtained by summing the predicted stress values ​​of each region and dividing by the number of load-bearing regions. The square of the difference between each predicted stress value and the average stress of the region is calculated. The square of the sum of the squared differences is divided by the number of load-bearing regions and the square root is obtained to obtain the standard deviation of the regional stress. The stress fluctuation value is obtained by dividing the standard deviation of the regional stress by the average stress of the region. When the average stress of the region is equal to 0, the stress fluctuation value is recorded as 0.

[0114] In this embodiment, generating process discrimination results and screening initial optimized configurations and corresponding structural parameters includes:

[0115] The process discrimination subnetwork includes a configuration coding branch, a parameter constraint coding branch, a feature fusion layer, and a process discrimination layer. It compares historical topological configurations and corresponding structural parameters with process constraints to generate process supervision labels. The construction of the process discrimination subnetwork specifically involves:

[0116] The system establishes a configuration encoding branch, a parameter constraint encoding branch, a feature fusion layer, and a process discrimination layer. The configuration encoding branch uses three sets of convolutional layers to encode the material distribution, structural boundaries, and connectivity of historical topological configurations. Each set of convolutional layers sequentially performs 3×3×3 convolution, batch normalization, GELU activation, and 2x downsampling to generate configuration encoding features. The parameter constraint encoding branch normalizes structural parameters and constraints such as minimum casting wall thickness, draft angle, machining limit, and structural continuity to 0 to 1, and generates parameter constraint encoding features through two fully connected transformations. The feature fusion layer performs global average aggregation on the configuration encoding features and concatenates them with the parameter constraint encoding features, generating process fusion features through a fully connected transformation. The process discrimination layer sets up four parallel output branches for minimum wall thickness, draft angle, machining limit, and structural continuity, and generates corresponding process discrimination values ​​through a Sigmoid function. Calculate the relative boundary spacing, the angle between the surface normal vector and the demolding direction, the distance from the processing area to the limit boundary, and the connectivity of the six adjacent material meshes of the historical topology configuration. Compare these with the corresponding process constraints. If the constraints are satisfied, the corresponding process supervision label is recorded as 1; otherwise, it is recorded as 0, thus forming a process supervision label.

[0117] The historical topology, corresponding structural parameters, and process constraints are input into the process discrimination subnetwork to generate process training discrimination values. The process training loss is calculated based on these values ​​and the process supervision labels. The parameters of the configuration encoding branch, parameter constraint encoding branch, feature fusion layer, and process discrimination layer are then updated to obtain the trained process discrimination subnetwork. Specifically, the trained process discrimination subnetwork is obtained as follows:

[0118] Historical topology configurations, corresponding structural parameters, process constraints, and process supervision labels are divided into training, validation, and test sets in an 8:1:1 ratio. Each training sample is input into the process discrimination subnetwork to generate process training discrimination values ​​for minimum wall thickness, draft angle, machining limits, and structural continuity. For each type of process supervision label, the number of samples with values ​​of 1 and 0 is counted. The positive sample loss weight is obtained by dividing the number of samples with values ​​of 0 by the number of samples with values ​​of 1. The natural logarithm of the process training discrimination value when the process supervision label is 1 is taken and multiplied by the positive sample loss weight. The natural logarithm of the process training discrimination value when the process supervision label is 0 is multiplied by the positive sample loss weight. Subtract the natural logarithm of the process training discriminant value, take the negative values ​​of the two results and average them to obtain the process discrimination loss for the corresponding category; take the arithmetic mean of the four process discrimination losses to generate the process training loss; use the Adam optimizer to update the parameters of the configuration coding branch, parameter constraint coding branch, feature fusion layer and process discrimination layer, with the initial learning rate set to 0.0001, and halving the learning rate every 10 rounds of training; stop training when the decrease in the process training loss on the validation set is less than 0.0001 for 10 consecutive rounds, save the network parameters of the round with the minimum process training loss on the validation set, and obtain the trained process discrimination sub-network;

[0119] The topology, structural parameters, and process constraints are input into the trained process discrimination subnetwork. Configuration process features and parameter constraint features are generated through configuration encoding and parameter constraint encoding branches, respectively. These features are then fused by a feature fusion layer and input into the process discrimination layer to generate the process discrimination result. Specifically, the generation of the process discrimination result involves:

[0120] The material distribution, structural boundaries, and connectivity of the topological configuration are input into the configuration encoding branch, and generated through convolution, batch normalization, GELU activation, and downsampling to produce configuration process features. Structural parameters and process constraints are normalized to 0-1 according to the minimum and maximum values ​​of the training set, and input into the parameter constraint encoding branch, generating parameter constraint features through two fully connected transformations. The configuration process features are then subjected to global averaging aggregation and concatenated with the parameter constraint features, followed by a fully connected transformation and GELU activation through a feature fusion layer to generate process fusion features. These process fusion features are input into the minimum wall thickness, draft angle, machining limit, and structural continuity discrimination branches, respectively, and corresponding process discrimination values ​​are generated using the Sigmoid function. Detection lines are extended from the center of each material boundary grid along the boundary normal into the material interior, and the distance between the intersection points of the detection lines and the relative material boundaries is calculated. The minimum value among all intersection distances is determined as the actual minimum wall thickness. The corresponding components of the unit normal vector of the casting surface and the unit vector of the demolding direction are multiplied and summed, and the result is used to calculate the minimum wall thickness. The direction angle is obtained by taking the inverse cosine of the value. The absolute value of the difference between 90° and the direction angle is determined as the local draft angle, and the minimum value among all local draft angles is determined as the actual minimum draft angle. The spatial distance between the center of the mesh in each machining area and the machining limit boundary point is calculated, and the minimum value among all spatial distances is determined as the actual minimum machining distance. Starting from the material mesh corresponding to the load application position, the material-preserved mesh is traversed according to the six-adjacency relationship of the shared mesh surface. When the support position can be reached, the actual structural continuity state is recorded as 1, and when the support position cannot be reached, the actual structural continuity state is recorded as 0. The process discrimination threshold is set to 0.5. When the process discrimination value is not less than 0.5, the corresponding process item is recorded as satisfying the constraint; when it is less than 0.5, the corresponding process item is recorded as not satisfying the constraint. The process discrimination result is composed of the four process discrimination values, the actual minimum wall thickness, the actual minimum draft angle, the actual minimum machining distance, the actual structural continuity state, the constraint satisfaction state, and the process items that do not satisfy the constraint.

[0121] The predicted performance results are compared with the performance constraints. Based on the process discrimination results, topological configurations and structural parameters that do not meet the performance or process constraints are eliminated. The remaining topological configurations are then arranged in ascending order of predicted structural quality values. The first topological configuration and its corresponding structural parameters are determined as the initial optimized configuration and its corresponding structural parameters. Specifically, the comparison of predicted performance results with performance constraints, combined with process discrimination results, is used to screen and determine the initial optimized configuration and its corresponding structural parameters.

[0122] Read the volume fraction, stress fluctuation value, and predicted stress and stiffness values ​​of each load-bearing region for each topological configuration. Compare the volume fraction with the upper limit, the stress fluctuation value with the stress fluctuation threshold, the predicted stress value of each load-bearing region with the stress threshold, and the predicted stiffness value of each load-bearing region with the lower limit. Record the performance constraint satisfaction status if all corresponding performance constraints are met; otherwise, record the performance constraint non-met. Read the constraint satisfaction status for minimum wall thickness, draft angle, machining limit, and structural continuity. Record the process constraint if all four are met. If the condition is met, record the condition where the process constraint is not met; delete the topology and its corresponding structural parameters that do not meet the performance constraint or process constraint; sum the predicted structural quality values ​​of each load-bearing region to obtain the predicted structural quality value of the corresponding topology, and arrange the remaining topologies in ascending order of predicted structural quality value; if the predicted structural quality values ​​are the same, arrange them in ascending order of predicted maximum stress value; if the predicted maximum stress values ​​are still the same, arrange them in ascending order of stress fluctuation value; determine the topology and its corresponding structural parameters that are ranked first as the initial optimized topology and its corresponding structural parameters.

[0123] In this embodiment, the establishment of the finite element model based on the initial optimized configuration and corresponding structural parameters, and the application of load boundaries include:

[0124] Based on the initial optimized configuration, the material retention area is determined. The wall thickness, stiffeners, hole system, transition fillets, and cavities are reconstructed according to the corresponding structural parameters to generate the engine structural geometric model. Specifically, generating the engine structural geometric model involves:

[0125] The material retention mesh is read according to the initial optimized configuration. Adjacent material retention meshes with shared mesh surfaces are merged, and the boundary surface between the material retention mesh and the material removal mesh is extracted to form the material retention region. The wall is formed by offsetting along the normal direction of the corresponding boundary surface according to the wall thickness parameter and closing the side. The center surface of the stiffener is determined according to the stiffener layout. The stiffener is stretched equidistantly to both sides of the center surface according to the stiffener thickness, and the stiffener entities are arranged according to the stiffener spacing. Cylindrical holes are established according to the hole system position, hole axis direction and hole diameter. Closed cavities are established according to the cavity boundary and cavity size. The wall and stiffener entities are Boolean unioned, and the holes and cavities are Boolean differenceed. Filleting is performed on the specified intersecting edges according to the transition fillet parameter. Adjacent boundaries with a distance of no more than 0.01 mm are merged and unconnected boundaries are closed to form a continuous and closed engine structure geometry model.

[0126] The engine structural geometric model is meshed using the region constraint diagram, the correspondence between finite element elements and load-bearing regions is established, and load boundaries are applied to the corresponding finite element meshes to generate the finite element model. Specifically, the generation of the finite element model involves:

[0127] Using one-third of the minimum casting wall thickness as the base mesh edge length, tetrahedral finite element elements are used to divide the engine structural geometry model. The mesh edge lengths of hole edges, transition fillets, cavity boundaries, and load-bearing area connection positions are reduced to half the base mesh edge length. The aspect ratio of each finite element is calculated, and finite element elements with aspect ratios greater than 5 are deleted and the corresponding regions are re-divided. The material mesh and load-bearing area nodes are found according to the center coordinates of the finite element, and the finite element numbers are written into the corresponding load-bearing areas. Finite element elements that cross the boundary of an area are assigned according to the maximum proportion of the element volume in each load-bearing area. The elastic modulus, Poisson's ratio, and density corresponding to the material grade are read from the engine design file and written into all finite element elements. Concentrated loads are applied to the finite element nodes closest to the load position. The distributed load values ​​are multiplied by the area of ​​the finite element mesh on the action surface and distributed to the mesh nodes along the load direction. The pressure load is applied to the corresponding surface mesh along the normal of the action surface. The displacement in the specified direction of the node corresponding to the support position is set to 0. A finite element model containing finite element elements, region correspondence, material properties, loads, and support constraints is generated.

[0128] The finite element model is solved using finite element analysis (FEM), with stress and stiffness collected by load-bearing region. The structural mass, volume fraction, and stress fluctuation values ​​are calculated to generate the true performance results for the region. Specifically, the FEM solution process involves collecting stress and stiffness by load-bearing region as follows:

[0129] The element stiffness matrix is ​​calculated based on the nodal coordinates, elastic modulus, and Poisson's ratio of the finite element elements. The element stiffness matrices are then superimposed according to the node numbers to form the overall stiffness matrix. Concentrated loads, distributed loads, and pressure loads are converted into nodal loads and a total load vector is formed. The nodal displacements corresponding to the support constraints are set to 0. The conjugate gradient method is used to solve the equation where the product of the total stiffness matrix and the nodal displacement vector equals the total load vector. The solution is stopped when the ratio of the residual to the length of the total load vector is less than 0.000001. The element strain is calculated based on the nodal displacements of the finite element elements, and the element stress and equivalent stress are calculated using the material elasticity matrix. According to the correspondence between the finite element elements and the bearing regions, the maximum equivalent stress in each bearing region is determined as the region stress. The region stiffness is obtained by dividing the sum of the region loads by the maximum nodal displacement in the load direction.

[0130] The calculation of structural mass, volume fraction, and stress fluctuation values ​​generates the actual performance results for the region, specifically as follows:

[0131] Based on the correspondence between finite element elements and load-bearing regions, the volume of each finite element element within each load-bearing region is multiplied by the material density, and the resulting products are summed to obtain the regional structural mass of each load-bearing region. The total structural mass of all regions is summed to obtain the structural mass. The total material retention volume is obtained by summing the volumes of all finite element elements, and the total design domain volume is obtained by summing the volumes of all meshes within the engine structure design domain. The volume fraction is obtained by dividing the total material retention volume by the total design domain volume. The regional average stress is obtained by summing the regional stresses of each load-bearing region and dividing by the number of load-bearing regions. The square of the difference between the regional stress and the regional average stress is calculated, and the average and square root of the squared differences are obtained to obtain the regional stress standard deviation. The stress fluctuation value is obtained by dividing the regional stress standard deviation by the regional average stress. When the regional average stress is 0, the stress fluctuation value is recorded as 0. The load-bearing region number, regional stress, regional stiffness, and regional structural mass are combined accordingly, and the structural mass, volume fraction, and stress fluctuation value are associated to generate the actual performance results of the region.

[0132] In this embodiment, the generation of the regional performance error and regional correction table based on the process discrimination results includes:

[0133] The actual performance results and predicted performance results of the region are compared according to the bearing area and performance type. The performance value difference is calculated to generate the regional performance error. Specifically, the calculation of the performance value difference and the generation of the regional performance error are as follows:

[0134] According to the actual and predicted performance results of the corresponding areas based on the load-bearing area number, and according to stress, stiffness, and structural mass, the actual and predicted performance values ​​are read respectively. The performance difference is obtained by subtracting the predicted performance value from the actual performance value, and the absolute value of the performance difference is taken to obtain the absolute performance error. The absolute values ​​of the actual performance values ​​that are not zero in the training supervision labels of the corresponding performance type are counted, and the absolute value of the actual performance value with the smallest value is determined as the zero value normalization benchmark. When the actual performance value is not zero, the relative performance error is obtained by dividing the absolute performance error by the absolute value of the actual performance value. When the actual performance value is zero, the relative performance error is obtained by dividing the absolute performance error by the zero value normalization benchmark. The corresponding combination of load-bearing area number, performance type, actual performance value, predicted performance value, performance difference, absolute performance error, and relative performance error is used to generate the regional performance error.

[0135] The actual performance results of the region are compared with the performance constraints, and the process discrimination results are compared with the process constraints. The correction targets and directions are determined based on the regional performance error. Specifically, determining the correction targets and directions based on the regional performance error is as follows:

[0136] According to the load-bearing area number, the regional stress, regional stiffness, and corresponding performance value difference are read. Areas where the regional stress exceeds the stress threshold and areas where the regional stiffness is below the lower stiffness limit are identified as load-bearing performance correction areas. When the stress performance value difference is greater than 0, the direction of lower predicted stress is recorded; when it is less than 0, the direction of higher predicted stress is recorded. Similarly, when the stiffness performance value difference is greater than 0, the direction of lower predicted stiffness is recorded; when it is less than 0, the direction of higher predicted stiffness is recorded. For load-bearing performance correction areas with excessive stress or insufficient stiffness, material distribution, wall thickness, stiffener thickness, stiffener spacing, hole diameter, transition fillet, and cavity size are identified as correction objects. Increasing material, wall thickness, stiffener thickness, decreasing stiffener spacing, decreasing hole diameter, increasing transition fillet, and decreasing cavity size are identified as correction directions. When the volume fraction exceeds the upper limit, the regional stress is lower than the average of all load-bearing areas. Regions with stress and stiffness exceeding the lower limit of stiffness are identified as weight reduction correction regions. The correction directions are defined as material reduction, wall thickness reduction, stiffener thickness reduction, stiffener spacing increase, hole diameter increase, and cavity size increase. When stress fluctuation exceeds the stress fluctuation threshold, regions with stress higher than the average stress are corrected by load-bearing enhancement, while regions with stress lower than the average stress are corrected by material reduction. When the minimum wall thickness requirement is not met, the wall thickness is identified as the correction target and the increase direction is determined. When the draft angle requirement is not met, the cavity boundary is identified as the correction target and the direction of increasing the angle along the demolding direction is determined. When the machining limit requirement is not met, the hole system location and cavity boundary are identified as correction targets and the correction direction away from the machining limit boundary is determined. When the structural continuity requirement is not met, the material distribution at the break point is identified as the correction target and the material increase direction is determined.

[0137] The correction amount is determined based on the regional performance error range and the deviation range between the process judgment results and the process constraints. The correction object, correction direction, and correction amount are recorded according to the load-bearing area to generate a regional correction table. Specifically, the correction amount is determined based on the regional performance error range and the deviation range between the process judgment results and the process constraints.

[0138] To obtain the corresponding performance deviation ratios, divide the absolute stress performance error by the stress threshold, divide the absolute stiffness performance error by the lower stiffness limit, divide the volume fraction exceeding the upper volume fraction limit by the upper volume fraction limit, and divide the stress fluctuation value exceeding the stress fluctuation threshold by the stress fluctuation threshold. For minimum casting wall thickness, draft angle, and machining limits, take the absolute value of the difference between the actual value when the constraint is not met and the constraint limit and divide it by the constraint limit to obtain the process deviation ratio. When the constraint limit is 0, the process deviation ratio is recorded as 0. For structural continuity, subtract the process discrimination value from 0.5 and divide by 0.5 to obtain the process deviation ratio. Take the maximum value of each performance deviation ratio and process deviation ratio within the same load-bearing area as the original correction coefficient. When the value is greater than 0, it is limited to between 0.05 and 0.20; when the original correction coefficient is 0, no correction is performed. For wall thickness, rib thickness, rib spacing, hole diameter, transition fillet, and cavity size, the upper limit of the parameter determined by process constraints and historical samples is subtracted from the lower limit of the parameter, and the resulting parameter range is multiplied by the correction coefficient to obtain the parameter correction amount. For material distribution, the correction coefficient is multiplied by 3 and rounded up to obtain the number of grid increases or decreases for layers 1 to 3. For hole system position and cavity boundary, the basic grid side length is multiplied by the number of grid increases or decreases to obtain the position correction amount. For draft angle, the difference between the minimum draft angle required by the process constraints and the current draft angle is determined as the angle correction amount. The correction object, correction direction, and correction amount are recorded according to the load-bearing area to generate a region correction table.

[0139] In this embodiment, updating the finite element mesh according to the configuration change region includes:

[0140] The region correction table is input into the improved Swin-Unet model. The features of the corresponding carrying regions are adjusted according to the correction object, correction direction, and correction amount to generate the updated topology and structural parameters. Specifically, adjusting the features of the corresponding carrying regions according to the correction object, correction direction, and correction amount to generate the updated topology and structural parameters is as follows:

[0141] According to the bearing area number in the area correction table, read the mesh index contained in the bearing area node, and determine the material distribution characteristic channel or structural parameter characteristic channel corresponding to the correction object. When the correction object is material distribution, according to the mesh increase / decrease amount recorded in the area correction table, select adjacent meshes of the shared mesh surface layer by layer starting from the material boundary of the corresponding bearing area. In the direction of material increase, set the material distribution characteristic value of the selected mesh to 1, and in the direction of material decrease, set the material distribution characteristic value of the selected mesh to 0. When the correction object is wall thickness, stiffener thickness, stiffener spacing, hole diameter, transition fillet, or cavity size, divide the correction amount by the difference between the upper limit and lower limit of the corresponding structural parameter to obtain... To obtain the normalized correction amount, the increasing direction adds the normalized correction amount to the corresponding structural parameter feature value, while the decreasing direction subtracts the normalized correction amount from the corresponding structural parameter feature value, and the adjusted feature value is limited to between 0 and 1. When the correction object is the hole system position or cavity boundary, the position correction amount is divided by the basic mesh side length and rounded to obtain the number of moving meshes, and the corresponding position feature is moved according to the correction direction. The adjusted material distribution features, structural parameter features, hole system position features, and cavity boundary features are written into the initial mapping feature map of the corresponding bearing area, input into the trained improved Swin-Unet model, and the updated topology and structural parameters are generated.

[0142] The updated topology configuration is compared with the original topology configuration by mesh correspondence, and the meshes where the material retention state changes and their adjacent meshes are identified as configuration change regions.

[0143] The finite element mesh of the configuration change region is redefined based on the updated topology and structural parameters, and then connected to the boundary of the finite element mesh outside the configuration change region to generate an updated finite element model. Specifically, the redefined finite element mesh of the configuration change region based on the updated topology and structural parameters is as follows:

[0144] Delete the original finite element elements in the configuration change zone, and retain the interface between the boundary of the configuration change zone and the external finite element mesh; extract the material boundary in the configuration change zone according to the updated material retention state, and reconstruct the local geometric entity based on the wall thickness, stiffener thickness, stiffener spacing, hole diameter, transition fillet, and cavity size; divide tetrahedral finite element elements with one-third of the minimum casting wall thickness as the basic mesh edge length, and reduce the mesh edge length of the hole system edge, transition fillet, cavity boundary, and material addition / reduction boundary to half of the basic mesh edge length; calculate the ratio of the longest side to the shortest side of each finite element element, delete finite element elements with a ratio greater than 5 and re-divide, until the ratio of the longest side to the shortest side of all finite element elements in the configuration change zone is no greater than 5;

[0145] The process involves connecting the boundary of the finite element mesh outside the configuration change region to generate an updated finite element model, specifically as follows:

[0146] Extract the new finite element nodes on the boundary of the configuration change zone and the boundary nodes of the finite element mesh outside the configuration change zone, and calculate the spatial distance between the new finite element nodes and each boundary node; when the distance is no greater than 0.01 mm, merge the corresponding nodes and unify the node numbers; when there are no boundary nodes that can be merged, project the new finite element nodes onto the boundary surface of the outer finite element element with the smallest distance, divide the corresponding boundary surface at the projection position and establish connection nodes; use tetrahedral elements to fill the gaps between the new finite element nodes and the boundary nodes, so that the finite element elements inside and outside the configuration change zone share the boundary nodes and boundary surfaces; re-associate the bearing area number according to the center coordinates of the finite element elements, and map the material properties, loads and support constraints from the original finite element elements or original finite element nodes at the same geometric position to the corresponding new finite element elements and new finite element nodes to generate an updated finite element model;

[0147] The updated topology and structural parameters are input into the performance evaluation subnetwork, and the updated topology, structural parameters, and process constraints are input into the process discrimination subnetwork to generate updated predicted performance results and process discrimination results.

[0148] The updated finite element model is recalculated using finite element analysis (FEM) to generate true performance results for the region. The true performance results are compared with the predicted performance results. The region performance error and region correction table are updated based on the process discrimination results. This process iteratively executes topology and structural parameter generation, performance evaluation, process discrimination, and FEM recalculation. Specifically, the FEM recalculation of the updated finite element model to generate true performance results for the region involves:

[0149] The element stiffness matrix is ​​calculated based on the nodal coordinates, elastic modulus, and Poisson's ratio of the new finite element elements within the configuration change region. The element stiffness matrices of the finite element elements outside the configuration change region are retained. All element stiffness matrices are superimposed according to node numbers to form an updated overall stiffness matrix. Loads and support constraints are mapped to updated node numbers to form an updated overall load vector, and the nodal displacements in the support constraint directions are set to 0. The conjugate gradient method is used to solve the equation where the product of the updated overall stiffness matrix and the nodal displacement vector equals the updated overall load vector. The solution is stopped when the ratio of the residual to the length of the updated overall load vector is less than 0.000001. Based on the nodal displacements, the strain, stress, and equivalent stress of each finite element are calculated. According to the bearing region associated with the finite element, the maximum equivalent stress within the region is determined as the regional stress. The regional load is then combined... The region stiffness is obtained by dividing the quantity by the maximum nodal displacement in the load direction. According to the load-bearing region associated with the finite element element, the volume of each finite element element within each load-bearing region is multiplied by the material density, and the sum of these products is obtained to get the regional structural mass of each load-bearing region. The total structural mass of all regions is then added together to obtain the structural mass. The volume fraction is obtained by dividing the sum of the volumes of all finite element elements by the total volume of the design domain. The square of the difference between the stress in each region and the average stress in the region is calculated. The average of the squared differences is taken, and the square root is obtained. The standard deviation of the regional stress is divided by the average stress in the region to obtain the stress fluctuation value. When the average stress in the region is 0, the stress fluctuation value is recorded as 0. The load-bearing region number, regional stress, regional stiffness, and regional structural mass are combined accordingly, and the structural mass, volume fraction, and stress fluctuation value are associated to generate the true performance results of the region.

[0150] In this embodiment, the output topology configuration and structural parameter optimization results include:

[0151] Set regional error thresholds, error stability thresholds, continuous judgment rounds, and maximum iteration rounds. Record the regional performance error of each carrying area according to the iteration rounds. Calculate the maximum absolute value of the regional performance error in each round, and calculate the absolute value of the difference between the regional performance errors of corresponding areas in adjacent rounds. Specifically, the calculation of the maximum absolute value of the regional performance error in each round and the calculation of the absolute value of the difference between the regional performance errors of corresponding areas in adjacent rounds are as follows:

[0152] Set both the regional error threshold and the error stability threshold to 0.01, set the continuous judgment rounds to 3 rounds, and set the maximum iteration rounds to 15 rounds. Establish error record items according to the iteration round, bearing area number, and performance type. Take the absolute values ​​of the stress relative performance error, stiffness relative performance error, and structural quality relative performance error of each bearing area within the same iteration round to form a set of absolute values ​​of regional performance error for the current iteration round. Take the first value in the set as the initial maximum value, and compare the remaining values ​​with the initial maximum value in turn. Replace the initial maximum value when the value is larger. After completing all value comparisons, determine the retained value as the maximum value of the absolute value of regional performance error for the current iteration round. Starting from the second iteration round, match the relative performance error of the current iteration round with the relative performance error of the previous iteration round item by item according to the bearing area number and performance type. Subtract the relative performance error of the previous iteration round from the relative performance error of the current iteration round, and take the absolute value of the difference to obtain the absolute value of the difference of regional performance error between adjacent rounds. The absolute value of the difference of regional performance error between adjacent rounds is not calculated in the first iteration round.

[0153] Automatic iteration terminates when the maximum absolute value of the regional performance error within consecutive evaluation rounds is not greater than the regional error threshold, and the absolute value of the difference between regional performance errors in adjacent rounds is not greater than the error stability threshold; automatic iteration terminates when the maximum number of iteration rounds is reached; otherwise, topology configuration and structural parameter generation, performance evaluation, and finite element recalculation continue, with the termination of automatic iteration specifically as follows:

[0154] Starting from the 3rd iteration, the maximum absolute value of the regional performance error in the most recent 3 iterations is read, along with the absolute value of the difference in regional performance error between adjacent iterations in the most recent 3 iterations. Automatic iteration terminates when the maximum absolute value of the regional performance error in the most recent 3 iterations is no greater than 0.01, and the absolute value of the difference in regional performance error between adjacent iterations in the most recent 3 iterations is no greater than 0.01. Automatic iteration terminates when the error termination condition is not met and the iteration count reaches 15. If the error termination condition is not met and the iteration count has not reached 15, the current regional correction table is input into the improved Swin-Unet model, and topology configuration and structural parameter generation, performance evaluation, process discrimination, and finite element recalculation continue.

[0155] From the topology configurations and structural parameters corresponding to consecutive decision rounds, select the round with the smallest absolute value of regional performance error, and determine the corresponding topology configuration and structural parameters as the optimization result of topology configuration and structural parameters.

[0156] Example 1: To verify the feasibility of this invention in practice, it was applied to the optimization of key structural parameters of a certain type of six-cylinder diesel engine block. The original three-dimensional model of the engine block includes cylinder bores, main bearing housings, crankcase, reinforcing ribs, and cooling water chambers. The system collected 126 structural parameters, including wall thickness, rib thickness, rib spacing, bore diameter, transition fillet radius, and cavity dimensions, as well as burst pressure, bolt preload, crankshaft and connecting rod load, valve train load, and external load. The maximum burst pressure was set to 18.6 MPa, the cylinder head bolt preload to 128 kN, and the peak load of the main bearing housing to 96 kN.

[0157] The system correlates structural parameters, load boundaries, performance constraints, process constraints, and historical topology configurations to form topology optimization data. In the performance constraints, the upper limit for volume fraction is set to 0.78, the equivalent stress threshold is set to 265 MPa, the stress fluctuation threshold is set to 18%, and the lower limit for stiffness is set to 1116.3 kN / mm, corresponding to a maximum displacement of the main bearing housing of no more than 0.086 mm under a peak load of 96 kN. In the process constraints, the minimum casting wall thickness is set to 4.5 mm, the lower limit for draft angle is set to 1.5°, and the lower limit for machining allowance is set to 2.0 mm.

[0158] Based on the load location and direction, the system divides the engine block into cylinder bore load area, main bearing housing load area, cylinder head bolt load area, crankcase load area, and peripheral support area, generating a total of 38 load area nodes. The system establishes 62 connecting edges based on the geometric contact relationship between adjacent areas and the load transfer direction, and associates various constraints with the corresponding load area nodes to form a region constraint diagram.

[0159] Topology optimization data and region constraint maps are input into the improved Swin-Unet model. The input mapping layer divides the 3D design domain into 128×96×96 mesh elements and maps structural parameters, loads, and constraint features into 64-dimensional initial features. The shift window encoder adopts a 4-level encoding structure with a window size of 4×4×4 and channel numbers of 64, 128, 256, and 512 respectively, extracting material distribution, structural boundary, and load-bearing region correlation features.

[0160] In the dual-constraint attention skip layer, the system aggregates encoded features by load-bearing region nodes and calculates performance constraint matching weights and process constraint matching weights. The two matching weights for the main bearing housing load-bearing region are 0.82 and 0.71, respectively, and the fusion weight for adjacent crankcase load-bearing regions is 0.63. After weighting, the feature response of the main bearing housing region increases from 0.54 to 0.79.

[0161] The shift-window decoder performs upsampling step by step and concatenates it with the double-constrained jump features at the corresponding scale to generate a topology decoding feature map. The parameter configuration co-output layer generates the material retention probability of each mesh, with the configuration generation threshold set to 0.5. Mesh with a material retention probability not lower than the threshold and connected to the existing material region are retained, ultimately generating 6 sets of candidate topology configurations and their corresponding structural parameters.

[0162] The system inputs candidate topologies and structural parameters into the performance evaluation sub-network. The configuration encoding branch and parameter encoding branch generate configuration encoding features and parameter encoding features, respectively. After processing by the region fusion layer, the performance regression layer outputs the predicted stress, stiffness, and structural mass values ​​for each load-bearing region. For the second group of candidate configurations, the maximum predicted stress is 247.8 MPa, the predicted stiffness of the main bearing pedestal is 1230.8 kN / mm, the corresponding predicted displacement is 0.078 mm, the predicted structural mass is 392.6 kg, the volume fraction is 0.742, and the stress fluctuation is 14.3%.

[0163] The candidate results were simultaneously input into the process discrimination sub-network. The second group of candidate configurations had a minimum wall thickness of 5.1 mm, a minimum draft angle of 1.8°, a machining allowance of 2.4 mm, and a structural continuity discrimination value of 0.93, satisfying the process constraints. After joint screening based on performance and process, the second group of candidate configurations was determined as the initial optimized configuration.

[0164] The system reconstructs the engine's structural geometric model based on the initial optimized configuration and divides it into approximately 1.86 million finite element elements. After load application and finite element solution, the maximum equivalent stress is 256.4 MPa, the true value of the main bearing housing stiffness is 1185.2 kN / mm, the corresponding maximum displacement is 0.081 mm, the structural mass is 396.1 kg, the volume fraction is 0.748, and the stress fluctuation value is 15.1%, forming the true performance results of the region.

[0165] The system compares the actual performance results with the predicted performance results for each region, categorized by load-bearing area and performance type. The stress performance difference in the main bearing housing load-bearing area is 8.6 MPa, the stiffness performance difference is -45.6 kN / mm, and the absolute stiffness performance error is 45.6 kN / mm; the stress performance difference in the cylinder bore load-bearing area is 5.9 MPa. Based on the process judgment results, the system identifies the main bearing housing stiffener thickness and transition fillet as correction targets, with correction amounts of 0.7 mm and 1.2 mm respectively, generating a region correction table.

[0166] The region correction table is input into the improved Swin-Unet model to generate updated topology and structural parameters. The system compares the material retention status of the meshes before and after the update, identifying a total of 12,460 changed meshes and classifying adjacent meshes into the configuration change region. The configuration change region contains 31,872 meshes, accounting for 1.71% of the total mesh. The system only re-meshes the finite element meshes in the configuration change region and connects them to the remaining meshes, reducing the single-round mesh update time from 47 minutes to 9 minutes.

[0167] The updated topology and structural parameters are re-entered into the performance evaluation subnetwork and process discrimination subnetwork, and the updated finite element model is recalculated. The maximum number of iterations is set to 15 rounds, and the regional error threshold and error stability threshold are both set to 0.01. After the first 6 iterations, the regional performance error continues to decrease. The maximum absolute values ​​of the regional performance error in the 7th, 8th, and 9th rounds are 0.0096, 0.0088, and 0.0080, respectively. The maximum absolute value of the difference between the regional performance errors in the 7th and 8th rounds is 0.0008, and the maximum absolute value of the difference between the regional performance errors in the 8th and 9th rounds is also 0.0008. The maximum absolute values ​​of the regional performance error from the 7th to the 9th rounds are all no greater than 0.01, and the absolute value of the difference between the regional performance errors in adjacent rounds is also no greater than 0.01. The system terminates automatic iteration in the 9th round.

[0168] In the final output topology, the structural mass was reduced from 421.8 kg to 389.7 kg, a weight reduction of 7.61%; the maximum equivalent stress was reduced from 279.5 MPa to 251.2 MPa; the maximum displacement of the main bearing housing was reduced from 0.094 mm to 0.079 mm; and the minimum casting wall thickness, draft angle, and machining allowance all met the process requirements. Compared with the traditional global topology optimization method, the single-round mesh update time of this invention is reduced by 80.9%, and the total iteration time is shortened from 18.6 h to 7.4 h.

[0169] As can be seen from Example 1, the present invention can collaboratively generate topology configuration and structural parameters based on the load transfer relationship and constraints of different load-bearing areas of the engine, and drive local mesh updates and finite element recalculation through regional performance errors. While meeting performance constraints and process constraints, it can improve optimization accuracy and iteration efficiency, thus verifying the engineering feasibility of the present invention in the automatic optimization of key structural parameters of the engine.

[0170] The above description is only a preferred embodiment 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 deep learning-based method for optimizing and automatically iterating key structural parameters of an engine, characterized in that, include: Collect structural parameters, load boundaries, performance constraints, process constraints, and historical topology configurations to generate topology optimization data. Divide the load-bearing areas according to load location and direction and associate constraints to generate a region constraint map. The topology optimization data and region constraint graph are input into the improved Swin-Unet model to generate topology configuration and structural parameters. The improved Swin-Unet model includes an input mapping layer, a shift window encoder, a double-constraint attention jump layer, a shift window decoder, and a parameter configuration collaborative output layer. Input the topology and structural parameters into the performance evaluation subnetwork and the process discrimination subnetwork to generate predicted performance results and process discrimination results, and select the initial optimized configuration and corresponding structural parameters. A finite element model is established based on the initial optimized configuration and corresponding structural parameters. Load boundaries are applied, and the actual performance results of the region are generated. Compare the actual performance results and predicted performance results of the region, and generate a regional performance error and regional correction table based on the process discrimination results; Input the region correction table into the improved Swin-Unet model, update the finite element mesh according to the configuration change region, and iteratively execute the topology configuration and structural parameter generation, performance evaluation and finite element recalculation. The process terminates after determining the regional performance error in multiple rounds, and outputs the optimization results of the topology configuration and structural parameters.

2. The method for optimizing and automatically iterating key structural parameters of an engine based on deep learning according to claim 1, characterized in that, The structural parameters include wall thickness, stiffener thickness, stiffener spacing, hole diameter, transition fillet, and cavity dimensions; the load boundaries include burst pressure, bolt preload, crankshaft and connecting rod load, valve train load, and external load corresponding to load values, load locations, load directions, and support constraints; the performance constraints include upper limit of volume fraction, stress threshold, stress fluctuation threshold, and lower limit of stiffness; the process constraints include minimum casting wall thickness, draft angle, machining limits, and structural continuity; the historical topology includes material distribution, stiffener layout, hole system location, and cavity boundaries.

3. The method for optimizing and automatically iterating key structural parameters of an engine based on deep learning according to claim 1, characterized in that, The generated region constraint map includes: By associating structural parameters, load boundaries, performance constraints, process constraints, and historical topology configurations, topology optimization data is generated, and the load location and load direction are determined based on the load boundaries. Based on the load location and load direction, combined with material distribution, stiffener layout, hole system location and cavity boundary, the load-bearing area is divided, and the load-bearing area nodes and area connection relationships are generated. Structural parameters, performance constraints, and process constraints are associated with nodes in the load-bearing region. Connection edges are established based on the region connection relationships and load directions to generate a region constraint diagram.

4. The method for optimizing and automatically iterating key structural parameters of an engine based on deep learning according to claim 1, characterized in that, The generated topology and structural parameters include: The topology optimization data and region constraint map are input into the input mapping layer of the improved Swin-Unet model. Mesh mapping and feature stitching are performed on structural parameters, load boundaries, performance constraints, process constraints, historical topology configurations and load-bearing region nodes. An initial mapping feature map is generated through linear transformation. The initial mapped feature map is input into the shift window encoder, and window attention, shift window attention and hierarchical downsampling are performed to extract the associated features of material distribution, structural boundary and load-bearing area, and generate multi-scale encoded feature groups; The multi-scale coded feature group, regional constraint map, performance constraint and process constraint are input into the dual-constraint attention jump layer. The coded features are aggregated according to the nodes of the carrying area. The matching weights of performance constraint and process constraint are calculated respectively. The matching weights of adjacent carrying areas are fused according to the connecting edges. The corresponding scale coded features are weighted to generate the multi-scale dual-constraint jump feature group. The deepest scale encoded features are input into the shift window decoder, upsampling is performed scale by scale, and concatenated with the double-constrained jump features of the corresponding scale. Window attention and shift window attention are then performed to generate a topological decoding feature map. The topology decoding feature map is input to the parameter configuration and output layer to generate the mesh material retention probability. The topology configuration is generated based on the configuration generation threshold and mesh connectivity. The topology configuration is used to perform region aggregation on the topology decoding feature map to generate configuration parameter feature vectors. The structural parameters are then generated from the configuration parameter feature vectors. Using topology optimization data and region constraint graphs as training inputs, and historical topology configurations and corresponding structural parameters as supervision labels, the topology configuration loss, structural parameter loss, and region constraint consistency loss are calculated. These are then weighted to generate training optimization objectives and update the parameters of the improved Swin-Unet model until the training optimization objectives converge, resulting in the trained improved Swin-Unet model.

5. The method for optimizing and automatically iterating key structural parameters of an engine based on deep learning according to claim 1, characterized in that, The generated prediction performance results include: The performance evaluation subnetwork includes a configuration coding branch, a parameter coding branch, a region fusion layer, and a performance regression layer. It performs finite element analysis on historical topological configurations and corresponding structural parameters, and generates stress monitoring labels, stiffness monitoring labels, and structural quality monitoring labels according to the load-bearing region. The historical topology and corresponding structural parameters are input into the performance evaluation subnetwork. The performance training loss is calculated based on the stress supervision label, stiffness supervision label and structural quality supervision label. The parameters of the configuration encoding branch, parameter encoding branch, region fusion layer and performance regression layer are updated to obtain the trained performance evaluation subnetwork. The topology and structural parameters are input into the trained performance evaluation subnetwork. Configuration coding features and parameter coding features are generated through configuration coding branch and parameter coding branch, respectively. After being fused by the region fusion layer, they are input into the performance regression layer to generate the stress prediction value, stiffness prediction value and structural mass prediction value of each load-bearing region. The volume fraction is calculated based on the topology and the stress fluctuation value is calculated based on the stress prediction value of each load-bearing region to form the predicted performance result.

6. The method for optimizing and automatically iterating key structural parameters of an engine based on deep learning according to claim 1, characterized in that, The generation process judgment results and the initial optimized configuration and corresponding structural parameters screening include: The process discrimination subnetwork includes a configuration coding branch, a parameter constraint coding branch, a feature fusion layer, and a process discrimination layer. It compares historical topological configurations and corresponding structural parameters with process constraints to generate process supervision labels. The historical topology, corresponding structural parameters, and process constraints are input into the process discrimination subnetwork to generate process training discrimination values. The process training loss is calculated based on the process training discrimination values ​​and process supervision labels. The parameters of the configuration encoding branch, parameter constraint encoding branch, feature fusion layer, and process discrimination layer are updated to obtain the trained process discrimination subnetwork. The topology configuration, structural parameters, and process constraints are input into the trained process discrimination subnetwork. Configuration process features and parameter constraint features are generated through configuration encoding branch and parameter constraint encoding branch, respectively. After being fused by the feature fusion layer, they are input into the process discrimination layer to generate the process discrimination result. The predicted performance results are compared with the performance constraints. Combined with the process discrimination results, topological configurations and structural parameters that do not meet the performance constraints or process constraints are screened out. The remaining topological configurations are arranged in ascending order of the predicted structural quality values. The first topological configuration and its corresponding structural parameters are determined as the initial optimized configuration and its corresponding structural parameters.

7. The method for optimizing and automatically iterating key structural parameters of an engine based on deep learning according to claim 1, characterized in that, The finite element model is established based on the initial optimized configuration and corresponding structural parameters, and the applied load boundaries include: Based on the initial optimized configuration, the material retention area is determined, and the wall thickness, stiffeners, hole system, transition fillets and cavities are reconstructed according to the corresponding structural parameters to generate the engine structural geometric model. The engine structure geometric model is meshed using the region constraint diagram, the correspondence between finite element elements and load-bearing regions is established, and the load boundary is applied to the corresponding finite element mesh to generate the finite element model. The finite element model is solved by performing finite element analysis, and the stress and stiffness are collected according to the load-bearing region. The structural mass, volume fraction and stress fluctuation value are calculated to generate the actual performance results of the region.

8. The method for optimizing and automatically iterating key structural parameters of an engine based on deep learning according to claim 1, characterized in that, The generation of the regional performance error and regional correction table based on the combined process discrimination results includes: The actual performance results and predicted performance results of the region are compared according to the bearing area and performance type, and the performance value difference is calculated to generate the regional performance error. The actual performance results of the region are compared with the performance constraints, and the process discrimination results are compared with the process constraints. The correction objects and correction directions are determined by combining the regional performance error. The correction amount is determined based on the regional performance error range and the deviation range between the process judgment result and the process constraint. The correction object, correction direction and correction amount are recorded according to the bearing area to generate a regional correction table.

9. The method for optimizing and automatically iterating key structural parameters of an engine based on deep learning according to claim 1, characterized in that, The finite element mesh update according to the configuration change region includes: Input the region correction table into the improved Swin-Unet model, adjust the characteristics of the corresponding bearing region according to the correction object, correction direction and correction amount, and generate the updated topology and structural parameters. The updated topology configuration is compared with the original topology configuration by mesh correspondence, and the meshes where the material retention state changes and their adjacent meshes are identified as configuration change regions. The finite element mesh of the configuration change region is redefined based on the updated topology and structural parameters, and the boundary is connected with the finite element mesh outside the configuration change region to generate an updated finite element model. The updated topology and structural parameters are input into the performance evaluation subnetwork, and the updated topology, structural parameters, and process constraints are input into the process discrimination subnetwork to generate updated predicted performance results and process discrimination results. The updated finite element model is recalculated using finite element analysis to generate the actual performance results of the region. The actual performance results of the region are compared with the predicted performance results. The region performance error and region correction table are updated in combination with the process discrimination results. The process of generating topology and structural parameters, performance evaluation, process discrimination and finite element analysis is repeated cyclically.

10. The method for optimizing and automatically iterating key structural parameters of an engine based on deep learning according to claim 1, characterized in that, The output topology configuration and structural parameter optimization results include: Set the regional error threshold, error stability threshold, continuous judgment rounds and maximum iteration rounds. Record the regional performance error of each carrying area according to the iteration rounds. Calculate the maximum value of the absolute value of the regional performance error in each round and calculate the absolute value of the difference between the regional performance errors of adjacent rounds. Automatic iteration terminates when the maximum absolute value of the regional performance error within consecutive judgment rounds is not greater than the regional error threshold, and the absolute value of the difference between regional performance errors in adjacent rounds is not greater than the error stability threshold; automatic iteration terminates when the number of iteration rounds reaches the maximum number of iteration rounds; otherwise, topology configuration and structural parameter generation, performance evaluation, and finite element recalculation continue. From the topology configurations and structural parameters corresponding to consecutive decision rounds, select the round with the smallest absolute value of regional performance error, and determine the corresponding topology configuration and structural parameters as the optimization result of topology configuration and structural parameters.

Images