Two-stage frequency selection method and device for microwave frequency sweep data

A two-stage frequency selection method for microwave frequency sweep data used to detect a material attribute, comprising:

• acquiring microwave frequency sweep data collected by a microwave frequency sweep detection system;
• normalizing the microwave frequency sweep data, and then partitioning an attenuation training data set and a phase shift training data set, wherein the two data sets exist in the form of a data table, data in a vertical axis of the data table represents a frequency domain {f1, f2, . . . , fi, . . . , fn}, data in a horizontal axis of the data table represents a sample domain {X1, X2, . . . , Xj, . . . , Xm}, and data elements corresponding to the data table are attenuation values A or phase shift values Phi;
• by using a random forest-recursive feature elimination algorithm, performing frequency selection on the microwave frequency sweep data, taking a preset parameter in the random forest-recursive feature elimination algorithm as a hyper-parameter, changing the value of the hyper-parameter, and generating a series of candidate frequency subsets within different frequencies sets, wherein the frequency selection step comprises:
• (2.1) training, by using the random forest algorithm, a material attribute prediction model on the attenuation training data set;
• (2.2) obtaining the importance of attenuation features corresponding to each frequency, sorting the frequencies according to the importance of features, and finding out frequencies with the lowest importance of the corresponding features;
• (2.3) removing attenuation feature data corresponding to the frequencies with the lowest importance of the corresponding attenuation features from the attenuation training data set, and retraining the material attribute prediction model on the updated attenuation training data set by using the random forest algorithm;
• (2.4) repeating steps (2.2) and (2.3) until only the data corresponding to PreNum frequencies remain in the attenuation training data set, and recording the set consisting of the PreNum frequencies as a frequency set FA, wherein PreNum is a number of features to be selected in the algorithm and may be taken as a hyper-parameter;
• (2.5) training a material attribute prediction model on the phase shift training data set by using the random forest algorithm;
• (2.6) obtaining the importance of phase shift features corresponding to each frequency, sorting the frequencies according to the importance of features, and finding out frequencies with the lowest importance of the corresponding features;
• (2.7) removing phase shift feature data corresponding to the frequencies with the lowest importance of the corresponding phase shift features from the phase shift training data set, and retraining the material attribute prediction model on the updated phase shift training data set by using the random forest algorithm;
• (2.8) repeating steps (2.6) and (2.7) until only the data corresponding to PreNum frequencies remains in the phase shift training data set, and recording the set consisting of the PreNum frequencies as a frequency set FP;
• (2.9) taking the intersection of the frequency set FA and the frequency set FP to obtain a candidate frequency subset Fsub;
• and
• (2.10) changing the value of the preset parameter PreNum of the random forest-recursive feature elimination algorithm, and repeating steps (2.1) to (2.9) to obtain a series of candidate frequency subsets within different frequencies sets;
• building the material attribute prediction models on the basis of the frequency sweep data corresponding to the candidate frequency subsets of different frequencies, wherein this step comprises:
• each candidate frequency subset corresponding to a frequency sequence number subset, extracting corresponding data from the attenuation training data set and the phase shift training data set respectively by using the frequency sequence number subsets, and combining the two parts of data into attenuation-phase shift frequency sweep data sets;
• and
• taking each attenuation-phase shift frequency sweep data set as input data and material attribute values as output data, and building prediction models for the material attribute values by using learning algorithms;
• wherein each candidate frequency subset corresponding to a frequency sequence number subset, extracting corresponding data from the attenuation training data set and the phase shift training data set respectively by using the frequency sequence number subsets, and combining the two parts of data into attenuation-phase shift frequency sweep data sets, which comprises:
• (4.1) searching for the sequence number of each frequency in the candidate frequency subset in the normalized attenuation frequency sweep data set or phase shift frequency sweep data set to form a frequency sequence number subset;
• (4.2) repeating step (4.1) until the frequency sequence number subset corresponding to each candidate frequency subset in step (3) is obtained;
• (4.3) extracting corresponding data from the attenuation training data set according to the frequency sequence number subset;
• (4.4) extracting corresponding data from the phase shift training data set according to the frequency sequence number subset;
• (4.5) concatenating the columns of the two data sets respectively to obtain an attenuation-phase shift frequency sweep data set corresponding to the candidate frequency subset;
• and
• (4.6) repeating steps (4.3)-(4.5) until a corresponding attenuation-phase shift frequency sweep data set is obtained for each candidate frequency subset;
• evaluating the performance of each prediction model by means of 10 fold cross validation, and calculating evaluation index values of model performance;
• and
• taking the evaluation indexes as a voting basis, and selecting an optimal frequency subset by using a majority voting method, which comprises:
• (6.1) using R2 as an index of the voting basis, selecting top k models with the maximum R2 value under each of T algorithms, obtaining a frequency subset corresponding to each model, and selecting a frequency subset with the most votes by using the majority voting method on the T×k candidate results, denoted as
• F
• opt
• R
• 2
• ;
• (6.2) using root mean square error (RMSE) as an index of the voting basis, selecting top k models with the minimum RMSE value under each of T algorithms, obtaining a frequency subset corresponding to each model, and selecting a frequency subset with the most votes by using the majority voting method on the T×k candidate results, denoted as
• F
• opt
• RMSE
• ;
• (6.3) using mean absolute error (MAE) as an index of the voting basis, selecting top k models with the minimum MAE value under each of T algorithms, obtaining a frequency subset corresponding to each model, and selecting a frequency subset with the most votes by using the majority voting method on the T×k candidate results, denoted as
• F
• opt
• MAE
• ;
• and
• (6.4) using the majority voting method to synthesize the optimal frequency sets
• F
• opt
• R
• 2
• ,
• F
• opt
• RMSE
• ⁢
• and
• ⁢
• F
• opt
• MAE
• respectively selected on the basis of the three regressive evaluation indexes, and selecting a final optimal frequency set Fopt, or selecting the frequency set within the least number of frequencies as the optimal frequency set Fopt in the event of a tie between optimal frequency values, and
• constructing a frequency sweep signal using the optimal frequency subset Fopt, and measuring the material attribute using the frequency sweep signal.