Casting process parameter optimization method and system based on machine learning

Abstract

The invention relates to the field of electric digital data processing, in particular to a casting process parameter optimization method and system based on machine learning, and the method comprises the steps: sampling in each preset range of process parameters and combining to generate different parameter groups, forming a candidate pool by the parameter groups, and selecting the parameter groups to carry out a simulation experiment, a training data set composed of the parameter set and the gas entrapment defect rate is obtained; training a Gaussian regression model based on the training data set; outputting a predicted gas entrapment defect rate and a predicted variance corresponding to the parameter group in the new candidate pool according to a Gaussian regression model; and constructing a weighted acquisition function, selecting a parameter group corresponding to a minimum weighted acquisition function value to carry out a simulation experiment, obtaining new training data, training a Gaussian regression model based on the new training data set until the minimum predicted gas entrapment defect rate does not change any more, and obtaining an optimal parameter group. According to the method, more attention can be paid to parameters which have great influence on the gas entrapment defect rate, fine search is carried out, and the optimization precision and efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric digital data processing. In particular, it relates to a casting process parameter optimization method and system based on machine learning. BACKGROUND

[0002] High-pressure die casting is an advanced metal forming process that rapidly injects molten metal into a mold cavity under high pressure, allowing it to solidify and form a casting with high precision, high strength, and good surface quality. The high-pressure and high-speed filling stage is a critical link in determining the quality of the casting. During this stage, the molten metal flows into the mold cavity at high pressure and high speed. If the air (including volatile gas from the coating) in the cavity cannot be timely removed through the exhaust groove, it will be entrapped in the molten metal, forming a "gas-entrapped defect". This defect is irreversible and can severely compromise the performance of the casting.

[0003] To reduce the occurrence of gas-entrapped defects, the process parameters of high-pressure die casting need to be precisely controlled. These parameters are strongly coupled, and the optimization of these parameters must consider both the performance limitations of the die casting machine and the physical properties of the metal. This makes the optimization of process parameters a complex multi-objective and multi-constrained problem. Existing optimization algorithms often only use known optimal solutions for local search or are prone to local optimal solutions when performing global search, lacking the ability to dynamically adjust the search direction. This results in low precision and poor optimization results. SUMMARY

[0004] To solve the technical problem of poor optimization result precision in the above optimization method, the present application provides solutions in the following aspects.

[0005] In a first aspect, a casting process parameter optimization method based on machine learning includes: sampling and combining different parameter groups within the preset ranges of the process parameters to form a candidate pool, selecting a parameter group for simulation experiments, and obtaining a training dataset composed of parameter groups and corresponding gas-entrapped defect rates; training a Gaussian regression model based on the training dataset, selecting a kernel function, inputting the parameter groups, and outputting the gas-entrapped defect rates; updating the candidate pool based on the weights of each process parameter to form a new candidate pool, and outputting the predicted gas-entrapped defect rates and prediction variances of each parameter group in the new candidate pool based on the Gaussian regression model; constructing a weighted collection function, traversing all parameter groups in the new candidate pool, selecting the parameter group corresponding to the minimum weighted collection function value for simulation experiments, obtaining new training data, and placing it into the training dataset, training the Gaussian regression model based on the new training dataset, until the minimum predicted gas-entrapped defect rate no longer changes, and taking the parameter group corresponding to the minimum predicted gas-entrapped defect rate as the optimal parameter group; The method for constructing the weighted acquisition function comprises the following steps: For a single parameter set in the new candidate pool, the similarity between the parameter set and the parameter set with the lowest volume defect rate is quantified according to the kernel function, the weight of each process parameter is fused with the similarity to form a weight adjustment term, the balance parameter is calculated based on the parameter set in the new candidate pool and the prediction variance of each parameter set, and the weighted acquisition function is constructed according to the weight adjustment term and the balance parameter.

[0006] Preferably, the kernel function is an ARD kernel function, and when the Gaussian regression model is trained, the signal variance of the kernel function and the length scale of each process parameter are calculated by using a maximum likelihood estimation method.

[0007] Preferably, the method for calculating the weight of the process parameter comprises the following steps: performing negative correlation normalization processing on the length scale of a single process parameter to obtain a normalized value, and calculating the ratio of the normalized value of the length scale of the single process parameter to the sum of the normalized values of the length scales of all process parameters to obtain the weight of the single process parameter.

[0008] Preferably, the method for updating the candidate pool comprises the following steps: outputting the predicted volume defect rate and the prediction variance of all parameter sets in the candidate pool which have not undergone simulation experiments by using the trained Gaussian regression model; selecting a preset number of process parameters corresponding to high weights as target process parameters, selecting a parameter set corresponding to the minimum predicted volume defect rate as a reference parameter set, and setting a sampling interval and performing sampling on the basis of the values of the target process parameters in the reference parameter set; combining the sampling points of the target process parameters and the values of other process parameters in the reference parameter set to form new parameter sets, removing the parameter sets which are repeated in the candidate pool, retaining a preset number of new parameter sets in the candidate pool, and removing the parameter sets which have undergone simulation experiments in the candidate pool to form a new candidate pool.

[0009] Preferably, before the weighted acquisition function is constructed, the method further comprises the following steps: inputting the new parameter sets added in the candidate pool into the trained Gaussian regression model to output the predicted volume defect rate and the prediction variance corresponding to each new parameter set, so as to obtain the predicted volume defect rate and the prediction variance of all parameter sets in the new candidate pool, and form a prediction result set.

[0010] Preferably, the method for calculating the weight adjustment term comprises the following steps: selecting a parameter set corresponding to the minimum predicted volume defect rate in the prediction result set as a reference parameter set, calculating the similarity between a single parameter set in the new candidate pool and the reference parameter set by using an ARD kernel function, calculating the product of the weight of each process parameter in the single parameter set and the similarity respectively, and adding the obtained products to obtain the weight adjustment term.

[0011] Preferably, the calculation method of the balance parameter comprises: performing standardization on each prediction variance in the prediction result set to obtain a prediction standard deviation, calculating a mean value of all prediction standard deviations, and performing negative correlation normalization on the mean value; determining a minimum balance parameter and a maximum balance parameter according to a range of the prediction standard deviation; calculating a difference value between the maximum balance parameter and the minimum balance parameter, calculating a difference value between 1 and the mean value after the negative correlation normalization, multiplying the two difference values, and adding the minimum balance parameter to obtain the balance parameter.

[0012] Preferably, the process of constructing the weighted acquisition function according to the weight adjustment term and the balance parameter comprises: taking a sum value of the weight adjustment term and the balance parameter as an adjustment coefficient of the prediction standard deviation of the parameter group, and subtracting the prediction standard gas defect rate corresponding to the parameter group from the prediction standard deviation after adding the adjustment coefficient to complete the construction of the weighted acquisition function.

[0013] Preferably, the process parameters comprise a slow pressing-in speed, a fast pressing-in speed, a metal liquid temperature and a mold temperature.

[0014] In a second aspect, a machine learning-based casting process parameter optimization system comprises a processor and a memory, and the memory stores computer program instructions which, when executed by the processor, implement the machine learning-based casting process parameter optimization method described above.

[0015] The present application has the following effects: 1. The present application assumes that the output variable (the gas defect rate) is subject to a Gaussian process distribution, learns the joint probability distribution between the input variable (the parameter group) and the output variable by using a training sample set, and further realizes the prediction of unknown inputs. This method can provide more accurate prediction results, reduce the dependence on experience, and the kernel function can accurately capture the differences in the influence of different process parameters on the gas defect rate, thereby improving the fitting ability and prediction accuracy of the model.

[0016] 2. The present application dynamically updates the candidate pool based on the weights of each process parameter, which can optimize the parameters that have the greatest impact on the gas defect rate, thereby improving the optimization accuracy.

[0017] 3. The present application constructs a weighted acquisition function by using the set weight adjustment term, so that the optimization process pays more attention to the parameters that have a greater impact on the gas defect rate, performs fine search, and improves the accuracy and efficiency of the optimization. Moreover, the weight adjustment term dynamically adjusts the weight to ensure that the optimization process not only focuses on the currently known optimal solution, but also dynamically adjusts the search direction to explore a wider parameter space, thereby avoiding being trapped in a local optimal solution and improving the possibility of finding a global optimal solution. In addition, by introducing the weight adjustment term, the model can better adapt to different input parameter combinations and better learn the influence of changes in the parameter group on the gas defect rate, thereby performing better on new data and enhancing the generalization ability of the model. Attached Figure Description

[0018] Figure 1 This is a flowchart of steps S1-S4 in a machine learning-based method for optimizing casting process parameters according to an embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating the construction method of the weighted acquisition function in steps S40-S42 of a machine learning-based casting process parameter optimization method according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] Reference Figure 1 A machine learning-based method for optimizing casting process parameters includes steps S1-S4, as detailed below: S1: Sample and combine different parameter groups within the preset range of process parameters. Different parameter groups form a candidate pool. Select parameter groups from the candidate pool to conduct simulation experiments and obtain a training dataset consisting of parameter groups and corresponding air entrapment defect rates.

[0023] Before sampling, the optimized working parameters are determined: High-pressure die casting is a molding process in which molten metal is forced into a mold cavity under pressure. During the high-speed filling stage, if air in the mold cavity is not expelled in time when the molten metal is filled at high pressure and high speed, it is easily trapped into the molten metal, forming an air entrapment defect. To solve this air entrapment defect, the target optimized casting process parameters include: slow injection speed (the speed at which molten metal is pushed into the barrel), fast injection speed (the speed at which the cavity is filled), molten metal temperature, and mold temperature.

[0024] Determine the preset range of process parameters: The range of slow injection speed and fast injection speed can be given according to the injection system capacity of the high pressure die casting machine (such as hydraulic pump power and injection rod accuracy); the range of molten metal temperature and mold temperature can be set based on the physical properties of the metal type (liquidothermal temperature, fluidity, solidification characteristics) to ensure the fluidity of molten metal filling and avoid oxidation or sticking to the mold.

[0025] Criteria for determining air entrapment defects: In one embodiment, based on the casting quality requirements and metal type, a casting is determined to be an air entrapment defect when any of the following conditions occur inside the casting: (1) the diameter of a single pore is greater than 0.2 mm; (2) the number of pores in a 100 mm² cross-sectional area is greater than 2.

[0026] For each process parameter, sample at equal intervals within its feasible range, ensuring that the values uniformly cover the entire interval. Combine all different sampling points of process parameters through Cartesian product to generate different parameter groups. Randomly select 50% of the parameter groups to form a candidate pool.

[0027] In one embodiment, a 3D simulation model is constructed based on simulation software, which is 1:1 with the actual casting. The boundary conditions of each process parameter are input, and 10 parameter groups are selected from the candidate pool for simulation of high-pressure die casting trial production. Each parameter group simulates 50 castings, detects internal porosity in the castings through a simulation post-processing module, identifies defective pieces according to the set gas entrapment defect judgment standard, and calculates the gas entrapment defect rate corresponding to each parameter group. The gas entrapment defect rate is the ratio of the total number of gas entrapment defective pieces to 50 castings.

[0028] The Min-Max normalization method is used to standardize each process parameter in the 10 parameter groups. When performing Min-Max normalization, the maximum and minimum values used are the boundary values of the feasible range of the corresponding process parameter. The normalized parameter group and the corresponding gas entrapment defect rate constitute the training data, forming a training data set containing 10 training data for subsequent Gaussian regression model training.

[0029] S2: Train the Gaussian regression model based on the training data set, select the kernel function, and input the parameter group. The output is the gas entrapment defect rate.

[0030] The Gaussian regression model assumes that the output variable (gas entrapment defect rate) follows a Gaussian process distribution. It learns the joint probability distribution between the input variable (parameter group) and the output variable using the training sample set, and then makes predictions for unknown inputs.

[0031] The kernel function is the core of the Gaussian regression model, used to measure the similarity between two input samples (parameter groups), and directly determines the model's fitting ability for nonlinear relationships. To accurately capture the differences in the effects of different process parameters on the gas entrapment defect rate (different weights of different process parameters on the gas entrapment defect rate), the ARD kernel (Automatic Relevance Determination Kernel) function is selected as the kernel function of the Gaussian regression model. The ARD kernel function is expressed as follows: In the formula, represents the similarity between the parameter group in the th training data and the th training data in the training data set; represents the signal variance, used to control the strength of the overall correlation between the process parameters and the gas entrapment defect rate; represents the th a process parameter; a length scale of a process parameter; a length scale of a process parameter; an exponential function with a natural number as a base number, for quantifying the distance between each process parameter between two training data, the closer the distance between each process parameter, the closer the value of the exponential function to 1, the higher the similarity.

[0032] In the process of training the Gaussian regression model, the normalized parameter group in the training data set is taken as the input, and the roll gas defect rate is directly taken as the model training label; the kernel parameters of the ARD kernel function, i.e. the signal variance and the length scale of each process parameter, are learned by using the maximum likelihood estimation method.

[0033] It is assumed that the outputs (roll gas defect rates) of all training data obey a multivariate Gaussian distribution, the marginal likelihood function of the current data set is determined, which represents the probability of observing the current training data given the kernel parameters, and the kernel parameters are adjusted by the gradient descent algorithm to maximize the above marginal likelihood function value, and the signal variance of the ARD kernel function and the length scale of the four process parameters are obtained.

[0034] S3: updating the candidate pool based on the weight of each process parameter, forming a new candidate pool, and outputting the predicted roll gas defect rate and the predicted variance corresponding to each parameter group in the new candidate pool according to the Gaussian regression model.

[0035] The smaller the length scale of the process parameter, the greater the influence of the process parameter on the roll gas defect rate, that is, a slight change in the process parameter can cause a significant fluctuation in the defect rate, and vice versa. Therefore, the calculation method of the weight of the process parameter includes: negatively correlating and normalizing the length scale of a single process parameter to obtain a normalized value, calculating the ratio of the normalized value of the length scale of the single process parameter to the sum of the normalized values of the length scales of all process parameters to obtain the weight of the single process parameter. The specific formula is as follows: In the formula, the weight of the i-th process parameter; the weight of the i-th process parameter; the length scale of the i-th process parameter; the length scale of the i-th process parameter; an exponential function with a natural number as a base number, for measuring the influence degree of the length scale of the i-th process parameter. The smaller the length scale , the smaller the length scale , the smaller the length scale , the smaller the length scale The greater the value of the process parameter is, the higher the sensitivity of the process parameter to the output (roll gas defect rate) is, that is, a slight change in the process parameter will cause a significant fluctuation in the roll gas defect rate, and therefore the weight of the influence of the process parameter on the roll gas defect rate should be greater.

[0036] The method for updating the candidate pool comprises: inputting all parameter groups in the candidate pool that have not undergone simulation experiments into the trained Gaussian regression model one by one, outputting the corresponding predicted roll gas defect rate (reflecting the potential defect level) and the predicted variance (the uncertainty of the model for the prediction, the greater the predicted variance is, the higher the uncertainty is); selecting a preset number (for example, two) of process parameters corresponding to high weights as target process parameters, selecting a parameter group corresponding to the minimum predicted roll gas defect rate as a reference parameter group, and dividing the sampling interval of the target process parameter based on the value of the target process parameter in the reference parameter group, densely sampling in the sampling interval at a set interval; combining the sampling points of the target process parameter and the values of the other process parameters in the reference parameter group to form new parameter groups, removing the parameter groups that are repeated in the candidate pool, and retaining a preset number of new parameter groups in the candidate pool, and removing the parameter groups that have undergone simulation experiments in the candidate pool to form a new candidate pool. The values of the target process parameters in the retained new parameter groups need to cover the values in different ranges of high, medium and low in the sampling interval.

[0037] By selecting the process parameters with high weights as the target process parameters, the optimization process can be focused on the parameters that have the greatest influence on the roll gas defect rate, which ensures that the optimization process is more efficient and avoids wasting computing resources on unimportant parameters. Dense sampling in the ±5% sampling interval of the target process parameter can more accurately explore the potential optimal values of these key parameters and improve the accuracy of optimization. Moreover, selecting the parameter group corresponding to the minimum predicted roll gas defect rate as the reference parameter group ensures that the optimization process is always directed towards reducing the roll gas defect rate, avoiding complete reliance on random exploration.

[0038] The new parameter groups added to the candidate pool are input into the trained Gaussian regression model one by one, and the predicted roll gas defect rate and the predicted variance corresponding to each new parameter group are output, so as to obtain the predicted roll gas defect rate and the predicted variance of all parameter groups in the new candidate pool, forming a prediction result set.

[0039] S4: Construct a weighted acquisition function, traverse all parameter groups in the new candidate pool, select a parameter group corresponding to the minimum weighted acquisition function value for simulation experiment, obtain new training data, and put the new training data into the training data set, train the Gaussian regression model based on the new training data set, until the minimum predicted roll gas defect rate no longer changes, and the parameter group corresponding to the minimum predicted roll gas defect rate is taken as the optimal parameter group.

[0040] In one embodiment, a lower confidence boundary function is selected as the basic acquisition function, and a weighted acquisition function is constructed based on the basic acquisition function. (Refer to...) Figure 2 The method for constructing the weighted acquisition function includes steps S40-S42, as detailed below: S40: For a single parameter group in the new candidate pool, the similarity between the parameter group and the parameter group with the lowest air entrapment defect rate is quantified according to the kernel function, and the weights of each process parameter are integrated with the similarity to form a weight adjustment term.

[0041] The calculation method for the weight adjustment term includes: selecting the parameter group corresponding to the minimum predicted air-entrainment defect rate in the prediction result set as the reference parameter group; using the ARD kernel function to calculate the similarity between the single parameter group in the new candidate pool and the reference parameter group; calculating the product of the weight and similarity of each process parameter in the single parameter group; and summing the resulting products to obtain the weight adjustment term. The specific formula is as follows: In the formula, Indicates the first in the new candidate pool Weighting adjustment terms for each parameter group; It represents the signal variance, used to control the strength of the overall correlation between the parameter group in the new candidate pool and the reference parameter group; Indicates the first The weight of each process parameter; Indicates the first in the new candidate pool The first parameter group The values ​​of each process parameter; Indicates the first reference parameter group The values ​​of each process parameter; Indicates the first The length of the scale of each process parameter; This indicates that the th candidate in the new candidate pool is calculated using the ARD kernel function. The similarity between each parameter group and the reference parameter group.

[0042] S41: Calculate the equilibrium parameters based on the parameter sets in the new candidate pool and the prediction variance of each parameter set.

[0043] The balancing parameter plays a role in the acquisition function in balancing exploration and utilization: when the balancing parameter is large, the acquisition function tends to select points with higher prediction uncertainty, i.e., explore unknown areas; when the balancing parameter is small, the acquisition function tends to select points with lower prediction values, i.e., utilize known high-quality areas.

[0044] The calculation method of the balance parameter comprises: performing standardization on the prediction variance of each parameter group in the new candidate pool (i.e. each prediction variance in the prediction result set) to obtain a prediction standard deviation, calculating the mean of all prediction standard deviations, and performing negative correlation normalization on the mean; determining the minimum balance parameter and the maximum balance parameter according to the range of the prediction standard deviation; calculating the difference between the maximum balance parameter and the minimum balance parameter, calculating the difference between 1 and the mean after the negative correlation normalization, multiplying the two differences, and adding the minimum balance parameter to obtain the balance parameter. The specific formula is as follows: In the formula, represents the balance parameter; represents the minimum balance parameter, represents the maximum balance parameter; represents the mean of all prediction standard deviations, , represents the number of all parameter groups in the new candidate pool, represents the prediction standard deviation of the i-th parameter group in the new candidate pool, represents the prediction standard deviation of the i-th parameter group in the new candidate pool, represents the exponential function with the natural number as the base number.

[0045] The minimum balance parameter and the maximum balance parameter are determined according to the range of the prediction standard deviation. For example, if the range of the prediction standard deviation is [0, 1], it means that the model is very certain about the prediction of some points (i.e. the prediction standard deviation is close to 0), and very uncertain about the prediction of other points (i.e. the prediction standard deviation is close to 1). When the model prediction is very certain (i.e. the prediction standard deviation is close to 0), the optimization process should be biased towards exploitation, which means fine search around the currently known optimal solution. A small value, for example 0.1, is usually chosen, which should be small enough to ensure that when the model prediction is very certain, the optimization process will not explore too much; when the model prediction is very uncertain (i.e. the prediction standard deviation is close to 1), the optimization process should be biased towards exploration, which means searching for potential optimal solutions in areas with high uncertainty. A large value, for example 1 or 2, is usually chosen, which should be large enough to ensure that when the model prediction is very uncertain, the optimization process can fully explore the unknown area.

[0046] By the positive correlation between the mean of the prediction standard deviation and the balance parameter is realized. The larger the mean, the lower the reliability of the model in predicting the roll gas defect rate of most parameter groups (high uncertainty), at which time the balance parameter is increased to expand the search range of high-quality process parameters; on the contrary, the smaller the mean, the clearer the model cognition (low uncertainty), at which time the balance parameter is decreased to focus on fine optimization around the process parameters corresponding to the current minimum roll gas defect rate.

[0047] S42: Construct a weighted acquisition function based on the weight adjustment term and the balancing parameter.

[0048] The adjustment coefficient, composed of the weighting adjustment term and the balancing parameter, is added to the basic acquisition function to form the weighted acquisition function, as shown in the following formula: In the formula, Indicates the first in the new candidate pool The weighted acquisition function values ​​of each parameter group; Indicates the first in the new candidate pool Predicted air-winding defect rate for each parameter set; Indicates the first in the new candidate pool Weighting adjustment terms for each parameter group; Indicates the balance parameters; Indicates the first in the new candidate pool The predictive standard deviation of each parameter group.

[0049] By standardizing the prediction variance for each parameter group, the prediction variance of all parameter groups is scaled to the same order of magnitude as the air-entraining defect rate. By introducing a weight adjustment term, the optimization process can focus more on process parameters that have a greater impact on the air-entraining defect rate, avoiding wasting computational resources on unimportant process parameters, thereby improving optimization efficiency and effectiveness. Moreover, by dynamically adjusting the weights, the weight adjustment term ensures that the optimization process not only focuses on the currently known optimal solution, but also dynamically adjusts the search direction to explore a wider parameter space, avoiding getting trapped in local optima.

[0050] Calculate the weighted acquisition function values ​​of all parameter groups in the new candidate pool, select the parameter group corresponding to the minimum weighted acquisition function value, and conduct a high-pressure die casting simulation experiment to simulate the production of 50 virtual castings. Calculate the air entrapment defect rate based on the experimental results. Normalize each process parameter in this parameter group using Min-Max. The normalized parameter group and the corresponding air entrapment defect rate form new training data, which is then added to the training dataset to obtain the updated training dataset.

[0051] The Gaussian regression model is retrained using the new training dataset. This involves returning to step S2 and iterating, recording the minimum air entrapment defect rate in the prediction result set during each iteration. If the minimum air entrapment defect rate remains unchanged for three consecutive iterations, the model is considered converged, and the iteration stops. The parameter set corresponding to the minimum air entrapment defect rate is then used as the optimal process parameter combination for the high-pressure die-casting high-speed filling stage in actual production.

[0052] This application also discloses a machine learning-based casting process parameter optimization system. The system includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the machine learning-based casting process parameter optimization method according to the above embodiments of the present invention is implemented.

[0053] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0054] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for optimizing casting process parameters based on machine learning, characterized in that, include: Different parameter groups are generated by sampling and combining within each preset range of process parameters. Different parameter groups form a candidate pool. The parameter groups are selected for simulation experiments to obtain a training dataset consisting of parameter groups and corresponding air entrainment defect rates. A Gaussian regression model is trained based on the training dataset, with a kernel function selected, the input being a set of parameters, and the output being the air-filling defect rate. The candidate pool is updated based on the weights of each process parameter to form a new candidate pool. The predicted air entrapment defect rate and prediction variance for each parameter group in the new candidate pool are output according to the Gaussian regression model. Construct a weighted acquisition function, traverse all parameter groups in the new candidate pool, select the parameter group corresponding to the minimum weighted acquisition function value for simulation experiments, obtain new training data, and put it into the training dataset. Train the Gaussian regression model based on the new training dataset until the minimum predicted vortex defect rate no longer changes, and take the parameter group corresponding to the minimum predicted vortex defect rate as the optimal parameter group. The method for constructing the weighted acquisition function includes: For a single parameter group in the new candidate pool, the similarity between the parameter group and the parameter group with the lowest air entrapment defect rate is quantified according to the kernel function. The weights of each process parameter are fused with the similarity to form a weight adjustment term. The balance parameter is calculated based on the parameter group in the new candidate pool and the prediction variance of each parameter group. A weighted acquisition function is constructed based on the weight adjustment term and the balance parameter.

2. The method for optimizing casting process parameters based on machine learning according to claim 1, characterized in that, The kernel function is the ARD kernel function. When training the Gaussian regression model, the maximum likelihood estimation method is used to calculate the signal variance of the kernel function and the length scale of each process parameter.

3. The method for optimizing casting process parameters based on machine learning according to claim 2, characterized in that, The method for calculating the weight of the process parameter includes: performing negative correlation normalization on the length scale of a single process parameter to obtain a normalized value; calculating the ratio of the normalized value of the length scale of a single process parameter to the sum of the normalized values ​​of the length scales of all process parameters to obtain the weight of the single process parameter.

4. The method for optimizing casting process parameters based on machine learning according to claim 3, characterized in that, The method for updating the candidate pool includes: using a trained Gaussian regression model to output the predicted air entrapment defect rate and prediction variance of all parameter groups in the candidate pool that have not undergone simulation experiments; selecting a preset number of high-weight process parameters as target process parameters, selecting the parameter group corresponding to the minimum predicted air entrapment defect rate as a reference parameter group, setting a sampling interval and sampling based on the value of the target process parameter in the reference parameter group; combining the sampling points of the target process parameter with the values ​​of other process parameters in the reference parameter group to form a new parameter group; removing parameter groups that are duplicated in the candidate pool, retaining a preset number of new parameter groups and placing them in the candidate pool, and removing parameter groups in the candidate pool that have participated in simulation experiments to form a new candidate pool.

5. The method for optimizing casting process parameters based on machine learning according to claim 4, characterized in that, Before constructing the weighted acquisition function, the following steps are also taken: inputting the new parameter groups added to the candidate pool into the trained Gaussian regression model, outputting the predicted vortex defect rate and prediction variance corresponding to each new parameter group, thereby obtaining the predicted vortex defect rate and prediction variance of all parameter groups in the new candidate pool, forming a prediction result set.

6. The method for optimizing casting process parameters based on machine learning according to claim 5, characterized in that, The calculation method of the weight adjustment term includes: selecting the parameter group corresponding to the minimum predicted air vortex defect rate in the prediction result set as the reference parameter group; using the ARD kernel function to calculate the similarity between the single parameter group in the new candidate pool and the reference parameter group; calculating the product of the weight and similarity of each process parameter in the single parameter group respectively; and adding the obtained products to obtain the weight adjustment term.

7. The method for optimizing casting process parameters based on machine learning according to claim 5, characterized in that, The method for calculating the balance parameter includes: standardizing the prediction variances in the prediction result set to obtain the prediction standard deviation; calculating the mean of all prediction standard deviations and performing negative correlation normalization on the mean; determining the minimum balance parameter and the maximum balance parameter based on the range of the prediction standard deviations; calculating the difference between the maximum balance parameter and the minimum balance parameter; calculating the difference between 1 and the mean after negative correlation normalization; multiplying the two differences and adding them to the minimum balance parameter to obtain the balance parameter.

8. The method for optimizing casting process parameters based on machine learning according to claim 7, characterized in that, The process of constructing a weighted acquisition function based on the weight adjustment term and the balance parameter includes: using the sum of the weight adjustment term and the balance parameter as the adjustment coefficient of the prediction standard deviation of the parameter group, and subtracting the prediction air defect rate of the corresponding parameter group from the prediction standard deviation after adding the adjustment coefficient to complete the construction of the weighted acquisition function.

9. The method for optimizing casting process parameters based on machine learning according to claim 1, characterized in that, The process parameters include slow injection speed, fast injection speed, molten metal temperature, and mold temperature.

10. A machine learning-based casting process parameter optimization system, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the machine learning-based casting process parameter optimization method according to any one of claims 1-9.

Images