General model building method for tobacco aroma components based on near infrared spectrum
By combining near-infrared spectroscopy with linear and nonlinear algorithms to establish a model of aroma components in tobacco leaves, the problem of insufficient variable selection in existing technologies is solved, enabling rapid and accurate assessment and control of tobacco leaf quality.
Patent Information
- Application Number
- CN202211649007.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-20
- Publication Date
- 2025-11-07
- Estimated Expiration
- 2042-12-20
AI Technical Summary
In existing technologies for modeling aroma components in tobacco leaves based on near-infrared molecular technology, the variable selection is merely the simplest spectral segmentation, which lacks theoretical and statistical significance.
Near-infrared spectroscopy was used, combined with linear algorithm PLS and nonlinear algorithm LSSVM, and a variable feature selection algorithm was used to optimize the model and establish a general model of aroma components in tobacco leaves. The linear algorithm PLS model was used first, and the nonlinear algorithm LSSVM model was used if the effect was not good. The RBF kernel function was used for nonlinear mapping to optimize the regularization parameters and kernel function parameters.
It enables rapid and accurate determination of aroma components in tobacco leaves, establishes a better model, improves the model's performance and robustness, simplifies computation, and provides a rapid and accurate means for tobacco quality control.
Smart Images
Figure CN115775606B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of cigarette manufacturing, and particularly relates to a general model modeling method for tobacco aroma components based on near-infrared spectroscopy. BACKGROUND
[0002] As a kind of increasingly popular consumer goods, the quality of cigarettes will directly affect the consumers' immediate interests, and tobacco leaf is the basis of cigarette production, and its inherent quality is the standard for measuring the usability of tobacco leaf. The chemical composition of cigarettes has a great correlation with its inherent quality, so the detection of the main components of tobacco leaf in the cigarette process has important theoretical significance and use value for the quality monitoring of the cigarette industry.
[0003] Tobacco aroma components are important factors affecting the sensory evaluation of tobacco flavor style quality, and studying tobacco aroma components is of great significance for revealing the causes of tobacco style quality.
[0004] Near-infrared molecular vibration spectroscopy mainly includes the frequency doubling and frequency combination absorption of X-H vibration of hydrogen-containing groups, and has the advantages of simplicity, rapidity (up to 3 minutes / sample), non-pollution, low cost, etc. The spectrum contains rich structural information and strong fingerprint characteristics, and is suitable for identification and identification of complex system samples. Near-infrared analysis technology has been widely used in the tobacco industry, such as origin traceability, tobacco quality determination, and rapid prediction of tobacco quality conventional physicochemical index values. Although near-infrared has been applied in the determination of tobacco aroma components, in these studies, the variable selection is only the simplest spectral band splitting, and lacks theoretical and statistical significance. Therefore, it is necessary to make improvements. SUMMARY
[0005] The technical problem solved by the present application is to provide a general model modeling method for tobacco aroma components based on near-infrared spectroscopy, which is proposed in view of the fact that in the prior art, the variable selection in the modeling research of tobacco aroma components based on near-infrared molecular technology is only the simplest spectral band splitting, and lacks theoretical and statistical significance.
[0006] To achieve the above purpose, the technical scheme adopted by the present application is as follows:
[0007] The general model modeling method for tobacco aroma components based on near-infrared spectroscopy comprises the following steps:
[0008] Step 1): Select different tobacco leaves to make tobacco leaf samples for storage;
[0009] Step 2): Collect near-infrared spectroscopy data of each tobacco leaf sample and optimize it into a near-infrared spectroscopy spectrum;
[0010] Step 3): detecting the aroma chemical values in the tobacco sample;
[0011] Step 4): setting a set of prediction set target evaluation parameters of tobacco aroma components; based on the near-infrared spectrum data collected in step 1), the spectrum data is processed by a suitable pretreatment method; then the pretreated spectrum data and the aroma component content data in step 3) are optimized by a variable characteristic selection algorithm model, the characteristic wavelength of the tobacco sample aroma component is extracted, and compared with the prediction set target evaluation parameters to determine the optimal feasible general model;
[0012] Wherein the variable characteristic selection algorithm adopts linear algorithm PLS and nonlinear algorithm lssvm, and the linear algorithm PLS model is preferred for modeling, and if the linear model is not good, the nonlinear algorithm lssvm model is used for modeling;
[0013] Step 5): for the tobacco sample to be measured, first collect its near-infrared spectrum data, and use the general model to predict the aroma components.
[0014] In the above step 1), when the tobacco sample is prepared, the selected different tobacco is ground and labeled and put into a sealed bag, the moisture is balanced, and the sample is stored for testing.
[0015] In the above step 2), when collecting the near-infrared spectrum data of each tobacco sample, the sample is placed in a rotating sampling cup, each sample is resampled three times, and each time the spectrum is collected and saved for subsequent data processing.
[0016] In the above step 2), the near-infrared spectrum measurement method is selected by diffuse reflection integral sphere measurement method; the near-infrared instrument working condition setting and spectrum collection: using Thermo Antaris II Fourier transform near-infrared spectrometer, after starting preheating for 30 minutes, the spectrum scanning range is 10000-4000cm -1 , the resolution is 8cm -1 , the scanning number is 64 times, the experimental environment temperature is (23±1)℃, and the humidity is (50±2)%; the rotating sample cup is determined as 0.5cm thick.
[0017] In the above step 3), the aroma components are propylene glycol, damascenone, farnesyl acetone, geranyl propyl ketone, and new ester diene.
[0018] In the above step 4), the specific process of each variable characteristic selection algorithm is as follows:
[0019] (1) The PLS algorithm steps are as follows:
[0020] Input sample matrix X, output value matrix Y, respectively, center the X, Y, and normalize the variance to obtain matrix: E0 and F;
[0021] Let n = 0;
[0022] Compute matrix E n 'FF'E n , compute E n 'FF'E n the unit eigenvector corresponding to the largest eigenvalue of E
[0023] Let:
[0024] t n = E n w n , principal components of the independent variables
[0025] p n = E n 't n / |t n | 2 , regression coefficients of the independent variables
[0026] r n = F't n / |t n | 2 , regression coefficients of the dependent variable
[0027] E n+1 = E n -t n p n , residuals of the independent variables
[0028] Also can be expressed as:
[0029] n++, repeat 3, 4, until residuals reach a satisfactory level
[0030] Final regression equation:
[0031] F = t0r0' + t1r1' +... t n r n ' + F n
[0032] where F n is the residual of the dependent variable
[0033] The prediction equation can be written as: F = E0B
[0034] where B = ∑k j r j '
[0035] In the above equation ( from j = 0 to j = h-1);
[0036] (2) Nonlinear algorithm lssvm:
[0037] The algorithm utilizes the nonlinear inner product function K(X) i X j Kernel functions, also known as kernel functions, nonlinearly map input low-dimensional data to high-dimensional space. They solve for the optimal linear classifier or multivariate linear model in high-dimensional space by minimizing structural risk. Since the kernel function in the original space replaces the inner product operation in the high-dimensional space, the algorithm complexity does not increase. In form, it is similar to artificial neural networks (ANN).
[0038] The LSSVM least squares support vector machine algorithm uses a least squares linear system as the loss function, replacing the quadratic programming method used in traditional support vector machines. Based on the principle of minimizing structural risk, LSSVM optimizes the following problems:
[0039]
[0040] λ is the regularization parameter, which plays a role in balancing fitting accuracy and model complexity;
[0041] The regression model can be described as:
[0042]
[0043] x i For the input vector; a k Let be the Lagrange multiplier, called the support value, and b be the bias kernel function, K(x, x) i This defines a method for making nonlinear mappings from the low-dimensional space containing the original variables to the high-dimensional space. Currently, there is no fixed method for choosing the kernel function in support vector machine research, but from the perspectives of kernel compactness and computational complexity, the radial basis function (RBF) is more advantageous.
[0044] K(x, x) k ) = exp(-||x-x k || 2 / σ 2 )
[0045] In the formula, σ 2 It is the width of the RBF kernel function.
[0046] Regularization parameter λ optimization, kernel function K(x, x) i The selection of kernel function parameters and their optimization are three key issues in LSSVM. The RBF kernel function was chosen, and simplex search and 10-fold cross-validation were used to optimize the parameters λ and σ. 2 These two parameters determine lssvm's learning ability, prediction ability, and generalization ability.
[0047] Compared with the prior art, the present application has the advantages of:
[0048] 1. In the present application, tobacco is selected as the research object, near-infrared molecular vibration spectroscopy technology is used to establish a tobacco quality rapid analysis and evaluation method and technology system, pretreatment and variable selection are combined to establish quantitative models of tobacco aroma components such as propylene glycol, damascenone, farnesyl acetone, geranyl propyl ketone and neolignan diene, thereby realizing rapid determination of the quality indexes of tobacco, providing a new method for rapid analysis and evaluation of the quality indexes of tobacco, and providing strong technical support for the rapidity, accuracy and digitization of tobacco quality control of cigarette enterprises;
[0049] 2. In the present application, the lssvm algorithm model is used to complete modeling of aroma components such as propylene glycol, damascenone, farnesyl acetone, geranyl propyl ketone and neolignan diene in tobacco, thereby establishing a more optimal and chemically meaningful model, improving model performance, simplifying the model, reducing calculation amount, improving the interpretability of the model, and improving the robustness of the model. BRIEF DESCRIPTION OF DRAWINGS
[0050] Figure 1 Fig. 1 is a modeling diagram of the lssvm model for aroma component propylene glycol in the embodiments of the present application;
[0051] Figure 2 Fig. 2 is a modeling diagram of the lssvm model for aroma component damascenone in the embodiments of the present application;
[0052] Figure 3 Fig. 3 is a modeling diagram of the lssvm model for aroma component farnesyl acetone in the embodiments of the present application;
[0053] Figure 4 Fig. 4 is a modeling diagram of the lssvm model for aroma component geranyl propyl ketone in the embodiments of the present application;
[0054] Figure 5 Fig. 5 is a modeling diagram of the lssvm model for aroma component neolignan diene in the embodiments of the present application;
[0055] Figure 6 Fig. 6 is a schematic diagram of the least squares support vector machine lssvm in the present application;
[0056] Figure 7 Fig. 7 is a specific flowchart of the least squares support vector machine lssvm in the present application. DETAILED DESCRIPTION
[0057] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the scope of protection of the present application.
[0058] Please refer to Figures 1-7 for detailed description of the embodiments of the present application.
[0059] Embodiment 1
[0060] The modeling method of the general model of tobacco aroma components based on near-infrared spectroscopy includes the following steps:
[0061] Step 1): Select more than 300 tobacco leaf samples in stock in a cigarette production area, grind and label them into sealed bags, balance the moisture, and save them for testing.
[0062] Step 2): Collect near-infrared spectroscopy data of each tobacco leaf sample and optimize the near-infrared spectroscopy spectrum;
[0063] The near-infrared spectroscopy measurement method selects the diffuse reflection integrating sphere measurement method; the near-infrared instrument working condition setting and spectrum collection: adopt Thermo Antaris II Fourier transform near-infrared spectrometer, preheat for 30 minutes after starting, the spectrum scanning range is 10000-4000 cm -1 , select resolution 8 cm -1 , scan 64 times, the experimental environment temperature is (23±1)℃, and the humidity is (50±2)%; the sample cup is determined to be 0.5 cm thick.
[0064] When collecting near-infrared spectroscopy data of each tobacco leaf sample, place the sample in a rotating sampling cup, take a sample three times for each sample, collect the spectrum each time, save it for subsequent data processing.
[0065] Step 3): Detect the aroma component chemical value in the tobacco leaf sample; the aroma component is propylene glycol, damascenone, farnesyl acetone, geranyl acetone, and neolignan diene.
[0066] Step 4): Set a group of prediction set target evaluation parameters of tobacco aroma components; based on the sample near-infrared spectroscopy data collected in step 1), process the spectrum data by using a suitable pretreatment method; then optimize the model of the variable characteristic selection algorithm for the pretreated spectrum data and the aroma component content data of the tobacco leaf sample in step 3), complete the extraction of the characteristic wavelength of the tobacco leaf sample, and compare it with the prediction set target evaluation parameters to determine the optimal feasible general model;
[0067] Wherein the variable feature selection algorithm adopts linear algorithm PLS and nonlinear algorithm lssvm, and the linear algorithm PLS model is preferentially adopted for modeling, and if the linear model is not good, the nonlinear algorithm lssvm model is adopted for modeling;
[0068] The specific process of each variable feature selection algorithm is as follows:
[0069] (1) The PLS algorithm steps are as follows:
[0070] The input sample matrix X and the output value matrix Y are respectively subjected to centering treatment and normalized variance to obtain matrix E0 and F;
[0071] Let n = 0:
[0072] Calculate the matrix E n ’FF’E n , calculate the maximum eigenvalue corresponding to the unit eigenvector of E n ’FF’E n
[0073] Let:
[0074] t n =E n w n , principal component of independent variable
[0075] p n =E n ’t n / t n | 2 , regression coefficient of independent variable
[0076] r n =F’t n / |t n | 2 , regression coefficient of dependent variable
[0077] E n+1 =E n -t n p n , residual of independent variable
[0078] It can also be expressed as:
[0079] n++, repeat 3, 4, until the residual reaches a satisfactory level
[0080] The final regression equation is:
[0081] F = t0r0’ + t1r1’ + …… t n r n ’ + Fn
[0082] where F n is the dependent variable residual
[0083] The prediction equation can be written as: F = E0B
[0084] where B =∑k j r j '
[0085] In the above equation ( from j = 0 to j = h-1);
[0086] (2) Nonlinear algorithm lssvm:
[0087] This algorithm maps the input low-dimensional space data to high-dimensional space through a nonlinear inner product function K(X i , X j ), also known as a kernel function, and solves the optimal linear classifier or multivariate linear model in high-dimensional space using the principle of structural risk minimization. Since the kernel function in the original space is used instead of the inner product operation in high-dimensional space, the algorithm complexity has not increased, and the form is similar to artificial neural network ANN. The structure is shown in Figure 6 . The specific process of lssvm is shown in Figure 7 .
[0088] The lssvm least squares support vector machine algorithm uses a least squares linear system as a loss function, instead of the quadratic programming method used by the traditional support vector machine, and based on the principle of structural risk minimization, lssvm optimizes the following problem:
[0089]
[0090] λ is a regularization parameter that plays a role in balancing fitting accuracy and model complexity;
[0091] The regression model can be described as:
[0092]
[0093] x i is the input vector; a k is the Lagrange multiplier, called the support value, and b is the bias kernel function, K(x, x i ) defines the way to do nonlinear mapping from the low-dimensional space where the original variable is located to the high-dimensional space; Currently, there is no fixed method for selecting the kernel function in the study of support vector machines, but from the compactness and computational complexity of the kernel function, the radial basis kernel function RBF has more advantages, and has:
[0094] K(x, x k)=exp(-||x-x k || 2 / σ 2 )
[0095] where σ 2 is the RBF kernel function width.
[0096] The regularization parameter λ optimization, kernel function K(x, x i ) selection and kernel function parameter optimization are three key problems of lssvm; select RBF kernel function, use simplex search, ten-fold cross-validation to optimize parameters λ and σ 2 , which determine the learning ability, prediction ability and generalization ability of lssvm.
[0097] The following is the specific modeling effect of the embodiment:
[0098] The PLS modeling results of the five aroma components are as follows:
[0099]
[0100] After rejection CVQ2 PREQ2 RMSECV RMSEP Rejection 10 Propylene glycol 0.5983 0.5046 0.941 0.9587 Rejection 14 Dammarenediolide 0.402 0.5567 1.5787 1.0976 Rejection 8 Neophytadiene 0.2397 0.2731 209.343 175.2491 Rejection 8 Geranylacetone 0.597 0.3415 1.9261 2.3727 Rejection 8 Farnesylacetone 05794 04378 34527 43264
[0101] Among them, the ipls, mwpls and uve models of wavelength selection in the pls model are as follows:
[0102] Ipls:
[0103] CVQ2 RMSECV PREQ2 RMSEP Propylene glycol 0.4600 0.8781 0.5096 0.9086 Dammarenediolide 0.2548 1.5699 0.3204 1.4736 Geranylacetone 0.4921 2.0310 0.5302 1.9594 Neophytadiene 0.3346 190.8100 0.3340 174.6200 Farnesylacetone 0.5067 3.7455 0.4930 4.7336
[0104] Mwpls:
[0105] mwpls CVQ2 PREQ2 RMSECV RMSEP Propylene glycol 0.563348 -0.02232 0.914242 0.923363 Dammarenediolide 0.242567 0.44885 1.571697 1.190716 Neophytadiene 0.303472 0.261156 190.0987 208.9042 Geranylacetone 0.462447 0.479087 2.283225 2.086884 Farnesylacetone 0.389559 0.390952 4.506249 4.141764
[0106] Uve:
[0107] nLV CVQ2 PREQ2 RMSECV RMSEP Propylene glycol 10 0.5529 0.7262 0.7990 0.6789 Dammarenediolide 7 0.4616 0.5110 1.3343 1.2500 Geranylacetone 8 0.5740 0.6610 1.8600 1.6645 Neophytadiene 5 0.3219 0.2680 192.6223 183.0664 Farnesylacetone 7 0.6068 0.4425 3.3432 4.9637
[0108] The lssvm modeling structure is as shown in the following table: Figures 1-5
[0109] CVQ2 Q2_pre RMSECV RMSEP Geranylacetone 1.0000 0.6260 0.0015 1.7609 Neophytadiene 1.0000 0.7030 0.5069 120.7261 Propylene glycol 1.0000 0.0018 0.8563 0.4827 Dammarenediolide 1.0000 0.8236 0.0054 0.7068 Farnesylacetone 1.0000 0.7805 0.0038 2.9693
[0110] The models are established by the chemical values and spectra of the above five aroma components, and pls, ipls, mwpls, uve models of wavelength selection and lssvm models are used. According to the above data, it can be seen that the effect of lssvm model is better.
[0111] Step 5): For the to-be-tested tobacco sample, first collect its near-infrared spectrum data, and use the general model to predict the aroma components.
[0112] In summary, the present application selects tobacco as the research object, establishes a tobacco rapid quality analysis and evaluation method and technology system by using near-infrared molecular vibration spectroscopy technology, combines pretreatment and variable selection to establish a quantitative model of tobacco aroma components propylene glycol, damascenone, farnesyl acetone, geranyl propyl ketone and neolignan diene, and realizes rapid determination of the quality index of tobacco.
[0113] It will be obvious to a person skilled in the art that the application is not limited to the details of the above-described exemplary embodiments, but that the application can be implemented in other concrete forms without departing from the spirit or essential characteristics of the application. The embodiments should therefore be considered in all respects as illustrative and not restrictive, the scope of the application being defined by the appended claims rather than by the above description, and it is therefore intended that all changes and modifications that fall within the meaning and range of equivalency of the elements of the claims are to be embraced by the application, and no picture reference in the claims should be considered limiting as to the claims concerned.
[0114] Furthermore, it should be understood that, although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the specification is described in this way only for the sake of clarity, and a person skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by a person skilled in the art.
Claims
1. A method for modeling a general model of tobacco aroma components based on near-infrared spectroscopy, characterized in that: It comprises the following steps: Step 1): select different tobaccos to make tobacco samples for storage; Step 2): collect near-infrared spectrum data of each tobacco sample and optimize the near-infrared spectrum spectrum; Step 3): detect the chemical value of the aroma components in the tobacco sample; Step 4): set a group of prediction set target evaluation parameters of tobacco aroma components; based on the sample near-infrared spectrum data collected in step 1), the spectrum data is processed by a pretreatment method; then the pretreated spectrum data and the chemical value data of the aroma components in the tobacco sample in step 3) are optimized by a variable characteristic selection algorithm model, the characteristic wavelength of the aroma components in the tobacco sample is extracted, and compared with the prediction set target evaluation parameters to determine the optimal feasible general model; The variable characteristic selection algorithm uses linear algorithm PLS and nonlinear algorithm lssvm, and the linear algorithm PLS model is preferred for modeling, and if the linear model is not good, the nonlinear algorithm lssvm model is used for modeling; Step 5): for the tobacco sample to be tested, first collect its near-infrared spectrum data, and use the general model to predict the aroma components. In the above step 3), the aroma components are propylene glycol, damascenone, farnesyl acetone, geranyl acetone, and neolignan diene.
2. The method according to claim 1, wherein the method is characterized by: In the above step 1), when making tobacco samples, the selected different tobaccos are ground, labeled and placed in sealed bags, balanced in moisture, and stored for testing.
3. The method according to claim 1, wherein the method is characterized by: In the above step 2), when collecting near-infrared spectrum data of each tobacco sample, the sample is placed in a rotating sampling cup, each sample is resampled three times, and each time the spectrum is collected and saved for subsequent data processing.
4. The method according to claim 2, wherein the method is characterized by: In the step 2), the near infrared spectrum measurement method is selected as the diffuse reflection integrating sphere measurement method; the near infrared instrument working condition setting and spectrum collection are as follows: a Thermo Antaris II Fourier transform near infrared spectrometer is used, after being preheated for 30 minutes, the spectrum scanning range is 10000-4000cm -1 , the resolution is selected as 8cm -1 , the scanning times are 64, the experimental environment temperature is (23±1)℃, the humidity is (50±2)%, and the thickness of the sample loaded in the rotating sample cup is determined as 0.5cm.
5. The method according to claim 1, wherein the method is characterized by: In the above step 4), the specific process of each variable characteristic selection algorithm is as follows: (1) PLS algorithm steps are as follows: (2) Nonlinear algorithm lssvm: The algorithm maps the input low-dimensional space data to high-dimensional space by a nonlinear inner product function K(X i , X j ), also called a kernel function, and solves an optimal linear classifier or a multivariate linear model in the high-dimensional space by using a structural risk minimization principle. Since the kernel function of the original space is used to replace the inner product operation in the high-dimensional space, the algorithm complexity does not increase. The least squares support vector machine algorithm lssvm uses a least squares linear system as a loss function, instead of the quadratic programming method used by traditional support vector machines, based on the principle of structural risk minimization, lssvm optimizes the following problem: MinJ(w, e) = MinJ(w, e) + 1 λ λ is a regularization parameter that plays a role in balancing fitting accuracy and model complexity; The regression model can be described as: y(x) = a x + b Ｋ (x, x ｉ ) + b x ｉ is an input vector; a k is a Lagrange multiplier, called support value, b is a bias kernel function, K(x, x ｉ ) defines the way of making a non-linear mapping from the low-dimensional space where the original variables lie to the high-dimensional space; there is: K(x, x ｋ ) = exp(-||x - x ｋ || / σ ２ ) ２ ) In the formula, σ ２ is the RBF kernel function width; The regularization parameter λ optimization, kernel function K(x, x ｉ ) selection and kernel function parameter optimization are three key problems of lssvm; the RBF kernel function is selected, the simplex search and ten-fold cross-validation are used to optimize parameters λ and σ ２ , and the two parameters determine the learning ability, prediction ability and generalization ability of lssvm.