Cigarette leaf group formula matching degree calculation method based on improved Gaussian mixture model

Abstract

The invention relates to a cigarette leaf group formula matching degree calculation method based on an improved Gaussian mixture model. The method comprises the following steps: acquiring characterization data of a historical leaf group formula of a target cigarette leaf specification; constructing a Gaussian mixture model, and enabling the Gaussian mixture model to fit probability distribution of characterization data of a historical leaf group formula; inputting the feature data of the to-be-selected test formula into the Gaussian mixture model, and recalculating Gaussian component weights for the feature data of the to-be-selected test formula through a dynamic weight adjustment method based on an adaptive attention mechanism to obtain an optimized Gaussian mixture model; and determining the comprehensive matching degree of the to-be-selected test formula and the historical formula by using the optimized Gaussian mixture model. By adopting the method, the tobacco grade replacement strategy and the adjustment measure of the formula structure can be evaluated by utilizing the data structure distribution of the historical tobacco formula.

Description

Technical Field

[0001] This application relates to the field of cigarette product design technology, and in particular to a method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model. Background Technology

[0002] In cigarette product design, formula maintenance is a crucial step in ensuring the stability of style, quality, and brand quality; its stability is the core of maintaining product quality and characteristics. During the formula maintenance design process, the volatility of raw material quality must be comprehensively considered, and appropriate tobacco grade substitution strategies and formula structure adjustments must be adopted to maximize the continuity of formula style and quality.

[0003] The development and optimization of cigarette leaf blends have long faced challenges such as the difficulty in accurately quantifying evaluation indicators and low design efficiency. This not only affects the rational allocation of tobacco resources but also limits the in-depth development of their potential value. Existing technologies are largely subjective, relying on sensory evaluation experience and supplemented by fluctuations in a few leaf blend characterization indicators to judge the sensory stability and production feasibility of the blend. A comprehensive and objective evaluation or selection mechanism is lacking. Furthermore, the data accumulated in cigarette product blend design, as an important reflection of style and quality characteristics, has not been systematically integrated and utilized, resulting in its underutilization in the optimization process. Although a series of explorations have been conducted on selecting alternative tobacco raw materials in conjunction with existing blend maintenance programs, these have mainly focused on tobacco substitution relationships and compatibility rules between single-grade tobacco leaves, lacking consideration of the internal data structure distribution, overall similarity, and comprehensive matching degree of various indicators of the blend leaf group with the chemical composition of the product blend.

[0004] Therefore, how to effectively combine historical formula characterization data to evaluate and optimize tobacco substitution strategies and formula structure adjustment measures in order to maximize the maintenance of formula style and quality continuity is an urgent problem to be solved. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model to address the above-mentioned technical problems, and to conduct an objective and accurate quantitative evaluation of the similarity of leaf blend formulation based on the internal structural characteristics of historical data.

[0006] Firstly, this application provides a method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model, including:

[0007] Obtain characterization data of historical leaf group formulations for the target tobacco leaf specifications and construct a historical formulation dataset;

[0008] Feature extraction is performed on the historical formula dataset to obtain a feature dataset;

[0009] A Gaussian mixture model consisting of a weighted superposition of multiple Gaussian components is constructed. Based on the feature dataset, the optimal parameters of the Gaussian mixture model are estimated so that the Gaussian mixture model fits the probability distribution of the characterization data of the historical leaf group formulation.

[0010] Feature extraction is performed on the characterization data of the candidate experimental formulation. The obtained candidate experimental formulation feature data is input into the Gaussian mixture model. Using a dynamic weight adjustment method based on an adaptive attention mechanism, the weights of the Gaussian components in the Gaussian mixture model are recalculated for the candidate experimental formulation feature data to obtain an optimized Gaussian mixture model.

[0011] The characteristic data of the candidate test formulations are input into the optimized Gaussian mixture model to calculate the probability density value. Based on the probability density value, the overall matching degree between the candidate test formulations and the historical formulations is determined.

[0012] In one embodiment, a Gaussian mixture model is constructed by weighted superposition of multiple Gaussian components. Optimal parameter estimation is performed on the Gaussian mixture model based on the feature dataset, enabling the Gaussian mixture model to fit the probability distribution of the characterization data of the historical leaf group formulation.

[0013] Establish the Gaussian mixture model, initialize the model parameters and the number of Gaussian components, wherein the model parameters include the weights, mean and covariance of the Gaussian components;

[0014] Based on the feature dataset, the model parameters are iteratively optimized using the expectation-maximization algorithm until the model convergence condition is met.

[0015] In one embodiment, iteratively optimizing the model parameters using the expectation-maximization algorithm includes:

[0016] For each feature dimension, calculate the posterior probability that each data point in the feature dataset belongs to each Gaussian component;

[0017] Update the weights, mean, and covariance of the Gaussian components based on the posterior probability.

[0018] Repeat the above steps until the model convergence condition is met.

[0019] In one embodiment, the feature data of the candidate experimental formulations are input into the Gaussian mixture model. Using a dynamic weight adjustment method based on an adaptive attention mechanism, the weights of the Gaussian components in the Gaussian mixture model are recalculated for the candidate experimental formulation feature data, resulting in an optimized Gaussian mixture model, including:

[0020] Input the characteristic data of the candidate test formulation into the Gaussian mixture model, and calculate the distance between the characteristic data of the candidate test formulation and the mean of each Gaussian component;

[0021] Based on the distance, the adaptive attention weights of the Gaussian components are calculated using an adaptive attention function;

[0022] The adaptive attention weights are used to replace the weights of the Gaussian components in the Gaussian mixture model to generate a Gaussian mixture model optimized for the feature data of the candidate experimental formulations.

[0023] In one embodiment, the adaptive attention function is calculated using the following formula:

[0024] ;

[0025] in, The attention weight of the j-th feature of the candidate experimental formulation feature data x in the k-th Gaussian component. Let K be the mean of the k-th Gaussian component, and K be the total number of Gaussian components. It is a hyperparameter that can be set or optimized through cross-validation, and .

[0026] In one embodiment, feature extraction from the historical formulation dataset and feature extraction from the characterization data of the candidate experimental formulations include:

[0027] The covariance matrix of the historical leaf group formula dataset is calculated, and eigenvalue decomposition is performed to obtain eigenvalues ​​and corresponding eigenvectors.

[0028] Calculate the cumulative variance contribution rate of the feature values, and retain the key feature values ​​and corresponding key feature vectors whose cumulative variance contribution rates reach a preset threshold.

[0029] The historical leaf group formula dataset is projected onto the subspace formed by the key feature vectors to generate a dimensionality-reduced dataset, which is used as the feature dataset.

[0030] The candidate test formulation data is projected onto the subspace formed by the key feature vectors to obtain the candidate test formulation feature data.

[0031] In one embodiment, the characteristic data of the candidate experimental formulation is input into the optimized Gaussian mixture model to calculate the probability density value. Based on the probability density value, the overall matching degree between the candidate experimental formulation and the historical formulation is determined, including:

[0032] Calculate the feature weight of each feature dimension in the candidate experimental formulation feature data for all features;

[0033] For each of the features, the local probability density value of each feature is obtained based on the optimized Gaussian mixture model;

[0034] Based on the feature weights, the local probability density values ​​are weighted and summed to obtain the overall matching degree between the experimental formula and the historical formula.

[0035] In one embodiment, the formula for calculating the overall matching degree is:

[0036] ;

[0037] ;

[0038] Where M represents the overall matching degree, Let be the feature weight of the j-th dimension feature, and p be the number of features. The j-th dimension feature represents the number of Gaussian components in the optimized Gaussian mixture model. Let the j-th feature be the weight of the k-th Gaussian component in the optimized Gaussian mixture model. The mean is The covariance matrix is The multidimensional Gaussian distribution.

[0039] Secondly, this application also provides a device for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model, comprising:

[0040] The historical data acquisition module is used to acquire characterization data of historical leaf group formulas for the target tobacco leaf specifications and construct a historical formula dataset.

[0041] The historical data feature extraction module is used to extract features from the historical formula dataset to obtain a feature dataset;

[0042] The model training module is used to construct a Gaussian mixture model composed of weighted superposition of multiple Gaussian components, and to perform optimal parameter estimation on the Gaussian mixture model based on the feature dataset, so that the Gaussian mixture model fits the probability distribution of the characterization data of the historical leaf group formula.

[0043] The model optimization module is used to extract features from the characterization data of the candidate experimental formulation. The obtained feature data of the candidate experimental formulation is input into the Gaussian mixture model. Through a dynamic weight adjustment method based on an adaptive attention mechanism, the weights of the Gaussian components in the Gaussian mixture model are recalculated for the feature data of the candidate experimental formulation to obtain the optimized Gaussian mixture model.

[0044] The matching degree calculation module is used to perform feature dimensionality reduction on the candidate experimental formula dataset and input it into the trained Gaussian mixture model. Through a dynamic weight adjustment method based on the attention mechanism, the module calculates the comprehensive matching degree between the candidate experimental formula and the historical formula.

[0045] The formulation selection module is used to input the feature data of the candidate experimental formulation into the optimized Gaussian mixture model, calculate the probability density value, and determine the comprehensive matching degree between the candidate experimental formulation and the historical formulation based on the probability density value.

[0046] Thirdly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any one of the first to second aspects above.

[0047] The aforementioned method, apparatus, and computer-readable storage medium for calculating the matching degree of cigarette leaf blend formulations based on an improved Gaussian mixture model, through in-depth analysis of the characterization data of historical feed leaf blend formulations, improves the Gaussian mixture model through an adaptive attention scoring mechanism, mines the internal data structure distribution of historical formulations, and conducts a multi-dimensional quantitative comprehensive matching degree evaluation of cigarette formulation design schemes. This maximizes the continuity of formulation style and quality, effectively solving the problem that existing technologies cannot scientifically evaluate and optimize tobacco grade substitution strategies and formulation structure adjustment schemes based on the characteristics of the leaf blend formulation itself. This provides a scientific basis and practical tools for the optimization and maintenance of cigarette leaf blend formulations. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is an application environment diagram of a method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model in one embodiment.

[0050] Figure 2 This is a flowchart illustrating a method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model in one embodiment.

[0051] Figure 3 This is a flowchart illustrating the feature extraction steps using PCA dimensionality reduction in one embodiment.

[0052] Figure 4This is a flowchart illustrating a dynamic weight adjustment method based on an adaptive attention mechanism in another embodiment.

[0053] Figure 5 This is a flowchart illustrating the steps for calculating the overall matching degree in one embodiment;

[0054] Figure 6 This is a flowchart illustrating a preferred embodiment of a method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model.

[0055] Figure 7(a) shows the principal components of the feature dataset based on a Gaussian mixture model in one embodiment. The probability distribution;

[0056] Figure 7(b) shows the principal components of the feature dataset based on a Gaussian mixture model in one embodiment. The probability distribution;

[0057] Figure 7(c) shows the principal components of the feature dataset based on a Gaussian mixture model in one embodiment. The probability distribution;

[0058] Figure 7(d) shows the principal components of the feature dataset based on a Gaussian mixture model in one embodiment. The probability distribution;

[0059] Figure 8 This is a structural block diagram of a device for calculating the matching degree of cigarette leaf blend formula based on an improved Gaussian mixture model in one embodiment. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0061] This application provides a method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model, which can be applied to, for example... Figure 1In the application environment shown, the method is executed by server 104. Server 104 receives formula data requests or candidate experimental formula data from terminal 102, retrieves historical leaf group formula characterization data of the target cigarette specification from the data storage system via the network, performs a series of calculations, and returns the comprehensive matching degree result to terminal 102. The data storage system stores formula characterization data, model parameters, and calculation results. It can be integrated on server 104 or deployed on a cloud platform or other network server. Terminal 102 communicates with server 104 via the network. Terminal 102 can be, but is not limited to, a personal computer, workstation, or professional terminal device used by cigarette formula designers. It is used to submit query requests, input candidate experimental formula data, and receive and display the formula matching degree calculation results returned by the server. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0062] In one exemplary embodiment, such as Figure 2 As shown, a method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model is provided, and this method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps S201 to S205. Wherein:

[0063] S201, Obtain characterization data of historical leaf group formulas for the target tobacco leaf specification and construct a historical formula dataset.

[0064] The target cigarette specification refers to a specific cigarette brand and specification that requires formula matching, typically defined by characteristics such as brand, tar content, and aroma style. Historical leaf blend formulas refer to proven and effective leaf blend formulas used in the past production or R&D processes of this specification; each formula consists of multiple tobacco leaf raw materials in specific proportions. Characterization data is a set of indicators used to quantitatively describe the characteristics of the leaf blend formula, including but not limited to conventional chemical component indicators, conventional chemical component derived indicators, near-infrared model prediction indicators, and aging time. After acquiring the characterization data, the characterization data of multiple historical leaf blend formulas are organized into a historical formula dataset according to a unified format.

[0065] S202, extract features from the historical formula dataset to obtain the feature dataset.

[0066] The feature extraction step involves extracting key information from the characterization data of historical leaf group formulations. The aim is to reduce data dimensionality, remove redundant information, and retain core features.

[0067] For example, based on the aforementioned indicator dataset, Principal Component Analysis (PCA) is used for dimensionality reduction, and key feature information is extracted. PCA is an unsupervised dimensionality reduction algorithm that can map high-dimensional data to a low-dimensional space while preserving the core information of the data, eliminating redundancy and multicollinearity between features, and reducing subsequent computational complexity. Simultaneously, a cumulative variance contribution rate threshold is set, and the number of principal components whose cumulative variance contribution rate reaches the preset threshold is selected as the key feature. The dimensionality-reduced indicator data is then used as the feature dataset.

[0068] S203. Construct a Gaussian mixture model consisting of a weighted superposition of multiple Gaussian components. Estimate the optimal parameters of the Gaussian mixture model based on the feature dataset so that the Gaussian mixture model fits the probability distribution of the characterization data of the historical leaf group formulation.

[0069] Among them, the Gaussian Mixture Model (GMM) is a modeling method that uses multiple Gaussian distributions to approximate complex data distributions. The multimodal distribution characteristics of the Gaussian Mixture Model have a better ability to represent complex data situations, and the robustness of the model can be ensured by setting appropriate parameters such as the Gaussian number, reducing overfitting. Therefore, it is suitable for capturing multiple different style clusters of formulations that may exist in historical formulation data.

[0070] Gaussian mixture models assume that all data points are composed of multiple Gaussian distributions mixed with certain weights, where each Gaussian distribution is also called a Gaussian component. Each Gaussian component has a probability density function defined by the mixture weights, mean, and covariance matrix. The mixture weights represent the importance of the component in the population, the mean represents the center point of the component in the feature space, and the covariance matrix represents the shape and distribution range of the component. The construction of Gaussian mixture models typically employs optimal parameter estimation methods to estimate the mixture weights, mean, and covariance matrix of each Gaussian component in the Gaussian mixture model, maximizing the probability of the Gaussian mixture model generating the observed data.

[0071] Once the model parameters are determined, they can represent the patterns in historical recipe data. When the known parameters are input into the model, the model outputs a corresponding probability density. This probability density represents the likelihood that the input variable is generated by the overall data distribution described by the Gaussian mixture model, i.e., the "similarity" between the recipe and the overall style of the historical recipe dataset.

[0072] S204. Feature extraction is performed on the characterization data of the candidate experimental formulation. The obtained feature data of the candidate experimental formulation is input into the Gaussian mixture model. Through a dynamic weight adjustment method based on an adaptive attention mechanism, the weights of the Gaussian components in the Gaussian mixture model are recalculated for the feature data of the candidate experimental formulation, and the optimized Gaussian mixture model is obtained.

[0073] Feature projection calculations are performed on the experimental formulation dataset based on the aforementioned dimensionality reduction method. Using the dimensionality-reduced dataset, an adaptive attention scoring mechanism is employed to dynamically optimize the component weight allocation in the Gaussian mixture model, achieving competitive weight adjustment and optimization among the Gaussian components.

[0074] S205. Input the characteristic data of the candidate test formulation into the optimized Gaussian mixture model, calculate the probability density value, and determine the comprehensive matching degree between the candidate test formulation and the historical formulation based on the probability density value.

[0075] Based on the optimized GMM model, the characterization indicators of the candidate experimental formulations are input, and the probability distribution values ​​are calculated. Combining the variance explained by PCA, the overall matching degree between the candidate experimental formulations and the characterization indicators of historical tobacco leaf group formulations is calculated.

[0076] The above-mentioned method for calculating the matching degree of cigarette leaf group formula based on the improved Gaussian mixture model obtains historical formula characterization data and constructs a Gaussian mixture model to fit the complex distribution characteristics within the formula data. The Gaussian mixture model is improved by a dynamic weight adjustment method based on an adaptive attention mechanism. The matching degree between the selected experimental formula and the historical distribution is dynamically evaluated, realizing the quantitative optimization of the selected formula. This solves the problems of existing technologies relying on subjective experience, difficulty in quantifying indicators, and lack of systematic utilization of historical data. It achieves objective quantification of formula matching degree, thereby maximizing the maintenance of the continuity of formula style and quality.

[0077] In one exemplary embodiment, such as Figure 3 As shown, the PCA dimensionality reduction method is used for feature extraction on the dataset, mainly including steps S301 to S304. Wherein:

[0078] S301, calculate the covariance matrix of the historical leaf group formula dataset, perform eigenvalue decomposition, and obtain the eigenvalues ​​and corresponding eigenvectors.

[0079] Before calculating the covariance matrix, the collected indicator dataset undergoes data preprocessing to ensure data integrity and reliability. Preprocessing includes data standardization and outlier removal, resulting in a preprocessed dataset. The covariance matrix is ​​then calculated from the dataset, and eigenvalue decomposition is performed to obtain the cumulative variance contribution rate and eigenvalues. and the corresponding feature vectors , where m represents the number of sample features.

[0080] S302, calculate the cumulative variance contribution rate of the feature values, and retain the key feature values ​​and corresponding key feature vectors whose cumulative variance contribution rate reaches the preset threshold.

[0081] For each of the above eigenvalues Calculate its cumulative variance contribution rate:

[0082] ;

[0083] It is understandable that the cumulative variance contribution rate can reflect the importance of this feature value to the main information of the original data. This is combined with a preset threshold for the cumulative variance contribution rate. , select to make The minimum value of j, denoted as p, represents the selection of the first p principal components. These principal components constitute the key features, including key feature values ​​and corresponding key feature vectors.

[0084] S303 projects the historical leaf group formula dataset onto a subspace composed of key feature vectors to generate a dimensionality-reduced dataset, which serves as the feature dataset.

[0085] By projecting the historical leaf group formula dataset onto the subspace formed by the selected first p feature vectors, the dimensionality-reduced dataset can be obtained. ,in ,and .

[0086] S304. Project the candidate test formulation data onto the subspace formed by the key feature vectors to obtain the candidate test formulation feature data.

[0087] Understandably, to maintain consistency in the data processing flow, if the candidate experimental formulation data needs to be input into a Gaussian mixture model fitted from the historical leaf group formulation dataset, it needs to undergo the same feature dimensionality reduction process as the historical leaf group formulation data. Therefore, when reducing the dimensionality of the candidate experimental formulation data, the same set of key feature vectors obtained from historical data needs to be used to project the candidate experimental formulation data into the same low-dimensional subspace, thereby obtaining the candidate experimental formulation feature data, which provides input for subsequent matching degree calculation.

[0088] In this embodiment, the PCA-based feature extraction process described above can convert high-dimensional original indices that may have multicollinearity into a few independent principal components, which greatly alleviates the curse of dimensionality and lays the foundation for the subsequent construction of Gaussian mixture models and the calculation of recipe matching degree.

[0089] In an exemplary embodiment, a Gaussian mixture model is constructed by weighted superposition of multiple Gaussian components. Optimal parameter estimation is performed on the Gaussian mixture model based on a feature dataset, enabling the Gaussian mixture model to fit the probability distribution of the historical leaf group formulation's characterization data. This can be implemented as follows:

[0090] First, a Gaussian mixture model is established, initializing the model parameters and the number of Gaussian components. The model parameters include the weights, mean, and covariance of the Gaussian components. Specifically, the probability density function of the Gaussian mixture model can be expressed as:

[0091] ;

[0092] Where x is each data point, and K is the number of Gaussian components in the Gaussian mixture model. Let be the weights of the k-th Gaussian distribution, and satisfy . . The probability density function of the k-th Gaussian distribution is defined as:

[0093] ;

[0094] in, Let be the mean of the k-th Gaussian distribution. For covariance.

[0095] Determine the range of Gaussian component numbers and randomly initialize parameters. For each feature dimension j, set the number of Gaussian components. For each possible The values ​​are randomly initialized for the GMM parameters of the j-th feature, including the mixture weights. ,satisfy and Mean Covariance Matrix In the one-dimensional case, the covariance matrix is ​​the variance. .

[0096] Next, based on the feature dataset, the expectation-maximization algorithm is used to iteratively optimize the model parameters until the model convergence condition is met.

[0097] The Expectation-Maximization (EM) algorithm is an optimization algorithm that performs maximum likelihood estimation through iterative optimization, consisting of alternating expectation calculation steps (E-steps) and parameter maximization steps (M-steps). Specifically, in an exemplary embodiment, iteratively optimizing the model parameters using the EM algorithm includes the following steps:

[0098] E-step: For each feature dimension, calculate the posterior probability that each data point in the feature dataset belongs to each Gaussian component.

[0099] Specifically, for the j-th feature, calculate the value of each data point. Posterior probabilities belonging to each Gaussian component The formula is:

[0100]

[0101] in, The mean is The covariance matrix is The multidimensional Gaussian distribution, in one dimension, is:

[0102] .

[0103] M-step: Update the weights, mean, and covariance of the Gaussian components based on the posterior probability.

[0104] The above parameters are calculated using the following formulas:

[0105] Update the mixed weights: ;

[0106] Update the mean: ;

[0107] Update the covariance matrix. For the one-dimensional case, the update method is as follows: The multidimensional case is as follows: .

[0108] Repeat the above steps until the model convergence condition is met. Specifically, the convergence condition can be set as the change in parameters being less than a preset threshold, for example... Where t represents the number of iterations. It is a very small positive number.

[0109] Preferably, some embodiments further include tuning the GMM model through cross-validation. For example, for each feature j dimension, 10-fold cross-validation is used in different... Evaluate model performance using the given values. Divide the dataset into training and validation sets, and for each... The log-likelihood value is calculated by training a GMM model on the training set and then calculating the log-likelihood value on the validation set. The model is selected to maximize the log-likelihood value on the validation set. The optimal number of Gaussian components for this dimension of features.

[0110] In one exemplary embodiment, such as Figure 4As shown, the characteristic data of the candidate experimental formulations are input into the Gaussian mixture model. Using a dynamic weight adjustment method based on an adaptive attention mechanism, the weights of the Gaussian components in the Gaussian mixture model are recalculated for the characteristic data of the candidate experimental formulations, resulting in an optimized Gaussian mixture model. This process includes the following steps S401 to S403:

[0111] S401, input the characteristic data of the candidate test formulation into the Gaussian mixture model, and calculate the distance between the characteristic data of the candidate test formulation and the mean of each Gaussian component.

[0112] For experimental candidate formulations First, we perform data standardization by taking the mean and variance of the data, which yields the following results. Next, following the same method as the feature extraction step in step S304, the first p feature vectors can be obtained. ,right Projecting, we obtain The formula for calculating projection is:

[0113] ;

[0114] in, , This refers to the sample size for the experimental formulation index data. , where p is the number of key features of the experimental formulation after feature extraction. (Projected) ,in, .

[0115] S402, based on the distance between the characteristic data of the candidate experimental formulation and the mean of each Gaussian component, the adaptive attention weight of the Gaussian component is calculated through the adaptive attention function.

[0116] For example, the adaptively adjusted attention weights can be calculated according to the following formula:

[0117] ;

[0118] in, The attention weight of the j-th feature of data point x in the candidate experimental formulation features in the k-th Gaussian component. Let be the mean of the k-th Gaussian component, and K be the total number of Gaussian components. It is a hyperparameter that can be set or optimized through cross-validation, and Attention weights can prevent certain "boundary points" from being evenly distributed among multiple Gaussian components, thus improving the model's representation performance.

[0119] S403 uses adaptive attention weights to replace the weights of the Gaussian components in the Gaussian mixture model, generating a Gaussian mixture model optimized for the feature data of the candidate experimental formulations.

[0120] Adaptively adjusted attention score As weights for the Gaussian components, attention scores are used in the trained Gaussian mixture model. Instead of fixed weights in traditional Gaussian mixture models It can achieve competitive weight adjustment between components, improve the generalization of the model, and reduce overfitting.

[0121] In the above embodiments, by introducing a dynamic weight adjustment method based on an adaptive attention mechanism, the problem of insufficient handling of boundary points in the traditional fixed weight model is avoided, the model's ability to represent complex distributed data is enhanced, the risk of overfitting the model to the training data is reduced, and the generalization performance is improved.

[0122] In one exemplary embodiment, the characteristic data of the candidate experimental formulation is input into an optimized Gaussian mixture model to calculate the probability density value. Based on the probability density value, the overall matching degree between the candidate experimental formulation and historical formulations is determined, such as... Figure 5 As shown, the process includes the following steps S501 to S503:

[0123] S501, calculate the feature weight of each feature dimension in the feature data of the candidate experimental formulation with respect to all features.

[0124] Calculate the GMM weight for each feature of the characteristic data points of the candidate experimental formulations. ,in is the feature value corresponding to the j-th feature, and p is the number of features after feature extraction and screening of candidate experimental formulations. Preferably, PCA dimensionality reduction algorithm is used during feature extraction, resulting in... is the eigenvalue corresponding to the j-th principal component, and p is the number of principal components selected within the threshold.

[0125] S502, for each feature dimension, the local probability density value of each feature dimension is obtained based on the optimized Gaussian mixture model.

[0126] By inputting each feature dimension into the Gaussian mixture model optimized using the adaptive attention mechanism, we can obtain the local probability density value of each feature dimension:

[0127] ;

[0128] in, It is the number of Gaussian components after the j-th dimension feature is optimized. For the optimized Gaussian component weights, The mean is The covariance matrix is The multidimensional Gaussian distribution.

[0129] S503, based on feature weights, weights and sums the local probability density values ​​to obtain the overall matching degree between the experimental formula and the historical formula.

[0130] The formula for calculating the overall matching degree M is:

[0131] ;

[0132] Where M represents the overall matching degree, Let be the feature weight of the j-th dimension, and p be the number of features. The overall matching degree M reflects the degree of matching between the experimental formulation and the historical formulation design of this specification in each feature dimension, with weights... This indicates the importance of each feature in the overall fit. The evaluation results of the overall fit are used to optimize the experimental formulation.

[0133] This embodiment ensures that principal components with higher effective information content have a higher proportion in the final decision by assigning weights based on variance explanation rate to different feature dimensions, and effectively improves the robustness and reliability of the evaluation under complex data distribution by obtaining local probability density values ​​based on the optimized Gaussian mixture model.

[0134] The following provides a preferred embodiment for detailed description, such as... Figure 6 As shown, this embodiment includes steps S1 to S6:

[0135] S1. Obtain characterization data of historical leaf group formulations for the target tobacco leaf specification and construct a historical formulation dataset.

[0136] We collected and organized 20 indicators, including conventional chemical and derived indicators, near-infrared model indicators, and aging time, from the continuous production batches of a certain cigarette specification A over the past three years, and established a characterization database for cigarette specification A.

[0137] The continuous formulation batch index data collected in this embodiment includes chemical composition indicators: water-soluble total sugar, total alkaloids, reducing sugar, chlorine content, potassium content, total nitrogen, starch, pH value, and chlorogenic acid; conventional chemical composition derived indicators: sugar-to-nitrogen ratio, sugar-to-alkali ratio, nitrogen-to-alkali ratio, and potassium-to-chlorine ratio; near-infrared model prediction indicators: part index, light aroma index, intermediate aroma index, strong aroma index, and elegance index; and aging time. The aging time is obtained by weighting the aging time of each formulation component tobacco leaf according to the leaf group formulation design ratio.

[0138] The collected indicator dataset undergoes data preprocessing to ensure data integrity and reliability. Preprocessing includes data standardization and outlier removal. The standardization method for the dataset is as follows. For each feature... Calculate its mean and standard deviation Then each data point Convert to To obtain the standardized dataset Table 1 shows the statistical indicators of the dataset based on historical leaf group formulations.

[0139] Table 1

[0140]

[0141] S2, extract features from the historical formula dataset to obtain the feature dataset.

[0142] PCA dimensionality reduction is performed on dataset Z to extract key feature information. Based on a set cumulative variance contribution rate threshold and the number of principal components selected, a new dimensionality-reduced index dataset is constructed.

[0143] Eigenvalue decomposition is performed on the covariance matrix S to obtain the cumulative variance contribution rate and eigenvalues. and the corresponding feature vectors The principal component eigenvalues ​​and variance contribution rates after feature extraction are shown in Table 2.

[0144] Table 2

[0145]

[0146] Furthermore, the selection of such that The minimum value of j, denoted as p, is the first p principal components selected, where the preset contribution rate threshold is... It is 90%.

[0147] Furthermore, the standardized dataset Project onto the selected front In the subspace formed by the eigenvectors, the dimensionality-reduced values ​​are obtained. Principal component datasets ,in ,and The first p feature components of the dataset Z after projection are shown in Table 3.

[0148] Table 3

[0149]

[0150] S3, based on the feature dataset, uses a Gaussian mixture model for modeling and optimal parameter estimation, so that the Gaussian mixture model fits the probability distribution of the characterization data of the historical leaf group formulation.

[0151] The specific calculation method is the same as that in the above embodiment, a Gaussian mixture model composed of multiple Gaussian components is constructed by weighted superposition. The optimal parameters of the Gaussian mixture model are estimated based on the feature dataset so that the probability distribution of the Gaussian mixture model fitting the characterization data of the historical leaf group formula is the same. It will not be repeated here.

[0152] The EM algorithm is used for iterative computation until the model convergence condition is met. Then, the GMM model is tuned using 5-fold cross-validation. For each feature j, under different... To evaluate model performance, select the value that maximizes the log-likelihood of the validation set. The optimal number of Gaussian components for this feature dimension is determined. After modeling and tuning the historical formula dataset, the parameters of the Gaussian mixture model are shown in Table 4 below.

[0153] Table 4

[0154]

[0155] S4. Based on the same feature dimensionality reduction method, feature projection calculation is performed on the dataset of the selected experimental formulations.

[0156] For the experimental candidate formulation, the projected result is obtained according to the calculation method in step S401 above. The characteristic projection calculation results of the experimental formulation are shown in Table 5.

[0157] Table 5

[0158]

[0159] S5 uses a dynamic weight adjustment method based on an adaptive attention mechanism to recalculate the weights of the Gaussian components in the Gaussian mixture model for the feature data of the candidate experimental formulations.

[0160] Based on the Gaussian mixture model parameters mentioned above, for each experimental formulation, the adaptive attention score under each principal component dimension is calculated. As weights for each Gaussian component. Adaptively adjusted attention weights: Table 6 shows the weight of each Gaussian component under each principal component dimension of the three samples, taking experimental formulations N1-N3 as examples. Figures 7(a), 7(b), 7(c), and 7(d) show the probability distribution of different principal components of the feature dataset based on the Gaussian mixture model in this embodiment, reflecting the internal data structure distribution of different comprehensive components after dimensionality reduction of the leaf group formulation characterization index of historical formulation data.

[0161] Table 6

[0162]

[0163] S6, calculate the GMM model weights for each feature based on the variance explanation ratio of PCA.

[0164] Based on the variance explanation ratio of PCA, the GMM model weights of the new features extracted after dimensionality reduction by PCA are calculated, and the weights are shown in Table 7.

[0165] Table 7

[0166]

[0167] In the improved Gaussian mixture model, the standardized candidate experimental formulation dataset, the new Gaussian component weights, and the GMM model weights for each feature are re-inputted to calculate the comprehensive matching degree between the candidate experimental formulations and the historical tobacco leaf group formulation characterization indicators. The calculation method is the same as steps S501 to S503, and will not be repeated here. Finally, the comprehensive matching degrees between multiple candidate experimental formulations and historical formulations are obtained, as shown in Table 8.

[0168] Table 8

[0169]

[0170] As can be seen from the results in Table 8, by mining the distribution of the internal data structure of the historical formula, the overall matching degree between the experimental formula and the leaf group of the cigarette formula of this brand and specification can be quantified. In addition, the adjustment measures of the leaf group formula in terms of tobacco grade substitution strategy and formula structure can be evaluated, which can maximize the maintenance of the continuity of formula style and quality.

[0171] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0172] Based on the same inventive concept, this application also provides an apparatus for implementing the above-mentioned method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model. The solution provided by this apparatus is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the apparatus for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model provided below can be found in the limitations of the method for calculating the matching degree of cigarette leaf blend formulation based on an improved Gaussian mixture model described above, and will not be repeated here.

[0173] In one exemplary embodiment, such as Figure 8 As shown, a tobacco leaf blend formulation matching degree calculation device 600 based on an improved Gaussian mixture model is provided, including: a historical data acquisition module 601, a historical data feature extraction module 602, a model training module 603, a model optimization module 604, a matching degree calculation module 605, and a formulation selection module 606, wherein:

[0174] The historical data acquisition module 601 is used to acquire characterization data of historical leaf group formulas for the target tobacco leaf specifications and construct a historical formula dataset.

[0175] The historical data feature extraction module 602 is used to extract features from the historical formula dataset to obtain a feature dataset;

[0176] The model training module 603 is used to construct a Gaussian mixture model composed of weighted superposition of multiple Gaussian components, and to perform optimal parameter estimation on the Gaussian mixture model based on the feature dataset, so that the Gaussian mixture model fits the probability distribution of the characterization data of the historical leaf group formula.

[0177] The model optimization module 604 is used to extract features from the characterization data of the candidate experimental formulation. The obtained candidate experimental formulation feature data is input into the Gaussian mixture model. Through a dynamic weight adjustment method based on an adaptive attention mechanism, the weights of the Gaussian components in the Gaussian mixture model are recalculated for the candidate experimental formulation feature data to obtain the optimized Gaussian mixture model.

[0178] The matching degree calculation module 605 is used to perform feature dimensionality reduction on the candidate experimental formula dataset and input it into the trained Gaussian mixture model. Through a dynamic weight adjustment method based on attention mechanism, the module calculates the comprehensive matching degree between the candidate experimental formula and the historical formula.

[0179] The formulation selection module 606 is used to input the feature data of the candidate experimental formulation into the optimized Gaussian mixture model, calculate the probability density value, and determine the comprehensive matching degree between the candidate experimental formulation and the historical formulation based on the probability density value.

[0180] The modules in the aforementioned tobacco leaf blend matching degree calculation device based on the improved Gaussian mixture model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0181] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0182] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0183] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0184] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for calculating the matching degree of a tobacco leaf group formula based on an improved Gaussian mixture model, characterized in that, The method comprises: obtaining the historical leaf group formula data of the target tobacco leaf product, and constructing a historical formula data set; extracting features from the historical formula data set to obtain a feature data set; constructing a Gaussian mixture model composed of multiple Gaussian components weighted and superimposed, and performing optimal parameter estimation on the Gaussian mixture model based on the feature data set, so that the Gaussian mixture model fits the probability distribution of the historical leaf group formula data; extracting features from the feature data of the to-be-selected test formula, inputting the obtained feature data of the to-be-selected test formula into the Gaussian mixture model, and recalculating the weights of the Gaussian components in the Gaussian mixture model for the feature data of the to-be-selected test formula by using a dynamic weight adjustment method based on an adaptive attention mechanism to obtain an optimized Gaussian mixture model; inputting the feature data of the to-be-selected test formula into the optimized Gaussian mixture model, calculating a probability density value, and determining the comprehensive matching degree between the to-be-selected test formula and the historical formula based on the probability density value.

2. The method of claim 1, wherein, The method comprises: establishing the Gaussian mixture model, initializing model parameters and the number of Gaussian components, and the model parameters include the weights, means and covariances of the Gaussian components; iteratively optimizing the model parameters based on the feature data set using an expectation maximization algorithm until the model convergence condition is met.

3. The method of claim 2, wherein, The method comprises: for each dimension of features, calculating the posterior probability of each data point in the feature data set belonging to each Gaussian component; updating the weights, means and covariances of the Gaussian components according to the posterior probability; repeating the above steps until the model convergence condition is met.

4. The method of claim 1, wherein, The method comprises: inputting the feature data of the to-be-selected test formula into the Gaussian mixture model, calculating the distance between the feature data of the to-be-selected test formula and the mean of each Gaussian component, calculating the adaptive attention weight of each Gaussian component through an adaptive attention function according to the distance, and using the adaptive attention weight to replace the weight of each Gaussian component in the Gaussian mixture model to generate an optimized Gaussian mixture model for the feature data of the to-be-selected test formula. The adaptive attention function is calculated by the following formula: The method comprises:

5. The method of claim 4, wherein, calculating the covariance matrix of the historical leaf group formula data set, performing eigenvalue decomposition to obtain eigenvalues and corresponding eigenvectors, and ； wherein, is the attention weight of the jth feature of the candidate test formulation feature data x in the kth Gaussian component, is the mean of the kth Gaussian component, K is the total number of Gaussian components, is a hyperparameter that can be set or optimized through cross-validation, and .

6. The method of claim 1, wherein, ​ ​ Calculate the cumulative variance contribution rate of the feature values, and retain the key feature values ​​and corresponding key feature vectors whose cumulative variance contribution rates reach a preset threshold. The historical leaf group formula dataset is projected onto the subspace formed by the key feature vectors to generate a dimensionality-reduced dataset, which is used as the feature dataset. The characterization data of the candidate test formulations are projected onto the subspace formed by the key feature vectors to obtain the feature data of the candidate test formulations.

7. The method of claim 1, wherein, The step of inputting the characteristic data of the candidate experimental formulation into the optimized Gaussian mixture model, calculating the probability density value, and determining the comprehensive matching degree between the candidate experimental formulation and the historical formulation based on the probability density value includes: Calculate the feature weight of each feature dimension in the candidate experimental formulation feature data for all features; For each of the features, the local probability density value of each feature is obtained based on the optimized Gaussian mixture model; Based on the feature weights, the local probability density values ​​are weighted and summed to obtain the overall matching degree between the experimental formula and the historical formula.

8. The method of claim 7, wherein, The formula for calculating the overall matching degree is: ； ； Wherein, M is the comprehensive matching degree, is the feature weight of the jth dimension feature, and p is the feature quantity, is the number of Gaussian components of the optimized Gaussian mixture model, is the weight of the jth dimension feature in the kth Gaussian component of the optimized Gaussian mixture model, is a multidimensional Gaussian distribution with mean and covariance matrix .

9. A device for calculating a degree of matching of a tobacco leaf blend recipe based on an improved Gaussian mixture model, characterized by, The device includes: The historical data acquisition module is used to acquire characterization data of historical leaf group formulas for the target tobacco leaf specifications and construct a historical formula dataset. The historical data feature extraction module is used to extract features from the historical formula dataset to obtain a feature dataset; The model training module is used to construct a Gaussian mixture model composed of weighted superposition of multiple Gaussian components, and to perform optimal parameter estimation on the Gaussian mixture model based on the feature dataset, so that the Gaussian mixture model fits the probability distribution of the characterization data of the historical leaf group formula. The model optimization module is used to extract features from the characterization data of the candidate experimental formulation. The obtained feature data of the candidate experimental formulation is input into the Gaussian mixture model. Through a dynamic weight adjustment method based on an adaptive attention mechanism, the weights of the Gaussian components in the Gaussian mixture model are recalculated for the feature data of the candidate experimental formulation to obtain the optimized Gaussian mixture model. The matching degree calculation module is used to perform feature dimensionality reduction on the candidate experimental formula dataset and input it into the trained Gaussian mixture model. Through a dynamic weight adjustment method based on the attention mechanism, the module calculates the comprehensive matching degree between the candidate experimental formula and the historical formula. The formulation selection module is used to input the feature data of the candidate experimental formulation into the optimized Gaussian mixture model, calculate the probability density value, and determine the comprehensive matching degree between the candidate experimental formulation and the historical formulation based on the probability density value.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images