Optimization Method for the Production Process of Irradiated Baijiu Based on PLS-LSBoost Gradient Boosting Tree

Abstract

The present invention discloses an optimization method for the production process of irradiated Baijiu based on the PLS-LSBoost gradient boosting tree. For the first time, the PLS-LSBoost gradient boosting tree model is used in the present invention to establish a model for predicting irradiation process parameters from volatile flavor components according to the change data of volatile flavor components during the irradiation process of Luzhou-flavor Baijiu, wherein the principal components can be extracted by partial least squares (PLS) as needed to improve the performance of the model; the above model is used to predict natural aged Baijiu of different years, and the irradiation technical parameters required to achieve the corresponding aging effect are obtained, so as to optimize the irradiation technical parameters. The irradiation aging parameter optimization method provided by the present invention can be used as a supplement to experimentally explore the irradiation process conditions, saving time and labor, and at the same time can provide a methodological reference for the process optimization of other artificial aging technologies.

Description

Technical Field

[0001] The present invention belongs to the fields of wine age detection and technology optimization, and relates to a method for optimizing the production process of irradiated wine based on a PLS-LSBoost gradient boosting tree. Background Art

[0002] Chinese liquor is a beverage liquor made from grain as the main raw material, with daqu, xiaoqu, bran koji, yeast, etc. as saccharifying and fermenting agents, and is made through cooking, saccharification, fermentation, and distillation. Aging is an important process in the production of Chinese liquor. During this process, many physical and chemical changes occur in the liquor body, resulting in changes in the components of Chinese liquor. The pungency, off-flavors, and irritating odors of freshly distilled Chinese liquor are reduced, the taste becomes mellow, and the quality is significantly improved. γ-ray irradiation can accelerate the aging of Chinese liquor, saving production costs and cycle time. A large number of studies have shown that Chinese liquor irradiated with γ-rays is safe for consumption, and has advantages such as simple process, low energy consumption, no environmental pollution and chemical drug residues, and can be processed with packaging.

[0003] However, there is a lack of an efficient and accurate method for optimizing the irradiation parameters of Chinese liquor. Currently, it is mainly achieved through orthogonal experiments combined with sensory evaluation. The optimization of the irradiation technology used for irradiated wine is inseparable from the study of the aging effect of wine samples. The results of sensory evaluation are affected by subjective factors, environmental conditions, etc., and there is also a certain degree of ambiguity and uncertainty. In addition, the orthogonal experiment method used to explore the optimal process conditions has certain limitations on factor levels and interactions, inevitably losing some information and unable to obtain the optimal solution in the multi-factor continuous region. Using the method of mathematical models, the optimal solution within a continuous range can be obtained, thereby calibrating the relationship between the relative content of flavor compounds in wine samples and irradiation parameters, and determining the combination of irradiation parameters required to achieve the target aging effect.

[0004] Currently, there have been many studies on the prediction methods of Chinese liquor age, mostly using techniques such as gas chromatography, high-performance liquid chromatography, and mass spectrometry combined with chemometric methods for analysis. Different combinations of Chinese liquor storage time and irradiation parameters have similar effects on the flavor quality of Chinese liquor. However, compared with the single variable of aging time, irradiation parameters will affect the results, so irradiation aging acceleration is more complex. Traditional chemometric methods such as partial least squares regression and multiple linear regression are too simple, and their fitting ability is slightly insufficient when applied to the irradiation aging acceleration system. The artificial neural network method has the ability of highly nonlinear operations, strong fault tolerance ability, and complex structure, and is also often used for process optimization, but it is a "black box" model and cannot well explain the contributions of the model and prediction features.

[0005] LSBoost is a gradient boosting tree with the mean squared error (MSE) as the loss function. Using decision trees as units, it continuously fits the error through the gradient descent method to make the model approximate the true value. Compared with the above methods, it can not only solve the problem that the number of predictor variables is much larger than the number of samples in the dataset, but also interpret the model to determine the biomarkers with greater contributions. Summary of the Invention

[0006] The object of the present invention is to overcome the deficiencies of the prior art and provide an optimization method for the production process of irradiated Baijiu based on the PLS-LSBoost gradient boosting tree. First, the present invention establishes LSBoost and PLS-LSBoost models between the volatile component content data and each irradiation parameter based on the irradiated Baijiu samples produced under different irradiation parameter combinations. The parameters of the models are optimized through the Bayesian optimization algorithm and variable selection method, and then the generalization abilities of PLS-LSBoost and LSBoost are compared to select the model type and model parameters that are applicable to the dataset and have the best prediction ability. The information of the naturally aged Baijiu samples is input into the optimal model to evaluate the best irradiation aging process parameter combination required to obtain a specific aging effect, achieving the purpose of process optimization.

[0007] An optimization method for the production process of irradiated Baijiu based on the PLS-LSBoost gradient boosting tree, comprising the following steps

[0008] 1) Obtain irradiated Baijiu samples with combinations of three parameters: irradiation dose, storage temperature, and ultrasonic time, and naturally aged Baijiu samples of different years. Use direct injection gas chromatography-mass spectrometry (GC-MS) and headspace solid-phase microextraction gas chromatography-mass spectrometry (HS-GC-MS) analysis techniques to determine the volatile flavor components of the samples. After using the ALS algorithm to complete the missing values, the volatile flavor component data is dimensionally reduced through partial least squares (PLS) to extract the scores of the irradiated Baijiu samples on the principal components.

[0009] 2) Respectively, with the irradiation dose, storage temperature, and ultrasonic time as the response variables, and the original volatile flavor component data or PLS principal component scores as the predictor variables, establish LSBoost models and PLS-LSBoost models optimized by the Bayesian algorithm.

[0010] 3) Calculate the importance scores of the predictor variables of the LSBoost model and the PLS-LSBoost model, screen the predictor variables of the models, and improve the prediction ability of the models.

[0011] 4) Calculate the importance scores of the compounds of the LSBoost model and the PLS-LSBoost model, and screen the characteristic compounds related to irradiation in the Baijiu system.

[0012] 5) Input the volatile components of naturally aged liquor into the optimal model, and output the combination of irradiation parameter values, which is the optimal irradiation aging process parameter combination required to obtain a specific aging effect, achieving the purpose of process optimization.

[0013] The irradiation liquor sample treatment parameters in step 1) are the orthogonal combination of irradiation source 60 Co with multiple irradiation doses, different storage temperatures, and ultrasonic time; use the ALS algorithm for matrix decomposition to supplement missing values; use the z-score method to standardize the data.

[0014] In step 1), use PLS to extract orthogonal principal components. The load of variables in the principal components is XL. Calculate the score XS = X0 * XL of the irradiated liquor sample on the principal components as the input variable for training the PLS-LSBoost model. Calculate the score XSp = Xp * XL of the naturally aged liquor sample on the principal components as the input variable for prediction using the PLS-LSBoost model.

[0015] In step 2), the training process of the LSBoost model is to add one learner in each round, fit the error in the previous round, and minimize the MSE:

[0016]

[0017]

[0018]

[0019] Among them, is the initial function, that is, the average value of the output data set, is a set of regression trees, v is the learning rate, and B m is the m-th regression tree, and α m ={α1, α2,...} is the parameter set of B m ; use the Bayesian optimization algorithm, with the cross-validated MSE as the objective function, update the three parameters of the learning rate, maximum number of splits, and number of learning cycles of the model, and find the optimal model parameter combination to minimize the objective function.

[0020] In step 3), the importance score of the prediction variable of the LSBoost model or the PLS-LSBoost model is

[0021]

[0022]

[0023] Among them, VIlsb j is the importance score of the j-th variable, are the compound or the main component x j in the boosting tree ensemble and the relative importance score in a single tree; M is the number of trees, J is the number of leaf nodes of tree T; v t is the splitting variable at node t, is the reduction in the squared error due to the split.

[0024] In step 3) described above, the predictive variables with small importance scores are filtered in sequence to obtain the learning curve of the change in the cross-validated MSE as the number of predictive variables decreases, and the optimal combination of predictive variables that minimizes the cross-validated MSE is determined. In step 4) described above, the compound importance score of the LSBoost model is the same as 5, and the compound importance score of the PLS-LSBoost model is

[0025]

[0026]

[0027]

[0028] where VIpls-lsb j is the importance score of the j-th variable, VIlsb b is the relative importance of the b-th principal component in LSBoost, w bj is the compound x b on the b-th principal component.

[0029] In step 4) described above, the compounds with importance scores greater than 0 in the best model and the compounds with importance scores greater than 0 and significantly correlated with the predictive variables in the non-best model are selected as the marker compounds during the irradiation process.

[0030] In step 5) described above, compare the cross-validated MSE with R 2 to determine the best model type, model parameters, and combination of predictive variables; substitute the scores of the natural-aged Baijiu samples on the principal components into the PLS-LSBoost model, and substitute the original data into the LSBoost model for process parameter prediction.

[0031] Compared with the prior art, the present invention has the following beneficial effects: The selection of conditional parameters during the irradiation process directly affects the final flavor of the irradiated wine. In the actual application process, it is necessary to conduct cumbersome experiments and sensory evaluations to determine, with low efficiency. Traditional chemometric methods such as partial least squares regression and multiple linear regression are too simple and have poor fitting ability, and are not sufficient for application in complex irradiation aging systems. Compared with traditional prediction models such as BP neural network, the PLS-LSBoost model and the LSBoost model are applicable to different data sets, can provide intuitive mathematical equations, help to deeply understand the mechanism behind the prediction results and data; can display the importance of factors and screen for signature compounds. Using PLS-LSBoost to establish a prediction model for irradiation process parameters to solve the technical problems of optimizing the irradiation wine process is beneficial to improving the production efficiency of irradiated wine, reducing technical costs, bringing economic benefits, and also conducive to promoting the standardized and regularized production of irradiated wine. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 are the prediction results of LSBoost and PLS-LSBoost for unknown samples at (A) irradiation dose, (B) ultrasonic time, and (C) storage temperature and the comparison with the true values.

[0033] Figure 2 are the normalized compound importance scores obtained by LSBoost, PLS-LSBoost, and correlation analysis. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The implementation process of the present invention will be described in detail with reference to examples. The example is to use the present invention to detect the Luzhou-flavor liquor samples provided by Sichuan *** Liquor Industry Co., Ltd. and optimize the irradiation process parameters according to the required aging degree of the irradiated wine. The instruments required for detection include a headspace sampler (7697A, Agilent), a gas chromatograph (7890B, Agilent), and a mass spectrometer (5977B, Agilent). The modeling processes of LSBoost and PLS-LSBoost are both completed in the software MATLAB R2020b.

[0035] 1. Determination of compound content.

[0036] Prepare 20 white liquor samples treated with different irradiation parameters and 5 samples aged naturally for different times. Among them, the irradiation doses have 4 gradients: 3 kGy, 4 kGy, 5 kGy, and 6 kGy; the storage temperatures used have 4 gradients: 20 °C, 25 °C, 30 °C, and 35 °C; the ultrasonic times used have 4 gradients: 0 min, 5 min, 10 min, and 15 min; the storage times of the naturally aged samples are 1, 2, 3, 4, and 5 years respectively. Determine the volatile compounds in them by GC-MS and HS-GC-MS, and conduct qualitative and quantitative analysis.

[0037] 2. Preprocess the original data. Complement the content matrix of 49 compounds in the white liquor samples using the ALS algorithm and perform standardization processing to remove the influence of dimensions. Take the compound content as the independent variable and the irradiation dose, storage temperature, and ultrasonic time as the dependent variables, and conduct PLS partial least squares analysis. For each process parameter, extract 19 principal components and 19 orthogonal principal components. Calculate the scores of the irradiated white liquor samples on the principal components XS = X0 * XL as the input variables for training the PLS-LSBoost model; calculate the scores of the naturally aged white liquor samples on the principal components XSp = Xp * XL as the input variables for predicting the naturally aged white liquor samples using the PLS-LSBoost model.

[0038] 3. Establish and optimize the LSBoost model.

[0039] 1) Selection of predictive variables. Respectively take the irradiation dose, storage temperature, and ultrasonic time as the response variables and the contents of 49 volatile compounds as the predictive variables, and establish an initial LSBoost model through Bayesian parameter optimization. Calculate the importance scores of each predictive variable according to the formula Screen the predictive variables according to the selection criteria in Table 1 to avoid overfitting and improve the generalization ability of the model. During this process, calculate the cross-validation error MSEcv and the determination function R2cv of each model through cross-validation to characterize the prediction ability of the model. According to Table 1, as the number of predictive variables decreases, MSEcv first decreases and then increases, and R2cv is the opposite. The model corresponding to the minimum value of MSEcv is the best model we want. The optimal numbers of predictive variables for the irradiation dose, storage temperature, and ultrasonic time models are 4, 4, and 5 respectively.

[0040] 2) For each group of predictive variables, use the Bayesian algorithm to establish a series of LSBoost models by searching for the best combination of learning rate, number of learning cycles, and maximum number of splits, and minimize the objective function, that is, the MSEcv of cross-validation. The optimization results are shown in Table 1.

[0041] Optimization Process of LSBoost Prediction Model for Irradiation Dose, Storage Temperature, and Ultrasonic Time

[0042]

[0043] 4. Establishment and Optimization of PLS-LSBoost Model

[0044] 1) Selection of prediction variables. Taking irradiation dose, storage temperature, and ultrasonic time as response variables respectively, and taking the scores of irradiated liquor samples on the 19 principal components extracted by PLS as prediction variables, an initial LSBoost model is established through Bayesian parameter optimization. According to the formula Calculate the importance scores of each prediction variable, and the results are as Figure 2 shown. Screen the prediction variables according to the selection criteria in Table 2 to avoid overfitting and improve the generalization ability of the model. During this process, calculate the cross-validation error MSEcv and the determination function R2cv of each model through cross-validation to characterize the prediction ability of the model. According to Table 2, as the number of prediction variables decreases, MSEcv first decreases and then increases, and R2cv is the opposite. The model corresponding to the minimum value of MSEcv is the best model we want. The optimal number of prediction variables for the irradiation dose, storage temperature, and ultrasonic time models are 2, 2, and 4 respectively.

[0045] 2) For each set of prediction variables, use the Bayesian algorithm to establish a series of LSBoost models by searching for the combination of the best learning rate, number of learning cycles, and maximum number of splits, and minimize the objective function, that is, the MSEcv of cross-validation. The optimization results are shown in Table 2.

[0046] Table 2 Optimization Process of PLS-LSBoost Prediction Model for Irradiation Dose, Storage Temperature, and Ultrasonic Time

[0047]

[0048]

[0049] 5. Model selection. For each parameter, calculate the MAE, MSE, and R2 of the LSBoost and PLS-LSBoost models on the training set and validation set to evaluate whether there is overfitting and underfitting in the model, and select a model with appropriate complexity. The results are shown in Table 3. The comparison between the predicted values and the true values of different models on the validation set is as Figure 1 shown. The model types suitable for irradiation dose, storage temperature, and ultrasonic time are PLS-LSBoost, LSBoost, and LSBoost respectively.

[0050] Table 3 Comparison of performance parameters of LSBoost and PLS-LSBoost in predicting irradiation dose, storage temperature, and ultrasonic time

[0051]

[0052] 6. Selection of characteristic compounds. Calculate the importance score of each compound according to the prediction variable importance formula of the model. Compounds with an importance score greater than 0 in the best model, and characteristic compounds with an importance score greater than 0 and significantly correlated with the prediction variables in non-best models are used to indicate the degree of irradiation aging and to understand the effects of irradiation dose, storage temperature, and ultrasonic time on the aging of Baijiu. The results are as Figure 2 shown.

[0053] 7. Model prediction. Input the principal component scores XSp obtained by PLS processing of the volatile components of naturally aged Baijiu and the volatile components Xp into the above model to obtain the predicted values of irradiation dose, storage temperature, and ultrasonic time, which are the irradiation process parameters required to achieve the corresponding aging effect. Finally, it is calculated that the irradiation doses required to achieve the effects of natural aging for 5 years, 4 years, 3 years, 2 years, and 1 year are 4.7772 kGy, 5.2353 kGy, 5.7832 kGy, 4.9593 kGy, and 5.7832 kGy respectively, the storage temperatures are 27.3127 °C, 27.3127 °C, 24.3618 °C, 28.6602 °C, and 26.6554 °C respectively, and the ultrasonic times are 4.2084 min, 10.7359 min, 11.6584 min, 7.1404 min, and 10.7945 min respectively.

[0054] The above-described embodiments merely represent several implementation manners of the present invention. Their descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. An optimization method for the production process of irradiated liquor based on PLS-LSBoost gradient boosting tree, characterized in that, It includes the following steps: 1) Obtain irradiated Baijiu samples with combinations of three parameters: irradiation dose, storage temperature, and ultrasonic time, and natural aging Baijiu samples of different years. Use direct injection gas chromatography-mass spectrometry and static headspace gas chromatography-mass spectrometry analysis techniques to determine the volatile flavor components of the samples. After using the ALS algorithm to complete the missing values, perform dimensionality reduction on the volatile flavor component data through partial least squares, and extract the scores of the irradiated Baijiu samples on the principal components; 2) Respectively take the irradiation dose, storage temperature, and ultrasonic time as response variables, and the original data of volatile flavor components or PLS principal component scores as predictive variables, and establish LSBoost models and PLS-LSBoost models optimized by the Bayesian algorithm; 3) Calculate the importance scores of the predictive variables of the LSBoost model and the PLS-LSBoost model, screen the predictive variables of the models, and improve the model prediction ability; 4) Calculate the importance scores of the compounds of the LSBoost model and the PLS-LSBoost model, and screen the characteristic compounds related to irradiation in the Baijiu system; 5) Input the volatile components of the naturally aged Baijiu into the best model, and output the combination of irradiation parameter values, which is the optimal irradiation aging process parameter combination required to obtain a specific aging effect, achieving the purpose of process optimization; The establishment and optimization of the PLS-LSBoost model include: a) Selection of predictors: Taking the irradiation dose, storage temperature, and ultrasonic time as response variables respectively, and the scores of the irradiated Baijiu samples on the principal components extracted by 19 PLSs as predictors, an initial LSBoost model was established through Bayesian parameter optimization; according to the formula Calculate the importance scores of each predictor, where, are the relative importance scores of the compound or principal component x j in the boosting tree ensemble and in a single tree, M is the number of trees, J is the number of leaf nodes of tree T; v t is the splitting variable at node t, is the reduction in the squared error caused by the split; screen the predictors to avoid overfitting and improve the generalization ability of the model; calculate the cross-validation error MSEcv and the determination function R2cv of each model through cross-validation to characterize the prediction ability of the model; as the number of predictors decreases, MSEcv first decreases and then increases, and R2cv is the opposite. The model corresponding to the minimum value of MSEcv is the best model; the optimal number of predictors for the irradiation dose, storage temperature, and ultrasonic time models are 2, 2, and 4 respectively; b) For each set of predictive variables, use the Bayesian algorithm to establish a series of LSBoost models by searching for the combination of the best learning rate, number of learning cycles, and maximum number of splits, and minimize the objective function, that is, the cross-validated MSEcv.

2. The method according to claim 1, wherein The processing parameters of the irradiated Baijiu sample in step 1) are the orthogonal combination of the irradiation source 60 60 Co with multiple irradiation doses, different storage temperatures, and ultrasonic time; matrix decomposition is performed using the ALS algorithm to supplement missing values; the z-score method is used to standardize the data.

3. The method according to claim 1, wherein In step 1), use PLS to extract orthogonal principal components. The load of the variables in the principal components is XL. Calculate the score XS of the irradiated Baijiu sample on the principal component as XS = X0 * XL, which is used as the input variable for training the PLS-LSBoost model. Calculate the score XSp of the naturally aged Baijiu sample on the principal component as XSp = Xp * XL, which is used as the input variable when using the PLS-LSBoost model for prediction.

4. The method according to claim 1, characterized in that In step 2), the training process of the LSBoost model is that in each round, add a learner to fit the error in the previous round, and update the model parameter set by minimizing the MSE: Among them, is the initial function, that is, the average value of the output data set, is a set of regression trees, v is the learning rate, B m is the m-th regression tree, α m ={α1,α2,…} is the parameter set of B m ; The Bayesian optimization algorithm is adopted, with the cross-validated MSE as the objective function, to update the three parameters of the learning rate, the maximum number of splits, and the number of learning cycles of the model, and to find the best model parameter combination to minimize the objective function.

5. The method according to claim 1, wherein In step 3), the importance scores of the predictive variables of the LSBoost model and the PLS-LSBoost model are Among them, VIlsb j is the importance score of the j-th variable, respectively for the compound or principal component x j in the boosting tree ensemble and the relative importance scores in a single tree; M is the number of trees, J is the number of leaf nodes of tree T; v t is the splitting variable at node t, is the reduction in the squared error due to the split.

6. The method according to claim 1, wherein In step 3), sequentially filter the predictive variables with small importance scores to obtain the learning curve of the change in the cross-validated MSE as the number of predictive variables decreases, and determine the optimal combination of predictive variables that minimizes the cross-validated MSE.

7. The method according to claim 1, characterized in that In step 4), the importance scores of the compounds of the PLS-LSBoost model are Among them, VIpls-lsb j is the importance score of the j-th variable in the PLS-LSBoost model, and VIlsb b is the relative importance score of the b-th principal component in the LSBoost model, and w bj is the weight of compound x j on the b-th principal component.

8. The method according to claim 1, wherein In step 4), select the compounds with importance scores greater than 0 in the best model, and the compounds with importance scores greater than 0 and significantly correlated with the predictive variables in the non-best model as the characteristic compounds during the irradiation process.

9. The method according to claim 1, wherein In step 5), compare the cross-validated MSE with R 2 , and determine the optimal model type, model parameters, and combination of predictor variables; substitute the scores of the naturally aged Baijiu samples on the principal components into the PLS-LSBoost model, substitute the original data into the LSBoost model, and perform process parameter prediction.

Images