A method and system for integrated die-casting rear floor reinforcement rib collaborative optimization

Abstract

The application provides a kind of integrated die-casting rear floor reinforcing rib collaborative optimization method and system, comprising: by the same division of real object and homologous process simulation, based on the mapping relationship of process, defect and mechanical property of the same division, the non-homogeneous material constitutive parameter is given to the finite element model division; the die-casting process and reinforcing rib structure are carried out full-parameterization linkage modeling, and the dynamic division rule is set, and the structure parameter adjustment is automatically updated when the threshold value is exceeded The constitutive of division is updated; a cascaded double prediction model composed of physical information neural network and graph attention network is constructed; based on the model, an agent environment is constructed, and a reinforcement learning is used to train a collaborative optimization agent, with the current process, structure, defect and performance as the state, with the parameter adjustment amount as the action, with the mechanical property, manufacturability and lightweight as the reward, and the optimal parameter combination is output by iterative optimization. The application realizes global collaborative optimization of process and structure, and considers crash performance, lightweight and manufacturability.

Description

Technical Field

[0001] This invention belongs to the field of die casting process and structure synergistic optimization technology, specifically relating to a synergistic optimization method and system for integrated die casting floor reinforcement. Background Technology

[0002] With the rapid development of the new energy vehicle industry, integrated die casting technology has become a core technology direction for new energy vehicle body manufacturing due to its advantages of significantly reducing the number of body parts, shortening the manufacturing cycle, and improving the rigidity and collision performance of the body structure. Among them, the integrated die-cast rear floor, as a core load-bearing component of the new energy vehicle body, directly determines the collision safety performance, NVH performance, and lightweight level of the body through its reinforcing rib structure design and die casting process parameters. Currently, the research and design process of integrated die-cast rear floor mainly faces the following unresolved core technical challenges: During the filling and solidification process of the integrated die-cast part, there are significant differences in the cooling rate and filling pressure in different regions, resulting in uneven distribution of defects such as porosity and shrinkage in different regions of the casting, and the material mechanical properties exhibit strong heterogeneity. However, existing technologies generally use ideal homogeneous material models for finite element simulation and structural optimization, which cannot reflect the material distribution characteristics of actual die-cast parts; most existing technologies adopt a serial research and development mode of first optimizing the structure and then verifying the process, first completing the reinforcing rib structure design based on the homogeneous model, and then optimizing the die casting process for the fixed structure. This model is prone to contradictions such as structural performance meeting standards but process failure, and process compliance but structural mechanical performance not meeting requirements. It cannot simultaneously consider the mechanical performance of the structure, the level of lightweighting, and the manufacturability of the die-casting process, nor can it achieve global collaborative optimization of process parameters and structural parameters. The pure data-driven process defect prediction model does not incorporate the physical laws of the die-casting process, has poor generalization to changes in process parameters, and traditional fully connected neural networks cannot accurately capture the impact of changes in the topology of the reinforcing ribs on the mechanical transmission path, resulting in low accuracy in mechanical performance prediction and easy optimization getting trapped in local optima. Summary of the Invention

[0003] The present invention aims to at least partially solve one of the technical problems in the related art.

[0004] Therefore, the first objective of this invention is to propose a method for compressing and storing historical data in a hydropower monitoring system.

[0005] The second objective of this invention is to propose a synergistic optimization system for integrated die-cast floor reinforcement.

[0006] The third objective of this invention is to provide a computer device.

[0007] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0008] To achieve the above objectives, a first aspect of the present invention proposes a method for co-optimization of floor reinforcement ribs after integral die casting, comprising: Based on the actual measurement of the integral die-cast floor in the physical object, the measured mechanical property parameters and measured defect parameters are obtained. The simulation model is then subjected to the same-source partition process simulation to obtain the process simulation parameters and simulation defect parameters. Based on the mapping relationship between process simulation parameters, simulation defect parameters and measured mechanical property parameters of the same partition, a process-mechanical property mapping model specific to the partition is constructed, and constitutive parameters of heterogeneous materials are assigned to the corresponding partition of the finite element model based on the mapping model, and a finite element calculation model of heterogeneous materials of the partition is established. Full-parameter linkage modeling is performed on the die casting process parameters and the stiffener structural parameters, and adaptive dynamic partitioning rules are set for the structure. When the adjustment range of the stiffener structural parameters exceeds the preset threshold, the dynamic adjustment of the partition is automatically triggered, and the constitutive parameters of the heterogeneous materials in each partition are updated synchronously. A cascaded dual prediction model is constructed, comprising a physical information neural network process defect prediction model and a graph attention network mechanical performance prediction model. An agent environment is constructed based on the cascaded dual prediction model, and a reinforcement learning algorithm is used to train a collaborative optimization agent. The collaborative optimization agent uses the current process parameters, structural parameters, defect parameters, and mechanical performance indicators as its state space, takes the adjustment amount of process parameters and structural parameters as its action, constructs a reward function with mechanical performance, process manufacturability, and lightweight objectives, and outputs the optimal combination of process parameters and stiffener structural parameters through interactive iterative optimization with the environment.

[0009] In one embodiment of the present invention, the partition measurement includes: Feature recognition is performed on the solid die-cast floor, extracting the melt filling flow direction, collision force transmission path and hot spot defect distribution characteristics. The floor is automatically partitioned according to the preset quantization gradient threshold to form multiple functional partitions, and the partition vector file with precise coordinate boundaries is output. Multiple parallel standard tensile specimens were taken from each partition for tensile testing to obtain the mechanical property parameters of each partition. Defect detection specimens taken from the same location were also tested to obtain the defect parameters of each partition.

[0010] In one embodiment of the present invention, the process simulation of homogeneous partitioning of the simulation model includes: A three-dimensional simulation model of the die casting process of the integrated die-casting post-die casting floor was established using die casting simulation software. The partition vector file was imported into the simulation model, and the simulation model was partitioned to ensure that the boundary coordinates, number of partitions, and partition functions of the simulation partitions and the physical partitions were consistent. Perform coupled simulation of melt filling and solidification process, and extract process simulation parameters and simulation defect parameters for each zone according to the corresponding zone.

[0011] In one embodiment of the present invention, the construction of a partition-specific process-mechanical performance mapping model includes: Using process simulation parameters and defect parameters of the same partition as input, and measured mechanical performance parameters of the same partition as output, a mapping model in the form of a neural network is constructed.

[0012] In one embodiment of the present invention, establishing the finite element calculation model of the partitioned heterogeneous material includes: A finite element model of the integrated die-cast floor collision was established. The partition vector file was imported into the finite element model and partitioned. For each partition, the process simulation parameters and defect parameters of that partition were input into the trained mapping model to obtain the material constitutive parameters of that partition. Each partition was assigned a corresponding differentiated material constitutive model, and collision simulation verification was performed.

[0013] In one embodiment of the present invention, the step of performing fully parametric linkage modeling and setting structural adaptive dynamic partitioning rules includes: Screening key process parameters that affect die casting defects and mechanical properties, as well as key structural parameters that affect impact performance and formability; When the adjustment range of the stiffener structure parameters exceeds the preset threshold, the dynamic adjustment of the partition is automatically triggered. Based on the adjusted stiffener structure, the partition boundary is recalculated, the process-mechanical property mapping model is updated synchronously, and the corresponding heterogeneous material constitutive parameters are assigned. The graph structure of the graph attention network mechanical property prediction model is also updated synchronously.

[0014] In one embodiment of the present invention, the construction of the cascaded dual prediction model includes: The physical information neural network process defect prediction model takes die casting process parameters as input, outputs defect parameters of each zone, and uses the physical conservation equation of the die casting process as the physical constraint embedding loss function. It is trained by weighted fusion of data fitting loss and physical constraint loss. The graph attention network mechanical performance prediction model transforms the stiffener topology into a graph structure, integrates node features with structural parameters, defect parameters, and heterogeneous material parameters, and outputs mechanical performance indicators.

[0015] To achieve the above objectives, a second aspect of the present invention provides a synergistic optimization system for integrated die-cast floor reinforcement, comprising: The physical homogeneous partitioning measurement module is used to perform partitioning measurements on the physical integral die-cast floor to obtain measured mechanical property parameters and measured defect parameters, and to perform homogeneous partitioning process simulation on the simulation model to obtain process simulation parameters and simulation defect parameters. The simulation module for the same origin partition process is used to construct a partition-specific process-mechanical property mapping model based on the mapping relationship between the process simulation parameters, simulation defect parameters and measured mechanical property parameters of the same partition, and to assign heterogeneous material constitutive parameters to the corresponding partition of the finite element model based on the mapping model, thereby establishing a partition heterogeneous material finite element calculation model. The partitioned heterogeneous material modeling module is used to perform fully parametric linkage modeling of die casting process parameters and stiffener structural parameters, and to set adaptive dynamic partitioning rules for the structure. When the adjustment range of stiffener structural parameters exceeds the preset threshold, the partition dynamic adjustment is automatically triggered, and the constitutive parameters of heterogeneous materials in each partition are updated synchronously. A parameterized fully coupled modeling module is used to construct a cascaded dual prediction model, which includes a physical information neural network process defect prediction model and a graph attention network mechanical performance prediction model. A cascaded dual prediction model module is used to construct an agent environment based on the cascaded dual prediction model and to train a collaborative optimization agent using a reinforcement learning algorithm. The collaborative optimization agent module is used to make the collaborative optimization agent use the current process parameters, structural parameters, defect parameters and mechanical performance indicators as the state space, the adjustment amount of process parameters and structural parameters as actions, and construct a reward function with mechanical performance, process manufacturability and lightweight objectives. Through interactive and iterative optimization with the environment, it outputs the optimal combination of process parameters and stiffener structural parameters.

[0016] The integrated die-cast floor reinforcement synergistic optimization method and system of this invention has the following beneficial effects: 1. This invention uses a completely consistent quantization partitioning scheme to divide the physical casting and the simulation model, realizing a one-to-one correspondence between the measured mechanical properties of the partition and the simulation data of the partition process, and establishing a precise partition-specific process-mechanical property mapping model, breaking through the accuracy bottleneck of traditional homogeneous material models.

[0017] 2. This invention embeds the physical conservation equations of the die-casting process into the model using the PINN physical information neural network, significantly reducing the reliance on large amounts of simulation data and improving the generalization and accuracy of process defect prediction. Furthermore, it accurately captures the mechanical transmission characteristics of the reinforcing rib topology through the GAT graph attention network, while simultaneously coupling the material heterogeneity effects caused by process defects, achieving for the first time a precise end-to-end mapping of the entire process link: process parameters, defect distribution, material heterogeneity, and structural mechanical properties.

[0018] 3. This invention uses a dual-network prediction model as a proxy model to build a PPO reinforcement learning collaborative optimization agent, which simultaneously optimizes die-casting process parameters and reinforcing rib structural parameters, while taking into account process manufacturability, collision mechanical properties and lightweight objectives.

[0019] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing a method for co-optimizing integrated die-cast floor reinforcement as described in the first aspect embodiment.

[0020] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for co-optimization of integrated die-cast floor reinforcement as described in the first aspect embodiment.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for coordinating and optimizing floor reinforcement ribs after integral die casting according to an embodiment of the present invention; Figure 2 This is a structural diagram of an integrated die-cast floor reinforcement synergistic optimization system according to an embodiment of the present invention; Figure 3 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] The following describes, with reference to the accompanying drawings, a method and system for coordinating optimization of integrated die-cast floor reinforcement ribs according to an embodiment of the present invention.

[0026] Example 1 Figure 1 This is a flowchart of a method for co-optimizing floor reinforcement ribs after integral die casting according to an embodiment of the present invention, as shown below. Figure 1 As shown, it includes: Based on the actual measurement of the integral die-cast floor in the physical object, the measured mechanical property parameters and measured defect parameters are obtained. The simulation model is then subjected to the same-source partition process simulation to obtain the process simulation parameters and simulation defect parameters. Based on the mapping relationship between process simulation parameters, simulation defect parameters and measured mechanical property parameters of the same partition, a process-mechanical property mapping model specific to the partition is constructed, and constitutive parameters of heterogeneous materials are assigned to the corresponding partition of the finite element model based on the mapping model, and a finite element calculation model of heterogeneous materials of the partition is established. Full-parameter linkage modeling is performed on the die casting process parameters and the stiffener structural parameters, and adaptive dynamic partitioning rules are set for the structure. When the adjustment range of the stiffener structural parameters exceeds the preset threshold, the dynamic adjustment of the partition is automatically triggered, and the constitutive parameters of the heterogeneous materials in each partition are updated synchronously. A cascaded dual prediction model is constructed, comprising a physical information neural network process defect prediction model and a graph attention network mechanical performance prediction model. An agent environment is constructed based on the cascaded dual prediction model, and a reinforcement learning algorithm is used to train a collaborative optimization agent. The collaborative optimization agent uses the current process parameters, structural parameters, defect parameters, and mechanical performance indicators as its state space, takes the adjustment amount of process parameters and structural parameters as its action, constructs a reward function with mechanical performance, process manufacturability, and lightweight objectives, and outputs the optimal combination of process parameters and stiffener structural parameters through interactive iterative optimization with the environment.

[0027] The present invention discloses a method for co-optimizing the reinforcing ribs of an integral die-cast floor, which can achieve global co-optimization of the process parameters and structural parameters of the integral die-cast floor, significantly improving the impact mechanical performance and lightweight level, while ensuring the manufacturability of the die-casting process.

[0028] This invention provides a coupled, partitioned, measured integrated die-cast floor reinforcement rib collaborative optimization system, comprising six core modules connected in sequence via communication, such as... Figure 2 As shown, it includes a physical homogeneous partitioning measurement module, a simulated homogeneous partitioning process simulation module, a partitioned heterogeneous material modeling module, a parameterized fully coupled modeling module, a PINN-GAT cascaded dual prediction model module, and a PPO reinforcement learning collaborative optimization agent module.

[0029] Specifically, The physical homogeneous partitioning measurement module is used to obtain the measured mechanical properties and defect parameters corresponding to each partition. Its setup steps are as follows: Step S101: Feature recognition is performed on the three-dimensional model of the die-cast integrated floor after die casting, and three major features are extracted: melt filling flow direction, collision force transmission path, and thermal defect distribution. Step S102: Automatically partition the area according to the quantitative gradient threshold: filling speed gradient ≥ 0.2 m / s, stress gradient ≥ 50 MPa, porosity gradient ≥ 0.3%. Divide the boundaries, merge overlapping areas, and form 6 functional zones: gate feeding zone, longitudinal beam collision reinforcement zone, installation point bearing zone, large plane thin wall zone, tail energy absorption zone, and far gate end zone. Step S103: Output a partition vector file with precise coordinate boundaries, which will serve as the sole reference for subsequent simulation partitioning and finite element partitioning. Step S104: Independently cut no less than 5 parallel standard tensile specimens in each partition. The specimens shall comply with the GB / T228.1-2021 standard. The parallel specimens shall be taken from the same position in the same partition, avoiding macroscopic defect areas such as gate residue, cold shut, and cracks. Step S105: Perform room temperature quasi-static tensile tests on the samples of each zone to obtain the measured mechanical property parameters of yield strength, tensile strength, elongation and elastic modulus of each zone. Step S106: Obtain a defect detection sample. The defect detection sample and the tensile sample are taken from the same location, and the dimensions are 10mm × 10mm × casting wall thickness. Step S107: Perform industrial CT scanning and metallographic testing on the defect detection sample, collect the average porosity and shrinkage volume fraction of each sample, and take the arithmetic mean of 5 parallel samples as the measured defect parameters of the zone. Step S108: Output the partition number + partition boundary + partition measured mechanical parameters + partition measured defect parameter dataset.

[0030] The simulated homogeneous partitioning process simulation module is used to obtain process simulation and defect parameters corresponding to the physical partitions. Its setup steps are as follows: Step S201: Use the industry-standard die casting simulation software Flow-3D to establish a three-dimensional simulation model of the integrated die casting post-floor die casting process. The casting material is AlSi9Mn aluminum alloy. The casting body adopts a hexahedral structured mesh with a mesh size of 2mm. The gating and overflow channels adopt a tetrahedral mesh with a mesh size of 1mm to ensure the simulation accuracy of the filling and solidification process. The pouring temperature, mold temperature, injection curve, and pressurization parameters adopt the benchmark process parameters produced by the enterprise. Step S202: Import the partition boundary vector map output by the physical homogeneous partitioning measurement module into the simulation model, and use Boolean operations to partition the simulation model to ensure that the boundary coordinates, number of partitions, and partitioning functions of the simulation partitions are completely consistent with those of the physical partitions, without any deviation, and achieve a one-to-one correspondence between points. Step S203: Set the initial process parameters for die casting simulation and execute coupled simulation of melt filling and solidification process; Step S204: According to the corresponding partition, extract the simulation output parameters for each partition, including the cooling rate, solid fraction, and filling pressure of each partition. The data extraction process is as follows: take the calculation results of no less than 1,000 grid nodes in each partition, and take the arithmetic mean as the simulation output parameter of that partition to ensure that it is consistent with the statistical caliber of the measured data. Step S205: According to the corresponding partition, extract the simulation defect parameters for each partition: porosity, shrinkage volume fraction, and hot spot location for each partition. The data extraction process is as follows: take the calculation results of no less than 1,000 grid nodes in each partition, and take the arithmetic mean as the simulation output parameter of that partition to ensure that it is consistent with the statistical caliber of the measured data. Step S206: Generate a dataset of process simulation parameters and defect parameters for the same partition, and achieve a one-to-one correspondence between the partitioned data and the measured data in step S108.

[0031] The partitioned heterogeneous material modeling module is used to construct a partitioned corresponding heterogeneous material finite element calculation model. The construction steps are as follows: Step S301: Using the process simulation parameters and simulation defect parameters of the same partition as model inputs, and the measured mechanical performance parameters corresponding to the same partition as model outputs, construct a partition-specific process-mechanical performance mapping model in the form of a three-layer BP neural network. Specifically, Network architecture: The input layer has 4 neurons, with the corresponding input variables being the zone cooling rate, zone porosity, zone shrinkage volume fraction, and zone solid fraction; the hidden layer has 1 layer with 12 neurons, and the activation function is ReLU; the output layer has 4 neurons, with the corresponding output variables being the zone yield strength, zone tensile strength, zone elongation, and zone elastic modulus, and the activation function of the output layer is a linear function. Training parameters: The optimizer used is Adam, the learning rate is 0.001, the batch size is 16, the number of iterations is 1000 epochs, the loss function is mean squared error (MSE), and the ratio of training set to test set is 8:2. Accuracy verification criteria: A test set prediction determination coefficient R² ≥ 0.95 is considered acceptable. If the accuracy does not meet the standard, at least 200 additional partitioned simulation-test samples should be added for retraining until the accuracy requirements are met.

[0032] Step S302: Establish a finite element model of the integrated die-cast floor collision. The mesh uses quadrilateral shell elements with a mesh size of 5mm, conforming to the LS-DYNA collision simulation standard. Step S303: Import the boundary vector map of the physical partition into the finite element model, and use component grouping to partition the finite element model to ensure complete consistency with the physical partition and simulation partition; Step S304: For each partition, input the process simulation parameters and defect parameters of that partition into the trained process-mechanical property mapping model to obtain the material constitutive parameters of that partition. Assign a corresponding differentiated material constitutive model to each partition to replace the traditional homogeneous material model. Step S305: Perform a frontal 100% rigid wall collision simulation using LS-DYNA. The simulation results are considered acceptable if the deviation from the actual collision test results is ≤10%. If the deviation exceeds the tolerance, the mapping model and partition material assignment rules need to be re-optimized.

[0033] The parametric fully coupled modeling module is used to achieve the linkage parameterization of die-casting process parameters and stiffener structural parameters, while also addressing the logical defect of zonal failure after structural changes. Its construction steps are as follows: Step S401: Screen eight key process parameters that affect die casting defects and mechanical properties: pouring temperature, mold temperature, low-speed injection speed, high-speed injection speed, pressurization start point, pressurization pressure, holding time, and cooling water temperature. The specific parameters are as follows: pouring temperature 660℃~720℃, mold preheating temperature 180℃~240℃, low-speed injection speed 0.2m / s~0.5m / s, high-speed injection speed 3m / s~6m / s, filling volume at the pressurization start point 60%~80%, pressurization pressure 80MPa~120MPa, holding time 15s~30s, and cooling water temperature 20℃~40℃. Step S402: Screen six key structural parameters that affect collision performance and formability: rib height, rib width, rib spacing, draft angle, transition fillet, and rib cross-section height-to-width ratio. Among them, the height of the main reinforcing rib is 15mm~30mm, the width of the rib is 2mm~5mm, the spacing between the ribs is 80mm~150mm, the draft angle is 1.5°~3°, the transition fillet is 2mm~5mm, and the height-to-width ratio of the rib section is 1.5~3. Step S403: Establish structural adaptive dynamic zoning rules: When the adjustment range of the stiffener structural parameters is ≥5%, the dynamic adjustment of the zoning is automatically triggered; The adaptive dynamic partitioning mechanism includes: re-executing the die-casting filling flow field simulation and collision force transmission path simulation based on the adjusted stiffener structure; recalculating the partition boundaries and updating the number and range of partitions according to the quantization rules of physical homogeneous partitioning; synchronously updating the process-mechanical property mapping model and assigning corresponding heterogeneous material constitutive parameters to the new partitions; and synchronously updating the graph structure nodes and edge features of the GAT mechanical property prediction model to ensure that the model matches the new structure. The PINN-GAT cascaded dual prediction model module is used to quickly predict process defect parameters and collision performance. Specifically, it includes a PINN physical information neural network process defect prediction model and a GAT graph attention network mechanical performance prediction model. The steps for building the PINN physical information neural network process defect prediction model are as follows: Step S501, define the model input as 8 normalized to The key process parameters for die casting in the specified range are defined in vector form as follows: These correspond to the pouring temperature, mold preheating temperature, low-speed injection speed, high-speed injection speed, boosting pressure, boosting start point, holding time, and cooling water temperature, respectively. Step S502, define the spatiotemporal coordinates as the independent variable of the PDE: The spatial coordinates of the casting cavity, The time coordinates are for filling and solidification.

[0034] The spatial domain perfectly matches the cavity boundaries of the six zones, covering the entire three-dimensional cavity of the integral die-cast floor; the total filling and solidification time domain The filling stage solidification stage .

[0035] Step S503: Define the network dual-branch output, including the main output branch and the physical constraint branch; The main output branch includes the average porosity and shrinkage volume fraction of 6 partitions; the physical constraint branch includes the melt velocity at any point in the spatiotemporal domain. ,pressure ,temperature solidity ; Step S504: Within the spatial and temporal domains of the casting cavity, generate N residual point sets using Latin hypercube sampling. The sampling rule is that the sampling density during the filling stage is 5 times that during the solidification stage to capture the transient changes during high-speed filling. The spatial sampling density in the gate feed area and the far gate end area is doubled to match the large gradient characteristics of the flow field / temperature field. The total number of sampling points N=20000 to ensure the coverage and accuracy of the residual calculation.

[0036] Step S505, Solving for partial derivatives of field variables: For the field variables output by the physical constraint branch, solve for the partial derivatives of each order required to save the conservation equations using the automatic differential operator, including the mass conservation equation, momentum conservation equation, energy conservation equation, and solidification phase transition control equation. Step S506: Substitute the obtained partial derivatives into the corresponding conservation equations and calculate the residuals for each sampling point. Residual = left side of the equation - right side of the equation. The smaller the absolute value of the residual, the more it conforms to physical laws, including the residuals of the mass conservation equation. residuals of the momentum conservation equation Energy conservation equation residual Solidification phase transition equation residual ; Step S507: Take the mean square error of the single-equation residuals for all sampling points to obtain the total residual loss of the four equations:

[0037] Step S508: The data fitting loss and physical constraint loss are weighted and fused to construct the total loss function of the neural network. During training, the total loss is minimized through backpropagation to achieve the physical conservation equation's deep constraint on the network. The complete form is:

[0038] in, The mean square error between the partition defect parameters predicted by the main output branch and the label data obtained from Flow-3D simulation is given by the following formula: Where M is the size of the training set. These are the defect parameters predicted by the model. For simulated label data, The total residual loss of the conservation equation, To balance the weights of data loss and physical loss, they are fixed at 0.1.

[0039] Step S509: Latin hypercube sampling generates 2000 sets of process parameter samples. Flow-3D simulation is used to obtain the corresponding sample partition defect labels, which are then divided into training and test sets in an 8:2 ratio. Step S510 involves training the PINN physical information neural network process defect prediction model, which includes the following steps: Initialize the PINN network weights, set the Adam optimizer, learning rate 1e-3, batch size 64, and total number of iterations 2000 epochs; Input process parameters, and the network outputs the defect prediction values ​​of the main branch and the field variables of the physical constraint branch; Calculate total loss Based on the total loss, the gradient of the network weights is solved by automatic differentiation, and the weight parameters of the entire network are updated. Test set prediction coefficient of determination and Stop training and output a qualified PINN model.

[0040] The steps for building the GAT graph attention network mechanical performance prediction model are as follows: Step S511: Transform the stiffener topology into an undirected graph. The specific definition is as follows: Vertex set Each stiffener centerline element is an independent node, with a total of nodes... The number of reinforcing ribs should be consistent with the number of reinforcing ribs. Edge set The necessary and sufficient condition for the existence of an edge between two nodes is that they are physically connected by reinforcing ribs and belong to the same collision force transmission path.

[0041] Node features Each node has 9 feature dimensions, integrating three major categories of information: structure, defects, and materials.

[0042] in: These are structural parameters, namely the height, width, spacing, draft angle, and fillet radius of the reinforcing ribs. The partition defect parameters output by PINN are the porosity and shrinkage volume fraction of the partition corresponding to the node. The parameters are the partitioned heterogeneous material parameters, which are the yield strength and elastic modulus of the partition corresponding to the node, respectively.

[0043] Step S512: Define the network architecture, which includes a 2-layer GAT graph attention network with 4 attention heads per layer, a hidden layer dimension of 128, and uses LeakyReLU (negative slope 0.2) as the activation function. The output layer has 4 neurons, which output 4 mechanical performance indicators: [collision energy absorption, maximum intrusion in key areas, maximum stress, and first-order torsional mode frequency]. Step S513: Using the Latin hypercube sampling method, 1500 sets of samples are generated within the range of structural parameter values. Based on the partitioned heterogeneous finite element model, collision simulation and modal simulation are completed using LS-DYNA to obtain the mechanical performance indicators of the corresponding samples. The dataset is constructed and divided into training set and test set in an 8:2 ratio. Step S514 involves training the GAT graph attention network mechanical performance prediction model, which includes the following steps: The optimizer used was Adam, with a learning rate of 5e-4, a batch size of 32, and 1500 epochs. Dropout (with a dropout rate of 0.3) was used to prevent overfitting. A test set prediction determination coefficient R² ≥ 0.93 is considered acceptable, and the relative error between the predicted and simulated values ​​is ≤ 10%, ensuring that it can replace simulation software for mechanical property prediction.

[0044] The PPO reinforcement learning collaborative optimization agent module is used to achieve collaborative optimization of process parameters and structural parameters. Its construction steps are as follows: Step S601: Define the state space, which is a 30-dimensional continuous vector. ,in: There are 8 die-casting process parameters. There are 6 stiffener structural parameters. There are 6 partition defect parameters. These are four mechanical performance indicators.

[0045] Step S602: Define the action space, which is a 14-dimensional continuous vector. ,in, The adjustment range for the eight die-casting process parameters is strictly limited to the upper and lower limits set by the patent. The adjustment range for the six reinforcing rib structural parameters is strictly limited to the upper and lower limits of the parameters set in the patent. The adjustment range for the corresponding 8 process parameters and 6 reinforcing rib structural parameters is strictly limited to the upper and lower limits set by the parametric modeling module. Step S603: Define a state transition function. The state transition is implemented by the PINN-GAT cascaded agent environment. The transition process is as follows: the agent outputs an action. The updated process parameters are obtained. With structural parameters When the structural parameters are adjusted by ≥5%, the adaptive dynamic partitioning mechanism is triggered, and the GAT model diagram structure is updated. New process parameters are input into the PINN model to obtain new partitioning defect parameters. The new structural parameters and defect parameters are input into the GAT model to obtain new mechanical performance indices. The combination yields the state at the next moment. Complete the state transition; Step S604: Build the PINN-GAT cascaded agent environment, including environment input interface, dynamic partitioning adaptation unit, PINN defect prediction unit, GAT mechanical performance prediction unit, and environment output interface; The environment input interface receives the 14-dimensional action vector output by the intelligent agent and parses it into updated die-casting process parameters and reinforcing rib structure parameters; The dynamic partitioning adaptation unit detects the magnitude of changes in structural parameters, triggers dynamic partitioning adjustments, and synchronously updates the graph structure nodes and edge features of the GAT model. The PINN defect prediction unit takes the updated process parameters as input and outputs the porosity and shrinkage volume fraction of each zone. The single-step inference time is ≤2ms. The GAT mechanical performance prediction unit takes the updated structural parameters and the defect parameters output by PINN as input, and outputs 4 mechanical performance indicators. The single-step inference time is ≤5ms. The environment output interface outputs the next moment's status, immediate reward, and round termination signal, completing one environment interaction.

[0046] Step S605: Build the PPO core network architecture, including a shared underlying feature extraction layer, an Actor policy network branch, and a Critic value network branch. The shared bottom-level feature extraction layer maps four types of heterogeneous data—process, structure, defect, and mechanics—into high-dimensional robust features. Specifically, the architecture is to input a 30-dimensional normalized state vector and output a 256-dimensional high-dimensional feature vector. The Actor policy network branch is based on the 256-dimensional output of the shared feature layer, constructing two fully connected branches to output the Gaussian distribution mean and variance of continuous actions, respectively. Based on the output mean... With variance Construct a Gaussian distribution 14-dimensional continuous motion was obtained from sampling. Simultaneously calculate the log probability of this action. ; The Critic value network branch is based on the 256-dimensional output of the shared feature layer, and constructs a 2-layer fully connected branch to output the state value scalar of the current state. ; Step S606: Construct a multi-objective reward function with a hierarchical structure, simultaneously balancing mechanical performance, manufacturability, and lightweight objectives, as shown in the following equation:

[0047] in, As a mechanical performance reward, it guides the agent to improve collision safety performance and is calculated according to the following formula:

[0048] in, To absorb energy during collisions This represents the maximum intrusion amount in the critical area.

[0049] As a manufacturability incentive, it guides the agent to reduce die-casting defects and is calculated using the following formula:

[0050] in The average porosity, This represents the average volume fraction of shrinkage.

[0051] As a lightweight reward item, it guides the agent to reduce the mass of the floor and is calculated according to the following formula:

[0052] in For the quality of the rear floor structure, The baseline quality is based on the original scheme.

[0053] To constrain the penalty term, it is calculated using the following formula:

[0054] Step S607: Train the PPO reinforcement learning collaborative optimization agent, and then solidify the model after training. Specifically, the training process includes: Initialize the weights of the policy network and value network, initialize the PINN-GAT cascaded agent environment, and set all hyperparameters; Sixteen parallel environments interact simultaneously. The policy network outputs actions based on the current state, the agent environment executes state transitions, and collects trajectory data (state, action, reward, next state, termination signal). Based on the collected trajectory data, the temporal difference error and the GAE generalized dominance value are calculated; The collected samples are divided into multiple batches and substituted into the total loss function, as shown in the following formula. The weights of the policy network and the value network are updated through backpropagation using the Adam optimizer. The training is repeated for 10 epochs for each batch of data.

[0055] Among them, value loss , Weighting of cumulative returns for discounts Entropy regularization term Used to encourage exploration, weight Strategy loss ,in The generalized advantage estimate is the pruning factor. Among them, the timing difference error The probability ratio of the new and old strategies ,in For the new strategy, This is the old strategy.

[0056] Every 50 iterations, the model weights are saved once to verify the constraint satisfaction rate and average reward of the model under 100 sets of random initial states; If the average reward fluctuation is ≤5% for 100 consecutive iterations and the constraint satisfaction rate reaches 100%, stop training and output the converged optimal policy model.

[0057] Example 2 This embodiment is only used to explain the present invention and is not intended to limit the scope of protection of the present invention. This embodiment takes a one-piece die-cast rear floor of a new energy vehicle as the optimization object. The rear floor is made of AlSi9Mn aluminum alloy and is integrally die-cast. The overall size is 1800mm×1500mm. It needs to meet the safety performance requirements of 100% rigid wall frontal collision and side collision, while also meeting the NVH modal and lightweight design goals.

[0058] Step S10: Same-source partitioning test and data acquisition: Based on the melt filling path, collision force transmission path, and heat distribution of the target post-floor, the physical integrated die-cast post-floor is divided into 6 quantitative feature zones. The partition boundaries are automatically defined based on a filling speed gradient threshold of ≥0.2 m / s, ensuring uniform distribution of process parameters and mechanical properties within each partition and significant gradient differences between partitions. Five GB / T228.1-2021 standard tensile specimens are taken from each partition, with parallel specimens taken from the same location within the same partition to ensure data consistency. Room temperature quasi-static tensile tests are conducted using a WDW-100 universal testing machine to obtain the average yield strength, tensile strength, elongation, and elastic modulus for each partition. Simultaneously, industrial CT scanning and metallographic microscopy analysis are used to obtain the average porosity and shrinkage volume fraction for each partition, serving as measured defect parameters.

[0059] Step S20 Simulation of homogeneous partitioning process: Using the same 6-zone partitioning method as step S10, a simulation model of the die-casting process of the target post-floor was established based on Flow-3D software. The material was set as AlSi9Mn aluminum alloy, and numerical simulation of the entire melt filling and solidification process was completed. The initial simulation parameters adopted the company's existing production process parameters: pouring temperature 690℃, mold preheating temperature 220℃, low-speed injection speed 0.3m / s, high-speed injection speed 4.5m / s, pressurization pressure 100MPa, and holding time 20s.

[0060] After the simulation is completed, the process simulation parameters and process defect parameters of each partition are output. The process simulation parameters include the average cooling rate, solid fraction, and filling pressure of each partition; the process defect parameters include the average porosity, shrinkage volume fraction, and hot spot location of each partition, so as to achieve a one-to-one correspondence between the simulation data and the measured data of each partition.

[0061] Step S30: Constructing a model of the heterogeneous material in the partitioned region: Using the cooling rate, porosity, and shrinkage volume fraction of each zone obtained in step S20 as input variables, and the yield strength, tensile strength, elongation, and elastic modulus of each zone obtained by actual measurement in step S10 as output variables, a three-layer BP neural network is used to construct a zone-specific process-mechanical property mapping model.

[0062] A finite element model of the target floor was established using HyperMesh software, with a mesh size of 5 mm and a total of approximately 1.2 million elements. The finite element model was partitioned according to the 6-partition scheme in step S10. Based on the aforementioned process-mechanical property mapping model, corresponding heterogeneous material constitutive parameters were assigned to each partition, and a partition-corresponding material heterogeneous finite element calculation model was established.

[0063] The LS-DYNA software was used to simulate the frontal collision of a 100% rigid wall, and the simulation results of the heterogeneous model and the traditional homogeneous model were compared.

[0064] Step S40: Fully Parametric Linkage Modeling The key process parameters of integrated die casting and the core structural parameters of the rear floor reinforcement are modeled in a fully parametric and linked manner. The value range of all parameters conforms to the die casting process specifications and vehicle body structure design standards. At the same time, a zoned dynamic adaptive adjustment rule is set: when the reinforcement structural parameters are adjusted, the filling flow field and collision force transmission path are recalculated. Using the filling speed gradient, temperature gradient, and stress gradient as thresholds, the zone boundaries and quantities are automatically updated, and the constitutive parameters of the heterogeneous materials in each zone are updated synchronously.

[0065] Step S50: Construction of the PINN-GAT cascaded dual prediction model: Using the eight process parameters from step S40 as input and the average porosity and shrinkage volume fraction of the six zones as output, a PINN physical information neural network is constructed. The mass conservation equation, momentum conservation equation, energy conservation equation, and solidification heat transfer equation of the die casting process are embedded as physical constraint terms into the loss function of the neural network.

[0066] The stiffener structure of the target rear floor is transformed into a graph structure, where nodes are the centerline elements of the stiffeners, and node features include stiffener height, width, spacing, and corresponding zone defect parameters. Edges represent the connection relationships of the stiffeners and the collision force transmission paths. Using the six stiffener structural parameters from step S40 and the porosity and shrinkage volume fraction of the six zones output by the PINN model as inputs, and the vehicle's energy absorption under frontal collision conditions, the maximum intrusion in the rear floor longitudinal beam connection area, the maximum stress at the seat mounting point, and the first-order torsional modal frequency as outputs, a GAT graph attention network model is constructed.

[0067] Step S60: PPO Collaborative Optimization Agent Construction and Global Optimization: A collaborative optimization agent is constructed using the PPO (Proximity Policy Optimization) algorithm. The trained PINN process defect prediction model and GAT (Gas-Attractive Abrasive) mechanical property prediction model are used as the environment model. The training model converges to the global optimum, outputting the collaborative optimization agent. This agent can generate optimal combinations of process parameters and reinforcing rib structural parameters based on the initial die-casting model, outputting the optimal combinations within seconds.

[0068] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 3 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0069] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0070] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0071] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for synergistic optimization of floor reinforcement ribs after integral die casting, characterized in that, include: Based on the actual measurement of the integral die-cast floor in the physical object, the measured mechanical property parameters and measured defect parameters are obtained. The simulation model is then subjected to the same-source partition process simulation to obtain the process simulation parameters and simulation defect parameters. Based on the mapping relationship between process simulation parameters, simulation defect parameters and measured mechanical property parameters of the same partition, a process-mechanical property mapping model specific to the partition is constructed, and constitutive parameters of heterogeneous materials are assigned to the corresponding partition of the finite element model based on the mapping model, and a finite element calculation model of heterogeneous materials of the partition is established. Full-parameter linkage modeling is performed on the die casting process parameters and the stiffener structural parameters, and adaptive dynamic partitioning rules are set for the structure. When the adjustment range of the stiffener structural parameters exceeds the preset threshold, the dynamic adjustment of the partition is automatically triggered, and the constitutive parameters of the heterogeneous materials in each partition are updated synchronously. A cascaded dual prediction model is constructed, comprising a physical information neural network process defect prediction model and a graph attention network mechanical performance prediction model. An agent environment is constructed based on the cascaded dual prediction model, and a reinforcement learning algorithm is used to train a collaborative optimization agent. The collaborative optimization agent uses the current process parameters, structural parameters, defect parameters, and mechanical performance indicators as its state space, takes the adjustment amount of process parameters and structural parameters as its action, constructs a reward function with mechanical performance, process manufacturability, and lightweight objectives, and outputs the optimal combination of process parameters and stiffener structural parameters through interactive iterative optimization with the environment.

2. The method as described in claim 1, characterized in that, The partition measurements include: Feature recognition is performed on the solid die-cast floor, extracting the melt filling flow direction, collision force transmission path and hot spot defect distribution characteristics. The floor is automatically partitioned according to the preset quantization gradient threshold to form multiple functional partitions, and the partition vector file with precise coordinate boundaries is output. Multiple parallel standard tensile specimens were taken from each partition for tensile testing to obtain the mechanical property parameters of each partition. Defect detection specimens taken from the same location were also tested to obtain the defect parameters of each partition.

3. The method as described in claim 2, characterized in that, The process simulation of homogeneous partitioning of the simulation model includes: A three-dimensional simulation model of the die casting process of the integrated die-casting post-die casting floor was established using die casting simulation software. The partition vector file was imported into the simulation model, and the simulation model was partitioned to ensure that the boundary coordinates, number of partitions, and partition functions of the simulation partitions and the physical partitions were consistent. Perform coupled simulation of melt filling and solidification process, and extract process simulation parameters and simulation defect parameters for each zone according to the corresponding zone.

4. The method as described in claim 1, characterized in that, The process-mechanical property mapping model specific to the construction partition includes: Using process simulation parameters and defect parameters of the same partition as input, and measured mechanical performance parameters of the same partition as output, a mapping model in the form of a neural network is constructed.

5. The method as described in claim 1, characterized in that, The establishment of the finite element calculation model for partitioned heterogeneous materials includes: A finite element model of the integrated die-cast floor collision was established. The partition vector file was imported into the finite element model and partitioned. For each partition, the process simulation parameters and defect parameters of that partition were input into the trained mapping model to obtain the material constitutive parameters of that partition. Each partition was assigned a corresponding differentiated material constitutive model, and collision simulation verification was performed.

6. The method as described in claim 1, characterized in that, The process of performing fully parametric linkage modeling and setting adaptive dynamic partitioning rules for the structure includes: Screening key process parameters that affect die casting defects and mechanical properties, as well as key structural parameters that affect impact performance and formability; When the adjustment range of the stiffener structure parameters exceeds the preset threshold, the dynamic adjustment of the partition is automatically triggered. Based on the adjusted stiffener structure, the partition boundary is recalculated, the process-mechanical property mapping model is updated synchronously, and the corresponding heterogeneous material constitutive parameters are assigned. The graph structure of the graph attention network mechanical property prediction model is also updated synchronously.

7. The method as described in claim 1, characterized in that, The construction of the cascaded dual prediction model includes: The physical information neural network process defect prediction model takes die casting process parameters as input, outputs defect parameters of each zone, and uses the physical conservation equation of the die casting process as the physical constraint embedding loss function. It is trained by weighted fusion of data fitting loss and physical constraint loss. The graph attention network mechanical performance prediction model transforms the stiffener topology into a graph structure, integrates node features with structural parameters, defect parameters, and heterogeneous material parameters, and outputs mechanical performance indicators.

8. A synergistic optimization system for integrated die-cast floor reinforcement ribs, characterized in that, include: The physical homogeneous partitioning measurement module is used to perform partitioning measurements on the physical integral die-cast floor to obtain measured mechanical property parameters and measured defect parameters, and to perform homogeneous partitioning process simulation on the simulation model to obtain process simulation parameters and simulation defect parameters. The simulation module for the same origin partition process is used to construct a partition-specific process-mechanical property mapping model based on the mapping relationship between the process simulation parameters, simulation defect parameters and measured mechanical property parameters of the same partition, and to assign heterogeneous material constitutive parameters to the corresponding partition of the finite element model based on the mapping model, thereby establishing a partition heterogeneous material finite element calculation model. The partitioned heterogeneous material modeling module is used to perform fully parametric linkage modeling of die casting process parameters and stiffener structural parameters, and to set adaptive dynamic partitioning rules for the structure. When the adjustment range of stiffener structural parameters exceeds the preset threshold, the partition dynamic adjustment is automatically triggered, and the constitutive parameters of heterogeneous materials in each partition are updated synchronously. A parameterized fully coupled modeling module is used to construct a cascaded dual prediction model, which includes a physical information neural network process defect prediction model and a graph attention network mechanical performance prediction model. A cascaded dual prediction model module is used to construct an agent environment based on the cascaded dual prediction model and to train a collaborative optimization agent using a reinforcement learning algorithm. The collaborative optimization agent module is used to make the collaborative optimization agent use the current process parameters, structural parameters, defect parameters and mechanical performance indicators as the state space, the adjustment amount of process parameters and structural parameters as actions, and construct a reward function with mechanical performance, process manufacturability and lightweight objectives. Through interactive and iterative optimization with the environment, it outputs the optimal combination of process parameters and stiffener structural parameters.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the integrated die-cast floor reinforcement collaborative optimization method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for co-optimization of integrated die-cast floor reinforcement as described in any one of claims 1-7.

Images