Large complex magnesium alloy component semi-solid injection one-key forming full-parameter self-optimization method, system and equipment

Abstract

The application belongs to the technical field of metal material advanced forming manufacturing and industrial artificial intelligence, and particularly relates to a large-scale complex magnesium alloy component semi-solid injection one-key forming full-parameter self-optimization method, system and equipment. The method comprises the following steps: constructing a digital twin model integrating physical mechanism and data-driven calibration; optimizing a multi-objective optimization objective function; loading a parameter set to a physical injection molding machine and collecting multi-dimensional time sequence sensing data flow; comparing the sensing data flow with a digital twin model prediction to generate a deviation index, which drives an actuator to adjust parameters online; storing data and performing incremental learning; when a preset period is completed or accumulated data reaches a threshold, triggering rolling re-optimization to generate an updated parameter set and performing smooth switching. The application realizes automatic optimization, online self-adaptation and continuous evolution of process parameters through the construction of an intelligent closed loop of sensing-decision-execution-learning, improves development efficiency and guarantees quality consistency, and reduces dependence on expert experience.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of advanced metal material forming and manufacturing and industrial artificial intelligence, specifically involving a method, system and equipment for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters. Background Technology

[0002] Semi-solid injection molding is a key process for manufacturing high-performance, lightweight, large, and complex magnesium alloy components. This technology is highly complex, involving multiple nonlinear and strongly coupled physical stages, including semi-solid slurry preparation, high-speed injection, and pressurized solidification. Its final quality is influenced by the interaction of dozens of high-dimensional process parameters, such as billet temperature, mold temperature field, injection speed curve, and pressure curve. For large and complex components, the process window is narrow, and traditional trial-and-error methods relying on engineer experience are time-consuming, costly, and difficult to obtain globally optimal parameters.

[0003] Currently, the intelligent upgrading of this field faces three major bottlenecks: the disaster of dimensionality in parameter optimization: traditional experimental design or single-point simulation is difficult to efficiently handle high-dimensional and nonlinear parameter spaces, resulting in low optimization efficiency; the lack of process "black box" and adaptiveness: the molding process occurs in a closed mold, key physical fields cannot be directly observed, and fixed parameter sets cannot cope with disturbances such as material fluctuations and equipment state drift, leading to quality fluctuations; and the lack of a knowledge loop: simulation, production, data, and optimization are isolated, and process knowledge cannot automatically iterate and evolve with the accumulation of production data.

[0004] While existing technologies utilize PID control, simple statistical models, or offline simulation for parameter setting, none have achieved a complete, closed-loop, autonomous optimization process encompassing the entire workflow and all parameters, from "intelligent initial parameter generation" to "online adaptive adjustment" and then to "self-evolution of process knowledge." Therefore, developing an intelligent method capable of automatically optimizing complex parameters with a single click and continuously maintaining stable production has become a critical challenge awaiting breakthroughs in the industry. Summary of the Invention

[0005] Therefore, the primary objective of this invention is to provide a method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully self-optimized parameters. This method aims to construct a complete intelligent closed loop of perception-decision-execution-learning, integrating digital twin models, artificial intelligence, and real-time control technology to achieve automatic optimization, online self-adaptation, and continuous evolution of process parameters without repeated manual adjustments, ultimately improving development efficiency, ensuring quality consistency, and reducing reliance on expert experience.

[0006] The second objective of this invention is to provide a fully parameter self-optimizing system for one-click semi-solid injection molding of large and complex magnesium alloy components.

[0007] A third objective of the present invention is to provide an electronic device for performing the above-described methods and systems.

[0008] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters, including the following steps:

[0010] S1. Obtain target component information and equipment constraints, and construct a high-fidelity digital twin model; the digital twin model is a hybrid model combining a physical mechanism model and a data-driven calibration model;

[0011] S2. Based on the multi-objective optimization objective function, optimize the high-dimensional process parameter space in the digital twin model to generate one or more sets of candidate parameters;

[0012] S3. Select a target parameter set from the candidate parameter set, load it into the physical injection molding machine, and execute the molding cycle;

[0013] S4. Acquire multi-dimensional time-series sensing data streams from multiple source sensors during the molding cycle;

[0014] S5. The multi-dimensional time-series sensing data stream is compared in real time with the theoretical prediction data synchronously output by the digital twin model to calculate the residual and generate a deviation index; the deviation index includes at least one of temperature deviation, pressure deviation and filling speed deviation.

[0015] S6. Input the deviation index into the edge artificial intelligence diagnostic model to output the process status identification result and defect risk prediction result, and call the adaptive control algorithm to calculate the control correction amount when the trigger condition is met, and drive the actuator to fine-tune the target parameter set online.

[0016] S7. Store the full process data, control correction amount and final product quality data of the current molding cycle into the process big data platform, and perform incremental learning or parameter calibration on the digital twin model based on the accumulated data of the process big data platform.

[0017] S8. When the preset cycle is completed or the accumulated data volume reaches the threshold, rolling re-optimization is triggered to generate an updated parameter set, and a smooth switching strategy is adopted to update the updated parameter set to the physical injection molding machine to continuously evolve the digital twin model.

[0018] Preferably, in step S1, the physical mechanism model is constructed based on the equations of mass conservation, momentum conservation and energy conservation, and coupled with a non-Newtonian fluid constitutive model and a phase transition-solidification model.

[0019] The data-driven calibration model is used to perform inversion calibration on the parameters to be identified in the physical mechanism model. The parameters to be identified include heat transfer coefficient, apparent viscosity model parameters or equipment dynamic parameters. The inversion calibration of the parameters to be identified adopts genetic algorithm, gradient descent method or Bayesian inversion method.

[0020] Preferably, in step S2, the high-dimensional process parameter space includes, but is not limited to: initial temperature of semi-solid billet, temperature of multiple zones of mold, screw speed and back pressure, multi-segment injection speed curve, injection to holding pressure switching point, injection pressure and holding pressure time;

[0021] The optimization process shall employ at least one of the following methods:

[0022] Deep reinforcement learning: The optimization process is constructed as a sequential decision-making process, wherein the state is composed of at least one of the temperature field, pressure field, flow front or solid fraction related features output by the digital twin model, the action is defined as the adjustment of the velocity curve, pressure curve, switching point or temperature setpoint, and the reward is calculated based on the multi-objective optimization objective function.

[0023] Bayesian optimization: By constructing a probabilistic surrogate model of the objective function and using the acquisition function to guide the selection of sampling points, a set of candidate parameters can be obtained under a finite number of twin evaluations.

[0024] More preferably, the multi-objective optimization objective function includes at least one of a quality objective and an efficiency objective; the quality objective includes at least one of a defect index, a density index, a warpage amount, a flow / filling uniformity, or a dimensional deviation; the efficiency objective includes at least one of a cycle time, energy consumption, or material utilization rate.

[0025] Preferably, in step S4, the multi-source sensors include at least: a temperature and pressure sensor embedded in the mold cavity, an ultrasonic sensor for online detection of solid fraction, a machine vision system for monitoring the surface of the component, and servo motor and hydraulic system sensors of the physical injection molding machine body.

[0026] Preferably, in step S6, the edge AI diagnostic model is a lightweight temporal learning model, which includes at least one of a one-dimensional convolutional neural network, a gated recurrent unit, or a temporal transformation model, and is used to extract features based on the multi-dimensional temporal sensing data stream and output process state identification results, wherein the process state includes at least one of filling imbalance, stagnation, jetting, short-shot risk, or over-pressure risk.

[0027] The adaptive control algorithm includes Model Predictive Control (MPC), which performs rolling optimization of the future control time domain based on a simplified predictive model to obtain control corrections, and forms a feedforward-feedback composite control with Proportional-Integral-Derivative (PID). The control corrections include online fine-tuning of servo valve opening, injection / delivery speed, injection / holding pressure, switching point, or heating power.

[0028] Preferably, in step S7, the incremental learning is based on online local fine-tuning of the current production batch data.

[0029] Preferably, in step S8, the preset cycle is triggered once every N units produced, where N is 50 to 5000; the smooth switching strategy includes at least one of amplitude and slope limiting switching, segmented interpolation switching, or parallel verification switching of new and old parameter sets.

[0030] Secondly, based on the above method, the present invention provides a fully parameter self-optimizing system for one-click semi-solid injection molding of large and complex magnesium alloy components, comprising:

[0031] The twin modeling and parameter calibration module is used to establish a digital twin model based on physical mechanisms and data-driven calibration models, and to perform inversion calibration on the parameters to be identified in the digital twin model using historical data.

[0032] The intelligent optimization module is used to perform multi-objective optimization in the digital twin model and output a set of candidate parameters;

[0033] The parameter loading and execution module is used to determine the target parameter set from the candidate parameter set, load it into the physical injection molding machine, and execute the molding cycle.

[0034] The multi-source sensing module is used to acquire multi-dimensional time-series sensing data streams;

[0035] The real-time comparison and deviation calculation module is used to generate twin theory prediction data and compare it with the sensing data stream to output deviation indicators;

[0036] The edge intelligent diagnosis and control module is used to output process status identification results and defect risk prediction results. When the triggering conditions are met, it outputs control correction quantities based on the MPC-based adaptive control algorithm and drives the actuator.

[0037] The data platform and model learning module are used to store molding cycle data and incrementally learn or calibrate the digital twin model;

[0038] The model evolution and parameter update module is used to optimize and update the parameter set and perform a smooth switch when a preset cycle is completed or the accumulated data volume reaches a threshold, so as to continuously evolve the digital twin model.

[0039] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a communication interface. The memory stores a computer program that can run on the processor. When the computer program is executed, it implements the above-mentioned method for one-click semi-solid injection molding of large and complex magnesium alloy components with full parameter self-optimization. The industrial communication interface is used to receive multi-source sensor data and output control correction quantities to the actuator.

[0040] The beneficial effects of this invention are:

[0041] 1. This invention constructs a high-fidelity digital twin model that combines a physical mechanism model with a data-driven calibration model. In the twin environment, a multi-objective optimization objective function is performed to find the optimal parameter set and output a set of candidate parameters. The target parameter set in the candidate parameter set is then automatically sent to the physical injection molding machine to complete the start-up and production cycle. This transforms the traditional serial trial-and-error process of "trial production - measurement - parameter adjustment - re-trial production" that relies on the experience of process experts into a parallel search and automatic configuration in virtual space. It can compress the process debugging work of large and complex components, which takes weeks or even months, to the level of a few days and significantly reduce the dependence on senior process experts to continuously adjust parameters on site.

[0042] 2. In the calibrated digital twin model, this invention employs deep reinforcement learning or Bayesian optimization artificial intelligence algorithms to globally explore the high-dimensional, strongly coupled process parameter space. Deep reinforcement learning can learn multi-stage optimal control sequences through a sequential decision-making mechanism of "state-action-reward," while Bayesian optimization can efficiently locate the global optimal region under a limited simulation budget through surrogate models and acquisition functions. Thus, it overcomes the limitations of traditional methods, such as easy local convergence, low sampling efficiency, and difficulty in capturing complex interaction effects. It can discover Pareto-optimal parameter combinations that are difficult to balance quality and efficiency with traditional experience.

[0043] 3. This invention acquires multi-dimensional time-series sensing data streams in real time through multi-source sensors, and aligns and fuses them with theoretical prediction data generated by a digital twin model to form a deviation index. Then, an edge artificial intelligence diagnostic model identifies the process status and defect risks in real time and triggers the model predictive control adaptive control algorithm to output control corrections, thereby forming a closed-loop correction mechanism of "online sensing-diagnosis-control". This enables the production line to dynamically resist internal and external disturbances such as material fluctuations, thermal field drift, and equipment status drift, significantly improving the cross-batch consistency and operational stability of product quality, and reducing abnormal events such as filling imbalance, stagnation, jetting, and short-shot caused by disturbances.

[0044] 4. This invention stores the entire process data, control corrections, and final product quality data of each molding cycle into a process big data platform. Based on the accumulated data, it performs incremental calibration or retraining on the key empirical parameters in the digital twin model using inversion calibration. At the same time, it periodically triggers rolling re-optimization to generate updated parameter sets and applies them to the production system through a smooth switching strategy. Therefore, it can continuously reduce the deviation between the digital twin model and reality, improve the credibility of optimization and control decisions, and enable the process capability to continuously iterate and improve with the accumulation of production data, forming a truly self-learning and self-optimizing intelligent system.

[0045] 5. This invention achieves a configurable trade-off between quality, cycle time, energy consumption, and material utilization by optimizing the objective function through multiple objectives. It also reduces defects and rework by combining online closed-loop diagnostic control with rolling re-optimization to continuously explore the potential for efficiency improvement. Therefore, it can improve material utilization and first-pass yield, reduce scrap rate and energy consumption per unit product, thus providing core technical support for the large-scale and stable production of high-end magnesium alloy large and complex components. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the self-optimization method for all parameters of semi-solid injection molding of large and complex magnesium alloy components provided in this embodiment of the invention.

[0048] Figure 2 This is a schematic diagram illustrating the construction principle of the digital twin hybrid model in an embodiment of the present invention.

[0049] Figure 3 This is a schematic diagram of the architecture of a semi-solid injection molding full-parameter self-optimization system for large and complex magnesium alloy components provided in an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Reference Figure 1This invention provides a method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters, including the following steps:

[0052] S1. Obtain target component information and equipment constraints, and construct a high-fidelity digital twin model; the digital twin model is a hybrid model that combines a physical mechanism model and a data-driven calibration model.

[0053] S2. Based on the multi-objective optimization objective function, optimize the high-dimensional process parameter space in the digital twin model to generate one or more sets of candidate parameters.

[0054] S3. Select the target parameter set from the candidate parameter set, load it into the physical injection molding machine, and execute the molding cycle.

[0055] S4. Acquire multi-dimensional time-series sensing data streams from multiple sources of sensors during the molding cycle.

[0056] S5. Compare the multidimensional time-series sensing data stream with the theoretical prediction data synchronously output by the digital twin model in real time, calculate the residuals and generate deviation indicators; the deviation indicators include at least one of temperature deviation, pressure deviation and filling speed deviation.

[0057] S6. Input the deviation index into the edge artificial intelligence diagnostic model to output the process status identification result and defect risk prediction result, and call the adaptive control algorithm to calculate the control correction amount when the trigger condition is met, driving the actuator to fine-tune the target parameter set online.

[0058] S7. Store the entire process data of the current molding cycle, control corrections, and final product quality data into the process big data platform, and perform incremental learning or parameter calibration on the digital twin model based on the accumulated data of the process big data platform.

[0059] S8. When the preset cycle is completed or the accumulated data volume reaches the threshold, rolling re-optimization is triggered to generate an updated parameter set, and a smooth switching strategy is adopted to update the updated parameter set to the physical injection molding machine to continuously evolve the digital twin model.

[0060] like Figure 2As shown, the physical mechanism model is constructed based on the equations of mass conservation, momentum conservation, and energy conservation, and coupled with a non-Newtonian fluid constitutive model and a phase transition-solidification model. For example, iterative solutions are performed based on the mass conservation, momentum conservation, and energy conservation equations. The finite element method (FEM) is used to discretize and simulate the temperature field, flow field, pressure field, and solidity evolution. The mesh size for the FEM discretization is controlled at 0.5 mm, and the entire solution domain is divided into 5 million elements. This mesh size and solution domain division effectively utilizes computational resources while capturing the details of the flow and temperature fields. Through this meshing strategy, the temperature field control accuracy can reach ±2°C, the flow field velocity error can be controlled within 5%, the pressure field error is less than 3%, and the prediction error for the solidity evolution is less than 5%.

[0061] To further improve the prediction accuracy of the physical mechanism model, a data-driven calibration model is introduced to achieve deep integration of physical mechanisms and data features. The constructed data-driven calibration model is an offline global calibration based on historical batch data. This model is used to perform inversion calibration on the parameters to be identified in the physical mechanism model. These parameters include interfacial heat transfer coefficients, apparent viscosity model parameters, or equipment dynamic parameters. The inversion calibration of these parameters employs genetic algorithms, gradient descent methods, or Bayesian inversion methods. A neural network compensator is trained using historical data to calibrate the residuals of the physical model. For example, a neural network compensator is trained using historical production data from 1000 sets of past process parameters and quality data. This compensator has a 3-layer fully connected structure: a 12-node input layer, a 64-node hidden layer, and a 6-node output layer. This neural network compensator is used to calibrate the residuals of the parameters to be identified in the physical mechanism model. For example, the initial value of the interfacial heat transfer coefficient is 500 W / m³. 2 After data-driven compensation, the error of K was reduced from ±10% to ±3%.

[0062] In step S1, the specific steps for acquiring target component information and equipment constraints and constructing a high-fidelity digital twin model are as follows: Based on the 3D CAD model of the target component, the equipment dynamics model, and the semi-solid rheological and thermophysical data of the magnesium alloy, a high-fidelity digital twin model covering the entire process of feeding, plasticizing, injection, pressure holding, and cooling is constructed by integrating physical mechanisms and data-driven mechanisms. This high-fidelity digital twin model can accurately simulate the evolution of temperature field, flow field, pressure field, and solid fraction. For example, the 3D CAD model can be a new energy vehicle battery pack shell with dimensions of 1500mm × 800mm and an average wall thickness of 2.5mm; the equipment dynamics model can be a screw diameter of 120mm and an injection capacity of 5000cm³. 3The injection molding machine was used; the semi-solid rheological and thermophysical data of the magnesium alloy can be selected from AZ91 magnesium alloy, which has a solidus point of 470°C and a liquidus point of 595°C. The rheological parameters were measured by a capillary rheometer, and the viscosity model was the Cross-WLF model. The parameters of the Cross-WLF model are as follows: zero-shear viscosity η0 is 250 Pa·s, and relaxation time λ is 0.5 s. Those skilled in the art will understand that the above parameters are only examples and can be adjusted according to the specific component size, alloy grade, and equipment model in actual applications, and do not constitute a limitation of the present invention. Of course, depending on the specific needs, other parameters and data can be selected to construct a high-fidelity digital twin model of the entire process.

[0063] It should be noted that, to achieve rapid response in actual production for the aforementioned high-precision simulation, the digital twin model runs on a high-performance computing platform. For example, using an NVIDIA A100 GPU computing platform for parallel acceleration, regardless of whether a neural network compensator or inversion calibration is employed, a single complete simulation takes 30 minutes. Of course, other computing platforms can be selected according to different needs, and this invention does not limit this choice.

[0064] To balance high fidelity and real-time performance, the digital twin model of this invention employs a layered implementation of offline and online methods. In the offline stage, a high-fidelity digital twin model is run for inversion calibration of the parameters to be identified. In the online stage, to meet the requirements of real-time comparison and closed-loop control within the molding cycle, the data generated by the offline digital twin model is trained to obtain a surrogate model, which outputs theoretical prediction data. For example, the surrogate model can be any one of a neural network regression model, a Gaussian process surrogate model, or a multinomial response surface model.

[0065] In step S2, based on a multi-objective optimization objective function, the high-dimensional process parameter space is optimized in the digital twin model to generate one or more sets of candidate parameters, aiming to simultaneously consider product quality and production efficiency. The multi-objective optimization objective function includes at least one of a quality objective and an efficiency objective; the quality objective includes at least one of defect index, density index, warpage, flow / fill uniformity, or dimensional deviation; the efficiency objective includes at least one of cycle time, energy consumption, or material utilization. For example, a comprehensive quality objective is set as the objective function: maximizing the target value of virtual CT scan predicted density > 99.5% and minimizing the target value of virtual warpage < 0.1 mm to ensure the internal quality and dimensional accuracy of the molded part.

[0066] It should be noted that the high-dimensional process parameter space includes, but is not limited to: initial temperature of semi-solid billet, temperature of multiple zones of mold, screw speed and back pressure, multi-segment injection speed curve, injection to holding pressure switching point, injection pressure and holding pressure time. To achieve high efficiency and optimization, the key parameters are set as follows: initial temperature of the semi-solid billet is 480–520°C; mold multi-zone temperature is 180–250°C; screw speed curve: three-segment control is adopted, with a speed range of 50–150 rpm. The first segment is a low-speed segment with a speed of 50 rpm, used for initial feeding to prevent raw material bridging or premature shearing; the second segment is a medium-speed segment with a speed of 100 rpm, used for stable feeding and plasticizing; the third segment is a high-speed segment with a speed of 150 rpm, used for final homogenization and pressure building. In semi-solid injection molding, back pressure is used to control the melt density and temperature uniformity during the screw plasticizing stage. Appropriate back pressure can expel gas from the slurry and improve density, but excessive back pressure will increase energy consumption and shear heat, and destroy the spherical grain structure of the semi-solid slurry. Multi-segment injection speed profile: A five-segment control system is employed, with a speed range of 60–120 mm / s. The first segment, a slow speed of 60 mm / s, is used for slow mold closing or initial filling, facilitating venting. The second segment, a medium-high speed of 80 mm / s, is used for rapid cavity filling, reducing cold material. The third segment, a medium-high speed of 100 mm / s, is used for rapid cavity filling, reducing cold material. The fourth segment, a medium-high speed of 120 mm / s, is used for rapid cavity filling, reducing cold material. The fifth segment, a low speed of 100 mm / s before holding pressure, serves as a buffer to prevent impact-induced flash or short-filling. The injection pressure at the V / P switching point is 40–80 MPa; the holding pressure time is 10–30 s. This parameter range covers a typical process window, providing ample room for subsequent optimization.

[0067] In step S2, the optimization employs at least one of the following methods: Deep reinforcement learning: The optimization process is constructed as a sequential decision-making process, wherein the state is composed of at least one of the temperature field, pressure field, flow front, or solid fraction-related features output by the digital twin model; the action is defined as an adjustment to the velocity curve, pressure curve, switching point, or temperature setpoint; and the reward is calculated based on the multi-objective optimization objective function. Bayesian optimization: A probabilistic surrogate model of the objective function is constructed, and the sampling point selection is guided by the acquisition function to obtain a set of candidate parameters under a finite number of twin evaluations.

[0068] For example, a deep reinforcement learning algorithm is used to automatically optimize the high-dimensional parameter space and construct the process parameter adjustment as a sequential decision process.

[0069] Specifically, the core elements of the PPO algorithm are set as follows:

[0070] State space: 15-dimensional, consisting of key physical field features output in real time by the digital twin model, including temperature field, pressure field, flow front position and solid fraction distribution, used to characterize the current process state.

[0071] Action space: 12 dimensions, corresponding to the process parameters to be optimized, namely, the continuous actions of adjusting the initial temperature of the semi-solid billet, the temperature of the mold multi-zone, the values ​​of each segment of the screw speed curve, the values ​​of each segment of the injection speed curve, the pressure at the V / P switching point, and the holding time.

[0072] Reward function: Based on the multi-objective optimization objective function calculation, the predicted density of virtual CT scan >99.5% and virtual warping deformation <0.1mm are converted into reward signals to guide the agent to explore high-quality, low-deformation parameter combinations.

[0073] Algorithm parameters: learning rate set to 0.001, discount factor set to 0.99, training iterations 1 million times to ensure the policy network fully converges.

[0074] Through the PPO algorithm described above, the agent continuously tries and fails in the digital twin environment, selects actions based on the current state, receives rewards and updates its strategy, and gradually approaches the optimal combination of process parameters.

[0075] After implementing the above optimization strategy, and following 72 hours of virtual testing (corresponding to 800,000 digital twin simulations), three sets of candidate process parameters were finally output. The optimal candidate parameter set yielded predictions of porosity below 0.5% and grain size below 50 μm, fully meeting the preset quality targets. This result verifies the effectiveness and efficiency of this method in optimizing high-dimensional process parameter spaces under multi-objective constraints. The output candidate parameter sets can be directly used for actual production debugging or as a benchmark for subsequent process optimization.

[0076] For example, the specific optimization process using Bayesian optimization is as follows: the initial blank temperature is set to 480-520°C. Through Bayesian optimization, the blank temperature is reduced to the high probability region of 490-510°C, which shortens the time of the entire molding cycle by 2.1%.

[0077] In step S3, the multiple candidate parameter sets output in step S2 are viewed through the human-machine interface. Each parameter set contains complete process parameter configurations and corresponding predicted quality indicators. The operator can select a target parameter set according to the real-time needs of the production site. For example, the parameters of the target parameter set are as follows: billet temperature is 500°C, mold temperature is 220°C, and injection speed curve is [60, 90, 110, 100, 80] mm / s. The selected target parameter set is loaded into the physical injection molding machine with one click through the intelligent control unit integrating Intel Xeon CPU and FPGA real-time module, automatically configuring the parameters and starting the molding cycle. Of course, other target parameter sets can be selected according to different needs, and this invention does not limit this.

[0078] In step S4, the multi-source sensors include at least: temperature and pressure sensors embedded in the mold cavity, ultrasonic sensors for online detection of solid fraction, machine vision systems for monitoring the surface of components, and servo motor and hydraulic system sensors of the physical injection molding machine body. For example, the miniature temperature sensor embedded in the mold cavity is a PT100 type with an accuracy of ±0.5°C and a sampling frequency of 100Hz; the pressure sensor embedded in the mold cavity is a piezoelectric type with a range of 0~100MPa and an accuracy of ±0.1MPa. The temperature and pressure sensors together form a spatially distributed monitoring network of 48 nodes; the ultrasonic sensor has a frequency of 5MHz and a resolution of ±2%, and is installed at a specific position in the mold to invert the solid fraction through the change in ultrasonic speed; the machine vision system is a CCD camera with a resolution of 1280×720 and a frame rate of 60fps, which captures image information of the visible area of ​​the mold cavity in real time, and extracts features such as the position of the flow front and surface defects through image processing algorithms; the servo motor encoder has a resolution of 1μm, and the hydraulic system pressure sensor has an accuracy of ±0.5%, which is used to provide real-time feedback on the screw position, speed, and hydraulic pressure.

[0079] For example, data collected from multiple sensors is transmitted via industrial Ethernet with a transmission delay of <1ms, ensuring data synchronization and timeliness. The collected data stream is organized by timestamps to form a multi-dimensional time-series sensing dataset containing temperature field, pressure field, solidity, image features, and equipment status.

[0080] In step S5, the digital twin model outputs theoretical prediction data in real time. The multi-dimensional time-series sensing data stream collected in step S4 is compared with the theoretical prediction data in real time to calculate the residual and generate a deviation index. The deviation index includes at least one of temperature deviation, pressure deviation, and filling speed deviation. For example, by comparing the multi-dimensional time-series sensing data stream with the theoretical prediction data in real time, a pressure deviation index with a phase difference tolerance of ±5ms is generated.

[0081] In step S6, the edge AI diagnostic model is a lightweight temporal learning model, comprising at least one of a one-dimensional convolutional neural network, a gated recurrent unit, or a temporal transformation model, used to extract features based on a multi-dimensional temporal sensing data stream and output process state identification results. For example, a lightweight one-dimensional convolutional neural network is used, with a model parameter size <1MB and a single inference time <1ms, meeting millisecond-level real-time diagnostic requirements. The lightweight one-dimensional convolutional neural network model, after training, can identify at least one of the following process states: filling imbalance, stagnation, jetting, short-shot risk, or over-pressure risk. Process states can be quantified into specific index outputs. For example, a normal threshold range for the filling balance index is set to >0.9; values ​​below this value are considered filling imbalance. When the edge AI diagnostic model identifies an abnormal process state, or when the deviation index in step S5 exceeds the preset tolerance, the system automatically triggers an adaptive control algorithm to calculate control corrections for closed-loop fine-tuning. For example, if the pressure at a thin rib drops by 10%, the adaptive control algorithm is automatically triggered. The adaptive control algorithm includes Model Predictive Control (MPC). MPC performs rolling optimization in the future control time domain based on a simplified predictive model to obtain control corrections, and forms a feedforward-feedback composite control with proportional-integral-derivative (PID) systems. For example, the MPC algorithm has a 10-step prediction time domain, predicting the system response at the next 10 sampling times; a 5-step control time domain, optimizing the next 5 control actions; and a weight matrix Q of diag[1,0.5], used to balance the weights of different control objectives. Upon receiving the trigger signal, the model predictive control completes the optimization calculation within 30ms and outputs the control correction. The control correction includes online fine-tuning of the servo valve opening, injection / delivery speed, injection / holding pressure, switching point, or heating power.

[0082] For example, the servo valve has a response time of less than 10ms, ensuring that control commands can be quickly translated into physical actions and drive the actuator. The specific form of the correction amount can be a relative adjustment or an absolute setpoint, such as increasing the second-stage injection speed by 5% from the current level. The latency of the entire diagnostic-decision-execution closed loop is controlled at the millisecond level, meeting the stringent real-time requirements of the injection molding process.

[0083] In step S7, the entire process data of the current molding cycle, control corrections, and final product quality data are stored in the process big data platform. For example, after each molding cycle, the entire process data, including sensor time-series data and product quality data, is stored in the process big data platform. Product quality data, such as X-ray porosity <0.8%, is used. The process big data platform is based on the Hadoop architecture, which employs a distributed file system and distributed computing framework, offering advantages such as high scalability, high fault tolerance, and low-cost storage. The platform is configured with a storage capacity of 100TB, capable of supporting long-term, large-scale data accumulation.

[0084] Furthermore, the digital twin model undergoes incremental learning or parameter calibration based on accumulated data from the process big data platform. Incremental learning involves online local fine-tuning based on current production batch data. For example, for every 100 products produced, the system automatically calls upon the most recently accumulated 50GB of new data for incremental learning.

[0085] For the calibration of the physical mechanism model parameters, gradient descent was used for optimization. For example, the learning rate for the interface heat transfer coefficient was set to 0.01, the batch size to 32, and the deviation between the measured and predicted temperature fields in the new data was used as the loss function. After iterative optimization, the prediction error of the interface heat transfer coefficient decreased from ±8% before calibration to ±3%, significantly improving the prediction accuracy of the physical mechanism model for the temperature field.

[0086] For updating the weights of the neural network compensator in the data-driven compensation part, for example, the EWC algorithm is used for incremental updates. The EWC algorithm introduces a regularization term into the loss function to constrain the magnitude of changes in network weights that are important to historical tasks. This effectively prevents catastrophic forgetting while absorbing new data features, ensuring that the data-driven calibration model maintains its memory of historical process windows during long-term operation.

[0087] In step S8, when a preset cycle is completed or the accumulated data volume reaches a threshold, rolling re-optimization is triggered to generate an updated parameter set. A smooth switching strategy is then used to update the updated parameter set to the physical injection molding machine, continuously evolving the digital twin model. Incremental learning and smooth switching together constitute the two-layer dynamic update mechanism of the digital twin model. Each batch of production automatically absorbs new data for local parameter correction; every certain period, the system automatically re-optimizes, generating better parameters and applying them to production through "smooth switching." This ensures both the global accuracy of the digital twin model and its rapid adaptability to the latest operating conditions. After the digital twin model completes incremental learning, rolling re-optimization is triggered according to a preset cycle. Based on the evolved digital twin model, the multi-objective optimization of step S2 is re-executed to discover a better combination of process parameters. The preset cycle is triggered once every N pieces produced, where N is 50 to 5000. For example, every 500 products produced, the system automatically initiates a re-optimization process: Bayesian optimization is used to construct a Gaussian process model. The kernel function of the Gaussian process model is Matern 5 / 2, and the data acquisition function is EI to balance exploration and utilization. Based on statistical analysis of historical production data, the optimization parameter space is appropriately narrowed, focusing on local regions with a high probability of containing the optimal solution. For example, the initial temperature range of the semi-solid billet is narrowed from the original 480–520°C to 490–510°C, and other parameters are also narrowed accordingly based on the characteristics of accumulated data to improve optimization efficiency and convergence speed.

[0088] Furthermore, to ensure production stability during the transition from the current operating parameter set to the updated parameter set, a smooth switching strategy is adopted to distribute the new parameter set to the physical injection molding machine. The smooth switching strategy includes at least one of the following: limited amplitude and slope switching, segmented interpolation switching, or parallel verification switching of the old and new parameter sets. The parallel verification switching strategy refers to running digital twin simulations of the old and new parameter sets in parallel on a background virtual machine based on historical data, comparing their predicted stability and quality indicators, and confirming that the new parameter set has no potential risks before performing a one-time switch in the next production cycle. For example, the transition time for the smooth switching strategy is set to 5 minutes. During this period, the control system gradually adjusts each process parameter from the old value to the new value according to the limited amplitude and slope strategy, and adopts the updated parameter set. The production cycle time is shortened by 2.1% compared to before optimization.

[0089] The present invention also provides a specific embodiment to verify the feasibility of a method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters.

[0090] Example

[0091] Taking the outer shell of a large AZ91 magnesium alloy battery pack for a new energy vehicle as an example, the size of the magnesium alloy is 1500mm×800mm×2.5mm.

[0092] Phase 1: Deployment and Offline Optimization of the Digital Twin Model

[0093] A digital twin model was deployed on a cloud supercomputing platform equipped with an AMD EPYC CPU, and multi-objective optimization in step S2 was performed. The PPO algorithm from deep reinforcement learning was used to complete 800,000 virtual trials within 72 hours to search for the optimal solution in a high-dimensional process parameter space. The final output set of optimal process parameters was: semi-solid billet temperature 505°C, mold temperature 225°C, injection speed curve [65,95,115,105,85] mm / s, V / P switching point pressure 68.5 MPa, and holding time 18.2 s.

[0094] Phase Two: Initial Production and Real-Time Control

[0095] The optimal parameter set described above is sent to the physical injection molding machine with a single click via an intelligent control unit integrating an Intel Xeon CPU and FPGA real-time module, initiating the production of the first batch of 100 products. During production, multi-source sensors deployed in step S4 collect data in real time: 48 temperature and pressure nodes embedded in the mold cavity sample at 100Hz, a 5MHz ultrasonic sensor monitors the solid fraction online, a CCD machine vision system tracks the flow front at 60fps, and a servo motor encoder with 1μm resolution and a hydraulic sensor with an accuracy of ±0.5% synchronously provide feedback on the equipment status. All data is transmitted to the intelligent control unit via the PROFINET protocol with a delay of <1ms.

[0096] Step S5 performs deviation analysis between real-time data and theoretical prediction data, while step S6 involves the edge AI diagnostic model continuously monitoring the process status. When the 23rd product is produced, the edge AI diagnostic model detects a risk of filling imbalance in a thin-walled area and immediately triggers the MPC algorithm. The MPC completes rolling optimization within 30ms, outputs control correction values, and drives the servo valve to perform adjustments via the FPGA real-time module, effectively avoiding short-shot defects. After this batch of products is completed, the CPK value of the critical dimension is measured to be 1.8, demonstrating excellent process stability.

[0097] Phase 3: Model Calibration and Rolling Optimization

[0098] After producing 500 products, the process big data platform has accumulated comprehensive data, including sensor time-series data, target parameter sets, control corrections, and quality inspection results. Based on steps S7 and S8, the new data is used to incrementally learn the digital twin model: gradient descent is employed to calibrate physical mechanism parameters such as the interface heat transfer coefficient, reducing the prediction error of cooling time to ±3%; simultaneously, the EWC algorithm is used to update the weights of the neural network compensator to prevent catastrophic forgetting.

[0099] Based on this, the rolling optimization in step S2 is re-executed using the evolved digital twin model. The parameter space is appropriately narrowed, and a new parameter set is searched using Bayesian optimization. The updated parameter set is applied to production through a smooth switching strategy, reducing the cycle time from the initial 60s to 58.7s, thus improving production efficiency by 2.1%.

[0100] By applying the method of this embodiment throughout the entire process, the development cycle of the magnesium alloy battery pack shell is shortened by 85% compared with the traditional method. The product qualification rate is stable at over 99.5% during the mass production stage, and the unit energy consumption is reduced by 7%, which fully verifies the significant effect of this method in improving product quality, production efficiency and reducing energy consumption.

[0101] At the same time, such as Figure 3As shown, this invention provides a fully parameter self-optimizing system for one-click semi-solid injection molding of large and complex magnesium alloy components based on the above method, comprising:

[0102] The twin modeling and parameter calibration module is used to establish a digital twin model based on physical mechanisms and data-driven calibration models, and to perform inversion calibration on the parameters to be identified in the digital twin model using historical data.

[0103] The intelligent optimization module is used to perform multi-objective optimization in the digital twin model and output a set of candidate parameters.

[0104] The parameter loading and execution module is used to determine the target parameter set from the candidate parameter set, load it into the physical injection molding machine, and execute the molding cycle.

[0105] The multi-source sensing module is used to collect multi-dimensional time-series sensing data streams.

[0106] The real-time comparison and deviation calculation module is used to generate twin theory prediction data and compare it with the sensing data stream to output deviation indicators.

[0107] The edge intelligent diagnosis and control module is used to output process status identification results and defect risk prediction results. When the triggering conditions are met, it outputs control correction quantities based on the MPC-based adaptive control algorithm and drives the actuator.

[0108] The data platform and model learning module are used to store molding cycle data and incrementally learn or calibrate the digital twin model.

[0109] The model evolution and parameter update module is used to optimize and update the parameter set and perform a smooth switch when a preset cycle is completed or the accumulated data volume reaches a threshold, so as to continuously evolve the digital twin model.

[0110] In addition, this embodiment of the invention also provides an electronic device, which includes a processor, a memory, and a communication interface. The memory stores a computer program that can run on the processor. When the computer program is executed, it implements the above-mentioned method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters. The industrial communication interface is used to receive multi-source sensor data and output control correction values ​​to the actuator.

[0111] For example, a computer program can be divided into one or more modules, one or more of which are stored in memory and executed by a processor to perform the present invention. The one or more modules can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in an electronic device.

[0112] The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters, characterized in that: Includes the following steps: S1. Obtain target component information and equipment constraints, and construct a high-fidelity digital twin model; the digital twin model is a hybrid model combining a physical mechanism model and a data-driven calibration model; S2. Based on the multi-objective optimization objective function, optimize the high-dimensional process parameter space in the digital twin model to generate one or more sets of candidate parameters; S3. Select a target parameter set from the candidate parameter set, load it into the physical injection molding machine, and execute the molding cycle; S4. Acquire multi-dimensional time-series sensing data streams from multiple source sensors during the molding cycle; S5. The multi-dimensional time-series sensing data stream is compared in real time with the theoretical prediction data synchronously output by the digital twin model to calculate the residual and generate a deviation index; the deviation index includes at least one of temperature deviation, pressure deviation and filling speed deviation. S6. Input the deviation index into the edge artificial intelligence diagnostic model to output the process status identification result and defect risk prediction result, and call the adaptive control algorithm to calculate the control correction amount when the trigger condition is met, and drive the actuator to fine-tune the target parameter set online. S7. Store the full process data, control correction amount and final product quality data of the current molding cycle into the process big data platform, and perform incremental learning or parameter calibration on the digital twin model based on the accumulated data of the process big data platform. S8. When the preset cycle is completed or the accumulated data volume reaches the threshold, rolling re-optimization is triggered to generate an updated parameter set, and a smooth switching strategy is adopted to update the updated parameter set to the physical injection molding machine to continuously evolve the digital twin model.

2. The method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters according to claim 1, characterized in that, In step S1, the physical mechanism model is constructed based on the equations of conservation of mass, conservation of momentum and conservation of energy, and is coupled with a non-Newtonian fluid constitutive model and a phase change-solidification model. The data-driven calibration model is used to perform inversion calibration on the parameters to be identified in the physical mechanism model. The parameters to be identified include heat transfer coefficient, apparent viscosity model parameters or equipment dynamic parameters. The inversion calibration of the parameters to be identified adopts genetic algorithm, gradient descent method or Bayesian inversion method.

3. The method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters according to claim 1, characterized in that, In step S2, the high-dimensional process parameter space includes, but is not limited to: initial temperature of semi-solid billet, temperature of multiple zones of mold, screw speed and back pressure, multi-segment injection speed curve, injection to holding pressure switching point, injection pressure and holding pressure time; The optimization process shall employ at least one of the following methods: Deep reinforcement learning: The optimization process is constructed as a sequential decision-making process, wherein the state is composed of at least one of the temperature field, pressure field, flow front or solid fraction related features output by the digital twin model, the action is defined as the adjustment of the velocity curve, pressure curve, switching point or temperature setpoint, and the reward is calculated based on the multi-objective optimization objective function. Bayesian optimization: By constructing a probabilistic surrogate model of the objective function and using the acquisition function to guide the selection of sampling points, a set of candidate parameters can be obtained under a finite number of twin evaluations.

4. The method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters according to claim 3, characterized in that, The multi-objective optimization objective function includes at least one of a quality objective and an efficiency objective; the quality objective includes at least one of a defect index, a density index, a warpage amount, a flow / filling uniformity, or a dimensional deviation; the efficiency objective includes at least one of a cycle time, energy consumption, or material utilization rate.

5. The method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters according to claim 1, characterized in that, In step S4, the multi-source sensors include at least: temperature and pressure sensors embedded in the mold cavity, ultrasonic sensors for online detection of solid fraction, machine vision systems for monitoring the surface of components, and servo motor and hydraulic system sensors of the physical injection molding machine body.

6. The method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters according to claim 1, characterized in that, In step S6, the edge AI diagnostic model is a lightweight temporal learning model, which includes at least one of a one-dimensional convolutional neural network, a gated recurrent unit, or a temporal transformation model. It is used to extract features based on the multi-dimensional temporal sensing data stream and output process state identification results. The process state includes at least one of filling imbalance, stagnation, jetting, short-shot risk, or over-pressure risk. The adaptive control algorithm includes Model Predictive Control (MPC), which performs rolling optimization of the future control time domain based on a simplified predictive model to obtain control corrections, and forms a feedforward-feedback composite control with Proportional-Integral-Derivative (PID). The control corrections include online fine-tuning of servo valve opening, injection / delivery speed, injection / holding pressure, switching point, or heating power.

7. The method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters according to claim 1, characterized in that, In step S7, the incremental learning is based on online local fine-tuning of the current production batch data.

8. The method for one-click semi-solid injection molding of large and complex magnesium alloy components with fully optimized parameters according to claim 1, characterized in that, In step S8, the preset cycle is triggered once every N units produced, where N is 50 to 5000; the smooth switching strategy includes at least one of amplitude and slope limiting switching, segmented interpolation switching, or parallel verification switching of new and old parameter sets.

9. A self-optimizing system for one-click semi-solid injection molding of large and complex magnesium alloy components, used to implement the self-optimizing method for one-click semi-solid injection molding of large and complex magnesium alloy components according to any one of claims 1 to 8, characterized in that, include: The twin modeling and parameter calibration module is used to establish a digital twin model based on physical mechanisms and data-driven calibration models, and to perform inversion calibration on the parameters to be identified in the digital twin model using historical data. The intelligent optimization module is used to perform multi-objective optimization in the digital twin model and output a set of candidate parameters; The parameter loading and execution module is used to determine the target parameter set from the candidate parameter set, load it into the physical injection molding machine, and execute the molding cycle. The multi-source sensing module is used to acquire multi-dimensional time-series sensing data streams; The real-time comparison and deviation calculation module is used to generate twin theory prediction data and compare it with the sensing data stream to output deviation indicators; The edge intelligent diagnosis and control module is used to output process status identification results and defect risk prediction results. When the triggering conditions are met, it outputs control correction quantities based on the MPC-based adaptive control algorithm and drives the actuator. The data platform and model learning module are used to store molding cycle data and incrementally learn or calibrate the digital twin model; The model evolution and parameter update module is used to optimize and update the parameter set and perform a smooth switch when a preset cycle is completed or the accumulated data volume reaches a threshold, so as to continuously evolve the digital twin model.

10. An electronic device, characterized in that, The device includes a processor, a memory, and an industrial communication interface. The memory stores a computer program that can run on the processor. When the computer program is executed, it implements the full parameter self-optimization method for semi-solid injection molding of large and complex magnesium alloy components as described in any one of claims 1 to 8. The industrial communication interface is used to receive multi-source sensor data and output control correction values ​​to the actuator.

Images