Model for representing smoke CO release amount based on cigarette microstructure and application thereof

Abstract

The invention provides a model for representing smoke CO release amount based on a cigarette microstructure and application of the model. According to the model, a 3D-mu CT scanning technology is adopted to construct a high-resolution fault sequence image in a cigarette sample, a cigarette three-dimensional structure model is established according to the fault sequence image, cigarette structure characteristic parameters are obtained, a cigarette three-dimensional pore throat ball-stick model is established according to the cigarette structure characteristic parameters, and the cigarette three-dimensional pore throat ball-stick model is established according to the cigarette structure characteristic parameters. And finally, regression processing is carried out on the model and the characteristic parameters, then a BP neural network model is established, neural network model training and optimization are carried out according to a CO release amount data set of the cigarette, and the method is obtained. The model has the advantages of being high in precision, high in speed, small in needed calculation force and the like, tests show that after the model is trained for 1000 times, the average absolute error of a test set is 0.0016, the average deviation of a training set is-0.0002, the average deviation of the test set is-0.0003, and the application requirement for adjusting the cigarette microstructure and quantitatively controlling the smoke CO release amount can be met.

Description

Technical Field

[0001] The present invention designs a model for the CO release amount of cigarette smoke, specifically relates to a model for characterizing the CO release amount of cigarette smoke based on the microscopic structure of cigarette sticks and its application, and belongs to the technical field of artificial intelligence tobacco design. Background Art

[0002] With the upsurge of the global anti-smoking movement, the impact of smoking on health has become one of the main challenges faced by the tobacco industry. How to reduce the release of harmful components such as CO in cigarette smoke while meeting the taste and physiological needs of consumers has become a difficult problem that urgently needs to be solved in the cigarette production and design industry. Cigarette smoke is a key factor affecting cigarette quality and harmfulness, and as the main part of cigarette combustion, the physical parameters of the microscopic structure of cut tobacco (such as pores, specific surface area, etc.) significantly affect the generation and transfer characteristics of harmful chemical components (mainly CO, etc.) in cigarette smoke.

[0003] Therefore, in-depth study of the relationship between the microscopic structure characteristics of cigarette sticks and the CO release amount of cigarette smoke is particularly important for improving cigarette structure and optimizing cigarette processing technology to achieve tar reduction and harm reduction. Summary of the Invention

[0004] Aiming at the problems existing in the prior art, the first object of the present invention is to provide a model for characterizing the CO release amount of cigarette smoke based on the microscopic structure of cigarette sticks. This model uses spiral CT tomographic image reconstruction technology to obtain the microscopic structure characteristic parameters of cigarette sticks, and after regression processing, combines with the BP neural network model to accurately describe the CO release amount of cigarette smoke, so as to achieve quantitative regulation through the CO release amount. This model has the advantages of high accuracy, fast speed and small required computing power, and has important guiding significance for improving cigarette design.

[0005] The second object of the present invention is to provide an application of a model for characterizing the CO release amount of cigarette smoke based on the microscopic structure of cigarette sticks, which is used to adjust the microscopic structure of cigarette sticks to quantitatively control the CO release amount of cigarette smoke. Based on the model provided by the present invention, quantitative controllable adjustment of the CO release amount of cigarette smoke by the microscopic structure of cigarette sticks is realized, and the technical purpose of tar reduction and harm reduction is achieved by changing the cigarette stick structure. After testing, the average absolute error of the test set is 0.0016, the average deviation of the training set is -0.0002, and the average deviation of the test set is -0.0003 after 1000 times of training of the model provided by the present invention.

[0006] To achieve the above technical object, the present invention provides a model for characterizing the CO release amount of cigarette smoke based on the microscopic structure of cigarette sticks, including:

[0007] S1. Using 3D-μCT scanning technology to construct a high-resolution tomographic sequence image inside the cigarette stick sample;

[0008] S2. Establish a three-dimensional structure model of the cigarette rod based on the tomographic sequence images, and obtain the characteristic parameters of the cigarette rod structure;

[0009] S3. Establish a three-dimensional structure model of the cigarette rod based on the tomographic sequence images, and establish a three-dimensional pore-throat ball-stick model of the cigarette rod according to the characteristic parameters of the cigarette rod structure;

[0010] S4. After performing regression processing on the models and characteristic parameters obtained in S1 to S3, establish a BP neural network model, and perform neural network model training and optimization according to the CO release amount data set of the cigarette rod, and that's it;

[0011] The characteristic structure parameters of the cigarette rod include: porosity, specific surface area, fractal dimension, and permeability.

[0012] As a preferred solution, the construction process of the high-resolution tomographic sequence images in S1 is as follows: Using a 3D-μCT system, taking the central rotation axis of the spiral trajectory as the z-axis, and taking the intersection point of the z-axis and the horizontal plane where the starting point of the spiral scan trajectory line is located as the origin, establish a Cartesian coordinate system o-xyz. At this time, the coordinates of the voxels to be reconstructed in the area to be reconstructed inside the spiral trajectory line are expressed as x = [x 0 , x 1 , x 2 , the distance from the X-ray source focus to the detector plane is D, the coordinates of the X-ray source focus are λ(s), and the rotation radius of the spiral scan trajectory is R. The calculation process of the spiral scan trajectory line C is as follows:

[0013] Equation 1:

[0014] As a preferred solution, the process of obtaining the porosity in S2 is as follows:

[0015] S2-1-1. According to the gray histogram of the tomographic sequence images, count the frequencies of each gray value in the gray levels of the images to obtain the occurrence probabilities p i and occurrence frequencies q i values, where q i = i × p i , i ∈ [0, 65535];

[0016] S2-1-2. Let the segmentation threshold be T. The proportions of the number of background pixels and foreground pixels in the tomographic sequence images in all the pixels of the image are respectively and ω 1 = 1 - ω 0 , and the average gray values of the background and foreground are respectively According to g = ω 0 (μ 0 - μ) 2 + ω 1 (μ 1-μ) 2 Calculate the between-class variance value g;

[0017] The S2-1-3 threshold T traverses the gray levels of all pixels in the image from 0 to 65535, and calculates the threshold T with the maximum between-class variance max , which is the required threshold;

[0018] S2-1-4 Determine whether the pixel points in the image belong to pores according to the obtained threshold, and calculate the ratio of the number of pixels belonging to pores to the total number of pixels, which is the porosity.

[0019] As a preferred solution, the process for obtaining the specific surface area in S2 is as follows:

[0020] S2-2-1 Use the Canny edge detection algorithm to obtain the edges of the extracted tomographic sequence images;

[0021] S2-2-2 Connect the pixel points considered to be edges through an edge connection method to form a closed edge path;

[0022] S2-2-3 Statistically calculate the percentage of the number of pixels belonging to the edges in the tomographic sequence image to the total number of pixels, which is the specific surface area of this layer of cut tobacco. The calculation process is as follows:

[0023] Equation 1:

[0024] In Equation 1: k is the specific surface area, M′ is the number of edge pixels in the cigarette image, and M is the total number of pixels in the cigarette image.

[0025] As a preferred solution, the calculation process of the fractal dimension in S2 is as follows:

[0026] S2-3-1 Divide the plane of the tomographic sequence image into grids of size R×R and arrange them as a surface with a length of R, a width of R, and a height of L in three-dimensional space. Also perform the same division on the coordinate axis corresponding to L, with the unit of R×L / M;

[0027] S2-3-2 In the R×R grid, find the maximum pixel value u and the minimum pixel value b, and calculate the number of boxes required to cover the entire area from the minimum value to the maximum value. The calculation process is n(i,j) = [(u - b + R - 1) / R]. Sum all n to get N, which is the number of boxes;

[0028] The calculation process of the fractal dimension F is as follows:

[0029] Equation 2: F = -logN / logR.

[0030] As a preferred solution, the calculation process of the permeability in S2 is as follows: In the three-dimensional structure model of the cigarette rod, simulate the pressure difference, state, and flow velocity of the fluid when flowing through the pores to obtain the statistical result of the cigarette rod permeability:

[0031] Equation 3:

[0032] In Equation 3: K represents the permeability, Q is the flow rate of the fluid through the porous medium, μ is the viscosity of the fluid, L represents the length of the fluid flow, A represents the cross-sectional area of the fluid flowing through the porous medium, t represents the flow-through time, and Δp is the pressure difference at both ends.

[0033] As a preferred solution, the three-dimensional pore-throat ball-stick model of the cigarette rod in S3 includes the following three characteristic indexes: pore-throat ratio h, average coordination number p, and three-dimensional shape factor G xyz , and the acquisition process is as follows:

[0034] In S3-1, the pore-throat ratio h is the ratio of the average pore diameter to the average throat diameter in the pore-throat ball-stick model, and its calculation process is as follows:

[0035] Equation 4:

[0036] In S3-2, the average coordination number p is the average value of the number of throats connected to a single pore in the pore-throat ball-stick model, and its calculation process is as follows:

[0037] Equation 5:

[0038] In S3-3, the three-dimensional shape factor G xyz is the degree of irregularity of the image shape, and its calculation process is as follows:

[0039] Equation 6:

[0040] In Equations 4 to 6: d t is the average diameter of the throats in the model; d p is the average diameter of the pores in the model; n t is the number of throats in the model; n p is the number of pores in the model; L is the average throat length; r is the average throat radius.

[0041] As a preferred solution, the process of the regression processing in S4 is as follows:

[0042] In S4-1-1, use Lasso regression to screen and compress all models and characteristic parameters in S1 to S3, and perform standardization transformation to make the mean equal to 0 and the variance equal to 1. At this time, the Lasso regression model is estimated:

[0043] Equation 7:

[0044] In Equation 7: For any harmonic parameter s, the estimation of α is The overall regression coefficient can be reduced by adjusting the harmonic parameter s. When s 0 = Σ j |β j |, s ≤ s 0 When this is the case, by screening out irrelevant variables and variables with extremely weak relationships, the accuracy of the regression model is improved;

[0045] In S4-1-2, the regression coefficients in the regression model are solved using the Mallows Cp statistic, and its calculation process is as follows:

[0046] Equation 8:

[0047] Equation 9:

[0048] In Equations 8 and 9: n represents the total number of sampling samples, p represents the number of independent variables in the subset regression model, δ 2 represents the mean variance during regression, and SSE p represents the sum of squared residuals after regression; when C p reaches the minimum value, the subset of variables taken by the model within the global scope is optimal, and the generated regression model has the best effect.

[0049] As a preferred solution, the establishment process of the BP neural network model in S4 is as follows: The establishment process of the BP neural network model in S4 is as follows: The characteristic parameters obtained after Lasso regression screening are used as the input layer of the BP neural network, and the BP neural network model is established through the newff function.

[0050] As a preferred solution, the number of hidden layers of the BP neural network model ≥ 5, the number of training times ≥ 1000, the minimum error ≤ 0.0001, and the learning rate is 0.01 - 0.1

[0051] The key of the model provided by the present invention lies in the combination of Lasso regression processing and the BP neural network to form a Lasso - BP neural network model. This model can screen a large number of parameters related to the model according to their influence relationships on the output results, thereby removing a large number of irrelevant or weakly related redundant parameters, reducing the input amount, and thus significantly reducing the calculation amount on the premise of ensuring the accuracy of the output results.

[0052] The present invention also provides an application of a model for characterizing the CO release amount of cigarette smoke based on the microstructure of cigarette rods, which is used to adjust the microstructure of cigarette rods to quantitatively control the CO release amount of cigarette smoke.

[0053] Compared with the prior art, the beneficial technical effects of the present invention are:

[0054] 1) The model provided by the present invention utilizes spiral CT tomographic image reconstruction technology to obtain the characteristic parameters of the microscopic structure of cigarette rods. After regression processing and combined with the BP neural network model, it accurately describes the CO release amount of cigarette smoke, thereby realizing the quantitative regulation of the CO release amount. This model has the advantages of high accuracy, fast speed, and small required computing power, and has important guiding significance for improving cigarette design.

[0055] 2) In the technical solution provided by the present invention, based on the model provided by the present invention, the quantitative controllable adjustment of the microscopic structure of cigarette rods on the CO release amount of cigarette smoke is realized, and the technical purpose of reducing tar and harm by changing the structure of cigarette rods is achieved. After testing, the average absolute error of the test set is 0.0016, the average deviation of the training set is -0.0002, and the average deviation of the test set is -0.0003 after the model provided by the present invention is trained 1000 times. Description of the Drawings

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0057] Figure 1 Schematic diagram of the overall structure of the CT acquisition and reconstruction experimental platform provided by the embodiment of the present invention;

[0058] Figure 2 Schematic diagram of the spatial coordinate system of the spiral scan trajectory CT system provided by the embodiment of the present invention;

[0059] Figure 3 Schematic diagram of the porosity statistics example of the cigarette rod sample provided by the embodiment of the present invention;

[0060] Figure 4 Schematic diagram of the specific surface area statistics process of the cigarette rod sample provided by the embodiment of the present invention;

[0061] Figure 5 Schematic diagram of the fractal dimension calculation process of the cigarette rod sample provided by the embodiment of the present invention;

[0062] Figure 6 Schematic diagram of the permeability calculation process of the cigarette rod provided by the embodiment of the present invention;

[0063] Figure 7 Schematic diagram of the generation process of the three-dimensional pore throat ball-and-stick model of the cigarette rod provided by the embodiment of the present invention;

[0064] Figure 8Schematic diagram of the Lasso-BP neural network structure provided by the embodiments of the present invention;

[0065] Figure 9 Schematic diagram of the prediction results of the training set of the Lasso-BP neural network model provided by the embodiments of the present invention;

[0066] Figure 10 Schematic diagram of the prediction results of the test set of the Lasso-BP neural network model provided by the embodiments of the present invention;

[0067] Figure 11 Schematic diagram of the regression line of the predicted value and the true value provided by the embodiments of the present invention. Detailed implementation manners

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0069] Embodiment 1

[0070] An embodiment of the present invention provides a model for characterizing the CO release amount of flue gas based on the microscopic structure of cigarette rods, including the following steps:

[0071] S1. Use 3D-μCT scanning technology to construct high-resolution tomographic sequence images inside the cigarette rod sample;

[0072] S2. Establish a characterization method for the porous structure of cigarette rods with parameters such as porosity, specific surface area, fractal dimension, and permeability for the tomographic sequence images;

[0073] S3. Use a three-dimensional pore-throat ball-and-stick model to visually characterize the pores and their connectivity of the cigarette rod;

[0074] S4. Establish a data set of the microscopic characteristics of the cigarette rod and the CO release amount of flue gas, construct a regression model based on the Lasso-BP neural network, and complete the effective regulation of CO in the flue gas.

[0075] In this embodiment, step S1 specifically includes:

[0076] Complete the global spiral scanning trajectory and high-resolution tomographic imaging of the cigarette rod sample to obtain the high-resolution tomographic sequence images of the cigarette rod sample.

[0077] Specifically, refer to Figure 1, the overall structure of the hardware system of the CT acquisition and reconstruction experimental platform mainly consists of an X-ray source, an area array detector, a motion control platform, and a host computer system.

[0078] Using the described CT system, refer to Figure 2 , taking the central rotation axis of the helical trajectory as the z-axis and the intersection point of the z-axis and the horizontal plane where the starting point of the helical scan trajectory line is located as the origin, a Cartesian coordinate system o-xyz is established. At this time, the coordinates of the voxels to be reconstructed in the area to be reconstructed inside the helical trajectory line are expressed as x = [x 0 , x 1 , x 2 , the distance from the X-ray source focus to the detector plane is D, the coordinates of the X-ray source focus are λ(s), and the rotation radius of the helical scan trajectory is R. Then the helical scan trajectory line C can be expressed.

[0079]

[0080] In this embodiment, the process of extracting the microscopic characteristics of the cigarette rod in step S2 is as follows:

[0081] A method for characterizing the porous structure of the cigarette rod represented by parameters such as porosity, specific surface area, fractal dimension, and permeability is established. A three-dimensional pore-throat ball-and-stick model is used to visually characterize the pores and their connectivity of the cigarette rod, and quantitative evaluation indexes are designed according to the model characteristics, including pore-throat ratio, average coordination number, three-dimensional shape factor, etc.

[0082] In this embodiment, refer to Figure 3 the porosity statistical example of the cigarette rod sample, and the process of extracting the porosity characteristics of the cigarette rod is as follows:

[0083] According to the gray histogram of the image, the frequency of each gray value in the gray level of the image is counted, so as to obtain the occurrence probability p i and the occurrence frequency q i values, where q i = i × p i , i ∈ [0, 65535];

[0084] Let the segmentation threshold be T, and the proportions of the number of background pixels and foreground pixels in the total number of pixels in the image are respectively and ω 1 = 1 - ω 0 , and the average gray values of the background and foreground are respectively According to g = ω 0 (μ 0 - μ) 2 + ω 1 (μ 1 - μ) 2 Calculate the between-class variance value g;

[0085] The threshold T traverses the gray levels of all pixels in the image from 0 to 65535, and calculates the threshold T with the largest between-class variance. max , which is the required threshold;

[0086] According to the threshold, it is determined whether the pixel points in the cigarette image belong to pores. The porosity is obtained by calculating the ratio of the number of pixels belonging to pores to the total number of pixels in the object area.

[0087] In this embodiment, referring to Figure 4 , the process of extracting the specific surface area feature of the cigarette is as follows:

[0088] To obtain the specific surface area of the cigarette, first, a preprocessing method similar to the porosity measurement is adopted to obtain the extracted object area of the cigarette;

[0089] The Canny edge detection algorithm is used to obtain the edge of the extracted cigarette tomographic image;

[0090] The pixel points considered to be edges are connected through an edge connection method to form a closed edge path;

[0091] The percentage of the number of pixels belonging to the edge in the cigarette area to the total number of pixels is the specific surface area of this layer of cut tobacco.

[0092] In this embodiment, referring to Figure 5 , the process of extracting the fractal dimension feature of the cigarette is as follows:

[0093] By calculating the fractal dimension, the pore structure complexity and filling performance of porous materials can be effectively evaluated, and the irregular distribution characteristics of the pore space can be comprehensively described.

[0094] First, the image plane is divided into grids of size R×R, and the corresponding coordinate axes of L are also divided in the same way, with the unit being R×L / M;

[0095] Within each R×R grid, find the maximum and minimum pixel values, and calculate how many boxes are needed to cover the entire area from the minimum value to the maximum value;

[0096] Assume that the current is the (i,j)th grid, and the number of boxes is denoted as n(i,j), that is, n(i,j) = [(u - b + R - I) / R]. The sum of the number of boxes for each R×R is obtained as N;

[0097] Theoretically, the fractal dimension F = -logN / logR. In fact, since R is a finite value, by changing the value of R, a set of N is obtained and the slope of the straight line is obtained by linear fitting, which is the fractal dimension F.

[0098] In this embodiment, referring to Figure 6, the process of extracting the cigarette penetration rate characteristics is as follows:

[0099] The penetration rate of a porous structure refers to the rate at which a fluid flows through a porous medium. It is a physical property of the medium and is related to factors such as the porosity, pore distribution, pore shape, and pore connection of the medium.

[0100] Use the tomographic image data to establish the three-dimensional volume data space and mask of the sample;

[0101] By constructing a pore network model, simulate the pressure difference, state, and flow rate of the fluid when flowing through the pores to obtain the statistical results of the cigarette penetration rate.

[0102] In this embodiment, the process of extracting the characteristics of the pore-throat stick model of the cigarette in step S3 is as follows:

[0103] The generation process of the three-dimensional pore-throat stick model of the cigarette is as Figure 7 shown.

[0104] The pore-throat stick model can simulate the movement of the fluid in the pore system, display the three-dimensional spatial distribution of the pores in the porous material and the connection between the pores, and facilitate understanding the texture and performance of the porous medium.

[0105] Define the three indicators to describe the pore-throat model of the cigarette sample.

[0106] Pore-throat ratio h: The ratio of the average pore diameter to the average throat diameter in the pore-throat stick model.

[0107]

[0108] Among them, d t is the average diameter of the throat in the model, and d p is the average diameter of the pores in the model.

[0109] Average coordination number p: The average value of the number of throats connected to a single pore in the pore-throat stick model.

[0110]

[0111] Among them, n t is the number of throats in the model, and n p is the number of pores in the model.

[0112] Three-dimensional shape factor G xyz : Quantitatively characterize the irregularity of the image shape.

[0113] In order to consider the three-dimensional structural characteristics of the throat, the three-dimensional shape factor G xyz is defined.

[0114]

[0115] Among them, L is the average throat length and r is the average throat radius.

[0116] Furthermore, referring to Figure 8 , step S4 specifically includes:

[0117] Use Lasso regression to screen and compress the initial variables. To eliminate the influence of variables with different dimensions on the results, in this model, all independent variables X i =(x i1 , x i2 , …, x im ) are subjected to standardization transformation, and let z i1 , z i2 ,..., z im have a mean equal to 0 and a variance equal to 1. At this time, the Lasso regression model is estimated:

[0118]

[0119] Among them, for any harmonic parameter s, the estimate of α is The overall regression coefficient can be reduced by adjusting the harmonic parameter s. When s 0 =∑ j |β j |, s ≤ s 0 , by screening out irrelevant variables and variables with extremely small relationships, the accuracy and scientificity of the model can be improved.

[0120] The regression coefficients in the regression model are solved using the Mallows Cp statistic:

[0121]

[0122]

[0123] In the formula, n represents the total number of sampling samples, p represents the number of independent variables in the subset regression model, δ 2 represents the mean variance during regression, and SSE p represents the sum of squared residuals after regression. When C p takes the minimum value, the subset of variables taken by the model within the global scope is the optimal one, and the generated regression model has the best effect.

[0124] Combine the Lasso regression and the BP neural network model;

[0125] Then, use the BP neural network model to train the significant factors, so as to obtain the relationship model between the microscopic structure characteristic parameters of the cigarette rod and the CO release amount in the flue gas.

[0126] Specifically, the establishment and performance testing of the Lasso-BP neural network model are implemented through the following steps:

[0127] Step 1: Extraction of cigarette microstructure characteristics

[0128] Five cigarette sticks (White Furongwang hard box) of the same kind of cut tobacco but with different structures were selected to exclude the interference of material basis differences on the regression model. The cigarette stick samples were numbered 1# to 5# respectively. The statistical data of the microstructure characteristics of each cigarette stick were used as independent variables, including the mean (A1), variance (A2), kurtosis (A3), skewness (A4) of the porosity curve, the specific surface area (A5), fractal dimension (A6), permeability (A7), pore throat ratio (A8), average coordination number (A9), and three-dimensional shape factor (A10) of the cigarette stick sample. The release amount of CO, the most important harmful substance in cigarette smoke, was selected as the dependent variable. The unit of the dependent variable is the CO release amount per unit weight of cut tobacco, that is, how many milligrams of CO are released per gram of cut tobacco. The CO release amount data are shown in Table 1.

[0129] Table 1

[0130]

[0131] Microscopic feature extraction was performed on the 5 cigarette sticks, and the statistical results are shown in Table 2.

[0132] Table 2

[0133]

[0134]

[0135] Step 2: Experiment on the influence of cigarette microstructure characteristics on CO release amount

[0136] 110 experimental samples were used as sample data for model establishment. The samples were randomly divided into two groups according to the ratio of 70% and 30%. The first group of training set contained 77 sample data, which were used for model establishment and optimization of related parameters. The second group of test set included 33 sample data, which were used to test the model prediction results and related evaluation indicators. The quantitative evaluation of the model was carried out by calculating the determination coefficient R 2 and root mean square error RMSE, and the accuracy of the model was tested by calculating indicators such as mean absolute error MAE and mean bias MBE.

[0137] A regression model based on Lasso regression and BP neural network was established. First, Lasso regression was used to screen the sample feature quantities, and the screening process was realized by the Lasso function in MATLAB software.

[0138] To improve the stability of the model, different groups were repeatedly set, and after 10-fold CV cross-validation, appropriate values were substituted into the Lasso model to fit the coefficients of each variable in the Lasso regression model. As the penalty strength increased, the variable coefficients continuously decayed until they reached 0. As shown in Table 3, the coefficients of variables A5, A6, and A9 had decayed to 0. These variables with less influence on the model were excluded, and 7 variables with significant influence in the model were selected, namely the mean (A1), variance (A2), kurtosis (A3), skewness (A4) of the porosity curve, the permeability (A7) of the sample, the pore throat ratio (A8), and the three-dimensional shape factor (A10).

[0139] Table 3

[0140]

[0141] The parameters screened by Lasso regression were used to build a BP neural network model. The BP neural network model was established using the newff function in MATLAB, and the optimal network structure obtained was 7-6-1-1, that is, there were 7 input variables, the model contained 6 hidden layers and 1 output layer.

[0142] The number of training times of the model was set to 1000, the minimum error of the training target was set to 0.00001, and the learning rate was set to 0.01. The prediction results of the model training set are as Figure 9 shown. It can be observed that the Lasso-BP neural network model has a relatively high fitting degree for predicting the flue gas CO release amount in the current data set.

[0143] The prediction results of the model test set are as Figure 10 shown. It can be observed that the model prediction results can be well fitted with the true results, and the model prediction ability is excellent.

[0144] The determination coefficient and root mean square error of the statistical model were used to evaluate the fitting accuracy of the model. The relevant indicators are shown in Table 4. Among them, the determination coefficient of the training set was 0.955, the determination coefficient of the test set was 0.951, the root mean square error of the training set was 0.0020, and the root mean square error of the test set was 0.0023, indicating that the fitting effect of the model was relatively ideal.

[0145] Table 4

[0146]

[0147] The mean absolute error and mean deviation of the statistical model were used to test the accuracy of the model. The results are shown in Table 5. The mean absolute error of the model training set was 0.0015, the mean absolute error of the test set was 0.0016, the mean deviation of the training set was -0.0002, and the mean deviation of the test set was -0.0003, indicating that the model had good prediction ability.

[0148] Table 5

[0149]

[0150] The fitting of the model for the prediction results of the training set, validation set, test set, and total set is represented by plotting the regression line, and the results are as Figure 11 shown. The slope of the ideal fitting line rises along a 45-degree angle, the model output result is equal to the true value, and at this time, the value of the correlation coefficient R is 1. Figure 11 The R values of the four lines in it are all above 0.95, meeting the requirements of the model accuracy.

[0151] To sum up, the determination coefficient of the regression model based on Lasso regression and BP neural network is 0.953, the root mean square error is 0.0022, the mean absolute error is 0.0016, the mean deviation is -0.0003, and the linear regression correlation coefficient between the predicted value and the true value is greater than 0.95. The model has a good fitting effect and high prediction accuracy. The regression model based on Lasso regression and BP neural network is suitable for predicting the CO release amount of mainstream smoke by the microscopic structure characteristics of cigarette rods. The error between the prediction result and the true value is within an acceptable range, which can accurately reflect the relationship between the microscopic structure characteristics of cigarette rods and the CO release amount of mainstream smoke, and can be applied in practice.

[0152] It can be seen from the above Lasso-BP neural network model that the specific surface area (A5) and fractal dimension (A6) characteristics in the cigarette rod parameters have the least influence on the CO release amount of mainstream smoke per unit weight of tobacco cut filler, especially the influence of the fractal dimension can be almost ignored. However, the mean value (A1) and variance (A2) characteristics of the porosity curve have an obvious influence on the CO release amount of mainstream smoke per unit weight of tobacco cut filler, especially the influence of the variance of the porosity curve is the most significant. A larger variance will cause more CO to be released per unit weight of tobacco cut filler. It can be considered to optimize the cigarette production process from this perspective to reduce the harm of cigarettes to the human body while meeting the needs of consumers.

[0153] The advantage of the present invention is to provide a study on the three-dimensional structure characterization of cigarette rods and its influence on the CO release amount, which can accurately obtain the projection images of cigarette rod samples and realize the tomographic image reconstruction under the spiral scanning trajectory; moreover, the microscopic feature characterization method of cigarette rods can clearly and accurately describe the internal microscopic structure of cigarette rod samples, and the CO release amount can be effectively regulated by modulating the porosity characteristics of cigarette rods, providing a new idea for cigarette designers to reduce tar and harm by improving the microscopic structure of cigarette rods.

[0154] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A model for characterizing the CO release amount of flue gas based on the microstructure of cigarette rods, characterized in that, it includes: S1. Using 3D-μCT scanning technology to construct a high-resolution tomographic sequence image inside the cigarette rod sample; S2. Establishing a three-dimensional structure model of the cigarette rod according to the tomographic sequence image, and obtaining the cigarette rod structure characteristic parameters; S3. Establishing a three-dimensional pore-throat ball-stick model of the cigarette rod according to the tomographic sequence image and the cigarette rod structure characteristic parameters; S4. After performing regression processing on the models and characteristic parameters obtained in S1 to S3, establishing a BP neural network model, and training and optimizing the neural network model according to the CO release amount data set of the cigarette rod, thus obtaining; The cigarette rod characteristic structure parameters include: porosity, specific surface area, fractal dimension, and permeability.

2. The model for characterizing the CO release amount of flue gas based on the microstructure of cigarette rods according to claim 1, characterized in that: The construction process of the high-resolution tomographic sequence images in S1 is as follows: Using a 3D-μCT system, taking the central rotation axis of the helical trajectory as the z-axis and the intersection point of the z-axis and the horizontal plane where the starting point of the helical scan trajectory line is located as the origin, a Cartesian coordinate system o-xyz is established. At this time, the coordinates of the voxels to be reconstructed in the region to be reconstructed inside the helical trajectory line are expressed as x = [x 0 , x 1 , x 2 . The distance from the X-ray source focus to the detector plane is D, the coordinates of the X-ray source focus are λ(s), and the calculation process of the helical scan trajectory line C is as follows: Formula 1:

3. The model for characterizing the CO release amount of flue gas based on the microstructure of cigarette rods according to claim 1, characterized in that: The process for obtaining the porosity in S2 is as follows: S2-1-1 statistically calculates the frequency of occurrence of each gray value in the gray level of the image according to the gray histogram of the tomographic sequence image, and obtains the occurrence probability p of each gray value i and the occurrence frequency q i value, where q i = i × p i , i ∈ [0, 65535]; Set the segmentation threshold as T. The proportions of the number of background pixels and the number of foreground pixels in the tomographic sequence image among all the pixels in the image are respectively and ω 1 = 1 - ω 0 . The average gray values of the background and the foreground are respectively According to g = ω 0 (μ 0 - μ) 2 + ω 1 (μ 1 - μ) 2 calculate the between-class variance value g; S2-1-3 The threshold T traverses the gray levels of all pixels in the image from 0 to 65535, and calculates the threshold T with the largest between-class variance max , which is the required threshold S2-1-4 Judging whether the pixel points in the image belong to pores according to the required threshold, and calculating the ratio of the number of pixel points belonging to pores to the total number of pixel points, which is the porosity.

4. The model for characterizing the CO release amount of flue gas based on the microstructure of cigarette rods according to claim 1, characterized in that: The process for obtaining the specific surface area in S2 is as follows: S2-2-1 Using the Canny edge detection algorithm to obtain the edges of the extracted tomographic sequence image; S2-2-2 Connecting the pixel points considered to be edges through an edge connection method to form a closed edge path; S2-2-3 Counting the percentage of the number of pixel points belonging to the edges in the tomographic sequence image in the total number of pixel points, which is the specific surface area of the tobacco shreds in this layer, and its calculation process is: Formula 1: In Equation 1: k is the specific surface area, M′ is the number of edge pixel points in the cigarette rod image, and M is the total number of pixel points in the cigarette rod image.

5. The model for characterizing the CO release amount of flue gas based on the microstructure of cigarette rods according to claim 1, characterized in that: The calculation process of the fractal dimension in S2 is as follows: S2-3-1 Divide the tomographic sequence image plane into grids of size R×R and arrange them in a three-dimensional space to form a curved surface with a length of R, a width of R, and a height of L, and perform the same division on the coordinate axis corresponding to L, with the unit of R×L / M; S2-3-2 Inside the R×R grid, find the maximum pixel value u and the minimum pixel value b, and calculate the number of boxes required to cover the entire area from the minimum value to the maximum value. Its calculation process is n(i,j) = [(u - b + R - 1) / R], and summing all n to get N, which is the number of boxes; S2-3-3 The calculation process of the fractal dimension F is as follows: Equation 2: F = -logN / logR.

6. The model for characterizing the CO release amount of flue gas based on the microstructure of cigarette rods according to claim 1, characterized in that: The calculation process of the permeability in S2 is: Simulating the pressure difference, state, and flow rate of the fluid flowing through the pores in the three-dimensional structure model of the cigarette rod to obtain the cigarette rod permeability statistical result: Formula 3: In Equation 3: K represents the permeability, Q is the flow rate of the fluid through the porous medium, μ is the viscosity of the fluid, L represents the length of the fluid flow, A represents the cross-sectional area of the fluid flowing through the porous medium, t represents the flow-through time, and Δp is the pressure difference at both ends.

7. A model for characterizing the CO release amount of cigarette smoke based on the microscopic structure of cigarette rods according to claim 1, characterized in that: The three-dimensional pore-throat stick model of the cigarette rod in S3 includes the following three characteristic indexes: pore-throat ratio h, average coordination number p, and three-dimensional shape factor G xyz , and the obtaining process is as follows: The pore-throat ratio h in S3-1 is the ratio of the average pore diameter to the average throat diameter in the pore-throat ball-and-stick model, and its calculation process is as follows: Formula 4: The average coordination number p in S3-2 is the average value of the number of throats connected to a single pore in the pore-throat ball-and-stick model, and its calculation process is as follows: Formula 5: The three-dimensional shape factor G described in S3-3 xyz is the degree of irregularity of the image shape, and its calculation process is as follows: Formula 6: In Formulas 4 to 6: d t is the average diameter of the throat in the model; d p is the average diameter of the pores in the model; n t is the number of throats in the model; n p is the number of pores in the model; L is the average throat length; r is the average throat radius.

8. A model for characterizing the CO release amount of cigarette smoke based on the microscopic structure of cigarette rods according to claim 1, characterized in that: The process of regression processing in S4 is as follows: S4-1-1 Use Lasso regression to screen and compress all models and characteristic parameters in S1 to S3, and perform standardization transformation to make the mean equal to 0 and the variance equal to 1. At this time, the Lasso regression model is estimated: Formula 7: In Equation 7: For any harmonic parameter s, the estimate of α is The overall regression coefficient can be reduced by adjusting the harmonic parameter s. When s 0 = Σ j |β j |, s ≤ s 0 At this time, by screening out irrelevant variables and variables with extremely weak relationships, the accuracy of the regression model is improved. S4-1-2 The regression coefficients in the regression model are solved using the Mallows Cp statistic, and its calculation process is as follows: Formula 8: Formula 9: In Equations 8 and 9: n represents the total number of sampling samples, p represents the number of independent variables in the subset regression model, and δ 2 represents the mean variance during regression, and SSE p represents the sum of squared residuals after regression; when C p reaches the minimum value, the subset of variables taken by the model within the global scope is the optimal one, and the generated regression model has the best effect.

9. A model for characterizing the CO release amount of cigarette smoke based on the microscopic structure of cigarette rods according to claim 8, characterized in that: The establishment process of the BP neural network model in S4 is as follows: Use the characteristic parameters obtained after Lasso regression screening as the input layer of the BP neural network, and establish a BP neural network model through the newff function; the number of hidden layers of the BP neural network model is ≥5, the number of training times is ≥1000, the minimum error is ≤0.0001, and the learning rate is 0.01 to 0.

1.

10. The application of a model for characterizing the CO release amount of cigarette smoke based on the microscopic structure of cigarette rods according to any one of claims 1 to 9, characterized in that: It is used to adjust the microscopic structure of cigarette rods to quantitatively control the CO release amount of cigarette smoke.

Images