Extrusion casting performance prediction method based on migration multilayer perceptron

Abstract

The invention provides an extrusion casting performance prediction method based on a migration multi-layer perceptron, and belongs to the field of extrusion casting, and the method comprises the steps: firstly, designing a data migration rule from the perspective of predicted casting performance and from the aspects of materials, process parameters, casting performance and the like, and screening out historical casting data most suitable for migration; on this basis, a pre-training multi-layer perceptron model of Bayesian optimization between process parameters and performance is established, then a small amount of target casting test data is introduced, and a new casting-oriented performance prediction model is finally established through a migration strategy of freezing-fine tuning of hidden layer parameters. The average prediction error of the migration model can be reduced by 80.46% to the maximum compared with that of the base model. Compared with an existing prediction model based on single casting data, the method has the advantages that knowledge migration between technological parameters and casting performance is achieved by applying historical cases, the requirement for training samples is further lowered, the number of times and cost of tests of new castings are reduced, and resources and energy are saved.

Description

Technical Field

[0001] The present invention relates to the field of squeeze casting, and in particular to a method for predicting the properties of squeeze castings based on a transfer multi-layer perceptron. Background Art

[0002] Squeeze casting is a near-net-shape manufacturing process that combines the characteristics of metal casting and die forging. It is widely used in fields such as aerospace, automotive manufacturing, and precision instruments because it can produce high-performance parts.

[0003] The properties of castings are directly related to the reliability and safety of products. Under the condition of determining the material composition, process parameters such as pouring temperature are the most critical factors affecting the properties of squeeze castings. Therefore, establishing the relationship between process parameters and the properties of squeeze castings, and then predicting the properties of squeeze castings under different process parameters, is the basic premise for optimizing the design to obtain higher-performance squeeze casting process parameters. Traditional casting property evaluation mainly relies on direct experimental testing. Although this method is accurate and can guide the selection of better process parameters, it is costly, time-consuming, and cannot be predicted. A small number of researchers have begun to explore using machine learning based on experimental data to predict the properties of castings. Such methods have established high-precision prediction models for casting properties based on experimental data, which can not only reduce the number of experiments, cost, and time consumption, but also provide theoretical guidance and data support for process optimization and product development. However, existing methods usually require a relatively large number of experiments. If the number of experiments and training samples is too small, the established model is difficult to accurately map the coupling relationship between influencing parameters and casting properties, resulting in low prediction accuracy, so the cost is still relatively high. At the same time, current these methods are only limited to using the experimental data of the single casting being studied, and do not involve the use of experimental data of existing (similar castings). Current data has become a new production resource, and using existing data and its related knowledge has become an important direction and inevitable trend of intelligent manufacturing. Therefore, there is an urgent need to establish a method for realizing high-precision prediction of squeeze casting properties based on existing data.

[0004] Transfer learning is a machine learning method that tries to utilize past existing knowledge. It transfers the knowledge of a similar domain (source domain) to a new domain (target domain) by using past identical (or similar) instances, features, model parameters, and relationships, so as to improve the generalization, accuracy, and learning efficiency of the target domain task. Compared with the modeling method that only uses target domain data, the modeling method based on transfer learning can obtain higher model accuracy with fewer training samples, thus reducing the need for training samples. In view of this, in order to use existing data and reduce the number of experiments to predict the properties of squeeze castings, this paper introduces a transfer learning method for model parameters. This method aims to meet the modeling accuracy between casting process parameters and quality or performance indicators, reduce the training data samples of new castings, lower the experimental cost of new castings, and achieve lower-cost and more accurate prediction of casting properties. Summary of the Invention

[0005] The object of the present invention is to provide a method for predicting the performance of extrusion castings based on a transfer multi-layer perceptron, so as to solve the technical problems mentioned in the background art. Transfer learning is a machine learning method that tries to utilize the existing knowledge in the past. It transfers the knowledge in a similar field (source domain) to a new field (target domain) by using the past same (or similar) instances, features, model parameters and relationships, so as to improve the generalization, accuracy and learning efficiency of the target domain task. Compared with the modeling method that only uses the data in the target domain, the modeling method based on transfer learning can obtain higher model accuracy with small training samples, thus reducing the demand for training samples.

[0006] The method first designs data transfer rules from aspects such as materials, process parameters, and casting performance from the perspective of the predicted casting performance, screens out the most suitable historical casting data for transfer, and on this basis, establishes a pre-trained multi-layer perceptron model with Bayesian optimization between process parameters and performance. Then, a small amount of target casting test data is introduced, and through the transfer strategy of freezing-fine-tuning the hidden layer parameters, a performance prediction model for new castings is finally established. The average prediction error of the transfer model can be reduced by up to 80.46% compared with the base model. Compared with the existing prediction models based on single casting data, the proposed method uses historical cases to realize the knowledge transfer between process parameters and casting performance, further reducing the demand for training samples, thereby reducing the number of tests and costs of new castings. It is not only applicable to the performance prediction of extrusion castings, but also applicable to the performance prediction of castings of other casting processes.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A method for predicting the performance of extrusion castings based on a transfer multi-layer perceptron, the method comprising the following steps:

[0009] Step 1: Construct target domain and source domain data;

[0010] Step 2: Construct a transfer multi-layer perceptron model;

[0011] Step 3: Construct transfer data measurement rules;

[0012] Step 4: Realize the performance prediction of extrusion castings based on Bayesian optimization of hyperparameter tuning.

[0013] Further, the specific process in Step 1 is:

[0014] Regard the test data of the target extrusion casting or material to be predicted as the target domain, denoted as D t ,D t involves d process parameters, forming a feature space X t ,and h performance indicators form a label space Yt , then D t The i-th sample data of where Y i t =(y1, y2,..., y h ), take the test data of existing similar squeeze castings as the source domain, denoted as D s , and the corresponding casting is called the source casting. Let D s involve u influencing process parameters, forming the feature space X s , with v performance indicators, the label space Y s , then D s The j-th data of is

[0015] Furthermore, the specific process of step 2 is as follows:

[0016] Take the numericalized squeeze casting process parameters as the input of the multi-layer perceptron MLP (Multi-Layer Perceptron) model, take each performance indicator of the casting as the output, select an MLP model with two hidden layers to build a model based on the test data. To reduce the demand for test samples of the target casting, first select and use the source casting data to establish a pre-trained MLP model, and then adopt the frozen fine-tuning strategy to achieve transfer learning and establish an MLP model for predicting the performance of the actual target casting.

[0017] Furthermore, in step 3, design data transfer rules from aspects of materials, process parameters, and casting performance, and screen out the historical casting data most suitable for transfer.

[0018] Furthermore, the specific process of step 3 is as follows:

[0019] Step 3.1, evaluate the similarity of the material composition of the source casting and the target casting, that is, whether they belong to the same type of alloy. If so, further evaluate the similarity of the test data; otherwise, exclude it.

[0020] Step 3.2, evaluate the similarity of the test process parameters of the source casting and the target casting. Due to the differences in the design experience of different castings, the process parameter test parameters of the source casting and the target casting have different data distributions. Therefore, to make more full use of historical data, it is necessary to evaluate the similarity of the test parameters of the source casting and the target casting. Use the maximum mean discrepancy (MMD) to measure the similarity of the test parameter distributions of the source casting and the target casting. The larger the MMD value, the less similar the two tests are. First, use the kernel function Φ to map the sample process parameters of the source casting and the target casting into the Hilbert space respectively, and then calculate the average difference, as shown in the formula:

[0021]

[0022] In the formula, Z s , Z t are the test process parameters of the source and target castings respectively. The linear kernel and Gaussian kernel are used to measure the similarity of the test parameters from the linear space and the non - linear space respectively. The linear kernel formula Φ l =k(X i , X j ) = X i ·X j , and the Gaussian kernel function formula where σ is the bandwidth, used to control the radiation range, and σ = 1 / d. Substituting the kernel function, the final MMD calculation formula is obtained, as shown in the formula:

[0023]

[0024] Denote the MMD calculated by the linear kernel as MMDL, and the MMD calculated by the Gaussian kernel as MMDG. Experiments show that the value of MMDL is usually greater than the set value, while the value of MMDG is less than 1. Therefore, MMD is transformed into a positive index, and the final MMDL + and MMDG + similarity scores are calculated respectively as:

[0025]

[0026] MMDG + =(1 - MMDG)·100

[0027] In the formula, P is the normalization constant value determined according to the value calculated by the linear kernel. Denote the similarity score of the test parameter distribution as Sim1:

[0028] Sim1 = MMGL + +MMDG + ;

[0029] Step 3.3, measure the similarity of the casting performance indicators;

[0030] Step 3.4, comprehensively obtain the overall similarity of the i - th same indicator and its test parameters of the final source casting and target casting:

[0031] Score = w1·Sim1+w2·Sim2+w3·Sim3

[0032] where w1, w2, and w3 are the weight coefficients of the three similarities, focusing on the similarity of the influence trend of the process parameters on the performance indicators, and secondly on the local numerical performance of the casting performance.

[0033] Furthermore, in Step 3.3, measure the similarity of the performance indicators from the following two aspects;

[0034] ① Measure the statistical characteristics of performance indicators, including six statistical characteristics: minimum value minval, first quartile q1, median q2, third quartile q3, mean value mean, and maximum value maxval. For calculation, it is required that the target casting needs at least 4 groups of tests to obtain statistical characteristics. Suppose there are m identical performance indicators for the source casting and the target casting, that is, Y s ∩Y t =y i (i = 1, 2... m), and form a statistical characteristic vector from the six local statistical characteristics, as shown in the formula:

[0035] SY (i) =(minval, q1, q2, q3, mean, maxval)

[0036] Then the statistical characteristic similarity score Sim2 of the i-th performance indicator between the source casting and the target casting is calculated as:

[0037]

[0038] Among them, SY s (i) and SY t (i) respectively represent the local statistical characteristic vectors of the i-th overlapping performance indicator of the source casting and the target casting;

[0039] ② Use the Pearson correlation coefficient to calculate the correlation between the performance indicators and process parameters of the source casting and the target casting respectively, and measure the similarity of the performance indicators and process parameters between the source casting and the target casting. First, according to the formula:

[0040]

[0041] Calculate the Pearson correlation coefficient between each performance indicator and each process parameter respectively. Since the value of the Pearson correlation coefficient is between -1 and 1, in order to obtain an explicit similarity score, the formula:

[0042]

[0043] is used to transform it to between 0 and 1; then form a correlation vector, as shown in the formula:

[0044] RRY (i) =(RY′ (i,1) ,RY′ (i,2) ,...,RY′ (i,j) )

[0045] In the formula, cov(y i ,x j ) is the i-th performance indicator yi and the jth process parameter x j The covariance of var(y i ) and var(x j ) are and y i x j variance, then the similarity between the correlation between the process parameters and performance indicators of the source casting and the target casting is:

[0046]

[0047] Furthermore, in step 4, in the pre-training stage of the multi-layer perceptron MLP model, it is necessary to determine the number of hidden layer neurons and the activation function, the learning rate, the number of iterations, the weight decay coefficient, and the optimizer. In the fine-tuning stage, the model parameters and optimizer of the training stage are used, and then the set learning rate and the set number of iterations are used to adapt the data performance of the target casting. The remaining training parameters are the default parameters of the optimizer. In order to obtain an intuitive training effect, a negative determination coefficient is used as the objective function.

[0048] Furthermore, in step 4, Bayesian optimization is performed in the given hyperparameter space to obtain the optimal solution θ * :

[0049] argminθ * =-f(θ)

[0050] Where f(θ) is the determination coefficient and θ is an arbitrary value in the hyperparameter space.

[0051] Furthermore, the specific process of Bayesian optimization is:

[0052] Step 4.1, build a proxy model based on the previous sampling points to obtain the probability density functions l(θ) and g(θ), that is, the conditional probability distribution p(θ|R) of the hyperparameter θ under the given objective function value R, as shown in the formula:

[0053]

[0054] Where R = -f(θ); on this basis, the marginal probability distribution p(θ) of the hyperparameter θ is obtained, as shown in the formula:

[0055] p(θ)=γl(θ)+(1-γ)g(θ)

[0056] Where γ = p(R < τ), τ represents the dynamic threshold value obtained based on the observed points;

[0057] Step 4.2, calculate the expected improvement function:

[0058]

[0059] Then combine the formula to get:

[0060]

[0061] Finally, the sampling point θ′ for the next iteration is obtained by maximizing EI, as shown in the formula:

[0062] θ′ = argmax θ EI τ (θ)

[0063] Step 4.3: Update the observed points, and add the new hyperparameter θ′ and its corresponding objective function to the previous sampling points;

[0064] Step 4.4: Repeat Steps 4.1 - 4.3 until the maximum number of iterations is completed;

[0065] Step 4.5: Determine whether the optimal objective function value |R| is greater than 0.95. If so, output the optimal objective function value |R| and its corresponding hyperparameter θ * , otherwise execute again.

[0066] Due to the adoption of the above - mentioned technical solution, the present invention has the following beneficial effects:

[0067] The present invention has successfully realized the application of transfer learning in squeeze casting and performance prediction based on small samples. Application examples prove that the present invention improves the accuracy of the performance prediction model for squeeze castings. Even in the case of small data samples, a high - performance performance prediction model for squeeze castings can be obtained. This can reduce the demand for test data of new castings, is conducive to reducing the number of physical tests, and thus greatly saves costs or time. The more similar the obtained source castings are, the better the performance of the prediction model. The present invention fully considers the similarity of the test data between the existing castings and the newly designed castings, ensuring the transfer application of the existing data and knowledge and the effect of transfer learning. As long as a certain performance index of the new casting has a high similarity with the source casting, the present invention can be used to improve the accuracy and reliability of its prediction model. It expands a new path for the performance prediction of squeeze castings, can also be used for reference in the performance prediction of other fields, and at the same time expands the direction for the use of existing data resources in fields such as squeeze casting, which is conducive to giving play to the value of the enterprise's data resources. Description of the Drawings

[0068] Figure 1 is the overall flowchart of the method of the present invention;

[0069] Figure 2 is the transfer modeling and transfer strategy diagram of the present invention;

[0070] Figure 3 is the data measurement rule diagram of the squeeze casting source castings of the present invention;

[0071] Figure 4It is the process diagram of Model1 and Model2 TPE-BO of the present invention;

[0072] Figure 5 It is the prediction performance result diagram of the pre-trained models Model1 and Model2 of the present invention;

[0073] Figure 6 It is the prediction performance result diagram of different algorithms under a certain sampling of the present invention;

[0074] Figure 7 It is the TPE-BO process diagram of Model3 of the present invention;

[0075] Figure 8 It is the prediction performance diagram of TModel3(6) for each index under a certain sampling of the present invention. Detailed implementation manners

[0076] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the following preferred embodiments are cited with reference to the accompanying drawings for further detailed description of the present invention. However, it should be noted that many details listed in the specification are only for enabling the reader to have a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be implemented even without these specific details.

[0077] As Figure 1-8 shown, a method for predicting the performance of extrusion castings based on a transfer multi-layer perceptron, the method includes the following steps:

[0078] Step 1: Construct target domain and source domain data. The test data of the target extrusion casting (or material) of the new design (to be predicted) is regarded as the target domain, denoted as D t . D t involves d process parameters (constituting the feature space X t ), h performance indicators (constituting the label space Y t ), then the i-th sample data of D t is expressed as where Y i t =(y1, y2,..., y h ); The test data of the existing similar extrusion castings is used as the source domain, denoted as D s , and the corresponding casting is called the source casting. Suppose D s involves u influencing process parameters, constituting the feature space X s , there are v performance indicators, the label space Y s , then the j-th data of D s is As shown in Table 1 - Table 3.

[0079] Table 1 is for the target casting D t Test data

[0080]

[0081] Table 2 is for the source casting D s1 Test data

[0082]

[0083]

[0084] Table 3 is for the source casting D s2 Test data

[0085]

[0086] Step 2: Construct a transfer multi-layer perceptron (MLP) model. Use the numericalized squeeze casting process parameters as the input of the MLP model, and each performance index of the casting as the output. Select an MLP model with two hidden layers to construct the model based on the test data; to reduce the demand for test samples of the target casting, first select and use the source casting data to establish a pre-trained MLP model, and then adopt the freeze-fine-tuning strategy to achieve transfer learning and establish an MLP model for predicting the performance of the actual target casting. The transfer model and transfer strategy are as Figure 2 shown.

[0087] Step 3: Transfer data measurement. The transfer data measurement rules are as Figure 3 shown. According to the designed source casting data measurement transfer rules, obtain the final source casting data similarity score Score as shown in Table 4. From the statistical results, it can be seen that the score of the tensile strength of D s1 is the highest with that of the target casting, followed by the elongation score of D s2 ; finally, it is the hardness index of D s1 . This indicates that the regularities between different performance indexes of the same casting and the process parameters are different. Therefore, transfer learning needs to be carried out with emphasis according to the performance of the performance indexes; secondly, even if they belong to the same type of material and the same process, due to the influence of factors such as materials, molds, and test methods, some performances shown by the castings may also be very different, showing dissimilarity. Therefore, in order to avoid the phenomenon of negative transfer, modeling analysis should be carried out by index. In this case, the modeling analysis will be mainly carried out for the tensile strength to prove the effectiveness of the transfer learning method.

[0088] Table 4 is the similarity of each index of the test data of the target casting and the source casting

[0089]

[0090] Step 4: Hyperparameter Tuning Based on Bayesian Optimization. In the pre-training stage of the MLP model, it is necessary to determine the number of neurons in the hidden layer, activation function, learning rate, number of iterations, weight decay coefficient, and optimizer; in the fine-tuning stage, the model parameters and optimizer used in the training stage are adopted, and then a smaller learning rate and fewer iteration steps are used to adapt to the data performance of the target casting. The remaining training parameters are the default parameters of the optimizer. In order to obtain an intuitive training effect, the negative coefficient of determination is used as the objective function.

[0091] Based on D s1 and D s2 Pre-training models for predicting the tensile strength are established, denoted as Model1 and Model2 respectively. The iterative process of TPE-BO is as Figure 4 shown. Figure 4 The results show that Model1 and Model2 reach the optimal fitting accuracy in the 27th and 17th optimization iterations respectively, which are 0.9952 and 0.9645 respectively, meeting the optimization requirements. The hyperparameters obtained by optimization are shown in Table 5. The prediction performance of Model1 for the tensile strength of the target casting is as Figure 5 , and the prediction performance is shown in Table 6.

[0092] Table 5 shows the TPE-BO optimization results of Model1 and Model2

[0093]

[0094] Table 6 shows the prediction performance of Model1 and Model2

[0095]

[0096] It can be seen that the performance of Model1 is better, and it shows that without fine-tuning using the training data of the target casting, it already has good generalization ability in the target domain. This is mainly because the tensile strength of D s1 has a high similarity with that of the target casting. While the predicted values of Model2 almost completely deviate from the true value curve. Comparing D s1 , the final similarity score of the tensile strength of D s2 is only 85.84% of that of D s1 ; secondly, since this paper pays more attention to the similarity performance of the correlation coefficient between process parameters and casting properties, comparing Sim2, D s2 is only 65.52% of that of D s1 . These differences result in a large prediction error for Model2. This verifies the effectiveness of the transfer rules proposed in this paper and also reflects the application value of the transfer learning method in the performance prediction of extrusion castings.

[0097] Table 7 shows the average results of the model performance obtained by randomly sampling the target casting sample data 5 times according to the requirements of Tasks 1 and 2, Figure 6 which is a comparison of the prediction performance of randomly selected single results. Denote the transfer models for learning Tasks 1 and 2 based on D s1 as TModel1(6) and TModel1(3) respectively, and the transfer models for the corresponding tasks based on D s2 are denoted as TModel2(6) and TModel2(3), and so on for the markings of other models. From Table 7 and Figure 6 it can be seen that, for Tasks 1 and 2, the transfer models established based on D s1 achieved high and the best prediction accuracies, with average errors of only 2.8% and 5.3% respectively, indicating that the transfer learning method in this paper can obtain a high-precision prediction model; while for Tmodel2, whether in Task 1 or Task 2, the prediction errors are relatively large, and its negative transfer effect is significant compared to the base model, indicating that the prediction results of the transfer models established based on dissimilar source domain data are very unreliable; in Task 1, the models based on the tree structure all achieved relatively high prediction accuracies, and the average error of the worst SVR model among the comparison models was only 18.24%. These models are all non-linear models, further proving that there is indeed a complex non-linear relationship between the casting process parameters and the performance parameters.

[0098] Table 2 shows the average results of the prediction performance of different algorithms

[0099]

[0100] To further prove the rationality of the modeling by sub-indicators in this paper, transfer modeling with multiple output indicators is carried out based on the source casting data D s1 During the TPE-BO process of the multi-output indicator transfer model, the objective function is the mean of the sum of the negative determination coefficients of all indicators, and the optimization process is as Figure 7 . It can be seen that the optimal average fitting accuracy of 0.9815 is obtained at the 14th iteration. The results of the hyperparameters are shown in Table 8, where |R1|, |R2|, and |R3| represent the optimal objective function values of the tensile strength, hardness, and elongation respectively. It can be seen that all the casting performance indicators meet the optimization requirements. Denote the pre-trained model as Model3. At this time, the results of Model3 predicting the casting performance indicators of the target casting are shown in Table 9. Only the tensile strength indicator still has a certain degree of prediction generalization, while the hardness and elongation indicators are obviously unreliable. Thus, it can be seen that for dissimilar casting performance indicators, the direct prediction error of the pre-trained model is large. To compare with the modeling effect of TModel1 in Section 2.3.1, Model3 is fine-tuned under the same sampling and TModel3 is obtained. And the prediction effects of TModel3 and TModel1 on the tensile strength are compared, and the results are shown in Table 10.

[0101] Table 8 shows the Bayesian hyperparameter tuning results for Model 3

[0102]

[0103] Table 9 shows the results of predicting all performance indicators of the target casting for Model 3

[0104]

[0105] Table 10 shows the average prediction performance and comparison of TModel3 for tensile strength

[0106]

[0107] As shown in Table 10, compared with the multi-output TModel3, the prediction performance of TModel1 has been significantly improved. That is, when modeling by integrating dissimilar casting performance indicators, it instead leads to a decrease in the modeling accuracy of similar indicators, causing a significant negative transfer effect. It should be noted that since in this case, the process parameters of the target casting have a low correlation with both hardness and elongation, the prediction error after fine-tuning is still relatively large, as shown in Figure 8 (b) and (c). This indicates that when performing transfer modeling between process parameters and performance indicators in the case of weak correlation, the source casting and the target casting need to have a higher degree of similarity.

[0108] Matters not covered by this invention are well-known techniques.

[0109] The above are only the preferred embodiments of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for predicting extrusion casting performance based on a transfer multi-layer perceptron, characterized in that: The method comprises the following steps: Step 1: Construct target domain and source domain data; Step 2: Construct and migrate several layers of perceptron models; Step 3: Construct migration data measurement rules; Step 4: Hyperparameter tuning based on Bayesian optimization to predict the performance of extruded castings.

2. The method for predicting extrusion casting performance based on a transfer multi-layer perceptron according to claim 1, characterized in that: The specific process in step 1 is: The test data of the target extrusion casting or material to be predicted is regarded as the target domain, denoted as D t , D t Involving d process parameters, forming the feature space X t , h performance indicators form the label space Y t , then D t The i-th sample data is expressed as in Y i t =(y1,y2,...,y h ), the experimental data of similar extrusion castings are used as the source domain, denoted as D s , the corresponding casting is called the source casting, let D s Involving u influencing process parameters, forming the feature space X s , there are v performance indicators, label space Y s , then D s The jth data is 3. The method for predicting extrusion casting performance based on a transfer multi-layer perceptron according to claim 1, characterized in that: The specific process of step 2 is: The numerical squeeze casting process parameters are used as the input of the multi-layer perceptron MLP model, and each performance indicator of the casting is used as the output. An MLP model with two hidden layers is selected to build the model based on the experimental data. In order to reduce the demand for test samples of target castings, the source casting data is first selected and used to establish a pre-trained MLP model. Then, the frozen fine-tuning strategy is used to realize transfer learning and establish an MLP model for performance prediction of the actual target casting.

4. The method for predicting extrusion casting performance based on a transfer multi-layer perceptron according to claim 1, characterized in that: In step 3, data migration rules are designed from the aspects of materials, process parameters, and casting performance to select the most suitable historical casting data for migration.

5. The method for predicting extrusion casting performance based on transfer multi-layer perceptron according to claim 4, characterized in that: The specific process of step 3 is: Step 3.1, evaluate the material composition similarity between the source casting and the target casting, that is, whether they belong to the same alloy. If yes, further evaluate the similarity of the test data, otherwise exclude them; Step 3.2, evaluate the similarity of the experimental process parameters of the source casting and the target casting. Due to the difference in casting design experience, the experimental process parameters of the source casting and the target casting have different data distributions. Therefore, in order to make full use of historical data, it is necessary to evaluate the similarity of the experimental parameters of the source casting and the target casting. The maximum mean difference is used to measure the similarity of the experimental parameter distribution of the source casting and the target casting. The larger the MMD value, the less similar the two experiments are. First, the kernel function Φ is used to map the process parameters of the source casting and the target casting samples to the Hilbert space respectively, and then the average difference is calculated, as shown in the formula: In the formula, Z s , Z t are the experimental process parameters of the source and target castings, respectively. The linear kernel and Gaussian kernel are used to measure the similarity of the experimental parameters from the linear space and nonlinear space respectively. The linear kernel formula Φ l = k(X i ,X j )=X i ·X j , Gaussian kernel function formula Among them, σ is the bandwidth, which is used to control the radiation range. Take σ = 1 / d, substitute the kernel function, and get the final MMD calculation formula, as shown in the formula: The MMD calculated by the linear kernel is MMDL, and the MMD calculated by the Gaussian kernel is MMDG. Experiments show that the value of MMDL is usually greater than the set value, while the value of MMDG is less than 1. Therefore, MMD is converted into a positive indicator, and the final MMDL + and MMDG + The similarity scores are calculated as: MMDG + =(1-MMDG)·100 Where P is the normalized constant value determined based on the linear kernel calculation value, and the test parameter distribution similarity score is Sim1: Sim1=MMGL + +MMDG + ; Step 3.3, measuring the similarity of casting performance indicators; Step 3.4, comprehensively obtain the overall similarity of the i-th identical index and its test parameters between the final source casting and the target casting: Score=w1·Sim1+w2·Sim2+w3·Sim3 Among them, w1, w2, and w3 are weight coefficients of three similarities, respectively, which focus on the similarity of the influence trend of process parameters on performance indicators, followed by the local numerical performance of casting performance.

6. The method for predicting extrusion casting performance based on transfer multi-layer perceptron according to claim 5, characterized in that: In step 3.3, the similarity of performance indicators is measured from the following two aspects; ① Statistical characteristics of performance indicators, including minimum minval, quarter quantile q1, median q2, three quarter median q3, mean, and maximum maxval. To achieve calculation, the target casting needs at least 4 groups of tests to obtain statistical characteristics. Suppose the source casting and the target casting have m identical performance indicators, that is, Y s ∩Y t =y i (i=1,2...m), the six local statistical features are combined into a statistical feature vector, as shown in the formula: <h2 style=";text-align:left;direction:ltr">SY<h2 style=";text-align:left;direction:ltr"> (i) <h2 style=";text-align:left;direction:ltr"> =(minval,q1,q2,q3,mean,maxval) Then the statistical feature similarity score Sim2 of the i-th performance index of the source casting and the target casting is calculated as: Among them, SY s (i) and SY t (i) denote the local statistical feature vectors of the i-th overlapping performance index of the source casting and the target casting respectively; ② Use the Pearson correlation coefficient to calculate the correlation between the performance indicators and process parameters in the source casting and the target casting, respectively, and measure the similarity of the performance indicators and process parameters in the source casting and the target casting. First, according to the formula: The Pearson correlation coefficient between each performance indicator and each process parameter is calculated respectively. Since the Pearson correlation coefficient value is between -1 and 1, in order to obtain an explicit similarity score, the formula is used: Convert it to between 0 and 1; then form a correlation vector, as shown in the formula: RRY (i) =(RY′ (i,1) ,RY′ (i,2) ,...,RY′ (i,j) ) In the formula, cov(y i ,x j ) is the i-th performance indicator y i and the jth process parameter x j The covariance of var(y i ) and var(x j ) are and y i x j variance, then the similarity between the correlation between the process parameters and performance indicators of the source casting and the target casting is:

7. The method for predicting extrusion casting performance based on transfer multi-layer perceptron according to claim 1, characterized in that: In step 4, during the pre-training stage of the multi-layer perceptron MLP model, it is necessary to determine the number of hidden layer neurons and the activation function, learning rate, number of iterations, weight decay coefficient, and optimizer. In the fine-tuning stage, the model parameters and optimizer of the training stage are used, and then the set learning rate and the set number of iterations are used to adapt the data performance of the target casting. The remaining training parameters are the default parameters of the optimizer. In order to obtain an intuitive training effect, a negative determination coefficient is used as the objective function.

8. The method for predicting extrusion casting performance based on transfer multi-layer perceptron according to claim 1, characterized in that: In step 4, Bayesian optimization is performed in the given hyperparameter space to obtain the optimal solution θ * : argminθ * =-f(θ) Where f(θ) is the determination coefficient and θ is an arbitrary value in the hyperparameter space.

9. The method for predicting extrusion casting performance based on transfer multi-layer perceptron according to claim 8, characterized in that: The specific process of Bayesian optimization is: Step 4.1, build a proxy model based on the previous sampling points to obtain the probability density functions l(θ) and g(θ), that is, the conditional probability distribution p(θ|R) of the hyperparameter θ under the given objective function value R, as shown in the formula: Where R = -f(θ); on this basis, the marginal probability distribution p(θ) of the hyperparameter θ is obtained, as shown in the formula: p(θ)=γl(θ)+(1-γ)g(θ) Where γ = p(R < τ), τ represents the dynamic threshold value obtained based on the observed points; Step 4.2, calculate the expected improvement function: Then combine the formula to get: Finally, the sampling point θ′ of the next iteration is obtained by maximizing EI, as shown in the formula: θ′=argmax θ IE τ (i) Step 4.3, update the observed points and add the new hyperparameter θ′ and its corresponding objective function to the previous sampling points; Step 4.4, repeat steps 4.1 to 4.3 until the maximum number of iterations is completed; Step 4.5: Determine whether the optimal objective function value |R| is greater than 0.

95. If so, output the optimal objective function value |R| and its corresponding hyperparameter θ * , otherwise re-execute.

Images