Hot stamping part conceptual design feasibility evaluation method based on deep learning

Abstract

The invention discloses a hot stamping part conceptual design feasibility evaluation method based on deep learning. The hot stamping part conceptual design feasibility evaluation method comprises the steps that a forming-service material model inheriting the forming history is constructed; constructing a deep learning agent model based on image representation; and evaluating the design feasibility based on the forming-service simulation model. The problem that in the prior art, in the hot stamping part development process, complex part design is difficult to deal with only relying on the experience of a designer, and the influence of the hot forming history on the part service performance is neglected, so that the structural design is unreasonable is solved, and by representing the degradation transfer and accumulation rule of the material performance from the forming stage to the service stage, the material performance is improved. The complex corresponding relation between part design parameters and forming response is defined, seamless connection and real-time linkage of the part forming process and the service process are achieved, the feasibility of all schemes can be rapidly and accurately evaluated in the conceptual design stage, guidance is provided for designers to select reasonable part design schemes, and design according to performance is achieved in the early development stage.

Description

Technical Field

[0001] This invention relates to the field of hot stamping parts design, and more specifically to a deep learning-based method for evaluating the feasibility of conceptual design of hot stamping parts. Background Technology

[0002] Lightweighting of automobiles is a crucial approach to achieving energy conservation and emission reduction in the transportation sector. Hot stamping, with its unique advantages, plays a vital role in this field. Its principle involves heating boron steel sheets with a microstructure of ferrite + pearlite to a temperature above Ac3, resulting in a uniform austenitic structure with low deformation resistance and high elongation. The sheet is then rapidly transferred to a stamping die equipped with a cooling system for rapid forming, followed by pressure holding and quenching to completely transform it into a full martensitic structure, significantly improving the strength of the part. Hot stamping can reduce the weight of the entire vehicle while ensuring passenger safety, and is widely used in the development of body parts for new energy vehicles. In the development process, the first step is to conduct a conceptual design of the part structure based on the required objectives, followed by a detailed design. Finally, the feasibility of the detailed design scheme is evaluated using numerical simulations of part forming and service operation, ultimately determining whether to initiate part manufacturing.

[0003] However, see Figure 1 Hot stamping is often used to manufacture complex-shaped parts with multiple interdependent geometric features. Relying solely on the product designer's experience can easily lead to unreasonable design solutions during the conceptual design phase. This results in repeated adjustments to multiple sets of geometric features and forming process parameters during detailed design, sometimes even requiring a complete overhaul of the conceptual design. Furthermore, multiple rounds of time-consuming, labor-intensive, and costly forming and service simulations are needed to find a feasible solution, severely impacting the quality and efficiency of part development. Although the rapid development of artificial intelligence has led to the use of scalar representation-based machine learning proxy models to establish the mapping relationship between part design parameters and forming responses, enabling rapid evaluation of the feasibility of design solutions, these models are only suitable for scenarios with limited and parameterizable inputs. Moreover, due to the dimensionality reduction process during training, these models often fail to extract bijective relationships and positional information between input and output, resulting in poor accuracy and generalization. Additionally, when using existing simulation software to design part structures, designers cannot determine the impact of hot forming history on part service performance, leading to inaccurate predictions of service performance. Some design solutions may not be feasible in actual service, significantly reducing the credibility of service simulations.

[0004] Therefore, a method is needed to characterize the degradation and accumulation of material properties from the forming stage to the service stage, clarify the complex correspondence between part design parameters and forming response, and achieve seamless connection and real-time linkage between the part forming process and the service process. This method (see...) Figure 2This allows for rapid and accurate evaluation of the feasibility of various schemes during the conceptual design phase, providing guidance for designers to select appropriate part design schemes and enabling "performance-based design" in the early stages of development. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the problem that in the development of hot stamped parts, the designer’s own experience alone is not enough to deal with the complex part design and the influence of hot forming history on the service performance of the part is ignored, resulting in unreasonable structural design. The present invention provides a feasibility evaluation method for the conceptual design of hot stamped parts based on deep learning.

[0006] To address the aforementioned technical problems, this invention provides a deep learning-based method for evaluating the feasibility of conceptual design of hot-stamped parts, comprising the following steps:

[0007] Construct a forming-service material model that inherits the forming history;

[0008] Construct a deep learning agent model based on image representation;

[0009] The feasibility of the design was evaluated based on the forming-service simulation model.

[0010] Preferably, the construction of the forming-service material model that inherits the forming history includes:

[0011] Constructing the forming part of the forming-service material model;

[0012] Construct the service portion of the forming-service material model.

[0013] Furthermore, the forming portion of the constructed forming-service material model includes:

[0014] Uniaxial tensile tests were conducted on specimens made of boron steel sheet under different deformation temperatures and strain rates to obtain the stress-strain curves of the boron steel sheet at high temperatures.

[0015] A model architecture was selected that can accurately characterize the high-temperature rheological behavior of the boron steel sheet and effectively transmit the hot forming history.

[0016] Select an appropriate objective function and optimization algorithm, and determine the model coefficients of the forming part of the forming-service material model based on the obtained stress-strain curve at high temperature.

[0017] The forming part of the forming-service material model is converted into a corresponding material subroutine using a high-level programming language and then embedded into numerical simulation software for verification.

[0018] The service portion of the constructed-service material model includes:

[0019] Uniaxial tensile tests were conducted on specimens made of boron steel sheet under different deformation temperatures, strain rates, and deformation amounts. All tests were completed without the specimens breaking. Tests under the same temperature, strain rate, and deformation amount conditions were repeated twice.

[0020] All specimens that have completed the uniaxial tensile test without breaking are quenched to room temperature and then divided into two groups; specimens that have undergone the same test conditions shall not be included in the same group.

[0021] Two sets of specimens were subjected to uniaxial tensile tests at room temperature and different strain rates to obtain the stress-strain curves of the two sets of specimens at room temperature.

[0022] A model architecture is constructed that can both reflect the influence of hot forming history on the subsequent mechanical properties of the material and accurately characterize the room temperature deformation behavior of the boron steel sheet.

[0023] Select an appropriate objective function and optimization algorithm, and determine the model coefficients of the service part of the forming-service material model based on the obtained stress-strain curve at room temperature.

[0024] The service portion of the formed-service material model was converted into a corresponding material subroutine using a high-level programming language and then embedded into numerical simulation software for verification.

[0025] Preferably, the construction of the deep learning agent model based on image representation includes:

[0026] Construct a simulation model of the part's thermoforming process;

[0027] Determine the input parameters and perform batch simulations of cloud computing.

[0028] Develop an automatic conversion program for input / output image representation;

[0029] Construct a neural network proxy model based on image representation.

[0030] Furthermore, the aforementioned simulation model for the thermoforming of the component includes:

[0031] Determine the geometry and expected forming conditions of hot stamped parts during the conceptual design phase;

[0032] A program A is developed using a programming language. The program A is able to automatically generate the geometric structure of the forming mold and the geometric structure of the sheet metal blank for hot stamping parts in computer-aided design software.

[0033] Program B is developed using a programming language. Program B can automatically generate the finite element mesh of each component of the thermoforming simulation model in the finite element preprocessing software.

[0034] A program C is developed using a programming language. Program C can automatically load the finite element mesh and the material subroutine of the forming-service material model in the finite element simulation software, and construct a thermoforming simulation model.

[0035] The aforementioned thermoforming simulation model was experimentally verified and its parameters were corrected.

[0036] The process of determining input parameters and performing batch simulations in cloud computing includes:

[0037] The input parameters that affect the forming response of the hot stamping parts include: some geometric parameters of the forming die, some geometric parameters of the sheet blank, and some process parameters, etc.

[0038] Based on engineering experience, the range and dispersion level of each input parameter are determined, and multiple combinations of input parameters are generated using experimental design methods.

[0039] Each set of input parameters is imported into the thermoforming simulation model described above.

[0040] Numerical simulation tasks are deployed using a cloud computing platform to achieve parallel offline solutions for multiple combinations of input parameters;

[0041] The output results of each numerical simulation task include: the displacement field, thickness reduction field, and material degradation field of the hot-stamped part after forming.

[0042] The aforementioned automatic conversion program for input / output image representation includes:

[0043] Define the graphical representation rules for the selected forming die geometry, sheet metal blank geometry, and process parameters as inputs;

[0044] A program D is developed using a programming language. The program D can automatically convert any combination of forming mold geometry, sheet metal blank geometry and process parameters into a corresponding input image combination.

[0045] Establish graphical representation rules for the displacement field, thickness reduction field, and material degradation field of the selected hot-stamped parts as output;

[0046] A program E is developed using a programming language. Program E can automatically convert the displacement field, thickness reduction field, and material degradation field of hot-stamped parts that match the input image combination into the corresponding output image combination.

[0047] The construction of the image representation-based neural network proxy model includes:

[0048] Establish a database with a unified format to record the storage addresses of input and output image combinations on the server;

[0049] Each set of input images and its corresponding output image are stored in pairs, and their address information is entered into the database.

[0050] Construct a deep neural network architecture with strong expressive power as a proxy model;

[0051] The training and test sets are divided. Based on the selected loss function, the optimized surrogate model is trained on the training set through cross-validation, and its prediction accuracy is evaluated on the test set.

[0052] Deploy the agent model that has been tested and is capable of rapid prediction and remote invocation to the cloud.

[0053] Preferably, the evaluation of design feasibility based on the forming-service simulation model includes:

[0054] Construct a part forming-service simulation model;

[0055] Evaluate the feasibility of forming the pre-selected part design scheme;

[0056] Evaluate the service feasibility of the formable solution and seek optimization.

[0057] Furthermore, the forming-service simulation model of the constructed part includes:

[0058] Based on relevant regulations and actual needs, determine the service conditions and performance indicators of hot-stamped parts;

[0059] Based on the service conditions and geometry of the hot-stamped parts and the material subroutine of the forming-service material model, a service simulation model is constructed.

[0060] The service simulation model was tested and its parameters were corrected.

[0061] The thermoforming simulation model, after being verified by experiments and with parameter corrections, is coupled with the service simulation model to form the forming-service simulation model.

[0062] The evaluation of the feasibility of the pre-selected part design scheme includes:

[0063] Construct a database of pre-selected design schemes to store the geometric parameters of forming molds, geometric parameters of sheet metal blanks, and process parameters corresponding to each design scheme to be evaluated.

[0064] The geometric structure of the forming mold and the geometric structure of the sheet blank corresponding to each design scheme to be evaluated are automatically generated in the computer-aided design software using the aforementioned program A.

[0065] The program D is used to automatically convert the geometric structure of the forming mold, the geometric structure of the sheet blank, and the combination of process parameters corresponding to each design scheme to be evaluated into the corresponding combination of input images to be evaluated.

[0066] The various input images to be evaluated are combined and input into the image representation-based neural network proxy model to filter out feasible design schemes.

[0067] The evaluation and optimization of the service feasibility of the formable solution includes:

[0068] The geometric parameters of the forming mold, the geometric parameters of the sheet metal blank, and the process parameters corresponding to the selected formable design schemes are input into the forming-service simulation model to solve and filter out all design schemes that meet the performance index requirements.

[0069] Using the maximum weight of the part in the conceptual design stage as a constraint, a suitable optimization algorithm is used to search for the design scheme with the optimal performance index that satisfies the constraint.

[0070] Compared with the prior art, the beneficial effects of the present invention are:

[0071] 1. The method of this invention achieves highly flexible characterization of material deformation behavior by constructing a forming-service material model that inherits the forming history. This model can not only describe the stress variation with temperature and strain rate during high-temperature deformation, but also inherit the influence of the forming history to describe the material's performance at different strain rates at room temperature, and is applicable to both forming and service states.

[0072] 2. The method of this invention achieves highly efficient evaluation of part forming performance by constructing a deep learning proxy model based on image representation. This model, through image representation, simultaneously considers the coupled influence of multiple geometric and process parameters on the part forming process, and can predict the forming performance of candidate design schemes in real time without design experience or simulation calculations, facilitating the rapid identification of feasible solutions.

[0073] 3. The method of this invention achieves high-precision evaluation of the service performance of parts by constructing a forming-service simulation model. This model, by calling a forming-service material model that inherits the forming history, can achieve seamless connection and real-time linkage between the forming and service simulation processes in a single calculation file, automatically transferring variables between the two processes without human intervention. Attached Figure Description

[0074] The invention will now be further described with reference to the accompanying drawings:

[0075] Figure 1 This is a flowchart illustrating the development process of traditional hot-stamped parts.

[0076] Figure 2A flowchart illustrating the development process of hot-stamped parts using the method of this invention;

[0077] Figure 3 This invention provides a flowchart of a deep learning-based method for evaluating the feasibility of conceptual design of hot-stamped parts.

[0078] Figure 4 This is a schematic diagram of the device used in the high-temperature uniaxial tensile test in the embodiment.

[0079] Figure 5 This is a comparison diagram of the stress-strain curves calculated from the formed part of the forming-service material model in the embodiment and those obtained from the high-temperature uniaxial tensile test.

[0080] Figure 6 This is a comparison chart of the stress-strain curves calculated from the service portion of the formed-service material model in the embodiment and those obtained from the room temperature uniaxial tensile test.

[0081] Figure 7 This is a schematic diagram of the structure of the part's thermoforming simulation model in the embodiment;

[0082] Figure 8 This is a schematic diagram of the geometric parameters of the forming die, which serve as input parameters for the proxy model in this embodiment.

[0083] Figure 9 This is a schematic diagram of the sheet metal blank geometric parameters used as input parameters for the proxy model in the embodiment;

[0084] Figure 10 This is a flowchart of the input image acquisition algorithm used for training and testing the proxy model in the embodiment;

[0085] Figure 11 This is a flowchart of the algorithm for obtaining the output image for training and testing the proxy model in the embodiment;

[0086] Figure 12 This is a schematic diagram of the structure of the neural network proxy model constructed in the embodiment;

[0087] Figure 13 This is a schematic diagram of the service simulation model of the part in the embodiment. Detailed Implementation

[0088] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0089] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0090] Example

[0091] See Figure 3 A deep learning-based method for evaluating the feasibility of conceptual design of hot-stamped parts includes:

[0092] S1: Construct a forming-service material model that inherits the forming history.

[0093] Step S1 includes:

[0094] S11: The forming part of constructing the forming-service material model includes:

[0095] S111: Uniaxial tensile tests were conducted on specimens made of boron steel sheet under different deformation temperatures and strain rates to obtain the stress-strain curves of the boron steel sheet at high temperatures.

[0096] In this embodiment, 22MnB5 boron steel plate was selected as the test object, and dumbbell-shaped tensile specimens conforming to GB / T 228.1-2010 standard were prepared, with a gauge length of 50mm × 15mm × 2mm. (See reference...) Figure 4 High-temperature uniaxial tensile tests were conducted using a Gleeble 3800 thermal simulation testing machine 1, which can precisely control temperature, strain rate and loading path.

[0097] During the experiment, all specimens were heated to 925℃ at a heating rate of 10℃ / s and held for 180s to ensure complete austenitization of the gauge length material. Subsequently, the specimens were cooled to different deformation temperatures (600℃, 700℃, and 800℃) at a cooling rate of 100℃ / s and held for 30s. Then, the specimens were subjected to isothermal tensile testing at different strain rates (0.01 / s, 0.1 / s, 1 / s, and 10 / s) until fracture. The entire experiment included 12 test conditions consisting of 3 deformation temperatures and 4 strain rate combinations. Displacement data was acquired in real-time using a CCD camera 2, and combined with the load sensor output, stress-strain curves at high temperatures were generated. Figure 5The data points in the figure show the stress-strain curves when the deformation temperature is 700℃ and the strain rates are 0.01 / s, 0.1 / s, 1 / s and 10 / s, respectively, as well as the stress-strain curves when the strain rate is 0.1 / s and the deformation temperatures are 600℃, 700℃ and 800℃.

[0098] S112: Select a model architecture that can accurately characterize the high-temperature rheological behavior of the boron steel sheet and effectively transmit the hot forming history.

[0099] In this embodiment, the model architecture of the forming part of the forming-service material model is specifically manifested as follows:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] In the formula: For stress; and and These are the degradation parameters of the forming material and their evolution rate; and These are the normalized dislocation density and its evolution rate, respectively. , and These are plastic strain, plastic strain rate, and total strain, respectively. and These are the model coefficients.

[0106] In addition, the other model parameters that vary with temperature are as follows:

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113]

[0114] In the formula: This is the universal gas constant; Absolute temperature; and These are the model coefficients.

[0115] S113: Select a suitable objective function and optimization algorithm, and determine the model coefficients of the forming part of the forming-service material model based on the obtained stress-strain curve at high temperature.

[0116] In this embodiment, the selected objective function and optimization algorithm are as follows: and genetic algorithms. Among them, It can be represented as:

[0117]

[0118] In the formula: The vector of model coefficient combinations to be determined; and These represent the number and index number of the stress-strain curves involved in the optimization; and They are the first ones participating in the optimization. The number of data points and their index numbers taken from the curve; This objective function calculates the weighted distance between the data points in the form and the experimental data. It can be used to accurately evaluate the difference between the calculated results of the formed portion of the formed-service material model and the results of high-temperature uniaxial tensile tests.

[0119] The source code for the forming part and objective function of the forming-service material model was written in Python. The genetic algorithm in Python's DEAP tool library was used to optimize the objective function and obtain the optimal objective function value. The set of model coefficients corresponding to the optimal objective function value is the set of optimal model coefficients for the forming part of the forming-service material model.

[0120] In this embodiment, The value is 6, representing 6 curves. The values ​​were all set to 20, and the resulting model coefficients were as follows: =0.4、 =5.2、 =1.55、 =3.1、 =17.6、 =12.5、 =30、 =0.007、 =80、 =55000、 =1.4e -4 , =1100、 =8400、 =8400、 =50000、 =8450、 =100000、 =10650、 =17500 and =8.314. Substituting all model coefficient values ​​into the forming part of the forming-service material model for calculation, the resulting stress-strain curve is compared with the stress-strain curve obtained from the high-temperature uniaxial tensile test, as shown below. Figure 5 As shown.

[0121] S114: The forming part of the forming-service material model is converted into a corresponding material subroutine using a high-level programming language and then embedded into numerical simulation software for verification.

[0122] In this embodiment, the forming part of the forming-service material model is converted into a corresponding material subroutine VUMAT_F using the Fortran language, which can be called by finite element simulation software such as LS-DYNA. This material subroutine mainly includes: a material parameter reading module, a state variable initialization module, a stress update module, a constitutive relation calculation module, and a thermodynamic calculation module.

[0123] To verify the correctness of the subroutine, a simulation model under the same conditions as the high-temperature uniaxial tensile test was established, and VUMAT_F was called to perform simulation calculations. Finally, by comparing the stress-strain curves obtained from the simulation and the experiment, it was proved that the simulation model established based on VUMAT_F can simulate the high-temperature rheological behavior of boron steel under uniaxial tension.

[0124] S12: The service portion of constructing the forming-service material model includes:

[0125] S121: Uniaxial tensile tests were conducted on specimens made of boron steel sheet under different deformation temperatures, strain rates and deformation amounts. All tests were completed without the specimens breaking. Tests under the same temperature, strain rate and deformation amount conditions were repeated twice.

[0126] In this embodiment, the tensile specimens, gauge length dimensions, and thermal simulation testing machine used are the same as in step S111, and will not be repeated here. During the test, all specimens were heated to 925°C at a heating rate of 10°C / s and held for 180s to ensure complete austenitization of the gauge length material. Subsequently, the specimens were cooled to different deformation temperatures (600°C, 700°C, and 800°C) at a cooling rate of 100°C / s and held for 30s. Then, the specimens were isothermally stretched to different deformation amounts (corresponding to the forming material degradation parameters) at different strain rates (0.01 / s, 0.1 / s, 1 / s, and 10 / s). The values ​​were 0, 0.1, 0.3, 0.5, and 0.7. The entire test included 60 test conditions consisting of 3 deformation temperatures, 4 strain rates, and 5 deformation amounts. The test was repeated twice under each condition.

[0127] Under given deformation temperature and strain rate conditions, for a specific value of the degradation parameter of the forming material, the deformation of the specimen is... The following formula can be used for calculation:

[0128]

[0129] In the formula: This is the original length of the gauge length of the specimen, which is 50 mm in this case. The fracture strain of the specimen under given deformation temperature and strain rate conditions can be obtained from the stress-strain curve obtained from the high-temperature uniaxial tensile test in step S111.

[0130] S122: All specimens that have completed the uniaxial tensile test without breaking shall be quenched to room temperature and the quenched specimens shall be divided into two groups; specimens that have undergone the same test conditions shall not be divided into the same group.

[0131] S123: Perform uniaxial tensile tests on the two sets of specimens at room temperature and different strain rates, and obtain the stress-strain curves of the two sets of specimens at room temperature.

[0132] In this embodiment, a room temperature uniaxial tensile test was conducted using an Instron 5982 electronic universal testing machine, which can precisely control the strain rate and loading path. The strain rates for the two sets of specimens tested were 1 / s and 10 / s, respectively. Figure 6 The data points in the figure show the stress-strain curves when the strain rate is 1 / s.

[0133] S124: Construct a model architecture that can both reflect the influence of hot forming history on the subsequent mechanical properties of the material and accurately characterize the room temperature deformation behavior of the boron steel sheet.

[0134] In this embodiment, the service portion of the forming-service material model is specifically represented by the following model architecture:

[0135]

[0136]

[0137]

[0138]

[0139]

[0140] In the formula: For stress; and These represent the work hardening term caused by dislocations and its evolution rate, respectively. and These are the degradation parameters and evolution rates of in-service materials; and These are the normalized dislocation density and its evolution rate, respectively. and These are plastic strain, plastic strain rate, and total strain, respectively. and These are the model coefficients.

[0141] S125: Select a suitable objective function and optimization algorithm, and determine the model coefficients of the service part of the forming-service material model based on the obtained stress-strain curve at room temperature.

[0142] In this embodiment, the same objective function and optimization algorithm as in step S113 are selected, and the corresponding source program is written using the same software to optimize the objective function. The entire process will not be described in detail here. Finally, the model coefficient values ​​of the service portion of the formed-service material model are as follows: =640、 =160、 =7.5、 =0.4、 =1100、 =100、 =105、 =1.55、 =0.09、 =0.2、 =10 and =207000. Substituting all model coefficient values ​​into the service part of the forming-service material model for calculation, the stress-strain curve obtained when the strain rate is 1 / s is compared with the stress-strain curve obtained from a room temperature uniaxial tensile test as follows: Figure 6 As shown.

[0143] S126: The service part of the forming-service material model is converted into a corresponding material subroutine using a high-level programming language and then embedded into numerical simulation software for verification.

[0144] In this embodiment, the service portion of the formed-service material model is converted into a corresponding material subroutine VUMAT_S using the Fortran language, which can be called by finite element simulation software such as LS-DYNA. This material subroutine mainly includes: a material parameter reading module, a state variable initialization module, a stress update module, and a constitutive relation calculation module.

[0145] To verify the correctness of the subroutine, a simulation model under the same conditions as the room temperature uniaxial tensile test was established, and VUMAT_S was called to perform simulation calculations. Finally, by comparing the stress-strain curves obtained from the simulation and the experiment, it was proved that the simulation model established based on VUMAT_S can simulate the room temperature deformation behavior of boron steel under uniaxial tension.

[0146] S2: Construct a deep learning agent model based on image representation.

[0147] Step S2 includes:

[0148] S21: Constructing a simulation model for the thermoforming of a part includes:

[0149] S211: Determine the geometry and expected forming conditions of hot-stamped parts during the conceptual design phase.

[0150] In this embodiment, a battery casing of a certain vehicle model is selected as the target hot-stamped part, and its geometric structure in the conceptual design stage is a square box with a flange.

[0151] S212: A program A is developed using a programming language. The program A is capable of automatically generating the geometric structure of the forming mold and the geometric structure of the sheet metal blank for hot stamping parts in computer-aided design software.

[0152] In this embodiment, the programming language and computer-aided design software used are VBA (Visual Basic for Applications) and CATIA, respectively. VBA achieves automated operation through macro programming.

[0153] S213: Program B is developed using a programming language. Program B is capable of automatically generating the finite element meshes of each component of the thermoforming simulation model in the finite element preprocessing software.

[0154] In this embodiment, the programming language and finite element preprocessing software used are TCL (Tool Command Language) and HyperMesh, respectively.

[0155] S214: A program C is developed using a programming language. The program C can automatically load the finite element mesh and the material subroutine of the forming-service material model in the finite element simulation software, and construct a thermoforming simulation model.

[0156] In this embodiment, the programming language and finite element simulation software used are Python and LS-DYNA, respectively. LS-DYNA supports automated operations through Python scripts, including generating and modifying input files, processing output data, parametric modeling, and post-processing analysis.

[0157] See Figure 7 The battery casing thermoforming simulation model includes four parts: sheet blank 3, punch 4, die 5, and blank holder 6. The initial dimensions of sheet blank 3 are 1300mm × 1300mm × 2mm, using a reduced integral shell element mesh, forming-service material model, and isotropic thermomaterial model. Punch 4, die 5, and blank holder 6 use a reduced integral shell element mesh, rigid material model, and isotropic thermomaterial model. Specific parameter settings are as follows: Poisson's ratio is 0.3 for all parts, Young's modulus is 210 GPa, and density is 7890 kg / m³. 3 All components have a thermal conductivity of 24 W / (m•K) and a specific heat capacity of 460 J / (kg•K). The punch 4, die 5, and blank holder 6 are all kept at a constant temperature of 25℃. The forming temperature of the sheet blank 3 is 800℃. All components are in surface-to-surface contact. The coefficient of friction is 0.1. The blank holder force is 20 kN.

[0158] Throughout the thermoforming simulation, the punch 4 remains stationary. The blank holder 6 presses down on the sheet blank 3 and moves downwards together with the die 5 at a stamping speed of 200 mm / s until the stroke ends. Subsequently, the punch 4, die 5, and blank holder 6 are used to hold the formed sheet blank 3 under pressure and quench it to obtain the final battery casing.

[0159] S215: Conduct experimental verification and parameter correction of the aforementioned thermoforming simulation model.

[0160] In this embodiment, the same parameters as the thermoforming simulation model were used for experimental verification, and the mechanical properties, thickness, and shape accuracy of the formed battery casing were measured. After three rounds of parameter correction, the errors between the simulation model's predicted indicators and the experimental results were all controlled within 8%.

[0161] S22: Determining input parameters and performing cloud computing batch simulation includes:

[0162] S221: The input parameters that affect the forming response of the hot stamping part include: some geometric parameters of the forming die, some geometric parameters of the sheet blank, and some process parameters.

[0163] See Figure 8 and Figure 9 In this embodiment, the final selected input parameters include: the upper corner radius of the forming mold. Lower fillet radius Transition fillet radius and sidewall height Length scaling factor for sheet metal blanks and corner radius scaling factor Die clearance Molding temperature and stamping speed .in, Indicates the length of the sheet metal blank. This indicates the radius of the fillet in the sheet blank.

[0164] S222: Based on engineering experience, determine the range and dispersion level of each input parameter, and use experimental design methods to generate multiple sets of input parameter combinations.

[0165] In this embodiment, and The value range is 5-30mm, with a horizontal value taken every 1mm. and The value range is 50-150mm, with a horizontal value taken every 10mm. The value ranges from 0.8 to 1.2, with a horizontal value taken every 0.1 mm. The value range is 0.2-1.2, with a horizontal value taken every 0.1 mm; The value range is 600-900℃, with a level taken every 25℃. The value range is 50-300 mm / s, and a horizontal value is taken every 25 mm / s. The value range is 2-8 mm, with a horizontal sample taken every 0.5 mm. Finally, the optimal Latin hypercube sampling method is used for the experimental design to ensure that the parameter space is uniformly covered and to avoid sample point clustering.

[0166] S223: Import each set of input parameter combinations into the thermoforming simulation model.

[0167] In this embodiment, a corresponding Python script was developed that can automatically and in batches import the various combinations of input parameters obtained from the experimental design into the thermoforming simulation model.

[0168] S224: Deploy numerical simulation tasks using a cloud computing platform to achieve parallel offline solutions for multiple sets of input parameter combinations.

[0169] In this embodiment, the AutoDL computing service is used, with the following specific configuration: Hardware environment consists of a single NVIDIA V100 32GB GPU, 6 vCPUs, 25GB of RAM, and 512GB of storage; Software environment includes Ubuntu 22.04, CUDA 12.1, PyTorch 2.1.0, Python 3.10, and LS-DYNA R13. Simulation task deployment includes: developing task scheduling scripts to automatically allocate parameter combinations; configuring a distributed file system to share input and result files; and setting up a monitoring system to track the running status of each task in real time. Parallel solution strategies include: employing a dynamic load balancing algorithm to ensure load balancing across computing nodes; and setting up a task timeout mechanism to automatically restart failed tasks.

[0170] S225: Collect the output results of each numerical simulation task, including the displacement field, thickness reduction field, and material degradation field of the hot-stamped part after forming is completed.

[0171] S23: The development of an automatic input / output image representation conversion program includes:

[0172] S231: Define the graphical representation rules for the selected forming die geometry, sheet metal blank geometry, and process parameters as inputs.

[0173] In this embodiment, the mold geometry is first converted into a finite element mesh, making it representable by a discrete node set. The mesh nodes are treated as point clouds in Cartesian coordinates in three-dimensional space, and the out-of-plane height values ​​are interpolated onto a uniform two-dimensional Cartesian mesh. Similarly, the sheet metal blank geometry is also interpolated onto the image, forming a binary image where pixel values ​​are "1" for areas with material and "0" for areas without material. The process parameters are then... , and Each image is multiplied by the blank image to form a new image representation. The pixel positions in the new image still represent the geometry of the blank, while the pixel values ​​represent the corresponding process parameter values.

[0174] S232: A program D is developed using a programming language. The program D can automatically convert any combination of forming mold geometry, sheet metal blank geometry, and process parameters into a corresponding input image combination.

[0175] In this embodiment, the programming language used is Python, and according to... Figure 10 Write program D for the algorithm shown.

[0176] S233: Define the graphical representation rules for the displacement field, thickness reduction field, and material degradation field of the selected hot-stamped part as output.

[0177] In this embodiment, similar to the processing of input images, finite element nodes are treated as point clouds, and the node data is converted into an image through interpolation. Specifically, based on the coordinates and displacements of each node in the formed part, the coordinates of the corresponding nodes in the sheet metal blank before forming are calculated. Subsequently, the point cloud data of the sheet metal blank before forming is interpolated onto a uniform two-dimensional Cartesian grid, where each pixel value represents the thickness reduction value and the material degradation value, respectively.

[0178] S234: A program E is developed using a programming language. The program E can automatically convert the displacement field, thickness reduction field, and material degradation field of the hot-stamped part that matches the input image combination into the corresponding output image combination.

[0179] In this embodiment, the programming language used is Python, and according to... Figure 11 Write a program for the algorithm shown. .

[0180] S24: Constructing a neural network proxy model based on image representation includes:

[0181] S241: Establish a database with a uniform format to record the storage address of the input and output image combinations on the server.

[0182] In this embodiment, a relational database MySQL is established, and a table structure containing fields such as input and output image addresses and unique identifiers is designed. The unified address format is "protocol: / / server information / path / filename", and the database initialization is completed.

[0183] S242: Store each set of input images and corresponding output images in pairs, and enter their address information into the database.

[0184] In this embodiment, input and output images are stored in pairs in corresponding directories according to the date + batch specification, associated with a unique identifier, and then the address information is entered into the database to ensure the consistency between storage and recording.

[0185] S243: Construct a deep neural network architecture with strong expressive power as a surrogate model.

[0186] In this embodiment, the mold geometry, sheet metal blank geometry, and forming process parameters are mapped to a two-dimensional image, and all inputs are normalized to ensure numerical stability during training. The output displacement field, thickness reduction field, and material degradation field are also normalized to ensure that the output physical quantities retain their actual physical meaning.

[0187] Based on the above preprocessing, SE-ResidualNet was selected as the surrogate model. (See also...) Figure 12 The deep neural network architecture includes: an input layer of... The multi-channel images correspond to the images obtained by mapping the mold geometry height, mold gap, forming temperature, and stamping speed, respectively. The encoder consists of four layers of convolution and downsampling units, which extract multi-scale spatial features through layer-by-layer convolution. The bottleneck layer introduces multiple cascaded residual modules, which, combined with residual connections and channel attention mechanisms, effectively enhance feature extraction capabilities. The decoder consists of four symmetrical layers of deconvolution and upsampling units, which fuse shallow features from the encoder with deep features from the decoder through skip connections, thereby restoring spatial resolution and improving prediction accuracy.

[0188] The output layer is configured differently depending on the prediction task: when predicting the displacement field, the output layer includes... Channels, respectively, correspond to the parts forming process during the process. Displacement distribution in three directions; when predicting the thickness reduction field, the output layer contains The channel corresponds to the thickness reduction distribution during the part forming process; when predicting the material degradation field, the output layer contains... The channels correspond to the material degradation distribution during the part forming process. Through the above design, the model can not only accurately capture local thinning and wrinkling during the part forming process, but also identify material degradation characteristics in real time, thereby enabling rapid evaluation of the forming feasibility of complex geometries under different process conditions.

[0189] S244: Divide the dataset into training and test sets, train and optimize the surrogate model on the training set using cross-validation based on the selected loss function, and evaluate its prediction accuracy on the test set.

[0190] In this embodiment, the data is randomly divided into training and test sets at a ratio of 90% and 10%, respectively. Mean squared error is selected as the loss function. Using the Adam optimizer within the PyTorch framework, iterative training is performed on the displacement neural network, thickness reduction neural network, and material degradation neural network, respectively. The prediction accuracy of the model is evaluated using the loss curve from the test set.

[0191] S245: Deploy the completed agent model, capable of rapid prediction and remote invocation, to the cloud.

[0192] In this embodiment, TensorFlow Serving is used to export the model to a deployable format and deploy it on the AutoDL cloud server. The TensorFlow Serving service is configured, and a RESTful API interface is developed. Through cloud deployment, the proxy model can be accessed by multiple users and platforms, enabling rapid evaluation of design parameters.

[0193] S3: Evaluate design feasibility based on forming-service simulation model.

[0194] Step S3 includes:

[0195] S31: The forming-service simulation model for building parts includes:

[0196] S311: Determine the service conditions and performance indicators of hot-stamped parts based on relevant regulations and actual needs.

[0197] In this embodiment, the service condition of the battery casing is set as a bottom impact condition, and its performance indicators include whether cracking occurs during service and the maximum intrusion amount.

[0198] S312: Based on the service conditions and geometry of the hot-stamped parts and the material subroutine of the forming-service material model, construct a service simulation model.

[0199] In this embodiment, service simulation models are established and solved in two finite element software programs, HyperMesh and LS-DYNA, respectively (see...). Figure 13 The model comprises two parts: the formed battery casing 7 and the punch 8. The actual geometry, mesh, and forming-service material model (including the material degradation history) of the battery casing 7 can be retrieved from the results file of the thermoforming simulation model. The punch 8 uses a reduced integral shell element mesh and a rigid material model. Specific parameter settings are as follows: both the battery casing 7 and the punch 8 have a Poisson's ratio of 0.3, a Young's modulus of 210 GPa, and a density of 7890 kg / m³. They are in surface-to-surface contact with each other, and the coefficient of friction is 0.3.

[0200] Throughout the entire service simulation, the outer edge of the battery casing 7 was completely fixed, and the punch 8 struck the battery casing 7 vertically with an initial velocity of 30 km / h until the end of the stroke.

[0201] S313: Conduct experimental verification and parameter correction of the aforementioned service simulation model.

[0202] In this embodiment, the same parameters as the service simulation model were used for experimental verification, and the occurrence of rupture and maximum intrusion during service were observed. After two rounds of parameter correction, the error between the simulation model's predicted indicators and the experimental results was controlled within 8%.

[0203] S314: Couple the thermoforming simulation model, which has been verified by experiments and corrected for parameters, with the service simulation model to form the forming-service simulation model.

[0204] S32: Evaluating the forming feasibility of the pre-selected part design includes:

[0205] S321: Construct a database of pre-selected design schemes, storing the geometric parameters of forming molds, geometric parameters of sheet metal blanks, and process parameters corresponding to each design scheme to be evaluated.

[0206] In this embodiment, a relational database MySQL is established to store the 500 pre-selected design schemes. and Values. Through a structured design that links the main table to sub-tables, the basic information of the solution and various detailed parameters are stored together.

[0207] S322: Using the program A described above, automatically generate the geometric structure of the forming mold and the geometric structure of the sheet metal blank corresponding to each design scheme to be evaluated in the computer-aided design software.

[0208] S323: Using the program D, the geometric structure of the forming mold, the geometric structure of the sheet metal blank, and the combination of process parameters corresponding to each design scheme to be evaluated are automatically converted into the corresponding input image combination to be evaluated.

[0209] S324: Combine the various input images to be evaluated and input them into the image representation-based neural network proxy model to filter out the formable design schemes.

[0210] In this embodiment, 420 feasible design schemes were ultimately selected.

[0211] S33: Evaluating and optimizing the service feasibility of formable solutions includes:

[0212] S331: Input the geometric parameters of the forming mold, the geometric parameters of the sheet metal blank, and the process parameters corresponding to the selected formable design schemes into the forming-service simulation model, solve and select all design schemes that meet the performance index requirements.

[0213] In this embodiment, the performance requirements are that the battery casing does not crack during service and the maximum intrusion does not exceed 10mm. Finally, 361 design schemes that meet the performance requirements were selected.

[0214] S332: Using the maximum weight of the part in the conceptual design stage as a constraint, a suitable optimization algorithm is used to search for the design scheme with the optimal performance index that satisfies the constraint.

[0215] In this embodiment, Isight is used as the optimization platform, integrating software such as HyperMesh and LS-DYNA. At the same time, Python scripts were developed to realize data transfer and process control between the various software programs.

[0216] The maximum weight constraint of the part, determined during the conceptual design phase, was set to 3 kg. The optimization objective was to minimize the maximum intrusion during service. A genetic algorithm was used for optimization: the population size was set to 50, the number of iterations was 50 generations, the crossover probability was 0.8, and the mutation probability was 0.05.

[0217] After 38 iterations, the optimal solution was finally obtained: and At this point, the weight of the battery casing is... Maximum intrusion volume reduced to .

[0218] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based hot stamping part conceptual design feasibility evaluation method, characterized by, The steps of the hot stamping part concept design feasibility evaluation method based on deep learning are as follows: 1) Construct a forming-service material model that inherits forming history, including constructing a forming part of the forming-service material model and constructing a service part of the forming-service material model; 2) Construct a deep learning agent model based on image representation, including constructing a part hot forming simulation model, determining input parameters and performing cloud computing batch simulation, developing an input-output image representation automatic conversion program, and constructing a neural network agent model based on image representation; 3) Evaluate design feasibility based on the forming-service simulation model, including constructing a forming-service simulation model of the part, evaluating the forming feasibility of the preselected part design scheme, evaluating the service feasibility of the formable scheme and optimizing.

2. The build-up of a forming-service material model inheriting a forming history according to claim 1, characterized in that, The steps of constructing the forming part and the service part of the forming-service material model include: 1) Perform uniaxial tensile test on the test piece made of boron steel sheet under different deformation temperatures and strain rates to obtain the stress-strain curve of the boron steel sheet at high temperature; 2) Select a model architecture that can accurately represent the high-temperature rheological behavior of the boron steel sheet and effectively transfer the hot forming history; 3) Select appropriate objective function and optimization algorithm, and determine the model coefficients of the forming part of the forming-service material model based on the obtained stress-strain curve at high temperature; 4) Convert the forming part of the forming-service material model into a corresponding material subroutine using a high-level programming language, and embed it in a numerical simulation software for verification; 5) Perform uniaxial tensile test on the test piece made of boron steel sheet under different deformation temperatures, strain rates and deformation amounts, all tests are completed on the premise that the test piece is not broken; wherein the test under the same temperature, strain rate and deformation amount conditions needs to be repeated twice; 6) Quench all test pieces that have completed the uniaxial tensile test to room temperature, and divide the quenched test pieces into two groups; wherein, for test pieces that have undergone the same test conditions, they must not be divided into the same group; 7) Perform uniaxial tensile test on the two groups of test pieces at room temperature and different strain rates to obtain the stress-strain curve of the two groups of test pieces at room temperature; 8) Construct a model architecture that can reflect the influence of hot forming history on the subsequent mechanical properties of the material and accurately represent the deformation behavior of the boron steel sheet at room temperature; 9) Select appropriate objective function and optimization algorithm, and determine the model coefficients of the service part of the forming-service material model based on the obtained stress-strain curve at room temperature; 10) Convert the service part of the forming-service material model into a corresponding material subroutine using a high-level programming language, and embed it in a numerical simulation software for verification. 3.The method of claim 1, wherein, The steps of constructing a part hot forming simulation model include: 1) Determine the geometric structure of the hot stamping part at the concept design stage and the expected forming working condition; 2) Develop program A using a programming language, which can automatically generate the geometric structure of the forming die of the hot stamping part and the geometric structure of the sheet blank in the computer-aided design software; 3) Program B is developed using a programming language. Program B can automatically generate the finite element mesh of each component of the thermoforming simulation model in the finite element preprocessing software. 4) A program C is developed using a programming language. The program C can automatically load the finite element mesh and the material subroutine of the forming-service material model in the finite element simulation software, and construct a thermoforming simulation model. 5) Conduct experimental verification and parameter correction of the aforementioned thermoforming simulation model.

4. The method of claim 1, wherein the constructing an image representation based deep learning agent model comprises: The process of determining input parameters and performing cloud computing batch simulation includes the following steps: 1) The input parameters that affect the forming response of the hot stamping parts include: some geometric parameters of the forming die, some geometric parameters of the sheet blank, and some process parameters, etc. 2) Based on engineering experience, determine the value range and dispersion level of each input parameter, and use experimental design methods to generate multiple sets of input parameter combinations; 3) Import each set of input parameters into the aforementioned thermoforming simulation model; 4) Utilize cloud computing platforms to deploy numerical simulation tasks and achieve parallel offline solutions for multiple combinations of input parameters; 5) Collect the output results of each numerical simulation task, including: the displacement field, thickness reduction field, and material degradation field of the hot stamped part after forming is completed.

5. The method of claim 1, wherein the constructing an image representation based deep learning agent model comprises: The development of the automatic input / output image representation conversion program includes the following steps: 1) Develop image representation rules for the selected forming die geometry, sheet metal blank geometry, and process parameters as inputs; 2) A program D is developed using a programming language. The program D can automatically convert any combination of forming mold geometry, sheet metal blank geometry and process parameters into a corresponding input image combination. 3) Develop graphical representation rules for the displacement field, thickness reduction field, and material degradation field of the selected hot-stamped parts as output; 4) A program E is developed using a programming language. The program E can automatically convert the displacement field, thickness reduction field, and material degradation field of the hot-stamped part that matches the input image combination into the corresponding output image combination.

6. The method of claim 1, wherein the constructing an image representation based deep learning agent model comprises: The construction of the image representation-based neural network proxy model includes the following steps: 1) Establish a database with a unified format to record the storage address of the input and output image combinations on the server; 2) Store each set of input images and their corresponding output images in pairs, and enter their address information into the database; 3) Construct a deep neural network architecture with strong expressive power as a proxy model; 4) Divide the data into training and test sets. Based on the selected loss function, train and optimize the surrogate model on the training set through cross-validation, and evaluate its prediction accuracy on the test set. 5) Deploy the completed agent model, which enables rapid prediction and remote invocation, to the cloud.

7. The method of claim 1, wherein the design feasibility is evaluated based on a shape-service simulation model. The forming-service simulation model of the constructed part includes the following steps: 1) Determine the service conditions and performance indicators of hot-stamped parts based on relevant regulations and actual needs; 2) Based on the service conditions and geometry of the hot-stamped parts and the material subroutine of the forming-service material model, a service simulation model is constructed; 3) Conduct experimental verification and parameter correction of the aforementioned service simulation model; 4) The thermoforming simulation model, after being verified by experiments and with parameters corrected, is coupled with the service simulation model to form the forming-service simulation model.

8. The method of claim 1, wherein the design feasibility is evaluated based on a shape-service simulation model. The evaluation of the feasibility of the pre-selected part design scheme includes the following steps: 1) Construct a database of pre-selected design schemes to store the geometric parameters of forming molds, geometric parameters of sheet metal blanks, and process parameters corresponding to each design scheme to be evaluated; 2) Using the program A described above, the geometric structure of the forming mold and the geometric structure of the sheet blank corresponding to each design scheme to be evaluated are automatically generated in the computer-aided design software. 3) Using the program D, the geometric structure of the forming mold, the geometric structure of the sheet blank, and the combination of process parameters corresponding to each design scheme to be evaluated are automatically converted into the corresponding combination of input images to be evaluated. 4) Combine the various input images to be evaluated and input them into the image representation-based neural network proxy model to filter out the feasible design schemes.

9. The method of claim 1, wherein the design feasibility is evaluated based on a shape-service simulation model. The evaluation and optimization of the service feasibility of the formable solution includes the following steps: 1) Input the geometric parameters of the forming mold, the geometric parameters of the sheet metal blank, and the process parameters corresponding to the selected formable design schemes into the forming-service simulation model, solve and select all design schemes that meet the performance index requirements. 2) Using the maximum weight of the part in the conceptual design stage as a constraint, a suitable optimization algorithm is used to search for the design scheme with the optimal performance index that satisfies the constraint.

Images