Methods, apparatus, equipment and storage media for screening tobacco leaves

By using a classification proxy model and constrained multi-objective optimization techniques, the problems of lost confidence differences and insufficient risk constraints in the screening of tobacco raw materials were solved, generating a traceable candidate list and improving the usability and stability of the screening results.

CN123100863APending Publication Date: 2026-11-17CHINA TOBACCO YUNNAN IND
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611320218.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, when screening tobacco raw materials using near-infrared chemical composition data, the confidence differences of candidates at different grades are lost, resulting in candidates at the grade boundary being misjudged or missed due to hard classification truncation. At the same time, screening based solely on a single core sensory indicator lacks a constraint mechanism on risk indicators such as irritation and off-odors, making it difficult to guarantee the overall usability of the screening results.

Method used

A classification proxy model is used to output the probability of ordered sensory levels. The core sensory performance and risk sensory confidence are quantified by the expected level and the probability of level satisfaction, respectively. Under the constrained multi-objective optimization framework, the optimal core sensory indicators are pursued simultaneously while the risk indicators are controlled to not exceed the limits. Offspring candidates that do not meet the constraints are gradually repaired through feasible anchor points to generate a traceable list of candidates to be reviewed.

Benefits of technology

While narrowing the scope of expert sensory review, it improves the stability and reliability of tobacco raw material screening, ensures the optimization goals of core sensory indicators and the constraints of risk indicators, and the output candidate list can be efficiently reviewed by experts.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123100863A_ABST
    Figure CN123100863A_ABST
Patent Text Reader

Abstract

This application discloses a method, apparatus, equipment, and storage medium for screening tobacco raw materials, relating to the field of tobacco technology. This method addresses the problems of hard classification truncation leading to lost confidence levels and lack of risk sensory constraints in tobacco raw material screening. It employs a classification surrogate model to output ordered sensory level probabilities, quantifying core sensory performance and risk sensory confidence levels through expected levels and level satisfaction probabilities, respectively. Under a constrained multi-objective optimization framework, it simultaneously pursues the optimal core sensory indicators while controlling risk indicators to avoid exceeding limits. For offspring candidates that do not meet the constraints, it gradually repairs and recovers some infeasible solutions through feasible anchor points. The optimization results are mapped to real historical tobacco samples to output a traceable list of candidates to be reviewed, thus narrowing the scope of expert sensory review while improving the stability of candidate screening under strict constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of tobacco technology, and in particular to a method, apparatus, equipment and storage medium for screening tobacco leaves. Background Technology

[0002] Tobacco raw materials are the foundation of the style and quality stability of cigarette products. Before tobacco leaves are put into storage and selected for formulation, quality evaluation is usually carried out by combining near-infrared detection and expert sensory evaluation.

[0003] Near-infrared spectroscopy (NIRS) acquires spectral information from tobacco leaf samples and, after conversion by a calibration model, yields numerical values ​​for chemical components such as total alkaloids, sugars, nitrogenous compounds, and organic acids. It features fast detection speed, non-destructive testing, and high-throughput operation, making it suitable for preliminary screening of a large number of candidate raw materials.

[0004] Sensory evaluation by experts assesses the actual use value of tobacco leaves based on indicators such as aroma quality, aroma quantity, aftertaste, irritation, and off-flavors. However, the evaluation is costly and time-consuming, and is affected by personnel condition and environmental conditions, making it difficult to conduct high-frequency verification of a large number of candidate raw materials.

[0005] Therefore, in the formulation and material selection process, candidate raw materials are usually screened first using near-infrared chemical composition data, and then a small number of candidates are reviewed by experts to narrow down the scope of sensory evaluation.

[0006] In existing technologies, when screening tobacco raw materials using near-infrared chemical composition data, a classification model is usually used to predict the sensory grade, and the grade with the highest probability is selected as the candidate sensory evaluation result for screening.

[0007] However, the maximum probability level output by the classification model only reflects the most likely category, losing the confidence difference of candidates at different levels. This leads to candidates at the level boundary being misjudged or missed due to hard classification truncation. At the same time, sorting and screening based on a single core sensory indicator lacks a constraint mechanism on risk indicators such as stimulation and off-gassing, making it easy to select candidates with better core indicators but higher risk indicators, making it difficult to guarantee the overall usability of the screening results. Summary of the Invention

[0008] The main objective of this application is to provide a method, apparatus, equipment, and storage medium for screening tobacco raw materials, in order to solve the problems in the prior art when screening tobacco raw materials using near-infrared chemical composition data. These problems include the loss of confidence differences in the judgment of candidates at different grades, which leads to the misjudgment or omission of candidates at the grade boundary due to hard classification truncation; at the same time, the screening based on a single core sensory index lacks a constraint mechanism on risk indicators such as irritation and off-odors, which easily selects candidates with better core indicators but higher risk indicators, making it difficult to guarantee the overall usability of the screening results.

[0009] To achieve the above objectives, this application provides the following technical solution: A method for screening tobacco leaves as raw material, the method comprising: Step S1: Obtain chemical composition data and sensory evaluation data from near-infrared spectroscopy of historical tobacco leaf samples, and match and preprocess the chemical composition data and the sensory evaluation data. Step S2: Generate ordered sensory level labels based on the sensory evaluation data, and train classification proxy models that output ordered sensory level probabilities for core sensory indicators and risk sensory indicators, respectively, using the chemical composition data as input. Step S3: Input the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model, calculate the expected level of the core sensory index based on the predicted probability and level order value of each ordered sensory level, and calculate the level satisfaction probability of the risk sensory index based on the sum of the predicted probabilities of not less than the preset minimum acceptable level. Step S4: Using the expected level of multiple core sensory indicators as the optimization objective and the probability of satisfying the level of the risk sensory indicator not being lower than the corresponding probability threshold as the constraint, construct a constrained multi-objective optimization problem, and generate an initial candidate population based on the chemical composition data, the sensory evaluation data, and the preset chemical composition value range. Step S5: Perform crossover and mutation on the initial candidate population to generate offspring candidates. For offspring candidates that do not meet the constraints, select the nearest feasible anchor point from the feasible file and repair them step by step according to the preset mixing coefficient sequence. Step S6: Perform non-dominated sorting on the candidates that meet the constraints to obtain a feasible Pareto candidate set, map the feasible Pareto candidate set to the historical tobacco leaf samples, and output a candidate list to be reviewed.

[0010] Beneficial effects of steps S1 to S6: This method addresses the issues of lost confidence levels and lack of risk sensory constraints in the screening of tobacco raw materials due to hard classification truncation. It employs a classification surrogate model to output ordered sensory level probabilities. The expected level and the level satisfaction probability are used to quantify the core sensory performance and the risk sensory confidence level, respectively. Under a constrained multi-objective optimization framework, the optimal core sensory indicators are pursued simultaneously while controlling the risk indicators to avoid exceeding the limits. For offspring candidates that do not meet the constraints, some infeasible solutions are gradually repaired and restored through feasible anchor points. The optimization results are mapped to real historical tobacco samples to output a traceable list of candidates to be reviewed. This reduces the scope of expert sensory review while improving the stability of candidate screening under strict constraints.

[0011] Step S1 provides a unified data foundation for training the surrogate model by matching and preprocessing the chemical composition data and sensory evaluation data of historical tobacco leaf samples; Step S2 preserves the confidence distribution of candidates at different levels by training a classification surrogate model that outputs ordered sensory level probabilities; Step S3 characterizes the core sensory performance and risk sensory constraints by using expected level and level satisfaction probability, respectively, to avoid misjudgment of boundary candidates due to hard classification truncation; Step S4 generates an initial candidate population by constructing a constrained multi-objective optimization problem, while constraining risk indicators while optimizing core sensory indicators; Step S5 gradually repairs infeasible offspring candidates by using feasible anchors, expanding feasible search paths and improving population utilization; Step S6 transforms virtual candidates into a traceable list of candidates to be reviewed by non-dominated sorting and real sample mapping.

[0012] As a further improvement to this application, step S1 involves acquiring chemical composition data and sensory evaluation data from near-infrared spectroscopy of historical tobacco leaf samples, and matching and preprocessing the chemical composition data and the sensory evaluation data, including: Step S1.1: Obtain near-infrared spectral data of historical tobacco leaf samples, and convert the near-infrared spectral data into chemical composition values ​​through a preset chemical composition correction model; Step S1.2: Obtain sensory evaluation data of the historical tobacco leaf samples, and match the chemical composition data with the sensory evaluation data according to the sample identifier; Step S1.3: Perform missing value detection on the matched data, fill in the missing chemical composition field with the median, and fill in the missing sensory evaluation field with the mean. Step S1.4: Detect outliers in the chemical composition data and mark values ​​that exceed the preset range of chemical composition values ​​as invalid. Step S1.5: Linear interpolation is used to compensate for fields marked as invalid, and Min-Max normalization is performed on the compensated chemical composition data and sensory evaluation data. Step S1.6: Arrange the normalized chemical composition data and sensory evaluation data according to the samples to obtain the preprocessed chemical composition data and sensory evaluation data.

[0013] Beneficial effects of steps S1.1 to S1.6: This series of steps addresses the issue of heterogeneous sources and inconsistent quality of historical tobacco sample chemical composition data and sensory evaluation data. Through near-infrared spectral conversion, sample label matching, missing value imputation, outlier detection, linear interpolation compensation, and normalization, preprocessed data of a uniform scale is formed, providing a reliable data foundation for surrogate model training.

[0014] The process includes the following steps: Step S1.1 converts near-infrared spectral data into chemical composition values ​​using a pre-defined chemical composition correction model, achieving lossless high-throughput data acquisition; Step S1.2 establishes a correspondence between chemical composition data and sensory evaluation data through sample identifier matching, eliminating the risk of cross-source data mismatch; Step S1.3 fills in missing fields for chemical composition and sensory evaluation using the median and mean, respectively, reducing the interference of missing values ​​on model training; Step S1.4 detects and marks outliers using a pre-defined value range, identifying abnormal records during data acquisition; Step S1.5 compensates for invalid fields using linear interpolation and performs Min-Max normalization to eliminate dimensional differences among different chemical components; and Step S1.6 arranges and organizes samples to form structured preprocessed data, providing a unified input for subsequent model training and candidate evaluation.

[0015] As a further improvement to this application, step S2 involves generating ordered sensory level labels based on the sensory evaluation data, and using the chemical composition data as input, training classification proxy models that output ordered sensory level probabilities for core sensory indicators and risk sensory indicators respectively, including: Step S2.1: Determine the level threshold based on the score distribution of each sensory index in the sensory evaluation data, and divide the sensory scores into three ordered sensory levels: low level, medium level, and high level. Step S2.2: The intermediate level is split into low-medium level and high-medium level to form four ordered levels. Then, the low-medium level and the high-medium level are merged into the intermediate level to obtain three ordered sensory level labels: Q1, Q2, Q3, and Q4. Step S2.3: Determine the core sensory indicators and risk sensory indicators. The core sensory indicators are aroma quality, aroma quantity, and aftertaste. The risk sensory indicators are irritation and off-odors. Step S2.4: Using the chemical composition data as input features, construct training sets and validation sets for the core sensory indicators and the risk sensory indicators, respectively. Step S2.5: Train a random forest classification model for each sensory indicator. The random forest classification model outputs the predicted probability of each ordered sensory level, thus obtaining a classification surrogate model.

[0016] Beneficial effects of steps S2.1 to S2.5: This series of steps addresses the issues of coarse sensory evaluation data classification and lack of targeted training of surrogate models. It divides sensory level labels into ordered categories based on score distribution, and trains random forest classification models for core sensory indicators and risk sensory indicators respectively, providing a probabilistic output basis for subsequent expected level calculation and level satisfaction probability assessment.

[0017] Specifically, step S2.1 determines the level threshold by rating distribution to divide ordered sensory levels, transforming continuous sensory ratings into learnable discrete labels; step S2.2 splits and merges intermediate levels through quartiles to form three ordered labels Q1, Q2, Q3, and Q4, retaining the level order information for subsequent expected level calculation; step S2.3 clarifies core sensory indicators and risk sensory indicators, providing modeling objects for objective and constraint functions for constrained multi-objective optimization; step S2.4 constructs training and validation sets respectively, providing data support for the independent training and validation of the classification proxy model; and step S2.5 trains a random forest classification model to output the predicted probability of each ordered sensory level, retaining the confidence distribution of candidates at different levels.

[0018] As a further improvement to this application, step S3 involves inputting the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model, calculating the expected level of the core sensory indicator based on the predicted probability and level ordinal value of each ordered sensory level, and calculating the level satisfaction probability of the risk sensory indicator based on the sum of predicted probabilities not lower than the preset minimum acceptable level, including: Step S3.1: Input the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model to obtain the predicted probability of each sensory index under each ordered sensory level; Step S3.2: For the core sensory index, multiply the predicted probability of each ordered sensory level by the corresponding level ordinal value and sum them to obtain the expected level of the core sensory index. Step S3.3: For the risk sensory index, determine a preset minimum acceptable level, and sum the predicted probabilities of each level that is not lower than the minimum acceptable level to obtain the level satisfaction probability of the risk sensory index. Step S3.4: When the probability of the grade is not lower than the corresponding probability threshold, the candidate corresponding to the chemical composition vector of the candidate tobacco raw material is marked as a feasible candidate; otherwise, it is marked as an infeasible candidate.

[0019] Beneficial effects of steps S3.1 to S3.4: This series of steps addresses the issues of confidence discrepancies in lost levels due to hard classification truncation and the lack of risk quantification methods. It outputs the predicted probability of ordered sensory levels through a classification surrogate model, calculates the expected level of core sensory indicators and the probability of meeting the level of risk sensory indicators, and determines the feasibility of candidates using probability thresholds, thus providing target values ​​and constraint values ​​for constrained multi-objective optimization.

[0020] Specifically, step S3.1 involves inputting candidate chemical component vectors into a classification proxy model to obtain the predicted probabilities of each ordered sensory level, preserving the complete confidence distribution of candidates in level judgment; step S3.2 involves calculating the expected level by weighted summation of the predicted probabilities and level ordinal values, using continuous values ​​to characterize core sensory performance and avoid hard classification truncation; step S3.3 involves calculating the level satisfaction probability by summing the predicted probabilities of levels not lower than the minimum acceptable level, quantifying the safety margin of risk sensory indicators; and step S3.4 involves determining whether a candidate is feasible or infeasible based on a probability threshold, transforming probability constraints into operable screening markers.

[0021] As a further improvement to this application, step S4 involves constructing a constrained multi-objective optimization problem with the desired levels of multiple core sensory indicators as the optimization objective and the constraint that the probability of satisfying the level of the risk sensory indicator is not lower than the corresponding probability threshold. An initial candidate population is then generated based on the chemical composition data, the sensory evaluation data, and a preset range of chemical composition values, including: Step S4.1: Take the expected level of the multiple core sensory indicators as the target to be maximized, and take the probability that the level of each risk sensory indicator is not lower than the corresponding probability threshold as the constraint condition to construct a constrained multi-objective optimization problem. Step S4.2: Extract a portion of samples from the historical samples corresponding to the chemical composition data and the sensory evaluation data as historical candidates; Step S4.3: Randomly generate restricted random candidates within a preset range of chemical composition values, and combine the historical candidates with the restricted random candidates to form an initial candidate population; Step S4.4: Input the chemical composition vector of each candidate in the initial candidate population into the classification proxy model, and calculate the expected level and level satisfaction probability of each candidate; Step S4.5: Sort the initial candidate population in a manner that feasible candidates take precedence over infeasible candidates, and perform non-dominated sorting and crowding distance calculation on the feasible candidates; Step S4.6: Select candidates from the initial candidate population to enter the next generation based on the non-dominated sorting results and crowding distance.

[0022] Beneficial effects of steps S4.1 to S4.6: This series of steps addresses the problem of mutual constraints between core sensory indicators and risk sensory indicators in multi-objective optimization. By constructing a constrained multi-objective optimization problem and generating a mixed initial candidate population, the steps involve probabilistic evaluation of the candidates and non-dominated ranking selection based on feasibility priority, thus providing a high-quality parent population for subsequent crossover and mutation.

[0023] Specifically, step S4.1 constructs a constrained multi-objective optimization problem by using the expected level as the target to be maximized and the probability of level satisfaction as the constraint, thus unifying the mathematical framework of core sensory pursuit and risk sensory control; step S4.2 extracts historical candidates from historical samples, retaining the real chemical composition combination structure as the starting point for the search; step S4.3 generates restricted random candidates within the value range and combines them with historical candidates, taking into account both the coverage and feasibility of the search space; step S4.4 uses a classification surrogate model to batch calculate the expected level and probability of level satisfaction of candidates, providing an evaluation basis for feasibility judgment and ranking; step S4.5 identifies the multi-objective optimal boundary among feasible candidates through feasible priority ranking, non-dominated ranking, and crowding distance calculation; step S4.6 selects candidates to enter the next generation based on the non-dominated ranking results and crowding distance, balancing convergence and diversity.

[0024] As a further improvement to this application, step S5 involves crossover and mutation of the initial candidate population to generate offspring candidates. For offspring candidates that do not meet the constraints, the nearest feasible anchor point is selected from the feasible file and repaired step by step according to a preset mixing coefficient sequence, including: Step S5.1: Perform a crossover operation on the selected candidates to generate intermediate candidates, and perform a mutation operation on the intermediate candidates to generate offspring candidates; Step S5.2: Based on the importance of the variables in the classification proxy model corresponding to the risk sensory index, determine the variation range of each chemical component, and constrain the variation range of the relevant chemical components to be less than the variation range of the other chemical components. Step S5.3: Input the offspring candidates into the classification proxy model and calculate the probability of each offspring candidate meeting the level requirement; Step S5.4: Determine whether the offspring candidate satisfies the constraint, and add the offspring candidate that satisfies the constraint to the feasible file; Step S5.5: For offspring candidates that do not meet the constraints, select the feasible candidate with the closest standardized distance from the feasible file as the feasible anchor point; Step S5.6: According to the preset mixing coefficient sequence arranged from largest to smallest, the offspring candidates are gradually mixed with the feasible anchor points. After each mixing, the level satisfaction probability is recalculated until a repair candidate that satisfies the constraint is obtained or the preset mixing coefficient sequence is completed.

[0025] Beneficial effects of steps S5.1 to S5.6: This series of steps addresses the problem of reduced search efficiency caused by directly discarding infeasible offspring in constrained multi-objective search. It generates offspring candidates by guiding differential mutation through variable importance. For offspring candidates that do not meet the constraints, the nearest feasible anchor point is selected from the feasible archive and repaired step by step according to the mixing coefficient sequence, thereby expanding feasible search paths and improving population utilization.

[0026] Specifically, step S5.1 generates offspring candidates by performing crossover and mutation on the selected candidates, introducing new candidate solutions into the search space; step S5.2 determines the variation amplitude of each chemical component by the importance of variables, reducing the perturbation amplitude of constraint-related components to reduce the feasibility of mutation destruction; step S5.3 calculates the level satisfaction probability of offspring candidates using a classification surrogate model, providing an evaluation basis for constraint determination; step S5.4 dynamically accumulates feasible reference points by judging whether the offspring meet the constraints and adding those that meet the constraints to the feasible file; step S5.5 selects the feasible candidate with the closest standardized distance from the feasible file as a feasible anchor point, providing a repair direction for infeasible offspring; and step S5.6 gradually mixes and re-evaluates the mixture by progressively mixing the mixing coefficient sequence from large to small, gradually approaching the feasible region while maintaining candidate diversity.

[0027] As a further improvement to this application, step S6 involves performing a non-dominated sort on the candidates that satisfy the constraints to obtain a feasible Pareto candidate set, mapping the feasible Pareto candidate set to the historical tobacco leaf samples, and outputting a list of candidates to be reviewed, including: Step S6.1: Perform non-dominated sorting on the candidates that satisfy the constraints to obtain feasible Pareto candidate sets at different non-dominated levels; Step S6.2: Calculate the mean of the expected levels of multiple core sensory indicators for each candidate in the feasible Pareto candidate set, and use it as the core sensory comprehensive score. Step S6.3: Select candidates whose core sensory comprehensive score reaches a preset percentile from the feasible Pareto candidate set to form a candidate pool; Step S6.4: Standardize the chemical composition vectors of each candidate in the candidate pool and the chemical composition data of the historical tobacco leaf samples, and calculate the Euclidean distance between each candidate and each historical tobacco leaf sample. Step S6.5: Select the historical tobacco leaf sample with the smallest Euclidean distance for each candidate as the mapping sample, and organize the desensitization identifier, core sensory expectation level, risk sensory level satisfaction probability, and review order of the mapping sample into a candidate list to be reviewed.

[0028] Beneficial effects of steps S6.1 to S6.5: This series of steps addresses the problem that virtual candidates in the optimized output are difficult to trace back to real tobacco leaf samples. It obtains a feasible Pareto candidate set through non-dominated sorting, filters the candidate pool according to the core sensory comprehensive score, maps the candidates to the historical tobacco leaf samples with the smallest standardized distance, and outputs a list of candidates to be reviewed with desensitization labels and review order, thus transforming the virtual optimization results into entity objects that can be reviewed by experts.

[0029] Specifically, step S6.1 involves performing non-dominated sorting on candidates that meet the constraints to obtain feasible Pareto candidate sets at different levels to identify the optimal boundary for multiple objectives; step S6.2 involves calculating the mean of the expected levels of multiple core sensory indicators as a comprehensive score to quantify the overall core sensory performance of the candidates; step S6.3 involves selecting candidates whose comprehensive scores reach a preset quantile to form a candidate pool, focusing on high-performing candidates to narrow the scope of review; step S6.4 involves standardizing the chemical composition data of candidates and historical tobacco samples and calculating Euclidean distance to establish a correlation metric between virtual candidates and real samples; and step S6.5 involves selecting the historical tobacco sample with the smallest distance as a mapping sample and organizing the desensitization labels, expected levels, satisfaction probabilities, and review order to output a traceable list of candidates to be reviewed.

[0030] To achieve the above objectives, this application also provides the following technical solutions: A raw material screening device for tobacco leaves, the raw material screening device being applied to the raw material screening method described above, the raw material screening device comprising: The tobacco leaf sample related data acquisition module is used to acquire chemical composition data and sensory evaluation data of historical tobacco leaf samples by near-infrared detection, and to match and preprocess the chemical composition data and the sensory evaluation data; The classification proxy model construction module is used to generate ordered sensory level labels based on the sensory evaluation data, and to train and output the probability of ordered sensory levels for the core sensory indicators and risk sensory indicators respectively, using the chemical composition data as input. The grade satisfaction probability calculation module is used to input the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model, calculate the expected grade of the core sensory index according to the predicted probability and grade order value of each ordered sensory grade, and calculate the grade satisfaction probability of the risk sensory index according to the sum of the predicted probabilities not lower than the preset minimum acceptable grade. The constrained multi-objective optimization problem construction module is used to construct a constrained multi-objective optimization problem with the expected level of multiple core sensory indicators as the optimization objective and the probability of satisfying the level of the risk sensory indicator not being lower than the corresponding probability threshold as the constraint, and to generate an initial candidate population based on the chemical composition data, the sensory evaluation data, and the preset chemical composition value range. The constrained multi-objective optimization problem solving module is used to perform crossover and mutation on the initial candidate population to generate offspring candidates. For offspring candidates that do not meet the constraints, the nearest feasible anchor point is selected from the feasible file and repaired step by step according to a preset mixing coefficient sequence. The candidate list output module is used to perform non-dominated sorting on the candidates that meet the constraints to obtain a feasible Pareto candidate set, map the feasible Pareto candidate set to the historical tobacco leaf samples, and output the candidate list to be reviewed.

[0031] To achieve the above objectives, this application also provides the following technical solutions: An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the raw material screening method for tobacco leaves as described above.

[0032] To achieve the above objectives, this application also provides the following technical solutions: A computer-readable storage medium storing program instructions that, when executed by a processor, can implement the tobacco leaf raw material screening method described above. Attached Figure Description

[0033] Figure 1 This is a schematic flowchart illustrating the steps of one embodiment of a method for screening tobacco leaves according to this application. Figure 2 This is a schematic diagram of the functional modules of a tobacco leaf raw material screening device according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an embodiment of the electronic device of this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0035] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (e.g., as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive having. For example, a process, method, system, product, or device having a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0036] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be present in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0037] It should be noted that, due to the limited types and number of symbols or letters that can represent specific meanings, for embodiments with many formulas or codes, there may be situations where symbols or letters cannot meet the usage requirements. Therefore, the interpretation of formula symbols in the steps or sub-steps of the embodiments is only valid for the current step or sub-step.

[0038] If the same symbol has different interpretations in different steps or sub-steps, the interpretation in the current step or sub-step shall prevail; if the same symbol appears in different steps or sub-steps, but no interpretation is given in subsequent steps or sub-steps after its first appearance, the interpretation in the first step or sub-step shall be used.

[0039] For example, "i=1,2,...,n" is a writing convention and a well-known meaning. If "i=1,2,...,n" has already appeared once in an embodiment and needs to be used again in formulas with different scenarios and meanings in the future, then the meaning of "i=1,2,...,n" should be interpreted according to the corresponding formulas. If "i=1,2,...,n" is not used for the first time and is changed to a symbol or letter with a non-well-known meaning, such as "p=1,2,...,q", to avoid repetition, it is more likely to cause confusion, ambiguity, and unclear problems.

[0040] like Figure 1 As shown, this embodiment provides an example of a method for screening tobacco leaves as raw material. In this embodiment, the raw material screening method includes the following steps: Step S1: Obtain chemical composition data and sensory evaluation data from near-infrared spectroscopy of historical tobacco leaf samples, and match and preprocess the chemical composition data and sensory evaluation data.

[0041] Furthermore, step S1 specifically includes the following steps: Step S1.1: Obtain near-infrared spectral data of historical tobacco leaf samples, and convert the near-infrared spectral data into chemical composition values ​​through a preset chemical composition correction model.

[0042] Preferably, the core of this step is to convert near-infrared spectral data into chemical composition values. A preset chemical composition correction model is used to quantitatively analyze the near-infrared spectral data of historical tobacco leaf samples, providing quantitative data at the chemical composition level for subsequent matching and preprocessing.

[0043] Near-infrared spectral data refers to the spectral curves obtained by scanning tobacco leaf samples using a near-infrared spectrometer, with the spectral wavelength range denoted as... to Spectral resolution is denoted as Historical tobacco leaf sample numbers are recorded as follows: , ,in The total number of historical tobacco leaf samples, the first The near-infrared spectral data of sample number is denoted as ,in For spectral data dimensions.

[0044] The chemical composition values ​​refer to the measured content of various chemical components in tobacco leaves, including total alkaloids, total sugars, reducing sugars, total nitrogen, protein, chlorine, potassium, starch, cellulose, and petroleum ether extract. Each chemical component is denoted as […]. ,in The number of chemical components, the first The numerical vector of chemical composition of sample number is denoted as: .

[0045] Among them, the preset chemical composition correction model refers to the mapping model from near-infrared spectral data to chemical composition values ​​established based on partial least squares regression. The correction model is denoted as... ,in For the partial least squares regression function, the calibration model is trained on the modeling set, and leave-one-out cross-validation is used to determine the optimal number of principal components.

[0046] Preferably, the near-infrared spectral wavelength range , spectral resolution spectral data dimensions Number of chemical components The principal component number of partial least squares regression is taken as follows Correction model correlation coefficient .

[0047] For example, the total number of historical tobacco leaf samples in a certain batch , No. Near-infrared spectral data of sample number After calibration model The converted chemical composition numerical vector is obtained. Total plant alkaloids Total sugar reducing sugars Total nitrogen ,protein ,chlorine potassium ,starch cellulose Petroleum ether extract The corresponding sample identifier is Sample-2026-0001.

[0048] Step S1.2: Obtain sensory evaluation data of historical tobacco leaf samples, and match the chemical composition data with the sensory evaluation data according to the sample identifier.

[0049] Preferably, the core of this step is to complete the cross-source matching of chemical composition data and sensory evaluation data, and establish the correspondence between the two types of data through sample identification, so as to provide complete paired training samples for subsequent surrogate model training.

[0050] The sensory evaluation data refers to the scores given by the expert evaluation panel after conducting sensory evaluations of the tobacco leaf samples. These sensory indicators include aroma quality, aroma intensity, aftertaste, irritation, and off-flavors. The scores for each sensory indicator are calculated using... to The system of divisions, the first The sensory evaluation data of sample number 1 is denoted as ,in Rate the fragrance's character. Rate the aroma intensity. Rate the aftertaste. To stimulate ratings, Scoring for impurities.

[0051] For sample identifier matching, the chemical composition numerical vector is used. Sensory evaluation data According to sample label Perform a one-to-one matching, and record the matched data as follows. ,in The number of successfully matched samples is specified, and samples that fail to match are removed.

[0052] Preferably, the number of members in the evaluation expert group is: For each person, the arithmetic mean of the scores from various expert evaluators is taken as the sensory index score for that sample, with the score rounded to one decimal place. The matching success rate is... .

[0053] For example, the Sample No. The fragrance was rated by experts, and the average score was [value missing]. Aroma lingering taste ,Stimulate Mixed gases Sensory evaluation data , and the numerical vector of chemical composition A match was successfully found using the sample identifier Sample-2026-0001, and the corresponding record identifier is Match-2026-0001.

[0054] Step S1.3: Perform missing value detection on the matched data, impute missing chemical component fields with median, and impute missing sensory evaluation fields with mean.

[0055] Preferably, the core of this step is to complete missing value detection and classification imputation. By using the median and mean to imput the missing fields of chemical composition and sensory evaluation respectively, the interference of missing values ​​on subsequent model training is reduced.

[0056] For missing value detection, the matched data is traversed sample by sample. Numerical vector detection of chemical components Does the data contain missing fields and sensory evaluation data? Does the data contain any missing fields? The locations of the missing values ​​are denoted as a set. ,in For sample number, Number the field. Identify the data type.

[0057] For median imputation of chemical composition fields, the median of each chemical composition field in all non-missing samples is calculated. The median of the chemical composition field is denoted as... ,in For the first The set of sample IDs for which the chemical composition field is not missing, and the set of missing values. use replace.

[0058] For the sensory evaluation field mean imputation, the arithmetic mean of each sensory index in all non-missing samples is calculated. The mean of the sensory index is denoted as . ,in For the first The set of sample IDs for which the sensory indicators are not missing, and the missing values. use replace.

[0059] Preferably, the pre-filling missing rate statistics show the missing rate of the chemical composition field. Missing rate of sensory evaluation field After imputation, all samples have no missing fields.

[0060] For example, the Chlorine content field of sample number Missing data; calculate the median chloride content of all non-missing samples. ,by Fill ;No. Stimulus rating field of sample number Missing data; calculate the mean stimulus score of all non-missing samples. ,by Fill The corresponding record identifiers are Fill-2026-0035 and Fill-2026-0048.

[0061] Step S1.4: Detect outliers in the chemical composition data and mark values ​​that exceed the preset range of chemical composition values ​​as invalid.

[0062] Preferably, the core of this step is to detect and mark abnormal values ​​of chemical components as invalid. By marking values ​​that exceed the preset range as invalid, abnormal records in the data acquisition or correction process are identified.

[0063] Among them, for the preset range of chemical component values, the valid value interval corresponding to each chemical component field is denoted as: ,in and The first The effective lower limit and effective upper limit of the chemical composition field.

[0064] For outlier detection, the imputed chemical composition data is traversed sample by sample and field by field. When the value of a certain field exceeds the corresponding valid value range, it is determined to be an outlier. The outlier criterion is as follows: .

[0065] Fields that meet the above criteria are marked as invalid, and the positions of invalid fields are denoted as a set. .

[0066] Preferably, the range of values ​​for each chemical component is set as follows: total alkaloids Total sugar reducing sugars Total nitrogen ,protein ,chlorine potassium ,starch cellulose Petroleum ether extract Outlier detection rate .

[0067] For example, the Total alkaloid value of sample number Exceeding the valid value range upper limit The value is identified as an outlier and marked as invalid. Invalid field location. Add to collection ;No. Chlorine content of sample number Exceeding the valid value range upper limit They are also marked as invalid, with the corresponding record identifiers being Invalid-2026-0072 and Invalid-2026-0105.

[0068] Step S1.5: Linear interpolation is used to compensate for fields marked as invalid, and Min-Max normalization is performed on the compensated chemical composition data and sensory evaluation data.

[0069] Preferably, the core of this step is to complete the linear interpolation compensation and full normalization of invalid fields. The invalid field data is restored by linear interpolation, and then the chemical composition and sensory evaluation data are normalized by Min-Max to eliminate dimensional differences.

[0070] For linear interpolation compensation, invalid fields are treated as missing values, and the same median imputation strategy as in S1.3 is used to compensate for invalid fields. Median of the chemical composition field Used to replace invalid fields The compensated chemical composition data is denoted as .

[0071] For Min-Max normalization, a linear mapping is performed on each field of the compensated chemical composition data, and the normalization formula is as follows: .

[0072] in, For the first Sample No. The normalized value of the chemical composition field, and its range. , and The first The effective value range of the chemical composition field is defined by its lower and upper limits.

[0073] For the sensory evaluation data normalization, the scores of each sensory index are also normalized using Min-Max normalization, and the normalization formula is as follows: .

[0074] in, For the first Sample No. The normalized value of the sensory index, and The first The lower and upper limits of the scoring range for the sensory indicators.

[0075] Preferably, the sensory evaluation scoring range is set as follows: , The normalized chemical composition data and sensory evaluation data both have a range of values. .

[0076] For example, the The invalid field of total alkaloids in sample number was determined by the median. After compensation Normalization ;No. Sample No. 1 fragrance quality rating Normalization After compensation and normalization, the first Normalized vector of chemical composition of sample number Sensory evaluation normalized vector The corresponding record identifier is Norm-2026-0001.

[0077] Step S1.6: Arrange the normalized chemical composition data and sensory evaluation data according to the samples to obtain the preprocessed chemical composition data and sensory evaluation data.

[0078] Preferably, the core of this step is to complete the structured arrangement of the normalized data, arranging the chemical composition and sensory evaluation normalized data in the order of sample numbers into a structured dataset, providing a unified format input for subsequent surrogate model training.

[0079] For sample permutations, all The normalized chemical composition vectors of each sample are arranged by sample number to form a chemical composition data matrix. The normalized sensory evaluation vectors of all samples are arranged by sample number to form a sensory evaluation data matrix. The pretreated chemical composition data and sensory evaluation data constitute a data pair. .

[0080] Preferably, the preprocessed dataset is ordered according to... The dataset is divided into training and test sets, with the number of samples in the training set being... Number of test set samples The classification uses stratified sampling to ensure that the proportion of each sensory level is consistent between the training set and the test set.

[0081] For example, the number of matched samples after preprocessing Number of training set samples Number of test set samples Chemical composition data matrix Sensory evaluation data matrix The pretreated chemical composition data and sensory evaluation data constitute a data pair. The corresponding record is identified as Dataset-2026-Preprocessed.

[0082] Beneficial effects of steps S1.1 to S1.6: This series of steps addresses the issue of heterogeneous sources and inconsistent quality of historical tobacco sample chemical composition data and sensory evaluation data. Through near-infrared spectral conversion, sample label matching, missing value imputation, outlier detection, linear interpolation compensation, and normalization, preprocessed data of a uniform scale is formed, providing a reliable data foundation for surrogate model training.

[0083] The process includes the following steps: Step S1.1 converts near-infrared spectral data into chemical composition values ​​using a pre-defined chemical composition correction model, achieving lossless high-throughput data acquisition; Step S1.2 establishes a correspondence between chemical composition data and sensory evaluation data through sample identifier matching, eliminating the risk of cross-source data mismatch; Step S1.3 fills in missing fields for chemical composition and sensory evaluation using the median and mean, respectively, reducing the interference of missing values ​​on model training; Step S1.4 detects and marks outliers using a pre-defined value range, identifying abnormal records during data acquisition; Step S1.5 compensates for invalid fields using linear interpolation and performs Min-Max normalization to eliminate dimensional differences among different chemical components; and Step S1.6 arranges and organizes samples to form structured preprocessed data, providing a unified input for subsequent model training and candidate evaluation.

[0084] Step S2: Generate ordered sensory level labels based on sensory evaluation data. Using chemical composition data as input, train classification proxy models that output ordered sensory level probabilities for core sensory indicators and risk sensory indicators respectively.

[0085] Furthermore, step S2 specifically includes the following steps: Step S2.1: Determine the level threshold based on the score distribution of each sensory indicator in the sensory evaluation data, and divide the sensory scores into three ordered sensory levels: low level, medium level, and high level.

[0086] Preferably, the core of this step is to complete the initial division of sensory scores into ordered sensory levels, and to determine the level threshold through the score distribution of each sensory indicator, so as to provide a basic classification framework for the subsequent refined generation of ordered sensory level labels.

[0087] Among them, for the score distribution statistics, the sensory evaluation data after preprocessing were statistically analyzed for each sensory index. The normalized score distribution of each indicator in the middle, the first The score distribution of the sensory index is denoted as the empirical cumulative distribution function. ,in This is the normalized score.

[0088] Specifically, the quantile method is used to determine the level threshold, and the boundary threshold between low and medium levels is denoted as... The threshold separating the medium and high levels is denoted as The formula for calculating the boundary threshold is: .

[0089] .

[0090] in, For the first The inverse function of the empirical cumulative distribution function of the sensory index. and This is the preset quantile.

[0091] Among them, for the orderly sensory hierarchy classification, the first Sample No. Normalized scores of sensory indicators They are divided into three ordered levels based on a threshold, and the division rules are as follows: .

[0092] in, For the first Sample No. Preliminary ordered level labels for sensory indicators. Corresponding to lower levels, Corresponding to intermediate level, Corresponds to a higher level.

[0093] Preferably, a preset quantile is used. , That is, the lower grades are the distribution of ratings. Below the quantile, the higher level is Above the percentile, the middle level is considered the middle. .

[0094] For example, the inverse function of the empirical cumulative distribution of the normalized score of the fragrance quality index. , Level threshold , , No. Sample No. 1 fragrance quality normalization score , in Intervals, preliminary ordered level labels The corresponding level is Grade-2026-0001.

[0095] Step S2.2: The intermediate level is split into low-intermediate and high-intermediate levels to form four ordered levels. The low-intermediate and high-intermediate levels are then merged into the intermediate level to obtain three ordered sensory level labels: Q1, Q2, Q3, and Q4.

[0096] Preferably, the core of this step is to refine the splitting and re-merging of the intermediate levels. After splitting into four quartiles, the two middle levels are merged to form a three-quartile ordered label, retaining the level sequence information for subsequent expected level calculation.

[0097] Among them, for the intermediate level segmentation, preliminary ordered level labels will be used. The sample is scored according to normalization. Further subdividing within the medium-level range, with the subdivision threshold denoted as... The calculation formula is: .

[0098] in, For the first under medium-level sample conditions The conditional median of the normalized scores for the sensory index.

[0099] Specifically, for the formation of the four levels, the medium level is divided into low-medium and high-medium levels, forming four ordered levels. The division rules are as follows: .

[0100] in, For the first Sample No. The four-level labeling of sensory indicators, Corresponding to lower levels, Corresponding to low and medium levels, Corresponding to medium to high levels, Corresponds to a higher level.

[0101] Among them, for the re-merging into third-place labels, the low and medium levels will be... With medium and high level Merging them into intermediate levels results in three ordered sensory level labels. , , The mapping relationship is as follows: .

[0102] in, For the first Sample No. The final ordered sensory level label of the sensory index, Corresponding to lower levels, Corresponding to the intermediate level, For higher levels, the level sequence values ​​are denoted as follows: , , .

[0103] Preferably, a third-position label , , The sample proportion is approximately , , Rank ordinal value Used for weighted summation in subsequent expected level calculations.

[0104] For example, the median of the grade conditions in the fragrance quality index. , No. Sample No. 1 fragrance quality normalization score , in Range, four-level labels Corresponding to medium and high levels, the final ordered sensory level labels are then merged. Corresponding to intermediate levels, level ordinal value The corresponding record identifier is Label-2026-0001.

[0105] Step S2.3: Determine the core sensory indicators and risk sensory indicators. The core sensory indicators are aroma quality, aroma quantity, and aftertaste, while the risk sensory indicators are irritation and off-flavors.

[0106] Preferably, the core of this step is to complete the functional classification of sensory indicators, distinguishing the core indicators that affect the use value of tobacco leaves from the indicators that need to control risk thresholds, thus providing modeling objects for objective functions and constraint functions for subsequent multi-objective optimization.

[0107] Among them, core sensory indicators refer to the main sensory evaluation dimensions that determine the use value of tobacco leaf formulations, including aroma quality, aroma intensity, and aftertaste. The set of core sensory indicators is denoted as... Corresponding sensory evaluation data In , , Three components.

[0108] Among them, risk sensory indicators refer to sensory evaluation dimensions that, when their performance is too low, affect the usability of tobacco leaves but do not need to be maximized. These include irritation and off-odors. The set of risk sensory indicators is denoted as […]. Corresponding sensory evaluation data In , Two components.

[0109] Preferably, the number of core sensory indicators Number of risk sensory indicators The two types of indicators do not overlap.

[0110] For example, the The core sensory indicator for sample number is the ordered sensory level label. , , Risk sensory indicators, ordered sensory level labels are , The corresponding record identifier is Classify-2026-0001.

[0111] Step S2.4: Using chemical composition data as input features, construct training sets and validation sets for core sensory indicators and risk sensory indicators, respectively.

[0112] Preferably, the core of this step is to prepare the training data for the classification agent model, using normalized chemical composition data as input features and ordered sensory level labels as output labels, and constructing independent training sets and validation sets for each sensory indicator.

[0113] Specifically, for the input features, the preprocessed chemical composition data matrix is ​​used. Each row vector in the vector is used as an input feature, and the first row vector is used as an input feature. The input features of sample number 1 are .

[0114] For the output label, take the first one. Ordered sensory level labels for sensory indicators As the first The output label for sample number, for each sensory indicator. Construct independent label vectors .

[0115] Among them, the division of the training set and the validation set will be The samples are divided into a training set and a validation set using stratified sampling. The set of sample IDs for the training set is denoted as . The set of validation sample numbers is denoted as The division ratio is the preset training ratio. The proportion of each ordered sensory level in the training and validation sets is consistent with the proportion of that level in the entire set.

[0116] Preferably, the training ratio Number of training set samples Number of validation set samples Stratified sampling , , The three levels are divided separately.

[0117] For example, for fragrance quality indicators, the training set input feature matrix Training set label vectors middle Label indivual, Label indivual, Label One, the validation set input feature matrix Validation set label vector middle Label indivual, Label indivual, Label There are 1, and the corresponding record is identified as Split-2026-Aroma.

[0118] Step S2.5: Train a random forest classification model for each sensory indicator. The random forest classification model outputs the predicted probability of each ordered sensory level, thus obtaining a classification surrogate model.

[0119] Preferably, the core of this step is to train the random forest classification model. By independently training the random forest classifier for each sensory index and outputting the predicted probability of each ordered sensory level, a probabilistic output basis is provided for subsequent calculations of expected level and level satisfaction probability.

[0120] For the random forest classification model, it refers to an ensemble classifier composed of multiple decision trees, where the first... The random forest classification model for sensory index number is denoted as . The model input is a normalized chemical composition vector. The output is the predicted probability vector for each ordered sensory level. The prediction probability vector satisfies .

[0121] Specifically, for decision tree construction, each decision tree is drawn with replacement from the training set. Each sample, from which a sample splits at each node. Randomly selected from the features The features are selected as candidate splitting features. The feature with the largest decrease in Gini impurity and the splitting point are selected for node splitting. The leaf node outputs the proportion of each level in the sample of that node as the prediction probability.

[0122] Among them, for the impurity of the ginni, the node The formula for calculating the impurity of Gini is: .

[0123] in, For nodes The total number of samples in the sample, For nodes Medium-level label The number of samples.

[0124] Among them, for the aggregation of prediction probabilities, in random forests Decision trees for input The predicted probabilities are taken as an arithmetic mean, and the aggregation formula is: .

[0125] in, For the first decision trees for the first Sample No. Sensory index level The predicted probability.

[0126] For the classification proxy model, the random forest classification model combination of each sensory index constitutes the classification proxy model set. ,common An independent random forest classification model.

[0127] Preferably, the number of decision trees in a random forest Candidate splitting feature number Maximum depth of decision tree Minimum number of samples for leaf nodes Sampling with replacement uses the Bootstrap method.

[0128] For example, training a random forest classification model based on fragrance quality indicators. training set and Post-input training The first decision tree, the Verification Sample No. enter The predicted probability vector is then obtained. ,in probability Highest, probability Secondly, validation set accuracy Macro-F1 value The corresponding record is identified as Model-2026-Aroma.

[0129] Beneficial effects of steps S2.1 to S2.5: This series of steps addresses the issues of coarse sensory evaluation data classification and lack of targeted training of surrogate models. It divides sensory level labels into ordered categories based on score distribution, and trains random forest classification models for core sensory indicators and risk sensory indicators respectively, providing a probabilistic output basis for subsequent expected level calculation and level satisfaction probability assessment.

[0130] Specifically, step S2.1 determines the level threshold by rating distribution to divide ordered sensory levels, transforming continuous sensory ratings into learnable discrete labels; step S2.2 splits and merges intermediate levels through quartiles to form three ordered labels Q1, Q2, Q3, and Q4, retaining the level order information for subsequent expected level calculation; step S2.3 clarifies core sensory indicators and risk sensory indicators, providing modeling objects for objective and constraint functions for constrained multi-objective optimization; step S2.4 constructs training and validation sets respectively, providing data support for the independent training and validation of the classification proxy model; and step S2.5 trains a random forest classification model to output the predicted probability of each ordered sensory level, retaining the confidence distribution of candidates at different levels.

[0131] Step S3: Input the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model, calculate the expected level of the core sensory index based on the predicted probability and level order value of each ordered sensory level, and calculate the level satisfaction probability of the risk sensory index based on the sum of the predicted probabilities that are not lower than the preset minimum acceptable level.

[0132] Furthermore, step S3 specifically includes the following steps: Step S3.1: Input the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model to obtain the predicted probability of each sensory index under each ordered sensory level.

[0133] Preferably, the core of this step is to complete the input and probability output of candidate tobacco raw materials to the classification proxy model. By inputting the candidate chemical component vector into the random forest classification model of each sensory index, the complete probability distribution of the candidate on all sensory indices is obtained.

[0134] Here, the chemical composition vector of the candidate tobacco raw materials refers to the normalized chemical composition vector obtained after the tobacco raw materials to be screened have undergone the same process as S1.1 to S1.5, denoted as... superscript Identify candidate samples.

[0135] Among them, for model input and probability output, Input the classification proxy model set respectively Random forest classification model corresponding to each sensory index , obtain the Predicted probability vector of sensory index The predicted probabilities of all sensory indicators constitute a probability matrix. ,in The total number of sensory indicators. For ordered sensory hierarchy numbers.

[0136] Preferably, the acquisition and preprocessing procedures for the chemical composition vectors of candidate tobacco raw materials are consistent with those for historical samples, and the near-infrared spectra are acquired and processed using the same calibration model. Transform and normalize.

[0137] For example, the chemical composition vector of a candidate tobacco raw material ,enter The probability of fragrance prediction was then obtained. ,enter The aroma quantity prediction probability was then obtained. ,enter The probability of predicting aftertaste was then obtained. ,enter The stimulus prediction probability is then obtained. ,enter Then the probability of mixed gas prediction was obtained. probability matrix The corresponding record identifier is Cand-2026-001.

[0138] Step S3.2: For the core sensory indicators, multiply the predicted probability of each ordered sensory level by the corresponding level ordinal value and sum them to obtain the expected level of the core sensory indicators.

[0139] Preferably, the core of this step is to perform a probability-weighted calculation of the expected level of the core sensory index. By weighted summing of the predicted probability and the level order value, a continuous value is used to characterize the expected performance of the candidate on the core sensory index, avoiding the loss of boundary candidate information caused by hard classification truncation.

[0140] Among them, for the calculation of the expected level, the first Expected Level of Core Sensory Indicators The calculation formula is: .

[0141] in, For the first The expected level of the core sensory indicators, For level ordinal value, , , , For candidates in Sensory index level The predicted probability.

[0142] Among them, regarding the range of values ​​for the expected level, since This is a probability-weighted summation, with values ​​ranging from 1 to 2. When the predicted probability is concentrated in hour Approaching Concentrated on hour Approaching When distributed across multiple levels It reflects the location of the probability centroid.

[0143] Preferably, the core sensory indicator set The expected level calculation is performed independently for each indicator in the set.

[0144] For example, the fragrance expectation level of candidate Cand-2026-001. Aroma level Aftertaste Expectation Level The corresponding record is identified as Expect-2026-001.

[0145] Step S3.3: For the risk sensory index, determine the preset minimum acceptable level, and sum the predicted probabilities of each level that is not lower than the minimum acceptable level to obtain the level satisfaction probability of the risk sensory index.

[0146] Preferably, the core of this step is to calculate the probability of meeting the risk sensory indicator level. By summing the predicted probabilities of not being lower than the minimum acceptable level, the confidence level of the candidate meeting the safety requirements in terms of risk sensory indicators is quantified.

[0147] The minimum acceptable level refers to the lowest level requirement that the risk sensory indicators must meet. The preset minimum acceptable level is denoted as... A unified setting is applied to all risk perception indicators. This means that candidates must reach at least the intermediate level in terms of risk perception indicators.

[0148] Among them, for the probability calculation of the level, the first Probability of meeting the level of risk perception indicator. The calculation formula is: .

[0149] in, For the first The probability of meeting the level of the risk perception indicator. This indicates that the rank ordinal value is not lower than The orderly hierarchy of senses, when hour, .

[0150] Preferably, the minimum acceptable level The level meets the probability range. A higher probability of meeting the requirements indicates that the candidate is safer in terms of risk perception indicators.

[0151] For example, the probability of satisfying the stimulus level of candidate Cand-2026-001. The probability of impurity level is met. The corresponding record identifier is Satisfy-2026-001.

[0152] Step S3.4: When the probability of the grade is not lower than the corresponding probability threshold, the candidate corresponding to the chemical composition vector of the candidate tobacco raw material is marked as a feasible candidate; otherwise, it is marked as an infeasible candidate.

[0153] Preferably, the core of this step is to determine the feasibility of candidates. By comparing the probability of meeting the level with the probability threshold, the candidates are divided into two categories: feasible and infeasible, providing a basis for constraint determination for subsequent multi-objective optimization.

[0154] Among them, the probability threshold refers to the minimum requirement for the risk perception index level to meet the probability. The probability threshold of the risk sensory indicator is denoted as... ,when This constraint is satisfied at that time.

[0155] For feasibility assessment, when a candidate meets all risk perception indicators... If the condition is met, the candidate is marked as feasible; otherwise, it is marked as infeasible. The feasibility criterion is as follows: .

[0156] in, This indicates a logical AND operation, meaning that a candidate is considered feasible when the probability of all risk sensory indicators being at least equal to the corresponding probability threshold.

[0157] For candidate labels, feasible candidates are denoted as the set. Unfeasible candidates are denoted as set. The feasibility markers of the candidates will be used for the evaluation and ranking of the initial candidate population in S4.

[0158] Preferably, the stimulus probability threshold Impurity probability threshold Both thresholds are higher than This is to ensure that the candidates have sufficient safety margin in terms of risk perception indicators.

[0159] For example, the probability of satisfying the stimulus level of candidate Cand-2026-001. The probability of impurity level is met. If both risk constraints are satisfied, it is marked as a feasible candidate and added to the set. The corresponding record is identified as Feasible-2026-001.

[0160] Beneficial effects of steps S3.1 to S3.4: This series of steps addresses the issues of confidence discrepancies in lost levels due to hard classification truncation and the lack of risk quantification methods. It outputs the predicted probability of ordered sensory levels through a classification surrogate model, calculates the expected level of core sensory indicators and the probability of meeting the level of risk sensory indicators, and determines the feasibility of candidates using probability thresholds, thus providing target values ​​and constraint values ​​for constrained multi-objective optimization.

[0161] Specifically, step S3.1 involves inputting candidate chemical component vectors into a classification proxy model to obtain the predicted probabilities of each ordered sensory level, preserving the complete confidence distribution of candidates in level judgment; step S3.2 involves calculating the expected level by weighted summation of the predicted probabilities and level ordinal values, using continuous values ​​to characterize core sensory performance and avoid hard classification truncation; step S3.3 involves calculating the level satisfaction probability by summing the predicted probabilities of levels not lower than the minimum acceptable level, quantifying the safety margin of risk sensory indicators; and step S3.4 involves determining whether a candidate is feasible or infeasible based on a probability threshold, transforming probability constraints into operable screening markers.

[0162] Step S4: With the expected level of multiple core sensory indicators as the optimization objective and the probability of satisfying the level of risk sensory indicators not being lower than the corresponding probability threshold as the constraint, a constrained multi-objective optimization problem is constructed, and an initial candidate population is generated based on chemical composition data, sensory evaluation data, and preset chemical composition value range.

[0163] Furthermore, step S4 specifically includes the following steps: Step S4.1: With the expected level of multiple core sensory indicators as the objective to be maximized, and with the probability that the level of each risk sensory indicator is not lower than the corresponding probability threshold as the constraint, a constrained multi-objective optimization problem is constructed.

[0164] Preferably, the core of this step is to complete the mathematical modeling of the constrained multi-objective optimization problem. By setting the core sensory expectation level as the maximization objective and the risk sensory level satisfaction probability as the inequality constraint, the mathematical framework for the pursuit of core sensory perception and the control of risk sensory perception are unified.

[0165] Among them, the normalized chemical composition vector of candidate tobacco raw materials is the optimization variable. To optimize the variables, the range of values ​​for each component is: .

[0166] Among them, for the objective function, the core sensory index set Expected levels of each indicator To maximize the objective, the objective function vector is denoted as . , among which Through classification proxy model The output prediction probability is calculated according to formula S3.2.

[0167] Among them, the set of risk perception indicators is used for the constraints. The probability of each indicator's level being satisfied Not lower than the corresponding probability threshold The constraint function vector is denoted as The constraints are .

[0168] Specifically, the constraint violation degree is denoted as follows: when a candidate does not satisfy a certain constraint, the constraint violation degree is denoted as... The total constraint violation degree of the candidate is denoted as , This indicates that the candidate is a feasible candidate. This indicates that the candidate is not feasible.

[0169] Preferably, the number of objective functions Number of constraints Optimize variable dimensions .

[0170] For example, the objective function vector of candidate Cand-2026-001 constraint function vector Both constraint components are greater than Total constraint violation The candidate is a feasible candidate, and the corresponding record identifier is Opt-2026-001.

[0171] Step S4.2: Extract a portion of samples from the historical samples corresponding to the chemical composition data and sensory evaluation data as historical candidates.

[0172] Preferably, the core of this step is to extract candidate individuals from historical samples as seed individuals for the initial population. By extracting real historical samples, the actual distribution characteristics of chemical composition combinations are preserved, providing a feasible starting point for the search process.

[0173] For historical sample extraction, the data is extracted from the preprocessed chemical composition data matrix. Randomly selected Rows are considered as historical candidates, and the set of historical candidates is denoted as . ,in From The sample number was randomly drawn without replacement.

[0174] Regarding the feasibility of historical candidates, since historical samples come from real tobacco leaves, their chemical composition combinations usually meet the risk sensory constraints, but feasibility still needs to be verified through a classification surrogate model. Feasible historical candidates after verification are denoted as follows: Not feasible is marked as .

[0175] Preferably, the number of historical candidates drawn The sampling method is random sampling without replacement, and the proportion of feasible candidates in the historical candidates is... .

[0176] For example, from Randomly drawn without replacement from a historical sample The nth sample, the th The historical candidate is It was verified as a feasible candidate by the classification proxy model, and the corresponding record is identified as Hist-2026-0127.

[0177] Step S4.3: Randomly generate restricted random candidates within the preset range of chemical composition values, and combine historical candidates with restricted random candidates to form an initial candidate population.

[0178] Preferably, the core of this step is to complete the hybrid construction of the initial candidate population by combining historical candidates with restricted random candidates, thus balancing the coverage of the search space and the feasibility of the starting point.

[0179] Among them, for restricted random candidate generation, within the normalized value range Internal random generation There are n candidates, and the restricted random candidate set is denoted as . Each component of each candidate is in Uniformly distributed within.

[0180] Among them, for the initial candidate population combination, the historical candidate set is used. With restricted random candidate set Merge into the initial candidate population Initial candidate population size .

[0181] Preferably, the number of random candidates is limited. Initial candidate population size Historical candidates and restricted random candidates each account for .

[0182] For example, the initial candidate population The Middle to Candidate number 1 is a historical candidate, and candidate number 2 is... to Candidate number 1 is a restricted random candidate, and candidate number 2 is a restricted random candidate. Candidate No. Population size The corresponding record is identified as Pop-2026-Init.

[0183] Step S4.4: Input the chemical composition vector of each candidate in the initial candidate population into the classification surrogate model, and calculate the expected level and level satisfaction probability of each candidate.

[0184] Preferably, the core of this step is to complete the batch evaluation of the initial candidate population, and to calculate the objective function value and constraint function value of each candidate through a classification proxy model, so as to provide an evaluation basis for subsequent feasible ranking and non-dominated ranking.

[0185] For batch evaluation, the chemical composition vector is input into the classification proxy model set for each candidate. Calculate the expected level of each core sensory indicator according to formula S3.2. Calculate the probability of meeting the risk sensory index level according to formula S3.3. Calculate the total constraint violation rate for each candidate. .

[0186] Among them, for the storage of evaluation results, the first The evaluation result of candidate number is denoted as ,in Let the objective function vector be... The total constraint violation rate is represented by the evaluation results of all candidates, which form an evaluation matrix.

[0187] Preferably, the batch evaluation is performed in the order of candidate numbers, and after the evaluation is completed, the number of feasible candidates in the initial candidate population is counted. Number of infeasible candidates , .

[0188] For example, the initial candidate population The Middle Evaluation results of candidate number (historical candidate) , As a feasible candidate; Evaluation results of candidate number (restricted random candidate) , Infeasible candidates, feasible candidates in the initial population Infeasible candidate The corresponding record is identified as Eval-2026-Init.

[0189] Step S4.5: Sort the initial candidate population in a manner that feasible candidates take precedence over infeasible candidates, and perform non-dominated sorting and crowding distance calculation on the feasible candidates.

[0190] Preferably, the core of this step is to prioritize and rank the candidates based on their feasibility. By placing feasible candidates before infeasible candidates and performing non-dominated ranking and crowding distance calculation among the feasible candidates, a fitness evaluation basis is provided for subsequent selection.

[0191] Specifically, for the feasibility priority ranking, feasible candidates in the initial candidate population are ranked before infeasible candidates, and feasible candidates are ranked according to their total constraint violation degree. Treat them equally, and among infeasible candidates, rank them according to the degree of total constraint violation. Sort in ascending order, the sorting rules are as follows: .

[0192] For non-dominated sorting, in the feasible candidate set In the middle, candidates Dominant Candidate If and only if And there exists at least one target component. The non-dominated hierarchy is denoted as ,in It is the set of candidates that is not dominated by any other feasible candidates.

[0193] For crowded distances, at the same non-dominated level In the middle, the first Crowding distance of candidate number The calculation formula is: .

[0194] in, and The first Candidate number 1 is in The values ​​of adjacent candidate targets on target number 1, arranged in ascending order. and For the first The maximum and minimum values ​​of target number in the current non-dominated hierarchy, and the boundary candidate crowding distance are set to... .

[0195] Preferably, the non-dominated sorting uses a fast non-dominated sorting algorithm, and the crowding distance is calculated independently within each non-dominated level.

[0196] For example, in the initial candidate population The feasible candidates were obtained through non-dominated sorting. hierarchy One candidate, hierarchy One candidate, hierarchy There are 1 candidate, and the rest are distributed at lower levels, the 1st... Candidate number is located at Hierarchy, crowded distance The corresponding record is identified as Sort-2026-Init.

[0197] Step S4.6: Select candidates from the initial candidate population to enter the next generation based on the non-dominated sorting results and crowding distance.

[0198] Preferably, the core of this step is to select candidate tournaments by using a dual-criteria selection strategy that prioritizes non-dominated levels and prioritizes crowding distance within the same level, thus balancing the convergence and diversity of the search process.

[0199] The selection criterion uses a tournament selection method, where a random sample is drawn from the population each time. One candidate participates in the tournament, with priority given to those with higher non-dominant tiers ( of When two candidates are at the same non-dominated level, the candidate with the larger crowding distance is selected.

[0200] For tournament selection, the selection rules are as follows: .

[0201] in, For random selection A set of candidate IDs For the first The candidate's non-dominated level number is selected when multiple candidates are at the same level. The largest one.

[0202] Among them, regarding the selection of scale, from Among the candidates, those who passed Second tournament selection Each candidate moves to the next generation, allowing the same candidate to be selected multiple times.

[0203] Preferably, the scale of the tournament This refers to binary tournament selection, where the number of selections equals the population size. .

[0204] For example, a tournament randomly selects candidates. and candidates Candidate lie in Hierarchy ( ), candidate Infeasible candidate (non-dominated level) ), candidate Those at a higher non-dominant level were selected, through The subsequent tournament selection forms the post-selection candidate set, in which... The tier candidate was selected a lot, and the corresponding record is identified as Select-2026-Init.

[0205] Beneficial effects of steps S4.1 to S4.6: This series of steps addresses the problem of mutual constraints between core sensory indicators and risk sensory indicators in multi-objective optimization. By constructing a constrained multi-objective optimization problem and generating a mixed initial candidate population, the steps involve probabilistic evaluation of the candidates and non-dominated ranking selection based on feasibility priority, thus providing a high-quality parent population for subsequent crossover and mutation.

[0206] Specifically, step S4.1 constructs a constrained multi-objective optimization problem by using the expected level as the target to be maximized and the probability of level satisfaction as the constraint, thus unifying the mathematical framework of core sensory pursuit and risk sensory control; step S4.2 extracts historical candidates from historical samples, retaining the real chemical composition combination structure as the starting point for the search; step S4.3 generates restricted random candidates within the value range and combines them with historical candidates, taking into account both the coverage and feasibility of the search space; step S4.4 uses a classification surrogate model to batch calculate the expected level and probability of level satisfaction of candidates, providing an evaluation basis for feasibility judgment and ranking; step S4.5 identifies the multi-objective optimal boundary among feasible candidates through feasible priority ranking, non-dominated ranking, and crowding distance calculation; step S4.6 selects candidates to enter the next generation based on the non-dominated ranking results and crowding distance, balancing convergence and diversity.

[0207] Step S5: Perform crossover and mutation on the initial candidate population to generate offspring candidates. For offspring candidates that do not meet the constraints, select the nearest feasible anchor point from the feasible file and repair them step by step according to the preset mixing coefficient sequence.

[0208] Furthermore, step S5 specifically includes the following steps: Step S5.1: Perform a crossover operation on the selected candidates to generate intermediate candidates, and perform a mutation operation on the intermediate candidates to generate offspring candidates.

[0209] Preferably, the core of this step is to complete the crossover and mutation in genetic operations, and to generate offspring candidates by simulating binary crossover and polynomial mutation, thereby introducing new candidate solutions into the search space.

[0210] Specifically, for the crossover operation, two parent candidates are randomly selected from the chosen candidate set each time. and Simulated binary crossover is performed on each component, and the crossover formula is: .

[0211] .

[0212] in, and For the first The two child values ​​after the cross of the number components For the first Cross-distribution parameters of the first component, Based on the preset cross-distribution index control: .

[0213] in, For the first On the number of components Uniformly distributed random numbers.

[0214] Among them, for the mutation operation, the intermediate candidates after crossover are... Each component undergoes polynomial mutation, and the mutation formula is: .

[0215] in, For the mutated first Component value, For the first The variation perturbation of the sign component Based on the preset variation distribution index control: .

[0216] in, For the first On the number of components Uniformly distributed random numbers, and For the first The normalized upper and lower limits of the chemical component number.

[0217] Among them, for boundary processing, the mutated component values ​​exceed When the range is defined, it is truncated to the boundary value.

[0218] Preferably, the crossover probability Cross-distribution index Probability of mutation Variation distribution index Each parent generation produces two offspring, and the size of the offspring is equal to the population size. .

[0219] For example, two parent candidates and , No. The number of random numbers in component number Cross-distribution parameters After crossing After polynomial mutation (probability of mutation) (No mutation triggered), the corresponding record is identified as Cross-2026-001.

[0220] Step S5.2: Based on the importance of variables in the classification proxy model corresponding to the risk sensory indicators, determine the variation range of each chemical component, and constrain the variation range of relevant chemical components to be less than the variation range of other chemical components.

[0221] Preferably, the core of this step is to set the differential variation amplitude, quantify the influence of each chemical component on the risk perception constraint by using the variable importance of random forest, and reduce the variation perturbation of the constraint-related components to reduce the probability of variation disrupting feasibility.

[0222] Among them, regarding the importance of variables, the first The random forest classification model corresponding to the sensory index. The Middle The importance of the variable for chemical component number is denoted as: The Gini impurity decrease was used as the metric. For the first The feature is in The arithmetic mean of the decrease in Gini impurity across all decision trees.

[0223] Among them, for the constraint of relevant chemical components, the risk sensory index set will be used. The importance of the variables corresponding to each indicator is aggregated by the maximum value. The constraint-related importance of chemical component number 1 is denoted as: .

[0224] in, For the first The number one chemical component is the most significant variable among all risk sensory indicators.

[0225] Among them, the determination of the variation magnitude is based on the importance of the constraint. The chemical components were divided into a constraint-related group and the remaining groups. The constraint-related group was... Top The components, constrained by the scaling factor of the variation perturbation of the relevant group, are denoted as The scaling factors for the remaining groups are denoted as ,satisfy .

[0226] Specifically, for the adjustment of the variation disturbance, the variation disturbance amount in S5.1 is... Scaling by group: .

[0227] in, For the first The scaling factor of the group, after adjustment, the mutation formula is replaced with .

[0228] Preferably, the number of relevant components is constrained. Constrain the scaling factor of the relevant group scaling factors for the remaining groups The importance of variables is measured by the decrease in Gini impurity.

[0229] For example, stimulus indicators The importance ranking of variables Its chemical composition is total alkaloids. Total nitrogen ,chlorine Potassium ,this Each component is grouped into a constraint-related group, and the variation perturbation scaling factor is used. , No. Original variation perturbation of component number After adjustment The amplitude of variation decreased Scaling factor for other components Remain unchanged, with the corresponding record identifier being Imp-2026-001.

[0230] Step S5.3: Input the offspring candidates into the classification proxy model and calculate the probability of each offspring candidate meeting the required level.

[0231] Preferably, the core of this step is to complete the constraint evaluation of the offspring candidates. The probability of each offspring candidate meeting the risk perception index is calculated through a classification proxy model, providing a basis for subsequent constraint determination.

[0232] For the evaluation of offspring candidates, the mutated chemical composition vector is generated for each offspring candidate. Input classification proxy model set Models corresponding to various risk sensory indicators Obtain the predicted probability Calculate the probability of meeting the grade according to formula S3.3. The total constraint violation degree is calculated according to the S3.4 criterion. .

[0233] Among them, for the storage of evaluation results, the first The evaluation results of the candidate for the position of offspring are recorded as follows: ,in The objective function vector is calculated using the model. For the total degree of constraint violation, This indicates that the offspring candidate is a feasible candidate.

[0234] Preferably, the evaluation of offspring candidates follows the same process as the batch evaluation in S4.4, and after the evaluation is completed, the number of feasible candidates in the offspring is counted. Number of infeasible candidates .

[0235] For example, after evaluation by the classification surrogate model, the probability of the stimulus level being satisfied for the offspring candidate Cross-2026-001 is... The probability of impurity level is met. Stimulus constraint Not satisfied, impurity gas constraint Satisfaction, total constraint violation degree This is an infeasible candidate, and the corresponding record is identified as EvalChild-2026-001.

[0236] Step S5.4: Determine whether the offspring candidates meet the constraints, and add the offspring candidates that meet the constraints to the feasible file.

[0237] Preferably, the core of this step is to complete the feasibility determination of offspring candidates and the dynamic accumulation of feasible files. By adding offspring candidates that meet the constraints to the feasible files, a reference point pool is provided for the subsequent repair of infeasible candidates.

[0238] Specifically, for feasibility determination, the offspring candidates are judged according to criterion S3.4 to determine whether they meet all risk perception constraints. Offspring candidates that meet the constraints are marked as feasible candidates and added to the feasibility file. .

[0239] Among them, for feasible archives, feasible archives It is a dynamically growing collection that stores feasible candidates confirmed in each evaluation, along with their chemical composition vectors and objective function values. The maximum file size is denoted as [file size]. When the number of feasible candidates in the file exceeds Eliminate candidates with the smallest congestion distance based on the congestion distance.

[0240] Preferably, the maximum feasible file size The feasible candidates in the file are arranged in order of their joining time. When eliminating candidates, the crowding distance of all candidates in the file is calculated, and the candidate with the smallest crowding distance is removed.

[0241] For example, offspring candidate Cross-2026-002 was evaluated as a viable candidate. Add feasible files The number of feasible candidates in the archives has increased from Updated to The capacity limit has not been exceeded. The corresponding record is identified as Archive-2026-002.

[0242] Step S5.5: For offspring candidates that do not meet the constraints, select the feasible candidate with the closest standardized distance from the feasible file as the feasible anchor point.

[0243] Preferably, the core of this step is to select feasible anchor points from infeasible candidates. By calculating the standardized Euclidean distance between the infeasible offspring candidates and each candidate in the feasible file, the feasible candidate with the closest distance is selected as the reference point for the repair direction.

[0244] Among them, for the standardized distance, the first Infeasible offspring candidates and feasible files The Middle The standardized Euclidean distance of the feasible candidate is denoted as . The calculation formula is: .

[0245] in, Candidate for infeasible offspring Component value, The first in the feasible file Candidate No. Component value, For the first The standard deviation of chemical component number in the feasibility file is used to eliminate differences in the scale of each component.

[0246] Among them, the selection of feasible anchor points is based on feasible archives. The feasible candidate with the smallest standardized distance from the infeasible offspring candidate is selected as the feasible anchor point, and the selection formula is as follows: .

[0247] in, Let be the number of the feasible anchor point in the feasibility file, and let be the chemical composition vector of the feasible anchor point. .

[0248] Preferably, in the standardized distance calculation Take the first one from the feasible files The standard deviation of chemical component number, when Time Alternatives prevent division by zero.

[0249] For example, the chemical composition vector of the infeasible offspring candidate EvalChild-2026-001 The first in the feasible file Candidate No. Standardized distance The minimum value in the archive is selected as the first value. Candidate number 1 is a feasible anchor point. The corresponding record is identified as Anchor-2026-047.

[0250] Step S5.6: According to the preset mixing coefficient sequence arranged from largest to smallest, the offspring candidates and feasible anchor points are gradually mixed. After each mixing, the level satisfaction probability is recalculated until a repair candidate that satisfies the constraint is obtained or the preset mixing coefficient sequence is completed.

[0251] Preferably, the core of this step is to gradually repair infeasible candidates. By using a mixed coefficient sequence from large to small, infeasible offspring candidates are gradually moved toward feasible anchor points and re-evaluated, thus gradually approaching the feasible region while maintaining candidate diversity.

[0252] For the mixed coefficient sequence, the preset mixed coefficient sequence is denoted as: The sequence is arranged in descending order, i.e. , Indicates the first The proportion of offspring candidates moving toward feasible anchor points during step mixing.

[0253] Among them, for gradual mixing, the first The formula for calculating the candidate chemical component vector after mixing is as follows: .

[0254] in, For the first The candidate chemical composition vector after step mixing For the first Step mixing coefficient, For the infeasible offspring candidate chemical composition vector, This is a feasible anchor point chemical composition vector.

[0255] Among them, for the step-by-step re-evaluation, after each mixing, The input classification proxy model calculates the probability of level satisfaction and the total constraint violation. ,when Repair successful. This means repairing candidate sequences and terminating the mixed sequence; after traversing the complete sequence... The constraint was still not met, the repair failed, and the offspring candidate was discarded.

[0256] Preferably, the mixing coefficient sequence Sequence length Arranging candidates from largest to smallest allows them to move significantly towards the feasible region before fine-tuning.

[0257] For example, infeasible child candidate EvalChild-2026-001 and feasible anchor Anchor-2026-047 are progressively repaired. Step mixing coefficient , Model evaluation Still not feasible; step , ,evaluate Still not feasible; step , ,evaluate Repair successful. To repair the candidate and terminate the mixed sequence, the corresponding record is identified as Repair-2026-001.

[0258] Beneficial effects of steps S5.1 to S5.6: This series of steps addresses the problem of reduced search efficiency caused by directly discarding infeasible offspring in constrained multi-objective search. It generates offspring candidates by guiding differential mutation through variable importance. For offspring candidates that do not meet the constraints, the nearest feasible anchor point is selected from the feasible archive and repaired step by step according to the mixing coefficient sequence, thereby expanding feasible search paths and improving population utilization.

[0259] Specifically, step S5.1 generates offspring candidates by performing crossover and mutation on the selected candidates, introducing new candidate solutions into the search space; step S5.2 determines the variation amplitude of each chemical component by the importance of variables, reducing the perturbation amplitude of constraint-related components to reduce the feasibility of mutation destruction; step S5.3 calculates the level satisfaction probability of offspring candidates using a classification surrogate model, providing an evaluation basis for constraint determination; step S5.4 dynamically accumulates feasible reference points by judging whether the offspring meet the constraints and adding those that meet the constraints to the feasible file; step S5.5 selects the feasible candidate with the closest standardized distance from the feasible file as a feasible anchor point, providing a repair direction for infeasible offspring; and step S5.6 gradually mixes and re-evaluates the mixture by progressively mixing the mixing coefficient sequence from large to small, gradually approaching the feasible region while maintaining candidate diversity.

[0260] Step S6: Perform non-dominated sorting on the candidates that meet the constraints to obtain a feasible Pareto candidate set, map the feasible Pareto candidate set to historical tobacco leaf samples and output a list of candidates to be reviewed.

[0261] Furthermore, step S6 specifically includes the following steps: Step S6.1: Perform non-dominated sorting on the candidates that meet the constraints to obtain feasible Pareto candidate sets at different non-dominated levels.

[0262] Preferably, the core of this step is to complete the non-dominated sorting of the final candidates. By performing non-dominated sorting on all candidates that meet the constraints, the optimal boundary of the multi-objective is identified, providing a hierarchical basis for the subsequent candidate pool selection.

[0263] Specifically, for the candidate set that satisfies the constraints, all feasible sub-generation candidates and successfully repaired candidates in S5 are merged to form the candidate set that satisfies the constraints. The total number of candidates is denoted as .

[0264] For non-dominated sorts, in The same non-dominated sorting is performed as in S4.5, with candidate... Dominant Candidate If and only if And there exists at least one target component. After sorting, the non-dominated hierarchy is obtained. .

[0265] Among them, for the feasible Pareto candidate set, the first Non-dominated hierarchy The candidates in the middle are not dominated by any other candidates, and constitute the feasible Pareto optimal front end, which is the feasible Pareto candidate set. The candidates of each non-dominated level constitute the whole of the feasible Pareto candidate set.

[0266] Preferably, the non-dominated sorting uses the fast non-dominated sorting algorithm, and the sorted items are arranged in ascending order of level. This is the optimal level.

[0267] For example, a candidate set that satisfies the constraints middle The candidates were obtained by non-dominated sorting. hierarchy One candidate, hierarchy One candidate, hierarchy One candidate, the rest are distributed at lower levels. middle The candidate frontends constitute the feasible Pareto optimal frontend, and the corresponding record is identified as Pareto-2026-Final.

[0268] Step S6.2: Calculate the mean of the expected levels of multiple core sensory indicators for each candidate in the feasible Pareto candidate set, and use it as the core sensory comprehensive score.

[0269] Preferably, the core of this step is to calculate the comprehensive core sensory score of the candidates. By taking the arithmetic mean of the expected levels of multiple core sensory indicators, the overall core sensory performance of the candidates is quantified, providing a scalar ranking basis for the candidate pool screening.

[0270] Among them, for the core sensory comprehensive score, the first Candidate No. 1's core sensory comprehensive score The calculation formula is: .

[0271] in, For the first The core sensory comprehensive score of candidate number 1. For the first Candidate No. The expected level of the core sensory indicators, The number of core sensory indicators.

[0272] Among them, regarding the range of score values, The range of values ​​is A higher score indicates that the candidate has a better overall expected performance on the core sensory metrics.

[0273] Preferably, the core sensory comprehensive score is stored by candidate number and recorded in association with the chemical composition vector and expected level vector of each candidate.

[0274] For example, the first in the feasible Pareto candidate set Fragrance expectation level of candidate number 1 Aroma lingering taste Core sensory comprehensive score , No. Candidate No. , No. Candidate No. The corresponding record is identified as Score-2026-Final.

[0275] Step S6.3: Select candidates whose core sensory comprehensive scores reach the preset percentile from the feasible Pareto candidate set to form a candidate pool.

[0276] Preferably, the core of this step is to complete the threshold screening of the candidate pool. By setting a preset quantile of the core sensory comprehensive score, high-performing candidates are screened to narrow the scope of review, and the optimization results are focused on the most valuable candidates.

[0277] Here, the preset quantile refers to the screening threshold of the core sensory comprehensive score, denoted as , Take all candidates from the feasible Pareto candidate set. The percentile.

[0278] Specifically, for candidate pool construction, candidates are selected from the feasible Pareto candidate set. The candidate pool consists of candidates. The size of the candidate pool is denoted as .

[0279] Preferably, the preset quantile is several hundredths. That is, selecting the top-ranked core sensory comprehensive scores The candidates constitute the candidate pool, and the size of the candidate pool is... The size of the feasible Pareto candidate set is adaptively determined.

[0280] For example, the feasible Pareto candidate set has a total of The candidate, with the highest overall core sensory score. percentile ,filter Candidates Each constitutes a candidate pool , , of which Candidate No. Selected, No. Candidate No. Not selected; the corresponding record is marked as Pool-2026-Final.

[0281] Step S6.4: Standardize the chemical composition vectors of each candidate in the candidate pool and the chemical composition data of historical tobacco leaf samples, and calculate the Euclidean distance between each candidate and each historical tobacco leaf sample.

[0282] Preferably, the core of this step is to complete the correlation measurement between candidates and historical samples. Through standardization and Euclidean distance calculation, a similarity measurement between virtual candidates and real historical tobacco leaf samples is established, providing a distance basis for mapping.

[0283] Among them, for the standardization process, the chemical composition vector of each candidate in the candidate pool Chemical composition data matrix of historical tobacco leaf samples row vectors in Min-Max normalization has been completed in S1.5. The range does not require additional standardization; the Euclidean distance is calculated directly.

[0284] Among them, for Euclidean distance, the candidate pool of the th Candidate No. and No. Euclidean distance between historical tobacco leaf samples The calculation formula is: .

[0285] in, For the first Candidate No. Normalized values ​​of chemical components, For the first Historical tobacco leaf sample No. Normalized values ​​of chemical components, This represents the number of chemical components.

[0286] Specifically, for the distance matrix, the candidate pool is calculated. One candidate and The total Euclidean distances between all historical tobacco leaf samples constitute a distance matrix. , of which Line 1 Column elements are .

[0287] Preferably, the Euclidean distance calculation is performed in a normalized space to avoid the influence of dimensional differences among chemical components on the distance, and the distance matrix... Arranged by candidate number.

[0288] For example, the candidate pool number Candidate No. Historical Sample No. Number Euclidean distance , is the minimum distance between the candidate and all historical samples, corresponding to the record identifier Dist-2026-001.

[0289] Step S6.5: Select the historical tobacco leaf sample with the smallest Euclidean distance for each candidate as the mapping sample, and organize the desensitization label, core sensory expectation level, risk sensory level satisfaction probability, and review order of the mapping sample into a list of candidates to be reviewed.

[0290] Preferably, the core of this step is to complete the mapping from virtual candidates to real samples and output the list. By selecting the historical tobacco leaf sample with the smallest Euclidean distance as the mapping sample, the optimization result is transformed into a traceable entity object and the list of candidates to be reviewed, including the review order, is output.

[0291] Among them, for the selection of mapping samples, the first one in the candidate pool Candidate number 1 is selected from historical tobacco leaf samples based on the sample with the smallest Euclidean distance as the mapping sample. The selection formula is as follows: .

[0292] in, For the first The candidate mapped sample number is denoted as and the chemical composition vector of the mapped sample is denoted as . The desensitization identifier of the mapped sample is denoted as .

[0293] The review order is based on the comprehensive score of core sensory perception. The candidates in the candidate pool are arranged from highest to lowest, and the sorting number is the review order, denoted as follows: , .

[0294] For the candidate list to be reviewed, each entry in the list includes: a de-identified identifier for the mapped sample. Candidate chemical composition vectors , Mapping sample chemical composition vector Core sensory expectation level , , Probability of meeting risk perception level , Core sensory comprehensive score Euclidean distance Review order .

[0295] Preferably, the candidate list to be reviewed is arranged in the review order, and the review order is... The candidates are ranked first and given priority for expert sensory evaluation. The list output format is structured text, with each item occupying a separate paragraph.

[0296] For example, the candidate pool number Candidate number 1 is mapped to historical sample number 2. Number, mapping sample desensitization identifier Core sensory expectation level , , Risk perception level meets probability , Core sensory comprehensive score Euclidean distance Review order Candidates pending review, No. The record is as follows: Desensitization label: Sample-2026-0127; Core sensory expectation level: Fragrance quality Aroma lingering taste Risk sensory level satisfaction probability: stimulus Mixed gases Overall score Mapping distance Review order .

[0297] The corresponding record is identified as List-2026-Final, all of them. The candidates are output in the order of review, forming a list of candidates to be reviewed.

[0298] Beneficial effects of steps S6.1 to S6.5: This series of steps addresses the problem that virtual candidates in the optimized output are difficult to trace back to real tobacco leaf samples. It obtains a feasible Pareto candidate set through non-dominated sorting, filters the candidate pool according to the core sensory comprehensive score, maps the candidates to the historical tobacco leaf samples with the smallest standardized distance, and outputs a list of candidates to be reviewed with desensitization labels and review order, thus transforming the virtual optimization results into entity objects that can be reviewed by experts.

[0299] Specifically, step S6.1 involves performing non-dominated sorting on candidates that meet the constraints to obtain feasible Pareto candidate sets at different levels to identify the optimal boundary for multiple objectives; step S6.2 involves calculating the mean of the expected levels of multiple core sensory indicators as a comprehensive score to quantify the overall core sensory performance of the candidates; step S6.3 involves selecting candidates whose comprehensive scores reach a preset quantile to form a candidate pool, focusing on high-performing candidates to narrow the scope of review; step S6.4 involves standardizing the chemical composition data of candidates and historical tobacco samples and calculating Euclidean distance to establish a correlation metric between virtual candidates and real samples; and step S6.5 involves selecting the historical tobacco sample with the smallest distance as a mapping sample and organizing the desensitization labels, expected levels, satisfaction probabilities, and review order to output a traceable list of candidates to be reviewed.

[0300] In summary, the overall beneficial effects of steps S1 to S6 in this embodiment are as follows: This method addresses the issues of lost confidence levels and lack of risk sensory constraints in the screening of tobacco raw materials due to hard classification truncation. It employs a classification surrogate model to output ordered sensory level probabilities. The expected level and the level satisfaction probability are used to quantify the core sensory performance and the risk sensory confidence level, respectively. Under a constrained multi-objective optimization framework, the optimal core sensory indicators are pursued simultaneously while controlling the risk indicators to avoid exceeding the limits. For offspring candidates that do not meet the constraints, some infeasible solutions are gradually repaired and restored through feasible anchor points. The optimization results are mapped to real historical tobacco samples to output a traceable list of candidates to be reviewed. This reduces the scope of expert sensory review while improving the stability of candidate screening under strict constraints.

[0301] Step S1 provides a unified data foundation for training the surrogate model by matching and preprocessing the chemical composition data and sensory evaluation data of historical tobacco leaf samples; Step S2 preserves the confidence distribution of candidates at different levels by training a classification surrogate model that outputs ordered sensory level probabilities; Step S3 characterizes the core sensory performance and risk sensory constraints by using expected level and level satisfaction probability, respectively, to avoid misjudgment of boundary candidates due to hard classification truncation; Step S4 generates an initial candidate population by constructing a constrained multi-objective optimization problem, while constraining risk indicators while optimizing core sensory indicators; Step S5 gradually repairs infeasible offspring candidates by using feasible anchors, expanding feasible search paths and improving population utilization; Step S6 transforms virtual candidates into a traceable list of candidates to be reviewed by non-dominated sorting and real sample mapping.

[0302] like Figure 2 As shown, this embodiment provides an example of a raw material screening device for tobacco leaves. In this embodiment, the raw material screening device is applied to the raw material screening method as described in the above embodiment.

[0303] Specifically, the raw material screening device includes a tobacco sample related data acquisition module 1, a classification proxy model construction module 2, a grade satisfaction probability calculation module 3, a constrained multi-objective optimization problem construction module 4, a constrained multi-objective optimization problem solving module 5, and a candidate list output module 6, which are connected electrically or through communication in sequence.

[0304] The module 1, which acquires relevant data from tobacco leaf samples, obtains chemical composition data and sensory evaluation data from near-infrared spectroscopy of historical tobacco leaf samples, and matches and preprocesses the chemical composition data and sensory evaluation data. The module 2, which constructs a classification proxy model, generates ordered sensory level labels based on the sensory evaluation data. Using chemical composition data as input, it trains and outputs ordered sensory level probabilities for core sensory indicators and risk sensory indicators respectively. The module 3, which calculates the level satisfaction probability, inputs the pre-acquired candidate tobacco leaf raw material chemical composition vectors into the classification proxy model, calculates the expected level of the core sensory indicators based on the predicted probability and level ordinal value of each ordered sensory level, and calculates the level satisfaction probability of the risk sensory indicators based on the sum of predicted probabilities not lower than the preset minimum acceptable level. The multi-objective optimization problem construction module 4 is used to construct a constrained multi-objective optimization problem with the expected level of multiple core sensory indicators as the optimization objective and the probability of satisfying the level of risk sensory indicators not being lower than the corresponding probability threshold as the constraint. It generates an initial candidate population based on chemical composition data, sensory evaluation data, and a preset range of chemical composition values. The constrained multi-objective optimization problem solving module 5 is used to cross and mutate the initial candidate population to generate offspring candidates. For offspring candidates that do not meet the constraints, the nearest feasible anchor point is selected from the feasible archive and repaired step by step according to a preset mixing coefficient sequence. The candidate list output module 6 is used to perform non-dominated sorting on the candidates that meet the constraints to obtain a feasible Pareto candidate set, map the feasible Pareto candidate set to historical tobacco leaf samples, and output a candidate list to be reviewed.

[0305] like Figure 3 As shown, the electronic device 7 includes a processor 71 and a memory 72 coupled to the processor 71.

[0306] The memory 72 stores program instructions for implementing the tobacco raw material screening method of any of the above embodiments.

[0307] The processor 71 is used to execute program instructions stored in the memory 72 to perform raw material screening of tobacco leaves.

[0308] The processor 71 can also be referred to as a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip with signal processing capabilities. The processor 71 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0309] Furthermore, Figure 4 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 4 The storage medium 8 in this embodiment stores program instructions 81 capable of implementing all the above methods. These program instructions 81 can be stored in the storage medium as a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods in each embodiment of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0310] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the raw material screening of tobacco leaves in a unit is only a logical function of tobacco leaf raw material screening. In actual implementation, there may be other methods of tobacco leaf raw material screening. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, signal, or other forms.

[0311] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for screening tobacco leaves as raw materials, characterized in that, The raw material screening method includes: Step S1: Obtain chemical composition data and sensory evaluation data from near-infrared spectroscopy of historical tobacco leaf samples, and match and preprocess the chemical composition data and the sensory evaluation data. Step S2: Generate ordered sensory level labels based on the sensory evaluation data, and train classification proxy models that output ordered sensory level probabilities for core sensory indicators and risk sensory indicators, respectively, using the chemical composition data as input. Step S3: Input the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model, calculate the expected level of the core sensory index based on the predicted probability and level order value of each ordered sensory level, and calculate the level satisfaction probability of the risk sensory index based on the sum of the predicted probabilities of not less than the preset minimum acceptable level. Step S4: Using the expected level of multiple core sensory indicators as the optimization objective and the probability of satisfying the level of the risk sensory indicator not being lower than the corresponding probability threshold as the constraint, construct a constrained multi-objective optimization problem, and generate an initial candidate population based on the chemical composition data, the sensory evaluation data, and the preset chemical composition value range. Step S5: Perform crossover and mutation on the initial candidate population to generate offspring candidates. For offspring candidates that do not meet the constraints, select the nearest feasible anchor point from the feasible file and repair them step by step according to the preset mixing coefficient sequence. Step S6: Perform non-dominated sorting on the candidates that satisfy the constraints to obtain a feasible Pareto candidate set, map the feasible Pareto candidate set to the historical tobacco leaf samples, and output a list of candidates to be reviewed.

2. The raw material screening method according to claim 1, characterized in that, Step S1: Obtain chemical composition data and sensory evaluation data from near-infrared spectroscopy of historical tobacco leaf samples; match and preprocess the chemical composition data and sensory evaluation data, including: Step S1.1: Obtain near-infrared spectral data of historical tobacco leaf samples, and convert the near-infrared spectral data into chemical composition values ​​through a preset chemical composition correction model; Step S1.2: Obtain sensory evaluation data of the historical tobacco leaf samples, and match the chemical composition data with the sensory evaluation data according to the sample identifier; Step S1.3: Perform missing value detection on the matched data, impute missing chemical component fields with median, and impute missing sensory evaluation fields with mean. Step S1.4: Detect outliers in the chemical composition data and mark values ​​that exceed the preset range of chemical composition values ​​as invalid. Step S1.5: Linear interpolation is used to compensate for fields marked as invalid, and Min-Max normalization is performed on the compensated chemical composition data and sensory evaluation data. Step S1.6: Arrange the normalized chemical composition data and sensory evaluation data according to the samples to obtain the preprocessed chemical composition data and sensory evaluation data.

3. The raw material screening method according to claim 1, characterized in that, Step S2 involves generating ordered sensory level labels based on the sensory evaluation data, and training classification proxy models that output ordered sensory level probabilities for core sensory indicators and risk sensory indicators, using the chemical composition data as input. This includes: Step S2.1: Determine the level threshold based on the score distribution of each sensory index in the sensory evaluation data, and divide the sensory scores into three ordered sensory levels: low level, medium level, and high level. Step S2.2: The intermediate level is split into low-medium level and high-medium level to form four ordered levels. Then, the low-medium level and the high-medium level are merged into the intermediate level to obtain three ordered sensory level labels: Q1, Q2, Q3, and Q4. Step S2.3: Determine the core sensory indicators and risk sensory indicators. The core sensory indicators are aroma quality, aroma quantity, and aftertaste. The risk sensory indicators are irritation and off-odors. Step S2.4: Using the chemical composition data as input features, construct training sets and validation sets for the core sensory indicators and the risk sensory indicators, respectively. Step S2.5: Train a random forest classification model for each sensory indicator. The random forest classification model outputs the predicted probability of each ordered sensory level, thus obtaining a classification surrogate model.

4. The raw material screening method according to claim 1, characterized in that, Step S3: Input the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model, calculate the expected level of the core sensory indicator based on the predicted probability and level order value of each ordered sensory level, and calculate the level satisfaction probability of the risk sensory indicator based on the sum of predicted probabilities not lower than the preset minimum acceptable level, including: Step S3.1: Input the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model to obtain the predicted probability of each sensory index under each ordered sensory level; Step S3.2: For the core sensory index, multiply the predicted probability of each ordered sensory level by the corresponding level ordinal value and sum them to obtain the expected level of the core sensory index. Step S3.3: For the risk sensory index, determine a preset minimum acceptable level, and sum the predicted probabilities of each level that is not lower than the minimum acceptable level to obtain the level satisfaction probability of the risk sensory index. Step S3.4: When the probability of the grade is not lower than the corresponding probability threshold, the candidate corresponding to the chemical composition vector of the candidate tobacco raw material is marked as a feasible candidate; otherwise, it is marked as an infeasible candidate.

5. The raw material screening method according to claim 1, characterized in that, Step S4 involves constructing a constrained multi-objective optimization problem, using the expected levels of multiple core sensory indicators as optimization objectives and the constraint that the probability of satisfying the level of the risk sensory indicator is not lower than the corresponding probability threshold. An initial candidate population is then generated based on the chemical composition data, the sensory evaluation data, and a preset range of chemical composition values, including: Step S4.1: Take the expected level of the multiple core sensory indicators as the target to be maximized, and take the probability that the level of each risk sensory indicator is not lower than the corresponding probability threshold as the constraint condition to construct a constrained multi-objective optimization problem. Step S4.2: Extract a portion of samples from the historical samples corresponding to the chemical composition data and the sensory evaluation data as historical candidates; Step S4.3: Randomly generate restricted random candidates within a preset range of chemical composition values, and combine the historical candidates with the restricted random candidates to form an initial candidate population; Step S4.4: Input the chemical composition vector of each candidate in the initial candidate population into the classification proxy model, and calculate the expected level and level satisfaction probability of each candidate; Step S4.5: Sort the initial candidate population in a manner that feasible candidates take precedence over infeasible candidates, and perform non-dominated sorting and crowding distance calculation on the feasible candidates; Step S4.6: Select candidates from the initial candidate population to enter the next generation based on the non-dominated sorting results and crowding distance.

6. The raw material screening method according to claim 1, characterized in that, Step S5: Perform crossover and mutation on the initial candidate population to generate offspring candidates. For offspring candidates that do not meet the constraints, select the nearest feasible anchor point from the feasible file and repair them step by step according to a preset mixing coefficient sequence, including: Step S5.1: Perform a crossover operation on the selected candidates to generate intermediate candidates, and perform a mutation operation on the intermediate candidates to generate offspring candidates; Step S5.2: Based on the importance of the variables in the classification proxy model corresponding to the risk sensory index, determine the variation range of each chemical component, and constrain the variation range of the relevant chemical components to be less than the variation range of the other chemical components. Step S5.3: Input the offspring candidates into the classification proxy model and calculate the probability of each offspring candidate meeting the level requirement; Step S5.4: Determine whether the offspring candidate satisfies the constraint, and add the offspring candidate that satisfies the constraint to the feasible file; Step S5.5: For offspring candidates that do not meet the constraints, select the feasible candidate with the closest standardized distance from the feasible file as the feasible anchor point; Step S5.6: According to the preset mixing coefficient sequence arranged from largest to smallest, the offspring candidates are gradually mixed with the feasible anchor points. After each mixing, the level satisfaction probability is recalculated until a repair candidate that satisfies the constraint is obtained or the preset mixing coefficient sequence is completed.

7. The raw material screening method according to claim 1, characterized in that, Step S6: Perform non-dominated sorting on the candidates that satisfy the constraints to obtain a feasible Pareto candidate set, map the feasible Pareto candidate set to the historical tobacco leaf samples, and output a list of candidates to be reviewed, including: Step S6.1: Perform non-dominated sorting on the candidates that satisfy the constraints to obtain feasible Pareto candidate sets at different non-dominated levels; Step S6.2: Calculate the mean of the expected levels of multiple core sensory indicators for each candidate in the feasible Pareto candidate set, and use it as the core sensory comprehensive score. Step S6.3: Select candidates whose core sensory comprehensive score reaches a preset percentile from the feasible Pareto candidate set to form a candidate pool; Step S6.4: Standardize the chemical composition vectors of each candidate in the candidate pool and the chemical composition data of the historical tobacco leaf samples, and calculate the Euclidean distance between each candidate and each historical tobacco leaf sample. Step S6.5: Select the historical tobacco leaf sample with the smallest Euclidean distance for each candidate as the mapping sample, and organize the desensitization identifier, core sensory expectation level, risk sensory level satisfaction probability, and review order of the mapping sample into a candidate list to be reviewed.

8. A raw material screening device for tobacco leaves, wherein the raw material screening device is applied to the raw material screening method as described in any one of claims 1 to 7, characterized in that, The raw material screening device includes: The tobacco leaf sample related data acquisition module is used to acquire chemical composition data and sensory evaluation data of historical tobacco leaf samples by near-infrared detection, and to match and preprocess the chemical composition data and the sensory evaluation data; The classification proxy model construction module is used to generate ordered sensory level labels based on the sensory evaluation data, and to train and output the probability of ordered sensory levels for the core sensory indicators and risk sensory indicators respectively, using the chemical composition data as input. The grade satisfaction probability calculation module is used to input the pre-acquired candidate tobacco raw material chemical composition vector into the classification proxy model, calculate the expected grade of the core sensory index according to the predicted probability and grade order value of each ordered sensory grade, and calculate the grade satisfaction probability of the risk sensory index according to the sum of the predicted probabilities not lower than the preset minimum acceptable grade. The constrained multi-objective optimization problem construction module is used to construct a constrained multi-objective optimization problem with the expected level of multiple core sensory indicators as the optimization objective and the probability of satisfying the level of the risk sensory indicator not being lower than the corresponding probability threshold as the constraint, and to generate an initial candidate population based on the chemical composition data, the sensory evaluation data, and the preset chemical composition value range. The constrained multi-objective optimization problem solving module is used to perform crossover and mutation on the initial candidate population to generate offspring candidates. For offspring candidates that do not meet the constraints, the nearest feasible anchor point is selected from the feasible file and repaired step by step according to a preset mixing coefficient sequence. The candidate list output module is used to perform non-dominated sorting on the candidates that meet the constraints to obtain a feasible Pareto candidate set, map the feasible Pareto candidate set to the historical tobacco leaf samples, and output the candidate list to be reviewed.

9. An electronic device, characterized in that, The method includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the raw material screening method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed by a processor, enable the raw material screening method as described in any one of claims 1 to 7.