A cross-scale prediction method and system for nine-grain flavor liquor

Abstract

The application discloses a kind of nine grain aroma type Baijiu flavor cross-scale prediction method and system, it is related to Baijiu fermentation process intelligent prediction, flavor chemistry and biological process engineering cross technical field.The application is aimed at the problem that there is systematic difference in laboratory scale and industrial scale Baijiu flavor, by establishing paired laboratory and industrial scale fermentation samples, the scale transfer coefficient specific to compound and its logarithmic form are calculated as prediction target;Formula perception prediction feature matrix is constructed, which contains laboratory flavor phenotype and daqu formula proportion simultaneously;Leave-one-formula cross-validation is used to train regression model;The laboratory flavor detection value of the formula to be predicted and the daqu ratio are input into the model, the transfer parameters are predicted and the industrial scale flavor concentration is calculated.The application can effectively avoid formula level information leakage, and realize reliable prediction of unseen formula industrial flavor.

Description

Technical Field

[0001] This invention relates to the fields of intelligent prediction of the fermentation process of baijiu, flavor chemistry and biological process engineering, and in particular to a cross-scale prediction method and system for the flavor of nine-grain baijiu. Background Technology

[0002] Nine-grain aroma baijiu employs a combination of different types of koji (fermentation starter) for co-fermentation. Its flavor profile is influenced by a multitude of factors, including the koji formula, fermentation microbial ecology, cellar mud microorganisms, temperature variations, mass transfer conditions, and fermentation history. While laboratory fermentation offers advantages such as a short cycle, controllable conditions, and suitability for multi-formula screening, there are systematic differences between laboratory and industrial fermentation pits in terms of microbial ecology and environmental conditions. Therefore, relying solely on the absolute content or simple similarity of a flavor compound in a laboratory sample is insufficient to reliably predict the actual flavor profile of that formula in industrial production.

[0003] Existing methods mainly suffer from the following problems: ① They typically compare the flavor concentrations of laboratory and industrial samples directly, without explicitly modeling the "scale transformation of flavor substances when migrating from the laboratory scale to the industrial scale"; ② When randomly dividing the training and test sets, information from the same formula may simultaneously enter both sets, easily leading to information leakage at the formula level and thus overestimating the model's generalization ability; ③ There is significant collinearity among multiple components and volatile flavor indices in Daqu formulas, and ordinary regression or single nonlinear models are prone to parameter instability under limited sample conditions; ④ Existing models rarely incorporate laboratory flavor phenotypes and Daqu formula backgrounds simultaneously, and lack rigorous hold-out validation for "unknown new formulas"; ⑤ Model prediction results usually lack experimental correspondence with industrial ecological factors such as pit mud and temperature.

[0004] To address the aforementioned issues, proposing a cross-scale prediction method and system for the flavor of nine-grain aroma-type baijiu to overcome the difficulties of existing technologies has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a cross-scale prediction method and system for the flavor of nine-grain aroma baijiu. Based on paired laboratory-industrial GC-MS flavor data, the method defines industrial-scale variation as a compound-specific scale migration coefficient and models its logarithmic value. Laboratory flavor phenotypes and the continuous proportions of various Daqu (fermented starter culture) are used as predictive variables. Leave-One-Formula-Out Cross-Validation (LOFO-CV) is employed to evaluate the model's generalization ability to unseen formulas grouped by complete formulas. Furthermore, the predicted migration direction can be mechanistically verified through cellar mud bridging and temperature bridging experiments. This achieves a quantitative transformation from "laboratory flavor detection" to "industrial flavor prediction," providing a technical means for screening new formulas for nine-grain aroma baijiu, prioritizing industrial trials, and scaling up production.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A cross-scale prediction method for the flavor of nine-grain aroma-type baijiu includes the following steps: S1: Establish multiple Daqu (fermentation starter) formulas and obtain paired laboratory-scale fermentation samples and industrial-scale fermentation samples for each formula; S2: Volatile flavor analysis was performed on laboratory-scale and industrial-scale fermentation samples to obtain concentration data of target flavor substances in each sample. S3: For each target flavor compound in each formulation, calculate the migration parameter from laboratory scale to industrial scale. The migration parameter is the logarithmic form of the scale migration coefficient calculated based on laboratory concentration and industrial concentration. S4: Construct a prediction feature matrix. Each row of the prediction feature matrix corresponds to a formula-compound observation sample. The feature variables of each row include: the concentration data of each target flavor substance at the laboratory scale corresponding to the current Daqu formula, and the proportion data of each component Daqu in the current Daqu formula; the label variable of the prediction feature matrix is ​​the migration parameter calculated in S3. S5: Construct at least one candidate regression model, use leave-one-out-of-one-formula cross-validation (LOFO-CV) to complete the training, tuning, and performance comparison of the candidate regression models, and select the model with the best generalization performance as the cross-scale prediction model. S6: Input the Daqu ratio data and laboratory-scale flavor concentration data of the Daqu formula to be predicted into the trained cross-scale prediction model to predict the migration parameters of each target flavor substance in the current Daqu formula. S7: Based on the migration parameters and the laboratory-scale flavor concentration data of the Daqu formula to be predicted, the predicted flavor concentration value of the current Daqu formula at the industrial scale is calculated.

[0007] Optionally, the target flavor compounds in S2 can be determined using the following three-layer screening strategy: S21. Screening for commercial product specificity: Compare the volatile flavor profiles of different commercial baijiu to obtain representative candidate characteristic substances of Jiuliangxiang-type baijiu. S22. Industrial Enrichment Screening: Calculating the concentration of candidate substances in industrial raw liquor. The ratio of the industrial enrichment coefficient to the average concentration in multiple laboratory Daqu (a type of starter culture) formulation samples is called the industrial enrichment coefficient. :

[0008] in, Candidate material Concentration in industrial bulk spirits Candidate material Average concentration in multiple laboratory Daqu (a type of starter culture) formulation samples; S23, Aroma Activity Screening: Calculation ,in For compounds concentration, For compounds Odor threshold.

[0009] Alternatively, the formula for calculating the S3 mesoscale migration coefficient can be used as follows: For the formula Target flavor compounds Define the scale migration coefficient :

[0010] in, Formula for Daqu Compounds at the industrial scale concentration, Compounds of the same Daqu formula on a laboratory scale concentration, A positive correction term is used to prevent zero values ​​from causing abnormalities in ratios or logarithmic operations; scale transfer coefficient Perform a logarithmic transformation to obtain the model response variables: .

[0011] Optionally, in the above method, the regression model in S5 can be selected from one or more of partial least squares regression, random forest regression, support vector regression, and ElasticNet regression.

[0012] Optionally, the regression model chosen in the above method is ElasticNet regression, whose objective function includes both L1 and L2 regularization terms.

[0013] Optionally, in the above method, the formula of the Daqu to be predicted in S7 can be... and compounds The predicted scale migration coefficient is obtained from the model. Then, the industrial-scale concentration is reconstructed using the following formula: ; in, This represents the cross-scale migration coefficient of flavor compounds predicted by the model. This indicates the concentration of flavor compounds in the Daqu (a type of starter culture) formula to be predicted under laboratory fermentation conditions. This represents the predicted flavor concentration under factory production conditions, reconstructed using a cross-scale migration model.

[0014] Optionally, the above method also includes S8: conducting pit mud microbial bridging and / or industrial temperature trajectory bridging experiments on representative Daqu formulas, and calculating the migration recovery index between the bridging flavor change vector and the model-predicted laboratory-to-industrial migration vector. TRI Cosine similarity was used to verify the rationality of the prediction direction mechanism; Migration Recovery Index: ; Cosine similarity: ; in, This is the model-predicted laboratory-to-industry migration vector. This is the change vector generated by the bridging process relative to the laboratory baseline.

[0015] A cross-scale prediction system for the flavor of nine-grain aroma baijiu, used to implement the cross-scale prediction method for the flavor of nine-grain aroma baijiu as described above, includes a data acquisition module, a feature construction module, a model training module, and a prediction module connected in sequence. The data acquisition module is used to acquire the concentration data of target flavor substances in paired laboratory-scale and industrial-scale fermentation samples corresponding to multiple Daqu (Chinese starter culture) formulas. The feature construction module is used to calculate the migration parameters of each target flavor substance in each Daqu formula based on the concentration data, and to construct a prediction feature matrix. The feature variables of the prediction feature matrix include laboratory flavor phenotype data and Daqu formula ratio data. The model training module is used to train the regression model by employing the leave-one-out cross-validation method, with the complete Daqu recipe as the smallest leave-out unit, taking all the observation samples corresponding to each Daqu recipe as the validation set in sequence, and the remaining Daqu recipes as the training set. The prediction module is used to predict migration parameters based on the Daqu ratio data of the Daqu formula to be predicted and its laboratory-scale flavor concentration data, and to back-calculate the predicted value of industrial-scale flavor concentration.

[0016] In the above system, the regression model can optionally be the ElasticNet regression model.

[0017] Optionally, the system described above also includes a mechanism verification module for conducting microbial bridging and / or industrial temperature trajectory bridging experiments on representative formulations, and calculating the migration recovery index and cosine similarity to verify the rationality of the predicted direction mechanism.

[0018] As can be seen from the above technical solution, compared with the prior art, the present invention provides a cross-scale prediction method and system for the flavor of nine-grain aroma-type baijiu, which has the following beneficial effects: (1) Transform the laboratory-industrial scale difference into a calculable compound-specific migration parameter. Using K=(C~factory~+δ) / (C~lab~+δ) and its logarithmic form lnK, the laboratory concentration is no longer assumed to be directly equivalent to the industrial concentration, but the magnitude of change of each target flavor substance during the scale-up process is explicitly described.

[0019] (2) Improve the ability to evaluate the true generalization of unknown Daqu formulas. Strict LOFO-CV is adopted, with the "complete formula" as the smallest leaveout unit, to avoid data leakage caused by the same formula information entering the training set and the test set at the same time.

[0020] (3) Taking into account both chemical phenotype and formulation background. The model simultaneously inputs laboratory flavor phenotype and continuous Daqu ratio, which is more suitable for describing the formulation-dependent scale transfer of nine-grain aroma baijiu.

[0021] (4) It has good stability for highly correlated, small-sample flavor data. ElasticNet is preferred because it uses L1 and L2 regularization to jointly constrain the relevant predictor variables, which effectively reduces the coefficient instability caused by multicollinearity.

[0022] (5) It can directly output industrial-scale flavor prediction results, providing technical means for screening new formulas of Jiuliang-flavored Baijiu, prioritizing industrial trials and scaling up.

[0023] (6) It can form a closed loop of mechanism verification with industrial ecology bridging experiments, avoiding the model performance being artificially high due to repeated use of samples. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a flowchart of a cross-scale prediction method for the flavor of nine-grain aroma-type baijiu disclosed in this invention. Figure 2 This is a block diagram of a cross-scale prediction system for the flavor of a nine-grain aroma-type baijiu disclosed in this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0028] This invention discloses a cross-scale prediction method and system for the flavor of nine-grain aroma-type Baijiu (Chinese liquor). Based on flavor phenotypes obtained from laboratory-scale fermentation and Daqu (fermentation starter) formula information, it predicts the characteristic flavor changes and flavor composition of the industrial liquor under the same formula at the industrial production scale. The specific technical solution is as follows: See Figure 1 As shown, this invention discloses a cross-scale prediction method for the flavor of nine-grain aroma-type baijiu, comprising the following steps: S1: Establish multiple Daqu (fermentation starter) formulas and obtain paired laboratory-scale fermentation samples and industrial-scale fermentation samples for each formula; S2: Volatile flavor analysis was performed on laboratory-scale and industrial-scale fermentation samples to obtain concentration data of target flavor substances in each sample. S3: For each target flavor compound in each formulation, calculate the migration parameter from laboratory scale to industrial scale. The migration parameter is the logarithmic form of the scale migration coefficient calculated based on laboratory concentration and industrial concentration. S4: Construct a prediction feature matrix. Each row of the prediction feature matrix corresponds to a formula-compound observation sample. The feature variables of each row include: the concentration data of each target flavor substance at the laboratory scale corresponding to the current Daqu formula, and the proportion data of each component Daqu in the current Daqu formula; the label variable of the prediction feature matrix is ​​the migration parameter calculated in S3. S5: Construct at least one candidate regression model, use leave-one-out-of-one-formula cross-validation (LOFO-CV) to complete the training, tuning, and performance comparison of the candidate regression models, and select the model with the best overall generalization performance as the cross-scale prediction model. S6: Input the Daqu ratio data and laboratory-scale flavor concentration data of the Daqu formula to be predicted into the trained cross-scale prediction model to predict the migration parameters of each target flavor substance in the current Daqu formula. S7: Based on the migration parameters and the laboratory-scale flavor concentration data of the Daqu formula to be predicted, the predicted flavor concentration value of the current Daqu formula at the industrial scale is calculated.

[0029] Further, predictor variable construction: Predictor variables include at least two categories: (1) Laboratory flavor phenotype: the concentration or transformation value of the industrial characteristic flavor substances to be modeled in laboratory fermented wine samples; (2) Daqu formula information: the continuous mass ratio of various types of Daqu in the formula.

[0030] This embodiment uses eight types of starter cultures as formulation variables and 14 industrial characteristic flavor indicators to construct the laboratory chemical phenotype. The eight starter cultures include vanilla medium-temperature starter culture, self-made medium-temperature starter culture, Luzhou medium-temperature starter culture, Guizhou high-temperature starter culture, white starter culture, aroma-producing yeast, bacterial starter culture, and high-yield starter culture. For ease of model implementation, they can be sequentially coded as DQ1-DQ8. The sum of the proportions of each starter culture is 100%.

[0031] Furthermore, the target flavor compounds in S2 were determined using the following three-layer screening strategy: S21. Screening for commercial product specificity: Compare the volatile flavor profiles of different commercial baijiu to obtain representative candidate characteristic substances of Jiuliangxiang-type baijiu. S22. Industrial Enrichment Screening: Calculating the concentration of candidate substances in industrial raw liquor. The ratio of the industrial enrichment coefficient to the average concentration in multiple laboratory Daqu (a type of starter culture) formulation samples is called the industrial enrichment coefficient. :

[0032] in, Candidate material Concentration in industrial bulk spirits Candidate material Average concentration in multiple laboratory Daqu (a type of starter culture) formulation samples; S23, Aroma Activity Screening: Calculation ,in For compounds concentration, For compounds Odor threshold, preferably retained Aroma-active compounds >1.

[0033] In this embodiment, 14 industrial characteristic flavor markers were obtained using the above method. These 14 markers are used for subsequent formulation evaluation and cross-scale prediction. For practical applications, the target flavor indicators can be added or removed according to the product aroma, detection platform, and industrial objectives, and are not limited to a fixed 14, as long as there are corresponding quantitative or semi-quantitative indicators between laboratory and industrial samples.

[0034] Furthermore, the formula for calculating the S3 mesoscale migration coefficient is as follows: For the formula Target flavor compounds Define the scale migration coefficient :

[0035] in, Formula for Daqu Compounds at the industrial scale concentration, Compounds of the same Daqu formula on a laboratory scale concentration, A positive correction term to prevent zero values ​​from causing abnormalities in ratios or logarithmic operations.

[0036] This embodiment is preferred. =1.

[0037] scale transfer coefficient Perform a logarithmic transformation to obtain the model response variables: .

[0038] Compared to directly predicting industrial concentrations, this treatment explicitly expresses cross-scale variations as the magnitude of the transformation of the same flavor compound from laboratory to industrial conditions, which can reduce the impact of differences in the absolute concentration magnitudes of different compounds on model training.

[0039] Furthermore, the regression model in S5 is selected from one or more of Partial Least Squares Regression (PLSR), Random Forest Regression, Support Vector Regression (SVR), and ElasticNet Regression.

[0040] Furthermore, the regression model chosen is ElasticNet regression, whose objective function includes both L1 and L2 regularization terms, which can compress redundant coefficients while retaining relevant effective variables.

[0041] Model evaluation employs strict leave-one-out-of-formula cross-validation (LOFO-CV), rather than randomly splitting all formulations × compound observations. In each outer validation trade-off, all target flavor observations corresponding to a complete formulation are simultaneously removed from the training set, with the remaining formulations used for model training. Model preprocessing, standardization, and hyperparameter selection are performed only within the training data, and then predictions are made on the removed complete formulations. This process iterates through all formulations, ensuring each formulation is predicted once as an unseen formulation.

[0042] Model evaluation metrics should include at least the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). It is preferable to simultaneously calculate both the pooled metric (which includes all retained observations) and the macro metric (which is calculated separately for each flavor compound and then averaged) to avoid excessive dominance of the overall metric by high-abundance or easily predictable compounds.

[0043] Furthermore, in S7, the formula for predicting Daqu (a type of starter culture) is... and compounds The predicted scale migration coefficient is obtained from the model. Then, the industrial-scale concentration is reconstructed using the following formula: ; in, These are the scale migration parameters predicted by the model. To determine the laboratory-scale flavor concentration of the Daqu (a type of starter culture) formula to be predicted. The predicted flavor concentration under factory production conditions is obtained by reconstructing the cross-scale migration model.

[0044] Therefore, by simply obtaining the proportion of Daqu (a type of starter culture) in the formula to be predicted and the laboratory-scale flavor test results, the predicted concentrations of each target flavor substance in the formula at the industrial scale can be output, and an industrial flavor prediction spectrum can be formed.

[0045] Furthermore, it also includes S8: conducting pit mud microbial bridging and / or industrial temperature trajectory bridging experiments on representative Daqu formulas, and calculating the migration recovery index between the bridging flavor change vector and the model-predicted laboratory-to-industrial migration vector. TRI Cosine similarity was used to verify the rationality of the prediction direction mechanism; Migration Recovery Index: ; Cosine similarity: ; in, This is the model-predicted laboratory-to-industry migration vector. This is the change vector generated by the bridging process relative to the laboratory baseline.

[0046] TRI The cosine similarity reflects the magnitude of the recovery along the predicted industrial migration axis after bridging processing, while the cosine similarity reflects the degree of consistency between the bridging change and the predicted migration direction. Bridging samples are not used as additional supervised training samples to avoid artificially increasing the training sample size.

[0047] and Figure 1 Corresponding to the method shown, this invention also discloses a cross-scale prediction system for the flavor of nine-grain aroma-type baijiu, used to achieve... Figure 1 The method for cross-scale prediction of the flavor of nine-grain aroma-type baijiu is shown below. The system structure diagram is shown in [link to system block diagram]. Figure 2 It includes a data acquisition module, a feature construction module, a model training module, and a prediction module connected in sequence; The data acquisition module is used to acquire the concentration data of target flavor substances in paired laboratory-scale and industrial-scale fermentation samples corresponding to multiple Daqu (Chinese starter culture) formulas. The feature construction module is used to calculate the migration parameters of each target flavor substance in each Daqu formula based on the concentration data, and to construct a prediction feature matrix. The feature variables of the prediction feature matrix include laboratory flavor phenotype data and Daqu formula ratio data. The model training module is used to train the regression model by employing the leave-one-out cross-validation method, with the complete Daqu recipe as the smallest leave-out unit, taking all the observation samples corresponding to each Daqu recipe as the validation set in sequence, and the remaining Daqu recipes as the training set. The prediction module is used to predict migration parameters based on the Daqu ratio data of the Daqu formula to be predicted and its laboratory-scale flavor concentration data, and to back-calculate the predicted value of industrial-scale flavor concentration.

[0048] Furthermore, the regression model is the ElasticNet regression model.

[0049] Furthermore, it also includes a mechanism verification module, which is used to conduct pit mud microbial bridging and / or industrial temperature trajectory bridging experiments on representative formulations, and calculate migration recovery index and cosine similarity to verify the rationality of the predicted direction mechanism.

[0050] Example: (I) Establishment of a cross-scale dataset of 47 sets of Daqu formulas for nine-grain aroma-type Baijiu Forty-seven Daqu (fermentation starter) formulations (Formula 1-Formula 47) were constructed using a constrained D-optimal mixing design. Each formulation consisted of eight different starter ingredients in a continuous ratio, with a total ratio of 100%. Laboratory fermentation samples and industrial production samples of the same formulations were obtained for each of the 47 formulations. Volatile flavor components were detected using a unified GC-MS method. Fourteen characteristic industrial flavor markers were identified through a three-layer screening process: "commercial product specificity—industrial enrichment—OAV". This resulted in a data structure of 47 formulations × 14 target flavor substances, comprising 658 cross-scale observations of "formula-compound".

[0051] .

[0052] Using laboratory flavor phenotypes and the continuous proportions of eight different types of Daqu (a type of starter culture) as inputs, Build a model for the prediction target.

[0053] (II) Comparison of candidate algorithms and determination of the optimal model

[0054] PLSR, RF, SVR, and ElasticNet models were built respectively. Strict LOFO-CV was employed: one complete formula and its 14 target observations were left out each time, and only the remaining 46 formulas were used for model training and internal parameter tuning; after 47 iterations, 658 true leaveout predictions from unseen formulas were obtained. The models were evaluated using R², RMSE, and MAE, and pooled and macro performance were also examined.

[0055] Experimental results show that ElasticNet performs best on this dataset, thus it is selected as the preferred cross-scale prediction algorithm. Under strict LOFO-CV, ElasticNet achieves a pooled R² of 0.8565 for the migration coefficients; after back-calculating the predicted migration coefficients into industrial flavor concentrations, the pooled R² for industrial flavor reconstruction reaches 0.8462. These results demonstrate that this method can learn the laboratory-industrial scale relationship from existing recipes and predict industrial flavors for recipes completely unseen during training.

[0056] (III) Reconstruction of Industrial Flavor

[0057] For any given recipe, the concentrations of 14 characteristic flavors from its laboratory and the proportions of 8 different Daqu (fermented starter cultures) are input into a trained ElasticNet model to obtain predictions for 14 compounds. Value. Then according to The industrial-scale concentrations are calculated one by one to form a predicted industrial flavor profile for the formulation. By comparing the predicted values ​​with the actual industrial GC-MS results, formulation-level R², RMSE, and MAE can be obtained to evaluate the industrial predictability of the formulation.

[0058] (iv) Verification of the bridging mechanism of representative formulations

[0059] Formulas 20, 35, and 47 were selected for bridging of pit mud microorganisms (B2) and industrial temperature trajectory (B3). The bridged samples were not included in the supervised model training of the 47 formulations. Instead, the TRI and cosine similarity between the bridged flavor change vector and the industrial migration vector predicted by the LOFO model were calculated independently.

[0060] For Formula 20, the TRI for the B2 treatment was 6.39% with a cosine similarity of 0.390; for the B3 treatment, the TRI was 1.75% with a cosine similarity of 0.139. For Formula 35, the TRI for the B2 treatment was 11.85% with a cosine similarity of 0.617; for the B3 treatment, the TRI was 19.28% with a cosine similarity of 0.765. For Formula 47, the TRI for the B2 treatment was 16.22% with a cosine similarity of 0.415; for the B3 treatment, the TRI was 19.54% with a cosine similarity of 0.500. The results indicate that different Daqu (a type of Chinese liquor) formulations have different responses to the same industrial ecological factors. Furthermore, the temperature bridging changes in Formulas 35 and 47 significantly follow the industrial migration direction predicted by the model, providing experimental support for cross-scale migration dependent on formulations.

[0061] Example 2: Comparative Example 1: Directly substituting laboratory concentration for industrial concentration If a scale migration coefficient is not established and the laboratory-detected concentration is directly regarded as the industrial production concentration, it is impossible to correct for the systematic bias between the laboratory and industrial environments caused by factors such as microbial ecology, temperature, mass transfer, and fermentation history. This invention establishes a scale migration coefficient for each compound. and By treating scale bias as an independent prediction object, industrial flavor concentration can be further reconstructed.

[0062] Comparative Example 2: Randomly splitting "formula × compound" for observation

[0063] If the 658 observations are randomly split into a training set and a test set, information about the same Daqu (a type of Chinese liquor) formula may exist in both sets simultaneously, resulting in information leakage at the formula level. Consequently, the accuracy obtained cannot truly reflect the model's predictive ability for new formulas. This invention employs strict LOFO-CV, leaving out an entire formula each time, which can directly test the generalization ability to formulas without prior knowledge.

[0064] Comparative Example 3: Different Regression Algorithms

[0065] Under the same cross-scale data and validation framework, PLSR, RF, SVR, and ElasticNet were compared. In the final data analysis, ElasticNet achieved the best overall generalization performance, indicating that for flavor data with multivariate correlations and limited sample size, L1 / L2 joint regularization can effectively reduce the coefficient instability problem caused by multicollinearity.

[0066] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cross-scale prediction method for the flavor of nine-grain aroma-type baijiu, characterized in that, Includes the following steps: S1: Establish multiple Daqu (fermentation starter) formulas and obtain paired laboratory-scale fermentation samples and industrial-scale fermentation samples for each formula; S2: Volatile flavor analysis was performed on laboratory-scale and industrial-scale fermentation samples to obtain concentration data of target flavor substances in each sample. S3: For each target flavor compound in each formulation, calculate the migration parameters from the laboratory scale to the industrial scale. The migration parameters are in the logarithmic form of the scale migration coefficients calculated based on laboratory and industrial concentrations. S4: Construct a prediction feature matrix. Each row of the prediction feature matrix corresponds to a formula-compound observation sample. The feature variables of each row include: the concentration data of each target flavor substance at the laboratory scale corresponding to the current Daqu formula, and the proportion data of each component Daqu in the current Daqu formula; the label variable of the prediction feature matrix is ​​the migration parameter calculated in S3. S5: Construct at least one candidate regression model, use leave-one-out-of-one-formula cross-validation (LOFO-CV) to complete the training, tuning, and performance comparison of the candidate regression models, and select the model with the best generalization performance as the cross-scale prediction model. S6: Input the Daqu ratio data and laboratory-scale flavor concentration data of the Daqu formula to be predicted into the trained cross-scale prediction model to predict the migration parameters of each target flavor substance in the current Daqu formula. S7: Based on the migration parameters and the laboratory-scale flavor concentration data of the Daqu formula to be predicted, the predicted flavor concentration value of the current Daqu formula at the industrial scale is calculated.

2. The method for cross-scale prediction of the flavor of nine-grain aroma-type baijiu according to claim 1, characterized in that, The target flavor compounds in S2 were determined using the following three-layer screening strategy: S21. Screening for commercial product specificity: Compare the volatile flavor profiles of different commercial baijiu to obtain representative candidate characteristic substances of Jiuliangxiang-type baijiu. S22. Industrial Enrichment Screening: Calculating the concentration of candidate substances in industrial raw liquor. The ratio of the industrial enrichment coefficient to the average concentration in multiple laboratory Daqu (a type of starter culture) formulation samples is called the industrial enrichment coefficient. : in, Candidate material Concentration in industrial bulk spirits Candidate material Average concentration in multiple laboratory Daqu (a type of starter culture) formulation samples; S23, Aroma Activity Screening: Calculation ,in For compounds concentration, For compounds Odor threshold.

3. The method for cross-scale prediction of the flavor of nine-grain aroma-type baijiu according to claim 1, characterized in that, The formula for calculating the S3 mesoscale migration coefficient is as follows: For the formula Target flavor compounds Define the scale migration coefficient : in, Formula for Daqu Compounds at the industrial scale concentration, Compounds of the same Daqu formula on a laboratory scale concentration, A positive correction term is used to prevent zero values ​​from causing abnormalities in ratios or logarithmic operations; scale transfer coefficient Perform a logarithmic transformation to obtain the model response variables: 。 4. The method for cross-scale prediction of the flavor of nine-grain aroma-type baijiu according to claim 1, characterized in that, In S5, the regression model is selected from one or more of partial least squares regression, random forest regression, support vector regression, and ElasticNet regression.

5. The method for cross-scale prediction of the flavor of nine-grain aroma-type baijiu according to claim 4, characterized in that, The regression model chosen is ElasticNet regression, whose objective function includes both L1 and L2 regularization terms.

6. The method for cross-scale prediction of the flavor of nine-grain aroma-type baijiu according to claim 1, characterized in that, S7 contains the Daqu recipe for prediction. and compounds The predicted scale migration coefficient is obtained from the model. Then, the industrial-scale concentration is reconstructed using the following formula: ； in, These are the scale migration parameters predicted by the model. To determine the laboratory-scale flavor concentration of the Daqu (a type of starter culture) formula to be predicted. The predicted flavor concentration under factory production conditions is obtained by reconstructing the cross-scale migration model.

7. The method for cross-scale prediction of the flavor of nine-grain aroma-type baijiu according to claim 1, characterized in that, This also includes S8: conducting cellar mud microbial bridging and / or industrial temperature trajectory bridging experiments on representative Daqu formulas, and calculating the migration recovery index between the bridging flavor change vector and the model-predicted laboratory-to-industrial migration vector. TRI Cosine similarity was used to verify the rationality of the prediction direction mechanism; Migration Recovery Index: ； Cosine similarity: ； in, Model-predicted laboratory-to-industry migration vectors This is the change vector generated by the bridging process relative to the laboratory baseline.

8. A cross-scale prediction system for the flavor of nine-grain aroma-type baijiu, characterized in that, A method for cross-scale prediction of the flavor of a nine-grain aroma-type baijiu as described in any one of claims 1-7 includes a data acquisition module, a feature construction module, a model training module, and a prediction module connected in sequence. The data acquisition module is used to acquire the concentration data of target flavor substances in paired laboratory-scale and industrial-scale fermentation samples corresponding to multiple Daqu (Chinese starter culture) formulas. The feature construction module is used to calculate the migration parameters of each target flavor substance in each Daqu formula based on the concentration data, and to construct a prediction feature matrix. The feature variables of the prediction feature matrix include laboratory flavor phenotype data and Daqu formula ratio data. The model training module is used to train the regression model by employing the leave-one-out cross-validation method, with the complete Daqu recipe as the smallest leave-out unit, taking all the observation samples corresponding to each Daqu recipe as the validation set in sequence, and the remaining Daqu recipes as the training set. The prediction module is used to predict migration parameters based on the Daqu ratio data of the Daqu formula to be predicted and its laboratory-scale flavor concentration data, and to back-calculate the predicted value of industrial-scale flavor concentration.

9. The cross-scale prediction system for the flavor of nine-grain aroma-type Baijiu according to claim 8, characterized in that, The regression model is the ElasticNet regression model.

10. The cross-scale prediction system for the flavor of nine-grain aroma-type Baijiu according to claim 8, characterized in that, It also includes a mechanism verification module, which is used to conduct pit mud microbial bridging and / or industrial temperature trajectory bridging experiments on representative formulations, and calculate migration recovery index and cosine similarity to verify the rationality of the predicted direction mechanism.

Images