Sintered neodymium-iron-boron magnet performance prediction method based on data driving

Abstract

The invention discloses a sintered neodymium-iron-boron magnet performance prediction method based on data driving, which is used for gradually capturing the influence of process parameters in the production process and gradually predicting the middle and final magnet performance. In the first stage, technological parameters are used as input, the characteristics of powder are predicted through the combined prediction model, and intermediate output characteristics of the first stage are obtained. And in the second stage, training and prediction are carried out again in combination with the process parameters and the powder performance parameters obtained through prediction in the first stage, so that the final performance of the magnet is obtained. Through the multi-stage mode, the model can gradually accumulate and extract the influence of each process link, which is beneficial to reveal the complex mass transfer mechanism of the process. The stage training not only improves the prediction precision, but also enhances the interpretability of the model, so that the key process links can be better identified in the production process, thereby improving the consistency of the product quality and the optimization efficiency of the process flow.

Description

Technical Field

[0001] The invention belongs to the technical field of production and quality control of sintered NdFeB magnets, and in particular is a method for predicting the performance of sintered NdFeB magnets based on a data-driven method. Background Art

[0002] With the booming domestic economy and technological progress, the requirements for the preparation of sintered NdFeB magnets in terms of process and performance are constantly improving, and higher standards are also put forward for the efficiency and environmental protection of the use of rare earth elements. The main components of sintered NdFeB magnets are neodymium (Nd), iron (Fe) and boron (B), and the typical chemical formula is Nd2Fe14B. Sintered NdFeB has high magnetic energy product, high coercivity and good temperature stability, which means that they can generate a large magnetic field in a small volume, can maintain their magnetism in a wide operating temperature range, and their operating temperature range can be increased by adding other elements such as cobalt (Co) and dysprosium (Dy). The production of NdFeB magnets usually adopts powder metallurgy, and the main steps include raw material preparation, smelting, powder making, orientation pressing, sintering, heat treatment, machining, surface treatment and magnetization. This series of processes ensures the high performance and reliability of the magnets.

[0003] At present, sintered NdFeB magnets are widely used in various fields due to their excellent magnetic properties, such as in the automotive industry for motors, sensors and other components of electric vehicles and hybrid vehicles; in the wind power industry for permanent magnet synchronous motors in wind turbines; in the medical device industry for magnets in magnetic resonance imaging (MRI) equipment. Although sintered NdFeB magnets have many advantages, there are still some problems in their production process. They are mainly divided into two categories: one is the cost problem, the price fluctuation of rare earth elements will affect the production cost; the other is the environmental impact, the mining and processing of rare earth ores may cause environmental pollution. Therefore, how to control the quality of the magnetism of sintered NdFeB magnets has become an important step. In the production process of sintered NdFeB magnets, quality control mainly relies on the following aspects: Raw material inspection: Strict quality inspection of purchased raw materials to ensure that they meet production standards. Process control: During the production process, key process parameters such as temperature and pressure are monitored in real time through online detection equipment. Finished product inspection: Comprehensive performance testing of the final product, including magnetic properties, dimensional accuracy, etc. Experience accumulation: Rely on the experience of engineers and technicians to adjust process parameters to achieve the best production results.

[0004] Although existing technologies can meet production needs to a certain extent, there are still some obvious limitations: Low data utilization: A large amount of data generated during the production process has not been effectively utilized, and data mining and analysis capabilities are insufficient. Strong dependence on manual labor: The production process is highly dependent on the experience of technicians and lacks systematic and automated process optimization methods. Difficulties in process optimization: Traditional methods are difficult to reveal the complex genetic laws of quality, which leads to greater uncertainty in optimizing production processes. Limited production efficiency: Due to the lack of effective tool support, there is limited room for improvement in production efficiency and product quality consistency. Summary of the invention

[0005] In view of the problems existing in the quality control of existing sintered NdFeB magnets, the present invention proposes a data-driven sintered NdFeB magnet performance prediction method, which integrates advanced data analysis technology and machine learning algorithms. It can monitor and analyze the key parameters in the production process of sintered NdFeB magnets in real time, thereby predicting product performance and optimizing the production process.

[0006] The data-driven sintered NdFeB magnet performance prediction method of the present invention comprises the following steps:

[0007] Step 1: Collect data including process parameters, powder characteristics and magnet properties. Powder characteristics include SMD and D99 / D10. Magnet properties include remanence, coercive force and maximum magnetic energy product.

[0008] Step 2: Establish a one-stage prediction model, including a D99 / D10 prediction model obtained by taking process parameters as input and powder characteristic data D99 / D10 as output. And an SMD prediction model obtained by taking process parameters as input and powder characteristic data SMD as output. The specific method is:

[0009] A. Data preprocessing to ensure data quality and consistency.

[0010] B. Divide the preprocessed data into a training set and a test set, and independently train the input data using the support vector regression model, the extreme gradient boosting model, and the elastic network model; and determine the prediction accuracy and generalization ability of each model;

[0011] C. Combine grid search and Bayesian optimization methods to fine-tune the parameters of the training model obtained in B.

[0012] D. For the prediction model after parameter adjustment, the weighted fusion method based on Shapley value is used to integrate the model and construct an integrated model based on Shapley value.

[0013] Step 3: Use the method in step 2 to establish a two-stage prediction model, including a remanence prediction model that takes process parameters and powder properties as input and remanence in magnet properties as output; a coercive force prediction model that takes process parameters and powder properties as input and coercive force in magnet properties as output; and a maximum energy product prediction model that takes process parameters and powder properties as input and maximum energy product in magnet properties as output.

[0014] Finally, a multi-stage magnet performance prediction model is obtained, in which the model input obtained in step 2 is the process parameters, and the output is the powder performance data. The process parameters are combined as the input of the model obtained in step 3, and the output is the magnet performance.

[0015] The advantages of the present invention are:

[0016] 1. The data-driven sintered NdFeB magnet performance prediction method of the present invention is a comprehensive data processing method. By integrating the relationship between multi-dimensional parameters in the production process (such as grinding chamber pressure, initial oxygen content, sorting wheel speed, etc.) and the final magnet performance (remanence, coercive force, maximum magnetic energy product), it realizes the efficient prediction of the quality inheritance law in the production process of sintered NdFeB. This method not only improves the accuracy of the prediction, but also simplifies the data processing process, so that non-professionals can also quickly understand and apply it.

[0017] 2. The present invention has developed a set of high-precision prediction models based on a data-driven sintered NdFeB magnet performance prediction method. By using advanced machine learning algorithms (XGBoost, ElsticNet, SVR, etc.), it can find hidden patterns in complex data sets and predict the performance of sintered NdFeB magnets. At the same time, the Shapley values ​​of the ElasticNet, SVR and XGBoost models are used to construct a multi-model fusion prediction model based on contribution analysis. Compared with traditional statistical methods, this model has significantly improved the prediction accuracy, which helps to discover potential quality problems in advance.

[0018] 3. The data-driven sintered NdFeB magnet performance prediction method of the present invention has built a multi-stage magnet performance prediction model based on multi-model fusion prediction, which realizes the online prediction of magnet performance and can automatically adjust the process parameters according to the real-time monitored production environment changes, so as to achieve the best production efficiency and product quality. This dynamic adjustment method can greatly reduce the need for manual intervention, reduce production costs, and improve the consistency and stability of finished products. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The figure is an overall flow chart of the data-driven sintered NdFeB magnet performance prediction method of the present invention. DETAILED DESCRIPTION

[0020] The present invention is further described in detail below with reference to the accompanying drawings.

[0021] The present invention is based on a data-driven sintered NdFeB magnet performance prediction method, such as Figure 1 As shown, the specific steps are:

[0022] Step 1: Data Collection

[0023] The acquired data include process parameters, powder characteristics and final magnet performance.

[0024] Among them, process parameters are important conditions that affect the production process and product quality, including mixing time, initial powder weight, silicone oil content, oxygen content, rotation speed, material weight, feeding current, grinding pressure and powder output rate.

[0025] Powder property data is the basis of sintered NdFeB magnet properties, including SMD (volume weighted mean diameter, that is, the average diameter of all particles calculated by volume), D99 / D10 (D99 represents the particle diameter corresponding to when the cumulative frequency reaches 99%, which means that 99% of the particles have a diameter less than or equal to this value. D10 represents the particle diameter corresponding to when the cumulative frequency reaches 10%, which means that 10% of the particles have a diameter less than or equal to this value. The D99 / D10 ratio describes the width or uniformity of the particle size distribution. It reflects the span from the smallest to the largest particle size.).

[0026] The final magnet performance parameters can directly reflect the quality and performance of the product, including remanence, coercive force and maximum magnetic energy product.

[0027] The above data collection provides extensive and detailed data support for the subsequent modeling process.

[0028] Step 2: Take the process parameters as input and the powder characteristic data D99 / D10 as output to obtain the best prediction model for D99 / D10; the specific method is as follows:

[0029] 201. Data Preprocessing

[0030] The data collected in step 1 often contain missing values, outliers, and features of different dimensions. Therefore, the original data of process parameters and powder properties collected in step 1 are cleaned, normalized, feature selected, and missing value processed to ensure data quality and consistency.

[0031] First, the raw data is cleaned and the statistical method Z-score is used to identify and remove noise and outliers, which helps to improve the credibility of the data and ensure the effectiveness of subsequent analysis. Z-score, also known as standard score or standardized value, converts data points into deviation measures relative to the mean. Z-score represents the number of standard deviations a data point differs from the mean.

[0032] Subsequently, the cleaned data is normalized, and all data features are adjusted to the same scale range by applying the Min-Max Scaling technique, reducing the scale differences between data features, so that each input variable can be treated more fairly in the subsequent model training process.

[0033] Further, the normalized data is screened out with features that are highly correlated with the output target variable (D99 / D10) through correlation coefficient analysis, recursive feature elimination (RFE), and LASSO regression methods, and the results obtained by the three methods are mutually verified to obtain the final features that are highly correlated with the target variable, specifically: the features selected by the above three methods are intersected to obtain a common feature set; the intersection represents the features that all methods believe have a greater impact on the target variable. For example: suppose that the correlation coefficient analysis selects features A, B, C, and D; RFE selects features B, C, E, and F; and LASSO regression selects features A, C, F, and G. The intersection of the three is feature C, which means that feature C is considered to have a strong correlation with the target variable in all three methods, and therefore can be used as the final selected feature. Among them, the correlation coefficient analysis method is used in statistics to measure the strength of the linear relationship between two variables, and the correlation coefficient in the present invention is the Pearson correlation coefficient. RFE is a general feature selection method that can be used in conjunction with any machine learning model that can assign importance scores or weights to features; the present invention uses recursive feature elimination (RFE) combined with a random forest method, which identifies and selects the most valuable features for prediction by repeatedly training the model and removing the least important features. LASSO (Least Absolute Shrinkage and Selection Operator) regression is a form of linear regression that adds an L1 regularization term to the loss function.

[0034] Correlation coefficient analysis can not only simplify the model structure and reduce the risk of overfitting, but also speed up training and improve prediction accuracy.

[0035] Finally, missing data is processed by inserting the mean to ensure the integrity and consistency of the data, thus laying a good foundation for subsequent model training.

[0036] 202. Model Training

[0037] The input data processed by step 201, including process parameter data and D99 / D10 data, are divided into a training set and a test set in a ratio of 8:2. Support vector regression (SVR), extreme gradient boosting (XGBoost), and elastic network (Elastic Net) models are used to independently train the input data processed by step 201 to capture the complex relationship between input features and output features; and the prediction accuracy and generalization ability of each model are judged by comparing the evaluation indicators of each model (i.e., R2 score, RMSE and MAE).

[0038] 203. Model Parameter Optimization

[0039] The performance of the model depends largely on the choice of hyperparameters. In order to find the optimal hyperparameter combination, the present invention combines the grid search and Bayesian optimization methods to fine-tune the parameters of the training model obtained in step 3 to obtain a single prediction model with the best performance.

[0040] First, a preliminary exploration of the hyperparameter space is performed through grid search, using a wide range of values ​​to select key hyperparameters.

[0041] Subsequently, a grid search is performed to evaluate the basic performance of each hyperparameter combination. The grid search method finds a better initial configuration by exhaustively enumerating all possible hyperparameter combinations. On this basis, Bayesian optimization is further used to fine-tune the hyperparameters.

[0042] Specifically:

[0043] All specified hyperparameter value combinations are exhaustively tested through grid search, and the best performing hyperparameter combination is selected based on the grid search results as the starting point for Bayesian optimization. Here, "best" is determined based on the selected evaluation metric, which is R. 2 Fraction.

[0044] Bayesian optimization is used to perform a more detailed search near the best hyperparameter combination selected by grid search to find the optimal solution. Bayesian optimization constructs a surrogate model (such as a Gaussian process) to predict the performance of different parameter combinations and conducts a detailed search in the area with better performance, thereby efficiently finding the global optimal solution.

[0045] The above combined optimization strategy can significantly improve the predictive performance and robustness of the model.

[0046] 204. Model Optimization

[0047] The sub-prediction models with better performance obtained in step 203 are integrated using a weighted fusion method based on Shapley values.

[0048] The Shapley value is derived from game theory and is used to quantify the contribution of each model in the overall prediction. By analyzing the Shapley value of each model in the prediction process, higher weights can be assigned to models with better performance. Therefore, weights are assigned to each sub-model based on the Shapley value of each model, and the model prediction results are weighted and summed according to these weights to construct an integrated model based on the Shapley value. This model can comprehensively utilize the advantages of each sub-model, reduce the uncertainty of a single model, and make the final prediction results more robust and accurate. In addition, the application of the Shapley value also improves the interpretability of the integrated model, can fairly evaluate the contribution of each feature to the prediction results, ensure that the importance of each model is reasonably reflected in the multi-model integration process, and provide a stronger basis for process optimization.

[0049] Step 3: With process parameters as input and powder property data SMD as output, repeat the above step 2 process to obtain the optimal SMD prediction model; this model is combined with the D99 / D10 prediction model established in step 2 as a one-stage prediction model, with process parameters as input and powder performance data as output, as an intermediate quantity for magnet performance prediction.

[0050] Step 4: With process parameters and powder properties as input and remanence in magnet properties as output, repeat the above step 2 to obtain the optimal remanence prediction model;

[0051] Step 5: With the process parameters and powder properties as input and the coercive force in the magnet properties as output, repeat the above step 2 to obtain the optimal coercive force prediction model;

[0052] Step 6: With the process parameters and powder properties as input and the maximum magnetic energy product among the magnet properties as output, repeat the above step 2 to obtain the optimal maximum magnetic energy product prediction model;

[0053] The model established in the above steps 4 to 6 is used as a two-stage fusion prediction model, and the input is the process parameters and the powder performance data obtained by the first-stage prediction to realize the magnet performance prediction.

[0054] Through the optimal sub-prediction models obtained from steps 2 to 6, a complete multi-stage magnet performance prediction model is constructed, which can further improve the accuracy and robustness of the prediction and ensure that the magnet performance can be stably predicted in practical applications.

[0055] Based on the above method, a real-time prediction system can be developed to receive process parameters as input and output predicted magnet performance. The online prediction function can realize real-time monitoring and optimization of the production process, improve production efficiency and product quality. By building a visual interactive platform, users can input production parameters through an intuitive graphical interface and view prediction results and optimization suggestions in real time, which enhances the user experience. This interactive method makes the operation easier, and even operators without professional background can quickly master the usage skills.

[0056] In summary, the present invention divides the experimental monitoring data into process parameters, powder performance parameters and magnet performance parameters. The powder performance is predicted with the process parameters as input, and the magnet performance is predicted with the process parameters and the predicted powder performance parameters as input. By comparing the powder performance and magnet performance in the actual detection, the model can provide highly reliable magnet performance parameter prediction results. From a diagnostic point of view, this method can not only predict the final performance of the magnet, but also help identify the key influencing factors in the production process. For example, by comparing the step-by-step prediction of the powder performance and the magnet performance with the actual values, possible process deviations or anomalies can be quickly discovered, providing a clear basis for process adjustments in the production process. This method enables the model to have a function similar to "virtual diagnosis". Through early prediction and comparison, problems can be discovered in the early stage, and measures can be taken to prevent production defects. This diagnostic prediction process significantly improves the real-time and accuracy of the prediction, and at the same time provides more targeted decision support for the production process of sintered NdFeB magnets.

[0057] Furthermore, the data-driven sintered NdFeB magnet performance prediction method of the present invention adopts a multi-model fusion weighted prediction method, trains input data based on three models: support vector regression, extreme gradient boosting, and elastic network, and uses grid search and Bayesian optimization methods to finely adjust the parameters of these models to obtain a single prediction model with the best performance;

[0058] Before fusion prediction, the single models (SVR, XGBoost, Elastic Net) are trained on the input data independently. By comparing the performance of each model, the advantages and disadvantages of each model can be identified. After the performance comparison and evaluation of each model, the grid search method and Bayesian optimization method are further used to fine-tune the parameters of these models. First, the hyperparameter space is roughly explored, and the key hyperparameters are selected using grid search and set a wide range of values ​​for them. On this basis, a grid search is performed to obtain the basic performance of each hyperparameter. Based on the grid search results, the hyperparameter configuration with better performance is selected as the starting point of Bayesian optimization, and then a more refined search is performed near these configurations through Bayesian optimization to find the optimal solution.

[0059] Furthermore, the present invention assigns weights to each sub-model based on the Shapley value of each sub-model, and weighted sums the model results according to these weights. The Shapley value method is used to intuitively display the contribution of each input feature in predicting magnet performance, and weights are assigned by analyzing the Shapley value of each model, thereby constructing an integrated model based on the Shapley value.

[0060] The advantage of weighted integration based on Shapley values ​​is that it can fully utilize the advantages of each sub-model. Different sub-models may perform differently when processing different features or samples. Through the weighted method based on Shapley values, models that perform better in some aspects can have a greater impact on the final results, thereby improving the overall prediction performance of the integrated model. This weighting method helps to build a more robust and accurate integrated model and overcome problems such as overfitting or bias that a single model may face. In addition, the introduction of Shapley values ​​makes the entire process more interpretable, and can clarify the contribution of each model and input feature, thereby providing a basis for optimizing and improving the model.

[0061] Furthermore, the present invention constructs a multi-stage (cross-process) magnet performance prediction method, using the constructed powder performance prediction sub-model to predict the powder performance and obtain the output characteristics of the first stage. These output characteristics contain information about the impact of the process on the powder performance, providing more representative input for subsequent stages.

[0062] In the multi-stage prediction process, the prediction model is not only used to obtain the key output of each stage, but also to quantify the impact of each input variable on the prediction result through variable importance analysis (such as feature importance score or Shapley value analysis). This method helps to discover the process parameters that have a greater impact on magnet performance, thereby providing a scientific basis for subsequent process optimization. Through variable importance analysis, the model can identify the most critical features, gradually capture the complex relationships that affect magnet performance, and help achieve more accurate predictions and more effective production control.

[0063] With process parameters and output characteristics of the first stage as input and magnet performance parameters as output, the magnet performance prediction sub-model is used to make predictions for the second stage. This method helps to better reveal the "heritability" of quality. In the production process, magnet performance is often affected by process operations and intermediate results at each stage. Therefore, stage-by-stage predictions can better capture the transfer process of these qualities. Through step-by-step analysis and prediction, the output of each stage can be effectively utilized, thereby constructing a quality prediction chain that is more in line with the actual production process, which helps to reveal the impact of different processes on the quality of the final product, thereby providing a more targeted basis for process optimization.

Claims

1. A data-driven sintered NdFeB magnet performance prediction method, characterized in that: The specific steps are: Step 1: Collect data including process parameters, powder characteristics and magnet properties; the powder characteristics are SMD and D99 / D10; the magnet properties include remanence, coercive force and maximum magnetic energy product; Step 2: Establish a one-stage prediction model, including a D99 / D10 prediction model obtained by taking process parameters as input and powder property data D99 / D10 as output; and an SMD prediction model obtained by taking process parameters as input and powder property data SMD as output; the specific method is: A. Data preprocessing to ensure data quality and consistency; B. Divide the preprocessed data into a training set and a test set, and use the support vector regression model, the extreme gradient boosting model, and the elastic network model to independently train the input data; And judge the prediction accuracy and generalization ability of each model; C. Combine grid search and Bayesian optimization methods to fine-tune the parameters of the training model obtained in 202; D. For the prediction models after parameter adjustment, the weighted fusion method based on Shapley value is used to integrate these models; and an integrated model based on Shapley value is constructed; Step 3: Use the method in step 2 to establish a two-stage prediction model, including a remanence prediction model that takes process parameters and powder properties as input and remanence in magnet properties as output; a coercive force prediction model that takes process parameters and powder properties as input and coercive force in magnet properties as output; and a maximum energy product prediction model that takes process parameters and powder properties as input and maximum energy product in magnet properties as output.

2. A data-driven sintered NdFeB magnet performance prediction method as claimed in claim 1, characterized in that: In step A of step 2, data preprocessing includes data cleaning, normalization, feature selection and processing of missing values; data cleaning uses the Z-score method to identify and remove noise and outliers; normalization processing uses the minimum-maximum scaling technique to adjust all data features to the same scale range; feature selection uses correlation coefficient analysis, recursive feature elimination, and LASSO regression methods to screen out features that are highly correlated with the output target variable, and the results obtained by the three methods are mutually verified to obtain the final features that are highly correlated with the target variable; Missing values ​​are handled by interpolating the mean.

3. A data-driven sintered NdFeB magnet performance prediction method as claimed in claim 1, characterized in that: In step C of step 2, the training model parameter adjustment method is: First, a rough exploration of the hyperparameter space is performed through grid search. The key hyperparameters are selected using grid search and a wide range of values ​​is set for them. Then a grid search is performed to obtain the baseline performance of each hyperparameter; Furthermore, Bayesian optimization was used to further fine-tune the hyperparameters; Through the grid search results, a hyperparameter combination is selected according to the evaluation index as the starting point of Bayesian optimization; then, a detailed search is performed in these hyperparameter combination areas through Bayesian optimization to find the optimal solution.

4. A data-driven sintered NdFeB magnet performance prediction method as claimed in claim 1, characterized in that: A real-time prediction system is used to build a visual interactive platform. Production parameters are input through an intuitive graphical interface, predicted magnet performance is output, and prediction results and optimization suggestions are viewed in real time.

5. A data-driven sintered NdFeB magnet performance prediction method as claimed in claim 1, characterized in that: The input of the two-stage prediction model is the process parameters and the powder performance data D99 / D10 and SMD obtained by the first-stage prediction, and the output is the magnet performance.

Images