Inductance material electromagnetic performance prediction method and system based on machine learning

Abstract

The invention discloses an inductance material electromagnetic performance prediction method and system based on machine learning, and the method comprises the steps: collecting inductance performance index data, carrying out the data preprocessing of an inductance material performance index optimization data set, and carrying out the visualization of the data set, and obtaining the linear relation between the variables of the quantized inductance material performance index optimization data set. Through inductance material magnetic conductivity prediction, saturated magnetic field intensity prediction and inductance material electromagnetic performance prediction, inductance material electromagnetic performance characterization reference data are finally obtained. According to the method, the potential nonlinear relationship among the inductance raw materials is mined by utilizing the powerful calculation learning capability of the machine learning model, so that the raw material selection and formula design of inductance products with different specifications are accurately guided.

Description

Technical Field

[0001] The present invention relates to the field of inductance performance prediction, and more specifically to a method and system for predicting the electromagnetic performance of inductance materials based on machine learning. Background Art

[0002] For inductance materials, permeability and saturation magnetic field intensity are key indicators characterizing their electromagnetic performance, which determine the upper limit and application scenarios of inductance materials. The permeability and saturation magnetic field intensity of inductance materials are determined by the types and proportions of raw materials used during material forming, that is, the raw material formula of inductance materials. The traditional inductance material formula design process is the same as that of many other material formula designs, and it is completed using the trial-and-error method. Specifically, it means continuously adjusting the formula according to experimental results and personnel experience to gradually approach the customer requirements and develop an inductance formula with the target permeability and saturation magnetic field intensity. However, this method not only consumes a large amount of trial-and-error time and development costs, but also has great uncertainty in the entire formula development process, which is completely determined by the experience of designers and the randomness of formula design. At the same time, there is a certain potential connection between the permeability prediction and saturation magnetic field intensity prediction of inductance materials. However, the traditional method predicts permeability and saturation magnetic field intensity as two independent targets respectively, resulting in insufficient comprehensiveness of the overall electromagnetic performance prediction. Moreover, the traditional inductance material formula design does not clearly explore the relationship between the inductance material formula and the permeability and saturation magnetic field intensity of inductance products, but only blindly adjusts and tries the formula, and the experimental results lack scientific basis. At the same time, due to the differences between experimental conditions and production reality, it is difficult to guarantee the reproducibility of experimental results in production reality. Summary of the Invention

[0003] The present invention aims to avoid the deficiencies of the above-mentioned prior art, and provides a method and system for predicting the electromagnetic performance of inductance materials based on machine learning, in order to utilize the machine learning model with powerful learning and computing capabilities to solve the blindness, randomness and uncertainty in the traditional inductance formula design process, deeply explore the relationship between the inductance material formula and its electromagnetic performance, reduce the research and development time and cost of inductance formula design, and accurately guide the selection of raw materials for inductance products.

[0004] The present invention adopts the following technical solutions to achieve the invention purpose:

[0005] The method for predicting the electromagnetic performance of inductance materials based on machine learning according to the present invention is characterized by including the following steps:

[0006] Step S1: Inductance performance index data collection refers to obtaining the original data A1 of the inductance material performance index through high-throughput experiments. Based on the original data A1 of the inductance material performance index and taking the inductance raw material ratio as the feature, a raw data set D1 of the inductance material performance index is established with the magnetic permeability and saturation magnetic field intensity as the target variables; the original data set D2 of the inductance raw material performance is obtained through independent single-component experiments;

[0007] Step S2: Data preprocessing is to perform basic data optimization processing and feature engineering on the original data set D1 of the inductance material performance index, and rearrange the data according to different formulation parameters to obtain an optimized data set D3 of the inductance material performance index; the basic data optimization processing and feature engineering include missing value processing, outlier processing, data feature transformation, and data standardization processing;

[0008] Step S3: Visualizing the data set is to perform exploratory data analysis on the optimized data set D3 of the inductance material performance index and create a correlation matrix to obtain the linear relationship R between the variables of the quantified optimized data set of the inductance material performance index;

[0009] The exploratory data analysis uses a scatter plot matrix to visually discover outliers, data distribution, and feature relationships;

[0010] The correlation matrix is a square matrix of Pearson product-moment correlation coefficients, which is used to measure the linear relationship between feature pairs;

[0011] Step S4: Inductance material magnetic permeability prediction is to use the optimized inductance electromagnetic performance data A2 in the optimized data set D3 of the inductance material performance index and the original data set D2 of the inductance raw material performance, and obtain the magnetic permeability prediction model Model according to the linear relationship R between the variables of the optimized data set of the inductance material performance index mp , and use the magnetic permeability prediction model Model mp to obtain the predicted magnetic permeability performance data B1;

[0012] The magnetic permeability prediction model Model mp is a random forest model combined with the CART algorithm and decision tree training;

[0013] The CART algorithm is a non-parametric classification and regression method used to construct a decision tree that maximizes information gain;

[0014] The decision tree training is to adjust the parameters of the decision tree through the feedback of the objective function;

[0015] The random forest model is an ensemble learning model, which is a classifier composed of multiple decision trees, and the final prediction result is determined by integrating the voting results of each decision tree;

[0016] Step S5: Prediction of saturation magnetic field intensity. Based on the optimized inductance electromagnetic performance data A2 in the optimized dataset D3 of inductance material performance indicators and the original dataset D2 of inductance raw material performance, combined with the predicted magnetic permeability performance data B1 and the linear relationship R between the variables in the optimized dataset of inductance material performance indicators, a saturation magnetic field intensity prediction model Model is obtained. sm , and using the saturation magnetic field intensity prediction model Model sm to obtain the saturation magnetic field intensity prediction data B2;

[0017] The saturation magnetic field intensity prediction model Model sm is a random forest model that combines an adaptive and self-iterative genetic algorithm and a gradient descent algorithm;

[0018] Step S6: Prediction of inductance material electromagnetic performance. Based on the predicted magnetic permeability performance data B1 and the saturation magnetic field intensity prediction data B2, integrating the characteristics and parameters of the magnetic permeability prediction model Model mp and the saturation magnetic field intensity prediction model Model sm , an inductance material electromagnetic performance prediction model is established, and the structure of the inductance material electromagnetic performance prediction model is adjusted and the algorithm is optimized to obtain an optimized prediction model Model of inductance material electromagnetic performance. ep , and using the optimized prediction model Model of inductance material electromagnetic performance ep to realize the synchronous prediction of magnetic permeability and saturation magnetic field intensity, and obtain the reference data B3 for characterizing the electromagnetic performance of the inductance material.

[0019] The characteristics of the inductance material electromagnetic performance prediction method based on machine learning in the present invention also lie in: predicting the magnetic permeability according to the following process:

[0020] Step S41: Dataset division;

[0021] The dataset division is to randomly divide the dataset into separate training datasets and test datasets based on the train_test_split tool of the sklearn.model_selection module. The training dataset is used to fit the model, and the test dataset is used to evaluate the model and the generalization performance when extended to new data that has never been seen, so as to obtain an unbiased estimate of the performance.

[0022] Step S42: Sample sampling;

[0023] The sample sampling is to sample from the original dataset with replacement through the Bootstrap sampling method to generate multiple different training subsets, and each subset is used to train a decision tree to ensure that the training data of each decision tree is different;

[0024] Step S43: Feature sampling;

[0025] The feature sampling is to randomly select some features during each split according to the random feature selection mechanism when creating each decision tree, which increases the difference between trees and the diversity of the model, and reduces the risk of overfitting.

[0026] Step S44: Decision tree construction;

[0027] The decision tree construction is to hierarchically select features, optimize the linear relationship between variables in the dataset according to the performance indicators of the inductive material, and iteratively split nodes until the stopping criterion for splitting is met. Then, based on the CART algorithm, the Gini coefficient and the information gain ratio are defined, and the root node is constructed with key features, and this process is repeated until the entire decision tree is constructed.

[0028] Step S45: Decision tree training;

[0029] The decision tree training is to adjust the distribution of the root node and the number of leaf nodes of the decision tree according to the dataset features and perform post-pruning on the decision tree, and reduce the risk of overfitting of the decision tree by restricting the depth of the tree, the number of leaf nodes, the number of samples in the leaf nodes, and the information increment.

[0030] Step S46: Result fusion;

[0031] The result fusion is to fuse the results of multiple decision trees, take the average of the predictions of all trees, smooth the noise of a single tree, and improve the stability and accuracy of the model.

[0032] Step S47: Permeability prediction model training;

[0033] The permeability prediction model training is specifically to perform permeability prediction model training and optimization according to the feedback through the dataset division, the decision tree construction, the decision tree result fusion, and the random forest ensemble learning, and obtain the optimal permeability prediction model Model mp ;

[0034] Step S48: Permeability prediction.

[0035] The permeability prediction is to use the permeability prediction model Model mp , and predict based on the optimized inductance electromagnetic performance data A2 and the original dataset D2 of the inductance raw material performance in the dataset D3 optimized according to the performance indicators of the inductive material, and obtain the predicted permeability performance data B1.

[0036] The characteristics of the electromagnetic performance prediction method of inductive materials based on machine learning in the present invention also lie in that: in step S5, the random forest model combining the adaptive and self-iterative genetic algorithm and the gradient descent algorithm includes binary coding, population initialization, elite selection, and multivariate function gradient setting;

[0037] The binary coding is used to map the independent variable from decimal to binary with the required precision;

[0038] The population initialization is used to generate an initial population for subsequent selection;

[0039] The elite selection is used to retain the optimal individual in each iteration process, so that it no longer undergoes mutation or crossover transformation, thus maximizing the inheritance of the optimal solution;

[0040] The multivariate function gradient setting is used to limit the gradient direction and learning rate.

[0041] The characteristics of the electromagnetic performance prediction method of inductive materials based on machine learning in the present invention also lie in that: in step S5, the saturation magnetic field strength prediction is carried out as follows:

[0042] Step S51: Perform feature processing and data set division, and complete the dimensionality reduction and denoising of the training set features;

[0043] Step S52: Construct a random forest regression model, and set the number of base learners and the optimal partitioning attribute criterion;

[0044] Step S53: Set the maximum depth of the decision tree, the minimum number of samples in the leaf node, and the maximum number of leaf nodes, optimize the model hyperparameters, and use the gradient descent algorithm to reduce the error;

[0045] Step S54: Train the saturation magnetic field strength prediction model, specifically, through the feature processing and data set division, the construction of the random forest regression model, and the optimization of the model hyperparameters, perform model training to obtain the saturation magnetic field strength model Model sm ;

[0046] Step S55: Saturation magnetic field strength prediction is to use the saturation magnetic field strength model Model sm , based on the optimized inductive electromagnetic performance data A2 and the original data set D2 of the inductive raw material performance in the optimized data set D3 of the inductive material performance indicators, combined with the predicted magnetic permeability performance data B1, perform saturation magnetic field strength prediction to obtain the saturation magnetic field strength data B2.

[0047] The electromagnetic performance prediction system of inductive materials based on machine learning of the present invention is used to implement the electromagnetic performance prediction method of inductive materials based on machine learning, and includes an electromagnetic performance index data collection module K1, a data preprocessing module K2, a magnetic permeability prediction module K3, a saturation magnetic field strength prediction module K4, a model parameter optimization module K5, and an inductive material electromagnetic performance prediction module K6.

[0048] The characteristics of the electromagnetic performance prediction system of inductive materials based on machine learning of the present invention also lie in:

[0049] The electromagnetic performance index data collection module K1 is used for collecting electromagnetic performance index data of inductive materials. Through high-throughput experimental data collection, the original data D1 of the performance indexes of inductive materials and the original data set D2 of the performance of inductive raw materials are obtained, and the original data set D1 of the performance indexes of inductive materials and the original data set D2 of the performance of inductive raw materials are sent to the data preprocessing module K2;

[0050] The data preprocessing module K2 is used for data preprocessing. Through data preprocessing and feature engineering, an optimized data set D3 of the performance indexes of inductive materials is obtained. By randomly dividing the data set, it is split into a training set D31 of the performance indexes of inductive materials and a test set D32 of the performance indexes of inductive materials, and the training set D31 of the performance indexes of inductive materials and the test set D32 of the performance indexes of inductive materials are jointly sent to the magnetic permeability prediction module K3 and the saturation magnetic field strength prediction module K4;

[0051] The magnetic permeability prediction module K3 is used for magnetic permeability prediction. Through magnetic permeability prediction, the predicted magnetic permeability data B1 is obtained, and the predicted magnetic permeability data B1 is sent to the saturation magnetic field strength prediction module K4 and the model parameter optimization module K5;

[0052] The saturation magnetic field strength prediction module K4 is used for saturation magnetic field strength prediction. Through saturation magnetic field strength prediction, the saturation magnetic field strength prediction data B2 is obtained, and the saturation magnetic field strength prediction data B2 is sent to the model parameter optimization module K5;

[0053] The model parameter optimization module K5 is used for hyperparameter adjustment and algorithm optimization of the model. Through receiving the feedback information of the model after the predicted magnetic permeability data and the saturation magnetic field strength prediction data, the hyperparameters of the model are adjusted and the algorithm is optimized to obtain the optimal model with the highest correlation coefficient, and the optimal model is sent to the inductive electromagnetic performance prediction module K6;

[0054] The inductive material electromagnetic performance prediction module K6 is used for inductive material electromagnetic performance prediction. Through the electromagnetic performance of inductive materials, the reference data B3 for characterizing the electromagnetic performance of inductive materials is obtained.

[0055] Compared with the existing technology, the beneficial effects of the present invention are reflected in:

[0056] 1. The present invention combines permeability prediction and saturation magnetic field strength prediction to achieve electromagnetic performance prediction of inductive materials. By adding optimized inductive electromagnetic performance data to the inductive electromagnetic performance prediction model and adjusting the model structure and algorithm logic, synchronous prediction of permeability and saturation magnetic field strength is realized, greatly improving the overall reliability and stability of electromagnetic performance prediction and reducing the dependence on data sources for electromagnetic performance prediction.

[0057] 2. The present invention uses machine learning methods to predict the electromagnetic performance of inductive materials. The optimized data set of inductive material performance indicators is fed into a machine learning model for continuous training and iteration to obtain the inductive material electromagnetic performance prediction model with the highest prediction accuracy, greatly reducing the cycle and cost of inductive material formula research and development and improving the efficiency, use value and production guiding ability of formula design.

[0058] 3. The present invention adopts high-throughput experiments for formula design experiments and data collection. Through high-throughput experiments, an original data set of inductive performance indicators and original data of inductive raw material performance that are all-round, three-dimensional and fully covered are obtained. It not only uses an efficient experimental platform to quickly accumulate data and improve experimental efficiency, but also reasonably utilizes all experimental sample data, avoiding repeated experiments and invalid experiments and extending the effective period of experimental data. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic flow chart of the method for predicting the electromagnetic performance of inductive materials of the present invention;

[0060] Figure 2 It is a schematic flow chart of the visual data set in the method of the present invention;

[0061] Figure 3 It is a schematic flow chart of the permeability prediction in the method of the present invention;

[0062] Figure 4 It is a schematic flow chart of the magnetic field strength prediction in the method of the present invention;

[0063] Figure 5 It is a schematic diagram of the system for predicting the electromagnetic performance of inductive materials of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] As Figure 1 shown, the method for predicting the electromagnetic performance of inductive materials based on machine learning in this embodiment includes the following steps:

[0065] Step S1: Inductance performance index data collection refers to obtaining the original data A1 of the inductance material performance index through high-throughput experiments. Based on the original data A1 of the inductance material performance index and taking the inductance raw material ratio as the feature, a raw data set D1 of the inductance material performance index is established with permeability and saturation magnetic field strength as the target variables; the original data set D2 of the inductance raw material performance is obtained through independent single-component experiments; high-throughput experiments refer to the experimental and testing work carried out through the integrated molded inductance raw material high-throughput experimental platform and the integrated molded inductance raw material high-throughput R & D test system.

[0066] Step S2: Data preprocessing is to perform basic data optimization processing and feature engineering on the original data set D1 of the inductance material performance index, and rearrange the data according to different formulation parameters to obtain the optimized data set D3 of the inductance material performance index; basic data optimization processing and feature engineering include missing value processing, outlier processing, data feature transformation, and data standardization processing.

[0067] Step S3: Visualize the data set, which is to perform exploratory data analysis on the optimized data set D3 of the inductance material performance index and create a correlation matrix to obtain the linear relationship R between the variables of the optimized data set of the inductance material performance index, which is used to present the data more intuitively and three-dimensionally.

[0068] Figure 2 The following visualization data set process is shown:

[0069] Step S31: Exploratory data analysis is to visually discover outliers, data distribution, and feature relationships in the optimized data set D3 of the inductance material performance index using a scatter plot matrix;

[0070] Step S32: Correlation matrix analysis. The correlation matrix is a square matrix of Pearson product-moment correlation coefficients, which is used to measure the linear relationship between feature pairs. Correlation matrix analysis is to perform a square matrix processing including Pearson product-moment correlation coefficients on the optimized data set D3 of the inductance material performance index to measure the linear dependence relationship between feature pairs. The calculation formula is:

[0071]

[0072] In the formula:

[0073] Let r represent the correlation, n represent the total number of samples, and i represent the i-th sample, i = 1, 2... n;

[0074] Let x (i) represent the true target value of sample x, and y (i) represent the true target value of sample y;

[0075] Let μ represent the mean of the samples, and μ x be the mean of sample x, and μy is the mean value of sample y;

[0076] Let σ x represent the standard deviation of sample x, and let σ y represent the standard deviation of sample y, and σ xy be the covariance between sample x and sample y;

[0077] Step S33: Overall image output. The MLxtend module is used to draw an image of the above analysis results, output the scatter plot matrix and heat map matrix related to the optimized data set of the inductive material performance indicators, and obtain the linear relationship R between the variables of the optimized data set of the inductive material performance indicators.

[0078] Step S4: Prediction of the magnetic permeability of the inductive material. Using the optimized inductive electromagnetic performance data A2 in the optimized data set D3 of the inductive material performance indicators and the original data set D2 of the inductive raw material performance, according to the linear relationship R between the variables of the optimized data set of the inductive material performance indicators, a magnetic permeability prediction model Model mp is obtained. Using the magnetic permeability prediction model Model mp to obtain the predicted magnetic permeability performance data B1; among them, the magnetic permeability prediction model Model mp is a random forest model combined with the CART algorithm and decision tree training; the CART algorithm is a non-parametric classification and regression method used to construct a decision tree that maximizes the information gain; decision tree training is to adjust the parameters of the decision tree through the feedback of the objective function; the random forest model is an ensemble learning model, a classifier composed of multiple decision trees, and the final prediction result is determined by synthesizing the voting results of each decision tree.

[0079] Step S5: Prediction of the saturation magnetic field strength. Based on the optimized inductive electromagnetic performance data A2 in the optimized data set D3 of the inductive material performance indicators and the original data set D2 of the inductive raw material performance, combined with the predicted magnetic permeability performance data B1 and the linear relationship R between the variables of the optimized data set of the inductive material performance indicators, a saturation magnetic field strength prediction model Model sm is obtained. Using the saturation magnetic field strength prediction model Model sm to obtain the saturation magnetic field strength prediction data B2; among them, the saturation magnetic field strength prediction model Model sm is a random forest model combined with the adaptive and self-iterative genetic algorithm and the gradient descent algorithm.

[0080] Step S6: Prediction of the inductive material electromagnetic performance. Based on the predicted magnetic permeability performance data B1 and the saturation magnetic field strength prediction data B2, integrating the magnetic permeability prediction model Model mp and the saturation magnetic field strength prediction model Model smBased on the characteristics and parameters, an electromagnetic performance prediction model of inductive materials is established, and the structure of the electromagnetic performance prediction model of inductive materials is adjusted and the algorithm is optimized to obtain an optimized prediction model Model of the electromagnetic performance of inductive materials ep Using the optimized prediction model Model of the electromagnetic performance of inductive materials ep Realize the synchronous prediction of permeability and saturation magnetic field strength, and obtain the reference data B3 for characterizing the electromagnetic performance of inductive materials.

[0081] In this embodiment, the electromagnetic performance of inductive materials is predicted by combining permeability prediction and saturation magnetic field strength prediction. By adding the optimized inductive electromagnetic performance data into the inductive electromagnetic performance prediction model, the model structure and algorithm logic are adjusted to realize the synchronous prediction of permeability and saturation magnetic field strength, improve the overall reliability and stability of electromagnetic performance prediction, and reduce the dependence on the data source of electromagnetic performance prediction.

[0082] Figure 3 Schematically shows the process of permeability prediction in this embodiment as follows

[0083] Step S41: Dataset division

[0084] Dataset division is to randomly divide the dataset into separate training datasets and test datasets based on the train_test_split tool of the sklearn.model_selection module. The training dataset is used to fit the model, and the test dataset is used to evaluate the model's generalization performance when applied to new data that has never been seen before, obtaining an unbiased estimate of the performance.

[0085] Step S42: Sample sampling

[0086] Sample sampling is to sample with replacement from the original dataset through the Bootstrap sampling method to generate multiple different training subsets, and each subset is used to train a decision tree to ensure that the training data for each decision tree is different.

[0087] Step S43: Feature sampling

[0088] Feature sampling is to randomly select some features at each split based on the random feature selection mechanism when creating each decision tree, increasing the difference of the tree and the diversity of the model, and reducing the risk of overfitting.

[0089] Step S44: Decision tree construction

[0090] Decision tree construction is to hierarchically select features, based on the linear relationship between the variables of the optimized dataset of inductive material performance indicators, and iteratively split nodes until the stopping criterion for splitting is met.

[0091] Define entropy as an indicator of impurity to determine which feature split can maximize the Information Gain (IG). The binary split is defined as:

[0092]

[0093] Where:

[0094] Let IG represent the information gain, N left and N right represent the number of samples in the left and right child nodes respectively. Let x i represent the sample feature to be split. N p is the number of samples in the parent node, I is the impurity function, D p is the subset of the training samples in the parent node, D left and D right are the training sample sets of the left and right child nodes after splitting. The goal is to find the feature split that can maximize the information gain and find the split feature that can reduce the impurity in the child nodes, so as to prevent the over-splitting of sample features and make the fitting process uncontrollable. However, whether it is the Gini impurity or the entropy measuring impurity, an impurity indicator suitable for continuous variables is required. Define the impurity indicator of node t as MSE(t):

[0095]

[0096] Where:

[0097] I(t) is the impurity function of node t;

[0098] N t is the number of training samples in node t, D t is the subset of the training data in node t;

[0099] y (i) is the true target value, is the predicted target value (sample mean):

[0100]

[0101] Then, based on the CART algorithm, define the Gini coefficient and the information gain ratio. Construct the root node with key features and repeat continuously until the entire decision tree is constructed;

[0102] Step S45: Decision tree training;

[0103] Decision tree training is to adjust the distribution of the root node and the number of leaf nodes of the decision tree according to the dataset features and perform post-pruning on the decision tree, so as to reduce the risk of overfitting of the decision tree by restricting the depth of the tree, the number of leaf nodes, the number of samples in the leaf nodes, and the information increment.

[0104] The measurement criterion for post-pruning processing is as follows:

[0105] L α = Gini(T) × |T| + α|T leaf |

[0106] In the formula: L α (T) represents the final loss, Gini(T) represents the entropy or Gini coefficient of the current node, |T| represents the number of data samples contained in the current node, and T leaf represents the number of leaf nodes generated after the current node is divided, and α is the preference coefficient.

[0107] Step S46: Result fusion;

[0108] Result fusion is achieved by fusing the results of multiple decision trees, taking the average of the predictions of all trees, smoothing the noise of a single tree, and improving the stability and accuracy of the model;

[0109] Step S47: Training of the permeability prediction model;

[0110] The training of the permeability prediction model is specifically to train and optimize the permeability prediction model according to the feedback through data set division, decision tree construction, decision tree result fusion, and random forest ensemble learning, and obtain the optimal permeability prediction model Model mp ;

[0111] Step S48: Permeability prediction.

[0112] Permeability prediction is to use the permeability prediction model Model mp , and predict based on the optimized inductance electromagnetic performance data A2 and the original data set D2 of the inductance raw material performance in the data set D3 optimized according to the performance indicators of the inductance material, and obtain the predicted magnetic permeability performance data B1.

[0113] The random forest model of this embodiment combining the adaptive and self-iterative genetic algorithm and the gradient descent algorithm includes binary coding, population initialization, elite selection, and multi-variable function gradient setting; among them, binary coding is used to map the independent variable from decimal to binary with the required precision; population initialization is used to generate an initial population for subsequent selection; elite selection is used to retain the optimal individual in each iteration process and prevent it from mutating or crossing, so as to maintain the inheritance of the optimal solution to the greatest extent; multi-variable function gradient setting is used to limit the gradient direction and learning rate.

[0114] Figure 4 It shows that the saturation magnetic field strength prediction in this embodiment is carried out according to the following process:

[0115] Step S51: Perform feature processing and dataset partitioning, and complete the dimensionality reduction and denoising of the training set features;

[0116] Step S52: Construct a random forest regression model and set the number of base learners and the optimal partitioning attribute criterion;

[0117] Step S53: Set the maximum depth of the decision tree, the minimum number of samples in the leaf nodes, and the maximum number of leaf nodes, optimize the model hyperparameters, and use the gradient descent algorithm to reduce the error. The gradient descent calculation formula is:

[0118]

[0119] In the formula: Θ 0 is the original position, Θ 1 is the next position, α is the step size, is the descending gradient, and J is a function of Θ;

[0120] Step S54: Train the saturation magnetic field intensity prediction model. Specifically, through feature processing and dataset partitioning, constructing a random forest regression model, and optimizing the model hyperparameters, select the mean squared error as the cost function to train the model, and obtain the saturation magnetic field intensity model Model sm , and the calculation formula of the mean squared error cost function is:

[0121]

[0122] In the formula:

[0123] Let m represent the number of samples, β (i) is the true value of the sample, γ (i) is the predicted value of the sample, and h 0 is the prediction function;

[0124] Step S55: Saturation magnetic field intensity prediction. Use the saturation magnetic field intensity model Model sm , optimize the optimized inductance electromagnetic performance data A2 in the dataset D3 and the original dataset D2 of the inductance raw material performance according to the inductance material performance index, combine the predicted magnetic permeability performance data B1, and perform saturation magnetic field intensity prediction to obtain the saturation magnetic field intensity data B2.

[0125] Figure 5 Figure shows the inductance material electromagnetic performance prediction system based on machine learning in this embodiment, including an electromagnetic performance index data collection module K1, a data preprocessing module K2, a magnetic permeability prediction module K3, a saturation magnetic field intensity prediction module K4, a model parameter optimization module K5, and an inductance material electromagnetic performance prediction module K6.

[0126] The electromagnetic performance index data collection module K1 is used for collecting the electromagnetic performance index data of inductive materials. Through high-throughput experimental data collection, the original data D1 of the performance index of inductive materials and the original data set D2 of the performance of inductive raw materials are obtained, and the original data set D1 of the performance index of inductive materials and the original data set D2 of the performance of inductive raw materials are sent to the data preprocessing module K2;

[0127] The data preprocessing module K2 is used for data preprocessing. Through data preprocessing and feature engineering, the optimized data set D3 of the performance index of inductive materials is obtained. Through random division of the data set, it is split into the training set D31 of the performance index of inductive materials and the test set D32 of the performance index of inductive materials, and the training set D31 of the performance index of inductive materials and the test set D32 of the performance index of inductive materials are jointly sent to the magnetic permeability prediction module K3 and the saturation magnetic field strength prediction module K4;

[0128] The magnetic permeability prediction module K3 is used for magnetic permeability prediction. Through magnetic permeability prediction, the predicted magnetic permeability data B1 is obtained, and the predicted magnetic permeability data B1 is sent to the saturation magnetic field strength prediction module K4 and the model parameter optimization module K5;

[0129] The saturation magnetic field strength prediction module K4 is used for saturation magnetic field strength prediction. Through saturation magnetic field strength prediction, the predicted saturation magnetic field strength data B2 is obtained, and the predicted saturation magnetic field strength data B2 is sent to the model parameter optimization module K5;

[0130] The model parameter optimization module K5 is used for hyperparameter adjustment and algorithm optimization of the model. Through the feedback information of the model after receiving the predicted magnetic permeability data and the predicted saturation magnetic field strength data, the hyperparameters of the model are adjusted and the algorithm is optimized to obtain the optimal model with the highest correlation coefficient and send the optimal model to the inductive electromagnetic performance prediction module K6;

[0131] The inductive material electromagnetic performance prediction module K6 is used for inductive material electromagnetic performance prediction. Through the inductive material electromagnetic performance, the reference data B3 for characterizing the inductive material electromagnetic performance is obtained.

[0132] The present invention uses a random forest model with genetic algorithm and gradient descent algorithm for saturation magnetic field strength prediction, optimizes the data set through the genetic algorithm experience selection rule, greatly reduces the error through the gradient descent algorithm, and finally improves the overall performance of the model through hyperparameter optimization, improving the reliability of the inductive material electromagnetic performance prediction process and the overall generalization ability, providing a good technical exploration and attempt for inductive material electromagnetic performance prediction.

Claims

1. A method for predicting electromagnetic properties of inductor materials based on machine learning, characterized by The following steps are involved: Step S1: collecting inductor performance index data, which refers to obtaining original data (A1) of inductor material performance indexes of different formulas through high-throughput experiments, establishing an original data set (D1) of inductor material performance indexes based on the original data (A1) of inductor material performance indexes and taking the ratio of inductor raw materials as features, and taking magnetic permeability and magnetic field strength as target variables; obtaining an original data set (D2) of inductor material performance through independent single-component experiments; Step S2: data preprocessing, which is to perform basic data optimization processing and feature engineering on the original data set (D1) of the inductor material performance index, and rearrange the data according to different recipe parameters to obtain an optimized data set (D3) of the inductor material performance index; the basic data optimization processing and feature engineering include missing value processing, outlier processing, data feature transformation and data standardization processing; Step S3: visualizing the data set, which is to perform exploratory data analysis on the inductor material performance index optimization data set (D3) and create a correlation matrix to obtain a quantified linear relationship (R) between the variables of the inductor material performance index optimization data set; The exploratory data analysis uses a scatter plot matrix to intuitively discover outliers, data distribution, and feature relationships; The correlation matrix is ​​a square matrix of Pearson product-moment correlation coefficients, which is used to measure the linear relationship between feature pairs; Step S4: Predicting the magnetic permeability of the inductor material is to use the optimized inductor electromagnetic performance data (A2) in the inductor material performance index optimization data set (D3) and the original inductor material performance data set (D2), and obtain the magnetic permeability prediction model Model according to the linear relationship (R) between the variables in the inductor material performance index optimization data set. mp , using the magnetic permeability prediction model Model mp Obtaining predicted magnetic permeability performance data (B1); The magnetic permeability prediction model mp It is a random forest model that combines the CART algorithm and decision tree training; The CART algorithm is a non-parametric classification and regression method used to construct a decision tree that maximizes information gain; The decision tree training is to adjust the parameters of the decision tree through objective function feedback; The random forest model is an integrated learning model, which is a classifier composed of multiple decision trees. The final prediction result is determined by integrating the voting results of each decision tree. Step S5: Saturation magnetic field intensity prediction is based on the optimized inductor electromagnetic performance data (A2) and the original inductor material performance data set (D2) in the inductor material performance index optimization data set (D3), combined with the predicted magnetic permeability performance data (B1) and the linear relationship (R) between the variables in the inductor material performance index optimization data set, to obtain a saturation magnetic field intensity prediction model Model sm , using the saturation magnetic field strength prediction model Model sm Obtain saturation magnetic field intensity prediction data B2; The saturation magnetic field strength prediction model Model sm It is a random forest model that combines an adaptive and self-iterative genetic algorithm with a gradient descent algorithm; Step S6: Predicting the electromagnetic properties of the inductor material, based on the predicted magnetic permeability performance data (B1) and the saturation magnetic field strength prediction data (B2), a fusion magnetic permeability prediction model Model mp And saturation magnetic field strength prediction model Model sm The electromagnetic performance prediction model of the inductor material is established based on the characteristics and parameters of the inductor material, and the electromagnetic performance prediction model of the inductor material is structurally adjusted and the algorithm is optimized to obtain the electromagnetic performance optimization prediction model of the inductor material. ep , using the electromagnetic performance optimization prediction model of the inductor material Model ep The simultaneous prediction of magnetic permeability and saturation magnetic field strength is achieved to obtain reference data for the electromagnetic performance characterization of inductor materials (B3).

2. The method for predicting electromagnetic properties of inductor materials based on machine learning according to claim 1, characterized in that: The magnetic permeability prediction is performed as follows: Step S41: data set division; The data set division is based on the train_test_split tool of the sklearn.model_selection module to randomly divide the data set into separate training data sets and test data sets. The training data set is used to fit the model, and the test data set is used to evaluate the generalization performance of the model when it is extended to new data that has never been seen, so as to obtain an unbiased estimate of the performance; Step S42: sample sampling; The sample sampling is performed by using the Bootstrap sampling method, sampling with replacement from the original data set to generate multiple different training subsets, each subset is used to train a decision tree, and ensures that the training data of each decision tree is different; Step S43: feature sampling; The feature sampling is to randomly select some features at each split according to the random feature selection mechanism when creating each decision tree, which increases the difference of the tree and the diversity of the model and reduces the risk of overfitting; Step S44: decision tree construction; The decision tree construction is to select features hierarchically, optimize the linear relationship between data set variables according to the performance index of the inductor material, iterate the node splitting until the criterion for stopping splitting is met, and then define the Gini coefficient and information gain rate based on the CART algorithm, construct the root node with the key features, and repeat until the entire decision tree is constructed; Step S45: decision tree training; The decision tree training is to adjust the distribution of the decision tree root nodes and the number of leaf nodes according to the characteristics of the data set and perform post-pruning on the decision tree, thereby reducing the risk of overfitting of the decision tree by limiting the depth of the tree, the number of leaf nodes, the number of samples contained in the leaf nodes, and the information increment; Step S46: result fusion; The result fusion is achieved by fusing the results of multiple decision trees and taking the predicted average of all trees, thereby smoothing the noise of a single tree and improving the stability and accuracy of the model. Step S47: training the magnetic permeability prediction model; The magnetic permeability prediction model training is specifically to perform magnetic permeability prediction model training and optimization according to feedback through the data set division, the decision tree construction, the decision tree result fusion and random forest ensemble learning, and obtain the optimal magnetic permeability prediction model Model mp ; Step S48: Magnetic permeability prediction. The magnetic permeability prediction is performed using the magnetic permeability prediction model Model mp , prediction is performed based on the optimized inductor electromagnetic performance data (A2) in the inductor material performance index optimization data set (D3) and the original inductor raw material performance data set (D2) to obtain predicted magnetic permeability performance data (B1).

3. The method for predicting electromagnetic properties of inductor materials based on machine learning according to claim 2, characterized in that: In step S5, the random forest model combining the adaptive and self-iterative genetic algorithm and the gradient descent algorithm includes binary coding, population initialization, elite selection and multivariable function gradient setting; The binary encoding is used to map the independent variable from decimal to binary with required precision; The population initialization is used to generate an initialization population for subsequent selection; The elite selection is used to retain the best individual in each iteration process, so that it will not mutate or crossover, thus maintaining the inheritance of the optimal solution to the greatest extent; The multivariable function gradient setting is used to limit the gradient direction and learning rate.

4. The method for predicting electromagnetic properties of inductor materials based on machine learning according to claim 3, characterized in that: In step S5, the saturation magnetic field strength prediction is performed according to the following process: Step S51: perform feature processing and data set division to complete feature dimension reduction and denoising of the training set; Step S52: construct a random forest regression model, and set the number of base learners and the optimal partitioning attribute criterion; Step S53: setting the maximum depth of the decision tree, the minimum number of leaf node samples and the maximum number of leaf nodes, optimizing the model hyperparameters, and using the gradient descent algorithm to reduce the error; Step S54: training the saturation magnetic field strength prediction model, specifically, training the model through the feature processing and data set division, building the random forest regression model, and optimizing the model hyperparameters to obtain the saturation magnetic field strength model Model sm ; Step S55: Saturation magnetic field strength prediction is performed using the saturation magnetic field strength model Model sm , based on the optimized inductor electromagnetic performance data (A2) in the inductor material performance index optimization data set (D3) and the original inductor raw material performance data set (D2), combined with the predicted magnetic permeability performance data (B1), a saturation magnetic field intensity prediction is performed to obtain saturation magnetic field intensity data (B2).

5. A system for predicting electromagnetic properties of inductor materials based on machine learning, used to implement the method for predicting electromagnetic properties of inductor materials based on machine learning as claimed in any one of claims 1 to 4, characterized in that: It includes an electromagnetic performance index data collection module (K1), a data preprocessing module (K2), a magnetic permeability prediction module (K3), a saturation magnetic field strength prediction module (K4), a model parameter optimization module (K5) and an inductor material electromagnetic performance prediction module (K6).

6. The electromagnetic properties prediction system of inductor materials based on machine learning according to claim 5, characterized in that: The electromagnetic performance index data collection module (K1) is used for collecting electromagnetic performance index data of inductor materials, and obtains original data of inductor material performance index (D1) and original data set of inductor raw material performance (D2) through high-throughput experimental data collection, and sends the original data set of inductor material performance index (D1) and original data set of inductor raw material performance (D2) to the data preprocessing module (K2); The data preprocessing module (K2) is used for data preprocessing, and obtains an inductor material performance index optimization data set (D3) through data preprocessing and feature engineering, and splits the data set into an inductor material performance index training set (D31) and an inductor material performance index training set test set (D32) by random partitioning of the data set, and sends the inductor material performance index training set (D31) and the inductor material performance index test set (D32) to the magnetic permeability prediction module (K3) and the saturation magnetic field strength prediction module (K4); The magnetic permeability prediction module (K3) is used for magnetic permeability prediction, obtains predicted magnetic permeability data (B1) through magnetic permeability prediction, and sends the predicted magnetic permeability data (B1) to the saturation magnetic field intensity prediction module (K4) and the model parameter optimization module (K5); The saturation magnetic field strength prediction module (K4) is used for saturation magnetic field strength prediction, obtains saturation magnetic field strength prediction data (B2) through saturation magnetic field strength prediction, and sends the saturation magnetic field strength prediction data (B2) to the model parameter optimization module (K5); The model parameter optimization module (K5) is used to adjust the hyperparameters and optimize the algorithm of the model, and adjust the hyperparameters and optimize the algorithm of the model by receiving the feedback information of the model after the predicted magnetic permeability data and the saturation magnetic field intensity predicted data, so as to obtain the optimal model with the highest correlation coefficient and send the optimal model to the inductance electromagnetic performance prediction module (K6); The inductor material electromagnetic performance prediction module (K6) is used to predict the electromagnetic performance of the inductor material, and obtains the electromagnetic performance characterization reference data (B3) of the inductor material through the electromagnetic performance of the inductor material.

Images