Secondary similar substitution method for cigarette formula maintenance

Abstract

The invention discloses a secondary similar substitution method for cigarette formula maintenance, which comprises the following steps of: firstly, calculating weighted values of q quality indexes of cigarettes in a plurality of seasons in the past to obtain an overall matrix; performing similarity evaluation on the current stock tobacco leaves to obtain a similar matrix of the stock tobacco leaves; s to-be-replaced tobacco leaves in the cigarette formula containing the M tobacco leaf groups are recognized, p tobacco leaves most similar to the to-be-replaced tobacco leaves are screened out from the similar matrix, and a new tobacco leaf group formula # imgabs0 of the cigarette is obtained; based on M leaf group weighted values of q quality indexes of a # imgabs 1 # new leaf group formula, obtaining # imgabs 2 # weighted values; respectively carrying out principal component analysis dimensionality reduction on the overall matrix and the # imgabs3 # weighted values to obtain an overall dimensionality reduction matrix and a # imgabs4 # dimensionality reduction vector; and calculating the mahalanobis distance from the dimensionality reduction vector of the # imgabs5 # seed to the overall dimensionality reduction matrix, and taking a new leaf group formula corresponding to the minimum mahalanobis distance as a similar alternative cigarette formula, so that the problems of low efficiency and difficulty in optimization of the cigarette leaf group formula are solved.

Description

Technical Field

[0001] The invention relates to a secondary similarity substitution method for maintaining cigarette formula, belonging to the technical field of tobacco formula analysis. Background Art

[0002] In the prior art, the maintenance of tobacco leaf group formula is mainly carried out by formula personnel based on objective data such as tobacco origin, grade, conventional chemical composition, etc., combined with years of formula experience and compatibility mode accumulation, to first design several preliminary tobacco alternatives, and then determine the final formula maintenance plan through artificial sensory evaluation of actual formula samples. However, this model has three major drawbacks: first, it relies heavily on previous formula experience and compatibility modes, the flexibility of cigarette formula is poor, and the quality is difficult to effectively improve, which has an adverse effect on formula maintenance and new product development; second, due to the large number of chemical and sensory indicators concerned, the tobacco leaf formula work is intensive and subjective, which is not only inefficient, but also often difficult to find the most suitable alternative tobacco leaves; third, as a raw material of agricultural products, the quality of tobacco leaves is greatly affected by the climate of the year, and it is difficult for tobacco industry enterprises to effectively predict this. The original formula of cigarettes must be appropriately adjusted based on the actual quality characteristics of tobacco leaf raw materials, which has created huge challenges for the field of cigarette formulas guided by experience. With the continuous development of cigarette brands, the imbalance in tobacco leaf inventory and purchasing regions, parts and categories has become more prominent. Although tobacco industry enterprises have continuously optimized leaf formulas and processing techniques in recent years, the bottlenecks in the supply and demand structure of tobacco leaf raw materials and the cost pressures they face have gradually become prominent.

[0003] Tobacco industry enterprises have realized new digital cigarette formulation solutions by introducing near-infrared analysis technology combined with traditional formulation experience. At present, tobacco leaf near-infrared spectral analysis technology can realize the identification of conventional chemical components and some sensory indicators of tobacco leaves and the analysis of style characteristics. By combining the similarity of near-infrared spectra with chemometric methods, the consistency of two tobacco leaves can be more comprehensively characterized from the overall perspective, thereby achieving effective substitution of tobacco leaves. However, if tobacco leaf substitution only considers the similarity of near-infrared spectra, the control range is too wide. On the one hand, it is easy to introduce some noise interference, and it also puts forward higher requirements for near-infrared spectral preprocessing; on the other hand, it lacks primary and secondary for the entire spectrum, and lacks attention to the bands involved in some important tobacco leaf quality indicators. If you only focus on the similarity of several quality indicators predicted by the near-infrared model, you will often generalize and fail to grasp the overall style quality characteristics of tobacco leaves. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a secondary similarity substitution method for cigarette formula maintenance to solve the problems of experience-based cigarette leaf group formula, low formula efficiency, difficulty in formula optimization and tight raw material supply.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] In a first aspect, the present invention provides a secondary similarity substitution method for maintaining cigarette formula, comprising:

[0007] Based on the near infrared spectrum-tobacco leaf quality index prediction model, the weighted values ​​of q quality indexes of cigarettes in the past several quarters are calculated to obtain the overall matrix;

[0008] According to the tobacco leaf similarity evaluation of the current tobacco leaves in stock, a similarity matrix of the tobacco leaves in stock is obtained;

[0009] By identifying s tobacco leaves to be replaced in a cigarette blend containing M tobacco leaf groups, p tobacco leaves that are most similar to the tobacco leaves to be replaced are selected from the similarity matrix to obtain p s A new leaf group formula of the cigarette;

[0010] Based on p s The M leaf group weighted values ​​of the q quality indicators of the new leaf group formula are obtained by s Weighted values;

[0011] For the overall matrix, p s The weighted values ​​are reduced by principal component analysis to obtain the overall dimension reduction matrix and p s A reduced dimensionality vector;

[0012] According to the calculation p s The Mahalanobis distance from the reduced dimension vector to the overall reduced dimension matrix is ​​obtained, and the new leaf group formula corresponding to the minimum Mahalanobis distance is the cigarette formula of similar substitution.

[0013] In combination with the first aspect, further, by evaluating the similarity of the current tobacco leaves in stock, a similarity matrix of the tobacco leaves in stock is obtained, including:

[0014] Multi-point sampling to collect near-infrared spectral data of current tobacco leaves in stock;

[0015] The near-infrared spectral data of the current stock tobacco leaves are subjected to scattering correction, synchronous smoothing derivation, feature selection and two-dimensional similarity calculation in sequence to obtain the similarity matrix of the stock tobacco leaves.

[0016] Furthermore, the synchronous smooth derivation includes:

[0017] The data in the moving window of the scattering-corrected spectral data are fitted with a least squares polynomial, and the fitting function is synchronously derived to obtain synchronous smoothed derivative spectral data.

[0018] Furthermore, the method also includes: using cubic spline interpolation to supplement the head and tail of the scattering-corrected spectral data with data of half the size of the window.

[0019] Furthermore, calculating the two-dimensional similarity includes:

[0020] Based on the shape differences and information content differences in the spectral data, the total similarity of the spectral data of any two tobacco leaves after feature selection is calculated;

[0021] The calculation formula for the total similarity is:

[0022]

[0023] Where: S represents the total similarity of the spectrum, D shape represents the shape similarity represented by the dynamic time warping distance; D Info It represents the difference in the amount of information contained in two spectra characterized by information divergence;

[0024]

[0025] Where: j = 1, 2, 3, .., n, x j ,y j are the data points of two different tobacco leaf samples after synchronous smoothing and derivation of near-infrared spectra; n is the number of spectral data points;

[0026] Furthermore, the q quality indicators include:

[0027] If the cigarette is a single recipe: select q quality indicators based on the style characteristics of the cigarette brand and the actual recipe;

[0028] If the cigarette is not a single recipe: according to the principle of feature selection, use relative standard deviation to select q quality indicators; or set at least five members to vote and select q quality indicators based on the total weight.

[0029] Furthermore, at least five members are set to vote and select q quality indicators based on the total weight, including:

[0030] Each member makes a primary and secondary judgment on any two chemical indicators to obtain a comparison matrix;

[0031] Convert the comparison matrix into a judgment matrix.

[0032] The objective function is constructed by the least square method, the auxiliary function is constructed by the Lagrange conditional extreme value, and the weight is obtained by solving the judgment matrix;

[0033] The obtained weights are averaged as the total weight, and q chemical indicators that are greater than the total weight are selected as quality indicators.

[0034] Furthermore, the method further comprises:

[0035] Calculate the correlation coefficient of the weights;

[0036] The quality index selected was eliminated by eliminating members whose correlation coefficient was lower than the critical correlation coefficient at 95% confidence level.

[0037] In a second aspect, the present invention provides a computer device, characterized in that it includes:

[0038] A memory for storing instructions;

[0039] The processor is used to execute the instructions so that the device performs operations to implement any one of the above-mentioned secondary similarity substitution methods for maintaining cigarette formula.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention adopts near-infrared spectroscopy for preliminary screening through a primary replacement algorithm, selects several tobacco leaves that are most similar to the replaced tobacco leaves as a whole to form a pre-selection pool, and adopts Mahalanobis distance selection through a secondary replacement algorithm. The formulator specifies several quality indicators to select tobacco leaves in the primary replacement pre-selection pool with the smallest Mahalanobis distance to the whole, thereby effectively solving the technical problems of low efficiency and difficult optimization of tobacco leaf group formulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flow chart of a secondary similarity substitution method for cigarette formula maintenance provided by an embodiment of the present invention;

[0043] Figure 2 A relationship diagram between the number of principal components and the contribution provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0045] Embodiment 1:

[0046] like Figure 1 As shown, a flowchart of a secondary similarity substitution method for cigarette formula maintenance provided in this embodiment is provided. Taking a cigarette brand L to be maintained (whose formula table contains M tobacco leaf groups) as an example, the method includes the following steps:

[0047] Step 1: Based on the tobacco leaf near-infrared spectrum sample data and the measured tobacco leaf quality index data, a near-infrared spectrum-tobacco leaf quality index prediction model is established.

[0048] Step 2: Select q quality indicators as the focus indicators based on the style characteristics and formula experience of L cigarettes, and use the prediction model established in step 1 to calculate the weighted values ​​of q quality indicators of L cigarettes in the past at least two years (8 quarters) to obtain the overall matrix U8×q .

[0049] Step 3: Collect the near-infrared spectra of the current stock tobacco leaves, build a tobacco leaf similarity evaluation method based on the near-infrared spectra, and obtain the similarity matrix of the stock tobacco leaves (first-level similarity sample pool).

[0050] Step 4: Identify the s tobacco leaves in the L cigarette formula that need to be replaced passively due to insufficient inventory or actively due to quality problems. According to the similarity matrix established in step 3, select the p tobacco leaves that are most similar to each of the s tobacco leaves, and obtain p s This L cigarette new leaf group formula.

[0051] Step 5: Use the same method as step 2 to obtain p s The M leaf group weighted values ​​of the q quality indicators of the new leaf group formula.

[0052] Step 6: For the overall matrix U 8×q Use principal component analysis (PCA) to reduce dimension, select the first t principal components with contribution > ratio1, and get the overall dimension-reduced matrix u 8×t .

[0053] Step 7: Step 5 s The weighted values ​​are reduced by the same t principal components as in step 6, and p s The dimension-reduced vector v t .

[0054] Step 8: Calculate p s v t to u 8×t The Mahalanobis distance of , where the leaf group formula (secondary similar sample) corresponding to the minimum Mahalanobis distance is the similar alternative cigarette formula.

[0055] Specifically, the formulator can select the most satisfactory one among several new leaf group formulas with the smallest Mahalanobis distance (such as the smallest 5 to 10 Mahalanobis distances) based on actual conditions such as production planning, inventory arrangement and sensory evaluation.

[0056] The above step 1 includes collecting tobacco leaf near-infrared spectrum sample data, measuring tobacco leaf quality index data, and establishing a near-infrared spectrum-tobacco leaf quality index prediction model. The specific implementation process is as follows:

[0057] Step 1.1: Collect tobacco leaf near-infrared spectrum sample data:

[0058] The tobacco leaf samples were dried at below 50°C for about 4 hours until they could be crushed with fingers. The samples were ground using a Foss Cyclotec 1093 cyclone mill (Foss, Denmark), and the sample powder was sieved through a 60-mesh (aperture 250 μm) sieve. An appropriate amount was placed in a sample cup, and a fixed mass weight was placed on top of the sample. After it was naturally compacted, it was sampled by a Fourier transform near-infrared spectrometer (Thermo Fisher Scientific China Co., Ltd.) at multiple points, and the response values ​​at 4000 to 10000 wave numbers (one point was taken every 6 wave numbers, the number of scans was 72 times, and a total of 1000 data for each near-infrared curve) were collected to obtain the tobacco leaf near-infrared spectrum sample data.

[0059] Step 1.2: Determine tobacco leaf quality index data:

[0060] According to GB / T 19616-2004 General Principles for Sampling of Tobacco Batch Raw Materials, tobacco leaf samples were extracted and samples were prepared according to YC / T31-1996 Preparation and Moisture Determination of Tobacco and Tobacco Products Samples;

[0061] According to "YC / T 159-2019 Determination of water-soluble sugars in tobacco and tobacco products", "YC / T 468-2021 Determination of total alkaloids in tobacco and tobacco products", "YC / T 162-2011 Determination of chlorine in tobacco and tobacco products", "YC / T 161-2002 Determination of total nitrogen in tobacco and tobacco products", "YC / T 217-2007 Determination of potassium in tobacco and tobacco products", the prepared samples were respectively determined for the contents of six conventional chemical components of tobacco leaves, namely total sugar (TS), reducing sugar (RS), total alkaloids (TA), chlorine (Cl), total nitrogen (TN) and potassium (K);

[0062] According to "YC / T 222-2007 Determination of pH in tobacco and tobacco products", "YC / T 202-2006 Determination of chlorogenic acid, scopoletin and rutin in tobacco and tobacco products", "YC / T 216-2013 Determination of starch in tobacco and tobacco products - continuous flow method", the prepared samples were tested for pH, chlorogenic acid and starch in tobacco leaves. Two parallel tests were performed for each sample, and the results were averaged.

[0063] Combined with the usage characteristics of tobacco industry companies and the accuracy of near-infrared spectroscopy modeling, the sensory style index was optimized and designed, including 6 items: light fragrance value, intermediate fragrance value, strong fragrance value, position value, elegance 1 and elegance 2. A 7-person professional team certified by the tobacco industry completed the evaluation and scoring of each tobacco leaf sample. The scores of the 6 indicators were all 5 to 10 points, and they were positive indicators. To ensure the reliability of the results, the two highest scores and the two lowest scores for each indicator were removed with reference to the scoring rules of competitive sports, and the average value of the remaining three was taken.

[0064] The 15 tobacco leaf quality indicators are composed of the above 9 tobacco leaf chemical indicators and 6 tobacco leaf sensory style scores.

[0065] Step 1.3: Establish near infrared spectroscopy-tobacco leaf quality index prediction model:

[0066] Based on the collected tobacco leaf near-infrared spectroscopy sample data and the above 15 tobacco leaf quality index data, a relatively stable near-infrared spectroscopy-tobacco leaf quality index prediction model was established using a chemometric method. The fitting determination coefficient between the true value and the predicted value of the general validation set samples reached R 2 (Generally required to be >0.9) or above can ensure that the model is usable.

[0067] Step 2 includes selecting q quality indicators as the focus indicators based on the style characteristics and formulation experience of L cigarettes, and using the prediction model established in step 1 to calculate the weighted values ​​of q quality indicators of L cigarettes at least two years (8 quarters) from now to obtain the overall matrix U 8×q The specific implementation process is as follows:

[0068] Step 2.1, selection of q quality indicators:

[0069] Since different formulators may pay attention to different tobacco quality indicators for different cigarette brands, this application arbitrarily selects from the following three different methods.

[0070] Method ①: A single formulation staff member independently selects q tobacco leaf quality indicators of concern based on the style characteristics of L cigarette brand and actual formulation needs.

[0071] Method ②: Select q items with a total weight > ratio2 as the focus indicators through expert group voting.

[0072] The expert group composed of at least 5 members is based on the understanding of the importance of chemical indicators. Since it is difficult to accurately weight multiple indicators at the same time and the confidence is not high, but it is easier to make a primary and secondary judgment on any two indicators, this application uses the 0-2 three-scale-hierarchy analysis method (3-ahp) to construct a comparison matrix g ij ,as follows:

[0073]

[0074] In order to reduce the arbitrariness of subjective decision-making, g ij Further transformed into the judgment matrix h ij :

[0075]

[0076] Where: h ij is the judgment matrix; If b i ≥ b j , T=1, if b i <b j , T = -1; i, j, k are the 15 tobacco leaf quality indicators in step 1.3;

[0077] Furthermore, the present application uses the least square method with better discrimination to calculate the weight w ahp :

[0078]

[0079] Where: (w ahp ) i >0.

[0080] The auxiliary function N is constructed using the Lagrange conditional extreme value method:

[0081]

[0082] Where: γ is an auxiliary parameter,

[0083] By ahp ) i Find partial derivatives, joint We get 16 linear equations with 16 variables, and after solving them we get w ahp .

[0084] The AHP method must be tested for consistency, and must meet the following requirements:

[0085] CR=CI / RI=(λ max -15) / (15-1) / RI<0.1,

[0086] Where: max is the judgment matrix q ij The maximum eigenvalue of; judgment matrix h ij When the order is 15, the parameter RI = 1.59.

[0087] In order to exclude the possibility that experts may have special "favor" or "discrimination" towards certain tobacco leaf quality indicators, an error correction mechanism is added, namely: the indicators of interest selected by experts whose average Pearson correlation coefficient (PR) with others is lower than the critical correlation coefficient at 95% confidence level are eliminated.

[0088] The remaining expert group members get w ahp Find the mean, and we get Select the q items whose total weight > ratio2 as the indicators of interest.

[0089] Method ③: Based on the principle of feature selection, the q item with relative standard deviation > ratio3 is used as the focus indicator

[0090]

[0091] Where: m represents the number of tobacco leaf samples in stock; y j Indicates tobacco leaf quality index.

[0092] Select RSD_ZB j >The q item of ratio3 is taken as the focus indicator, and ratio3 is determined to be between 0.01 and 0.10.

[0093] Step 2.2, taking the quality of L cigarettes in the past two years (8 quarters) as the target, the established near-infrared prediction model is used to calculate the q-item quality index values ​​of the M tobacco leaves in the tobacco leaf group formula, and the weighted values ​​of the M tobacco leaves are calculated according to the proportion w of the L cigarette leaf group, and the overall matrix U is obtained. 8×q :

[0094]

[0095] Where: Q 8×q×k represents the q-item quality index value of the k-th leaf group of L cigarettes in 8 quarters; w i Indicates the usage ratio of the kth leaf group.

[0096] In step 3, the near infrared spectra of the current stock tobacco leaves are collected, and a tobacco leaf similarity evaluation method is constructed based on the near infrared spectra to obtain a similarity matrix of the stock tobacco leaves. The calculation of the similarity matrix of the stock tobacco leaves is divided into five steps: near infrared spectrum measurement of stock tobacco leaves, near infrared spectrum scattering correction, near infrared spectrum synchronous smoothing derivation, near infrared spectrum feature selection, and near infrared spectrum two-dimensional similarity algorithm. The specific process is as follows:

[0097] Step 3.1, Near infrared spectroscopy measurement of stock tobacco leaves:

[0098] Tobacco leaves, as biomass raw materials, are often not uniform enough. When sampling stock tobacco leaves, it should be ensured that they are sampled at multiple points and mixed as one sample, and then the near-infrared spectrum of the stock tobacco leaves should be determined according to the method in step 1.1.

[0099] Step 3.2, near infrared spectroscopy scatter correction:

[0100] Standard normal variable transformation (SNV) is mainly used to eliminate the influence of tobacco powder particle size, surface scattering and optical path change on near-infrared diffuse reflectance spectrum. The new spectral data X after SNV processing ij for:

[0101]

[0102] Where: X' represents the original spectrum measured in step 3.1; X represents the new spectrum after processing; n represents the number of spectral points, which is 1000 here, and X ij represents the jth spectral point of the i-th sample of the new spectrum, X i ' j Represents the jth spectral point of the i-th sample of the original spectrum.

[0103] Step 3.3, synchronous smoothing derivation of near-infrared spectrum:

[0104] Derivation of near-infrared spectra can eliminate baseline drift or reduce background interference, distinguish overlapping peaks, and improve the resolution and sensitivity of the spectrum. However, for some high-frequency noise, differentiation will further amplify these noise signals, and the original spectrum generally needs to be smoothed before differentiation. Although excessive smoothing is conducive to enhancing the signal-to-noise ratio after differentiation, the spectral information loss is large. If the smoothing is insufficient, it is easy to increase the high-frequency noise after differentiation.

[0105] This application adopts the synchronous smoothing derivation method (SSD), which performs the two operations of derivation and smoothing simultaneously, effectively avoiding the tedious operation of repeated debugging and optimization. The method is to perform least squares fitting of the data in a moving window through a polynomial, and synchronously derive the fitting function to obtain the weighted average expression of the spectral derivative signal. Select a cubic polynomial with a moving window size of 7, then the polynomial to be fitted is:

[0106]

[0107] in, For X ij The fitting value of ; eps is the fitting error; a0, a1, a2, a3 are all parameters to be fitted; i = 1, 2, 3, ..., m, are different inventory cigarette strips; m is the number of cigarette strip samples; j = -3, -2, -1, 0, 1, 2, 3, are the 7 relative position points in the moving window, then the polynomial to be fitted can be expressed as: in According to the least squares fitting method, the values ​​of a0, a1, a2, and a3 can be obtained:

[0108]

[0109] For a cubic polynomial Taking the derivative of both sides, we get:

[0110]

[0111] where x ij for The derivative value of , then the above polynomial can be expressed as:

[0112]

[0113] The midpoint x of the moving window i0 As the signal value of the new spectrum after processing, then:

[0114] x i0 =0.087X i-3 -0.266X i-2 -0.230X i-1 +0.230X i1 +0.266X i2 +0.087X i3 ,

[0115] Change the position of the moving window to obtain the new spectral data x after synchronous smoothing derivative preprocessing ij In order to keep the number of spectral points the same before and after processing, since the window size is 7, cubic spline interpolation is used to interpolate 3 points at the beginning and end of the spectrum before synchronous smoothing and derivation of the near-infrared spectrum.

[0116] Step 3.4, near infrared spectroscopy feature selection:

[0117] In addition to the information of the sample itself, the near-infrared spectrum also contains a lot of irrelevant and repeated information. It is necessary to perform feature selection (FS) on the spectrum to screen the appropriate bands and improve the modeling effect. The spectral feature selection of this application adopts the coefficient of variation method. The relative standard deviation of the absorbance at each wavelength point of all inventory sample spectra is calculated according to the following formula to obtain the variance spectrum:

[0118]

[0119] Where: m represents the number of tobacco leaf samples in stock.

[0120] Choose RSD-NIR j >The spectral point of ratio4, ratio4 is generally determined between 0.01 and 0.10.

[0121] Step 3.5, near infrared spectrum two-dimensional similarity algorithm:

[0122] The differences between the two near-infrared spectra include shape differences and information content differences. This application establishes a two-dimensional similarity evaluation method that combines the two, namely:

[0123]

[0124] Where: S is the total similarity of the spectra. The larger S is, the higher the similarity between the two tobacco samples is, and vice versa. D shape The shape similarity represented by the dynamic time warping distance is expressed as:

[0125] D shape (n,n)=|xn -y n |+min[D shape (n-1,n),D shape (n,n-1),D sshape (n-1,n-1)],

[0126] Where: x n ,y n are the data points of two different tobacco leaf samples after spectral preprocessing in step 3.3; n is the number of spectral data points, which is 1000 here;

[0127] D Info The information divergence is the difference in the amount of information contained in the two spectra, and its expression is:

[0128]

[0129] in:

[0130] Calculate the similarity of all the spectra of the stocked tobacco strips, and arrange the names of the tobacco strips from large to small according to the similarity, and obtain the tobacco strip similarity matrix, which is recorded as S mm-1 :

[0131]

[0132] Where: s i1 ,s i2， s i3 ,…,s im-1 It indicates the names of the tobacco leaves, excluding itself, arranged in descending order of similarity to tobacco leaf i.

[0133] In step 4, identify the s tobacco leaves in the L cigarette formula that need to be passively replaced due to insufficient inventory or actively replaced due to quality problems. According to the similarity matrix established in step 3, select the top p tobacco leaves that are most similar to each of the s tobacco leaves, and get p s Using the same method as step 2, p s The q quality indicators of the new leaf group formula (same as in step 2.1) are calculated, and the weighted values ​​of the M leaf groups are obtained.

[0134]

[0135] Where: Indicates L cigarette p s The kth leaf group of a new recipe has the qth quality index value of interest; w i Indicates the usage ratio of the kth leaf group.

[0136] Steps 6 and 7 perform PCA dimensionality reduction:

[0137] For the overall matrix U 8×q Use principal component analysis (PCA) to reduce dimension, select the first t principal components with contribution > ratio1, and get the overall dimension-reduced matrix u 8×t .right Using the same t principal components for dimensionality reduction, we get p s The dimension-reduced vector v t .

[0138] Step 8 performs secondary optimization and calculates the minimum Mahalanobis distance, specifically calculating p s v t to u 8×t The Mahalanobis distance d is:

[0139]

[0140] in: for u 8×t The eight-quarter average of for The transpose of -1 for u 8×t The inverse of the covariance matrix of .

[0141] The smaller the d value, the higher the similarity. The tobacco leaf with the minimum d value is the target replacement tobacco leaf combination. When using it, the formulator can select the most satisfactory one among several new leaf group formulas with the smallest d value (such as the smallest 5 to 10 d values) as the new tobacco leaf group formula based on actual conditions such as production planning, inventory arrangement and sensory evaluation.

[0142] For specific data processing, this application uses Matlab 2020b to write auxiliary calculation scripts according to the above steps to complete the calculation process of all processes.

[0143] Embodiment 2:

[0144] This embodiment takes the formula maintenance of Liqun Longzui cigarette brand as an example to perform secondary similarity substitution of formula maintenance. The specific process is as follows:

[0145] Taking 12000Kg of feed as an example, its usage ratio and leaf group feed amount are shown in Table 1. As shown in Table 1, Liqun Long Mouth Cigarettes are composed of 20 leaf groups, and the leaf group ratio is 5%, so the required feed amount is 600Kg. According to the current inventory quantity, the six types of tobacco leaves GZ3120CFL, HA4118CCH, HN1121CEA, SC4119CEA, SC4120XEY, and YN8119CKH are insufficient in stock and need to be replaced with other similar tobacco leaves. The other tobacco leaves are sufficient in stock and have no quality problems, so they do not need to be replaced.

[0146] Table 1: Liqun Longzui formula

[0147] Recipe number Tobacco leaf name Leaf group ratio w(%) Stock quantity (Kg) Leaf group feed amount (Kg) 1 CA18GLH 5 7330 600 2 GZ20CFP 5 3404 600 3 GZ20CFL 5 448 600 4 HA18CCH 5 522 600 5 HA19XCS 5 3028 600 6 HB18CCQ 5 2288 600 7 HN19XDG 5 1327 600 8 HN21CEA 5 356 600 9 HN21CEB 5 6802 600 10 SC19CDY 5 7273 600 11 SC19CEA 5 371 600 12 SC20XEY 5 292 600 13 YN19XFL 5 1120 600 14 YN19BES 5 3340 600 15 YN19CKB 5 1637 600 16 YN21CLM 5 1020 600 17 YN21CLN 5 7263 600 18 YN19CKN 5 656 600 19 YN19CKH 5 65 600 20 ZW21ILJ 5 1650 600

[0148] Note: ① The first two letters of the tobacco strip code represent the place of origin, CA for Canada, GZ for Guizhou, HA for Henan, HB for Hubei, HN for Hunan, SC for Sichuan, YN for Yunnan, and ZW for Zimbabwe; the numbers in the tobacco strip code represent the year of production; the fifth digit of the tobacco strip code is the part code; the sixth and seventh digits are identification codes and have no clear meaning.

[0149] The parameter selection involved in the secondary similarity substitution method is shown in Table 2.

[0150] Table 2: Parameter selection

[0151]

[0152] According to step 1.1, collect tobacco leaf near infrared spectroscopy data; according to step 1.2, complete tobacco leaf quality index data collection; according to step 1.3, build a near infrared spectroscopy-tobacco leaf quality index prediction model; and optimize the model so that the true value and predicted value fitting determination coefficient of the validation set samples reach R 2 =0.9 or above.

[0153] In this case, method ② in step 2.1 is considered. Five tobacco leaf quality experts (A, B, C, D and E) establish a comparison matrix g according to method ② in step 2.1. ij Table 3 lists the comparison matrix of expert A.

[0154] Table 3: Comparison matrix of expert A

[0155]

[0156]

[0157] Note: z1~z15 represent 15 tobacco leaf quality indicators: total sugar, reducing sugar, total alkaloids, chlorine, total nitrogen, potassium, pH, chlorogenic acid, starch, light aroma value, intermediate aroma value, strong aroma value, part value, elegant 1 and elegant 2.

[0158] Five experts’ judgment matrix h ij The maximum eigenvalue λ max They are 15.75, 15.21, 15.35, 16.37 and 15.21 respectively, and the consistency test CR values ​​are all less than 0.1. The objective function Z is constructed by the least square method, and the auxiliary function N is constructed by the Lagrangian conditional extreme value. The weighting results of the five experts are solved, as shown in Table 4.

[0159] Table 4: Weighting results of five experts

[0160] A B C D E <![CDATA[w ahp ]]> z1 0.076 0.09 0.089 0.034 0.058 0.069 z2 0.136 0.09 0.089 0.127 0.122 0.113 z3 0.207 0.234 0.209 0.158 0.234 0.208 z4 0.015 0.015 0.015 0.03 0.015 0.018 z5 0.045 0.046 0.041 0.043 0.041 0.043 z6 0.021 0.028 0.021 0.021 0.026 0.023 z7 0.021 0.019 0.021 0.021 0.019 0.02 z8 0.015 0.015 0.015 0.015 0.015 0.015 z9 0.015 0.015 0.015 0.015 0.015 0.015 z10 0.031 0.028 0.041 0.041 0.041 0.036 z11 0.031 0.028 0.028 0.028 0.026 0.028 z12 0.031 0.028 0.028 0.028 0.026 0.028 z13 0.076 0.09 0.089 0.049 0.122 0.085 z14 0.076 0.09 0.089 0.089 0.122 0.093 z15 0.207 0.184 0.209 0.302 0.122 0.205

[0161] The correlation coefficient of the weighted results of the five experts was further calculated. When there were five experts participating in the evaluation, the LPR was 0.88. Table 5 shows that all experts met the requirements. However, considering that the mean correlation coefficient between expert D and other experts was obviously small, the result of expert D was excluded, and the mean weight of the four experts A, B, C and E was used as the final weight, and w was obtained. ahp , see Table 4. Finally, the six items with relatively large weights, namely, total sugar, reducing sugar, nicotine, part value, elegance 1 and elegance 2, were selected as tobacco leaf quality indicators. At this time, the total weight = 0.773, q = 6.

[0162] Table 5: PR of five experts

[0163]

[0164]

[0165] Taking the quality of Liqun Long Mouth cigarettes in the past eight quarters as the target, the established near-infrared prediction model was used to calculate the six quality index values ​​of total sugar, reducing sugar, nicotine, part value, elegance 1 and elegance 2 of 20 tobacco leaves (M=20) in the tobacco leaf group formula table, and the weighted values ​​of the 20 tobacco leaves were calculated according to the proportion w (all 5%) of the Liqun Long Mouth cigarette leaf group, and the overall matrix U was obtained. 8×6 :

[0166]

[0167] Where: Q 8×6×k represents the values ​​of six quality indicators of concern for the kth leaf group of Liqun Long Mouth cigarettes in eight quarters; w i Indicates the usage ratio of the kth leaf group.

[0168] According to the five-step method in step 3, the similarity matrix of tobacco leaves in stock is calculated. The calculation process is as follows:

[0169] ① Take samples from multiple locations of the stock tobacco leaves and mix them as one sample, then measure the spectrum according to the method in step 1.

[0170] ②SNV method to process the new spectral data X after scatter correction ij :

[0171]

[0172] ③SSD method synchronously smooths and derives the new spectral data x ij :

[0173] x ij =0.087X ij-3 -0.266X ij-2 -0.230X ij-1 +0.230Xij+1 +0.266X ij+2 +0.087X ij+3

[0174] ④RSD method to select characteristic spectral points:

[0175]

[0176] Select RSD j The number of spectral points with ratio >0.05 (ratio2=0.05) is calculated to be 430 in total, that is, n=430 at this time.

[0177] ⑤ Similarity matrix of tobacco leaves in stock:

[0178] D shape (430,430)=|x 430 -y 430 |+min[D shape (429,430),D shape (430,429),D sshape (429,429)];

[0179]

[0180] Calculate the similarity of all the spectra of the stocked tobacco strips, and arrange the names of the tobacco strips from large to small according to the similarity, and obtain the tobacco strip similarity matrix, which is recorded as S 425×424 :

[0181]

[0182] The tobacco similarity matrix S obtained by the method of this application 425×424 Compared with several traditional near-infrared preprocessing and similarity calculation methods, the average relative deviations of the sensory quality indicators and chemical quality indicators of the 10 alternative tobacco strips closest to the target tobacco strips were statistically analyzed. The results are shown in Table 6.

[0183] Table 6: Comparison of results

[0184]

[0185] It can be seen from Table 6 that the relative deviations of the position, elegance 1, elegance 2, total sugar, reducing sugar and nicotine indicators of the pretreatment (SNV+SSD+FS) of the present application and the similarity method (2D-CORR) of the present application are relatively small, showing obvious advantages, and the combination of the two, namely SNV+SSD+FS+2D-CORR, has the most obvious advantage, indicating that the near-infrared spectral pretreatment and similarity evaluation method established in the present application is better.

[0186] According to S 425×424The top 10 most similar tobacco leaves of GZ3120CFL, HA4118CCH, HN1121CEA, SC4119CEA, SC4120XEY and YN8119CKH (s=6) were selected. The results are shown in Table 7.

[0187] Table 7: The top 10 most similar tobacco leaves for each of the six alternative tobacco leaves ①

[0188]

[0189]

[0190] Note: ①FJ Fujian, SD Shandong, GX Guangxi.

[0191] As shown in Table 7, although the similarity-based screening did not specify the consistency of the replacement tobacco leaves with the original tobacco leaves in terms of position and origin, the results still showed a certain similarity of origin and position, among which the average matching degree of origin reached 50% and the average matching degree of position reached 70%. This is because tobacco leaves are agricultural products, which are greatly affected by the ecological environment and soil conditions and show strong regional characteristics. However, similarity screening can help tobacco leaf formulators provide new ideas and use the matching schemes or origins that have never been used in the past in the existing formulas. For example, the origin matching degree of the three tobacco leaves GZ20CFL, SC19CEA and SC20XEY is not too high. Among them, GZ20CFL has 6 Sichuan tobacco leaves that are highly similar to it, SC19CEA has 7 Yunnan tobacco leaves that are highly similar to it, and SC20XEY has 3 Guizhou tobacco leaves and 3 Yunnan tobacco leaves that are highly similar to it. These new ideas have greatly expanded the flexibility of the formula and facilitated the better and more reasonable allocation of current tobacco leaf resources. On the other hand, the Yunnan, Guizhou and Sichuan regions are geographically adjacent to each other, and their climate environment and soil conditions are highly similar, which results in a high degree of similarity in their tobacco leaves.

[0192] The top 10 most similar tobacco leaves of GZ3120CFL, HA4118CCH, HN1121CEA, SC4119CEA, SC4120XEY, and YN8119CKH (s=6) each consist of 10 6 The established near infrared model was used to further obtain 10 6 The six quality indicators of concern, namely, the part of the new leaf group formula, elegant 1, elegant 2, total sugar, reducing sugar, and nicotine, were calculated, and the weighted values ​​of 20 leaf groups were obtained.

[0193]

[0194] Where: Indicates Liqun long-mouth cigarettes p sThe kth leaf group of a new recipe has the qth quality index value of interest; w i Indicates the usage ratio of the kth leaf group.

[0195] For the overall matrix U 8×6 Principal component analysis (PCA) is used for dimensionality reduction, and the first t principal components with contribution > 0.9 (ratio1 = 0.9) are selected. Figure 2 The contribution-principal component number results shown in the figure show that when t = 5, the contribution reaches 92%, and the overall dimension reduction matrix u is obtained 8×5 .

[0196] right Using the same t principal components for dimensionality reduction, we get 10 6 The dimension reduction vector v5. Secondary optimization, calculation 10 6 indivual to u 8×5 The Mahalanobis distance is:

[0197]

[0198] Matlab 2020b was used to write auxiliary calculation scripts according to the above steps to complete the calculation process of all processes. The leaf group formula corresponding to the minimum value is the final new tobacco leaf group formula. The formula maker can select the most satisfactory one from several new leaf group formulas with the smallest d value (such as the smallest 5 to 10 d values) based on actual conditions such as production planning, inventory arrangement and sensory evaluation.

[0199] As shown in Table 8, Among the top five best combinations, combination 1 (SC20CEE, HA18CCK, HN21XDP, YN19CKN, GZ20CFM, YN19CKJ) was the best replacement combination found, and none of the best combinations selected their most similar tobacco leaves, among which GZ20CFL, SC19CEA and SC20XEY did not select tobacco leaves from the same origin, and HN21CEA and SC20XEY did not select tobacco leaves from the same part.

[0200] Table 8: Top 5 best combinations

[0201] Target GZ20CFL HA18CCH HN21CEA SC19CEA SC20XEY YN19CKH Combination 1 SC20CEE HA18CCK HN21XDP YN19CKN GZ20CFM YN19CKJ Combination 2 SC20CEE HA18XCQ HN21CEC YN19CKN GZ20CFM YN19CKJ Combination 3 SC20CEE HA18CCK HN21CEC YN19CKN SC19CEK YN19CKJ Combination 4 SC20CEE HA18XCQ HN21CEC YN20CKL SC20CEK YN19CKJ Combination 5 SC20CEE HA20CCH HN21CEC YN20CKL GZ20CFM YN19CKJ

[0202] Table 9 shows the differences between the 16 quality index values ​​of the original formula and the quality indexes of the five new combinations and the original formula. The results show that the differences among the five combinations are small. From the total difference results, combination 1 and combination 2 are closest to the original formula.

[0203] Table 9: Differences in quality indicators between the five new combinations and the original formula

[0204]

[0205]

[0206] Table 10 further shows the results obtained through manual sensory evaluation, indicating that combination 1 and combination 2 are close to the original formula and meet the requirements of formula maintenance. The formulator can choose combination 1 or combination 2 as an alternative.

[0207] Table 10: Manual sensory evaluation results

[0208] sample Fragrance Aroma volume Irritant Aftertaste Agglomeration Miscellaneous gas Softness total Original recipe 7.3 7.6 7.4 7.3 7.5 7.9 7.5 52.5 Combination 1 7.4 7.7 7.4 7.3 7.6 7.8 7.5 52.7 Combination 2 7.5 7.6 7.3 7.5 7.6 7.7 7.6 52.8 Combination 3 7.5 7.5 7.3 7.2 7.3 7.5 7.3 51.6 Combination 4 7.3 7.3 7.2 7.4 7.5 7.5 7.3 51.5 Combination 5 7.1 7.3 7.2 7.1 7.5 7.3 7.3 50.8

[0209] Embodiment three:

[0210] This embodiment provides a computer device, including:

[0211] A memory for storing instructions;

[0212] The processor is used to execute the instructions so that the device performs operations to implement any one of the above-mentioned secondary similarity substitution methods for maintaining cigarette formula.

[0213] The method comprises:

[0214] Based on the near infrared spectrum-tobacco leaf quality index prediction model, the weighted values ​​of q quality indexes of cigarettes in the past several quarters are calculated to obtain the overall matrix;

[0215] According to the tobacco leaf similarity evaluation of the current tobacco leaves in stock, a similarity matrix of the tobacco leaves in stock is obtained;

[0216] By identifying s tobacco leaves to be replaced in a cigarette blend containing M tobacco leaf groups, p tobacco leaves that are most similar to the tobacco leaves to be replaced are selected from the similarity matrix to obtain p s A new leaf group formula of the cigarette;

[0217] Based on p s The M leaf group weighted values ​​of the q quality indicators of the new leaf group formula are obtained by s Weighted values;

[0218] For the overall matrix, p s The weighted values ​​are reduced by principal component analysis to obtain the overall dimension reduction matrix and p s A reduced dimensionality vector;

[0219] According to the calculation p s The Mahalanobis distance from the reduced dimension vector to the overall reduced dimension matrix is ​​obtained, and the new leaf group formula corresponding to the minimum Mahalanobis distance is the cigarette formula of similar substitution.

[0220] The present invention provides a two-level similarity substitution process method for cigarette formula maintenance. The first-level substitution algorithm first uses near-infrared spectroscopy for preliminary screening, and selects several tobacco leaves that are most similar to the tobacco leaves to be replaced as a whole to form a pre-selection pool. The second-level substitution algorithm uses Mahalanobis distance selection, and selects the tobacco leaves with the smallest Mahalanobis distance from the overall pool in the first-level substitution pre-selection pool through the experience of the formulator to specify several quality indicators. This method combines the advantages of both near-infrared analysis technology and cigarette formulators to create a new digital cigarette formula maintenance solution that combines man and machine. It provides an effective solution to the long-standing problems in the field of cigarette group formula, such as experience-based, low formula efficiency, difficult formula optimization, and tight raw material supply.

[0221] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A secondary similarity substitution method for cigarette formula maintenance, characterized in that: include: Based on the near infrared spectrum-tobacco leaf quality index prediction model, the weighted values ​​of q quality indexes of cigarettes in the past several quarters are calculated to obtain the overall matrix; According to the tobacco leaf similarity evaluation of the current tobacco leaves in stock, a similarity matrix of the tobacco leaves in stock is obtained; By identifying s tobacco leaves to be replaced in a cigarette blend containing M tobacco leaf groups, p tobacco leaves that are most similar to the tobacco leaves to be replaced are selected from the similarity matrix to obtain A new leaf group formula of the cigarette; based on The M leaf group weighted values ​​of the q quality indicators of the new leaf group formula are obtained weighted values; For the overall matrix, The weighted values ​​are reduced by principal component analysis to obtain the overall dimension reduction matrix and A reduced dimensionality vector; According to calculation The Mahalanobis distance from the reduced dimension vector to the overall reduced dimension matrix is ​​obtained, and the new leaf group formula corresponding to the minimum Mahalanobis distance is the cigarette formula of similar substitution.

2. The method according to claim 1, characterized in that Based on the tobacco leaf similarity evaluation of the current stock tobacco leaves, the similarity matrix of the stock tobacco leaves is obtained, including: Multi-point sampling to collect near-infrared spectral data of current tobacco leaves in stock; The near-infrared spectral data of the current stock tobacco leaves are subjected to scattering correction, synchronous smoothing derivation, feature selection and two-dimensional similarity calculation in sequence to obtain the similarity matrix of the stock tobacco leaves.

3. The method according to claim 2, characterized in that The synchronous smooth derivation comprises: The least squares fitting of the data in the moving window of the scattering-corrected spectral data is performed by polynomial. The fitting function is differentiated synchronously to obtain synchronous smoothed derivative spectral data.

4. The method according to claim 3, characterized in that The method further comprises: using cubic spline interpolation to supplement the head and tail of the spectral data after scattering correction with data of half the size of the window.

5. The method according to claim 2, characterized in that: Calculating two-dimensional similarity involves: Based on the shape differences and information content differences in the spectral data, the total similarity of the spectral data of any two tobacco leaves after feature selection is calculated; The calculation formula for the total similarity is: ， Where: S represents the total similarity of the spectrum, Represents shape similarity characterized by dynamic time warping distance; It represents the difference in the amount of information contained in two spectra characterized by information divergence; ， ， Where: j = 1, 2, 3,.., n, , are two different tobacco leaf sample data points after synchronous smoothing derivation of near-infrared spectra; n is the number of spectral data points; ; .

6. The method according to claim 1, characterized in that The q quality indicators include: If the cigarette is a single recipe: select q quality indicators based on the style characteristics of the cigarette brand and the actual recipe; If the cigarette is not a single recipe: according to the principle of feature selection, use relative standard deviation to select q quality indicators; or set at least five members to vote and select q quality indicators based on the total weight.

7. The method according to claim 6, characterized in that At least five members are set to vote and select q quality indicators based on the total weight, including: Each member makes a primary and secondary judgment on any two chemical indicators to obtain a comparison matrix; Convert the comparison matrix into a judgment matrix. The objective function is constructed by the least square method, the auxiliary function is constructed by the Lagrange conditional extreme value, and the weight is obtained by solving the judgment matrix; The obtained weights are averaged as the total weight, and q chemical indicators that are greater than the total weight are selected as quality indicators.

8. The method according to claim 7, characterized in that The method further comprises: Calculate the correlation coefficient of the weights; The quality index selected by eliminating members whose correlation coefficient is lower than the critical correlation coefficient at 95% confidence level.

9. A computer device, characterized in that: include: A memory for storing instructions; A processor is used to execute the instructions so that the device performs operations to implement a secondary similarity substitution method for cigarette recipe maintenance as described in any one of claims 1 to 8.

Images