Method for rapidly predicting slice proportion in leaf group formula module and storage medium
Through near-infrared spectral analysis technology, predict the chemical composition and proportional relationship of cigarette sheets, establish a prediction model, and realize automated monitoring of sheet content, solving the problem of inefficient traditional manual monitoring and improving the production efficiency and product quality stability.
Patent Information
- Application Number
- CN202510194376.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-21
- Publication Date
- 2025-05-23
AI Technical Summary
In the traditional cigarette silk production process, monitoring of the flake content depends on manual sampling and manual selection, which is inefficient, labor-intensive and susceptible to human factors, resulting in instability of the results.
Near-infrared spectroscopy analysis technology is used to predict various chemical components in the sheet, and the proportional relationships of sugar-base ratio, two sugar-base ratio, nitrogen-base ratio and potassium-chlorine ratio are calculated. Combined with these components and the sheet ratio, a prediction model is established to achieve automated monitoring of the sheet content.
It significantly improves production efficiency, reduces dependence on labor, ensures consistency and stability of product quality, and the prediction accuracy reaches 88% and the average accuracy is 89%.
Smart Images
Figure CN120028286A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of monitoring the content of thin slices in cigarette shreds, and in particular to a method and a storage medium for rapidly predicting the proportion of thin slices in a leaf group formula module. Background Art
[0002] As an important component of cigarettes, flakes play a key role in improving the quality of cigarettes, improving the taste characteristics and reducing the release of harmful substances. Flakes are usually made from tobacco raw materials through special processing. They can not only improve the filling performance of tobacco, but also optimize the smoke characteristics by adjusting its chemical composition to meet the needs of different consumers.
[0003] In the traditional tobacco production process, the monitoring of flake content mainly relies on manual sampling and hand selection. Specifically, the operator needs to randomly extract a certain amount of tobacco samples from the production line, then identify and separate the flakes and pure tobacco by naked eyes, and then calculate the proportion of flakes. Although this method is intuitive and reliable, it has obvious shortcomings, such as low efficiency, high labor intensity and susceptibility to human factors, resulting in unstable results. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for quickly predicting the proportion of flakes in a leaf group formula module. The present invention uses near-infrared spectroscopy analysis technology to predict multiple chemical components in flakes, calculates four proportion relationships such as sugar-alkali ratio, two sugar ratio, nitrogen-alkali ratio and potassium-chlorine ratio, and combines these components with the proportion of flakes to establish a prediction model to achieve automatic monitoring of the flake content, thereby significantly improving production efficiency, reducing dependence on manual labor, and ensuring the consistency and stability of product quality.
[0005] The technical solution of the present invention is:
[0006] A method for rapidly predicting the proportion of flakes in a leaf group recipe module, comprising:
[0007] (1) Sample processing
[0008] The mixed tobacco slices before shredding are manually selected into tobacco slices and thin slices, and the tobacco slices and thin slices are shredded using a sample shredder respectively.
[0009] The moisture of the tobacco slices, thin slices and cooled tobacco was balanced in a constant temperature and humidity chamber, and the moisture content after balance was 7-9%.
[0010] The dried samples were crushed using a cyclone mill with a particle size of ≤40 mesh.
[0011] (2) Preparation of standard samples
[0012] The tobacco powder and thin flake powder were mixed according to the experimental design ratio using a laboratory balance to prepare standard samples with different thin flake blending ratios.
[0013] (3) Chemical composition testing of standard samples
[0014] A near-infrared spectrometer was used to collect the sample powder spectrum, and the aged tobacco leaf model of the industry's "Tobacco Near-Infrared Analysis System Platform" was used to predict the chemical composition.
[0015] (4) Building a prediction model
[0016] The prediction model is constructed using the XGBoost (eXtreme Gradient Boosting) library.
[0017] The functions provided by the library are used to evaluate the important features of the model and to assess the prediction accuracy of the model through metrics such as Root Mean Square Error (RMSE), Precision, Recall, and F1-csore.
[0018] (6) Discriminant analysis model
[0019] The prediction was made using the three-category (0.0-4.9%, 5.0-8.0%, 8.1-11.0%) or four-category (0.0-2.0%, 2.1-5.4%, 5.5-8.0%, 8.1-11.0%) discriminant model classification method of flake content, and the XGBoost classification model was used for modeling and prediction.
[0020] (7) Use the same pre-processing method to process and scan the samples to be tested.
[0021] (8) The 17 chemical component characteristics of the sample to be tested are brought into the classification model for calculation.
[0022] (9) Realize fast prediction of flake ratio in leaf group formula module.
[0023] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for rapidly predicting the proportion of slices in a leaf group recipe module of the present invention.
[0024] Beneficial effects of the present invention:
[0025] The prediction accuracy of the invention is 88%, and the average accuracy of the model prediction is 89%, which can be used as an auxiliary means for monitoring the blending ratio of thin slices. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 : It is a schematic diagram of the process of the present invention.
[0027] Figure 2: Feature weight map of the discriminant model - three-classification model.
[0028] Figure 3 : Confusion matrix diagram of the discriminant model - three-classification model.
[0029] Figure 4 : Feature weight map of the discriminant model - four-classification model.
[0030] Figure 5 : Confusion matrix diagram of the discriminant model - four-classification model. DETAILED DESCRIPTION
[0031] Example 1
[0032] According to the leaf group formula design of a certain cigarette brand, the design proportion of thin slices is 6.5%, the production tolerance is ±1.5%, and the maximum proportion of thin slices in actual production sampling does not exceed 11%, so the design prediction model uses a three-classification model to meet the actual production requirements. The design classification model has category one of 0.0-4.9%, category two of 5.0-8.0%, and category three of 8.1-11.0%.
[0033] A number of standard samples with different thin slice ratios are designed for Category 1, Category 2 and Category 3 respectively. The design ratios of each sample are shown in Table 1.
[0034] Table 1 Proportion of thin slices in categories 1, 2 and 3
[0035]
[0036]
[0037]
[0038] Pure tobacco flakes and pure flakes are collected at the production site.
[0039] The sample shredder (QS-I type) was used for shredding, and a constant temperature and humidity chamber (brand: Shanghai Yiheng Scientific Instrument Co., Ltd. BPS-100CL, temperature parameter: 22°C, humidity parameter: 35%, balancing time: 14h) was used to balance the moisture of the tobacco slices and thin slices, and the moisture content after balancing was 7-9%. The tobacco slices and thin slices with balanced moisture content were crushed by cyclone mills, and the particle size was ≤40 mesh.
[0040] According to the slice ratio designed in Table 1, standard samples were prepared by manual weighing.
[0041] A near-infrared spectrometer (ThermoFisher AntarisⅡ) was used to collect and preprocess the spectra of standard samples.
[0042] The spectra of 54 standard samples were calculated using the tobacco leaf chemical composition prediction model, and 13 chemical components were predicted (total alkaloids %, total nitrogen %, total sugars %, reducing sugars %, potassium %, chlorine %, starch %, neophytodienes mg / g, polyphenols mg / g, polyacids mg / g, higher fatty acids mg / g, amino acids mg / g, Amadori compounds mg / g), and 4 proportional relationships were calculated (sugar-alkali ratio, two sugars ratio, nitrogen-alkali ratio, potassium-chlorine ratio).
[0043] An XGBoost classification model based on 17 chemical component features was established.
[0044] Check the influence of each component (feature) on the result (see Figure 2 ), the top four in importance are nitrogen-base ratio, starch, Amadori compounds and higher fatty acids.
[0045] Verification of the model: Use the remaining tobacco powder and thin flake powder to configure 22 standard samples of category one, 19 standard samples of category two, and 9 standard samples of category three. After predicting 17 chemical components, bring them into the established XGBoost model to predict their classification.
[0046] The model is evaluated by the statistical prediction results, and the statistical table is shown in Table 2. From the perspective of precision, that is, from the perspective of thin film abnormality processing, the precision of the abnormal interval (0.0-4.9% and 8.1-11.0%) is greater than 0.90, that is, if the detection shows that the thin film ratio is abnormal, there is a probability of more than 90% that it is abnormal. From the perspective of recall, that is, from the perspective of quality monitoring, the recall rates of the abnormal interval (0.0-4.9% and 8.1-11.0%) are 0.91 and 0.78 respectively, that is, in the production process, 91% of the abnormalities in the 0.0-4.9% interval will be detected and 78% of the abnormalities in the 8.1-11.0% interval will be detected.
[0047] Table 2. Evaluation results of the three-classification test set model
[0048] Classification Accuracy Recall F1-csore Number of samples 0.0-4.9％ 0.91 0.91 0.91 22 5.0-8.0％ 0.81 0.89 0.85 19 8.1—11.0％ 1.00 0.78 0.88 9 Accuracy 0.88 50 Average value (comprehensive) 0.89 0.88 0.88 50
[0049] From Figure 3 In the confusion matrix diagram with the X-axis as the true category and the Y-axis as the predicted category, it can be found that the prediction effect is good, and the prediction error will only deviate to the adjacent category. For example, there are 22 samples with an actual value of 0.0-4.9%, 20 of which are predicted correctly and 2 are predicted incorrectly. The 2 wrong samples are both predicted to be 5.0-8.0%, which are both categories adjacent to the correct classification.
[0050] Example 2
[0051] According to the leaf group formula design of a certain cigarette brand, the design proportion of thin slices is 5.1%, the production tolerance is ±3%, and the positive and negative deviations between the actual proportion and the design proportion need to be detected. The maximum proportion of thin slices in the actual production sampling is no more than 11%, so the design prediction model uses a four-classification model to meet the actual production requirements. The design classification model has category one of 0.0-2.0%, category two of 2.1-5.4%, category three of 5.5-8.0%, and category four of 8.1-11.0%.
[0052] A plurality of standard samples with different thin slice ratios are designed for category 1, category 2, category 3 and category 4 respectively.
[0053] Pure tobacco flakes and pure flakes are collected at the production site.
[0054] The sample shredder (QS-I type) was used for shredding, and a constant temperature and humidity chamber (brand: Shanghai Yiheng Scientific Instrument Co., Ltd. BPS-100CL, temperature parameter: 22°C, humidity parameter: 35%, balancing time: 14h) was used to balance the moisture of the tobacco slices and thin slices, and the moisture content after balancing was 7-9%. The tobacco slices and thin slices with balanced moisture content were crushed by cyclone mills, and the particle size was ≤40 mesh.
[0055] According to the designed slice ratio, standard samples are prepared by manual weighing.
[0056] A near-infrared spectrometer (ThermoFisher AntarisⅡ) was used to collect and preprocess the spectra of standard samples.
[0057] The spectra of 54 standard samples were calculated using the tobacco leaf chemical composition prediction model, and 13 chemical components were predicted (total alkaloids %, total nitrogen %, total sugars %, reducing sugars %, potassium %, chlorine %, starch %, neophytodienes mg / g, polyphenols mg / g, polyacids mg / g, higher fatty acids mg / g, amino acids mg / g, Amadori compounds mg / g), and 4 proportional relationships were calculated (sugar-alkali ratio, two sugars ratio, nitrogen-alkali ratio, potassium-chlorine ratio).
[0058] An XGBoost classification model based on 17 chemical component features was established.
[0059] Check the influence of each component (feature) on the result (see Figure 4 ), the top four in importance are starch, higher fatty acids, potassium-chlorine ratio, polyphenols and total sugars.
[0060] Verification of the model: Use the remaining tobacco powder and thin flake powder to configure 17 standard samples of category one, 8 standard samples of category two, 16 standard samples of category three, and 9 standard samples of category four. After predicting 17 chemical components, bring them into the established XGBoost model to predict their classification.
[0061] The model is evaluated by the statistical prediction results, and the statistical table is shown in Table 3. From the perspective of precision, that is, from the perspective of thin film abnormality processing, the precision of the abnormal interval (0.0-2.0% and 8.1-11.0%) is greater than 0.90, that is, if the detection shows that the thin film ratio is abnormal, there is a probability of more than 90% that it is abnormal. From the perspective of recall, that is, from the perspective of quality monitoring, the recall rates of the abnormal interval (0.0-2.0% and 8.1-11.0%) are 0.82 and 0.78 respectively, that is, in the production process, 82% of the abnormalities in the 0.0-2.0% interval will be detected and 78% of the abnormalities in the 8.1-11.0% interval will be detected.
[0062] Table 3. Four-classification test set model evaluation results
[0063] Classification Accuracy Recall F1-csore Number of samples 0.0-2.0％ 0.93 0.82 0.87 17 2.1-5.4％ 0.55 0.75 0.63 8 5.5—8.0％ 0.82 0.88 0.85 16 8.1-11.0％ 1.00 0.78 0.88 9 Accuracy 0.82 50 Average value (comprehensive) 0.85 0.82 0.83 50
[0064] From Figure 4 In the confusion matrix diagram with the X-axis as the true category and the Y-axis as the predicted category, it can be found that the prediction effect is good, and the prediction error will only deviate to the adjacent category. For example, there are 8 samples with an actual value of 2.1-5.4%, 6 of which are predicted correctly and 2 are predicted incorrectly. Among the 2 wrong samples, one is predicted to be 0-2.0% class and the other is predicted to be 5.5-8.0% class, both of which are classes adjacent to the correct classification.
[0065] In both Example 1 and Example 2, a discrimination model was designed based on actual production, and 17 chemical components were used to predict the flake ratio, with good results, meeting the production detection requirements.
Claims
1. A method for rapidly predicting the proportion of flakes in a leaf group formula module, characterized in that: The following steps are involved: (1) Sample processing The mixed tobacco slices before shredding are manually selected into tobacco slices and thin slices, and the tobacco slices and thin slices are shredded by using a sample shredder respectively; Use a constant temperature and humidity chamber to balance the moisture of tobacco slices, sheets and cooled tobacco; The dried samples were pulverized using a cyclone mill; (2) Preparation of standard samples Use a laboratory balance to mix tobacco powder and thin sheet powder according to the experimental design ratio to prepare standard samples with different thin sheet blending ratios; (3) Chemical composition testing of standard samples Use a near-infrared spectrometer to collect the spectrum of sample powder, and use the alcoholized tobacco leaf model of the industry's "Tobacco Near-Infrared Analysis System Platform" to predict the chemical composition; (4) Building a prediction model Use XGBoost library to build prediction models; Use the functions provided by the library to evaluate the important features of the model and evaluate the prediction accuracy of the model through the root mean square error, precision, recall and F1-csore indicators; (5) Discriminant analysis model Use the three-class or four-class discriminant model classification method to perform modeling and prediction through the XGBoost classification model; (6) using the same pretreatment method to process and scan the sample to be tested; (7) The 17 chemical component characteristics of the sample to be tested are brought into the classification model for calculation; (8) Realize fast prediction of flake ratio in leaf group formula module.
2. The method according to claim 1, characterized in that In step (5), the steps of distinguishing the thin slices from the conventional tobacco leaves include: ① The total alkaloid value of conventional middle tobacco is generally 2.0-3.0, and the design value of thin slices is generally 0.9-1.
3. Adding thin slices will reduce the total alkaloid value of the entire module and affect the nitrogen-alkali ratio; ② The potassium value of conventional middle smoke is generally 1.0-2.2, and the design value of thin slices is generally 2.4-3.
4. Adding thin slices will increase the potassium value of the entire module.
3. The method according to claim 1, characterized in that In step (1): The constant temperature and humidity chamber used includes temperature parameters: 22° C., humidity parameters: 35%, and equilibrium time: 14 h.
4. The method according to claim 1, characterized in that In step (1): The moisture content of tobacco sheets, flakes and cooled cut tobacco after moisture balance is 7-9%; The sample particle size is ≤40 mesh.
5. The method according to claim 1, characterized in that In step (5), the step of importing data into the XGBOOST regression model includes: 20% of the 240 data were randomly selected as the validation set and 80% were randomly selected as the training set for model training.
6. The method according to claim 1, characterized in that In step (5), the importance of each component feature includes: The importance from high to low is total alkaloids, potassium, reducing sugars, nitrogen-base ratio and amino acids.
7. The method according to claim 1, characterized in that In step (6), the three-class discrimination model classification includes: Classification is performed using flake contents of 0.0-4.9%, 5.0-8.0%, and 8.1-11.0%.
8. The method according to claim 1, characterized in that In step (6), the four-category discrimination model classification includes: Classification is performed using flake contents of 0.0-2.0%, 2.1-5.4%, 5.5-8.0%, and 8.1-11.0%.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of a method for quickly predicting the proportion of thin slices in a leaf group formula module as described in any one of claims 1 to 8 are implemented.