Slush molding mold performance prediction method based on machine learning

Abstract

The invention provides a machine learning-based slush mold performance prediction method, which comprises the following steps of: firstly, collecting process parameters and mold performance parameters of heat treatment and electroplating processes in a manufacturing process of a slush mold, constructing a basic database and carrying out data preprocessing; then taking P20 steel as an example, constructing a plurality of CNN models and LSTM models by using the performance parameters and the process parameters of the mold, and training the CNN models and the LSTM models; carrying out weighted fusion through a bagging algorithm to obtain a mold performance prediction model made of P20 steel; the method is used for predicting the incidence relation between process parameters and performance parameters. For materials homologous with P20 steel, a mold performance basic prediction model made of 718 steel and NAK80 steel is obtained through transfer learning. According to the method, prediction results of a plurality of algorithms are fused, and better generalization performance and robustness are achieved after fusion. And by utilizing a transfer learning algorithm, longitudinal transfer among different mold materials is realized, the workload of machine learning is greatly reduced, and the time is saved.

Description

Technical Field

[0001] The present invention relates to a method for predicting the performance of a slush mold based on machine learning, and belongs to the technical field of machine learning. Background Art

[0002] Slush molds are made using the electroforming process and are primarily used in the production of automotive dashboards, interior trims, and slush toys. Slush molds are widely used in the production of plastic products, particularly high-end automotive dashboards. However, mold manufacturing cycles are long, lifespans are short, costs are high, and mold repairs are difficult. Furthermore, the slush molding process requires rapid mold heating and cooling, resulting in high energy consumption. Therefore, predicting slush mold performance is of great significance for the slush molding process. In recent years, machine learning has become a cutting-edge research area, offering significant advantages in predictive efficiency and cost control. Convolutional neural networks, a deep feedforward neural network with convolutional operations, are highly capable of handling complex nonlinear problems and offer excellent feature extraction capabilities. Using convolutional neural networks, models are developed to correlate slush mold material type, mold design parameters, molding parameters, and mold properties, such as mold fatigue life. The model's predicted values are compared with actual values through visualization. Furthermore, the model's predictions on a test set are also analyzed. This allows users to flexibly customize the model to meet their specific needs and adapt it to different scenarios, reducing production costs and improving efficiency. Based on the above reasons, there is an urgent need to develop a prediction method for slush mold performance parameters based on machine learning to save R&D and testing costs. Summary of the Invention

[0003] Based on this, the present invention provides a machine learning-based method for predicting the performance of slush molds. This method utilizes an ensemble learning bagging algorithm to fuse the prediction results of multiple convolutional neural network (CNN) and long short-term memory (LSTM) networks to obtain a prediction model for slush mold performance parameters. Furthermore, a transfer learning algorithm (TCA) is used to achieve longitudinal transfer between different mold materials, based on the different correspondences between feature inputs and mold performance parameter outputs.

[0004] A method for predicting the performance of a slush mold based on machine learning, characterized by comprising the following steps:

[0005] S1 Data collection and data preprocessing: Collect the process parameters of heat treatment and electroplating process and mold performance parameters in the manufacturing process of slush molds, build a basic database; and preprocess the basic data set;

[0006] S2 builds and trains predictive models related to mold performance and manufacturing processes:

[0007] Taking P20 steel as an example, multiple CNN models and multiple LSTM models were constructed based on the correlation data of mold strength and heat treatment temperature, stiffness and electroplating solution type, fatigue life and temperature fluctuation time series, and melt density thickness and electroplating solution type.

[0008] CNN model: The input is the process parameter matrix. The convolution layer extracts the spatial correlation features between the process parameters, the pooling layer reduces the dimension, and the fully connected layer outputs the performance parameters.

[0009] LSTM model: The input is a time series of process parameters, the memory unit captures the long-term dependencies between parameters, and outputs performance parameters;

[0010] By adjusting the hyperparameters of each CNN model and LSTM model, and using the basic dataset for model training;

[0011] S3 constructs the source network through the integrated algorithm: in the trained CNN model of strength, stiffness, fatigue life, and melt density thickness, the R 2 The a-group CNN algorithm and the b-group LSTM algorithm with a value of more than 0.98 are weightedly fused through the bagging algorithm, and the performance parameter prediction value is the weighted average of the outputs of each model; finally, the basic prediction model of the mold performance based on P20 steel material is obtained.

[0012] Furthermore, the process parameters of the electroplating process include the type of electroplating solution, pH value, and temperature; the heat treatment process parameters include the heat treatment temperature; and the mold performance parameters include strength, stiffness, fatigue life, and melt-densified thickness.

[0013] Furthermore, the data preprocessing in S1 specifically includes:

[0014] S1.1 Detect and handle missing values and outliers in datasets: by filling missing values or deleting sample data containing missing values; detect outliers through statistical or visualization methods and handle them by deletion, replacement, or transformation;

[0015] S1.2 Standardize the data set using formula (1).

[0016] Min-Max normalization: X' = (X - X_min) / (X_max - X_min) (1)

[0017] Where X' is the normalized value, X is the original value, X_min is the minimum value of the data, and X_max is the maximum value of the data;

[0018] S1.3: To handle imbalanced datasets, use oversampling, undersampling, or synthesizing minority classes to balance the data, transform the data, improve its distribution, or increase the complexity of features.

[0019] S1.4: Divide the dataset into training set and test set.

[0020] Furthermore, in S2, the hyperparameters of CNN that need to be adjusted are the number of convolutional layers, pooling window, and number of filters; the hyperparameters of LSTM are the number of hidden layers, number of neurons, and dropout rate; the hyperparameter adjustment method is:

[0021] CNN network: Grid search optimization for 2-5 convolutional layers, 16-128 filters, and pooling window size from 2×2 to 4×4;

[0022] LSTM: Bayesian optimization was used to adjust the number of hidden layers to 1-3, the number of neurons to 32-256, and the dropout rate to 0.2-0.5.

[0023] Furthermore, in S2, the hyperparameter adjustment method is:

[0024] CNN network: Grid search optimization: 3 convolution layers, 64 filters, and 3×3 pooling window size;

[0025] LSTM: Bayesian optimization was used to adjust the number of hidden layers to 2, the number of neurons to 128, and the dropout rate to 0.3.

[0026] Furthermore, in S3, the model evaluation after model training uses the determination coefficient R 2 The calculation formula is:

[0027]

[0028] is the model performance prediction result; y m is the performance parameter value in the test set; is the average value of the performance parameter in the test set; q is the total number of data in the test set;

[0029] Using K-fold cross validation, we calculate the weights of each CNN algorithm and LSTM algorithm. The weight formula is as follows:

[0030]

[0031] Where: k represents the kth performance parameter; i = 1, 2, ..., a; j = 1, 2, ..., b; is the determination coefficient of the i-th CNN algorithm for the k-th performance parameter, w CNNik is the weight of the kth performance parameter of the i-th CNN algorithm, is the determination coefficient of the j-th LSTM algorithm for the k-th performance parameter, w LSTMjk is the weight of the j-th CNN algorithm for the k-th performance parameter;

[0032] The formula of the prediction model of the abrasive tool performance based on P20 steel obtained by the weighted average voting method is as follows:

[0033]

[0034] Where: y predk is the predicted performance parameter of the kth performance parameter, y CNNik is the predicted value of the i-th CNN algorithm for the k-th performance parameter, y LSTMjk The j-th LSTM algorithm predicted value for the k-th performance parameter.

[0035] Furthermore, the machine learning-based slush mold performance prediction method also includes step S4 transfer learning optimization: through the TCA (transfer component analysis) algorithm, the source domain model knowledge based on P20 material is migrated to the target domain of 718 steel and NAK80 steel mold materials to obtain the basic prediction model of mold performance with materials of 718 steel and NAK80 steel.

[0036] Furthermore, the specific steps of transfer learning in S4 are: inputting the data of the source material and the target material into the TCA algorithm, finding the common subspace of the source space and the target space, and mapping the original features to the subspace;

[0037] Map the source domain and target domain to the new feature space so that the distribution distance between the mapped source domain and target domain is minimized; the objective function is:

[0038]

[0039] Where: X s is the source domain data feature matrix, Where d is the feature dimension, n s is the number of source domain samples;

[0040] X t is the target domain data feature matrix, where n t is the number of samples in the target domain;

[0041] H is the centering matrix, I is the identity matrix, 1 is the all-one vector, n=n s +n t ;

[0042] A is the mapping matrix A∈R d×k , map the original features to k-dimensional space;

[0043] λ is the regularization parameter, and the optimal value is determined by grid search combined with cross-validation; Frobenius norm: ||.|| F represents the Frobenius norm of the matrix;

[0044] Data centering: The global mean is subtracted from the source domain and the target domain to eliminate the offset.

[0045]

[0046] Construct the joint covariance matrix and calculate the difference between the source domain and the target domain:

[0047]

[0048] By solving the mapping matrix A through generalized eigenvalue decomposition, the eigenvectors corresponding to the first K largest eigenvalues are selected to form A;

[0049] (MM T +λI)A=γA

[0050] The formula for mapping the source domain and target domain to the common subspace is:

[0051] Z S =A T X S ,Z t =A T X t

[0052] In subspace Z S The model is trained on Z t Fine-tune the parameters to achieve performance migration of CNN and LSTM on the target material; use the migrated model to predict the mold performance in the target domain; if the predicted performance is qualified, add the new data to the basic dataset; otherwise, re-optimize λ or adjust the mapping dimension k.

[0053] Ensemble learning algorithms (bagging) achieve diversity across base learners by creating multiple CNN and LSTM model instances and training them using different subsets of sample data. After training, the predictions of these base learners can be combined using voting averaging (a regression task). In this scenario, different CNN and LSTM models capture different features and dependencies; therefore, by combining the predictions of multiple models, the ensemble model achieves better generalization and robustness. Transfer learning (TCA) algorithms are used to achieve vertical transfer between different mold materials, with different correspondences between feature inputs and mold performance parameter outputs. Transfer learning can complete predictions using a small amount of data from the target material, significantly reducing the machine learning workload and saving time. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of the machine learning-based slush mold performance prediction method described in the present invention.

[0055] Figure 2 Flowchart of transfer learning between different materials. DETAILED DESCRIPTION

[0056] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0057] like Figure 1 As shown, the method for predicting the performance of a slush mold based on machine learning according to the present invention is characterized in that it includes the following steps:

[0058] S1 Data Collection and Data Preprocessing: Collect process parameters and mold performance parameters of the heat treatment and electroplating processes during the manufacturing process of the P20 steel slush mold to build a basic data set. The electroplating process parameters include the type of electroplating solution, pH value, and temperature; the heat treatment process parameters include the heat treatment temperature; and the mold performance parameters include strength, stiffness, fatigue life, and melt density thickness. The basic data set is then preprocessed. The data preprocessing specifically includes:

[0059] S1.1 Detect and handle missing values and outliers in datasets: by filling missing values or deleting sample data containing missing values; detect outliers through statistical or visualization methods and handle them by deletion, replacement, or transformation;

[0060] S1.2 Standardize the data set using formula (1).

[0061] Min-Max Normalization: X' = (X - X_min) / (X_max - X_min)

[0062] Where X' is the normalized value, X is the original value, X_min is the minimum value of the data, and X_max is the maximum value of the data;

[0063] S1.3: To handle imbalanced datasets, use oversampling, undersampling, or synthesizing minority classes to balance the data, transform the data, improve its distribution, or increase the complexity of features.

[0064] S1.4: Divide the dataset into a training set and a test set. Select 80% of the data as the training set and the remaining 20% as the test set.

[0065] S2 builds and trains predictive models related to mold performance and manufacturing processes:

[0066] Heat treatment temperature directly affects material phase transitions and is therefore a key parameter for strength. The type and concentration of the electroplating solution affect coating uniformity and density, respectively, so stiffness and melt thickness are separately matched. The temperature fluctuation time series reflects thermal cycling stress and is strongly correlated with fatigue life. Using P20 steel as an example, we constructed multiple CNN and LSTM models based on correlation data between mold strength and heat treatment temperature, stiffness and electroplating solution type, fatigue life and temperature fluctuation time series, and melt density thickness and electroplating solution type.

[0067] The CNN model takes as input a matrix of process parameters, such as the local interaction between temperature and pH. The convolutional layer extracts the spatial correlation features between these parameters, the pooling layer performs dimensionality reduction, and the fully connected layer outputs performance parameters. The LSTM model takes as input a time series of process parameters, such as the fluctuation of heat treatment temperature over time. The memory cells capture the long-term dependencies between these parameters and output performance parameters. This is achieved by adjusting the hyperparameters of each CNN and LSTM model and training them using a basic dataset. The CNN hyperparameters that require adjustment are the number of convolutional layers, pooling window, and number of filters; the LSTM hyperparameters are the number of hidden layers, number of neurons, and dropout rate.

[0068] The relationship between the mold performance and process parameters and the key points of hyperparameter adjustment of the CNN model and LSTM model are shown in Table 1.

[0069] Table 1 Relationship between mold performance and process parameters and key points of hyperparameter adjustment for CNN model and LSTM model

[0070]

[0071]

[0072] Specifically, the hyperparameters that need to be adjusted for CNN are the number of convolutional layers, pooling window, and number of filters; the hyperparameters for LSTM are the number of hidden layers, number of neurons, and dropout rate. The hyperparameter adjustment method is:

[0073] CNN networks: Grid search optimizes for 2-5 convolutional layers, 16-128 filters, and a pooling window size of 2×2 to 4×4. Preferably, grid search optimizes for 3 convolutional layers, 64 filters, and a pooling window size of 3×3.

[0074] LSTM: Bayesian optimization is used to adjust the number of hidden layers to 1-3, the number of neurons to 32-256, and the dropout rate to 0.2-0.5. Preferably, Bayesian optimization is used to adjust the number of hidden layers to 2, the number of neurons to 128, and the dropout rate to 0.3.

[0075] Table 2 Hyperparameter adjustment of CNN and LSTM

[0076]

[0077] Both CNN and LSTM algorithms use the Relu (Reinforced Luminance) function, which is characterized by fast convergence and no vanishing gradients. The Relu activation function helps improve network performance, accelerate training convergence, conserve computing resources, and reduce storage costs when dealing with large amounts of prediction data. As shown in the formula, the Relu function has a value of 0 when X is less than 0 and a value of X when X is greater than 0.

[0078] Relu(x)=max{0,x}

[0079] In regression tasks, the coefficient of determination (R 2 ) is used as an indicator to evaluate the accuracy of the model. 2 The range is between 0 and 1. The closer the value is to 1, the higher the degree of explanation of the dependent variable by the explanatory variable (i.e., the degree of fit) and the better the model fitting effect. 2 The algorithm proceeds to step S3 when the coefficient of determination R is above 0.98. 2 The calculation formula is:

[0080]

[0081] is the model performance prediction result; y m is the performance parameter value in the test set; is the average value of the performance parameter in the test set; q is the total number of data in the test set;

[0082] For example, the CNN algorithm is used to predict the heat treatment holding time alone, with 376 sets of yield strength, tensile strength and hardness as input values and heat treatment holding time as output value. The CNN algorithm hyperparameters are shown in the table. Training set R 2 The test set R 2 It is 0.982.

[0083] Input layer [4,1] Number of convolutional layers 3 Convolutional layer [3,1] Maximum number of training sessions 1200 Pooling Window [2,1] Initial learning rate 0.01 step length [1,1] Learning rate after 800 epochs 0.001

[0084] S3 constructs the source network through an integrated algorithm:

[0085] In the integration stage, the prediction robustness is improved by weight distribution complementation. In the trained CNN model of strength, stiffness, fatigue life, and melt density thickness, R 2The a-group CNN algorithm and the b-group LSTM algorithm with a value of more than 0.98 are weightedly fused through the bagging algorithm, and the performance parameter prediction value is the weighted average of the outputs of each model; finally, the basic prediction model of the mold performance based on P20 steel material is obtained.

[0086] Specifically, K-fold cross validation is used to calculate the weights of each CNN algorithm and LSTM algorithm. The weight formula is as follows:

[0087]

[0088] Where: k represents the kth performance parameter; i = 1, 2, ..., a; j = 1, 2, ..., b; is the determination coefficient of the i-th CNN algorithm for the k-th performance parameter, w CNNik is the weight of the kth performance parameter of the i-th CNN algorithm, is the determination coefficient of the j-th LSTM algorithm for the k-th performance parameter, w LSTMjk is the weight of the kth performance parameter of the jth CNN algorithm.

[0089] The formula of the prediction model of the abrasive tool performance based on P20 steel obtained by the weighted average voting method is as follows:

[0090]

[0091] Where: y predk is the predicted performance parameter of the kth performance parameter, y CNNik is the predicted value of the i-th CNN algorithm for the k-th performance parameter, y LSTMjk The j-th LSTM algorithm predicted value for the k-th performance parameter.

[0092] For 718 steel and NAK80 steel, materials similar to P20 steel, transfer learning optimization is performed in step S4 to achieve cross-material feature distribution adaptation. Specifically, the TCA (Transfer Component Analysis) algorithm is used to transfer the source domain model knowledge based on P20 material to the target domain of 718 steel and NAK80 steel mold materials, thereby obtaining a basic prediction model for mold performance of 718 steel and NAK80 steel.

[0093] The specific steps of transfer learning are: input the data of source material and target material into the TCA algorithm, find the common subspace of the source space and target space, and map the original features to the subspace;

[0094] Map the source domain and target domain to the new feature space so that the distribution distance between the mapped source domain and target domain is minimized; the objective function is:

[0095]

[0096] Where: X s is the source domain data feature matrix, Where d is the feature dimension, n s is the number of source domain samples;

[0097] X t is the target domain data feature matrix, where n t is the number of samples in the target domain;

[0098] H is the centering matrix, I is the identity matrix, 1 is the all-one vector, n=n s +n t ;

[0099] A is the mapping matrix A∈R d×k , map the original features to k-dimensional space;

[0100] λ is the regularization parameter, and the optimal value is determined by grid search combined with cross-validation; Frobenius norm: ||.|| F represents the Frobenius norm of the matrix;

[0101] Data centering: The global mean is subtracted from the source domain and the target domain to eliminate the offset.

[0102]

[0103] Construct the joint covariance matrix and calculate the difference between the source domain and the target domain:

[0104]

[0105] By solving the mapping matrix A through generalized eigenvalue decomposition, the eigenvectors corresponding to the first K largest eigenvalues are selected to form A;

[0106] (MM T +λI)A=γA

[0107] The formula for mapping the source domain and target domain to the common subspace is:

[0108] Z S =A T X S ,Z t =A T X t

[0109] In subspace Z S The model is trained on Z tFine-tune the parameters to achieve performance migration of CNN and LSTM on the target material; use the migrated model to predict the mold performance in the target domain; if the predicted performance is qualified, add the new data to the basic dataset; otherwise, re-optimize λ or adjust the mapping dimension k.

[0110] The embodiments described are preferred implementations of the present invention, but the present invention is not limited to the above implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention are within the scope of protection of the present invention.

Claims

1. A method for predicting the performance of slush molds based on machine learning, characterized in that: The following steps are involved: S1 Data collection and data preprocessing: Collect the process parameters of heat treatment and electroplating process and mold performance parameters in the manufacturing process of slush molds, build a basic database; and preprocess the basic database; S2 builds and trains predictive models related to mold performance and manufacturing processes: Taking P20 steel as an example, multiple CNN models and multiple LSTM models were constructed based on the correlation data of mold strength and heat treatment temperature, stiffness and electroplating solution type, fatigue life and temperature fluctuation time series, and melt density thickness and electroplating solution type. CNN model: The input is the process parameter matrix. The convolution layer extracts the spatial correlation features between the process parameters, the pooling layer reduces the dimension, and the fully connected layer outputs the performance parameters. LSTM model: The input is a time series of process parameters, the memory unit captures the long-term dependencies between parameters, and outputs performance parameters; By adjusting the hyperparameters of each CNN model and LSTM model, and using the basic dataset for model training; S3 constructs the source network through the integrated algorithm: in the trained CNN model of strength, stiffness, fatigue life, and melt density thickness, the R 2 The a-group CNN algorithm and the b-group LSTM algorithm with a value of more than 0.98 are weightedly fused through the bagging algorithm, and the performance parameter prediction value is the weighted average of the outputs of each model; finally, the basic prediction model of the mold performance based on P20 steel material is obtained.

2. The method for predicting the performance of a slush mold based on machine learning according to claim 1, wherein: The process parameters of the electroplating process include the type of electroplating solution, pH value, and temperature; the heat treatment process parameters include the heat treatment temperature; and the mold performance parameters include strength, stiffness, fatigue life, and melt-densified thickness.

3. The method for predicting the performance of a slush mold based on machine learning according to claim 1, wherein: The data preprocessing in S1 specifically includes: S1.1 Detect and handle missing values and outliers in datasets: by filling missing values or deleting sample data containing missing values; detect outliers through statistical or visualization methods and handle them by deletion, replacement, or transformation; S1.2 Standardize the data set using formula (1). Min-Max normalization: X' = (X - X_min) / (X_max - X_min) (1) Where X' is the normalized value, X is the original value, X_min is the minimum value of the data, and X_max is the maximum value of the data; S1.3: To handle imbalanced datasets, use oversampling, undersampling, or synthesizing minority classes to balance the data, transform the data, improve its distribution, or increase the complexity of features. S1.4: Divide the dataset into training set and test set.

4. The method for predicting the performance of a slush mold based on machine learning according to claim 1, wherein: In S2, the hyperparameters of CNN that need to be adjusted are the number of convolutional layers, pooling window, and number of filters; the hyperparameters of LSTM are the number of hidden layers, number of neurons, and dropout rate. The hyperparameter adjustment method is: CNN network: Grid search optimization for 2-5 convolutional layers, 16-128 filters, and pooling window size from 2×2 to 4×4; LSTM: Bayesian optimization was used to adjust the number of hidden layers to 1-3, the number of neurons to 32-256, and the dropout rate to 0.2-0.

5.

5. The method for predicting the performance of a slush mold based on machine learning according to claim 4, wherein: In S2, the hyperparameter adjustment method is: CNN network: Grid search optimization: 3 convolution layers, 64 filters, and 3×3 pooling window size; LSTM: Bayesian optimization was used to adjust the number of hidden layers to 2, the number of neurons to 128, and the dropout rate to 0.

3.

6. The method for predicting the performance of a slush mold based on machine learning according to claim 4, wherein: In S3, the model evaluation after model training uses the determination coefficient R 2 The calculation formula is: is the model performance prediction result; y m is the performance parameter value in the test set; is the average value of the performance parameter in the test set; q is the total number of data in the test set; Using K-fold cross validation, we calculate the weights of each CNN algorithm and LSTM algorithm. The weight formula is as follows: Where: k represents the kth performance parameter; i = 1, 2, ..., a; j=1,2,…,b; is the determination coefficient of the i-th CNN algorithm for the k-th performance parameter, w CNNik is the weight of the kth performance parameter of the i-th CNN algorithm, is the determination coefficient of the j-th LSTM algorithm for the k-th performance parameter, w LSTMjk is the weight of the j-th CNN algorithm for the k-th performance parameter; The formula of the prediction model of the abrasive tool performance based on P20 steel obtained by the weighted average voting method is as follows: Where: y predk is the predicted performance parameter of the kth performance parameter, y CNNik is the predicted value of the i-th CNN algorithm for the k-th performance parameter, y LSTMjk The j-th LSTM algorithm predicted value for the k-th performance parameter.

7. The method for predicting the performance of a slush mold based on machine learning according to claim 1, wherein: It also includes step S4 transfer learning optimization: through the TCA migration component analysis algorithm, the source domain model knowledge based on P20 material is transferred to the target domain of 718 steel and NAK80 steel mold materials to obtain the basic prediction model of mold performance made of 718 steel and NAK80 steel.

8. The method for predicting the performance of a slush mold based on machine learning according to claim 1, wherein: The specific steps of transfer learning in S4 are: input the data of source material and target material into the TCA algorithm, find the common subspace of source space and target space, and map the original features to the subspace; Map the source domain and target domain to the new feature space so that the distribution distance between the mapped source domain and target domain is minimized; the objective function is: Where: X s is the source domain data feature matrix, Where d is the feature dimension, n s is the number of source domain samples; X t is the target domain data feature matrix, where n t is the number of samples in the target domain; H is the centering matrix, I is the identity matrix, 1 is the all-one vector, n=n s +n t ; A is the mapping matrix A∈R d×k , map the original features to k-dimensional space; λ is the regularization parameter, and the optimal value is determined by grid search combined with cross-validation; Frobenius norm: ||.|| F represents the Frobenius norm of the matrix; Data centering: The global mean is subtracted from the source domain and the target domain to eliminate the offset. Construct the joint covariance matrix and calculate the difference between the source domain and the target domain: By solving the mapping matrix A through generalized eigenvalue decomposition, the eigenvectors corresponding to the first K largest eigenvalues are selected to form A; (MM T +λI)A=γA The formula for mapping the source domain and target domain to the common subspace is: Z S =A T X S ,Z t =ATX t In subspace Z S The model is trained on Z t Fine-tune parameters to achieve performance transfer of CNN and LSTM on the target material; Use the migrated model to predict the target domain mold performance; if the predicted performance is qualified, add the new data to the basic database; otherwise, re-optimize λ or adjust the mapping dimension k.

Images