A Die Casting Process Parameter Optimization Method and System Based on Digital Twin

Abstract

The present invention relates to the field of die-casting optimization technology, and discloses a die-casting process parameter optimization method and system based on digital twin. The method includes: arranging sensors to collect die-casting equipment operation parameters and die-casting part quality data in real time to form multi-source die-casting production data. According to the multi-source die-casting production data, establish a multi-physics field simulation model and construct a digital twin model. Compare the virtual prediction results with the measured quality data, and construct a virtual-real difference compensation network to correct the parameters of the digital twin model. Based on the compensated digital twin model, perform a multi-objective reinforcement learning method to generate the optimal die-casting process parameters. Apply the optimal die-casting process parameters to the die-casting equipment for verification, and update the virtual-real difference compensation network according to the verification results. This application realizes the intelligent optimization and continuous self-evolution of die-casting process parameters, and improves the casting forming accuracy, energy efficiency utilization rate and production stability.

Description

Technical Field

[0001] This invention relates to the field of die casting optimization technology, and more specifically, to a method and system for optimizing die casting process parameters based on digital twins. Background Technology

[0002] Die casting is a key metal forming process that involves filling molten metal into a metal mold cavity under high pressure and high speed, and then rapidly solidifying it into shape. It is widely used in the manufacturing of automotive parts, aerospace components, and precision instrument housings. The filling, heat transfer, and solidification processes in die casting are extremely complex. Fluctuations in any process parameter (such as injection speed, pouring temperature, mold temperature, and holding time) directly affect the density, surface quality, and internal defect distribution of the casting. Therefore, precise setting and dynamic optimization of process parameters are the core issues in die casting process control.

[0003] Currently, the industry often determines process parameters through empirical methods or offline experiments. For example, the existing patent CN119066817A, "Optimization and Processing Method and System for Die Casting Parameters Based on Orthogonal Experiments," proposes using orthogonal experimental design to optimize die casting process parameters. By experimenting with combinations of parameters such as injection speed, mold temperature, and alloy temperature, a mapping relationship between parameters and quality indicators is established, and the process parameter configuration scheme is determined accordingly. This scheme achieves a certain degree of systematization based on traditional manual experimental optimization, but some problems still exist: For example, it cannot update process parameters in real time according to equipment operating status or material fluctuations, resulting in parameter lag and difficulty in coping with production batch differences. Parameter optimization is mainly based on experimental data statistics, without combining multi-physics simulation processes such as filling flow, heat transfer, and solidification, lacking a mechanistic understanding of the dynamic behavior inside die casting, making accurate prediction difficult. Furthermore, it cannot achieve multi-objective adaptive optimization, and has not established a dynamic mapping relationship between the virtual simulation model and the actual equipment, making the optimization results difficult to maintain effectiveness.

[0004] Therefore, it is necessary to design a method and system for optimizing die-casting process parameters based on digital twins to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a method and system for optimizing die casting process parameters based on digital twins, aiming to solve the problems of lack of real-time performance and dynamic adaptive capability, and the need to consider the multi-physics coupling effect in the die casting process in the existing technology.

[0006] In one aspect, this invention proposes a method for optimizing die-casting process parameters based on digital twins, comprising:

[0007] Temperature sensors, pressure sensors, and flow sensors are installed to collect real-time operating parameters of the die-casting equipment and quality data of the die-cast parts, forming multi-source die-casting production data.

[0008] Based on the multi-source die casting production data, a multi-physics simulation model coupling filling, heat transfer and solidification is established, and a digital twin model is constructed with the multi-physics simulation model as the core. The digital twin model is used to reflect the dynamic behavior of the actual die casting process.

[0009] The virtual prediction results output by the digital twin model are compared with the measured quality data. A virtual-real difference compensation network is constructed based on a multi-temporal Transformer neural network to correct the parameters of the digital twin model and obtain the compensated digital twin model.

[0010] Based on the compensated digital twin model, with casting defect rate, energy consumption and production cycle as multi-objective functions, process parameters are mapped to action space, quality indicators are mapped to reward functions, and a multi-objective reinforcement learning method based on Bayesian belief optimization algorithm is used to search for and iteratively train the combination of die casting process parameters, and to search for and optimize the die casting process parameters to generate the optimal die casting process parameters.

[0011] The optimal die-casting process parameters are applied to the die-casting equipment for verification, and the virtual-real difference compensation network is updated based on the verification results.

[0012] Furthermore, when collecting real-time operating parameters of die-casting equipment and quality data of die-cast parts to form multi-source die-casting production data, this includes:

[0013] Temperature sensors, pressure sensors, and flow sensors are respectively installed on the gating system, injection system, and mold temperature control circuit of the die casting equipment, and timestamps are synchronized.

[0014] The collected temperature, pressure, and flow signals are filtered for noise and feature extracted to generate operating parameter data that includes process characteristics, mold temperature distribution characteristics, and filling flow rate characteristics.

[0015] The quality data output by the die-casting quality inspection equipment is correlated with the operating parameter data in a time series to form the multi-source die-casting production data, which includes surface defects, internal shrinkage cavities, and density distribution.

[0016] Furthermore, when establishing a multiphysics simulation model coupling filling, heat transfer, and solidification based on the aforementioned multi-source die-casting production data, it includes:

[0017] Based on the multi-source die casting production data, boundary conditions are identified and material parameters are assimilated. The gate volume flow rate, cavity heat transfer coefficient and alloy thermal properties are taken as the quantities to be identified. The initial values ​​are obtained by the least squares inversion with Tikhonov regularization and updated recursively in the time series using extended Kalman filtering.

[0018] The incompressible Navier-Stokes equation and energy equation are solved separately and coupled with a phase change solidification model. The flow-heat-solidification simultaneous iteration is achieved through partitioned coupling and staggered time steps.

[0019] Mesh refinement and step size control are applied to the gate, thin-walled, and hot-spot regions. The convergence criterion is that the simulation-measured residual is less than a preset threshold, thus obtaining the multiphysics simulation model.

[0020] Furthermore, when constructing a digital twin model based on the aforementioned multiphysics simulation model, the following steps are included:

[0021] Based on the multiphysics simulation model, a mapping between state vectors and observation vectors is generated, a state observer is constructed, and real-time multi-source die-casting production data is input into the state observer to achieve virtual-real state synchronization.

[0022] The multiphysics simulation model is subjected to POD-Galerkin order reduction, and a Gaussian process surrogate model is added to the time-consuming local subprocesses to form the digital twin model.

[0023] Furthermore, when constructing a virtual-real difference compensation network based on a multi-temporal Transformer neural network, the following steps are included:

[0024] The virtual prediction results and measured quality data are time-stamp aligned and dimensionally normalized, and a multimodal time-series feature vector is constructed using temperature, pressure, flow rate and quality detection indicators.

[0025] Each modality is assigned an independent encoder and positional encoding and missing mask are added. Differential representations are extracted based on a shared temporal self-attention layer and a cross-modal attention fusion layer.

[0026] Output parameter correction vector, which corresponds to a subset of boundary condition parameters and material parameters of the digital twin model;

[0027] Uncertainty weights are calculated based on sensor confidence and quality detection confidence and used for weighting the difference representation to obtain the virtual-real difference compensation network.

[0028] Furthermore, when modifying the parameters of the digital twin model to obtain the compensated digital twin model, the following steps are included:

[0029] The virtual-real difference compensation network is trained with a combined loss consisting of residual regression loss and physical consistency regularization term, wherein the physical consistency regularization term constrains the corrected heat transfer coefficient, volumetric flow rate and solidification parameters to satisfy the stability region of the multiphysics equation.

[0030] The robust Huber loss is used to suppress outlier samples, and a sliding time window and early stopping strategy are used to control training stability.

[0031] The parameter correction vector is constrained within a preset feasible region by a projection operator, and the compensated digital twin model is obtained by incrementally updating it with a step size decay coefficient.

[0032] Parameter updates are triggered only if both the virtual / real residual threshold and the prediction confidence threshold are met; otherwise, the previously valid parameters are retained.

[0033] Furthermore, when performing a multi-objective reinforcement learning method based on the compensated digital twin model, it includes:

[0034] A state space is constructed using the state vector of the compensated digital twin model, and an action space is constructed by making the adjustable process parameters continuous. The adjustable process parameters include pouring temperature, injection speed, mold temperature, and holding time.

[0035] The defect rate of the casting, energy consumption and production cycle are used to form a reward vector, which is then processed to be dimensionless and directional consistent.

[0036] The reward vector is weighted using a Bayesian confidence optimization algorithm, and the weights are determined based on the confidence coefficients calculated from the mean and variance of the rewards.

[0037] Using the physical feasible region output by the digital twin model as the set of security constraints, penalties are applied to samples that violate the feasible region and then removed, thereby obtaining a constraint-based single-objective reward.

[0038] Furthermore, strategy search and iterative training are performed on the combination of die-casting process parameters. When generating the optimal die-casting process parameters, the following steps are taken:

[0039] The test sample with the highest potential reward is generated based on Bayesian confidence optimization in the candidate parameter set, and the reward value and constraint satisfaction of each sample are calculated in parallel in the compensated digital twin model.

[0040] During policy training, a proximal policy optimization method with convergence constraints is used to update the policy network, and an advantage estimation method is used to correct the value network.

[0041] The confidence coefficients used for multi-objective weight allocation are updated in the outer layer, and the policy network and value network are updated in the inner layer. When the improvement rate of the model is lower than the improvement rate threshold and the constraints are met, the training is considered to have converged, and the resulting combination of process parameters is determined as the optimal die-casting process parameters.

[0042] Furthermore, the optimal die-casting process parameters are applied to the die-casting equipment for verification. When updating the virtual-real difference compensation network based on the verification results, the following steps are included:

[0043] Based on the optimal die-casting process parameters, operational parameter data and quality data are collected according to the timestamp alignment rules consistent with the training phase to construct a verification dataset.

[0044] The compensated digital twin model generates corresponding virtual prediction results under the same working conditions, and calculates the virtual-real difference between the virtual prediction results and the measured quality data. When the virtual-real difference exceeds a preset difference threshold, the virtual-real difference compensation network is incrementally learned and updated. The incremental learning update uses the validation data within the sliding time window as training samples and adopts a hierarchical freezing strategy to update only the high-sensitivity parameter layer. After the update is completed, it is reviewed on an independent reserved subset. If the review passes, the updated virtual-real difference compensation network is synchronized to the digital twin model. If the review fails, the parameters of the previous version remain unchanged.

[0045] Compared with existing technologies, the advantages of this invention are as follows: By constructing a digital twin system with a multi-physics simulation model as its core, multi-source operational data such as temperature, pressure, and flow rate of the actual die-casting process are dynamically coupled with the virtual simulation model, realizing real-time digital mapping and self-learning optimization of the die-casting process. A virtual-real difference compensation network is established using a multi-temporal Transformer neural network to adaptively correct the parameters of the digital twin model, improving the consistency between virtual prediction and actual production. Based on the compensated high-precision twin model, a multi-objective reinforcement learning mechanism incorporating Bayesian belief optimization algorithm is introduced, enabling trade-off optimization among multiple indicators such as casting defect rate, energy consumption, and production cycle, and autonomously generating optimal die-casting process parameters. By applying the optimal parameters to equipment verification and dynamically updating the virtual-real compensation network based on the verification results, a closed-loop optimization system integrating virtual and real is formed, thereby achieving intelligent optimization and continuous self-evolution of die-casting process parameters, improving casting forming accuracy, energy efficiency, and production stability.

[0046] On the other hand, this application also provides a digital twin-based die-casting process parameter optimization system for applying the above-mentioned digital twin-based die-casting process parameter optimization method, including:

[0047] The sensor unit includes a temperature sensor, a pressure sensor, and a flow sensor. The sensor unit is used to collect real-time operating parameters of the die-casting equipment and quality data of the die-cast parts to form multi-source die-casting production data.

[0048] The acquisition unit is configured to establish a multi-physics simulation model of filling, heat transfer and solidification coupling based on the multi-source die casting production data, and to construct a digital twin model with the multi-physics simulation model as the core. The digital twin model is used to reflect the dynamic behavior of the actual die casting process.

[0049] The processing unit is configured to compare the virtual prediction results output by the digital twin model with the measured quality data, construct a virtual-real difference compensation network based on a multi-temporal Transformer neural network, correct the parameters of the digital twin model, and obtain a compensated digital twin model.

[0050] The optimization unit is configured to use the compensated digital twin model as a multi-objective function, with casting defect rate, energy consumption and production cycle as the multi-objective function, map process parameters to action space, map quality indicators to reward function, and perform strategy search and iterative training on die casting process parameter combinations based on a multi-objective reinforcement learning method that integrates Bayesian belief optimization algorithm, and perform iterative search and strategy optimization on die casting process parameters to generate optimal die casting process parameters.

[0051] The verification unit is configured to apply the optimal die-casting process parameters to the die-casting equipment for verification, and update the virtual-real difference compensation network based on the verification results.

[0052] It is understandable that the above-mentioned methods and systems for optimizing die-casting process parameters based on digital twins have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0053] 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:

[0054] Figure 1 A flowchart illustrating the die-casting process parameter optimization method based on digital twin provided in this embodiment of the invention;

[0055] Figure 2 A functional block diagram of a die-casting process parameter optimization system based on digital twins provided in an embodiment of the present invention. Detailed Implementation

[0056] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0057] In traditional die-casting process parameter optimization, offline testing methods rely on static mapping relationships of fixed parameter combinations, failing to perceive real-time equipment operating status and material property fluctuations. This results in process parameter adjustments lagging behind changes in actual production conditions. The lack of mechanistic models makes it difficult to accurately predict filling flow and solidification phase transformation processes, and the multiphysics coupling effects are not effectively quantified, leading to parameter optimization results deviating from real-world conditions. Furthermore, the mismatch between virtual simulation models and the dynamic behavior of actual production lines prevents the establishment of a continuously updated bidirectional mapping mechanism, causing the effectiveness of optimization strategies to gradually diminish in continuous production.

[0058] For example, in the die-casting of automotive transmission housings, batch variations in molten aluminum alloy lead to fluctuations in thermophysical parameters. Traditional methods, using orthogonal experiments to determine fixed injection speeds and mold temperature parameters, cannot adapt to real-time changes. During the casting process, temperature monitoring data in the mold's hot spots deviates continuously from simulation predictions, and the predicted flow rate at the end of the filling exceeds the allowable threshold, causing localized shrinkage cavities. Mold wear during the production cycle leads to changes in the cavity heat transfer coefficient, resulting in a decrease in casting yield and an increase in energy consumption in continuous production, despite the original combination of process parameters.

[0059] If the above problems are not addressed, the dynamic mismatch between process parameters and real-time operating conditions will lead to periodic fluctuations in scrap rates, resulting in continuous waste of raw materials and energy. Insufficient prediction accuracy of mechanistic models will mask the formation mechanisms of critical defects, causing process optimization to fall into a cycle of trial and error. The decoupling between virtual models and actual production lines hinders the closed-loop optimization capability of process parameters, making it difficult to meet the stringent requirements for process stability in high-precision die-casting parts, ultimately restricting the consistency and economy of mass production of complex structural parts.

[0060] For this, please refer to Figure 1 As shown, this application proposes a method for optimizing die-casting process parameters based on digital twins, including:

[0061] S100: Equipped with temperature, pressure and flow sensors, it collects real-time operating parameters of the die-casting equipment and quality data of the die-cast parts, forming multi-source die-casting production data.

[0062] S200: Based on multi-source die casting production data, a multi-physics simulation model coupling filling, heat transfer and solidification is established, and a digital twin model is constructed with the multi-physics simulation model as the core. The digital twin model is used to reflect the dynamic behavior of the actual die casting process.

[0063] S300: The virtual prediction results output by the digital twin model are compared with the measured quality data. A virtual-real difference compensation network is constructed based on a multi-temporal Transformer neural network to correct the parameters of the digital twin model and obtain the compensated digital twin model.

[0064] S400: Based on the compensated digital twin model, it uses casting defect rate, energy consumption and production cycle as multi-objective functions, maps process parameters to action space, and quality indicators to reward functions. Based on a multi-objective reinforcement learning method that integrates Bayesian belief optimization algorithm, it performs strategy search and iterative training on die casting process parameter combinations, and iterative search and strategy optimization on die casting process parameters to generate optimal die casting process parameters.

[0065] S500: The optimal die-casting process parameters are applied to the die-casting equipment for verification, and the virtual-real difference compensation network is updated based on the verification results.

[0066] Specifically, multi-source die-casting production data refers to the operating parameters of the die-casting equipment and the quality data of the die-cast parts collected by temperature sensors, pressure sensors, and flow sensors. This can be achieved by arranging sensors in the gating system, injection system, and mold temperature control loop, synchronizing them with timestamps, and then correlating them with quality inspection data after noise filtering and feature extraction. This comprehensively reflects the dynamic characteristics of the die-casting process. The multiphysics simulation model refers to a numerical model coupling the filling, heat transfer, and solidification processes. This can be achieved by solving the incompressible Navier-Stokes equations, energy equations, and phase change solidification models, combined with boundary condition identification and material parameter assimilation, to simulate the physical behavior of the die-casting process. The digital twin model is a virtual mapping model built around the multiphysics simulation model. This can be achieved by synchronizing virtual and real states using a state observer and accelerating calculations using POD-Galerkin order reduction and Gaussian process proxy models, to reflect the actual dynamic behavior of die casting in real time. The virtual-real difference compensation network refers to a correction module built based on a multi-temporal Transformer neural network. Specifically, it can be implemented by encoding multi-modal temporal feature vectors, extracting difference representations through shared self-attention layers and cross-modal fusion layers, and combining this with uncertainty weight calculation to reduce prediction errors in the digital twin model. The multi-objective reinforcement learning method refers to a strategy search algorithm with casting defect rate, energy consumption, and production cycle as optimization objectives. Specifically, it can be implemented by mapping process parameters to an action space, quality indicators to a reward function, and combining this with a Bayesian confidence optimization algorithm to determine weight allocation, thus balancing multi-objective optimization needs. The process parameter verification and update refers to applying optimal parameters to the equipment and then using feedback data to correct the model. Specifically, it can be implemented by collecting verification data with timestamp alignment, calculating the virtual-real difference threshold to trigger incremental learning updates, and adjusting network parameters using a hierarchical freezing strategy to maintain the accuracy of the digital twin model.

[0067] This application constructs a dynamic optimization framework that combines multiphysics simulation, temporal difference compensation, and multi-objective reinforcement learning to achieve closed-loop optimization of die-casting process parameters. This framework drives real-time correction of the digital twin model using sensor data, efficiently searches for optimal parameters in a virtual environment using reinforcement learning, and continuously updates the model based on actual verification results. This solves the problems of traditional methods being unable to dynamically respond to production fluctuations and lacking multi-objective trade-offs.

[0068] The working process and principle of this application are as follows: First, temperature sensors, pressure sensors, and flow sensors are installed on the die-casting equipment to collect real-time operating parameters of the die-casting equipment and quality data of the die-cast parts, forming multi-source die-casting production data. These sensors are used to monitor changes in temperature, pressure, and flow rate during the die-casting process.

[0069] Based on the collected multi-source die-casting production data, a multiphysics simulation model coupling filling, heat transfer, and solidification was established. This model comprehensively considers the complex physical processes such as fluid flow, heat conduction, and phase change in the die-casting process, enabling a more accurate simulation of the actual die-casting process. Using this multiphysics simulation model as the core, a digital twin model was constructed to reflect the dynamic behavior of the actual die-casting process.

[0070] By comparing the virtual predictions output by the digital twin model with the measured quality data, discrepancies were found. To compensate for these discrepancies, a virtual-to-real discrepancy compensation network was constructed based on a multi-temporal Transformer neural network. This network can learn and compensate for the differences between the virtual model and the actual process, thereby correcting the parameters of the digital twin model and obtaining a more accurate digital twin model after compensation.

[0071] Based on the compensated digital twin model, with casting defect rate, energy consumption, and production cycle as multi-objective functions, process parameters are mapped to the action space, and quality indicators are mapped to the reward function. A multi-objective reinforcement learning method integrating Bayesian belief optimization algorithm is employed to search for and iteratively train the combination of die-casting process parameters. This process continuously tries different parameter combinations, evaluates their effects, and optimizes them based on feedback, ultimately generating the optimal die-casting process parameters.

[0072] Finally, the generated optimal die-casting process parameters are applied to actual die-casting equipment for verification. Based on the verification results, the virtual-to-real difference compensation network is updated again, forming a closed-loop optimization process. This method can continuously improve the model accuracy and optimization effect, adapting to dynamic changes in actual production.

[0073] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0074] Temperature sensors, pressure sensors, and flow sensors are installed on the gating system, injection system, and mold temperature control circuit of the die-casting equipment, respectively. These sensors are connected to the central processing unit via a data acquisition system to collect temperature, pressure, and flow data in real time during the die-casting process. Simultaneously, quality inspection equipment is used to detect surface defects, internal shrinkage cavities, and density distribution in the die-cast parts.

[0075] The collected data undergoes preprocessing, including noise filtering and feature extraction, to generate operational parameter data containing process characteristics, mold temperature distribution characteristics, and filling flow rate characteristics. This data is then correlated with quality inspection data over time to form a multi-source die-casting production dataset.

[0076] A multiphysics simulation model coupling filling, heat transfer, and solidification was established using the finite element method and computational fluid dynamics principles. This model considers physical processes such as molten metal flow, heat conduction, and the release of latent heat of phase change. A simultaneous iterative solution for the flow-heat-solidification process was achieved through partitioned coupling and staggered time-stepping methods.

[0077] A digital twin model is constructed based on a multiphysics simulation model. The POD-Galerkin method is used to reduce the order of the model, and a Gaussian process surrogate model is introduced for computationally time-consuming local processes to improve computational efficiency.

[0078] A multi-temporal Transformer neural network is constructed as a virtual-real difference compensation network. This network contains multiple independent encoders to process temporal data from different modalities. Discrete features are extracted through a self-attention mechanism and a cross-modal attention fusion layer, outputting a parameter correction vector.

[0079] A multi-objective reinforcement learning algorithm incorporating Bayesian belief optimization is employed to optimize process parameters. Process parameters such as pouring temperature, injection speed, mold temperature, and holding time are used as the action space, while a reward function is constructed based on casting defect rate, energy consumption, and production cycle. Through iterative training, the policy network and value network are continuously updated to ultimately generate the optimal combination of die-casting process parameters.

[0080] The optimized process parameters were applied to an actual die-casting production line for verification. Based on the verification results, an incremental learning method was used to update the virtual-real difference compensation network, further improving the prediction accuracy of the digital twin model.

[0081] Through the above scheme, this application achieves real-time dynamic optimization of die-casting process parameters. The digital twin model accurately reflects the dynamic behavior of the actual die-casting process, and the virtual-to-real difference compensation network effectively improves the model's prediction accuracy. A multi-objective reinforcement learning method achieves comprehensive optimization of casting defect rate, energy consumption, and production cycle. This method overcomes the parameter lag problem of traditional offline experimental methods, enabling real-time adjustment of process parameters based on equipment operating status and material fluctuations. Simultaneously, by integrating multiphysics simulation and data-driven methods, it improves the mechanistic understanding of the internal dynamic behavior of die casting, achieving more accurate prediction and optimization. A closed-loop verification and update mechanism ensures the continuous effectiveness of the optimization results in continuous production, improving the production consistency of complex die-cast parts.

[0082] This application further proposes arranging temperature sensors, pressure sensors, and flow sensors on the gating system, injection system, and mold temperature control circuit of the die-casting equipment, and synchronizing them with timestamps. The collected temperature, pressure, and flow signals are filtered for noise and feature extracted to generate operating parameter data containing process characteristics, mold temperature distribution characteristics, and filling flow rate characteristics. The quality data output from the die-casting quality inspection equipment is then correlated with the operating parameter data in a time series to form multi-source die-casting production data. The quality data includes surface defects, internal shrinkage cavities, and density distribution.

[0083] Specifically, sensors are placed at key locations in the gating system, injection system, and mold temperature control circuit. Timestamp synchronization employs either hardware triggering or software interpolation alignment. Noise filtering combines wavelet thresholding with moving average filtering, and feature extraction utilizes time-frequency analysis and principal component analysis to extract process features. Time series correlation is based on a dynamic time warping algorithm to match operating parameters with quality data.

[0084] Specifically, temperature sensors are installed at the sprue sleeve and runner cone of the gating system, pressure sensors are integrated at the front end of the piston rod of the injection cylinder, and flow sensors are embedded in the oil circuit of the injection punch. Timestamp synchronization errors are controlled within ±5ms to ensure time consistency of multi-source data. During noise filtering, wavelet decomposition is performed at 5 levels, using the sym4 wavelet basis function, and soft thresholding is applied to high-frequency coefficients. The moving average window length is 10 sampling points to filter out high-frequency noise. For feature extraction, short-time Fourier transform is performed on the temperature signal to extract mold temperature distribution features, and empirical mode decomposition is performed on the pressure signal to obtain filling velocity features. Operating parameter data and quality data are aligned using a dynamic time warping algorithm. A time-series mapping relationship is established using the mold closure signal as a reference point, forming a multidimensional dataset containing process parameters, process characteristics, and quality indicators.

[0085] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0086] Temperature sensors, pressure sensors, and flow sensors are installed in the gating system, injection system, and mold temperature control circuit of the die-casting equipment, and timestamped synchronously. Specifically, thermocouple temperature sensors are installed at the barrel, sprue, and gate of the gating system; pressure sensors are installed at the hydraulic cylinder and injection chamber of the injection system; and turbine flow sensors are installed on the inlet and outlet water pipes of the mold temperature control circuit. These sensors are synchronously collected through a data acquisition system with a sampling frequency set to 100Hz, and a millisecond-level timestamp is added to each data point.

[0087] The collected temperature, pressure, and flow signals undergo noise filtering and feature extraction to generate operational parameter data that includes process characteristics, mold temperature distribution characteristics, and filling flow rate characteristics. Furthermore, wavelet denoising algorithms are applied to the original signals to remove high-frequency noise, and then time-domain and frequency-domain features are extracted. For example, features such as average value, maximum value, and rise time are extracted from the temperature signal; peak value, duration, and rise rate are extracted from the pressure signal; and average flow rate and pulsation amplitude are extracted from the flow signal.

[0088] The quality data output from the die-casting quality inspection equipment was correlated with the operating parameter data in a time series to form multi-source die-casting production data. This quality data included surface defects, internal shrinkage cavities, and density distribution. X-ray inspection was then used to perform non-destructive testing on the die-castings to obtain images of the internal shrinkage cavity distribution. A 3D scanner was used to acquire the surface contour data of the die-castings, and surface defects were analyzed by comparing it with the CAD model. The Archimedes method was used to measure the density distribution of the die-castings. These quality data were then time-aligned with the corresponding operating parameter data to construct a multi-source dataset containing process parameters, process characteristics, and quality indicators.

[0089] Through the above technical solution, this application achieves high-precision acquisition and correlation analysis of multi-source heterogeneous data in the die-casting process. By deploying sensors at key locations and synchronizing them in time, comprehensive process parameters and characteristics are obtained. Noise reduction and feature extraction of the raw signals improve data quality. A process-quality mapping relationship is established by temporally correlating operating parameters with quality inspection data. This lays a data foundation for subsequent digital twin model construction and parameter optimization, contributing to a deeper understanding of the physical mechanisms in the die-casting process and improving the accuracy and real-time performance of process parameter optimization.

[0090] This application further proposes a method for boundary condition identification and material parameter assimilation based on multi-source die-casting production data. The gate volumetric flow rate, cavity heat transfer coefficient, and alloy thermal properties are used as the variables to be identified. Initial values ​​are obtained using Tikhonov-regularized least-squares inversion, and updated recursively over time using extended Kalman filtering. The incompressible Navier-Stokes equations and energy equations are solved separately and coupled with a phase change solidification model. Iterative flow-heat-solidification processes are achieved through partitioned coupling and staggered time steps. Mesh refinement and step size control are applied to the gate, thin-walled, and hot-spot regions. A multiphysics simulation model is obtained with the simulation-measured residual less than a preset threshold as the convergence criterion.

[0091] In this process, Tikhonov-regularized least-squares inversion addresses the ill-posedness of parameter inversion by introducing a regularization term to suppress noise interference and ensure the stability of the initial parameter estimates. Extended Kalman filtering, through a recursive update mechanism, dynamically integrates real-time multi-source data into the parameter estimation process, enabling online correction of boundary conditions and material parameters. The flow-heat-solidification simultaneous iteration employs a partitioned coupling strategy, decomposing the solution processes of the Navier-Stokes equations and energy equations into independent subdomains and reducing computational complexity through staggered stepping. Local mesh refinement is used in the gating region, with dynamic step size control balancing computational accuracy and efficiency. The convergence criterion with a preset threshold is dynamically adjusted based on the residuals of measured data to ensure consistency between the simulation model and the physical process.

[0092] Specifically, in constructing the multiphysics simulation model, the real-time operating parameters of the gating system, injection system, and mold are first extracted from multi-source die-casting production data. Initial estimates are obtained by inverting the gate volumetric flow rate, cavity heat transfer coefficient, and alloy thermophysical parameters using the Tikhonov regularization method. Subsequently, the extended Kalman filter algorithm is used to recursively update the initial parameters over time, dynamically correcting boundary conditions and material parameters by utilizing the temporal correlation between sensor data and quality inspection results. In the physics solution stage, the flow equation and heat transfer equation are decomposed into three subdomains: the gate region, the cavity region, and the mold region. Multiphysics coupling calculations are achieved by alternately solving the equations of each subdomain and passing boundary conditions. For thin-walled and hot-spot regions, adaptive mesh refinement technology is used to improve local resolution, while the time step is dynamically adjusted according to the residual change rate. When the root mean square value of the residual between the simulation results and the measured data is lower than a preset threshold three times consecutively, the model is considered converged, and the final parameters are output. The above methods effectively solve the problems of parameter drift and convergence difficulties in traditional simulation models, providing a high-precision physical field prediction foundation for digital twin models.

[0093] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0094] Boundary condition identification and material parameter assimilation are performed based on multi-source die-casting production data. The gating volumetric flow rate, cavity heat transfer coefficient, and alloy thermal properties are used as the parameters to be identified. Initial values ​​are obtained using Tikhonov regularized least squares inversion, and then updated recursively over time using an extended Kalman filter. Specifically, the Tikhonov regularization method is first used to preliminarily estimate the gating volumetric flow rate, cavity heat transfer coefficient, and alloy thermal properties. Then, the extended Kalman filter algorithm dynamically updates these parameters over time.

[0095] The incompressible Navier-Stokes equations and energy equations are solved separately and coupled with a phase change solidification model. Iterative solutions to the flow, heat, and solidification processes are achieved through partitioned coupling and staggered time steps. Furthermore, the Navier-Stokes equations and energy equations are discretized using the finite volume method, and the phase change solidification model is coupled with the flow and heat transfer equations. The computational domain is divided into flow, heat transfer, and solidification regions using a partitioned coupling strategy, and a staggered time step method is employed to solve these three physical fields simultaneously.

[0096] Mesh refinement and step size control were implemented for the gate, thin-walled, and hot-point regions. The convergence criterion was that the simulation-to-measured residual was less than a preset threshold, resulting in a multiphysics simulation model. Adaptive mesh refinement technology was employed for key regions such as the gate, thin-walled, and hot-point regions to improve computational accuracy. Simultaneously, the computational step size was adjusted according to the physical characteristics of different regions to ensure numerical stability. Finally, by comparing the residuals between the simulation results and measured data, the multiphysics simulation model was considered converged when the residual was less than a preset threshold.

[0097] Through the above technical solutions, this application achieves boundary condition and material parameter identification based on multi-source die-casting production data, improving the accuracy of the simulation model. By coupling the solutions for flow, heat transfer, and solidification processes, the complex physical phenomena in the die-casting process are accurately captured. Mesh refinement and step size control strategies for key areas improve the accuracy and computational efficiency of the simulation results. Therefore, this application constructs a high-precision, high-efficiency multiphysics simulation model for die casting, laying the foundation for subsequent digital twin models and process parameter optimization.

[0098] This application further proposes a method to generate a mapping between state vectors and observation vectors based on a multiphysics simulation model, construct a state observer, and input real-time acquired multi-source die-casting production data into the state observer to achieve synchronization between virtual and real states. The multiphysics simulation model is reduced to the order of POD-Galerkin and Gaussian process surrogate models are added to time-consuming local subprocesses to form a digital twin model.

[0099] The state observer maps state vectors to observation vectors, inputting real-time collected temperature, pressure, and flow data into the model to dynamically adjust the initial conditions and boundary parameters of the simulation model, synchronizing the virtual simulation results with actual production data in the time dimension. The POD-Galerkin order reduction method extracts the principal modes of the multiphysics simulation model, projecting high-dimensional partial differential equations into a low-dimensional space, reducing computational degrees of freedom; for example, reducing the degrees of freedom of the original model from millions to hundreds. For time-consuming subprocesses such as gating flow and thin-walled heat transfer, a Gaussian process surrogate model is used to replace the local calculations of the original model. The surrogate model is trained using historical data to predict changes in local physical quantities; for example, replacing the iterative solution required for local flow processes with regression prediction based on input parameters. The combination of the order reduction model and the surrogate model reduces the overall computation time to less than 10% of the original model, while ensuring that the computational accuracy error in key areas is less than 5%.

[0100] Specifically, the state observer dynamically corrects model parameters through feedback from real-time data and simulation output residuals. For example, it uses the difference between the mold temperature distribution characteristics and simulation predictions as a correction signal to adjust the cavity heat transfer coefficient. During the order reduction process, transient solution sets from multiphysics simulations under different operating conditions are collected. Dominant modes are extracted using eigenorthogonal decomposition to construct reduced-order basis functions, and the original equations are projected into a lower-dimensional space. For example, the temperature field in the energy equation is represented as a linear combination of the first 50 modes. For locally time-consuming subprocesses, such as turbulence calculations in the gating region, a Gaussian process surrogate model is used to establish a mapping relationship between input parameters and output flow rate. Input parameters include injection velocity and alloy viscosity, and the output is the gating volumetric flow rate. The surrogate model fits nonlinear relationships using kernel functions and updates hyperparameters based on Bayesian optimization, reducing local computation time from seconds to milliseconds. Thus, the reduced-order model and the surrogate model work together to improve computational efficiency while preserving the multi-physics coupling characteristics, enabling the digital twin model to achieve virtual-real synchronization with a response speed of minutes. At the same time, the temperature prediction error in key areas is controlled within ±3℃, meeting the requirements for real-time optimization.

[0101] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0102] Based on a multiphysics simulation model, a mapping between state vectors and observation vectors is generated. A state observer is constructed, and real-time multi-source die-casting production data is input into the state observer to achieve synchronization between virtual and real states. Specifically, the multiphysics simulation model is first expressed in state space, with internal state variables such as temperature, velocity, and pressure fields forming the state vector, and observable physical quantities such as cavity pressure and mold surface temperature forming the observation vector. Then, a Kalman filter is designed as the state observer, using real-time sensor data as the observation input, and state estimation and synchronization are achieved through prediction-correction iteration.

[0103] Furthermore, a POD-Galerkin reduction model is applied to the multiphysics simulation model, and Gaussian process surrogate models are added to time-consuming local subprocesses to form a digital twin model. The POD-Galerkin reduction is achieved by extracting the dominant modes and projecting them onto a low-dimensional subspace. For example, the velocity field of the filling process is decomposed using POD, retaining the first k modes with 95% energy, and the Navier-Stokes equations are projected onto the subspace spanned by these k basis functions. For locally highly nonlinear regions such as solidification phase transition processes, Gaussian process regression is used to establish a surrogate model for the input-output mapping. This results in a computationally efficient and accuracy-controllable digital twin model.

[0104] Through the above technical solutions, this application achieves efficient and accurate digital twin model construction. The introduction of POD-Galerkin order reduction and Gaussian process surrogate models improves computational efficiency, enabling the digital twin model to respond in real time. Simultaneously, the design of the state observer ensures the synchronization between the virtual model and the actual system, enhancing the model's predictive ability. This lays the foundation for subsequent optimization of die-casting process parameters based on digital twins, helping to improve the accuracy and efficiency of optimization.

[0105] This application further proposes a virtual-real difference compensation network based on a multi-temporal Transformer neural network, including: aligning virtual prediction results with measured quality data using timestamps and normalizing dimensions; constructing multimodal temporal feature vectors using temperature, pressure, flow rate, and quality inspection indicators; setting independent encoders for each modality and adding positional encoding and missing masking; and extracting difference representations based on a shared temporal self-attention layer and a cross-modal attention fusion layer. The output parameter correction vector corresponds to a subset of boundary condition parameters and material parameters of the digital twin model. Uncertainty weights are calculated based on sensor confidence and quality inspection confidence and used for weighting the difference representation.

[0106] In this process, timestamp alignment employs a sliding window matching algorithm to eliminate transmission delay, and dimensional normalization uses Z-score standardization to eliminate magnitude differences. The multimodal temporal feature vector includes the mean temperature signal, peak pressure signal, and integral value of the flow rate signal. Quality inspection indicators include the proportion of surface defect area and the number of shrinkage cavities. Independent encoders use one-dimensional convolutional layers to extract local features of each mode, and position encoding embeds time-series sequence information using a sine function. A shared temporal self-attention layer calculates cross-time-step dependencies, and a cross-modal attention fusion layer aggregates the associated features between different modes through a multi-head mechanism. Parameter correction vectors are mapped to the gate volumetric flow rate, cavity heat transfer coefficient, and alloy thermophysical parameters of the digital twin model. Uncertainty weights are calculated based on the sensor calibration error covariance matrix and the repeatability test results of the quality inspection equipment.

[0107] Specifically, during the timestamp alignment stage, a dynamic time warping algorithm is used to match the temporal phase offset between virtual predictions and measured data, ensuring that data within the same time window corresponds to the same physical process. Dimensional normalization converts temperature signals into Celsius offsets, pressure signals into megapascals per unit, and flow signals into liters per minute. When constructing multimodal temporal feature vectors, temperature modal features include temperature gradients in different regions of the mold, pressure modal features include the pressure rise rate of the injection cylinder, and flow modal features include the instantaneous flow fluctuation amplitude of the gating system. An independent encoder performs a one-dimensional convolution operation with a kernel of 5 on each modal data, extracts local temporal patterns, and retains temporal order information through positional encoding. A shared temporal self-attention layer calculates the correlation weights between different time steps to identify temporal difference patterns between virtual and measured data. A cross-modal attention fusion layer performs cross-attention calculations on the hidden state vectors of temperature, pressure, and flow modes to capture the coupling relationships between multi-source data. The parameter correction vector maps the difference representation to the adjustable parameter space of the digital twin model through a fully connected layer. The correction amount for boundary condition parameters is limited to ±10%, and the correction amount for material parameters is limited to ±5%. The uncertainty weight is calculated using Bayesian inference based on the ±0.5℃ error range of the temperature sensor and the ±2% repeatability error of the quality inspection equipment, to suppress the interference of low-confidence data on parameter correction.

[0108] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0109] The virtual prediction results and measured quality data are time-stamp aligned and dimensionally normalized, and a multimodal temporal feature vector is constructed using temperature, pressure, flow rate, and quality inspection indicators. An independent encoder is set up for each modality, and positional encoding and missing data masks are added. Difference representations are extracted based on a shared temporal self-attention layer and a cross-modal attention fusion layer. A parameter correction vector is output, corresponding to a subset of boundary condition parameters and material parameters of the digital twin model. Uncertainty weights are calculated based on sensor confidence and quality inspection confidence and used for weighting the difference representations, thus obtaining a virtual-real difference compensation network.

[0110] Specifically, the virtual prediction results are first aligned with the measured quality data according to timestamps, and the data of different dimensions are normalized. Then, a multimodal temporal feature vector containing temperature, pressure, flow rate, and quality inspection indicators is constructed. An independent encoder is set up for each modality of data, and positional encoding information and missing data masks are added during the encoding process. Next, a shared temporal self-attention layer is used to extract features from each modality of data, and then a cross-modal attention fusion layer is used to fuse the features from different modalities, thereby extracting the differences between the virtual and real data.

[0111] Furthermore, based on the extracted difference representations, a parameter correction vector is output to refine the digital twin model. This vector corresponds to a subset of boundary condition parameters and material parameters in the digital twin model. This allows for targeted adjustments to key parameters within the model.

[0112] Finally, considering the potential differences in reliability between different sensors and quality inspection methods, uncertainty weights are introduced. By calculating the confidence levels of the sensors and quality inspection, the difference representation is weighted. This highlights the influence of highly reliable data sources and reduces the impact of data with greater uncertainty, thereby obtaining a more accurate virtual-real difference compensation network.

[0113] Through the above technical solution, this application achieves accurate comparison and difference extraction between virtual prediction results and measured quality data. By employing multimodal data fusion and attention mechanisms, complementary information from different types of data can be fully utilized to improve the accuracy of difference representation. Introducing uncertainty weights can effectively handle reliability differences in data sources and improve the robustness of the compensation network. Therefore, this solution can more accurately capture the differences between the virtual model and actual production, providing a reliable basis for subsequent model parameter correction, thereby improving the simulation accuracy of the digital twin model for the actual die-casting process.

[0114] This application further proposes a combined loss method consisting of residual regression loss and physical consistency regularization term to train a virtual-real difference compensation network. The physical consistency regularization term constrains the corrected heat transfer coefficient, volumetric flow rate, and solidification parameters to satisfy the stability region of the multiphysics equations. Robust Huber loss is used to suppress outlier samples, and a sliding time window and early stopping strategy are employed to control training stability. The parameter correction vector is constrained within a preset feasible region using a projection operator, and incremental updates are performed using a step-size decay coefficient to obtain the compensated digital twin model. Parameter updates are triggered only when both the virtual-real residual threshold and the prediction confidence threshold are simultaneously satisfied; otherwise, the previously valid parameters are retained.

[0115] The combined loss includes residual regression loss and physical consistency regularization. Residual regression loss minimizes the difference between virtual predictions and measured data, while physical consistency regularization introduces stability domain constraints from multiphysics equations to ensure that corrected parameters such as heat transfer coefficient and volumetric flow rate conform to fundamental laws of fluid mechanics and thermodynamics. Robust Huber loss reduces parameter fluctuations during training by decreasing gradient contributions from outliers. A sliding time window uses a fixed-length historical data sequence for training, and an early stopping strategy terminates training early by monitoring changes in the validation set loss to avoid overfitting. The projection operator maps the parameter correction vector to a preset feasible region, such as ensuring the heat transfer coefficient is not lower than the lower limit of material thermal conductivity and the volumetric flow rate does not exceed the maximum flow rate of the equipment. The step size decay coefficient dynamically adjusts the parameter update magnitude based on the number of iterations, using a larger step size to accelerate convergence in the initial stage and a smaller step size for fine-tuning in the later stages. The joint triggering condition for the virtual / real residual threshold and the prediction confidence threshold is calculated by the average residual of the current batch of data and the model prediction confidence; parameter updates are only performed when both are below the set threshold.

[0116] Specifically, in the training process of the virtual-real difference compensation network, the residual regression loss is first combined with the physical consistency regularization term, and the network parameters are optimized through backpropagation. The residual regression loss uses the mean squared error to calculate the difference between the prediction and the measurement, while the physical consistency regularization term calculates the impact of parameter correction on model stability through the Jacobian matrix of the multiphysics equations, ensuring that the corrected parameters do not cause the simulation model to diverge. The robust Huber loss uses a linear loss function when the residual exceeds a set threshold to suppress the influence of outliers on the gradient. The sliding time window uses historical data of a fixed length as training samples, for example, selecting data from the most recent 50 time steps. The early stopping strategy terminates training when the validation set loss does not decrease for 5 consecutive cycles. After the parameter correction vector is processed by the projection operator, for example, limiting the heat transfer coefficient to the range of 100-500 W / (m·K), it is incrementally updated using a step size decay coefficient, with the step size decreasing exponentially from 0.1 to 0.01. The threshold for virtual and real residuals is set to 1.5 times the historical average residual, and the prediction confidence threshold is set to 0.9. When the real-time calculated residual is lower than the threshold and the confidence level is higher than the threshold, parameter updates are triggered. Otherwise, the previously valid parameters are retained to avoid invalid or erroneous corrections. These measures work together to ensure the stability, physical rationality, and update reliability of the parameter correction process, thereby improving the dynamic tracking accuracy of the digital twin model.

[0117] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0118] A virtual-to-real difference compensation network is trained using a combined loss consisting of residual regression loss and physical consistency regularization term. The physical consistency regularization term constrains the corrected heat transfer coefficient, volumetric flow rate, and solidification parameters to satisfy the stability region of the multiphysics equations. Specifically, the residual regression loss uses the mean square error function to calculate the difference between the virtual prediction results and the measured mass data, while the physical consistency regularization term introduces the stability constraint of the multiphysics equations into the loss function using the Lagrange multiplier method.

[0119] Furthermore, robust Huber loss is used to suppress outliers, and a sliding time window and early stopping strategy are employed to control training stability. The robust Huber loss function uses squared loss for small errors and linear loss for large errors, thereby reducing the impact of outliers on model training. The sliding time window is set to 24 hours, and the early stopping strategy halts training if there is no improvement for five consecutive epochs on the validation set.

[0120] Therefore, the parameter correction vector is constrained within a preset feasible region by a projection operator, and the compensated digital twin model is obtained by incrementally updating it with a step size decay coefficient. For example, the projection operator uses gradient projection to project the correction vector onto the convex set defined by the physical constraints, and the initial value of the step size decay coefficient is set to 0.1, which decays by 10% every 50 epochs.

[0121] Parameter updates are triggered only when both the virtual / real residual threshold and the prediction confidence threshold are met; otherwise, the previously valid parameters are retained. As a preferred implementation, the virtual / real residual threshold is set to 0.05, and the prediction confidence threshold is set to 0.9; parameter updates are only performed when both thresholds are met simultaneously.

[0122] Through the above technical solutions, this application achieves adaptive correction of digital twin model parameters. By designing a combined loss function, both the model's fitting accuracy to the measured data and the physical rationality of the corrected parameters are ensured. The introduction of a robust loss function and training strategy improves the model's robustness to outliers and noise. The parameter update mechanism design ensures model stability while enabling rapid response to fluctuations in process parameters. Therefore, the digital twin model can more accurately reflect the dynamic behavior of the actual die-casting process, laying the foundation for subsequent process parameter optimization.

[0123] This application further proposes constructing a state space using the state vector of a compensated digital twin model, and constructing an action space by making adjustable process parameters continuous. These adjustable process parameters include casting temperature, injection speed, mold temperature, and holding time. A reward vector is formed using casting defect rate, energy consumption, and production cycle, and then processed to be dimensionless and directionally consistent. A Bayesian confidence optimization algorithm is used to weight the reward vector, with the weights determined based on confidence coefficients calculated from the reward mean and variance. The physical feasible region output by the digital twin model is used as the set of safety constraints. Samples that violate the feasible region are penalized and removed, thereby obtaining a constraint-based single-objective reward.

[0124] The state space consists of a multi-dimensional vector composed of the temperature field, flow velocity field, and solidification phase transformation rate output by the digital twin model. The action space is normalized to map the casting temperature to the range of [650℃, 720℃] and the injection velocity to the range of [2.5m / s, 5.0m / s]. In the reward vector, the casting defect rate is expressed as a percentage, energy consumption is in kilowatt-hours, and production cycle is in seconds, all unified to the range of [0,1] after range standardization to eliminate dimensional differences. The Bayesian confidence optimization algorithm dynamically adjusts the weight allocation by calculating the mean and variance of the historical rewards for each objective; for example, it reduces the weight coefficient when the defect rate variance exceeds 0.15. The safety constraint set includes the mold temperature gradient threshold, the upper limit of the filling pressure, and the lower limit of the solidification time. If the combination of sampling parameters causes the temperature gradient to exceed 8℃ / mm, a constraint penalty is triggered.

[0125] Specifically, during the injection stage, the digital twin model outputs the temperature and flow rate distribution inside the cavity in real time, forming a state vector containing 128 dimensions. The action space is discretized with the casting temperature in 10℃ increments and the injection speed in 0.2m / s increments, resulting in 72 candidate parameter combinations. The initial weights for the defect rate, energy consumption, and production cycle in the reward vector are set to 0.6, 0.3, and 0.1, respectively. Bayesian confidence optimization adjusts these weights to 0.55, 0.35, and 0.10 based on the defect rate standard deviation reaching 0.12 in the first 50 iterations. The safety constraint set is pre-calculated using the digital twin model; for example, if the holding time is less than 3.2 seconds, an incomplete solidification warning is triggered, and that parameter combination is directly eliminated. Through a constrained single-objective reward mechanism, the optimal parameter combination of 680℃ casting temperature, 4.2m / s injection speed, 220℃ mold temperature, and 4.8 seconds holding time was obtained after 100 iterations. The defect rate was reduced to 1.2%, energy consumption was controlled at 18.3kWh, and the production cycle was shortened to 86 seconds.

[0126] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0127] When using a multi-objective reinforcement learning method based on a compensated digital twin model, the following steps are included:

[0128] First, a state space is constructed using the state vectors of the compensated digital twin model. Then, an action space is constructed by continuously converting the adjustable process parameters. These adjustable process parameters include casting temperature, injection speed, mold temperature, and holding time. For example, the casting temperature range is set to 600-750℃, the injection speed range is 0.1–5 m / s, the mold temperature range is 150-300℃, and the holding time range is 0.01–1 s.

[0129] Secondly, a reward vector is formed using casting defect rate, energy consumption, and production cycle, and then processed to be dimensionless and aligned in direction. Specifically, the defect rate, energy consumption, and production cycle are normalized to the 0-1 interval, and the defect rate and energy consumption are negative to align their directions.

[0130] Furthermore, the reward vector is weighted using a Bayesian confidence optimization algorithm, with the weights determined based on the confidence coefficients calculated from the reward mean and variance.

[0131] Finally, using the physically feasible region output by the digital twin model as the set of safety constraints, penalties are applied to samples that violate the feasible region and then eliminated, thereby obtaining a constrained single-objective reward. Thus, the multi-objective optimization problem is transformed into a single-objective optimization problem while ensuring the physical feasibility of the solution.

[0132] Through the above technical solution, this application realizes the application of multi-objective reinforcement learning methods in the optimization of die-casting process parameters. This allows for the simultaneous consideration of multiple objectives such as casting quality, energy consumption, and production efficiency, avoiding the limitations that may arise from single-objective optimization. Using Bayesian confidence optimization for weight allocation allows for dynamic adjustment of the optimization direction based on the uncertainty of each objective, improving the robustness of the optimization. Introducing physical feasible region constraints ensures the practical feasibility of the optimization results, avoiding inconsistencies that may arise from purely data-driven methods. Compared to traditional single-objective or weighted summation multi-objective methods, this method can more comprehensively and flexibly balance multiple performance indicators, thereby obtaining a better combination of die-casting process parameters.

[0133] This application further proposes generating test samples with the highest potential reward based on Bayesian confidence optimization from the candidate parameter set, and calculating the reward value and constraint satisfaction for each sample group in parallel within the compensated digital twin model. During policy training, a proximal policy optimization method with convergence constraints is used to update the policy network, and an advantage estimation method is used to correct the value network. The confidence coefficients used for multi-objective weight allocation are updated at the outer layer, and the policy network and value network are updated at the inner layer. When the model's improvement rate is lower than the improvement rate threshold and the constraint conditions are met, training convergence is determined, and the resulting combination of process parameters is determined as the optimal die-casting process parameters.

[0134] In this process, Bayesian confidence optimization models the potential reward distribution using a Gaussian process. When generating test samples, a confidence upper bound criterion is used to screen high-potential parameter combinations, with the confidence coefficient dynamically calculated from the historical reward mean and variance. Parallel computing utilizes a distributed computing architecture to achieve synchronous simulation of multiple parameter sets in a digital twin model, where each parameter set corresponds to an independent thread and shares model cache data. Proximal policy optimization with convergence constraints introduces KL divergence constraints in the policy update step size to prevent training oscillations caused by policy mutations; the KL divergence threshold is set to 0.01-0.03. The advantage estimation method uses the generalized advantage estimation formula to calculate the temporal difference error, with a discount factor set to 0.95-0.99. The outer layer update of the confidence coefficient uses a sliding window to statistically analyze the distribution characteristics of multi-objective rewards, with a window length set to 50-100 iterations. The improvement rate threshold is set based on the statistical standard deviation of the objective function's rate of change; for example, convergence is triggered when the objective rate of change is less than 15% of the standard deviation for 10 consecutive iterations.

[0135] Specifically, Bayesian confidence optimization first generates a set of candidate parameters, predicts the potential reward distribution of each parameter combination based on Gaussian process regression, and selects the top 5%–10% of samples with the highest confidence upper limit as the experimental group. The digital twin model rapidly evaluates the reward value and constraint satisfaction status of each parameter group through multi-threaded parallel computation, with each thread allocated independent memory space to avoid data conflicts. The proximal policy optimization algorithm limits the difference between old and new policies by using KL divergence constraints when updating policy network parameters; when the KL divergence exceeds 0.02, it automatically reduces the learning step size. The value network correction uses a generalized advantage estimation method, combining the weighted average of the current state value and subsequent state values ​​to calculate the advantage function, with a discount factor set to 0.97 to balance short-term and long-term rewards. The confidence coefficient outer layer update module recalculates the multi-objective weights after every 20 iterations, adjusting the weight allocation ratio according to the variance of each objective reward, with objectives with higher variance receiving higher weights. When the objective function improvement rate is below 1.5% for 10 consecutive iterations and the satisfaction rate of all constraints exceeds 98%, the training process is considered converged, and the current optimal combination of process parameters is output. For example, in the scenario of mold temperature optimization, this method can reduce the defect rate from 2.1% to 0.7% within 200 iterations, while reducing energy consumption by 12%, and shortening the time of each iteration from 45 minutes in the traditional method to 8 minutes.

[0136] As a preferred embodiment, the specific implementation of this application is as follows: In the candidate parameter set generation stage, the Bayesian confidence optimization algorithm constructs a Gaussian process surrogate model based on historical training data. By maximizing the confidence upper bound function, it selects the 20 combinations of process parameters with the highest potential reward value from the casting temperature range [650℃, 720℃], injection speed range [3.5m / s, 5.2m / s], mold temperature range [180℃, 240℃], and holding time range [8s, 15s]. These 20 samples are input into the compensated digital twin model for parallel computation. Each computing node is allocated 3 samples. Through a distributed computing framework, the predicted value of casting defect rate, the estimated value of energy consumption, and the simulated value of production cycle corresponding to each sample are obtained synchronously, and it is verified whether they meet the physical feasible domain constraint that the peak pressure of the mold cavity is lower than 85MPa.

[0137] During the policy network update phase, the proximal policy optimization method uses a KL divergence threshold of 0.02 as a limit on the update magnitude. Each iteration samples 200 time-step interaction data from the experience pool, and the advantage function value is calculated using generalized advantage estimation. The value network adopts a dual-network structure: the main network outputs the state value estimate, and the target network synchronizes its parameters every 10 iterations to stabilize the training process. The outer confidence coefficient update module reallocates weights based on the mean-variance matrix of the current reward vector, increasing the defect rate target weight to 0.45, adjusting the energy consumption weight to 0.35, and decreasing the production cycle weight to 0.2. Training is considered converged when the reward increase over 5 consecutive iterations is less than 2% and all samples meet the mold temperature gradient constraint. The optimal parameter combination of 680℃ casting temperature, 4.8m / s injection speed, 210℃ mold temperature, and 12s holding time is then output.

[0138] Through the above technical solution, this application effectively solves the problem of traditional methods struggling to balance conflicts between objectives in multi-objective optimization. By employing a synergistic mechanism of Bayesian confidence optimization and constrained proximal policy optimization, the efficiency of process parameter search is improved. Safety constraints based on the physically feasible region ensure the practical feasibility of the optimization results, while the hierarchical update strategy, while maintaining training stability, achieves dynamic adaptation of multi-objective weights, enabling the final combination of process parameters to achieve an optimal balance between controlling casting defect rates, reducing energy consumption, and shortening production cycles.

[0139] This application further proposes generating test samples with the highest potential reward based on Bayesian confidence optimization from the candidate parameter set, and calculating the reward value and constraint satisfaction for each sample group in parallel within the compensated digital twin model. During policy training, a proximal policy optimization method with convergence constraints is used to update the policy network, and an advantage estimation method is used to correct the value network. The confidence coefficients used for multi-objective weight allocation are updated at the outer layer, and the policy network and value network are updated at the inner layer. When the model's improvement rate is lower than the improvement rate threshold and the constraint conditions are met, training convergence is determined, and the resulting combination of process parameters is determined as the optimal die-casting process parameters.

[0140] Among these methods, Bayesian confidence optimization models the probability distribution of the reward function and parameter space using a Gaussian process, generating experimental samples that comprehensively evaluate the expected and variance of potential rewards, thus improving the exploration efficiency of high-value regions. Parallel computing employs a distributed architecture to synchronously simulate multiple sets of parameter samples in a digital twin model, shortening the time per iteration. Convergence-constrained proximal policy optimization avoids training oscillations caused by abrupt changes in policy network parameters by limiting the step size range of policy updates. The advantage estimation method calculates action value bias based on temporal difference error, correcting the prediction error of the value network on state value. Outer layer updates dynamically adjust the confidence coefficient based on the historical reward mean and variance, adapting the multi-objective weight allocation to the non-uniform characteristics of the parameter space.

[0141] Specifically, Bayesian confidence optimization first generates a probability distribution model of parameter samples. It then calculates the expected potential reward and variance for each candidate sample using a collection function, selecting the sample with the highest overall score to join the experimental queue. A compensated digital twin model performs parallel simulations on multiple sets of parameters in the queue, outputting the defect rate, energy consumption, and production cycle indicators for each set of parameters. After normalization, the reward value is calculated, and the constraints are verified. Proximal policy optimization with convergence constraints ensures the stability of policy updates by limiting the KL divergence threshold between the old and new policies during policy network updates, while simultaneously adjusting the action selection probability based on the advantage estimate of the value network. The outer loop updates the confidence coefficient based on the reward statistics within the sliding window, while the inner loop alternately trains the policy network and the value network. When the reward improvement after several consecutive iterations falls below a preset threshold and all constraints are met, training terminates, and the current optimal combination of process parameters is output. This process achieves adaptive allocation of multi-objective weights and stable updating of policy parameters through a hierarchical optimization mechanism, improving parameter search efficiency and the reliability of optimization results.

[0142] As a preferred embodiment, the solution of this application is implemented as follows: In the die-casting production process of automobile engine cylinder blocks, after the optimal die-casting process parameters are deployed to the die-casting equipment, the system collects data on pouring temperature, injection pressure, mold cooling water flow rate, and surface defect detection of the casting at the same sampling frequency as the training phase, forming a verification dataset containing timestamps. The compensated digital twin model generates virtual prediction results based on the current operating parameters, and calculates the difference between the predicted and measured values ​​of filling time, surface porosity, and mold temperature distribution. When the porosity difference exceeds a preset threshold, the system extracts the verification data within the last 30 minutes as a sliding time window sample, freezes the low-sensitivity layer parameters in the virtual-real difference compensation network, and only performs backpropagation training on the cross-modal attention fusion layer. The updated network parameters are reviewed on an independently retained 15-minute verification subset. If the porosity prediction error decreases to within the threshold range, the updated parameters are synchronized to the digital twin model. If the error does not meet the standard, the system reverts to the previous version of parameters and extends the sliding time window to 45 minutes for retraining.

[0143] Through the above technical solution, this application realizes a dynamic closed-loop optimization mechanism for the virtual-real difference compensation network, solving the model prediction drift problem caused by mold wear and material batch fluctuations. By using time window sliding sampling and hierarchical freezing strategies, local parameters are quickly calibrated while ensuring model stability, ensuring that the digital twin model maintains a high degree of consistency with the actual production process in the long term, thereby continuously supporting the accurate optimization of process parameters.

[0144] In the above embodiments, a digital twin system centered on a multiphysics simulation model is constructed, dynamically coupling multi-source operational data such as temperature, pressure, and flow rate from the actual die-casting process with the virtual simulation model. This achieves real-time digital mapping and self-learning optimization of the die-casting process. A virtual-real difference compensation network is established using a multi-temporal Transformer neural network to adaptively correct the parameters of the digital twin model, improving the consistency between virtual predictions and actual production. Based on the compensated high-precision twin model, a multi-objective reinforcement learning mechanism incorporating Bayesian belief optimization is introduced. This mechanism can achieve trade-off optimization among multiple indicators such as casting defect rate, energy consumption, and production cycle, autonomously generating optimal die-casting process parameters. By applying the optimal parameters to equipment verification and dynamically updating the virtual-real compensation network based on the verification results, a closed-loop optimization system integrating virtual and real elements is formed. This achieves intelligent optimization and continuous self-evolution of die-casting process parameters, improving casting forming accuracy, energy efficiency, and production stability.

[0145] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a die-casting process parameter optimization system based on digital twins, used to apply the above-described die-casting process parameter optimization method based on digital twins, including:

[0146] The sensor unit includes a temperature sensor, a pressure sensor, and a flow sensor. The sensor unit is used to collect real-time operating parameters of the die-casting equipment and quality data of the die-cast parts, forming multi-source die-casting production data.

[0147] The acquisition unit is configured to establish a multiphysics simulation model that couples filling, heat transfer and solidification based on multi-source die casting production data, and to build a digital twin model based on the multiphysics simulation model. The digital twin model is used to reflect the dynamic behavior of the actual die casting process.

[0148] The processing unit is configured to compare the virtual prediction results output by the digital twin model with the measured quality data, construct a virtual-real difference compensation network based on a multi-temporal Transformer neural network, correct the parameters of the digital twin model, and obtain the compensated digital twin model.

[0149] The optimization unit is configured as a compensated digital twin model, with casting defect rate, energy consumption and production cycle as multi-objective functions. It maps process parameters to action space and quality indicators to reward function. Based on a multi-objective reinforcement learning method that integrates Bayesian belief optimization algorithm, it performs strategy search and iterative training on the combination of die casting process parameters, and iteratively searches and optimizes the die casting process parameters to generate the optimal die casting process parameters.

[0150] The verification unit is configured to apply the optimal die-casting process parameters to the die-casting equipment for verification, and update the virtual-real difference compensation network based on the verification results.

[0151] Specifically, the sensor unit collects real-time operating parameters of the die-casting equipment, which are then processed by the acquisition unit to form time-series data. This data is then transmitted to the processing unit to construct a digital twin model. This model achieves second-level simulation speeds through order reduction and surrogate model techniques. The processing unit inputs the virtual prediction results and measured quality data into a multi-time-series Transformer neural network to generate boundary condition corrections and update the model parameters. The optimization unit performs multi-objective reinforcement learning based on the updated digital twin model, searching for the optimal combination of process parameters under physical feasible region constraints. For example, it automatically excludes parameter combinations when the mold temperature is below 200℃. The verification unit applies the optimization results to the actual equipment and extracts verification data through a sliding time window. If the deviation between the surface defect rate and the simulation prediction exceeds a threshold, incremental learning is initiated to update the virtual-real difference compensation network, ensuring the model continuously adapts to changes in the production environment. This system solves the problems of parameter lag, multi-objective conflict, and model mismatch in traditional methods through data closure and collaborative optimization between units, achieving real-time dynamic optimization of die-casting process parameters.

[0152] As a preferred embodiment, the solution of this application is implemented as follows: The die-casting process parameter optimization system consists of a sensor unit, a data acquisition unit, a processing unit, an optimization unit, and a verification unit. The sensor unit includes a K-type thermocouple arranged at the outlet of the die-casting machine's injection cylinder, a piezoelectric pressure sensor integrated on the mold cavity wall, and an electromagnetic flowmeter installed in the cooling water circuit, used to acquire parameters such as injection speed, mold temperature, and cooling water flow rate. The data acquisition unit uses ANSYS Fluent software to establish a three-dimensional unsteady filling flow model, couples the alloy solidification latent heat model through a user-defined function, and uses dynamic layered mesh technology to locally refine the gate area, generating a multiphysics simulation model kernel. The processing unit constructs a Transformer network with a 12-layer encoder. The input layer receives the shrinkage cavity distribution data predicted by the digital twin model and the measured data from industrial CT scans. The cross-modal attention layer calculates the correlation characteristics between temperature gradient and solidification rate, and the output layer generates the correction amount for the mold heat transfer coefficient. The optimization unit deploys a multi-objective PPO algorithm based on the TensorFlow framework, continuously discretizing the casting temperature setpoint from 620-680℃ into 60 action nodes. In each iteration, the five sets of parameters with the highest upper bound of the confidence interval are selected and input into the digital genetic model for parallel simulation. The verification unit initiates an online monitoring mode after parameter application. When X-ray inspection detects that the number of pores exceeds a threshold, the incremental training module of the difference compensation network is triggered, freezing the parameters of the first eight layers and updating only the weights of the last four fully connected layers.

[0153] Through the above technical solutions, this application achieves closed-loop dynamic optimization of die-casting process parameters, solving the problem that traditional methods cannot respond to production fluctuations in real time. Continuous calibration of the multiphysics simulation model and measured data ensures the synchronization of the virtual model and physical equipment states, and the multi-objective optimization mechanism under the reinforcement learning framework balances quality, efficiency, and energy consumption indicators. The incremental update function of the difference compensation network enables the system to have online adaptive capabilities, suppressing model drift caused by mold wear or changes in alloy composition, and ensuring the continuous effectiveness of optimization parameters.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for optimizing die-casting process parameters based on digital twins, characterized in that, include: Temperature sensors, pressure sensors, and flow sensors are installed to collect real-time operating parameters of the die-casting equipment and quality data of the die-cast parts, forming multi-source die-casting production data. Based on the multi-source die casting production data, a multi-physics simulation model coupling filling, heat transfer and solidification is established, and a digital twin model is constructed with the multi-physics simulation model as the core. The digital twin model is used to reflect the dynamic behavior of the actual die casting process. The virtual prediction results output by the digital twin model are compared with the measured quality data. A virtual-real difference compensation network is constructed based on a multi-temporal Transformer neural network to correct the parameters of the digital twin model and obtain the compensated digital twin model. Based on the compensated digital twin model, with casting defect rate, energy consumption and production cycle as multi-objective functions, process parameters are mapped to action space, quality indicators are mapped to reward functions, and a multi-objective reinforcement learning method based on Bayesian belief optimization algorithm is used to search for and iteratively train the combination of die casting process parameters, and to search for and optimize the die casting process parameters to generate the optimal die casting process parameters. The optimal die-casting process parameters are applied to the die-casting equipment for verification, and the virtual-real difference compensation network is updated based on the verification results.

2. The method for optimizing die-casting process parameters based on digital twins according to claim 1, characterized in that, When collecting real-time operating parameters of die-casting equipment and quality data of die-cast parts to form multi-source die-casting production data, the following should be included: Temperature sensors, pressure sensors, and flow sensors are respectively installed on the gating system, injection system, and mold temperature control circuit of the die casting equipment, and timestamps are synchronized. The collected temperature, pressure, and flow signals are filtered for noise and feature extracted to generate operating parameter data that includes process characteristics, mold temperature distribution characteristics, and filling flow rate characteristics. The quality data output by the die-casting quality inspection equipment is correlated with the operating parameter data in a time series to form the multi-source die-casting production data, which includes surface defects, internal shrinkage cavities, and density distribution.

3. The method for optimizing die-casting process parameters based on digital twins according to claim 1, characterized in that, When establishing a multiphysics simulation model of mold filling, heat transfer, and solidification coupling based on the aforementioned multi-source die casting production data, the following are included: Based on the multi-source die casting production data, boundary conditions are identified and material parameters are assimilated. The gate volume flow rate, cavity heat transfer coefficient and alloy thermal properties are taken as the quantities to be identified. The initial values ​​are obtained by the least squares inversion with Tikhonov regularization and updated recursively in the time series using extended Kalman filtering. The incompressible Navier-Stokes equation and energy equation are solved separately and coupled with a phase change solidification model. The flow-heat-solidification simultaneous iteration is achieved through partitioned coupling and staggered time steps. Mesh refinement and step size control are applied to the gate, thin-walled, and hot-spot regions. The convergence criterion is that the simulation-measured residual is less than a preset threshold, thus obtaining the multiphysics simulation model.

4. The method for optimizing die-casting process parameters based on digital twins according to claim 3, characterized in that, When constructing a digital twin model based on the aforementioned multiphysics simulation model, the following steps are included: Based on the multiphysics simulation model, a mapping between state vectors and observation vectors is generated, a state observer is constructed, and real-time multi-source die-casting production data is input into the state observer to achieve virtual-real state synchronization. The multiphysics simulation model is subjected to POD-Galerkin order reduction, and a Gaussian process surrogate model is added to the time-consuming local subprocesses to form the digital twin model.

5. The method for optimizing die-casting process parameters based on digital twins according to claim 1, characterized in that, When constructing a virtual-real difference compensation network based on a multi-temporal Transformer neural network, the following are included: The virtual prediction results and measured quality data are time-stamp aligned and dimensionally normalized, and a multimodal time-series feature vector is constructed using temperature, pressure, flow rate and quality detection indicators. Each modality is assigned an independent encoder and positional encoding and missing mask are added. Differential representations are extracted based on a shared temporal self-attention layer and a cross-modal attention fusion layer. Output parameter correction vector, which corresponds to a subset of boundary condition parameters and material parameters of the digital twin model; Uncertainty weights are calculated based on sensor confidence and quality detection confidence and used for weighting the difference representation to obtain the virtual-real difference compensation network.

6. The method for optimizing die-casting process parameters based on digital twins according to claim 5, characterized in that, When correcting the parameters of the digital twin model to obtain a compensated digital twin model, the process includes: The virtual-real difference compensation network is trained with a combined loss consisting of residual regression loss and physical consistency regularization term, wherein the physical consistency regularization term constrains the corrected heat transfer coefficient, volumetric flow rate and solidification parameters to satisfy the stability region of the multiphysics equation. The robust Huber loss is used to suppress outlier samples, and a sliding time window and early stopping strategy are used to control training stability. The parameter correction vector is constrained within a preset feasible region by a projection operator, and the compensated digital twin model is obtained by incrementally updating it with a step size decay coefficient. Parameter updates are triggered only if both the virtual and real residual thresholds and the prediction confidence thresholds are met; otherwise, the previously valid parameters are retained.

7. The method for optimizing die-casting process parameters based on digital twins according to claim 1, characterized in that, When performing a multi-objective reinforcement learning method based on the compensated digital twin model, it includes: A state space is constructed using the state vector of the compensated digital twin model, and an action space is constructed by making the adjustable process parameters continuous. The adjustable process parameters include pouring temperature, injection speed, mold temperature, and holding time. The defect rate of the casting, energy consumption and production cycle are used to form a reward vector, which is then processed to be dimensionless and directional consistent. The reward vector is weighted using a Bayesian confidence optimization algorithm, and the weights are determined based on the confidence coefficients calculated from the mean and variance of the rewards. Using the physical feasible region output by the digital twin model as the set of security constraints, penalties are applied to samples that violate the feasible region and then removed, thereby obtaining a constraint-based single-objective reward.

8. The method for optimizing die-casting process parameters based on digital twins according to claim 7, characterized in that, When performing strategy search and iterative training on combinations of die-casting process parameters, and iteratively searching and optimizing the die-casting process parameters to generate the optimal die-casting process parameters, the following steps are taken: The test sample with the highest potential reward is generated based on Bayesian confidence optimization from the candidate parameter set, and the reward value and constraint satisfaction of each sample are calculated in parallel in the compensated digital twin model. During policy training, a proximal policy optimization method with convergence constraints is used to update the policy network, and an advantage estimation method is used to correct the value network. The confidence coefficients used for multi-objective weight allocation are updated in the outer layer, and the policy network and value network are updated in the inner layer. When the improvement rate of the model is lower than the improvement rate threshold and the constraints are met, the training is considered to have converged, and the resulting combination of process parameters is determined as the optimal die-casting process parameters.

9. The method for optimizing die-casting process parameters based on digital twins according to claim 1, characterized in that, When applying the optimal die-casting process parameters to die-casting equipment for verification, and updating the virtual-real difference compensation network based on the verification results, the following steps are included: Based on the optimal die-casting process parameters, operational parameter data and quality data are collected according to the timestamp alignment rules consistent with the training phase to construct a verification dataset. The compensated digital twin model generates corresponding virtual prediction results under the same working conditions, and calculates the virtual-real difference between the virtual prediction results and the measured quality data. When the virtual-real difference exceeds a preset difference threshold, the virtual-real difference compensation network is incrementally learned and updated. The incremental learning update uses the validation data within the sliding time window as training samples and adopts a hierarchical freezing strategy to update only the high-sensitivity parameter layer. After the update is completed, it is reviewed on an independent reserved subset. If the review passes, the updated virtual-real difference compensation network is synchronized to the digital twin model. If the review fails, the parameters of the previous version remain unchanged.

10. A digital twin-based die-casting process parameter optimization system, used to apply the digital twin-based die-casting process parameter optimization method according to any one of claims 1-9, characterized in that, include: The sensor unit includes a temperature sensor, a pressure sensor, and a flow sensor. The sensor unit is used to collect real-time operating parameters of the die-casting equipment and quality data of the die-cast parts to form multi-source die-casting production data. The acquisition unit is configured to establish a multi-physics simulation model of filling, heat transfer and solidification coupling based on the multi-source die casting production data, and to construct a digital twin model with the multi-physics simulation model as the core. The digital twin model is used to reflect the dynamic behavior of the actual die casting process. The processing unit is configured to compare the virtual prediction results output by the digital twin model with the measured quality data, construct a virtual-real difference compensation network based on a multi-temporal Transformer neural network, correct the parameters of the digital twin model, and obtain a compensated digital twin model. The optimization unit is configured to use the compensated digital twin model as a multi-objective function, with casting defect rate, energy consumption and production cycle as the multi-objective function, map process parameters to action space, map quality indicators to reward function, and perform strategy search and iterative training on die casting process parameter combinations based on a multi-objective reinforcement learning method that integrates Bayesian belief optimization algorithm, and perform iterative search and strategy optimization on die casting process parameters to generate optimal die casting process parameters. The verification unit is configured to apply the optimal die-casting process parameters to the die-casting equipment for verification, and update the virtual-real difference compensation network based on the verification results.

Images