Wave-absorbing agent material component design method based on feature enhancement and multi-target regression

Abstract

The invention discloses a wave-absorbing agent material component design method based on feature enhancement and multi-target regression. The method comprises the following steps: acquiring wave-absorbing agent material data; constructing a regression chain initial sequence based on the wave-absorbing agent material data; predicting related components of the wave-absorbing agent material data, and obtaining a prediction result and a prediction error; correcting the regression chain initial sequence by using the prediction error to obtain a regression chain sequence; and based on the prediction result and in combination with the regression chain sequence, obtaining the wave-absorbing agent material components. According to the method, the prediction error coefficient is introduced, the correlation coefficient and the prediction error are integrated, link dependence of correlation and prediction precision from high to low is realized, and the component design of the wave absorbing agent is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of material composition design, and in particular relates to a method for designing absorber material composition based on feature enhancement and multi-objective regression. Background Art

[0002] In recent years, absorbing materials have been widely used in military and civilian fields. In the military field, absorbing materials can absorb electromagnetic waves emitted by radar detection devices, thereby reducing the radar scattering surface to reduce the probability of being detected. In the civilian field, with the development of communication technology, more and more microwave frequencies are being used. Absorbing materials can absorb most of the electromagnetic waves into the material, and finally convert them into heat energy or other forms after attenuation to eliminate the negative effects of electromagnetic radiation. Traditional absorbing materials are prepared by the process of base materials and absorbers. Limited by the characteristics of the absorber itself, a single-layer absorbing material can only show good absorbing performance in certain specific frequency bands when the thickness is determined. Under the conditions of measuring data of different frequencies and material ratios, how to accelerate the two-way prediction of material properties and component frequencies with a data-driven method has become a research hotspot combining materials and artificial intelligence. According to the performance of absorber materials, directly designing the composition of materials has the following two problems:

[0003] 1. Limitation of features: There are relatively few features used in the prediction of absorbing materials. Taking the composition-performance prediction as an example, under certain conditions, there are only operating frequency, thickness, ratio, aspect ratio, tube diameter and operating frequency. It is urgent to use data analysis or model prediction to expand the potential features for more effective model prediction.

[0004] 2. Sequential characterization of variables: The prediction method of regression chain improves the prediction performance of multiple targets. In the process of building the regression chain, the traditional method is based on the order of the dependent variables defined by the user. Because there is a potential relationship between material components, such as the dielectric constant tangent is affected by the material thickness and material ratio.

[0005] In order to solve the above two problems and improve the intelligent design performance of absorber materials, the present invention proposes a method for designing absorber material composition based on feature enhancement and multi-objective regression. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a method for designing absorber material composition based on feature enhancement and multi-objective regression, which can improve the intelligent design performance of absorber materials.

[0007] The present invention proposes a method for designing absorber material composition based on feature enhancement and multi-objective regression, comprising:

[0008] Obtain absorber material data;

[0009] Based on the absorber material data, constructing a regression chain initial sequence;

[0010] Predicting relevant components of the absorber material data, and obtaining prediction results and prediction errors;

[0011] Correcting the initial order of the regression chain using the prediction error to obtain a regression chain order;

[0012] Based on the prediction result and the regression chain sequence, the material composition of the absorber is obtained.

[0013] Optionally, obtaining the absorber material data includes: obtaining the absorber performance data under different absorber material composition conditions;

[0014] Among them, the absorber material composition includes: thickness, ratio, operating frequency, aspect ratio and tube diameter; the absorber performance data includes: real part of dielectric constant, imaginary part of dielectric constant, tangent value of dielectric constant, real part of magnetic permeability, imaginary part of magnetic permeability and tangent value of magnetic permeability.

[0015] Optionally, based on the absorber material data, constructing the initial sequence of the regression chain includes:

[0016] Calculating the correlation of the absorber material data using the Pearson correlation coefficient analysis method to obtain a correlation coefficient value;

[0017] The correlation coefficient values ​​are sorted in descending order to obtain the initial order of the regression chain.

[0018] Optionally, predicting the relevant components of the absorber material data and obtaining the prediction result and the prediction error includes:

[0019] Inputting the absorber material data into a base learner to obtain the prediction result, wherein the base learner is used to output a single component corresponding to the absorber material;

[0020] According to the prediction result, a prediction error is obtained.

[0021] Optionally, using the prediction error to correct the initial order of the regression chain, and obtaining the regression chain order includes:

[0022] Sum the correlation coefficient values ​​to obtain a correlation sum;

[0023] Setting weights to fuse the sum of the correlations and the prediction error to obtain a fusion score;

[0024] The initial order of the regression chain is modified using the fusion score to obtain the regression chain order.

[0025] Optionally, the method to obtain the fusion score is:

[0026] S l =α*S yd -β*e d

[0027] Among them, S l is the fusion score, α and β are weights, S yd is the sum of correlations, e d is the prediction error.

[0028] Optionally, obtaining the absorber material composition based on the prediction result and the regression chain sequence includes:

[0029] Using the prediction results as cumulative features of the regression chain sequence;

[0030] The cumulative features and the target variable of the regression chain sequence are input into a regression link model to obtain the absorber material composition, wherein the regression link model is obtained through a training set, and the training set includes a regression chain sequence.

[0031] Compared with the prior art, the present invention has the following advantages and technical effects:

[0032] In terms of feature enhancement of absorber data, the present invention introduces the prediction value of a single feature under a multi-objective framework to reduce the disturbance of the accumulated prediction error on the result in the multi-objective process. In terms of the variable prediction chain of absorber components, a regression chain method based on the Pearson coefficient is constructed to strengthen the prediction order according to the degree of prediction target dependence. The prediction error coefficient is introduced, and the correlation coefficient and prediction error are integrated to achieve the link dependence of correlation and prediction accuracy from high to low, so as to realize the design of absorber components. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0034] Figure 1 It is a flow chart of a method for designing absorber material composition based on feature enhancement and multi-objective regression according to an embodiment of the present invention;

[0035] Figure 2 is a schematic diagram of the data collection results of the wave absorbing agent according to an embodiment of the present invention;

[0036] Figure 3 Schematic diagram of a regression link model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0038] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] The present invention proposes a method for designing absorber material composition based on feature enhancement and multi-objective regression. Figure 1 As shown, the specific steps include:

[0040] Obtain absorber material data;

[0041] Based on the absorber material data, the initial sequence of the regression chain is constructed;

[0042] Predict the relevant components of the absorber material data and obtain the prediction results and prediction errors;

[0043] The initial order of the regression chain is corrected using the prediction error to obtain the order of the regression chain;

[0044] Based on the prediction results and the regression chain sequence, the absorber material composition is obtained.

[0045] Specifically, the present invention relates to a method for designing absorber material components based on feature enhancement and multi-objective regression, which designs absorber material components in a data-driven manner. The first step is to collect required data through experiments. Typical absorber properties include the real part and imaginary part of dielectric constant; the real part and imaginary part of magnetic permeability, etc. Typical absorber material components include working frequency, thickness, ratio, aspect ratio, pipe diameter, etc. The second step is to use the collected data and use the Pearson correlation coefficient to construct the basic sequence of the regression chain, that is, the initial sequence of the regression chain. The third step is to train the base learner for each component as a predictor variable. The fourth step is to introduce the error coefficient of the third step model to adjust the regression chain construction order based on the initial construction of the regression chain in the second step, and introduce the error coefficient adjustment mechanism in the regression chain construction to reduce the interference of variables with large errors on the model. The fifth step is to introduce the single variable prediction result of the third step model as a cumulative feature, and use the prediction result of the previous target as a cumulative feature to enhance the feature expression ability of the model. The sixth step is to use the constructed model to input the required absorber material properties, intelligently generate the relevant components of the absorber material, and accelerate the material design efficiency.

[0046] This method uses the experimentally measured typical absorber properties, such as the real and imaginary parts of the dielectric constant; the real and imaginary parts of the magnetic permeability, etc., and the absorber material components, such as operating frequency, thickness, ratio, aspect ratio, and tube diameter, to design a multi-objective regression model. Using the collected data, the Pearson correlation coefficient is used to construct the basic order of the regression chain: by analyzing the correlation between the targets, the construction order of each target in the regression chain is optimized. The error coefficient is introduced to adjust the order of regression chain construction, and the error coefficient adjustment mechanism is introduced in the regression chain construction to reduce the interference of variables with large errors on the model. In terms of feature enhancement, the prediction results of a single variable are introduced as cumulative features, and the prediction results of a single target are used as cumulative features to enhance the feature expression ability of the model. Finally, using the constructed model, the required absorber material properties are input, and the relevant components of the absorber material are intelligently generated to accelerate the efficiency of material design.

[0047] Further, obtaining the absorber material data includes: obtaining the absorber performance data under different absorber material composition conditions;

[0048] Among them, the absorber material composition includes: thickness, ratio, operating frequency, aspect ratio and tube diameter; the absorber performance data includes: real part of dielectric constant, imaginary part of dielectric constant, tangent value of dielectric constant, real part of magnetic permeability, imaginary part of magnetic permeability and tangent value of magnetic permeability.

[0049] Specifically, the performance data of the absorber (real part, imaginary part, tangent value of dielectric constant; real part, imaginary part, tangent value of magnetic permeability, etc.) under different composition conditions (thickness, ratio, aspect ratio, tube diameter, etc.) are collected. It can be stored in electronic documents or using database software such as Mysql.

[0050] Furthermore, based on the absorber material data, the basic sequence of constructing the regression chain includes:

[0051] The Pearson correlation coefficient analysis method is used to calculate the correlation of the absorber material data and obtain the correlation coefficient value;

[0052] Sort the correlation coefficient values ​​in descending order to obtain the regression chain basis order.

[0053] Specifically, the present invention applies a correlation analysis method, including but not limited to the Pearson correlation coefficient, to analyze the correlation relationship of the variables based on the data collected in the first step. Taking the Pearson correlation coefficient as an example:

[0054]

[0055] Among them, x i and i are the data points of the two variables, and is the average of the two variables. r is the Pearson correlation coefficient, where +1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear correlation. In the present invention, x and y represent the composition of the absorber material, such as thickness, ratio, aspect ratio, tube diameter, etc.

[0056] The present invention assumes that the d-1th variable has a higher contribution to the prediction of the dth variable, and uses the Pearson correlation coefficient of the above formula to calculate the variable y d The sum of the absolute values ​​of the correlation coefficients with other variables is used as the variable sorting criterion:

[0057]

[0058] The present invention uses the coefficients of the above formula to describe the linear relationship between the two variables of the absorber components, and then sorts them in descending order to form the largest correlation chain. Placing the variable with the highest sum of correlation coefficients at the front of the regression chain helps to predict other variables based on this variable, highlighting the advantage of its correlation coefficient.

[0059] Further, predicting the relevant components of the absorber material data, obtaining the prediction results and prediction errors includes:

[0060] Inputting the absorber material data into the base learner to obtain a prediction result, wherein the base learner is used to output a single component corresponding to the absorber material;

[0061] According to the prediction results, the prediction error is obtained.

[0062] Specifically, the base learner, including but not limited to using a gradient boosting train (GBDT) etc. for learning, takes the absorber material properties collected in the first step as input and the components as output.

[0063] If we assume that there are D components, then there are D base learners and the corresponding prediction errors e d .

[0064] Furthermore, the prediction error is used to correct the basic order of the regression chain, and the order of the regression chain is obtained, including:

[0065] Sum the correlation coefficient values ​​to obtain the total correlation;

[0066] Set weights to fuse the sum of correlations and prediction errors to obtain a fusion score;

[0067] The fusion score is used to modify the basic order of the regression chain and obtain the regression chain order.

[0068] The present invention uses the third step to perform basic prediction of components and finds that there are differences in the prediction error range. In order to reduce the interference of variables with large prediction errors on the regression chain and reduce the prediction performance of global variables, the present invention introduces the performance metric (error or determination coefficient) of the third step base learner on the basis of the second step to calculate the fusion score S l , to adjust the link order construction.

[0069] Set the weight to sum the correlation S yd The error e with the base learner d The fusion is performed as shown in the following formula. α and β are custom weights. The weight fusion score is used to update the regression chain order to reduce the interference of variables with large errors on the regression chain order.

[0070] S l =α*S yd -β*e d

[0071] Among them, S l is the fusion score, α, β are weights, S yd is the sum of correlations, e d is the prediction error.

[0072] Furthermore, based on the prediction results and the regression chain sequence, the absorber material composition is obtained including:

[0073] The prediction results are used as cumulative features of the regression chain sequence;

[0074] The target variables of the cumulative features and the regression chain sequence are input into the regression link model to obtain the absorber material composition, wherein the regression link model is obtained through a training set, and the training set includes the regression chain sequence.

[0075] The present invention adds the prediction result of the single component in the third step to the regression link. is the prediction of the d-1th dimension regression chain, This variable depends only on the predicted result of material properties, so the d-th dimension prediction variable is:

[0076]

[0077] Under the condition of limited characteristics of absorber data collection, the present invention uses a single model to predict the results as the cumulative part of the regression chain. The cumulative design can not only mine y d Yu d-1 The relationship between the target variable and the absorber performance can also be directly analyzed. When learning along the chain, the estimate of the target variable is cascaded into the subsequent model as an additional feature, and the training dataset is enhanced in the order of the regression chain.

[0078] The present embodiment is described in detail below with reference to the accompanying drawings:

[0079] Step 1. Absorber Material Data Collection Module

[0080] First, identify the main raw materials of the absorber, including but not limited to carbon nanotubes. Collect the performance data of the absorber (real part, imaginary part, tangent value of dielectric constant; real part, imaginary part, tangent value of magnetic permeability, etc.) under different composition conditions (thickness, ratio, operating frequency, aspect ratio, tube diameter, etc.). Figure 2 The data collection results can be stored in electronic documents or using database software such as MySQL.

[0081] Step 2. Build modules based on correlation regression chain

[0082] In this embodiment, based on the data collected in the first step, the Pearson correlation coefficient analysis method can be used to analyze the correlation relationship of the absorber components. The Pearson correlation coefficient is shown as (1):

[0083]

[0084] Among them, x i and i are the data points of the two variables, and is the average of the two variables. r is the Pearson correlation coefficient, where +1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear correlation. In this embodiment, x and y represent the composition of the absorber material, such as thickness, ratio, aspect ratio, tube diameter, etc.

[0085] This embodiment assumes that the d-1th variable has a higher contribution to the prediction of the dth variable, and uses the Pearson correlation coefficient of formula (1) to calculate the variable y d The sum of the absolute values ​​of the correlation coefficients with other variables is used as the variable sorting criterion:

[0086]

[0087] This embodiment uses the coefficients of formula (2) to describe the linear relationship between the two variables of the absorber components, and then sorts them in descending order to form the largest correlation chain. Placing the variable with the highest sum of correlation coefficients at the front of the regression chain helps to predict other variables based on this variable, highlighting the advantage of its correlation coefficient.

[0088] Step 3. Base learner module

[0089] This embodiment uses a gradient boosted tree (GBDT) as a base learner to continuously improve the prediction performance based on local features and errors.

[0090] Taking the performance prediction component of the absorber material as an example, the data composition is {(x i ,y i )} n ,i=1,2,...,n, where n represents the number of samples, x i represents the material properties of the i-th sample, y i The material composition loss function to be predicted is L(y,F m (x)), which can be the mean square error in regression problems, F m (x) represents the mth weak classifier.

[0091] In the initialization stage, c can be the mean value of the training set material index, and F0(x) is expressed as:

[0092]

[0093] In the M-th iteration stage, the input of the tree is optimized for m=1,2,...,M. For the i-th sample, the negative gradient in the m-th tree is expressed as:

[0094]

[0095] Current decision tree h m (x)Use (x) i ,r m,i ) to train the weak classifier and obtain the corresponding leaf stage region R m,j , where j = 1, 2, ..., J m , J m is the number of leaf nodes of the mth regression tree. For each leaf node, its fitting value is obtained:

[0096]

[0097] Get the strong learner F M The expression of (x) is:

[0098]

[0099] Where I represents the indicator function, indicating whether the sample x is in the leaf node j. The last item of formula (6) can be added with a decay coefficient to gradually enhance the influence of the subsequent tree model.

[0100] Step 4. Correction module based on error coefficient regression chain

[0101] This example uses step 3 to perform basic prediction of components and finds that there are differences in the prediction error range. Based on step 2, the performance metric (error or determination coefficient) of the base learner in step 3 is introduced to calculate the fusion score S l , to adjust the link order construction.

[0102] Set the weight to sum the correlation S yd The error e with the base learner d Fusion is performed, as shown in formula (7), the error e d It can be the mean square error, the smaller it is, the better the prediction effect is. Among them, α and β are custom weights. Use weight fusion scores to update the order of regression chains to reduce the interference of variables with large errors on the order of regression chains.

[0103] S l =α*S yd -β*e d (7)

[0104] Among them, α and β can be set to 0.5 respectively, or set according to subsequent requirements.

[0105] Step 5. Accumulate feature enhancement modules

[0106] In this embodiment, the prediction result of the single component in step 3 is added to the regression link. is the prediction of the d-1th dimension regression chain, This variable depends only on the predicted result of material properties, so the d-th dimension prediction variable is:

[0107]

[0108] Under the condition of limited features of absorber data collection, this embodiment uses a separate model prediction result as the cumulative part of the regression chain. When learning along the chain, the estimate of the target variable is cascaded to the subsequent model as an additional feature, and the training data set is enhanced in the order of the regression chain. In this process, this embodiment can obtain D-1 models using the regression chain (the first prediction value does not participate in the regression chain construction).

[0109] Comprehensive steps 2-5, the schematic diagram of the model of this embodiment is as follows Figure 3 As shown:

[0110] Step 6. Model application intelligent material design module

[0111] Step 6: Use the model building results of steps 2-5 to make data predictions. The process is as follows:

[0112] 6.1 Determine the required material performance values ​​(thickness, ratio, operating frequency, aspect ratio, pipe diameter, etc.) and construct a complete data input.

[0113] 6.2 Using the corresponding GBDT classifier trained in step 3, obtain the composition prediction values ​​of the materials one by one according to the performance values.

[0114] 6.3 Using the regression chain constructed in step 4, obtain the link prediction order of the components, and obtain the link-related GBDT model that has been trained in this step.

[0115] 6.4 According to step 5, in the link prediction sequence, add the component prediction value of 6.2, and use the link-related model corresponding to 6.3 to design the material composition.

[0116] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for designing absorber material composition based on feature enhancement and multi-objective regression, characterized in that: include: Obtain absorber material data; Based on the absorber material data, constructing a regression chain initial sequence; Predicting relevant components of the absorber material data, and obtaining prediction results and prediction errors; Correcting the initial order of the regression chain using the prediction error to obtain a regression chain order; Based on the prediction result and the regression chain sequence, the material composition of the absorber is obtained.

2. The method for designing absorber material composition based on feature enhancement and multi-objective regression according to claim 1, characterized in that: Acquiring the absorber material data includes: acquiring the absorber performance data under different absorber material composition conditions; Among them, the absorber material composition includes: thickness, ratio, operating frequency, aspect ratio and tube diameter; the absorber performance data includes: real part of dielectric constant, imaginary part of dielectric constant, tangent value of dielectric constant, real part of magnetic permeability, imaginary part of magnetic permeability and tangent value of magnetic permeability.

3. The method for designing absorber material composition based on feature enhancement and multi-objective regression according to claim 1, characterized in that: Based on the absorber material data, the initial sequence of constructing a regression chain includes: Calculating the correlation of the absorber material data using the Pearson correlation coefficient analysis method to obtain a correlation coefficient value; The correlation coefficient values ​​are sorted in descending order to obtain the initial order of the regression chain.

4. The method for designing absorber material composition based on feature enhancement and multi-objective regression according to claim 1, characterized in that: Predicting the relevant components of the absorber material data and obtaining the prediction results and prediction errors include: Inputting the absorber material data into a base learner to obtain the prediction result, wherein the base learner is used to output a single component corresponding to the absorber material; According to the prediction result, a prediction error is obtained.

5. The method for designing absorber material composition based on feature enhancement and multi-objective regression according to claim 3, characterized in that: Correcting the initial order of the regression chain by using the prediction error, and obtaining the regression chain order includes: Sum the correlation coefficient values ​​to obtain a correlation sum; Setting weights to fuse the sum of the correlations and the prediction error to obtain a fusion score; The initial order of the regression chain is modified using the fusion score to obtain the regression chain order.

6. The method for designing absorber material composition based on feature enhancement and multi-objective regression according to claim 5, characterized in that: The method to obtain the fusion score is: S l =α*S yd -b*e d Among them, S l is the fusion score, α and β are weights, S yd is the sum of correlations, e d is the prediction error.

7. The method for designing absorber material composition based on feature enhancement and multi-objective regression according to claim 1, characterized in that: Based on the prediction result and the regression chain sequence, obtaining the absorber material composition includes: Using the prediction results as cumulative features of the regression chain sequence; The cumulative features and the target variable of the regression chain sequence are input into a regression link model to obtain the absorber material composition, wherein the regression link model is obtained through a training set, and the training set includes a regression chain sequence.

Images