An integrated learning-based squeeze casting process parameter design method

Abstract

The application provides an extrusion casting process parameter design method based on integrated learning, and belongs to the field of extrusion casting process parameter design. 2 The method comprises the following steps: constructing an extrusion casting process data set of a target casting; establishing a unit model based on the data set; integrating the unit model by using a weight assignment method and an improved model integration strategy R The integrated model is constructed, and finally, the integrated model is solved by using an ant colony algorithm to obtain optimal process parameters and perform verification. The optimization design level of the process parameters is improved, a framework is provided for more accurately establishing a complex relationship model between the process parameters and performance, and the fitting and generalization abilities of a single model are effectively improved. Meanwhile, the integrated model makes the method more suitable for parallel optimization design of the extrusion casting process parameters under small sample data conditions, thereby reducing the time and cost of specific casting process parameter design. The method is not only suitable for the design of extrusion casting process parameters, but also suitable for parameter design in other similar fields.

Description

Technical Field

[0001] This invention relates to the field of extrusion casting process parameter design, and in particular to an extrusion casting process parameter design method based on ensemble learning. Background Technology

[0002] Extrusion casting is a near-net-shape manufacturing process that combines the advantages of casting and forging. It boasts a short process route, high efficiency, and high forming accuracy, offering benefits such as improved casting performance, reduced casting defects, production of complex-shaped parts, and energy conservation and environmental protection. Optimal process parameters are key to achieving these advantages. Therefore, extensive research has been conducted both domestically and internationally on extrusion casting process parameters, resulting in several design methods. However, a unified and highly efficient design method has yet to be established. Current extrusion casting production processes often utilize physical experiments or numerical simulations to obtain optimal process parameters. However, experimental methods are not only affected by the analytical and optimization capabilities of designers but also cannot guarantee globally optimal parameters because they are directly selected from discrete experimental results. Therefore, some recent studies have constructed fitting models based on experimental data to optimize extrusion casting process parameters, achieving globally optimal values. However, due to the unknown and complex multivariate relationships between extrusion casting process parameters and casting performance, the generalization ability of a single model is insufficient, especially when the relevant fitting data sample is small, weakening its ability to optimize process parameters.

[0003] Ensemble learning is a machine learning method that improves the generalization ability of a model. Compared with using a single model, combining multiple machine learning models can achieve better predictive performance. To date, no ensemble model has been developed for designing squeeze casting process parameters. This patent addresses this background by introducing ensemble learning into squeeze casting and, based on the principles of squeeze casting, proposes an ensemble learning-based method for designing squeeze casting process parameters. Summary of the Invention

[0004] The purpose of this invention is to provide a method for designing extrusion casting process parameters based on ensemble learning, thereby solving the technical problems mentioned in the background art. According to the method for designing extrusion casting process parameters of this invention, parallel optimization design of extrusion casting process parameters is achieved.

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

[0006] A method for designing extrusion casting process parameters based on ensemble learning, the method comprising the following steps:

[0007] Step 1: Construct a squeeze casting process dataset for the target casting. First, obtain squeeze casting process data through physical experiments, numerical simulations, or literature reviews. The data includes two parts: process parameter data and casting performance index data. Let the dataset be DB, with n samples, v process parameters, and m performance indices, S... i Let X represent the i-th sample, and let X be the set of process parameters. h Let the h-th process parameter be denoted as YA, where h = 1, 2, ..., v, and the performance index set be denoted as YJ. j Let j = 1, 2, ..., m be attributes of DB, then sample S i The corresponding attribute data values ​​are denoted as x. ih and y ij Therefore, DB = {S} i (x ih y ij )|i=1, 2,...,n; h=1, 2,..., v; j=1, 2,..., m};

[0008] Step 2: Construct unit models based on the extrusion casting process dataset. Normalize the dataset DB and divide it into training and testing sets. Use process parameters XA as input and casting performance index YA as output. Based on DB, fit the data using several different types of models to obtain o unit models. The output of each unit model is denoted as T1, T2, ..., T o ;

[0009] Step 3: Based on the improved model ensemble strategy R 2 A weighted ensemble model is constructed by combining the unit models obtained in step 2 into an ensemble model through addition. The performance of the ensemble model is then evaluated. If its accuracy does not meet the requirements, it is optimized. Considering the differences in the fitting performance of the unit models, an R-squared algorithm is proposed. 2 Weighting method for integrating unit models;

[0010] Step 4: Use the Ant Colony Optimization Algorithm (ACO) to solve the ensemble model, obtain the optimal extrusion casting process parameters, and verify them. If they meet the requirements, the extrusion casting process parameter design is completed; otherwise, adjust the parameters of the ant colony algorithm and re-optimize and solve the problem.

[0011] In step 1, a squeeze casting process dataset for the target casting is constructed. The dataset is represented by DB and its format is shown in the table below.

[0012] Squeeze casting process dataset DB

[0013]

[0014] Furthermore, in step 2, the process of normalizing the dataset DB and dividing it into training and test sets is as follows:

[0015] Step 2.1: Normalize the dataset DB using the following formula:

[0016] f = (uu min ) / (u max -u min )

[0017] Where f represents the normalized data, and u represents the original value of each attribute in the database. max and u min These represent the maximum and minimum values ​​of each attribute in the database;

[0018] Step 2.2: Divide the preprocessed dataset samples into a training set and a test set, with the ratio of the training set to the test set between 1.5 and 3.

[0019] Furthermore, in step 3, the improved model ensemble strategy R... 2 The process of constructing an ensemble model using the weight assignment method is as follows:

[0020] Step 3.1, according to R 2 The weights of each unit model in the ensemble model are defined as follows:

[0021]

[0022] Where o is the number of unit models, Representing R for each model 2 ω1, ω2, ω3…ωo are the corresponding weights assigned to each model;

[0023] Because of R 2 Since there are many decimal places in the model, allocating weights according to their size would lead to more complex weight values. Therefore, the weight allocation is based on the following principles: (1) round the allocated weights approximately; (2) the weights of the optimal model are the largest.

[0024] The individual unit models are combined using an additive approach, with their respective weights, to form an ensemble model. The input to the ensemble model is expressed as follows:

[0025] T0 = ​​ω1*T1 + ω2*T2 + ... + ω o *T o

[0026] Where T1, T2....T o This represents the output of each unit model;

[0027] Step 3.2: After establishing the ensemble model, in order to comprehensively evaluate the performance of the ensemble model, the coefficient of determination R is selected. 2 Mean squared error (MSE) and root mean squared error (RMSE) are used as evaluation metrics for the model. 2 The values ​​are between [0, 1], and the closer to 1 the better. The smaller the MSE and RMSE, the better. The expressions for the three model evaluation indicators are as follows:

[0028]

[0029]

[0030]

[0031] Where n′ is the number of test samples, Let y be the model prediction value of the j-th performance metric in the r-th test sample. rj Let j be the true value of the performance metric in the r-th test sample. Let i be the true average value of the j-th performance metric in the n-th test sample, where i and j are both positive integers;

[0032] Step 3.3: If the performance of the ensemble model cannot meet the requirements, it needs to be optimized. The optimization method consists of two steps: (1) Adjusting the hyperparameters of each unit model to improve its performance, and (2) According to the new R of each unit model... 2 Recalculate the weights of each model and reconstruct the ensemble model.

[0033] Furthermore, in step 4, the process of selecting the Ant Colony Algorithm (ACO) to solve the ensemble model and obtaining the extrusion casting process parameters is as follows:

[0034] The optimization variable is set as a set of extrusion casting process parameters. Each ant represents a set of extrusion casting process parameters, containing v process parameters to be designed. The value range of each process parameter is normalized and set to [0, 1]. Therefore, the ants are denoted as X = {(X1, X2, ..., X...}. v )|X h ∈[0,1], h=1,2,…v), the optimal casting performance output by the integrated model is used as the target of ant colony optimization, and the range of process parameters is the optimization range of the ant colony;

[0035] Ants search for paths based on pheromone concentration. The pheromone concentration along the path determines the probability that ant e will move from its current position i to the next position j′ at time t, as shown in the following formula:

[0036]

[0037] In the formula, μ i′j′(t) represents heuristic information. d represents the magnitude of the ant's expected transfer degree. i′j′ τ represents the distance from the current position i′ to the next position j′; i′j′ (t) represents the concentration of pheromones along the path; μ i′s (t) represents the heuristic information from the previous position i to the final target position s, τ i′s (t) represents the amount of information an ant has from its current position i to its final target position s, α represents the information heuristic factor, β represents the expected heuristic factor, and the set allowed e This represents all possible locations where ant e can move;

[0038] The pheromone concentration is constantly updated locally; the pheromone concentration along the path evaporates and decreases over time, while ants continuously release new pheromones after passing through the path, thus constantly updating the pheromone concentration. After k seconds, the pheromone concentration from the current position i′ to the next position j′ is shown below:

[0039] τ i′j′ (t+k)=(1-ρ)*τ i′j′ (t)+Δτ i′j′ (t), ρ∈(0,1)

[0040]

[0041] In the formula, Δτ i′j′ (t) represents the total amount of pheromone left by all ants m′ in the colony from the current position i′ to the next position j′; ρ represents the pheromone evaporation coefficient; at the initial time t=0, τ i′j′ (t)=c,Δτ i′j′ (t) = 0;

[0042] The maximum number of iterations n′ is set to range from [50, 200], the number of ants m′ is 1 to 3 times the number of training samples, the evaporation coefficient c′ ranges from (0, 0.2], the pheromone increase intensity p′ ranges from [1, 4], and the ant crawling speed a′ ranges from [0.5, 2].

[0043] The ant colony algorithm for solving the ensemble model is as follows:

[0044]

[0045]

[0046] The present invention, by adopting the above-described technical solution, has the following beneficial effects:

[0047] This invention provides a novel data-driven design method for squeeze casting process parameter design, improving the optimization level of squeeze casting process parameters. It provides a framework for more accurately establishing a complex relationship model between squeeze casting process parameters and casting properties, effectively improving the fitting and generalization capabilities of single models. The ensemble model makes this method more suitable for parallel optimization design of process parameters under small sample data conditions, thereby reducing the time and cost of designing process parameters for specific castings and promoting the application and development of squeeze casting. This method is not only applicable to the design of squeeze casting process parameters but also has reference value for the design of process parameters in other similar fields. Attached Figure Description

[0048] Figure 1 This is a flowchart of the overall method of the present invention.

[0049] Figure 2 This is a dimensional drawing of the high-speed train drive shaft, an example of the present invention.

[0050] Figure 3 This is a verification filling diagram of the high-speed train drive shaft, an example of the present invention.

[0051] Figure 4 This is a verification temperature field diagram of a high-speed train drive shaft, an example of the present invention.

[0052] Figure 5 This is a verification diagram of the shrinkage hole of a high-speed train drive shaft, an example of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the present invention, and these aspects of the invention can be implemented even without these specific details.

[0054] like Figure 1-5 As shown, a method for designing extrusion casting process parameters based on ensemble learning is described, and the method includes the following steps:

[0055] Step 1: Construct a squeeze casting process dataset for the target casting. First, obtain squeeze casting process data through physical experiments, numerical simulations, or literature reviews. This data includes two parts: process parameter data and casting performance index data. Let the dataset be DB, with n samples, v process parameters, and m performance indices, and S... i Let X represent the i-th sample, and let X be the set of process parameters. h Let the h-th process parameter be denoted as YA, and let the j-th performance parameter be denoted as Y.j Let j = 1, 2, ..., m be attributes of DB, then sample S i The corresponding attribute data values ​​are denoted as x. ih and y ij Therefore, DB = {S} i (x ih ,y ij )|i=1,2,…,n; h=1,2,…,v; j=1,2,…,m}.

[0056] Taking the extrusion casting process parameter design of the 7075 aluminum alloy high-speed train drive shaft as an example, its three-dimensional model and dimensional drawing are as follows: Figure 2 As shown in Table 1, a dataset DB for the extrusion casting process of high-speed train drive shafts was constructed through numerical simulation experiments using ProCAST.

[0057] Table 1 shows the extrusion casting process dataset DB for high-speed train drive shafts.

[0058]

[0059]

[0060] Step 2: Construct unit models based on the extrusion casting process dataset. Normalize the dataset DB and divide it into training and testing sets. Using process parameters XA as input and casting performance index YA as output, fit the data using various models of different types according to DB, thus obtaining o unit models. The output of each unit model is denoted as T1, T2, ..., T... o .

[0061] Theoretically, this method can integrate multiple arbitrary unit models, but integrating too many unit models can lead to long modeling time and coordination difficulties. Based on the methodological framework and considering the existing research on the optimization design models for squeeze casting process parameters, this example selects Support Vector Machine (SVM), Multivariable Linear Regression (MLR), and Random Forest (RF) as unit models, with process parameter XA as input and casting performance index YA as output, denoted as T1, T2, and T3 respectively.

[0062] Specific steps:

[0063] Step 2.1: Divide the 25 sets of sample data in Table 1 into 15 training sets and 10 test sets in a 3:2 ratio, and perform data preprocessing.

[0064] Step 2.2: Constructing the SVM model. The steps for constructing the SVM model of squeeze casting process parameters and casting performance indicators are as follows:

[0065] 1) Using casting performance indicators as the dependent variable and process parameters as the independent variables, train the model in a high-dimensional space H, as shown in the following formula:

[0066]

[0067] In the formula, Φ(X) is the nonlinear correlation function between the casting performance index and the process parameters; b is the coefficient factor. To obtain the expression for Φ(X), the following relationship is introduced:

[0068]

[0069] In the formula, ω is the weighting coefficient, and ξ is the weighting coefficient. i and is a non-negative relaxation coefficient; c is a penalty factor. The following relationship satisfies the constraint condition of the nonlinear correlation function Φ(X) with the casting performance index:

[0070]

[0071] In the formula, ε is the training bias of the SVM model.

[0072] 2) Set the penalty factor c and the radial basis function kernel parameter g, and train the SVM model using the training set through steps 1)-2), finally obtaining the model with respect to the Lagrange operator α. i , The relevant SVM model expression is as follows:

[0073]

[0074] Based on the above SVM model construction steps, an SVM unit model is constructed using the training set data and MATLAB.

[0075] Step 2.3: Construct the MLR model. The MLR multiple linear regression model for the squeeze casting process parameters is shown in the following equation.

[0076] T2 = b0 + b1X1 + ... + b v X v

[0077] Where b0 is a constant term, b1, b2, ..., b v is the regression coefficient.

[0078] The optimal fitting line to the training set is selected to fit the training set to the multiple linear regression model shown in the above formula, and the regression coefficients are obtained to construct an MLR model of extrusion casting process parameters and casting performance.

[0079] Based on the above steps for building the MLR model, an MLR unit model is constructed using the training set data and MATLAB.

[0080] Step 2.4: Constructing the RF model. The steps for constructing the RF model of squeeze casting process parameters and casting performance indicators are as follows:

[0081] 1) Suppose the RF model consists of N decision trees. Each decision tree model is constructed using XA as input and YA as output. The output of the k-th decision tree is denoted as {DT}. k ,k=1,…,N;

[0082] 2) Using the sampling with replacement method (Bootsrap), select the same number of samples as the training set from the divided extrusion casting process data training set as a new training set. Use the newly extracted training set to train a decision tree model, and set a minimum number of leaves l, which represents the lower limit of the number of samples contained in each node of the decision tree.

[0083] 3) Repeat step 2) N times to obtain N decision trees, and then integrate them according to the mean method strategy to form an RF model, as shown in the following formula, thereby completing the RF model training.

[0084]

[0085] Based on the above steps for constructing the RF model, an RF unit model is constructed using the training set data and MATLAB.

[0086] Step 3: Based on the improved model ensemble strategy R 2 An ensemble model is constructed using a weighted assignment method. The unit models obtained in step 2 are combined into an ensemble model through addition. The performance of the ensemble model is then evaluated, and if its accuracy does not meet the requirements, it is optimized. Considering the differences in fitting performance between unit models, an R-squared algorithm is proposed. 2 The weighting method is used to integrate unit models.

[0087] This example uses R. 2 The SMR (SVM-MLR-RF) ensemble model is constructed using the weight assignment method, and the output of the SMR ensemble model is denoted as T0.

[0088] Specific steps:

[0089] Step 3.1: According to R 2 The weights of each unit model in the ensemble model are defined as follows:

[0090]

[0091] Where o is the number of unit models, Representing R for each model 2 ω1, ω2, ω3…ω o The corresponding weights assigned to each model.

[0092] Because of R 2 Since there are many decimal places in the model, allocating weights according to their size would lead to more complex weight values. Therefore, the weight allocation ratio is approximated according to the following principles: (1) round the allocated ratio approximately; (2) allocate as much of the ratio as possible to the optimal model.

[0093] The individual unit models are combined using an additive approach, with their respective weights, to form an ensemble model. The input to the ensemble model is expressed as follows:

[0094] T0 = ​​ω1*T1 + ω2*T2 + ... + ω o *T o

[0095] Based on the above steps, as shown in Table 3, the R values ​​of the three basic models, SVM, MLR, and RF, are... 2 If the values ​​are 0.424, 0.440, and 0.359, then the weights ω1, ω2, and ω3 of each basic model are assigned as 4 / 11, 4 / 11, and 3 / 11, respectively, to construct the SMR ensemble model.

[0096] Step 3.2: After establishing the ensemble model, in order to comprehensively evaluate the performance of the ensemble model, the coefficient of determination R is selected. 2 Mean Square Error (MSE) and Root Mean Squared Error (RMSE) are used as evaluation metrics for the model. 2 The values ​​should be between [0,1], with closer to 1 being better; the smaller the MSE and RMSE, the better. The expressions for the three model evaluation indicators are as follows:

[0097]

[0098]

[0099]

[0100] Where n' is the number of test samples, Let y be the model prediction value of the j-th performance metric in the r-th test sample. rj Let j be the true value of the performance metric in the r-th test sample. It represents the true average value of the j-th performance metric in the n-th test sample.

[0101] Based on the above steps, the evaluation indices of the SMR ensemble model are calculated, as shown in Table 2, and its R-value is obtained. 2 The performance of the SMR ensemble model was improved compared to the basic model, with an R² value of 0.536, an MSE of 0.043, and an RMSE of 0.208. However, the performance evaluation metric R² remained low. 2 The performance is not high enough; further optimization of the ensemble model is needed.

[0102] Table 2 shows the performance evaluation index values ​​for each model.

[0103]

[0104]

[0105] Step 3.3: If the performance of the ensemble model cannot meet the requirements, the ensemble model needs to be optimized. The optimization method consists of two steps: (1) Adjust the hyperparameters of each unit model to improve its performance; (2) Recalculate the weights of each model based on the new R2 of each unit model and reconstruct the ensemble model.

[0106] Based on the above steps, the SMR ensemble model was optimized. After optimization, the penalty factor c of the SVM model was 4.0, and the radial basis function kernel parameter g was 0.8; the number of decision trees N in the RF model was 500, and the minimum number of leaves l was 2. After optimizing the performance of the basic model, the R-values ​​of each unit model after optimization were compared with those shown in Table 3. 2 Size, set the weights of the three basic models SVM, MLR, and RF to be... The optimized SMR ensemble model is shown in the following equation, and its performance is shown in Table 3.

[0107]

[0108] Step 4: Use the Ant Colony Optimization (ACO) algorithm to solve the ensemble model, obtain the optimal extrusion casting process parameters, and verify them. If they meet the requirements, the extrusion casting process parameter design is completed; otherwise, adjust the parameters of the ant colony algorithm and re-optimize and solve the problem.

[0109] The specific steps are as follows:

[0110] Step 4.1: The process of using the Ant Colony Algorithm (ACO) to solve the ensemble model and obtain the extrusion casting process parameters is as follows:

[0111] The optimization variable is set as a set of extrusion casting process parameters. Each ant represents a set of extrusion casting process parameters, containing v process parameters to be designed. The value range of each process parameter is normalized to [0,1]. Therefore, the ant is denoted as X={(X1,X2,…,X…)}. v )|Xh ∈[0,1],h=1,2…v}, the optimal casting performance output by the integrated model is used as the target of ant colony optimization, and the range of process parameters is the optimization range of the ant colony.

[0112] Ants search for paths based on pheromone concentration. The pheromone concentration along the path determines the probability that ant e will move from the current position i to the next position j′ at time t, as shown in the following formula.

[0113]

[0114] In the formula, μ i′j′ (t) represents heuristic information. d represents the magnitude of the ant's expected transfer degree. i′j′ τ represents the distance from the current position i′ to the next position j′; i′j′ (t) represents the concentration of pheromones along the path; μ i′s (t) represents the heuristic information from the previous position i to the final target position s, τ i′s (t) represents the amount of information the ant has from its current position i to its final target position s; α represents the information heuristic factor, β represents the expected heuristic factor, and the set allowed... e This represents all possible locations where ant e can move.

[0115] The pheromone concentration is constantly updated locally; the pheromone concentration on the path will continuously evaporate and decrease over time, while the ant releases new pheromones after passing through the path, and the pheromone concentration is constantly updated. After k seconds, the pheromone concentration from the current position i′ to the next position j′ is shown below.

[0116] τ i′j′ (t+k)=(1-ρ)*τ i′j′ (t)+Δτ i′j′ (t), ρ∈(0,1)

[0117]

[0118] In the formula, Δτ i′j′ (t) represents the total amount of pheromone left by all ants m′ in the colony from the current position i′ to the next position j′; ρ represents the pheromone evaporation coefficient; at the initial time t=0, τ i′j′ (t)=c,Δτ i′j′ (t) = 0.

[0119] The maximum number of iterations n′ is set to range from [50, 200], the number of ants m′ is 1 to 3 times the number of training samples, the evaporation coefficient c′ is set to range from (0, 0.2], the pheromone increase intensity p′ is set to range from [1, 4], and the ant crawling speed a′ is set to range from [0.5, 2].

[0120] The ant colony algorithm for solving the ensemble model is as follows:

[0121]

[0122]

[0123] Based on the above steps, the SMR ensemble model was solved using the ant colony algorithm to obtain the process parameters. The ACO parameter settings are shown in Table 3. The optimal casting temperature was 735.2℃, the extrusion pressure was 192.77MPa, and the die temperature was 345.4℃.

[0124] Table 3 shows the ant colony algorithm parameters.

[0125]

[0126] Step 4.2: Numerical simulation experiments were conducted in ProCAST to verify the optimized combination of process parameters. Under these process parameter conditions, such as... Figure 3 and Figure 4 As shown, the workpiece has a good filling effect. When the filling is finished, the temperature of the molten metal in each area of ​​the casting is between 709.23℃ and 735.20℃. The molten metal has good fluidity and few internal defects when it solidifies. Figure 5 The simulation results for shrinkage porosity and cavities in the castings are shown in the figure. As can be seen, shrinkage porosity and cavities in the castings are minimal, mainly occurring at the thick-walled junction of the drive shaft chassis and shaft body, consistent with prior engineering knowledge. Table 1 shows the maximum and minimum shrinkage porosity and cavities in the experimental samples, which are 4.095% and 2.979%, respectively. The process parameters designed based on this method reduce the shrinkage porosity and cavities in the castings to 1.316%, significantly reducing them compared to process parameters directly based on experimental results, resulting in a significant improvement in casting quality. This demonstrates that the process parameters designed using this method effectively meet the process requirements.

[0127] Matters not covered in this invention are common knowledge.

[0128] The above description is only a preferred embodiment of 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 principle 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 designing extrusion casting process parameters based on ensemble learning, characterized in that: The method includes the following steps: Step 1: Construct a squeeze casting process dataset for the target casting. First, obtain squeeze casting process data through physical experiments, numerical simulation experiments, or literature extraction. The data includes two parts: process parameter data and casting performance index data. Let the dataset be... DB ,have n One sample v Each process parameter and m One performance metric, S i Indicates the first i The set of process parameters for each sample is represented as follows: XA , X h For the first h One process parameter, The performance index set is represented as YA , No. j Each performance metric is expressed as Y j , They are all called DB The attribute of the sample S i The corresponding attribute data values ​​are respectively denoted as and Therefore, there is ; Step 2: Construct a unit model based on the extrusion casting process dataset, and analyze the dataset. DB Normalization was performed, and the dataset was divided into training and testing sets, based on process parameters. XA For input, casting performance indicators YA For output, according to DB Several different types of models are used to fit the data to obtain... o There are 1 unit model, and the output of each unit model is denoted as follows: T 1, T 2, ..., T o ; Step 3: Based on the improved model ensemble strategy R 2 A weighted method is used to construct an ensemble model. The unit models obtained in step 2 are combined into an ensemble model through addition. The performance of the ensemble model is then evaluated. If its accuracy does not meet the requirements, it is optimized. Considering the differences in the fitting performance of the unit models, a... R 2 Weighting method for integrating unit models; Step 4: Use the Ant Colony Optimization Algorithm (ACO) to solve the ensemble model, obtain the optimal extrusion casting process parameters, and verify them. If they meet the requirements, the extrusion casting process parameter design is completed; otherwise, adjust the parameters of the ant colony algorithm and re-optimize and solve the problem. Step 3 is based on the improved model ensemble strategy R 2 The process of constructing an ensemble model using the weight assignment method is as follows: Step 3.1, according to R 2 The weights of each unit model in the ensemble model are defined as follows: in, o The number of unit models, Each model represents R 2 , The corresponding weights assigned to each model; because R 2 Since there are many decimal places in the model, allocating weights according to their size would lead to more complex weight values. Therefore, the weight allocation is based on the following principles: (1) rounding the allocated weights; (2) maximizing the weights of the optimal model. The individual unit models are combined using an additive approach, with their respective weights, to form an ensemble model. The input to the ensemble model is expressed as follows: in T 1. T 2…. T o This represents the output of each unit model; Step 3.2: After establishing the ensemble model, in order to comprehensively evaluate the performance of the ensemble model, the coefficient of determination is selected. R 2 Mean square error MSE Root mean square error RMSE As an evaluation metric for the model, R 2 The value should be between [0, 1], and the closer to 1, the better. MSE , RMSE The smaller the two model error evaluation metrics, the better. The expressions for the three model evaluation metrics are as follows: in, n' The number of test samples, For the first r In the test sample, the first j The model prediction values ​​for each performance metric, For the first r In the test sample, the first j The true value of each performance metric For the first n In the test sample, the first j The true average of each performance metric i and j All represent positive integers; Step 3.3: If the performance of the ensemble model cannot meet the requirements, it needs to be optimized. The optimization method consists of two steps: (1) adjusting the hyperparameters of each unit model to improve its performance, and (2) adjusting the hyperparameters of each unit model according to its new performance. R 2 Recalculate the weights of each model and reconstruct the ensemble model.

2. The method for designing extrusion casting process parameters based on ensemble learning according to claim 1, characterized in that: In step 2, the dataset is processed. DB The process of normalizing and dividing the dataset into training and test sets is as follows: Step 2.1, for the dataset DB The normalization process is performed using the following formula: in, f This represents the data after normalization. u Indicates each attribute in DB The original value in u max and u min Each attribute represents the value in the following ways. DB The maximum and minimum values ​​in; Step 2.2: Divide the preprocessed dataset samples into training set and test set, with the ratio of training set to test set between 1.5 and 3.

3. The method for designing extrusion casting process parameters based on ensemble learning according to claim 1, characterized in that: In step 4, the process of using the Ant Colony Algorithm (ACO) to solve the ensemble model and obtain the extrusion casting process parameters is as follows: The optimization variable is set as a set of extrusion casting process parameters, with each ant representing a set of extrusion casting process parameters, including... v For each process parameter to be designed, the value range of each parameter is normalized and set to [0,1]. Therefore, the ant is denoted as... The optimal casting performance output by the integrated model is used as the target of ant colony optimization, and the range of process parameters is the optimization range of the ant colony. Ants search for paths based on pheromone concentrations; the pheromone concentration along the path determines the ant's location. e At any moment t From current location i' Move to the next location j' The probability of is shown in the following formula: In the formula, For inspirational information, This indicates the magnitude of the ant's expected transfer degree. Indicates starting from the current position i' To the next position j' The distance; Indicates the concentration of pheromones along the path; For the previous position i To the final target location s Inspirational information, This indicates that the ant starts from its current position. i To the final target location s The amount of information, α Indicates the information heuristic factor, β Represents the expected heuristic factor, set Ants e All possible locations to be moved; The pheromone concentration is constantly updated locally; the pheromone concentration along the path evaporates and decreases over time, while ants continuously release new pheromones after passing through the path, thus constantly updating the pheromone concentration. k Current position in seconds i' To the next position j' The pheromone concentrations are shown below: In the formula, This refers to all the ants in the ant colony. m' At the current location i' To the next position j' The total amount of pheromones left on the surface; Indicates the evaporation coefficient of pheromones; starting time t When =0, , ; Set the maximum number of iterations n' The value range is [50, 200], and the number of ants m' The evaporation coefficient is 1 to 3 times that of the training set samples. c' The value range is (0, 0.2], and the pheromone increases in intensity. p' The value range is [1,4], representing the ant crawling speed. a' The range is [0.5, 2].

Images