A feature selection method for a cured tobacco threshing and redrying sheeting rate prediction model

Abstract

The present application provides a feature selection method for a flue-cured tobacco threshing and redrying sheeting rate prediction model, collects features of tobacco and corresponding threshing and redrying sheeting rate data through processing multiple batches of tobacco; performs feature selection through Pearson correlation analysis, Kendall correlation analysis and Spearman correlation analysis respectively and establishes corresponding prediction models through regression analysis; calculates comprehensive correlation and comprehensive significance of the features of the tobacco, and finally selects features used for the flue-cured tobacco threshing and redrying sheeting rate model. The feature selection method reduces the number of selected features and the preparation time for predicting the threshing and redrying sheeting rate, reduces the work burden of the detection personnel, effectively reduces the detection cost, and can more accurately predict the threshing and redrying sheeting rate result, providing faster and more accurate decision support for cigarette factories and redrying enterprises, which can more accurately adjust the procurement plan, prepare subsequent packaging materials, logistics capacity and storage space, and is beneficial to cost reduction and efficiency improvement of enterprises.

Description

TECHNICAL FIELD

[0001] The present application relates to the feature selection of tobacco threshing and redrying sheet rate, and in particular to a feature selection method for improving the accuracy of a tobacco threshing and redrying sheet rate prediction model, including physicochemical indexes, correlation analysis and feature selection. BACKGROUND

[0002] The sheet rate refers to the percentage of the weight of sheet tobacco obtained after processing of tobacco raw materials in the standard weight of the raw materials. Threshing and redrying is an essential part of the processing and production of flue-cured tobacco, which not only separates sheet tobacco from tobacco stems, but also ensures the quality of the product. The sheet rate is an important economic and technical indicator for processing enterprises. The main factors affecting the sheet rate include the physical properties of tobacco, the performance of threshing equipment and the process parameters during threshing.

[0003] The physical properties of tobacco, such as leaf structure, moisture content, thickness, density, shear force and penetration force, have a significant impact on the sheet rate. The chemical properties of tobacco also have some impact on the physical properties, such as total nitrogen and ash content, reducing sugar content, and so on.

[0004] The performance of key equipment, such as vacuum moisture regenerator, leaf moistener, threshing machine and redrying machine, directly affects the sheet rate. For example, different manufacturers, equipment usage time and maintenance quality have a significant impact on the sheet rate. The process parameters during threshing include the speed of the threshing roller, the impact force and the moisture content of the tobacco. Moderate increase in the speed of the threshing roller is beneficial to improving the sheet rate, but excessive speed will cause excessive crushing of the tobacco and reduce the sheet rate. The moisture content of the tobacco refers to the moisture content before threshing. Reasonable moisture content is beneficial to improving the sheet rate, but excessive or insufficient moisture content is not conducive to improving the sheet rate.

[0005] For a redrying plant, the performance of the threshing equipment and the process parameters during threshing are relatively constant, so it is necessary to find the features affecting the sheet rate from the physical and chemical indexes of the tobacco raw materials.

[0006] The application with the application number 201911228338.1 discloses a kind of detection method of tobacco leaf threshing and redrying data accuracy, comprising: according to the origin and flavor of tobacco leaf, tobacco leaf is divided into light flavor, strong flavor and intermediate flavor, and the tobacco leaf of each flavor is divided into upper leaf, middle leaf and lower leaf according to part;The mathematical model of the strip rate of different flavor, different part tobacco leaf and raw tobacco stem rate is established to represent the corresponding relationship between the strip rate of tobacco leaf and raw tobacco stem rate;According to the mathematical model, the upper limit value and the lower limit value of the strip rate corresponding to different flavor and different part tobacco leaf are calculated, and the strip rate interval corresponding to different flavor and different part tobacco leaf is established;The strip rate data of the tobacco leaf to be detected in threshing and redrying is obtained, whether the strip rate data is in the corresponding strip rate interval is judged, if not, it is judged that the strip rate data is abnormal.The above-mentioned application can improve the intelligence of enterprise production management and improve the accuracy of product data analysis.But the above-mentioned application does not select the characteristics used for flue-cured tobacco threshing and redrying strip rate prediction model, cannot more accurately predict the strip rate of tobacco leaf threshing and redrying, is not conducive to the accurate prediction of the strip rate of tobacco raw material by cigarette factory and redrying enterprise, and cannot accurately prepare subsequent packaging material, logistics capacity and storage site. SUMMARY

[0007] In view of the technical problem that it is difficult to select the characteristics used for flue-cured tobacco threshing and redrying strip rate prediction model, the present application provides a feature selection method for flue-cured tobacco threshing and redrying strip rate prediction model, which detects the physical and chemical indexes of tobacco raw material, innovatively carries out correlation analysis of tobacco physical and chemical indexes and threshing and redrying strip rate, selects indexes with significant characteristics from numerous physical and chemical indexes, establishes a threshing and redrying strip rate prediction model, and thus realizes accurate prediction of flue-cured tobacco threshing and redrying strip rate.

[0008] To achieve the above-mentioned purpose, the technical scheme of the present application is as follows:

[0009] A feature selection method for flue-cured tobacco threshing and redrying strip rate prediction model, comprising:

[0010] Step 1: Collect the characteristics of Q tobacco leaves and the corresponding threshing and redrying strip rate data by processing multiple batches of tobacco leaves;

[0011] Step 2: Select Q1 characteristics related to the threshing and redrying strip rate from the characteristics of Q tobacco leaves by Pearson correlation analysis, and establish a Pearson prediction model by regression analysis;

[0012] Step 3: Select Q2 characteristics related to the threshing and redrying strip rate from the characteristics of Q tobacco leaves by Kendall correlation analysis, and establish a Kendall prediction model by regression analysis;

[0013] Step 4: select Q3 features from Q features of the tobacco leaves by Spearman correlation analysis, and establish a Spearman prediction model by regression analysis;

[0014] Step 5: calculate the error proportion of the Pearson prediction model, the Kendall prediction model and the Spearman prediction model respectively, and obtain the credibility of the Pearson correlation analysis, the Kendall correlation analysis and the Spearman correlation analysis respectively;

[0015] Step 6: according to the credibility obtained in step 5, calculate the corresponding weight of the Pearson correlation analysis, the Kendall correlation analysis and the Spearman correlation analysis respectively, and calculate the comprehensive correlation and comprehensive significance of the Q features of the tobacco leaves, set the significance threshold of the significance test to 0.05, and finally select Q4 features for the tobacco redrying and leaf stripping rate model.

[0016] The comprehensive correlation R = W1r + W2τ + W3ρ, and the comprehensive significance R p = W1r p + W2τ p + W3ρ p , r p is the significance of the Pearson correlation analysis, τ p is the significance of the Kendall correlation analysis, ρ p is the significance of the Spearman correlation analysis, W1 is the weight of the Pearson correlation analysis, W2 is the weight of the Kendall correlation analysis, and W3 is the weight of the Spearman correlation analysis.

[0017] The implementation method of the step 2 is:

[0018] Step 2.1: calculate the Pearson correlation coefficient of Q features of the tobacco leaves and the tobacco redrying and leaf stripping rate respectively;

[0019] Step 2.2: perform significance test on the Pearson correlation coefficient to obtain the significance, select Q1 features which are significantly correlated with the tobacco redrying and leaf stripping rate from the Q features of the tobacco leaves according to the Pearson coefficient, establish a Pearson prediction model by regression analysis, and calculate the mean square error of the Pearson prediction model;

[0020] The significance level threshold of the Pearson chi-square test is set to 0.05.

[0021] The implementation method of the step 3 is:

[0022] Step 3.1: calculate the Kendall correlation coefficient of Q features of the tobacco leaves and the tobacco redrying and leaf stripping rate respectively;

[0023] Step 3.2: Significance test of Kendall correlation coefficient is significant, according to the Kendall correlation coefficient in Q tobacco leaf characteristics, select out Q2 characteristics significantly related to the threshing and redrying leaf out rate, establish the Kendall prediction model by regression analysis and calculate the mean square error of the Kendall prediction model;

[0024] The significance level threshold of Kendall correlation coefficient chi-square test is set to 0.05.

[0025] The implementation method of step 4 is:

[0026] Step 4.1: Calculate the Spearman correlation coefficient of Q tobacco leaf characteristics and threshing and redrying leaf out rate respectively;

[0027] Step 4.2: Significance test of Spearman correlation coefficient is significant, according to the Spearman correlation coefficient in Q tobacco leaf characteristics, select out Q3 tobacco leaf characteristics significantly related to the threshing and redrying leaf out rate, establish the Spearman prediction model by regression analysis and calculate the mean square error of the Spearman prediction model;

[0028] The significance level of Spearman correlation coefficient chi-square test is set to 0.05.

[0029] The Pearson correlation coefficient calculation method is: r k Pearson correlation coefficient of the kth characteristic, Indicates the kth characteristic value of the ith tobacco sample, y i Indicates the threshing and redrying leaf out rate of the ith tobacco sample, N is the number of samples, Indicates the average value of all tobacco samples on the kth characteristic, Indicates the average value of all sample threshing and redrying leaf out rate.

[0030] The Kendall correlation coefficient calculation method is: In the formula, τ g Kendall correlation coefficient of the gth characteristic, e is the number of observations, The rank of the lth observation of the gth characteristic, rank(y j ) is the rank of the jth observation of the label, Compare all possible feature and label combinations, calculate the sum of absolute value of rank difference.

[0031] The Spearman correlation coefficient calculation method is: The rank difference between the rank of the s th observation of the f th characteristic and the rank of the label, e is the number of observations, ρ fThe Kendall correlation coefficient between the fth feature and the threshing and redrying leaf stripping rate.

[0032] The error ratio is the mean square error of the prediction model / total error, the total error is the mean square error of the Pearson prediction model+the mean square error of the Kendall prediction model+the mean square error of the Spearman prediction model, and the reliability is 100%-error ratio.

[0033] The calculation method of the weight is: W u The weight corresponding to the correlation analysis, h u The reliability corresponding to the correlation analysis, u=1, 2, and 3 respectively correspond to Pearson correlation analysis, Kendall correlation analysis, and Spearman correlation analysis.

[0034] Compared with the prior art, the present application has the beneficial effects that the present application reduces the detection indexes of physical and chemical indexes, significantly reduces the work burden of the detection personnel, and effectively reduces the detection cost. The preparation work time for predicting the threshing and redrying leaf stripping rate is also greatly reduced, thereby greatly improving the efficiency of the threshing and redrying leaf stripping rate prediction, providing faster and more accurate decision support for cigarette factories and redrying enterprises, and enabling more accurate adjustment of the purchase plan, preparation of subsequent packaging materials, logistics capacity, and storage space, which is beneficial to cost reduction and efficiency improvement of enterprises. The leaf stripping rate prediction model can also be used to guide tobacco cultivation and conditioning to improve the leaf stripping rate by adjusting the characteristic indexes of tobacco raw materials. The features selected by the three analysis methods are fused to form a new feature selection scheme, and a regression equation is established to predict the threshing and redrying leaf stripping rate. Compared with single use of Pearson correlation analysis, Kendall correlation analysis, or Spearman correlation analysis, the feature selection method of the present application not only reduces the number of selected features, but also can more accurately predict the threshing and redrying leaf stripping rate result. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0036] Figure 1 The figure is a schematic diagram of the feature selection method of the present application for the tobacco threshing and redrying leaf stripping rate prediction model. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0038] As shown in the figure, a feature selection method for a cured tobacco primary processing leaf stripping rate prediction model includes the following steps: Figure 1

[0039] Step 1: Collect the characteristics of Q batches of tobacco leaves and the corresponding primary processing leaf stripping rate data through processing.

[0040] The shear force, penetration force, tensile force, adhesion force, thickness, total sugar, reducing sugar, total nitrogen, total potassium, total chlorine, nicotine, starch, sulfur ion, anti-dry crushing property, dichloromethane extract, oxalic acid, succinic acid, malonic acid, neochlorogenic acid, anisatin, rutin, malic acid, asparagine, glycine, phenylalanine, histidine, Glu-An (glutamine), Fru-Asn (fruit aspartic acid), Fru-Ile (fruit isoleucine), Fru-Tyr (fruit tyrosine), neophytadiene, and more than seventy physical and chemical indicators of a batch of cured tobacco of Hebei Tobacco Industry Co., Ltd. are determined.

[0041] Table 1 Physical and chemical characteristics

[0042]

[0043] Specifically, the physical indicators such as shear force, penetration force, tensile force, and adhesion force of the tobacco leaves are detected according to the provisions and requirements of YQ-GY / T 7.2-2023 Tobacco Physical Property Determination. The total sugar, reducing sugar, total nitrogen, total potassium, total chlorine, nicotine, and starch content in the tobacco leaves are determined according to the provisions and requirements of the national standard YC / T 159-2019 Continuous Flow Determination. The chemical indicators such as dichloromethane extract, oxalic acid, succinic acid, malonic acid, neochlorogenic acid, anisatin, rutin, malic acid, asparagine, glycine, phenylalanine, histidine, Glu-An, Fru-Asn, Fru-Ile, Fru-Tyr, and neophytadiene are detected by near-infrared spectroscopy technology. The primary processing leaf stripping rate is expressed as the percentage of the weight of the obtained cut tobacco to the standard weight of the raw tobacco leaves after processing the raw tobacco leaves.

[0044] Step 2: Select Q1 characteristics related to the primary processing leaf stripping rate from the characteristics of the Q tobacco leaves through Pearson correlation analysis, and establish a Pearson prediction model through regression analysis.

[0045] ​Step 2.1: Calculate the Pearson correlation coefficient of Q characteristics of tobacco leaves and the threshing and redrying sheeting rate respectively.

[0046] Pearson correlation coefficient, also known as Pearson product-moment correlation coefficient, is a statistical measure of the degree of linear correlation between two variables, used to assess the linear correlation between two continuous variables. Its value ranges between -1 and 1, the closer to 1 or -1, the stronger the linear relationship between the two variables; the closer to 0, the weaker the linear relationship between the two variables.

[0047] Features: Measures the strength and direction of linear relationship between two variables.

[0048] Advantages: When there is a linear relationship between variables, Pearson correlation coefficient can effectively capture the strength of the relationship. It is the most commonly used correlation measure, and the results are easy to understand.

[0049] Applicable scenarios: Suitable for continuous data and variables that meet the normal distribution.

[0050] The formula for calculating Pearson correlation coefficient is as follows:

[0051]

[0052] r k Pearson correlation coefficient of the kth feature, yk represents the kth feature value of the ith tobacco leaf sample, y i N represents the threshing and redrying sheeting rate of the ith tobacco leaf sample, and N represents the number of samples. yk represents the average value of all tobacco leaf samples on the kth feature, N represents the average value of all sample threshing and redrying sheeting rates.

[0053] Step 2.2: Perform significance test on Pearson correlation coefficient to obtain significance, select Q1 features that are significantly correlated with threshing and redrying sheeting rate from Q characteristics of tobacco leaves according to Pearson coefficient, establish Pearson prediction model through regression analysis and calculate the mean square error of Pearson prediction model.

[0054] In order to judge the statistical significance of Pearson correlation coefficient, the purpose of significance test is to judge whether there is significant difference between tobacco characteristics and label (threshing and redrying sheeting rate), so as to judge the correlation between tobacco characteristics and threshing and redrying sheeting rate. The significance level threshold of Pearson chi-square test is set to 0.05.

[0055] The significance P-value of Pearson correlation coefficient is usually calculated by hypothesis testing, which is based on the t-distribution of Pearson coefficient. The t-statistic is calculated, whose formula is:

[0056]

[0057] t k is the t-statistic of the kth feature, r k is the Pearson correlation coefficient of the kth feature, and n is the number of trials. The degrees of freedom (df) of the t-statistic is determined, which is n-2 for Pearson correlation coefficient. The P-value corresponding to the t-statistic, i.e., the significance, is determined using the t-distribution table or computer software (such as R, Python, SPSS, etc.).

[0058] The selected features are F1…F Q1 A production experiment is conducted on multiple batches of tobacco raw materials, and a Pearson prediction model is established based on regression analysis (such as linear regression model, neural network, etc., determined according to the number of samples) on the selected Q1 features. The mean square error MSE r of the Pearson prediction model is calculated to evaluate the prediction accuracy of the Pearson prediction model.

[0059] Step 3: Select Q2 features that are related to the threshing and redrying leaf stripping rate from the Q tobacco leaf features through Kendall correlation analysis, and establish a Kendall prediction model through regression analysis.

[0060] Step 3.1: Calculate the Kendall correlation coefficient of the Q tobacco leaf features and the threshing and redrying leaf stripping rate.

[0061] Kendall correlation analysis is a statistical method for quantifying the correlation of rank or order between two variables, which is a kind of rank correlation coefficient, usually represented by the Greek letter τ. Kendall correlation coefficient is mainly applied to ordered categorical variables, which can measure the similarity of the ordering of two variables. The value of Kendall correlation coefficient τ is between -1 and 1. When τ = 1, it means that the two variables are completely consistent; when τ = -1, it means that the two variables are completely opposite; when τ = 0, it means that there is no correlation between the two variables.

[0062] Features: Non-parametric method based on rank, measures the correlation between two variables.

[0063] Advantages: Strong robustness to outliers, does not require data to conform to a specific distribution, suitable for handling small sample data.

[0064] Applicable scenarios: Suitable for cases where data does not conform to normal distribution, sample size is small, or there are outliers.

[0065] The calculation formula of Kendall correlation coefficient is as follows:

[0066]

[0067] where τ g is the Kendall correlation coefficient of the gth feature, e is the number of observations (the number of observations is the number of features of Q tobacco leaves measured), is the rank of the lth observation of the gth feature; rank(y j ) is the rank of the jth observation of the label (primary processing flake rate). (The rank of the first observation is 1, the rank of the second observation is 2, and so on.) is the sum of the absolute values of the rank differences for all possible combinations of features and labels (primary processing flake rate).

[0068] Step 3.2: Perform a significance test on the Kendall correlation coefficient to obtain significance, select Q2 features that are significantly correlated with the primary processing flake rate from the Q features of the tobacco leaves according to the Kendall correlation coefficient, establish a Kendall prediction model through regression analysis, and calculate the mean square error of the Kendall prediction model.

[0069] To determine the statistical significance of the Kendall correlation coefficient, the purpose of the significance test is to determine whether there is a significant difference between the features of the tobacco leaves and the label (primary processing flake rate), and thus to determine the correlation between the features of the tobacco leaves and the primary processing flake rate. The significance level threshold for the Kendall correlation coefficient chi-square test is set to 0.05.

[0070] The significance test of the Kendall correlation coefficient is usually performed by calculating its corresponding t-statistic, which is as follows:

[0071]

[0072] t g is the t-statistic of the gth feature, τ g is the Kendall correlation coefficient of the gth feature, and n is the number of trials. After calculating the t value, determine the degrees of freedom (df) of the t-statistic. For the Kendall correlation coefficient, the degrees of freedom is n-2. Use the t-distribution table or computer software (such as R, Python, SPSS, etc.) to determine the P value corresponding to the t-statistic, i.e., the significance.

[0073] The selected features are F1…F Q2 , and through production experiments on multiple batches of tobacco raw materials, a Kendall prediction model is constructed based on the obtained data through linear regression analysis, and the mean square error MSE τ of the Kendall prediction model is calculated to evaluate the prediction accuracy of the Kendall prediction model.

[0074] Step 4: Select Q3 features from Q features of tobacco leaves that are correlated with the redrying strip rate by Spearman correlation analysis, and establish a Spearman prediction model by regression analysis.

[0075] Step 4.1: Calculate the Spearman correlation coefficient of Q features of tobacco leaves and the redrying strip rate respectively.

[0076] Spearman's rank correlation coefficient, also known as Spearman's rank correlation coefficient, is a non-parametric statistical index that measures the dependence between two variables, suitable for evaluating the correlation between the rank or order data of two variables. Spearman's correlation coefficient can also be used to evaluate the correlation between two continuous variables, but it does not require them to be normally distributed. When p is close to 1, it indicates that there is a strong positive correlation between the two variables. When p is close to -1, it indicates that there is a strong negative correlation between the two variables. When p is close to 0, it indicates that there is no or weak correlation between the two variables.

[0077] Features: It is also a non-parametric method based on rank, but compared with Kendall correlation coefficient, it pays more attention to the ordering and hierarchical structure of data.

[0078] Advantages: It also has good resistance to outliers and does not require data distribution, suitable for processing ordered categorical data and non-linear relationship data.

[0079] Applicable scenarios: It is suitable for variables that may have non-linear relationships or are ordered categorical data.

[0080] The calculation formula of Kendall correlation coefficient is as follows:

[0081]

[0082] is the rank difference between the rank of the s-th observation of the f-th feature and the rank of the label (redrying strip rate), e is the number of observations, p f is the Kendall correlation coefficient between the f-th feature and the redrying strip rate.

[0083] Step 4.2: Perform significance test on the Spearman correlation coefficient to obtain significance, select Q3 features of tobacco leaves that are significantly correlated with the redrying strip rate from Q features of tobacco leaves according to the Spearman correlation coefficient, establish a Spearman prediction model by regression analysis, and calculate the mean square error of the Spearman prediction model.

[0084] For the significance test of Spearman correlation coefficient, the purpose of significance test is to determine whether there is a significant difference between the characteristics of tobacco leaves and the label (primary processing yield), so as to determine the correlation between the characteristics of tobacco leaves and the primary processing yield. The significance level of Spearman correlation coefficient chi-square test is set to 0.05.

[0085] Convert p to t statistics. The formula for calculating t statistics is:

[0086]

[0087] t f t statistics of the fth feature, p f Spearman correlation coefficient of the fth feature, n is the number of trials.

[0088] Determine the degrees of freedom (df) of t statistics. For Spearman correlation coefficient, the degrees of freedom is n-2. Use t distribution table or computer software (such as R, Python, SPSS, etc.) to determine the P value corresponding to t statistics, i.e. significance.

[0089] The selected features are F1…F Q3 Through production experiments on multiple batches of tobacco raw materials, and based on the obtained data, a Spearman prediction model is constructed by linear regression analysis, and the mean square error MSE ρ of the Spearman prediction model is calculated to evaluate the prediction accuracy of the Spearman prediction model.

[0090] Step 5: Calculate the error proportion of Pearson prediction model, Kendall prediction model, and Spearman prediction model respectively, and obtain the reliability of Pearson correlation analysis, Kendall correlation analysis, and Spearman correlation analysis respectively.

[0091] According to the mean square error MSE r of the Pearson prediction model, the mean square error MSE τ of the Kendall prediction model, and the mean square error MSE ρ of the Spearman prediction model, the total error is:

[0092] Total error = MSE r + MSE τ + MSE ρ (7)

[0093] Then the error proportion of the Pearson prediction model is:

[0094]

[0095] Then the error proportion of the Kendall prediction model is:

[0096]

[0097] The error proportion of the Spearman prediction model is:

[0098]

[0099] According to the error proportions of the Pearson prediction model, the Kendall prediction model and the Spearman prediction model, the reliability = 100% - error proportion, it is concluded that the reliability of Pearson correlation analysis is (1-a), the reliability of Kendall correlation analysis is (1-b), and the reliability of Spearman correlation analysis is (1-c).

[0100] Step 6: According to the reliability obtained in step 5, the weights corresponding to Pearson correlation analysis, Kendall correlation analysis and Spearman correlation analysis are calculated respectively, so as to calculate the comprehensive correlation and comprehensive significance of the characteristics of Q tobacco leaves. The significance critical value of significance test is set to 0.05, and finally Q4 characteristics for flue-cured tobacco redrying and leaf threshing rate model are selected.

[0101] It is found from steps 2, 3 and 4 that the features selected by the three correlation analysis methods are inconsistent, and the number of features is large. Pearson correlation analysis selects Q1 features, Kendall correlation analysis selects Q2 features, and Spearman correlation analysis selects Q3 features.

[0102] Compared with single use of Pearson correlation analysis, Kendall correlation analysis or Spearman correlation analysis, the three correlation analysis methods have certain limitations. In statistical analysis, Pearson correlation analysis requires data to be normally distributed, while Spearman correlation analysis is suitable for non-normal distributed data. Pearson correlation analysis may not reveal the significant relationship between them. While Spearman correlation analysis as a non-parametric test method does not require the distribution characteristics of data, so in this case it can show the significant correlation between the two variables and the leaf threshing rate. According to practical experience, although Pearson correlation analysis does not show that there is a significant relationship between a certain feature and the leaf threshing rate, in fact the role of these chemical components in the production process may be very critical. Therefore, in practical application, even if the statistical data does not completely support, the potential influence of these factors still needs to be concerned, and appropriate consideration should be given in process optimization and production control.

[0103] To solve this problem, the application considers the Pearson correlation analysis, Kendall correlation analysis and Spearman correlation analysis in combination, and deeply fuses the correlation of each feature and the label (primary tobacco processing flake rate). The weight-based fusion strategy and feature selection strategy are as follows: the weight of the corresponding correlation method is calculated according to the credibility proportion, the corresponding correlation is weighted, the three different correlation coefficients and the significance are recalculated to obtain the comprehensive correlation and comprehensive significance through the weight, although they may not be significant in the Pearson correlation analysis, Kendall correlation analysis or Spearman correlation analysis, but they show significance in other types of correlation analysis. This may be because different correlation analysis methods have different requirements for the distribution, type and relationship assumption of data, and thus different correlation patterns are revealed.

[0104] The weight-based fusion strategy and feature selection strategy are as follows: in order to fuse the features, the weight of the corresponding correlation method is calculated according to the credibility proportion, the corresponding correlation method is weighted, and the calculation formula of the weight W is as shown in formula (10):

[0105]

[0106] W u is the weight of the corresponding correlation analysis, h u is the credibility of the corresponding correlation analysis, u = 1, 2, 3 respectively correspond to the Pearson correlation analysis, Kendall correlation analysis and Spearman correlation analysis.

[0107] The weights of the Pearson correlation analysis, Kendall correlation analysis and Spearman correlation analysis are calculated through formula (10) to be W1, W2 and W3 respectively, and the comprehensive correlation R and the comprehensive significance R p of the features of the Q tobaccos are calculated through formula (12) and formula (13) respectively.

[0108] R = W1r + W2τ + W3ρ (12)

[0109] R p = W1r p + W2τ p +W3ρ p (13)

[0110] In the formula, r p is the significance of the Pearson correlation analysis, τ p is the significance of the Kendall correlation analysis, and ρ p is the significance of the Spearman correlation analysis.

[0111] The significance level threshold of the significance test of the application is set to 0.05, and the comprehensive correlation R p≤0.05 as the threshold value to select Q4 features for the flue-cured tobacco threshing and redrying strip rate model, and the selected features are F1…F Q4 A comprehensive prediction model is established by linear regression analysis.

[0112] The following is a specific embodiment of the feature selection method for the flue-cured tobacco threshing and redrying strip rate prediction model of the present application.

[0113] Hebei China Tobacco Industry Co., Ltd. purchased a batch of tobacco raw materials in Liancheng, Fujian, and then sent the raw materials to Fujian Longyan Jinye Redrying Factory for further processing. During the processing, the tobacco was comprehensively detected for physical and chemical indicators, and relevant data such as threshing and redrying strip rate were recorded. The Pearson correlation coefficient and significance, Kendall correlation coefficient and significance, and Spearman correlation coefficient and significance of each feature indicator and the threshing and redrying strip rate were calculated. Through Pearson correlation analysis, 21 features related to the threshing and redrying strip rate were selected, through Kendall correlation analysis, 25 features related to the threshing and redrying strip rate were selected, and through Spearman correlation analysis, 26 features related to the threshing and redrying strip rate were selected. Prediction models of the three correlation analysis methods were established, the mean square error of each prediction model was calculated, the weight of the corresponding correlation method was calculated according to the weight fusion strategy and feature selection strategy of the present application according to the credibility ratio, the corresponding correlation method was weighted, and the fused correlation coefficient and significance were calculated.

[0114] Specifically, 21 features related to the threshing and redrying strip rate were selected through Pearson correlation analysis, as shown in Table 2.

[0115] Table 2 Pearson correlation coefficient of physical and chemical indicators and threshing and redrying strip rate

[0116]

[0117] Through the Pearson correlation coefficient and its significance, it can be known that shear force, thickness, brittleness, nicotine, total nitrogen, dichloromethane extract, solanesol, tetradecanoic acid, linoleic acid, asparagine, glutamic acid, methionine, phenylalanine, arginine, Glu-An, Fru-Leu, and sixteen indicators have a positive correlation with the threshing and redrying strip rate, i.e. the larger the values of these indicators, the higher the threshing and redrying strip rate, among which the shear force and thickness have the largest positive correlation with the threshing and redrying strip rate and the highest significance; potassium, magnesium, succinic acid, Fru-Ala, and neophytadiene, five indicators, have a negative correlation with the threshing and redrying strip rate, i.e. the larger the values of these indicators, the lower the threshing and redrying strip rate, among which succinic acid has the largest negative correlation with the threshing and redrying strip rate and the highest significance.

[0118] A Pearson prediction model of the features selected by the Pearson correlation coefficient and significance and the threshing and redrying strip rate is established:

[0119] Y p = 10.30 + 1.66 x 10 -2 X1+ 4.27 X2+…+ 1.34 x 10 -2 X 21 .

[0120] Specifically, 25 features related to the threshing and redrying leaf out sheet rate were selected by Kendall correlation analysis, as shown in Table 3.

[0121] Table 3 Kendall correlation coefficient of physical and chemical indexes and threshing and redrying leaf out sheet rate

[0122]

[0123] Through the Kendall correlation coefficient and its significance, it can be known that the shear force, tensile force, thickness, brittleness, tetradecanoic acid, oleic acid + linolenic acid, asparagine, glutamic acid, glutamine, glycine, alanine, lysine, tryptophan, arginine, Glu-An, Fru-Val, Fru-Gly, Fru-Asp eicosanoic acid, anisatin, 20 indexes are positively correlated with the threshing and redrying leaf out sheet rate, that is, the larger the values of these indexes, the higher the threshing and redrying leaf out sheet rate, among which the tensile force has the largest positive correlation with the threshing and redrying leaf out sheet rate and the highest significance; potassium, succinic acid, cystine, Fru-Ala, neophytadiene, five indexes are negatively correlated with the threshing and redrying leaf out sheet rate, that is, the larger the values of these indexes, the lower the threshing and redrying leaf out sheet rate, among which the potassium has the largest negative correlation with the threshing and redrying leaf out sheet rate and the highest significance.

[0124] A Kendall prediction model of the features selected by Kendall correlation coefficient and significance and the threshing and redrying leaf out sheet rate was established:

[0125] Y k = 3.78 + 4.12 x 10 -1 X1+ 9.28 x 10 -2 X2+…+ 2.16 x 10 -2 X 24 .

[0126] Specifically, 26 features related to the threshing and redrying leaf out sheet rate were selected by Spearman correlation analysis, as shown in Table 4.

[0127] Table 4 Spearman correlation coefficient of physical and chemical indexes and threshing and redrying leaf out sheet rate

[0128]

[0129] The correlation coefficient and significance of Spearman showed that the shear force, tensile force, thickness, penetration force, adhesion force, brittleness, nicotine, solanols, tetradecanoic acid, oleic acid + linolenic acid, asparagine, glutamic acid, glycine, alanine, methionine, phenylalanine, 4-aminobutyric acid, arginine, Glu-An, Fru-His, and twenty indicators were positively correlated with the threshing and redrying leaf output rate, that is, the larger the values of these indicators, the higher the threshing and redrying leaf output rate, among which the tensile force, asparagine, and arginine had the largest positive correlation with the threshing and redrying leaf output rate and the highest significance; potassium, succinic acid, octadecanoic acid, tryptophan, Fru-Ala, and neophytadiene had a negative correlation with the threshing and redrying leaf output rate, that is, the larger the values of these indicators, the lower the threshing and redrying leaf output rate, among which potassium had the largest negative correlation with the threshing and redrying leaf output rate and the highest significance.

[0130] A Spearman prediction model of the features selected by the correlation coefficient and significance of Spearman and the threshing and redrying leaf output rate was established.

[0131] Y (斯皮尔曼) = 25.62 + 4.12 x 10 -1 X1+ 2.62 x 10 -2 X2+ … + 1.38 x 10 -1 X 26 .

[0132] The three prediction models, i.e., the Pearson prediction model, the Kendall prediction model, and the Spearman prediction model, were verified in production. The data of multiple batches of raw tobacco were collected, the three prediction models were deployed to the production environment to ensure that the input and output of the models matched the production data interface. The same batch of raw materials was predicted using the three prediction models, and the prediction results were recorded. The differences between the prediction results of the three prediction models and the actual production results were compared to evaluate the accuracy and reliability of each prediction model. The weights were calculated according to the method of step 6, and the results are shown in Table 5.

[0133] Table 5 Error of each model

[0134] Pearson Kendall Spearman number of choices 21 24 26 mean squared error 1.09 2.48 0.97 weight 0.24 0.55 0.21

[0135] According to the weights of each feature indicator, the correlation and significance of each feature were recalculated based on the weights, the features were fused, the significance level critical value of the significance test was set to 0.05, and new features were selected. Finally, 13 features for the flue-cured tobacco threshing and redrying leaf output rate model were selected, as shown in Table 6. The correlation coefficients of the features were iteratively calculated based on the weights, and a comprehensive prediction model was established for production verification.

[0136] Table 6 Correlation coefficient of physical and chemical indicators and threshing and redrying leaf output rate based on weight calculation

[0137] feature thickness tensile force shear force brittleness potassium succinic acid new correlation coefficient based on weights 0.786 0.851 0.763 0.730 -0.756 -0.725 significance 0.017 0.005 0.025 0.029 0.021 0.036 feature myristic acid asparagine glutamic acid arginine Glu-An Fru-His new correlation coefficient based on weights 0.706 0.751 0.785 0.805 0.754 0.701 significance 0.043 0.029 0.015 0.012 0.022 0.040 feature Fru-Ala new correlation coefficient based on weights -0.688 significance 0.047

[0138] It can be known by calculating the feature correlation coefficient and its significance through weight reiteration that the shear force, the tension, the thickness, the brittleness, the tetradecanoic acid, the asparagine, the glutamic acid, the arginine, the Glu-An and the Fru-His ten indexes are in positive correlation with the threshing and redrying sheeting rate, that is, the greater the values of these indexes are, the higher the threshing and redrying sheeting rate is, wherein the tension has the greatest positive correlation with the threshing and redrying sheeting rate and the highest significance; the potassium, the succinic acid and the Fru-Ala three indexes are in negative correlation with the threshing and redrying sheeting rate, that is, the greater the values of these indexes are, the lower the threshing and redrying sheeting rate is, wherein the potassium has the greatest negative correlation with the threshing and redrying sheeting rate and the highest significance.

[0139] Establishing a comprehensive prediction model of the threshing and redrying sheeting rate based on the selected features after weight fusion

[0140] Y (发明) = -165.54 + 4.12 x 10 -1 X1+ 5.34 x 10 -3 X2… 5.14 x 10 -2 X 13 .

[0141] Feature selection comparison: the Pearson correlation analysis selects 21 features, the Kendall correlation analysis selects 24 features, the Spearman correlation analysis selects 26 features, and the feature selection method of the present application selects 13 features, that is, the feature selection of the present application reduces 8 features compared with the Pearson correlation analysis, reduces 11 features compared with the Kendall correlation analysis, and reduces 13 features compared with the Spearman correlation analysis.

[0142] The four prediction models are deployed in the production environment to ensure that the input and output of the model match the production data interface. The same batch of raw materials is predicted using the four prediction models, and the prediction results are recorded. The differences between the prediction results of the four prediction models and the actual production results are compared to evaluate the accuracy and reliability of each prediction model. The mean square error of the Pearson prediction model established by Pearson correlation analysis of feature selection is 1.13; the mean square error of the Kendall prediction model established by Kendall correlation analysis of feature selection is 2.31; the mean square error of the Spearman prediction model established by Spearman correlation analysis of feature selection is 1.01; and the average error of the comprehensive prediction model established by the feature selection method of the application is 0.31. The comprehensive prediction model established by the feature selection method used in the application reduces the error by 0.88 compared with the Pearson prediction model established by Pearson correlation analysis of feature selection, reduces the error by 2.00 compared with the Kendall prediction model established by Kendall correlation analysis of feature selection, and reduces the error by 0.70 compared with the Spearman prediction model established by Spearman correlation analysis of feature selection.

[0143] Table 7 Comparison of feature selection of Pearson, Kendall, Spearman and the application

[0144] Pearson Kendall Spearman the invention number of choices 21 24 26 13 model error 1.29 2.31 1.01 0.31

[0145] In summary, the feature selection method proposed in the application has the following advantages:

[0146] Comprehensiveness: By combining Pearson correlation coefficient, Kendall correlation coefficient and Spearman correlation coefficient, the linear relationship of data, rank correlation and sensitivity to outliers can be considered comprehensively. This can obtain reliable feature correlation evaluation under different types of data relationship and distribution conditions.

[0147] Robustness: When facing outliers or non-normal distribution data, Kendall and Spearman correlation coefficients provide additional stability, which can help improve the quality of feature selection and the final performance of the model.

[0148] Flexibility: Suitable for various data types and structures, whether linear or nonlinear, continuous or categorical data.

[0149] The application reduces the detection indicators of physical and chemical indicators, significantly reduces the work burden of detection personnel, and effectively reduces the detection cost. It also greatly reduces the preparation time for predicting the threshing and redrying leaf stripping rate, thereby greatly improving the efficiency of predicting the threshing and redrying leaf stripping rate, and providing faster and more accurate decision support for cigarette factories and redrying enterprises.

[0150] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A feature selection method for a cured tobacco threshing and redrying leaf yield prediction model, characterized in that, The method comprises the following steps: Step 1: Collecting Q characteristics of tobacco leaves and corresponding threshing and redrying sheeting rates of the tobacco leaves by processing multiple batches of tobacco leaves; Step 2: Selecting Q1 characteristics that are correlated with the threshing and redrying sheeting rate from the Q characteristics of the tobacco leaves by Pearson correlation analysis, and establishing a Pearson prediction model by regression analysis; Step 3: Selecting Q2 characteristics that are correlated with the threshing and redrying sheeting rate from the Q characteristics of the tobacco leaves by Kendall correlation analysis, and establishing a Kendall prediction model by regression analysis; Step 4: Selecting Q3 characteristics that are correlated with the threshing and redrying sheeting rate from the Q characteristics of the tobacco leaves by Spearman correlation analysis, and establishing a Spearman prediction model by regression analysis; Step 5: Calculating the error proportion of the Pearson prediction model, the Kendall prediction model and the Spearman prediction model respectively, and obtaining the reliability of the Pearson correlation analysis, the Kendall correlation analysis and the Spearman correlation analysis respectively; Step 6: According to the reliabilities obtained in step 5, calculating the corresponding weights of the Pearson correlation analysis, the Kendall correlation analysis and the Spearman correlation analysis respectively, and thus calculating the comprehensive correlation and comprehensive significance of the Q characteristics of the tobacco leaves, setting the significance threshold of the significance test as 0.05, and finally selecting Q4 characteristics for the flue-cured tobacco threshing and redrying sheeting rate model; The comprehensive correlation R = W1r + W2τ + W3ρ, the comprehensive significance R p = W1r p + W2τ p + W3ρ p , r p is the significance of Pearson correlation analysis, τ p is the significance of Kendall correlation analysis, ρ p is the significance of Spearman correlation analysis, W1 is the weight of Pearson correlation analysis, W2 is the weight of Kendall correlation analysis, and W3 is the weight of Spearman correlation analysis; According to the integrated saliency R p ≤0.05Select out Q4 features for the tobacco threshing and redrying sheeting rate model.

2. The feature selection method for the flue-cured tobacco primary processing yield prediction model according to claim 1, characterized in that, The implementation method of step 2 is as follows: Step 2.1: Calculating the Pearson correlation coefficient of the characteristics of the Q tobacco leaves and the threshing and redrying sheeting rate respectively; Step 2.2: Performing significance test on the Pearson correlation coefficient to obtain significance, selecting Q1 characteristics that are significantly correlated with the threshing and redrying sheeting rate from the Q characteristics of the tobacco leaves according to the Pearson correlation coefficient, establishing a Pearson prediction model by regression analysis and calculating the mean square error of the Pearson prediction model; The significance level threshold of the Pearson correlation coefficient chi-square test is set as 0.

05.

3. The feature selection method for the flue-cured tobacco primary processing yield prediction model according to claim 2, characterized in that, The implementation method of step 3 is as follows: Step 3.1: Calculating the Kendall correlation coefficient of the characteristics of the Q tobacco leaves and the threshing and redrying sheeting rate respectively; Step 3.2: Performing significance test on the Kendall correlation coefficient to obtain significance, selecting Q2 characteristics that are significantly correlated with the threshing and redrying sheeting rate from the Q characteristics of the tobacco leaves according to the Kendall correlation coefficient, establishing a Kendall prediction model by regression analysis and calculating the mean square error of the Kendall prediction model; The significance level threshold of the Kendall correlation coefficient chi-square test is set as 0.

05.

4. The feature selection method for the flue-cured tobacco primary processing yield prediction model according to claim 3, characterized in that, The implementation method of step 4 is as follows: Step 4.1: Calculating the Spearman correlation coefficient of the characteristics of the Q tobacco leaves and the threshing and redrying sheeting rate respectively; Step 4.2: Performing significance test on the Spearman correlation coefficient to obtain significance, selecting Q3 characteristics that are significantly correlated with the threshing and redrying sheeting rate from the Q characteristics of the tobacco leaves according to the Spearman correlation coefficient, establishing a Spearman prediction model by regression analysis and calculating the mean square error of the Spearman prediction model; The significance level threshold of the Spearman correlation coefficient chi-square test is set as 0.

05.

5. The feature selection method for the flue-cured tobacco primary processing yield prediction model according to claim 4, characterized in that, The Pearson correlation coefficient calculation method is: r k is the Pearson correlation coefficient of the kth feature, represents the kth feature value of the ith tobacco sample, y i represents the threshing and redrying sheeting rate of the ith tobacco sample, N is the sample number, represents the average value of all tobacco samples on the kth feature, represents the average value of all sample threshing and redrying sheeting rates.

6. The feature selection method for the flue-cured tobacco primary processing yield prediction model according to claim 5, characterized in that, The Kendall correlation coefficient calculation method is: In the formula, τ g is the Kendall correlation coefficient of the gth feature, e is the number of observation values, is the rank of the lth observation value of the gth feature, rank(y j ) is the rank of the jth observation value of the label, is to compare all possible feature and label combinations, calculate the sum of the absolute value of the rank difference.

7. The feature selection method for the flue-cured tobacco primary processing yield prediction model according to any one of claims 4-6, characterized in that, The Spearman correlation coefficient calculation method is: is the rank difference of the s-th observation value of the f-th feature and the label rank, e is the number of observation values, and p f is the Kendall correlation coefficient between the f-th feature and the leaf cleaning and threshing outsheet rate.

8. The feature selection method for the flue-cured tobacco primary processing yield prediction model according to claim 7, characterized in that, The error proportion is equal to the mean square error of the prediction model / total error, the total error is equal to the mean square error of the Pearson prediction model+the mean square error of the Kendall prediction model+the mean square error of the Spearman prediction model, and the reliability is equal to 100%-the error proportion. The error proportion is equal to the mean square error of the prediction model / total error, the total error is equal to the mean square error of the Pearson prediction model+the mean square error of the Kendall prediction model+the mean square error of the Spearman prediction model, and the reliability is equal to 100%-the error proportion. The error proportion is equal to the mean square error of the prediction model / total error, 9. The feature selection method for the flue-cured tobacco primary processing yield prediction model according to claim 1 or 8, characterized in that, The calculation method of the weight is: W u h is the weight corresponding to the correlation analysis, u u=1, 2, 3 respectively correspond to Pearson correlation analysis, Kendall correlation analysis, and Spearman correlation analysis.

Images