Construction method of PREP powder particle size prediction model

Abstract

The invention discloses a construction method of a PREP powder particle size prediction model, and relates to the field, and the technical scheme of the invention comprises the following steps: firstly, collecting supplementary characteristic parameters through literature data collection and thermodynamic calculation; then, data preprocessing is carried out; secondly, selecting an optimal model by evaluating the performance of different machine learning algorithms, and optimizing model parameters by utilizing grid search to further improve the model performance; and finally, verifying the generalization ability of the model on unseen data through comparison with experimental powder making data. According to the method, the PREP technology with higher precision and efficiency can be provided, the quality of products produced by the constructed model has better stability and consistency, the prediction accuracy is remarkably improved, and a scientific basis is provided for optimization of process parameters.

Description

Technical Field

[0001] The present invention relates to a field of application, and in particular to a method for constructing a PREP powder particle size prediction model. Background Art

[0002] Plasma Rotating Electrode Process (PREP), an advanced metal powder preparation method, has attracted widespread attention for its ability to produce powders with high sphericity, excellent flowability, and minimal satellite and hollow powder content. These powder properties are particularly critical for metal additive manufacturing (AM), as they directly impact the quality and performance of the final part. PREP technology has been widely applied in strategic sectors such as aerospace, biomedicine, and automotive manufacturing.

[0003] However, despite the outstanding performance of PREP technology in metal powder preparation, its complex system of process parameters presents challenges in accurately predicting powder particle size. Traditional particle size prediction methods are often based on empirical formulas. While these formulas account for factors such as alloy density and surface tension, the accuracy and applicability of their predictions are often limited by the diversity and complexity of PREP process parameters. This limitation increases the difficulty of process control, which in turn affects the stability and consistency of product quality. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a method for constructing a PREP powder particle size prediction model, which can provide PREP technology with higher accuracy and efficiency. The constructed model outputs product quality with good stability and consistency, significantly improves the accuracy of the prediction, and provides a scientific basis for the optimization of process parameters.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a method for constructing a PREP powder particle size prediction model, comprising the following steps:

[0006] First, supplementary characteristic parameters were collected through literature data collection and thermodynamic calculations. Then, data preprocessing was performed. Then, the optimal model was selected by evaluating the performance of different machine learning algorithms, and the model parameters were optimized using grid search to further improve the model performance. Finally, the model's generalization ability on unseen data was verified by comparison with experimental milling data.

[0007] Preferably, relevant data are extracted from literature on PREP technology research, including alloy composition, key process parameters and powder performance indicators;

[0008] With alloy composition as input parameters, thermodynamic properties including surface tension, viscosity, and density were calculated using thermodynamic calculations and empirical formulas from the literature for characteristic parameters.

[0009] About the dataset:

[0010]

[0011] Q is the melting rate; I is the current intensity; ΔH is the melting heat; α is the thermal efficiency of the plasma gun, which is taken as 0.35; U is the plasma gun voltage, which is taken as 55V.

[0012] Preferably, regarding data preprocessing,

[0013] The formula to eliminate the difference in magnitude between different parameters is:

[0014]

[0015] Where, v norm is the normalized eigenvalue, v represents the original value of a feature in the sample, ∑ i v i Represents the sum of the feature in all samples, that is, the accumulation of the feature values in all samples;

[0016] Correlation of parameters:

[0017]

[0018] X and Y represent the sample values of two variables, E represents the mean, σ represents the standard deviation, and ρ X,Y Represents the correlation coefficient between two variables.

[0019] Preferably, a comparative evaluation of the models is performed:

[0020] The following parameters are selected as evaluation indicators:

[0021] Coefficient of determination R 2 :

[0022]

[0023] Mean Absolute Error (MAE):

[0024]

[0025] Root mean square error (RMSE)

[0026]

[0027] y i is the true value, is the predicted value, is the average of the true values, and m represents the total number of samples.

[0028] The method for constructing a PREP powder particle size prediction model according to claim 1, characterized in that:

[0029] Dataset division and preliminary evaluation:

[0030] Use the train_test_split function: randomly divide the dataset into training and test sets, use candidate models to train the training sets separately, and use the test sets to verify their prediction performance.

[0031] Preferably, the k-fold cross-validation method is used: the dataset is divided into K subsets, each subset is independent and of the same size; in each round of iteration, K-1 subsets are used as training data, and the remaining subsets are used as validation data; through K iterations, a different subset is used as the validation set in each iteration, and finally the performance of the model is evaluated by calculating the average of all iteration results.

[0032] Beneficial effects of the present invention:

[0033] The construction method of this solution can provide PREP technology with higher accuracy and efficiency. The constructed model has good stability and consistency in product quality, significantly improves the accuracy of prediction, and provides a scientific basis for the optimization of process parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only three of the drawings of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 The data set of an embodiment of the present invention; (a) data distribution radar chart (b) data heat map;

[0036] Figure 2 The performance of four models in the train_test_split of the present invention: (a) XGboost model; (b) LightGBM model; (c) KNN model; (d) GBDT model;

[0037] Figure 3 The performance of four models in the embodiment of the present invention in k-fold: (a) XGboost model; (b) LightGBM model; (c) KNN model; (d) GBDT model; DETAILED DESCRIPTION

[0038] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0039] Example

[0040] like Figure 1 As shown, the following steps:

[0041] First, supplementary characteristic parameters are collected through literature data collection and thermodynamic calculations;

[0042] Literature data, with the help of Origin software, relevant data were extracted from the literature, which covered alloy composition, key process parameters (such as electrode rod speed, rod diameter, current intensity, feed rate, etc.), and powder performance indicators (such as average particle size D 50 ). The constructed dataset brings together more than 100 sets of experimental data ] , providing a basis for subsequent data analysis and model building.

[0043] Thermodynamic parameters combine empirical models with thermodynamic theory. Using alloy composition as input, thermodynamic calculations and empirical formulas from the literature are used to calculate thermodynamic properties such as surface tension, viscosity, and density for use as characteristic parameters. These thermodynamic properties serve as supplementary input features to improve the model's predictive accuracy.

[0044] Dataset,

[0045]

[0046] Q is the melting rate; I is the current intensity; ΔH is the melting heat; α is the thermal efficiency of the plasma gun, which is taken as 0.35; U is the plasma gun voltage, which is taken as 55V.

[0047] Next, data preprocessing is performed;

[0048] When dealing with missing values in data,

[0049] For key parameters such as electrode rod rotation speed and bar diameter, a strategy of deleting data containing these missing values was adopted.

[0050] For missing values of parameters such as current intensity and feed rate, the hot-calculation method based on similar observations was used to improve the completeness of the dataset.

[0051] Data distribution:

[0052] Combine Figure 1The radar chart shown in (a) illustrates the characteristics of the dataset, which covers key PREP process parameters, including rod diameter (15mm to 80mm), electrode rod speed (6000rpm to 30000rpm), current intensity (50A to 2600A), and feed rate (0.5mm / s to 1.8mm / s). The dataset also includes the average powder particle size D50 (45μm to 650μm), reflecting typical PREP process conditions and product characteristics. Thermodynamic parameters such as melting point, surface tension, density, and heat of fusion obtained by thermodynamic calculations show diversity, revealing the essential differences between different alloys. However, the orders of magnitude of the data vary greatly. To eliminate the problem that the machine learning model may be overly biased towards characteristic parameters with larger values due to the order of magnitude differences between different parameters, the sum normalization method is used. The following formula processes the values of the characteristic parameters.

[0053]

[0054] Where, v norm is the normalized eigenvalue, v represents the original value of a feature in the sample, ∑ i v i Represents the sum of the feature in all samples, that is, the accumulation of the feature values in all samples;

[0055] The Pearson correlation coefficient is used to express the correlation of each parameter, as shown in the following formula:

[0056]

[0057] X and Y represent the sample values of two variables, E represents the mean, σ represents the standard deviation, and ρ X,Y Represents the correlation coefficient between two variables.

[0058] The closer the absolute value of the correlation coefficient is to 1, the stronger the correlation is. The heat map in Figure (b) shows the correlation between the parameters in the data set. The results show that there is a strong correlation between the current intensity I and parameters such as the melting rate Q and the electrode rod diameter d; in addition, the average specific heat capacity C also shows a strong correlation with the viscosity μ and the melting point density ρ. Considering that data multicollinearity may have an adverse effect on model training, such as causing overfitting or model instability, the current intensity I and the average specific heat capacity C are considered to be potential collinear variables during the model construction process and are excluded to improve the stability and prediction accuracy of the model. Therefore, the final model input parameters have 10 independent variables, and the output parameter is the powder particle size D 50 .

[0059] Then, the optimal model is selected by evaluating the performance of different machine learning algorithms, and the model parameters are optimized using grid search to further improve the model performance;

[0060] 3 Model selection and comparison

[0061] This example uses the JupyterNotebook program to compare and evaluate the performance of four machine learning algorithms, XGBoost, LightGBM, KNN, and GBDT, on a data set. The XGBoost algorithm effectively controls the complexity of the model by integrating regularization terms, thereby effectively preventing overfitting and enhancing the generalization ability of the model. The LightGBM algorithm uses a gradient boosting framework to continuously improve prediction accuracy and performs well in regression problems. The KNN algorithm, because of its characteristic of not assuming data distribution, is suitable for processing nonlinear classification and regression problems, and shows great potential for processing complex data sets. GBDT gradually improves the accuracy of the model by iteratively fitting the residuals, and can usually achieve higher accuracy than traditional regression methods.

[0062] In order to objectively evaluate the performance of each model in the dataset, this paper uses the following parameters as evaluation indicators:

[0063] Coefficient of determination R 2 :

[0064]

[0065] Mean Absolute Error (MAE):

[0066]

[0067] Root mean square error (RMSE)

[0068]

[0069] y i is the true value, is the predicted value, is the average of the true values, and m represents the total number of samples.

[0070] R 2 The closer the value is to 1 and the lower the MAE and RMSE values are, the better the model fitting effect is.

[0071] Finally, the model’s generalization ability on unseen data was verified by comparing it with experimental milling data.

[0072] The dataset is randomly divided into a training set (70%) and a test set (30%) using the train_test_split function in order to compare the performance of different machine learning models on the dataset. Subsequently, four candidate models, namely XGBoost, LightGBM, KNN, and GBDT, are trained on the training set respectively, and their prediction performance is verified using the test set. The experimental results are shown in Figure 2 , where the horizontal axis represents the true value and the vertical axis represents the predicted value. The results show that after training, the XGBoost model performs best on the test set, with an R 2 The R value of the LightGBM model is 0.9386, the MAE is 16.46, and the RMSE is 25.36. 2 The value is 0.7032, MAE is 36.94, and RMSE is 49.66, which shows relatively weak performance.

[0073] Using only a single training set and test set to compare model effects is accidental, and this method may not fully reflect the true performance of the machine learning model. Therefore, this embodiment also adopts the k-fold cross-validation method. The data set is divided into K subsets, each subset is independent of each other and of the same size (ie, "fold"). In each round of iteration, K-1 subsets are used as training data, and the remaining subsets are used as verification data. Through K iterations, each iteration uses a different subset as a verification set, and finally the performance of the model is evaluated by calculating the average of all iteration results. In this study, the K value is set to 4, that is, the data set is divided into four parts. In each iteration, three parts of the data are used for training, and the remaining part is used for verification, and the entire verification process is repeated four times. The verification results are as follows Figure 3 As shown in Figure 2, the horizontal axis represents the true value and the vertical axis represents the predicted value. The results show that the XGBoost model has a better performance in cross-validation than the 2 The value is 0.908, MAE is 8.12, and RMSE is 19.76, which is the best performance. 2 The value is 0.7032, MAE is 43.01, and RMSE is 57.73, which is relatively insufficient.

[0074] Comprehensive analysis results show that the XGBoost model performs well, outperforming other candidate models, whether using the basic train_test_split method or the more rigorous k-fold cross-validation method. Therefore, it can be concluded that the XGBoost model is more suitable for the characteristics of the current dataset and was selected as the primary algorithm for the powder particle size prediction model.

[0075] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0076] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for constructing a PREP powder particle size prediction model, characterized in that: The steps are as follows: First, supplementary characteristic parameters were collected through literature data collection and thermodynamic calculations. Then, data preprocessing was performed. Then, the optimal model was selected by evaluating the performance of different machine learning algorithms, and the model parameters were optimized using grid search to further improve the model performance. Finally, the model's generalization ability on unseen data was verified by comparison with experimental milling data.

2. The method for constructing a PREP powder particle size prediction model according to claim 1, characterized in that: Relevant data including alloy composition, key process parameters and powder performance indicators were extracted from the literature of PREP technology research; With alloy composition as input parameters, thermodynamic properties including surface tension, viscosity, and density were calculated using thermodynamic calculations and empirical formulas from the literature for characteristic parameters. About the dataset: Q is the melting rate; I is the current intensity; ΔH is the melting heat; α is the thermal efficiency of the plasma gun, which is taken as 0.35; U is the plasma gun voltage, which is taken as 55V.

3. The method for constructing a PREP powder particle size prediction model according to claim 1, wherein: Regarding data preprocessing, The formula to eliminate the difference in magnitude between different parameters is: Where, v norm is the normalized eigenvalue, v represents the original value of a feature in the sample, ∑ i v i Represents the sum of the feature in all samples, that is, the accumulation of the feature values in all samples; Correlation of parameters: X and Y represent the sample values of two variables, E represents the mean, σ represents the standard deviation, and ρ X,Y Represents the correlation coefficient between two variables.

4. The method for constructing a PREP powder particle size prediction model according to claim 1, wherein: Perform comparative evaluation of the models: The following parameters are selected as evaluation indicators: Coefficient of determination R 2 : Mean Absolute Error (MAE): Root mean square error (RMSE) y i is the true value, is the predicted value, is the average of the true values, and m represents the total number of samples.

5. The method for constructing a PREP powder particle size prediction model according to claim 1, wherein: Dataset division and preliminary evaluation: Use the train_test_split function: randomly divide the dataset into training and test sets, use candidate models to train the training sets separately, and use the test sets to verify their prediction performance.

6. The method for constructing a PREP powder particle size prediction model according to claim 1, characterized in that: Use the k-fold cross-validation method: the dataset is divided into K subsets, each subset is independent and of the same size; in each round of iteration, K-1 subsets are used as training data, and the remaining subsets are used as validation data; through K iterations, a different subset is used as the validation set in each iteration, and finally the performance of the model is evaluated by calculating the average of all iteration results.

Images