A press free forging large-scale drop hammer multi-parameter multi-field intelligent optimization method and system

Abstract

The present application relates to a kind of press free forging large-scale drop hammer multi-parameter multi-field intelligent optimization method and system, through the three-field coupling modeling of heat-force-microstructure, depth nuclear learning agent model and the multi-objective evolutionary algorithm guided by digital twin, realize the collaborative optimization of process parameters, die geometry and microstructure, improve optimization efficiency, reduce the number of trial die;While building real-time closed-loop control system, realize the grain size uniformity of forging, forming load reduction and production cycle shortening, improve process stability and die life.

Description

Technical Field

[0001] This invention relates to the field of design of large-scale die for free forging presses, and in particular to a multi-parameter, multi-field intelligent optimization method and system for large-scale die for free forging presses. Background Technology

[0002] With the rapid development of high-end equipment manufacturing, large forgings are increasingly widely used in nuclear power, aerospace and other fields. Press free forging, as a core forming process, directly affects the performance and production efficiency of forgings due to the design quality of its die. In the field of press free forging large die design, traditional methods often rely on numerical simulation combined with manual experience for fine-tuning. For example, patent CN119203652A discloses an online optimization method and device for hot forging process parameters. It optimizes process parameters through multi-scale simulation of hot forging, but its dynamic recrystallization model relies on empirical formulas, resulting in large errors in grain size prediction and the lack of hard constraint optimization. The problems include: 1. Lack of multi-field coupling: Only considering the thermo-mechanical dual fields, without constructing a dynamic feedback mechanism between the microstructure field and macroscopic stress, leading to grain coarsening (standard deviation ≥ 20%) and low pass rate of forging mechanical properties. 2. Low optimization efficiency: Relying on empirical formulas and simple biomimetic algorithms, it lacks adaptive variation strategies for temperature gradients. Isolated optimization of process parameters, die geometry, and microstructure parameters is inefficient and prolongs the design cycle. 3. Control lag: Lack of real-time data acquisition and closed-loop feedback results in anomaly response times exceeding 2 seconds and a yield rate of less than 80%. Furthermore, in actual production, the high coefficient of variation in grain size uniformity significantly shortens mold life. Therefore, a mold design method integrating multi-physics coupled modeling, intelligent optimization algorithms, and real-time feedback control is urgently needed. Summary of the Invention

[0003] (a) Technical problems to be solved In view of the problems of incomplete multi-field coupling, low intelligent optimization efficiency, and lack of closed-loop control in existing large-scale die design methods, this invention provides a multi-parameter, multi-field intelligent optimization method and system for large-scale die design in press free forging. It breaks through the limitations of traditional thermodynamic simulation, accurately predicts grain evolution through three-field coupling modeling, and realizes the synergistic optimization of process parameters, die geometry and microstructure, thereby improving optimization efficiency and reducing the number of die trials. A real-time closed-loop control system is constructed to improve the uniformity of forging grain size, reduce forming load, shorten production cycle, and enhance process stability and mold life.

[0004] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a multi-parameter, multi-field intelligent optimization method for large-scale free forging dies in presses, comprising the following steps: Multi-field modeling: Input the target shape of the forging, material properties and process constraints, construct a multi-physics coupled simulation model, and realize real-time interaction of multiple fields through a unified time step; Parameter pre-screening: An initial parameter set is generated using Latin hypercube sampling. A deep kernel learning surrogate model is then used to simultaneously predict macroscopic molding accuracy, microscopic grain size, and residual stress. The range of variables covered by Latin hypercube sampling includes temperature gradient coefficients, strain rate thresholds, and mold heat transfer correction terms. Specifically, the temperature gradient coefficient range is... The strain rate threshold is The mold heat exchange correction item is .

[0005] Objective optimization: A Pareto optimal solution set is generated using a digital twin-guided multi-objective evolutionary algorithm (DT-Guided EA) that satisfies a dynamic recrystallization volume fraction >85%. Mold generation: Based on optimal parameters, the parametric CAD template is used to generate the mold geometry, and the grain size distribution map and thermal parameter mapping database are called to automatically verify the design compliance. The parametric CAD generation is based on B-spline surface reconstruction of the mold cavity, outputting a STEP format model, and then automatically generating a compliance report and triggering CAD template updates.

[0006] Processing monitoring: Temperature and strain field data are collected in real time using a high frame rate infrared thermal imager and an embedded strain sensor, and dynamic recrystallization calculations are performed at the edge computing node; Closed-loop feedback: Based on LSTM network analysis of actual production data and simulation errors, the digital twin and surrogate model are updated, and cross-production line knowledge transfer is achieved through a federated learning framework; LSTM error modeling takes time-series data (temperature, strain rate, grain size) as input and outputs the recrystallization activation energy correction amount. The federated learning framework uses the FedAvg algorithm to aggregate parameters. Clients upload model parameters encrypted with AES-256, and the central server aggregates them, performs integrity verification using SHA-256, and distributes them. Local data is not shared.

[0007] The multiphysics coupling simulation model is a three-field coupling model of thermo-mechanical-microstructure, in which: Thermodynamic field modeling: Establish unsteady-state heat transfer equations, simultaneously establish the mold-workpiece interface heat transfer model, and integrate plastic work-heat source terms. in, For the efficiency of plastic work to heat conversion, For interfacial frictional heat generation, For contact heat dissipation; here ; Furthermore, The coefficient of friction; Furthermore, .

[0008] Stress-strain field modeling: Based on the modified Johnson-Cook constitutive equation, piecewise correction of flow stress is performed when... Activate softening correction at time: Where A is the initial yield strength (unit: MPa), B is the strain hardening coefficient (unit: MPa), and n is the hardening exponent (dimensionless). Equivalent plastic strain (dimensionless). This is the volume fraction of dynamic recrystallization (dimensionless). C is the softening coefficient, and C is the strain rate sensitivity coefficient. The strain rate is dimensionless. is a dimensionless temperature, and m is the thermal softening index.

[0009] Microstructure field modeling: Dynamic recrystallization and grain evolution are simulated using cellular automata; nucleation density is correlated with strain excess; grain boundary migration is determined by curvature. and storage energy gradient Co-driving: in, For grain boundary mobility, , For energy storage weighting coefficients, .

[0010] The three fields achieve real-time interaction through a unified time step: the temperature field drives the recrystallization activation energy update, the strain field triggers grain evolution, and grain size inversely corrects macroscopic flow stress, all within a unified time step. .

[0011] The implementation steps of the deep kernel learning surrogate model in parameter pre-screening include: The input variables are defined as process parameters (temperature gradient coefficient, strain rate threshold, mold heat transfer correction term) and three-field coupled state variables; The network structure is designed to use a 3-layer fully connected neural network, where the output layer embeds the Matern kernel function. The joint probability distribution of the three indicators was established through Bayesian optimization to screen constraint parameters, including the dynamic recrystallization initiation threshold. ) and grain size standard deviation (<15%); The model is trained using historical data and cross-validated to evaluate its generalization ability.

[0012] Digital twin-guided multi-objective evolutionary algorithms (DT-Guided EA) in objective optimization include: Digital twin pre-verification: A reduced-order model is used to quickly simulate candidate solutions, retaining only solutions with a dynamic recrystallization volume fraction > 85%; that is, a lightweight digital twin is constructed, integrating a thermo-mechanical-microstructure reduced-order model; candidate solutions with a dynamic recrystallization volume fraction > 85% are screened, eliminating invalid solutions and improving efficiency by 3 times.

[0013] Gradient-sensitive variation: based on the temperature gradient field ( The mutation probability is dynamically adjusted, and the formula for the mutation intensity coefficient is: in, , , For high gradient regions ( Apply Gaussian variation (standard deviation) ).

[0014] Target hard constraint optimization: The NSGA-III algorithm is used to minimize forming load and average grain size. The target is to rigidly constrain the standard deviation of grain size to be <15%.

[0015] Automatic verification of design compliance during mold generation includes: Access the historical thermal parameter database. Match the grain size variation coefficient of the current mold cavity; For regions where the coefficient of variation exceeds the limit (>15%), adjust the fillet radius of the mold based on the strain gradient distribution (compensation amount). ,in, ).

[0016] The federated learning framework in closed-loop feedback includes: The central server aggregates the model parameters of each production line and uses the FedAvg algorithm to allocate weights. The client only uploads model parameters encrypted with AES-256. The central server aggregates the data, performs integrity verification using SHA-256, and distributes the data. Local data is not shared, and the model update cycle is 24 hours.

[0017] A system for a multi-parameter, multi-field intelligent optimization method for large-scale free forging dies in presses includes: Multi-field coupling modeling module: used to construct a three-field coupling model of thermo-mechanical-microstructure field, including temperature field modeling unit, stress field modeling unit and microstructure evolution unit, integrating plastic work heat source and dynamic recrystallization evolution equation; Parameter pre-screening module: Deploys Latin hypercube sampling and deep kernel learning models, and outputs the initial parameter set; Intelligent optimization module: Runs the DT-Guided EA algorithm to generate the Pareto optimal solution set; Mold generation module: Generates mold geometry based on parametric CAD templates and triggers local geometric compensation; Processing monitoring module: connects an infrared thermal imager and a strain sensor to calculate the dynamic recrystallization volume fraction in real time; Closed-loop feedback module: The digital twin is updated through an LSTM network, and the federated learning framework enables cross-production line collaboration.

[0018] The system runs on a distributed architecture, including: High-performance simulation cluster: Employs MPI parallel computing for three-field coupled models, with load difference <5%; Edge computing node: Data is acquired at a frequency of 1kHz, noise is reduced by an FIR filter, and the response time is <2ms; Cloud platform: Stores thermal parameter mapping database (MongoDB sharded cluster) and federated learning model library (AES-256 encryption).

[0019] (III) Beneficial Effects (1) By using a three-field coupled modeling approach involving thermo-mechanical-microstructure, the limitations of traditional thermo-mechanical simulation are overcome. Through real-time interactive feedback of temperature, stress, and grain evolution, the grain size distribution and mechanical properties are accurately predicted. At the same time, deep kernel learning and the DT-Guided EA algorithm are combined with multi-objective optimization to simultaneously minimize forming load and grain size while ensuring the dynamic recrystallization volume fraction. This achieves dual optimization of macroscopic forming accuracy and microstructure quality, improving grain size uniformity, reducing residual stress, and increasing the pass rate of forging mechanical properties. Experimental data show that: (2) The LSTM-Federated learning mechanism enables dynamic updating of digital twin parameters and breaks down data silos through cross-device knowledge transfer, effectively reducing errors. The entire system achieves a closed-loop control and feedback of "perception-analysis-decision-execution" from design to production, promoting the forging industry to upgrade to intelligent manufacturing and improving the consistency of the entire industrial chain process. Attached Figure Description

[0020] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart of a multi-parameter, multi-field intelligent optimization method and system for a large-scale free forging die in a press, as described in this application. Figure 2This is a schematic diagram of the multi-parameter, multi-field intelligent optimization method and system for large-scale free forging die of a press, as described in this application; Figure 3 This is a schematic diagram illustrating the multi-parameter, multi-field intelligent optimization method and system for large-scale free forging dies in a press, based on this application; (arrows indicate the direction of data transmission). Figure 4 This is a flowchart of the DT-Guided EA optimization algorithm in the system of a multi-parameter, multi-field intelligent optimization method for large-scale free forging dies in a press, as described in this application; (three stages: pre-validation, mutation, and optimization). Figure 5 This application presents a multi-parameter, multi-field intelligent optimization method for a large-scale free forging die in a press and a timing diagram of closed-loop feedback in the system. Figure 6 This is a compliance verification interface diagram for a multi-parameter, multi-field intelligent optimization method and system for large-scale free forging dies in a press, as described in this application; (showing grain size distribution and geometric compensation scheme). Figure 7 This application presents a multi-parameter, multi-field intelligent optimization method and system federated learning architecture diagram for a large-scale free forging die in a press; (data interaction between the central server and multiple production line clients). Figure 8 This is a diagram showing the module composition under the distributed architecture support of the intelligent optimization method and system for large-scale free forging die of a press according to this application; Detailed Implementation

[0021] The following will refer to the appendix in the examples of this invention. Figure 1 - Appendix Figure 8 The technical solutions in the embodiments of the present invention are clearly and completely described. 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 are within the scope of protection of the present invention.

[0022] A multi-parameter, multi-field intelligent optimization method for large-scale free forging dies in presses includes the following steps: Multi-field modeling: Inputting the target shape of the forging, material properties, and process constraints, a multi-physics coupled simulation model is constructed, enabling real-time interaction across multiple fields through a unified time step. The target shape of the forging is input in STL format, material properties include but are not limited to elastic modulus and thermal conductivity, and process constraints include but are not limited to press tonnage and allowable stress of the die. A thermo-mechanical-microstructure three-field coupled model is constructed, and the model is implemented through a unified time step. Enable real-time interaction.

[0023] Parameter pre-screening: Latin hypercube sampling is used to generate an initial parameter set, and a deep kernel learning surrogate model is used to simultaneously predict macroscopic forming accuracy, microscopic grain size and residual stress; Objective optimization: A Pareto optimal solution set is generated using a digital twin-guided multi-objective evolutionary algorithm (DT-Guided EA) that satisfies a dynamic recrystallization volume fraction >85%. Mold generation: Based on optimal parameter-driven parametric CAD templates, mold geometry is generated, and the design compliance is automatically verified by calling the grain size distribution map and thermodynamic parameter mapping database; grain size mapping and compensation, that is, calling the historical database to match the grain size variation coefficient: Then, for In the region where the strain gradient is used, the fillet radius is adjusted using the formula: Processing monitoring: Temperature and strain field data are collected in real time using a high frame rate infrared thermal imager and an embedded strain sensor, and dynamic recrystallization calculations are performed at the edge computing node; Closed-loop feedback: Based on LSTM network analysis of actual production data and simulation errors, update digital twins and agent models, and realize cross-production line knowledge transfer through federated learning framework.

[0024] The multiphysics coupling simulation model is a three-field coupling model of thermo-mechanical-microstructure, in which: Thermodynamic field modeling: Establish unsteady-state heat transfer equations, simultaneously establish the mold-workpiece interface heat transfer model, and integrate plastic work-heat source terms. in, For the thermal conductivity of the material, For the efficiency of plastic work to heat conversion, For interfacial frictional heat generation, For contact heat dissipation; here ; Furthermore, The coefficient of friction, To contact pressure, Relative sliding speed; Furthermore, The interfacial heat transfer coefficient is denoted as .

[0025] Stress-strain field modeling: Based on the modified Johnson-Cook constitutive equation, piecewise correction of flow stress: Where A is the initial yield strength, B is the strain hardening coefficient, and n is the hardening exponent. This refers to the volume fraction of dynamic recrystallization. The softening coefficient is given; the dynamic recrystallization volume fraction here is calculated using cellular automata simulation; when The softening correction will be activated at that time.

[0026] Microstructure field modeling: Dynamic recrystallization and grain evolution are simulated using cellular automata; nucleation density is correlated with strain excess; grain boundary migration is determined by curvature. and storage energy gradient Co-driving: in, For grain boundary mobility, This is the weighting coefficient for energy storage.

[0027] The three fields achieve real-time interaction through a unified time step: the temperature field drives the recrystallization activation energy update, the strain field triggers grain evolution, and grain size inversely corrects macroscopic flow stress, all within a unified time step. .

[0028] The implementation steps of the deep kernel learning surrogate model in parameter pre-screening include: The input variables are defined as process parameters (temperature gradient coefficient, strain rate threshold, and mold heat transfer correction term) and three-field coupled state variables. Further explanation is needed here: parameter pre-screening uses Latin hypercube sampling to generate an initial parameter set (covering the temperature gradient coefficient). strain rate threshold The mold heat transfer correction term is 0.8-1.2), and it is predicted by a deep kernel learning surrogate model. The network structure is designed to use a 3-layer fully connected neural network, where the output layer embeds the Matern kernel function. By establishing the joint probability distribution of the three indicators through Bayesian optimization, constraint parameters are screened, retaining those that meet the dynamic recrystallization initiation threshold. ) and parameters for grain size standard deviation (<15%); The model is trained using historical data and cross-validated to evaluate its generalization ability.

[0029] Digital twin-guided multi-objective evolutionary algorithms (DT-Guided EA) in objective optimization include: Digital twin pre-verification: A reduced-order model is used to quickly simulate candidate solutions, retaining only those with a dynamic recrystallization volume fraction >85%, with an elimination rate ≥60%. Gradient-sensitive variation: based on the temperature gradient field ( The mutation probability is dynamically adjusted, and the formula for the mutation intensity coefficient is: in, , , Apply Gaussian variation to the high gradient region, with a standard deviation of ; ; Target hard constraint optimization: The NSGA-III algorithm is used to minimize forming load and average grain size. The target is to rigidly constrain the standard deviation of grain size to be <15%. Automatic verification of design compliance during mold generation includes: Access the historical thermal parameter database. Match the grain size variation coefficient of the current mold cavity; To further explain, mold generation is based on determining the parameterized CAD template using optimal parameters, i.e., B-spline surface reconstruction, with a tolerance of ±0.1mm. The grain size distribution map is checked against the thermal parameter database for compliance and regularity. For regions where the coefficient of variation exceeds the limit (>15%), based on the strain gradient distribution ( ,in, Adjust the corner radius of the mold.

[0030] The federated learning framework in closed-loop feedback includes: The central server aggregates the model parameters of each production line and uses the FedAvg algorithm to allocate weights. The client only uploads model parameters encrypted with AES-256. The central server aggregates the data, performs integrity verification using SHA-256, and distributes the data. Local data is not shared, and the model update cycle is 24 hours.

[0031] A system for a multi-parameter, multi-field intelligent optimization method for large-scale free forging dies in presses includes: Multi-field coupling modeling module: used to construct a three-field coupling model of thermo-mechanical-microstructure field, including temperature field modeling unit, stress field modeling unit and microstructure evolution unit, integrating plastic work heat source and dynamic recrystallization evolution equation; Parameter pre-screening module: Deploys Latin hypercube sampling and deep kernel learning models, and outputs the initial parameter set; Intelligent optimization module: Runs the DT-Guided EA algorithm to generate the Pareto optimal solution set; Mold generation module: Generates mold geometry based on parametric CAD templates and triggers local geometric compensation; Processing monitoring module: Connects an infrared thermal imager and a strain sensor to calculate the dynamic recrystallization volume fraction in real time; that is, it collects data in real time through a high frame rate infrared thermal imager (accuracy ±1℃) and an embedded strain sensor (sampling rate 1kHz), and the edge computing node performs dynamic recrystallization calculation at a frequency of 1kHz, with a response time of <2ms.

[0032] Closed-loop feedback module: Updates the digital twin through an LSTM network, and achieves cross-production line collaboration through a federated learning framework. Specifically, it inputs time-series data (temperature, strain rate, grain size) into the LSTM network, outputs recrystallization activation energy correction, and updates the digital twin. Then, it aggregates model parameters from each production line using a federated learning framework (FedAvg algorithm). Clients upload AES-256 encrypted model parameters, and the central server aggregates them, performs integrity verification using SHA-256, and distributes them. Local data is not shared, and the model update cycle is 24 hours.

[0033] The system runs on a distributed architecture, including: High-performance simulation cluster: Employs MPI parallel computing for three-field coupled models, with load difference <5%; Edge computing node: Data is acquired at a frequency of 1kHz, noise is reduced by an FIR filter, and the response time is <2ms; Cloud platform: Stores thermal parameter mapping database (MongoDB sharded cluster) and federated learning model library (AES-256 encryption).

[0034] Example 1: Referring to the accompanying drawings, this example aims to intelligently optimize a certain aero-engine turbine disk forging through a multi-parameter, multi-field intelligent optimization method and system. That is, to design a high-precision die to ensure the uniformity of grain size in the blade tenon groove area (target coefficient of variation ≤15%), and avoid fatigue fracture caused by grain coarsening during service.

[0035] The implementation steps are as follows: Figure 1 As shown, the parameter settings for the forging must first be clarified, including its material properties (Ti-6Al-4V alloy, initial temperature 950℃, strain rate...). Process constraints (maximum hydraulic press tonnage 10000 kN, allowable mold stress 800 MPa) and optimization objectives (forming load <1000 kN, average grain size) The dynamic recrystallization volume fraction is ≥85%. This means that parameters must be input before constructing the three-field coupling model. After the parameter settings are clarified, the multi-field coupling modeling module begins to construct a three-field coupling model, including the thermodynamic field, the stress-strain field, and the microstructure field. The thermodynamic field is established by setting the heat transfer coefficient at the mold-workpiece interface by establishing an unsteady heat transfer equation. And through plastic work heat source item Modeling was implemented; the stress-strain field was modeled using a modified Johnson-Cook model with key parameters A = 850 MPa, B = 680 MPa, n = 0.4, C = 0.015, and m = 1.2; and the dynamic recrystallization softening coefficient was also used. The microstructure field is calculated using cellular automata, and its grid resolution is... Initial grain size .

[0036] After the thermo-mechanical-microstructure three-field coupling modeling is completed, the parameter pre-screening module is entered to perform parameter pre-screening and surrogate model training. That is, 200 sets of initial parameters are first generated through Latin hypercube sampling, covering the temperature gradient coefficient ( ), strain rate threshold ( Subsequently, the dataset was divided into a training set (70%), a validation set (15%), and a test set (15%) using a deep kernel learning model. After the dataset was divided, the error was predicted using a surrogate model, resulting in an MSE < 0.05 and a grain size standard deviation prediction accuracy > 90%.

[0037] Subsequently, the intelligent optimization module utilizes a digital twin-guided multi-objective evolutionary algorithm (DT-Guided EA) to generate the optimal solution set. First, digital twin pre-validation is performed, initially with 100 candidate solutions. After screening using a lightweight twin, 35 solutions are retained. These 35 solutions simultaneously satisfy a requirement, namely… After that, the temperature gradient was... In that region, the mutation probability increased to 0.4, with a standard deviation of [missing value]. Then, the NSGA-III algorithm was used for 50 iterations to obtain the final Pareto solution set containing 20 sets of parameters, with a forming load range of 820-950 kN and an average grain size. .

[0038] After the target intelligent optimization is completed, the module proceeds to the mold generation module. In this module, the mold cavity geometry is generated based on the optimal parameters (B-spline surface reconstruction, tolerance within ±0.1mm). Subsequently, the grain size variation coefficient is checked, specifically the out-of-limit region (CV=17%), which is compensated by adjusting the fillet radius. ).

[0039] The processing monitoring module collects data in real time during these processes. The collected data includes the temperature field collected by the infrared thermal imager (accuracy ±1℃) and the strain field collected by the embedded sensor (sampling rate 1kHz).

[0040] The closed-loop feedback module updates the digital twin parameters through the LSTM network, reducing the recrystallization activation energy prediction error from 12% to 3.2%; at the same time, federated learning aggregates three production line models, reducing the global grain size standard deviation prediction error from 15% to 7%.

[0041] The specific optimization results are shown in the table below: Example 2: Optimization of forging die for low-pressure rotor in nuclear power plants. First, input the parameters, specifically: material properties (SA508-3 steel, initial temperature 850℃, strain rate). Process constraints (mold preheating temperature 300℃, maximum pressing speed) ), optimization objective (avoiding abnormal grain growth, i.e., grain size) (Mold life > 1000 cycles); then the multi-field coupling modeling module will perform modeling, first the thermodynamic field, where the heat transfer coefficient at the mold-workpiece interface is... Friction generates heat The pressure here ,speed Following this is the stress-strain field, where the Johnson-Cook parameters are A = 600 MPa, B = 550 MPa, n = 0.35, C = 0.02, m = 1.0, and the dynamic recrystallization softening coefficient is... The initial grain size in the microstructure field is 50 μm, and the stored energy is calculated based on the dislocation density model. After establishing the thermo-mechanical-microstructure three-field coupling model, the model enters the parameter pre-screening module. In the parameter pre-screening module, 150 sets of parameters are generated by Latin hypercube sampling, and candidate solutions with a grain size standard deviation of <15% are screened out. At the same time, the surrogate model predicts a maximum grain size error of <8%. Then, the model enters the intelligent optimization module, where solutions with a dynamic recrystallization volume fraction >80% are screened through digital twin pre-verification, with an elimination rate of 65%. Gradient-sensitive variation is used to optimize solutions in high strain rate regions ( The mutation probability is increased to 0.5, and the maximum grain size is minimized through the objective function optimized by hard constraints. The system generates 15 Pareto optimal solution sets. After intelligent optimization, it enters the mold generation module. In the mold generation module, the draft angle is increased by 2° for the region with a grain size >90μm through mold cavity geometry compensation, and a compliance report is generated, triggering CAD template update. Finally, it enters the closed-loop feedback template, and the recrystallization connection energy parameters are corrected through LSTM network, reducing the prediction error from 18% to 5%. At the same time, the federated learning update cycle is shortened to 12 hours, and the cross-site model synchronization error is <3%. The processing monitoring module monitors the grain size online in real time through EBSD, with a monitoring rate of 10Hz.

[0042] The specific optimization results are shown in the table below:

Claims

1. A multi-parameter, multi-field intelligent optimization method for large-scale free forging dies in presses, characterized in that, Includes the following steps: Multi-field modeling: Input the target shape of the forging, material properties and process constraints, and construct a three-field coupled simulation model of thermo-mechanical-microstructure. The three fields achieve real-time interaction through a unified time step; among them, the temperature field drives the recrystallization activation energy update, the strain field triggers grain evolution, and the grain size inversely corrects the macroscopic flow stress. The construction of the thermo-mechanical-microstructure three-field coupled simulation model includes thermodynamic field modeling, stress-strain field modeling, and microstructure field modeling; The thermodynamic field modeling establishes unsteady-state heat transfer equations, combines the mold-workpiece interface heat transfer model, and integrates plastic work-heat source terms. in, For material density, The specific heat capacity at constant pressure of the material. For temperature, Thermal conductivity of the material, in units of , For the efficiency of plastic work to heat conversion, For stress tensor, For strain rate tensor, the physical characterization is strain rate. This indicates the term related to interfacial frictional heat generation. This refers to the contact heat dissipation item; The stress-strain field modeling is based on the modified Johnson-Cook constitutive equation, with piecewise correction of the flow stress, taking into account the dynamic recrystallization volume fraction. The softening modification will be activated at that time. in, The initial yield strength, The strain hardening coefficient is... The hardening index, For equivalent plastic strain, This refers to the volume fraction of dynamic recrystallization. The softening coefficient is... The strain rate sensitivity coefficient, The strain rate is dimensionless. The temperature is dimensionless. The thermal softening index; The microstructure field modeling simulates dynamic recrystallization and grain evolution through cellular automata, with nucleation density correlated with strain excess, and grain boundary migration driven by curvature and storage energy gradient. Parameter pre-screening: Latin hypercube sampling is used to generate an initial parameter set, and a deep kernel learning surrogate model is used to simultaneously predict macroscopic forming accuracy, microscopic grain size and residual stress; Objective optimization: A Pareto optimal solution set satisfying dynamic recrystallization volume fraction > 85% is generated using a digital twin-guided multi-objective evolutionary algorithm (DT-Guided EA). The digital twin-guided multi-objective evolutionary algorithm (DT-Guided EA) includes three stages: digital twin pre-validation, gradient-sensitive mutation, and objective hard constraint optimization. Among them, digital twin pre-verification: a reduced-order model is used to quickly simulate candidate solutions, and only solutions with a dynamic recrystallization volume fraction >85% are retained; Gradient-sensitive variability: The variability probability is dynamically adjusted based on the temperature gradient field. The formula for the variability intensity coefficient is: in, Represents the basic mutation probability. This represents the influence coefficient of the temperature gradient. This represents the strain gradient influence coefficient, which is dimensionless after normalization. Represents the temperature gradient field. Represents the strain gradient field; Target hard constraint optimization: The NSGA-III algorithm is used to minimize forming load and average grain size. The target is a hard constraint on grain size standard deviation <15%; Mold generation: The mold geometry is generated based on the optimal parameter-driven parametric CAD template, and the design compliance is automatically verified by calling the grain size distribution map and thermodynamic parameter mapping database; Processing monitoring: Temperature and strain field data are collected in real time using a high frame rate infrared thermal imager and an embedded strain sensor, and dynamic recrystallization calculations are performed at the edge computing node; Closed-loop feedback: Time-series data, including temperature, strain rate, and grain size, are input through an LSTM network, and the recrystallization activation energy correction is output to update the digital twin. The federated learning framework uses the FedAvg algorithm to aggregate parameters. The client uploads model parameters encrypted with AES-256, and the central server aggregates them and performs integrity verification using SHA-256 before distributing them. Local data is not shared.

2. The intelligent optimization method for large-scale free forging die of a press according to claim 1, characterized in that, The term for interfacial frictional heat generation is: Among them, the coefficient of friction Contact pressure relative sliding speed ; Contact heat dissipation items are: Among them, the interfacial heat transfer coefficient ; The initial temperature of the forging. Preheating temperature for the mold.

3. The intelligent optimization method for large-scale free forging die of a press according to claim 1, characterized in that, The implementation steps of the deep kernel learning proxy model in the parameter pre-screening include: The input variables are defined as process parameters and three-field coupling state variables; The network structure is designed to use a 3-layer fully connected neural network, where the output layer embeds the Matern kernel function. By establishing the joint probability distribution of the three indices through Bayesian optimization, constraint parameters are screened, including the dynamic recrystallization initiation threshold and the standard deviation of grain size. The model is trained using historical data and cross-validated to evaluate its generalization ability.

4. The intelligent optimization method for large-scale free forging die of a press according to claim 1, characterized in that, The automatic verification of design compliance during mold generation includes: Call the historical thermal parameter database to match the grain size variation coefficient of the current mold cavity; For regions where the coefficient of variation exceeds the limit, the fillet radius of the mold is adjusted based on the strain gradient distribution. The fillet radius compensation amount is: in, strain gradient Obtained through finite element analysis.

5. The intelligent optimization method for large-scale free forging die of a press according to claim 1, characterized in that, The federated learning framework in the closed-loop feedback includes: The central server aggregates the model parameters of each production line and uses the FedAvg algorithm to allocate weights. The client only uploads model parameters encrypted with AES-256. The central server aggregates the data, performs integrity verification using SHA-256, and then distributes the data. Local data is not shared.

6. A system for implementing the multi-parameter, multi-field intelligent optimization method for large-scale free forging die in a press as described in any one of claims 1-5, characterized in that, include: Multi-field coupling modeling module: used to construct a three-field coupling model of thermo-mechanical-microstructure field, including temperature field modeling unit, stress field modeling unit and microstructure evolution unit, integrating plastic work heat source and dynamic recrystallization evolution equation; Parameter pre-screening module: Deploys Latin hypercube sampling and deep kernel learning models, and outputs the initial parameter set; Intelligent optimization module: Runs the DT-Guided EA algorithm to generate the Pareto optimal solution set; Mold generation module: Generates mold geometry based on parametric CAD templates and triggers local geometric compensation; Processing monitoring module: Connects an infrared thermal imager and a strain sensor to calculate the dynamic recrystallization volume fraction in real time; Closed-loop feedback module: The digital twin is updated through an LSTM network, and the federated learning framework enables cross-production line collaboration.

7. The system for implementing a multi-parameter, multi-field intelligent optimization method for large-scale free forging die in a press, as described in claim 6, is characterized in that... The system operates in a distributed architecture, including: High-performance simulation cluster: Employs MPI parallel computing for three-field coupled models, with load difference <5%; Edge computing node: Data is acquired at a frequency of 1kHz, noise is reduced by an FIR filter, and the response time is <2ms; Cloud platform: MongoDB sharded cluster database for storing thermal parameter mappings, and AES-256 encrypted federated learning model library.

Images