Magnesium-aluminum alloy semi-solid slurry rheoforming method and device

Abstract

The invention relates to the technical field of alloy forming, and discloses a magnesium-aluminum alloy semi-solid slurry rheoforming method and device. Cooling the alloy liquid to a semi-solid state interval through PID closed-loop temperature control, and synchronously executing a closed-loop regulation process of multi-sensor data acquisition, CNN-LSTM model prediction and GA-fuzzy PID parameter optimization until spherical / nearly spherical grain slurry is formed; after core sensitive parameters are screened through a GPR parameter mapping model, an injection speed curve is dynamically adjusted through a DDPG reinforcement learning algorithm, slurry is conveyed to a mold cavity along the optimized injection speed curve, and after pressure maintaining, the temperature is lowered to 150 DEG C or below through a cooling water channel arranged in a mold; and post-treatment and quality control. Through the full-chain technical innovation of composite shearing and stirring, multi-algorithm fusion closed-loop regulation and control, high-precision simulation and dynamic mold filling control, the technical breakthrough of magnesium-aluminum alloy semi-solid rheoforming is realized.

Description

Technical Field

[0001] This invention relates to the field of alloy forming technology, and more specifically, to a method and apparatus for rheoforming a semi-solid slurry of magnesium-aluminum alloy. Background Technology

[0002] Magnesium-aluminum alloy semi-solid rheoforming technology, with its advantages of high casting density, excellent mechanical properties, and short forming cycle, has become a key technology for manufacturing complex structural parts in the automotive, aerospace, and other fields. Its core principle is to utilize the solid particles formed in the solid-liquid coexistence region of the alloy to mix with the liquid metal slurry, achieving precise forming through rheological filling. Compared to traditional liquid die casting, it can significantly reduce defects such as shrinkage cavities and cracks, and lower energy consumption and subsequent processing costs.

[0003] In existing semi-solid rheoforming technologies, the industry generally uses a single mechanical or electromagnetic stirring method to prepare the slurry, combined with simple PID temperature control and fixed-parameter filling processes. Specifically, the slurry preparation stage relies heavily on experience to set parameters such as stirring speed and cooling rate, lacking real-time monitoring and dynamic adjustment of solid fraction and grain morphology. Furthermore, existing technologies suffer from a disconnect between sensor data acquisition and algorithm control, failing to form a closed-loop system of data acquisition, quality prediction, parameter optimization, and execution feedback. This makes it difficult to adapt to the non-Newtonian fluid characteristics of magnesium-aluminum alloy semi-solid slurries and the complexity of the molding process. Therefore, there is an urgent need to design a method and apparatus for rheoforming magnesium-aluminum alloy semi-solid slurries to address these issues. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and apparatus for semi-solid slurry rheoforming of magnesium-aluminum alloys. Through a full-chain technological innovation involving composite shearing and stirring, multi-algorithm fusion closed-loop control, high-precision simulation, and dynamic filling control, a technological breakthrough in semi-solid rheoforming of magnesium-aluminum alloys has been achieved, providing a reliable solution for the large-scale production of high-quality magnesium-aluminum alloy complex structural parts.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A method for rheoforming semi-solid slurry of magnesium-aluminum alloy includes the following steps:

[0007] S1: Raw material pretreatment and melting: Alloy ingots with surface oxide scale removed according to the composition ratio are heated to 50-80°C above the liquidus after argon gas is introduced to isolate air. After refining, degassing and slag removal, the temperature is held for 15-20 minutes to obtain pure alloy liquid.

[0008] S2: Semi-solid slurry preparation: The alloy liquid is cooled to the semi-solid range at 5-8℃ / min by PID closed-loop temperature control. A composite shear stirring system consisting of mechanical stirring component, electromagnetic stirring component and ultrasonic stirring component is started. Simultaneously, a closed-loop control process of multi-sensor data acquisition, CNN-LSTM model prediction and GA-fuzzy PID parameter optimization is executed until spherical / near-spherical grain slurry is formed.

[0009] S3: Rheological filling and molding. A flow field-temperature field coupled simulation model is constructed based on the improved Bingham-Casson viscosity model. After screening the core sensitive parameters through the GPR parameter mapping model, the injection speed curve is dynamically adjusted through the DDPG reinforcement learning algorithm. The optimized injection speed signal is transmitted to the servo-driven injection system. The servo-driven injection system delivers the slurry to the mold cavity at a preset temperature of 180-220℃ along the optimized injection speed curve according to the set filling pressure. After holding the pressure for 10-25s, the temperature is cooled to below 150℃ through the mold's built-in cooling water channel.

[0010] S4: Post-processing and quality control: eject the casting and remove the gating system and burrs. After ultrasonic testing, qualified castings are aged according to specifications, while unqualified castings are broken and remelted for recycling.

[0011] As a preferred embodiment of the present invention, the starting method of the composite shear stirring system in S2 is as follows: first, the mechanical stirring component and the electromagnetic stirring component are started; when the alloy liquid is cooled to the upper limit of the semi-solid range, the ultrasonic stirring component is started. The three work together to achieve grain breakage and spheroidization. The stirring paddle of the mechanical stirring component has a double helix structure, the electromagnetic stirring component has a circumferential coil structure, and the transducer of the ultrasonic stirring component is attached to the outer wall of the crucible.

[0012] As a preferred embodiment of the present invention, the implementation method of the multi-sensor data acquisition, CNN-LSTM model prediction, and GA-fuzzy PID parameter optimization closed-loop control process in S2 is as follows:

[0013] S21. Multi-sensor data acquisition: Through integrated infrared thermometer, ultrasonic flaw detector, and laser particle size analyzer, 12 key data dimensions such as temperature, solidity, and grain size are acquired at a frequency of 1 time / second to construct an input data matrix.

[0014]

[0015] in, For temperature data, For solid fraction data, For grain size data, For the remaining 9 dimensions of auxiliary process parameter data, This refers to the number of data collection sessions.

[0016] S22, CNN-LSTM model prediction: The data matrix... The CNN-LSTM model is input in real time to extract features and identify the slurry quality status, and output a quality grade signal.

[0017] S23, GA - Fuzzy PID parameter optimization: Receives quality level signals and dynamically adjusts ultrasonic stirring power, cooling rate, and mechanical stirring speed to form a closed-loop feedback.

[0018] As a preferred embodiment of the present invention, the specific implementation of the CNN-LSTM model in S22 is as follows:

[0019] S22a, CNN Feature Extraction: Texture features are extracted from the grain images acquired by the laser grain size analyzer using three convolutional layers. The kernel sizes are 3×3, 5×5, and 3×3, respectively. The ReLU activation function is used, and the expression is as follows:

[0020]

[0021] Output 64-dimensional feature vector ;

[0022] S22b, LSTM time series modeling: combining temperature time series data with feature vectors After concatenation, the data is input into a 4-layer LSTM network. Temporal correlation is handled through a gating mechanism. The core expression is:

[0023]

[0024] in, For input gate output, Output for the forget gate For output gate output, For cell state, For output of the hidden layer, This is the weight matrix. For bias vectors, For the sigmoid function, For Hadamah accumulation;

[0025] S22c, the output layer outputs quality level signals through the Softmax function, including: qualified, grain agglomeration, and abnormal solidity.

[0026] As a preferred embodiment of the present invention, the specific implementation of the GA-fuzzy PID parameter optimization in S23 is as follows:

[0027] S23a. The genetic algorithm constructs a fitness function with the objective of minimizing the standard deviation of the solid fraction:

[0028]

[0029] In the formula, For the first Solid fraction of the second sample. The average solid fraction This refers to the number of data collections.

[0030] S23b, The optimized variable is ultrasonic stirring power. Cooling rate The number of iterations is ≥50, the crossover probability is 0.6-0.8, the mutation probability is 0.01-0.03, and the optimal parameter combination is output and fed back to the ultrasonic stirring component and temperature control system.

[0031] S23c, fuzzy PID controller with solidity deviation Deviation change rate Input, output stirring speed correction amount PID parameters are dynamically adjusted using 25 fuzzy rules. , , And feed it back to the mechanical stirring component.

[0032] As a preferred embodiment of the present invention, the improved Bingham-Casson viscosity model described in S3 is used to improve the accuracy of flow-temperature field coupled simulation, and its expression is:

[0033]

[0034] in, For slurry viscosity, Zero shear viscosity For yield stress, Shear rate, The critical shear rate. This is the consistency coefficient. This is a liquidity behavior index.

[0035] As a preferred embodiment of the present invention, the GPR parameter mapping model in step S3 is based on simulation data of the improved Bingham-Casson viscosity model to screen key filling parameters, specifically as follows:

[0036] By pouring temperature Injection speed For input vectors With defect rate To output, construct a Gaussian process regression model:

[0037]

[0038] In the formula, the mean function The kernel function uses a squared exponential kernel:

[0039]

[0040] in, For signal variance, For length scale, For noise variance, Given the Dirac function, the kernel function parameters are solved by maximum likelihood estimation, and the casting temperature weight is determined to be >0.6.

[0041] As a preferred embodiment of the present invention, the optimization logic of the DDPG reinforcement learning algorithm in S3 is as follows: receiving the key parameter weights output by the GPR parameter mapping model, and dynamically optimizing the injection speed curve with the goal of minimizing the defect rate, specifically implemented as follows:

[0042] 1) State Space ,in For the real-time temperature of the slurry, For mold cavity pressure, For the filling progress;

[0043] 2) Action Space , For the number of segments in the filling type, For the injection speed adjustment, the motion constraint is 0.1 ≤ ≤0.5m / s;

[0044] 3) The reward function expression is:

[0045]

[0046] In the formula, Basic rewards, , The penalty coefficient is... The optimal injection velocity is output from the flow field simulation.

[0047] 4) An experience replay mechanism and target network soft update are adopted. Both the Actor / Critic network contain 3 fully connected layers, which improves the algorithm convergence speed by 40% and outputs an optimized injection velocity curve signal.

[0048] As a preferred embodiment of the present invention, the execution logic of the servo-driven injection system in S3 is as follows: receiving the injection speed curve signal output by the DDPG reinforcement learning algorithm, and combining it with the set filling pressure, the heat-insulating screw conveyor is driven to stably transport the slurry to the mold cavity through the pressure-speed coordinated control strategy. During the transportation process, the slurry temperature is maintained at the lower limit of the semi-solid range +5℃, and the transportation rate fluctuation is ≤±5%.

[0049] A magnesium-aluminum alloy semi-solid slurry rheoforming device includes a raw material pretreatment unit, a slurry preparation unit, a precise temperature control unit, a conveying and forming unit, a quality control and recycling unit, and an intelligent control module. The intelligent control module is connected to the other units by signals and incorporates a CNN-LSTM prediction model, a GA-fuzzy PID hybrid optimization framework, a GPR parameter mapping model, and a DDPG reinforcement learning algorithm.

[0050] The slurry preparation unit includes a multi-sensor array and a composite shear stirring system. The multi-sensor array consists of an infrared thermometer, an ultrasonic flaw detector, and a laser particle size analyzer, with its data output terminal connected to the CNN-LSTM prediction model data input terminal of the intelligent control module. The composite shear stirring system includes a double-helix mechanical stirring component, a coil-type electromagnetic stirring component, and an ultrasonic stirring component attached to the outer wall of the crucible. The control input terminals of all three components are connected to the GA-fuzzy PID hybrid optimization framework signal output terminal of the intelligent control module.

[0051] The conveying and forming unit includes an insulated screw conveyor, a servo-driven injection system, a mold, and a vision recognition module. The control input of the servo-driven injection system is connected to the signal output of the DDPG reinforcement learning algorithm of the intelligent control module, integrating a pressure-speed co-controller with an injection speed adjustment error of <±0.02m / s and a pressure control error of <±0.05MPa. The mold has built-in cooling water channels and a cavity pressure sensor, and its pressure signal output is connected to the state space data acquisition terminal of the intelligent control module. The filling progress signal output of the vision recognition module is connected to the state space data acquisition terminal of the intelligent control module.

[0052] The intelligent control module includes a PLC controller, an algorithm execution server, and a data interaction interface. The signal acquisition terminal of the PLC controller is connected to the sensor signal output terminal of each unit, and the control signal output terminal is connected to the actuator of each unit. The algorithm execution server deploys a CNN-LSTM prediction model, a GA-fuzzy PID hybrid optimization framework, a GPR parameter mapping model, and a DDPG reinforcement learning algorithm. It realizes real-time transmission of model input and output data through the data interaction interface to complete closed-loop control.

[0053] The beneficial technical effects of this invention are:

[0054] This invention achieves real-time monitoring and dynamic optimization of slurry quality through the synergistic effect of a mechanical, electromagnetic, and ultrasonic composite shearing and stirring system, combined with a closed-loop control process involving multi-sensor data acquisition, CNN-LSTM model prediction, and GA-fuzzy PID parameter optimization. The CNN-LSTM can accurately identify defects such as grain agglomeration and abnormal solid fraction; the GA-fuzzy PID framework aims to minimize the standard deviation of solid fraction, dynamically adjusting ultrasonic stirring power, cooling rate, and mechanical stirring speed, resulting in minimal stirring speed control error.

[0055] This invention constructs a coupled flow-temperature field simulation model based on an improved Bingham-Casson viscosity model. It then uses a GPR parameter mapping model to select core sensitive parameters such as pouring temperature, and dynamically optimizes the injection speed curve using the DDPG reinforcement learning algorithm. The DDPG algorithm uses the lowest defect rate as the reward function and responds in real-time to changes in slurry temperature, cavity pressure, and filling progress, resulting in minimal errors in injection speed adjustment and filling pressure control. The final casting exhibits extremely low porosity and incomplete filling rate for complex structures, improving fatigue life and eliminating the need for secondary processing or repair.

[0056] This invention integrates multi-dimensional intelligent algorithms to form a closed-loop control system for the entire process. Multiple sensors collect 12-dimensional key data at a frequency of once per second, providing comprehensive support for the algorithms; the CNN-LSTM model captures spatiotemporal coupling relationships, GA-fuzzy PID achieves fine-grained parameter optimization, and the DDPG algorithm enhances the dynamic response capability of the filling process. Compared with existing technologies, the parameter control accuracy is significantly improved, it can adapt to different alloy types and casting structures, and maintains the optimal forming state without manual intervention. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0058] Figure 2 This is a schematic diagram of the device architecture of the present invention. Detailed Implementation

[0059] In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0060] Combination Figure 1 The present invention provides the following embodiments:

[0061] A method for rheoforming semi-solid slurry of magnesium-aluminum alloy includes the following steps:

[0062] S1: Raw material pretreatment and melting: The alloy ingot with surface oxide scale removed according to the composition ratio is purged with argon gas to isolate it from air, with an argon gas flow rate of 0.3-0.5 m³ / h, and heated to 50-80℃ above the liquidus line. After refining, degassing and slag removal, it is held at the temperature for 15-20 minutes to obtain pure alloy liquid.

[0063] S2: Semi-solid slurry preparation: The alloy liquid is cooled to the semi-solid range at a rate of 5-8℃ / min using PID closed-loop temperature control. Taking AZ91 magnesium alloy (580-620℃) and A356 aluminum alloy (600-640℃) as examples, a composite shear stirring system consisting of mechanical stirring, electromagnetic stirring, and ultrasonic stirring components is started. Simultaneously, a closed-loop control process of multi-sensor data acquisition, CNN-LSTM model prediction, and GA-fuzzy PID parameter optimization is executed until spherical / near-spherical grain slurry is formed. The solid phase of the slurry is 40%-50% for AZ91 magnesium alloy and 35%-45% for A356 aluminum alloy.

[0064] The implementation method of the multi-sensor data acquisition, CNN-LSTM model prediction, and GA-fuzzy PID parameter optimization closed-loop control process is as follows:

[0065] S21. Multi-sensor data acquisition: Through integrated infrared thermometer, ultrasonic flaw detector, and laser particle size analyzer, 12 key data dimensions such as temperature, solidity, and grain size are acquired at a frequency of 1 time / second to construct an input data matrix.

[0066]

[0067] in, For temperature data, For solid fraction data, For grain size data, For the remaining 9 dimensions of auxiliary process parameter data, This represents the number of times data was collected.

[0068] S22, CNN-LSTM model prediction: The data matrix... The CNN-LSTM model is input in real time to extract features and identify the slurry quality status, and output a quality grade signal.

[0069] The specific implementation of the CNN-LSTM model is as follows:

[0070] S22a, CNN Feature Extraction: Texture features are extracted from the grain images acquired by the laser grain size analyzer using three convolutional layers. The kernel sizes are 3×3, 5×5, and 3×3, respectively. The ReLU activation function is used, and the expression is as follows:

[0071]

[0072] Output 64-dimensional feature vector ;

[0073] S22b, LSTM time series modeling: combining temperature time series data with feature vectors After concatenation, the input is a 4-layer LSTM network with 128 hidden neurons. Temporal correlations are handled through a gating mechanism. The core expression is:

[0074]

[0075] in, For input gate output, Output for the forget gate For output gate output, For cell state, For output of the hidden layer, This is the weight matrix. For bias vectors, For the sigmoid function, For Hadamah accumulation;

[0076] S22c, the output layer outputs quality level signals through the Softmax function. The quality level signals include qualified, grain agglomeration, and abnormal solidity.

[0077] Three convolutional layers combined with a batch normalization (BN) layer can progressively extract grain features from local edges to global texture. The BN layer accelerates model convergence and suppresses overfitting. The max pooling layer reduces feature dimensionality while preserving key texture information, such as the circular outline of spherical grains and the irregular blocky features of clustered grains. The four-layer LSTM network, through the synergy of forget gates, input gates, and output gates, can effectively memorize long-term temperature change trends and avoid short-term temperature noise interfering with quality judgment. The probability threshold setting of the Softmax function can balance the false positive rate and the false negative rate, ensuring that no serious defects are missed.

[0078] S23, GA - Fuzzy PID parameter optimization: Receives quality level signals and dynamically adjusts ultrasonic stirring power, cooling rate, and mechanical stirring speed to form a closed-loop feedback.

[0079] The specific implementation of GA-fuzzy PID parameter optimization is as follows:

[0080] S23a. The genetic algorithm constructs a fitness function with the objective of minimizing the standard deviation of the solid fraction:

[0081]

[0082] In the formula, For the first Solid fraction of the second sample. The average solid fraction This refers to the number of data collections.

[0083] S23b, The optimized variable is ultrasonic stirring power. , The value range can be 200≤ ≤800W, cooling rate , The range of values ​​is 1≤ ≤5℃ / s, iterations ≥50 generations, crossover probability 0.6-0.8, mutation probability 0.01-0.03, output the optimal parameter combination and feed it back to the ultrasonic stirring component and temperature control system;

[0084] S23c, fuzzy PID controller with solidity deviation Deviation change rate Input, output stirring speed correction amount PID parameters are dynamically adjusted using 25 fuzzy rules. , , And feed it back to the mechanical stirring component.

[0085] S3: Rheological filling and molding. A flow-temperature field coupled simulation model is constructed based on the improved Bingham-Casson viscosity model. After screening the core sensitive parameters through the GPR parameter mapping model, the injection speed curve is dynamically adjusted through the DDPG reinforcement learning algorithm. The optimized injection speed signal is transmitted to the servo-driven injection system. The servo-driven injection system delivers the slurry to the mold cavity at a preset temperature of 180-220℃ along the optimized injection speed curve according to the set filling pressure, which is 1.0-2.0MPa for AZ91 magnesium alloy and 0.8-1.8MPa for A356 aluminum alloy. After holding the pressure for 10-25s, the temperature is cooled to below 150℃ through the mold's built-in cooling water channel.

[0086] The improved Bingham-Casson viscosity model is used to enhance the accuracy of flow-temperature field coupled simulations. The expression is as follows:

[0087]

[0088] in, For slurry viscosity, Zero shear viscosity For yield stress, Shear rate, The critical shear rate. This is the consistency coefficient. For liquidity behavior index, 0.8 ≤ ≤1.2.

[0089] The traditional Bingham model only considers yield flow and cannot describe the shear-thinning characteristics of semi-solid slurries at high shear rates. The improved model, however, uses piecewise functions to... At that time, the yield-dominant flow is described by zero shear viscosity plus yield stress / shear rate; in At that time, additional power-law terms are added. It accurately captures the shear-thinning effect.

[0090] The GPR parameter mapping model, based on simulation data from the improved Bingham-Casson viscosity model, selects key filling parameters, specifically:

[0091] By pouring temperature Injection speed , is the input vector With defect rate To output, construct a Gaussian process regression model:

[0092]

[0093] In the formula, the mean function The kernel function uses a squared exponential kernel:

[0094]

[0095] Among them, 620≤ ≤680℃, 0.1≤ ≤0.5m / s; For signal variance, For length scale, For noise variance, Given the Dirac function, the kernel function parameters are solved by maximum likelihood estimation, and the casting temperature weight is determined to be >0.6.

[0096] The optimization logic of the DDPG reinforcement learning algorithm is as follows: It receives the key parameter weights output by the GPR parameter mapping model, and dynamically optimizes the injection velocity curve with the goal of minimizing the defect rate. Specifically, this is implemented as follows:

[0097] 1) State Space ,in For the real-time temperature of the slurry, For mold cavity pressure, For filling progress, 0 ≤ ≤1;

[0098] 2) Action Space , For the number of segments in the filling type, For the injection speed adjustment, the motion constraint is 0.1 ≤ ≤0.5m / s;

[0099] 3) The reward function expression is:

[0100]

[0101] In the formula, Basic rewards, , The penalty coefficient is... The optimal injection velocity is output from the flow field simulation.

[0102] 4) Employ an experience replay mechanism and target network soft updates, with an update rate of... Both the Actor / Critic network contain 3 fully connected layers, improving the algorithm's convergence speed by 40% and outputting an optimized injection velocity curve signal.

[0103] S4: Post-processing and quality control: Eject castings and remove gates, risers, and burrs. After ultrasonic testing, qualified castings are aged according to specifications: AZ91 magnesium alloy 160℃ / 4h, A356 aluminum alloy 180℃ / 6h. Unqualified castings are broken and remelted for recycling.

[0104] Furthermore, the startup method of the composite shear stirring system described in S2 is as follows: first, the mechanical stirring component and the electromagnetic stirring component are started. Once the alloy liquid cools to the upper limit of the semi-solid range, the ultrasonic stirring component is started. The three components work together to achieve grain breakage and spheroidization. The mechanical stirring component has a double-helix structure for its stirring paddle, the electromagnetic stirring component has a coiled structure, and the transducer of the ultrasonic stirring component is attached to the outer wall of the crucible. Specifically, starting the mechanical and electromagnetic stirring first allows the mechanical shear force to initially break the growth trend of dendritic grains during the initial transition of the alloy liquid from a fully liquid to a semi-solid state. Simultaneously, the annular magnetic field formed by the electromagnetic stirring drives slurry convection, preventing localized temperature buildup. Once the temperature drops to the upper limit of the semi-solid range, a small number of solid particles have formed in the slurry. At this point, ultrasonic stirring is started. The high-frequency vibration energy can be precisely applied to the surface of the solid particles, further refining the grains and promoting spheroidization. This avoids energy waste due to starting the ultrasonic stirring too early or ineffective grain morphology correction due to starting it too late. Composite stirring increases the proportion of spherical grains and reduces grain size uniformity errors, laying the foundation for subsequent filling stability.

[0105] Furthermore, the execution logic of the servo-driven injection system described in S3 is as follows: receiving the injection speed curve signal output by the DDPG reinforcement learning algorithm, and combining it with the set filling pressure (1.0-2.0MPa for AZ91 magnesium alloy and 0.8-1.8MPa for A356 aluminum alloy), the system drives the heat-insulating screw conveyor to stably transport the slurry to the mold cavity through a pressure-speed coordinated control strategy. During the transportation process, the slurry temperature is maintained at the lower limit of the semi-solid range +5℃, and the transportation rate fluctuation is ≤±5%.

[0106] The servo-driven injection system uses a permanent magnet synchronous servo motor, paired with a ball screw transmission mechanism; the pressure-speed coordinated control adopts a cross-coupled PID algorithm, with speed control as the main control and pressure control as the auxiliary control. When the actual pressure exceeds the set value by 10%, it automatically switches to pressure priority control.

[0107] A magnesium-aluminum alloy semi-solid slurry rheoforming device includes a raw material pretreatment unit, a slurry preparation unit, a precise temperature control unit, a conveying and forming unit, a quality control and recycling unit, and an intelligent control module. The intelligent control module is connected to the other units by signals and incorporates a CNN-LSTM prediction model, a GA-fuzzy PID hybrid optimization framework, a GPR parameter mapping model, and a DDPG reinforcement learning algorithm.

[0108] The slurry preparation unit includes a multi-sensor array and a composite shear stirring system. The multi-sensor array consists of an infrared thermometer, an ultrasonic flaw detector, and a laser particle size analyzer, with its data output terminal connected to the CNN-LSTM prediction model data input terminal of the intelligent control module. The composite shear stirring system includes a double-helix mechanical stirring component, a coil-type electromagnetic stirring component, and an ultrasonic stirring component attached to the outer wall of the crucible. The control input terminals of all three components are connected to the GA-fuzzy PID hybrid optimization framework signal output terminal of the intelligent control module.

[0109] The conveying and forming unit includes an insulated screw conveyor lined with graphite and an insulation layer thickness of 50mm; it also includes a servo-driven injection system, a mold, and a vision recognition module; the control input of the servo-driven injection system is connected to the signal output of the DDPG reinforcement learning algorithm of the intelligent control module, integrating a pressure-speed co-controller with an injection speed adjustment error of <±0.02m / s and a pressure control error of <±0.05MPa; the mold has built-in cooling water channels and a cavity pressure sensor, and its pressure signal output is connected to the state space data acquisition terminal of the intelligent control module; the filling progress signal output of the vision recognition module is connected to the state space data acquisition terminal of the intelligent control module.

[0110] The intelligent control module includes a PLC controller, an algorithm execution server, and a data interaction interface. The signal acquisition terminal of the PLC controller is connected to the sensor signal output terminal of each unit, and the control signal output terminal is connected to the actuator of each unit. The algorithm execution server deploys a CNN-LSTM prediction model, a GA-fuzzy PID hybrid optimization framework, a GPR parameter mapping model, and a DDPG reinforcement learning algorithm. It realizes real-time transmission of model input and output data through the data interaction interface to complete closed-loop control.

[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for rheoforming semi-solid slurry of magnesium-aluminum alloy, characterized in that, Includes the following steps: S1: Raw material pretreatment and melting: Alloy ingots with surface oxide scale removed according to the composition ratio are heated to 50-80°C above the liquidus after argon gas is introduced to isolate air. After refining, degassing and slag removal, the temperature is held for 15-20 minutes to obtain pure alloy liquid. S2: Semi-solid slurry preparation: The alloy liquid is cooled to the semi-solid range at 5-8℃ / min by PID closed-loop temperature control. A composite shear stirring system consisting of mechanical stirring component, electromagnetic stirring component and ultrasonic stirring component is started. Simultaneously, a closed-loop control process of multi-sensor data acquisition, CNN-LSTM model prediction and GA-fuzzy PID parameter optimization is executed until spherical / near-spherical grain slurry is formed. S3: Rheological filling and molding. A flow field-temperature field coupled simulation model is constructed based on the improved Bingham-Casson viscosity model. After screening the core sensitive parameters through the GPR parameter mapping model, the injection speed curve is dynamically adjusted through the DDPG reinforcement learning algorithm. The optimized injection speed signal is transmitted to the servo-driven injection system. The servo-driven injection system delivers the slurry to the mold cavity at a preset temperature of 180-220℃ along the optimized injection speed curve according to the set filling pressure. After holding the pressure for 10-25s, the temperature is cooled to below 150℃ through the mold's built-in cooling water channel. S4: Post-processing and quality control: eject the casting and remove the gating system and burrs. After ultrasonic testing, qualified castings are aged according to specifications, while unqualified castings are broken and remelted for recycling.

2. The method for rheoforming semi-solid slurry of magnesium-aluminum alloy according to claim 1, characterized in that, The startup method of the composite shear stirring system described in S2 is as follows: first, start the mechanical stirring component and the electromagnetic stirring component. When the alloy liquid cools down to the upper limit of the semi-solid range, start the ultrasonic stirring component. The three work together to achieve grain breakage and spheroidization. The stirring paddle of the mechanical stirring component has a double helix structure, the electromagnetic stirring component has a circumferential coil structure, and the transducer of the ultrasonic stirring component is attached to the outer wall of the crucible.

3. The method for rheoforming semi-solid slurry of magnesium-aluminum alloy according to claim 2, characterized in that, The implementation method of the multi-sensor data acquisition, CNN-LSTM model prediction, and GA-fuzzy PID parameter optimization closed-loop control process described in S2 is as follows: S21. Multi-sensor data acquisition: Through integrated infrared thermometer, ultrasonic flaw detector, and laser particle size analyzer, 12 key data dimensions such as temperature, solidity, and grain size are acquired at a frequency of 1 time / second to construct an input data matrix. in, For temperature data, For solid fraction data, For grain size data, For the remaining 9 dimensions of auxiliary process parameter data, This refers to the number of data collection sessions. S22, CNN-LSTM model prediction: The data matrix... The CNN-LSTM model is input in real time to extract features and identify the slurry quality status, and output a quality grade signal. S23, GA - Fuzzy PID Parameter Optimization: Receives quality level signals and dynamically adjusts ultrasonic stirring power, cooling rate, and mechanical stirring speed to form a closed-loop feedback.

4. The method for rheoforming semi-solid slurry of magnesium-aluminum alloy according to claim 1, characterized in that, The specific implementation of the CNN-LSTM model described in S22 is as follows: S22a, CNN Feature Extraction: Texture features are extracted from the grain images acquired by the laser grain size analyzer using three convolutional layers. The kernel sizes are 3×3, 5×5, and 3×3, respectively. The ReLU activation function is used, and the expression is as follows: Output 64-dimensional feature vector ; S22b, LSTM time series modeling: combining temperature time series data with feature vectors After concatenation, the data is input into a 4-layer LSTM network. Temporal correlation is handled through a gating mechanism. The core expression is: in, For input gate output, Output for the forget gate For output gate output, For cell state, For output of the hidden layer, This is the weight matrix. For bias vectors, For the sigmoid function, For Hadamah accumulation; S22c, the output layer outputs quality level signals through the Softmax function, including: qualified, grain agglomeration, and abnormal solidity.

5. The method for rheoforming semi-solid slurry of magnesium-aluminum alloy according to claim 1, characterized in that, The specific implementation of the GA-fuzzy PID parameter optimization described in S23 is as follows: S23a. The genetic algorithm aims to minimize the standard deviation of the solid fraction, and constructs the fitness function as follows: In the formula, For the first Solid fraction of the second sample. The average solid fraction This represents the number of data collections. S23b, The optimized variable is ultrasonic stirring power. Cooling rate The number of iterations is ≥50, the crossover probability is 0.6-0.8, the mutation probability is 0.01-0.03, and the optimal parameter combination is output and fed back to the ultrasonic stirring component and temperature control system. S23c, fuzzy PID controller with solidity deviation Deviation change rate Input, output stirring speed correction amount PID parameters are dynamically adjusted using 25 fuzzy rules. , , And feed it back to the mechanical stirring component.

6. The method for rheoforming semi-solid slurry of magnesium-aluminum alloy according to claim 1, characterized in that, The improved Bingham-Casson viscosity model described in S3 is used to enhance the accuracy of flow-temperature field coupled simulations, and its expression is: in, For slurry viscosity, Zero shear viscosity For yield stress, Shear rate, The critical shear rate. This is the consistency coefficient. This is a liquidity behavior index.

7. The method for rheoforming semi-solid slurry of magnesium-aluminum alloy according to claim 1, characterized in that, The GPR parameter mapping model described in step S3 is based on simulation data from the improved Bingham-Casson viscosity model, and selects key filling parameters, specifically: By pouring temperature Injection speed For input vectors With defect rate To output, construct a Gaussian process regression model: In the formula, the mean function The kernel function uses a squared exponential kernel: in, For signal variance, For length scale, For noise variance, Given the Dirac function, the kernel function parameters are solved by maximum likelihood estimation, and the casting temperature weight is determined to be >0.

6.

8. The method for rheoforming semi-solid slurry of magnesium-aluminum alloy according to claim 7, characterized in that, The optimization logic of the DDPG reinforcement learning algorithm described in S3 is as follows: It receives the key parameter weights output by the GPR parameter mapping model, and dynamically optimizes the injection speed curve with the goal of minimizing the defect rate. Specifically, this is implemented as follows: 1) State Space ,in For the real-time temperature of the slurry, For mold cavity pressure, For the filling progress; 2) Action Space , For the number of segments in the filling type, For the injection speed adjustment, the motion constraint is 0.1 ≤ ≤0.5m / s; 3) The reward function expression is: In the formula, Basic rewards, , The penalty coefficient is... The optimal injection velocity is output from the flow field simulation. 4) An experience replay mechanism and target network soft update are adopted. Both the Actor / Critic network contain 3 fully connected layers, which improves the algorithm convergence speed by 40% and outputs an optimized injection velocity curve signal.

9. The method for rheoforming semi-solid slurry of magnesium-aluminum alloy according to claim 1, characterized in that, The execution logic of the servo-driven injection system described in S3 is as follows: receiving the injection speed curve signal output by the DDPG reinforcement learning algorithm, and combining it with the set filling pressure, the system drives the heat-insulating screw conveyor to stably transport the slurry to the mold cavity through a pressure-speed coordinated control strategy. During the transportation process, the slurry temperature is maintained at the lower limit of the semi-solid range +5℃, and the transportation rate fluctuation is ≤±5%.

10. A magnesium-aluminum alloy semi-solid slurry rheoforming device, characterized in that, It includes a raw material pretreatment unit, a slurry preparation unit, a precise temperature control unit, a conveying and molding unit, a quality control and recycling unit, and an intelligent control module. The intelligent control module is connected to the signals of the other units and has built-in CNN-LSTM prediction model, GA-fuzzy PID hybrid optimization framework, GPR parameter mapping model, and DDPG reinforcement learning algorithm. The slurry preparation unit includes a multi-sensor array and a composite shear stirring system. The multi-sensor array consists of an infrared thermometer, an ultrasonic flaw detector, and a laser particle size analyzer, with its data output terminal connected to the CNN-LSTM prediction model data input terminal of the intelligent control module. The composite shear stirring system includes a double-helix mechanical stirring component, a coil-type electromagnetic stirring component, and an ultrasonic stirring component attached to the outer wall of the crucible. The control input terminals of all three components are connected to the GA-fuzzy PID hybrid optimization framework signal output terminal of the intelligent control module. The conveying and molding unit includes an insulated screw conveyor, a servo-driven injection system, a mold, and a vision recognition module. The control input of the servo-driven injection system is connected to the signal output of the DDPG reinforcement learning algorithm of the intelligent control module, integrating a pressure-speed co-controller with an injection speed adjustment error of <±0.02m / s and a pressure control error of <±0.05MPa. The mold has built-in cooling water channels and a cavity pressure sensor, and its pressure signal output is connected to the state space data acquisition terminal of the intelligent control module. The filling progress signal output of the vision recognition module is connected to the state space data acquisition terminal of the intelligent control module. The intelligent control module includes a PLC controller, an algorithm execution server, and a data interaction interface. The signal acquisition terminal of the PLC controller is connected to the sensor signal output terminal of each unit, and the control signal output terminal is connected to the actuator of each unit. The algorithm execution server deploys a CNN-LSTM prediction model, a GA-fuzzy PID hybrid optimization framework, a GPR parameter mapping model, and a DDPG reinforcement learning algorithm. It realizes real-time transmission of model input and output data through the data interaction interface to complete closed-loop control.

Images