Method for training and applying a baijiu quality prediction model and related device

Abstract

The application discloses a liquor quality prediction model training and application method and related devices, relates to the liquor evaluation technical field, and comprises the following steps: acquiring a data set, wherein the data set comprises the content of each flavor compound and the score of each sensory characteristic descriptor of each liquor sample; for each sensory characteristic descriptor, all the flavor compounds are screened to obtain screened flavor compounds; a plurality of initial prediction models are trained to obtain a plurality of trained models and the model performance of each trained model in the plurality of trained models; finally, the trained model with the best model performance is selected as a sensory characteristic prediction model, and all the sensory characteristic prediction models form a quality prediction model, which is used for predicting the score of each sensory characteristic descriptor of the liquor and determining the quality of the liquor. The application can efficiently and accurately complete the quality prediction of the liquor.

Description

Technical Field

[0001] This application relates to the field of liquor tasting technology, and in particular to a method and apparatus for training and applying a liquor quality prediction model. Background Technology

[0002] As a traditional distilled spirit, baijiu is mainly composed of water and ethanol. However, the flavor compounds (which constitute a very small percentage) are the core source of baijiu's aroma. These flavor compounds not only determine the sensory characteristics of baijiu but are also the core elements determining its quality. Due to the differences in raw materials, regional environments, microbial communities, and brewing techniques, baijiu possesses complex sensory characteristics, resulting in the unique styles of different aroma types. At the same time, the complex sensory characteristics of baijiu also pose a great challenge to accurate perception and precise control, thus seriously affecting the stability and even improvement of baijiu quality.

[0003] The recipients and the subjects of perception in baijiu (Chinese liquor) evaluation are both human beings. Human cognitive behavior is easily influenced by multiple factors, including external environmental factors and internal emotional states, leading to significant fluctuations in baijiu quality evaluation results. A quantitative sensory characteristic description system, constructed based on multivariate statistical analysis techniques and incorporating multidimensional sensory feature descriptive vocabulary, can effectively mitigate the uncertainty caused by individual differences among evaluators, improving the accuracy and objectivity of baijiu quality evaluation. However, for baijiu producers, while a scientific quantitative sensory characteristic description system optimizes evaluation accuracy, it suffers from low evaluation efficiency. Therefore, how to simultaneously improve evaluation efficiency while ensuring accuracy has become a key technical challenge that the baijiu industry urgently needs to address. Summary of the Invention

[0004] The purpose of this application is to provide a method and related apparatus for training and applying a baijiu quality prediction model, which can efficiently and accurately predict the quality of baijiu.

[0005] To achieve the above objectives, this application provides the following solution.

[0006] Firstly, this application provides a method for training a baijiu (Chinese liquor) quality prediction model, the method comprising: Obtain a dataset; the dataset includes feature data and label data for each of the multiple baijiu samples. The feature data includes the content of each flavor compound among multiple flavor compounds. The label data includes the score of each sensory descriptor among multiple sensory descriptors. The sensory descriptors include aroma sensory descriptors and taste sensory descriptors. For each of the sensory characteristic descriptors, based on the content of each flavor compound in each baijiu sample and the score of the sensory characteristic descriptor, all the flavor compounds are screened to obtain the screened flavor compounds corresponding to the sensory characteristic descriptor. For each of the sensory feature descriptors, the content of each screened flavor compound corresponding to the sensory feature descriptor and the score of the sensory feature descriptor of each baijiu sample are used as inputs to train multiple initial prediction models, resulting in multiple trained models and the model performance of each trained model; the initial prediction model is a model that can perform the prediction function. For each sensory feature descriptor, the trained model with the best performance is selected as the sensory feature prediction model corresponding to the sensory feature descriptor; all sensory feature prediction models constitute a quality prediction model, which is used to predict the score of each sensory feature descriptor of baijiu and determine the quality of baijiu.

[0007] Optionally, based on the content of each flavor compound in each baijiu sample and the score of the sensory characteristic descriptor, all the flavor compounds are screened to obtain the screened flavor compounds corresponding to the sensory characteristic descriptor, specifically including: Using the content of each flavor compound and the score of the sensory feature descriptor for each baijiu sample as input, the correlation coefficient between each flavor compound and the sensory feature descriptor is calculated using a correlation analysis algorithm, thus obtaining the correlation coefficient corresponding to each flavor compound. Using the content of each flavor compound and the score of the sensory feature descriptor for each baijiu sample as input, the VIP value between each flavor compound and the sensory feature descriptor is calculated using the variable projection importance algorithm, thus obtaining the VIP value corresponding to each flavor compound. For each flavor compound, if the absolute value of the correlation coefficient corresponding to the flavor compound is greater than a first preset value and the VIP value corresponding to the flavor compound is greater than a second preset value, then the flavor compound is selected as a screened flavor compound to screen all the flavor compounds and obtain the screened flavor compounds corresponding to the sensory feature descriptor.

[0008] Optionally, the initial prediction model includes: a random forest model, a support vector machine model, and a partial least squares model; The model performance is characterized by performance evaluation metrics, including the coefficient of determination and root mean square error.

[0009] Optionally, using the content of each screened flavor compound corresponding to the sensory feature descriptor of each liquor sample and the score of the sensory feature descriptor as input, multiple initial prediction models are trained to obtain multiple trained models and the model performance of each trained model, specifically including: All the liquor samples were divided into training liquor samples and test liquor samples. For each of the initial prediction models, the content of each screened flavor compound corresponding to the sensory feature descriptor and the score of the sensory feature descriptor of each training baijiu sample are used as inputs to train the initial prediction model and obtain the trained model corresponding to the initial prediction model. For each of the initial prediction models, the content of each screened flavor compound corresponding to the sensory feature descriptor of each test liquor sample is used as input. The predicted score of the sensory feature descriptor of each test liquor sample is predicted by the trained model corresponding to the initial prediction model. Based on the score of the sensory feature descriptor and the predicted score of each test liquor sample, the model performance of the trained model corresponding to the initial prediction model is determined.

[0010] Optionally, if the sensory feature prediction model corresponding to the sensory feature descriptor is a random forest model, after obtaining the sensory feature prediction model corresponding to the sensory feature descriptor, the training method for the liquor quality prediction model further includes: For each of the selected flavor compounds corresponding to the sensory feature descriptor, the average Gini impurity reduction of the selected flavor compound is calculated based on the sensory feature prediction model corresponding to the sensory feature descriptor. If the average Gini impurity reduction of the selected flavor compound is greater than a third preset value, then the selected flavor compound is regarded as the key flavor compound corresponding to the sensory feature descriptor.

[0011] Optionally, the aroma sensory characteristics descriptors include soy sauce aroma, grain aroma, yeast aroma, fruit aroma, sweet aroma, sour aroma, floral aroma, and baking aroma, and the taste sensory characteristics descriptors include mellow sweetness, mellowness, and long-lasting flavor.

[0012] Secondly, this application provides a method for applying a liquor quality prediction model, the method comprising: For each sensory feature descriptor, the content of each screened flavor compound corresponding to the sensory feature descriptor in the liquor to be predicted is used as input, and the predicted score corresponding to the sensory feature descriptor is determined by the sensory feature prediction model corresponding to the sensory feature descriptor in the quality prediction model; the quality prediction model is a model trained using the above-mentioned liquor quality prediction model training method. The quality of the liquor to be predicted is determined based on the predicted scores corresponding to all the sensory feature descriptors of the liquor to be predicted.

[0013] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the above-described method for training a liquor quality prediction model or the above-described method for applying a liquor quality prediction model.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for training or applying the liquor quality prediction model.

[0015] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for training or applying the liquor quality prediction model.

[0016] According to the specific embodiments provided in this application, this application has the following technical effects.

[0017] This application provides a method and related apparatus for training and applying a baijiu (Chinese liquor) quality prediction model. The method involves acquiring a dataset, which includes the content of each flavor compound in each of multiple baijiu samples and the score of each sensory descriptor. For each sensory descriptor, all flavor compounds are first screened to obtain the screened flavor compounds. Then, multiple initial prediction models are trained to obtain multiple trained models and the model performance of each trained model. Finally, the trained model with the best performance is selected as the sensory feature prediction model. All sensory feature prediction models constitute a quality prediction model, which is used to predict the score of each sensory descriptor of baijiu to determine the quality of the baijiu. This application introduces sensory feature descriptors, which can improve the accuracy of tasting. By training a quality prediction model and using it to determine the quality of baijiu, no human intervention is required, thus improving the efficiency of tasting. At the same time, the best-performing trained model is selected as the sensory feature prediction model during training, which is then used to form the quality prediction model. This ensures the performance of the quality prediction model and further improves the accuracy of tasting. Thus, while improving the accuracy of tasting, the efficiency of tasting is also improved, enabling efficient and accurate prediction of baijiu quality. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an application environment diagram for a method of training and applying a liquor quality prediction model, as provided in Embodiment 1 of this application.

[0020] Figure 2 This is a flowchart illustrating a method for training a liquor quality prediction model, as provided in Embodiment 1 of this application.

[0021] Figure 3 This is a schematic diagram of the technical route for a liquor quality prediction model training method provided in Embodiment 1 of this application.

[0022] Figure 4 This is a schematic diagram showing the comparison of model performance of different trained models when the performance evaluation index is the coefficient of determination, as provided in Embodiment 1 of this application.

[0023] Figure 5 This is a flowchart illustrating the application method of a liquor quality prediction model provided in Embodiment 2 of this application.

[0024] Figure 6 This is a schematic diagram of the structure of a computer device provided in Embodiment 3 of this application. Detailed Implementation

[0025] 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 some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Example 1 The liquor quality prediction model training method provided in this application embodiment can be applied to, for example... Figure 1 The application environment shown depicts a scenario where the terminal communicates with the server via a network. The data storage system stores the data the server needs to process. This data storage system can be configured independently, integrated into the server, or located in the cloud or on another server. The terminal can send a training request to the server. Upon receiving the request, the server retrieves the dataset, which includes feature data and label data for each of the multiple baijiu samples. The feature data includes the content of each flavor compound, and the label data includes the score of each sensory descriptor. For each sensory descriptor, based on the content of each flavor compound and the score of the sensory descriptor for each baijiu sample, all flavor compounds are screened to obtain the screened flavor compounds corresponding to the sensory descriptor. For each sensory descriptor, using the content of each screened flavor compound and the score of the sensory descriptor for each baijiu sample as input, multiple initial prediction models are trained to obtain multiple trained models and the model performance of each trained model. For each sensory descriptor, the trained model with the best performance is selected as the sensory feature prediction model corresponding to the sensory descriptor. All sensory feature prediction models constitute the quality prediction model, which is used to predict the score of each sensory descriptor of baijiu to determine the quality of the baijiu. The server can feed back the training result of the quality prediction model obtained in response to the training request to the terminal.

[0027] In addition, in some embodiments, the training method for the liquor quality prediction model can also be implemented by the server or the terminal alone. For example, the terminal can directly process the training request to be processed, or the server can obtain the training request to be processed from the data storage system and process it.

[0028] The terminals can be, but are not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices, while portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. Servers can be implemented using independent servers, server clusters composed of multiple servers, or cloud servers.

[0029] In one exemplary embodiment, such as Figure 2 As shown, a method for training a liquor quality prediction model is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 The following steps, S1-S4, are used as an example to illustrate the process of using a server in the example.

[0030] Step S1: Obtain the dataset; the dataset includes feature data and label data for each of the multiple baijiu samples. The feature data includes the content of each flavor compound among multiple flavor compounds. The label data includes the score of each sensory feature descriptor among multiple sensory feature descriptors. The sensory feature descriptors include aroma sensory feature descriptors and taste sensory feature descriptors.

[0031] Step S2: For each of the sensory feature descriptors, based on the content of each flavor compound in each baijiu sample and the score of the sensory feature descriptor, all the flavor compounds are screened to obtain the screened flavor compounds corresponding to the sensory feature descriptor.

[0032] Step S3: For each of the sensory feature descriptors, the content of each screened flavor compound corresponding to the sensory feature descriptor and the score of the sensory feature descriptor of each liquor sample are used as inputs to train multiple initial prediction models, thereby obtaining multiple trained models and the model performance of each trained model; the initial prediction model is a model that can perform the prediction function.

[0033] Step S4: For each sensory feature descriptor, select the trained model with the best model performance as the sensory feature prediction model corresponding to the sensory feature descriptor; all sensory feature prediction models constitute a quality prediction model, which is used to predict the score of each sensory feature descriptor of the liquor and determine the quality of the liquor.

[0034] By implementing steps S1 to S4 above, this embodiment can train a quality prediction model composed of sensory feature prediction models corresponding to each sensory feature descriptor, thereby enabling efficient and accurate prediction of the sensory features (using the sensory feature prediction model) and quality (using the quality prediction model) of baijiu.

[0035] The flavor (i.e., sensory characteristics) of baijiu is its core characteristic that distinguishes it from other distilled spirits. It directly determines the quality grade, style characteristics, and consumer acceptance of baijiu. It is a key carrier of the value and cultural connotation of baijiu products. Studying the flavor of baijiu is to clarify the chemical nature of baijiu and solve the problem of poor batch stability of flavor and quality of baijiu products caused by the complexity of solid-state brewing process. This embodiment first uses gas chromatography to determine major and trace components, and gas chromatography-mass spectrometry to determine trace components, accurately quantifying a total of 208 flavor compounds. Then, a quantitative description system of the sensory characteristics of baijiu is constructed, identifying eight aroma sensory characteristic descriptors: soy sauce aroma, grain aroma, yeast aroma, fruit aroma, sweet aroma, sour aroma, floral aroma, and roasted aroma; and three taste sensory characteristic descriptors: mellow sweetness, mellowness, and long-lasting flavor. Professionals taste and score the baijiu, determining the scores for 11 sensory characteristic descriptors. Based on this, a dataset is constructed. Then, for each sensory characteristic descriptor, the correlation between the content of the 208 flavor compounds and the score of the sensory characteristic descriptor is analyzed, specifically using correlation analysis algorithms and variable importance in the... The Projection (VIP) algorithm screens differential flavor compounds that are significantly correlated with sensory feature descriptors, resulting in screened flavor compounds. Multiple initial prediction models are constructed, using the content of the screened flavor compounds as input variables and the scores of the sensory feature descriptors as output variables. Each initial prediction model is trained, yielding the trained model and its performance. The best-performing trained model is selected as the sensory feature prediction model corresponding to the sensory feature descriptor. Finally, the sensory feature prediction models corresponding to each sensory feature descriptor are combined to form a quality prediction model. This quality prediction model enables a quantitative mapping from chemical components to sensory experience, breaking through the limitations of traditional sensory evaluation and providing data support for the standardization of baijiu flavor.

[0036] The following, combined with Figure 3 This paper provides a detailed introduction to a training method for a liquor quality prediction model used in this embodiment.

[0037] (I) Dataset Construction This embodiment selects baijiu (Chinese liquor) as the research object, and selects 42 baijiu samples with sensory differences (other numbers can be used according to user needs) for the construction of a quantitative description system of baijiu sensory characteristics, dataset construction, screening of important related components of baijiu sensory characteristics (i.e., determining the flavor compounds corresponding to each sensory characteristic descriptor), and construction of baijiu sensory characteristics and quality prediction models (i.e., determining the sensory characteristic prediction model corresponding to each sensory characteristic descriptor and the quality prediction model composed of all sensory characteristic prediction models). Among them, in the construction of baijiu sensory characteristics and quality prediction models, 36 baijiu samples can be selected for model training and 6 baijiu samples can be selected for model validation.

[0038] In the process of constructing a quantitative description system for the sensory characteristics of baijiu and building a dataset, gas chromatography was used to determine the major and trace components of baijiu samples, and gas chromatography-mass spectrometry was used to determine the trace components of baijiu samples. The content of each flavor compound in each of the multiple baijiu samples was obtained. The sensory quantitative description analysis of baijiu was carried out to construct a quantitative description system for the sensory characteristics of baijiu. Professionals scored the baijiu samples to obtain the score of each sensory descriptive term in the multiple sensory descriptive terms of each baijiu sample.

[0039] (1) Component determination To complete the component determination, the instruments and equipment required in this embodiment include: 1) a TriPlus RSH SMART autosampler for automated sample pretreatment and injection; 2) a GC-MS / MS system (Thermo Scientific TRACE1600-TSQ 9610), where the TRACE 1600 is a gas chromatograph (GC) and the TSQ 9610 is a triple quadrupole tandem mass spectrometer (MS / MS). The two are coupled to form a highly sensitive and selective GC-MS / MS system. Of course, other models of autosamplers and GC-MS / MS systems can also be used according to user requirements.

[0040] Based on the above-mentioned instruments and equipment, the conditions used for determining the major and trace components of baijiu samples by gas chromatography and for determining the trace components of baijiu samples by gas chromatography-mass spectrometry are as follows: 1) Gas chromatography conditions: The column was a WAX (60 m × 0.25 mm × 0.25 μm); the temperature program was as follows: initial temperature 35 ℃, hold for 1 min, increase to 230 ℃ at a rate of 3.5 ℃ / min, hold for 10 min; detector temperature was 250 ℃; injection port temperature was 250 ℃; carrier gas was high-purity nitrogen, flow rate was 1.0 mL / min; injection volume was 1.0 μL; split mode was used, split ratio was 20:1. 2) Gas chromatography-mass spectrometry conditions: The chromatographic column was WAXMS (60 m × 0.25 mm × 0.25 μm); the injection port temperature was 250℃; splitless injection; the injection volume was 1 μL; the carrier gas was He (helium); the flow rate was 1 mL / min; the initial temperature of the column oven was 35℃, held for 1 min, increased to 150℃ at a rate of 2℃ / min, held for 10 min, increased to 230℃ at a rate of 5℃ / min, held for 15 min, and the total programmed temperature time was 99.5 min; the transfer line temperature was 250℃, the ion source temperature was 260℃, the ion source was EI (Electron Ionization), the collision gas was Ar (Ar), the ionization voltage was 70 eV, the electron energy was 70 eV, and the scan mode was SRM (Selected Reaction Monitoring).

[0041] Based on the aforementioned instruments, equipment, and conditions, for each baijiu sample, the content (specifically, concentration, which can be mass concentration) of flavor compounds in the baijiu sample was accurately quantified using gas chromatography and gas chromatography-mass spectrometry. The specific steps included: 1) Gas chromatography pretreatment of baijiu samples: Accurately pipette 900 μL of baijiu sample filtered through a 0.22 μm organic phase microporous membrane and place it in a 1.5 mL liquid chromatography vial. Then add 100 μL of 200 mg / L tert-amyl alcohol, n-amyl acetate, and 2-ethylbutyric acid as internal standards to the liquid chromatography vial. Quickly tighten the cap of the liquid chromatography vial, vortex for 2 min, and wait for the instrument to inject the sample for analysis.

[0042] 2) Pretreatment of Baijiu sample by gas chromatography-mass spectrometry: Accurately pipette 900 μL of Baijiu sample into a 1.5 mL liquid chromatography vial, then add 100 μL of 10 mg / L 2-propylpyrazine and 3,4-dimethylphenol as internal standards to the vial, then quickly tighten the vial cap, vortex for 2 min, and wait for the instrument to inject the sample for analysis.

[0043] 3) Instrumental detection: The above-mentioned liquor samples were detected using gas chromatography (GC) and gas chromatography-tandem mass spectrometry (GC-MS / MS system) to obtain chromatograms.

[0044] 4) Qualitative and quantitative analysis: The mass spectrometry information of each flavor compound corresponding to the chromatographic peaks separated on the chromatogram is retrieved and analyzed using the NIST 2020 spectral library in the computer to determine the type of each flavor compound. The peak area of ​​each flavor compound is calculated using GC-MS chromatography workstation software. The calculated peak area of ​​each flavor compound is then substituted into the standard curve of each flavor compound to calculate the accurate concentration of each flavor compound and obtain the content of each flavor compound.

[0045] In this embodiment, 208 flavor compounds can be quantified, including alcohols, aldehydes, esters, phenols, pyrazines, ketones and other substances, as shown in Table 1 below.

[0046] Table 1 208 flavor compounds

[0047] Before accurate quantification, standard curves for 208 flavor compounds were first plotted. The methods for plotting standard curves were divided into those for gas chromatography-mass spectrometry and those for gas chromatography. The method for plotting standard curves for gas chromatography-mass spectrometry was used to plot standard curves for trace components determined by gas chromatography, while the method for plotting standard curves for gas chromatography was used to plot standard curves for major and trace components determined by gas chromatography.

[0048] The method for constructing a standard curve for gas chromatography-mass spectrometry includes the following steps: 1) Preparation of single standard solutions (single standard stock solutions) for various flavor compounds: Accurately weigh 0.1 g (accurate to 0.0001 g) of various flavor compound standards into 10 mL volumetric flasks, dissolve them in anhydrous ethanol and dilute to 10 mL to prepare single standard solutions of various flavor compounds with a concentration of 10000 mg / L. Store at 0-4 ℃, and bring to room temperature before use. Shake well before use.

[0049] 2) Preparation of mixed internal standard solution for gas chromatography-mass spectrometry: Accurately weigh 100 mg of 2-propylpyrazine and 3,4-dimethylphenol, dissolve them in anhydrous ethanol and bring the volume to 100 mL to obtain a 1000 mg / L mixed internal standard solution. Store at 0–4 °C and bring to room temperature before use. Shake well before use.

[0050] 3) Preparation of gas chromatography-mass spectrometry (GC-MS) mixed standard solutions: Take 0.1 mL of each of the single standard solutions (10000 mg / L) of various flavor compounds determined by GC-MS in Table 1 above into 50 mL volumetric flasks, and make up to 50 mL with anhydrous ethanol. Then, serially dilute to prepare a series of mixed standard solutions from 0 mg / L to 5 mg / L. At the same time, add a mixed internal standard solution to obtain mixed standard solutions of different concentrations. The concentration gradient can be designed by the user.

[0051] 4) Gas chromatography-mass spectrometry was used to determine the standard curves of various flavor compounds by plotting the peak area ratio of each flavor compound to the internal standard as the ordinate and the mass concentration ratio of each flavor compound to the internal standard as the abscissa.

[0052] The method for constructing a standard curve for gas chromatography specifically includes the following steps: 1) Preparation of single standard solutions (single standard stock solutions) for various flavor compounds: Accurately weigh 0.1 g (accurate to 0.0001 g) of various flavor compound standards into 10 mL volumetric flasks, dissolve them in anhydrous ethanol and dilute to 10 mL to prepare single standard solutions of various flavor compounds with a concentration of 10000 mg / L. Store at 0-4 ℃, and bring to room temperature before use. Shake well before use.

[0053] 2) Preparation of gas chromatography mixed internal standard solution: Accurately weigh 100 mg of tert-amyl alcohol, n-amyl acetate and 2-ethylbutyric acid respectively, dissolve them in anhydrous ethanol and make up to 100 mL to obtain a 1000 mg / L mixed internal standard solution. Store at 0-4 ℃. Before use, bring the solution to room temperature and shake well.

[0054] 3) Preparation of gas chromatography mixed standard solutions: Take 0.1 mL of each of the single standard solutions (10000 mg / L) of various flavor compounds determined by gas chromatography in Table 1 above into 50 mL volumetric flasks, and make up to 50 mL with anhydrous ethanol. Then, serially dilute to prepare a series of mixed standard solutions ranging from 100 mg / L to 2000 mg / L. At the same time, add a mixed internal standard solution to obtain mixed standard solutions of different concentrations. The concentration gradient can be designed by the user.

[0055] 4) Gas chromatography was used to determine the flavor compounds according to the experimental method. The peak area ratio of each flavor compound to the internal standard was used as the ordinate, and the mass concentration ratio of each flavor compound to the internal standard was used as the abscissa to plot the standard curves of each flavor compound.

[0056] (2) Screening and quantification of sensory feature descriptive words This embodiment constructs a quantitative description system for the sensory characteristics of baijiu, obtaining typical sensory characteristic descriptive terms for baijiu, as shown in Table 2 and Table 3 below. Table 2 contains sensory characteristic descriptive terms for the aroma of baijiu, and Table 3 contains sensory characteristic descriptive terms for the taste of baijiu. Thus, the construction of the quantitative description system for sensory characteristics is completed.

[0057] Table 2. Descriptive terms for the sensory characteristics of Baijiu aroma.

[0058] Table 3. Descriptive terms for the sensory characteristics of Baijiu (Chinese liquor)

[0059] After determining the sensory characteristic descriptors, the sensory characteristic descriptors are further quantified. Specifically, professional sensory evaluators (i.e., professionals) are first selected: 20 professionally certified wine tasters (or other numbers depending on user needs) are chosen, an evaluation team leader is appointed, the wine tasters' evaluation motivations are identified, and their ability to repeat evaluations is assessed. Then, each baijiu sample is evaluated: a blind tasting method is used, with the above 42 baijiu samples presented randomly in multiple rounds, using standard baijiu tasting glasses, and approximately 20 ml of wine is poured into each glass. In mL, tasters evaluated baijiu samples using a sensory quantitative descriptive analysis method, employing a 5-point scale (0 = no sensation, 1 = weak, 2 = slightly weak, 3 = average, 4 = slightly strong, 5 = strong). Tasters scored each sensory characteristic descriptor of the baijiu sample. For each sample, multiple tasters provided scores for each sensory characteristic descriptor. Finally, the geometric mean of these scores was used to quantify the sensory characteristic descriptors. However, due to individual differences among tasters, directly quantifying sensory characteristic descriptors using a 5-point scale introduces uncertainty. Therefore, [further details are needed]. The method quantifies sensory descriptive words from the perspectives of frequency and sensory intensity. Specifically, for each sensory descriptive word of each baijiu sample, all tasters' scores for that sensory descriptive word are statistically analyzed. Outliers are first removed (this can be done manually), and then the average value is calculated. Specifically, the geometric mean of all tasters' scores for that sensory descriptive word is calculated to obtain the score of that sensory descriptive word. Thus, the geometric mean of the sensory descriptive words is used to quantify the sensory descriptive words, obtaining the score of each sensory descriptive word for each baijiu sample. The scores of sensory descriptive words such as aroma and taste can be directly used for the overall sensory quantitative evaluation of the baijiu.

[0060] In this embodiment, a dataset is obtained, which includes feature data and label data for each of the multiple baijiu samples. The feature data includes the content of each flavor compound among the multiple flavor compounds, and the label data includes the score of each sensory feature descriptor among the multiple sensory feature descriptors. The sensory feature descriptors include aroma sensory feature descriptors and taste sensory feature descriptors.

[0061] Among them, the aroma sensory characteristics descriptions include soy sauce aroma, grain aroma, yeast aroma, fruit aroma, sweet aroma, sour aroma, floral aroma and baking aroma, and the taste sensory characteristics descriptions include mellow sweetness, mellowness and long-lasting taste.

[0062] (II) Screening of flavor compounds When screening important components related to the sensory characteristics of baijiu (Chinese liquor), using accurate quantitative data (i.e., datasets), the content of flavor compounds is used as the input variable, and the scores of sensory feature descriptors are used as the output variable. The correlation between the content of flavor compounds and the scores of sensory feature descriptors is analyzed. For each sensory feature descriptor, the correlation coefficient and VIP value between each flavor compound and that sensory feature descriptor are calculated based on correlation analysis and variable projection importance algorithms. Further screening of all flavor compounds is then performed based on the correlation coefficient and VIP value to obtain the screened flavor compounds corresponding to that sensory feature descriptor. The correlation coefficient can be the Pearson correlation coefficient. When screening all flavor compounds based on the correlation coefficient and VIP value, the screening conditions can be |r|>0.6 and VIP value>1, where r is the correlation coefficient, and 0.6 and 1 can be other values ​​depending on user needs. By screening flavor compounds with an absolute correlation coefficient greater than 0.6 and a VIP value greater than 1, the screened flavor compounds significantly correlated with each sensory feature descriptor can be identified, effectively reducing model complexity and avoiding overfitting.

[0063] In this embodiment, for each sensory characteristic descriptor, based on the content of each flavor compound in each baijiu sample and the score of the sensory characteristic descriptor, all flavor compounds are screened to obtain the screened flavor compounds corresponding to the sensory characteristic descriptor.

[0064] The process involves screening all flavor compounds based on the content of each flavor compound and the score of each sensory characteristic descriptor for each baijiu sample. This results in the selected flavor compounds corresponding to the sensory characteristic descriptors. The specific steps include: (1) Using the content of each flavor compound and the score of each sensory characteristic descriptor of each baijiu sample as input, the correlation coefficient between each flavor compound and the sensory characteristic descriptor is calculated by using the correlation analysis algorithm, and the correlation coefficient corresponding to each flavor compound is obtained.

[0065] The correlation analysis algorithm can be any existing correlation analysis algorithm, which is a mature technology and will not be elaborated here.

[0066] (2) Using the content of each flavor compound and the score of each sensory feature descriptor of each baijiu sample as input, the VIP value between each flavor compound and sensory feature descriptor is calculated using the variable projection importance algorithm, and the VIP value corresponding to each flavor compound is obtained.

[0067] The variable projection importance algorithm is a mature existing technology and will not be elaborated here.

[0068] (3) For each flavor compound, if the absolute value of the correlation coefficient corresponding to the flavor compound is greater than the first preset value (e.g., 0.6) and the VIP value corresponding to the flavor compound is greater than the second preset value (e.g., 1), then the flavor compound is used as the screened flavor compound to screen all flavor compounds and obtain the screened flavor compounds corresponding to the sensory feature descriptors.

[0069] (III) Model Construction When constructing a sensory characteristic and quality prediction model for baijiu (Chinese liquor), machine learning and statistics are combined to establish the quality prediction model. For each sensory characteristic descriptor, multiple initial prediction models are established. These initial prediction models can employ machine learning models, such as Random Forest (RF) and Support Vector Machine (SVM), or statistical models, such as Partial Least Squares (PLS). Specifically, three initial prediction models can be constructed: a Random Forest model, a Support Vector Machine model, and a Partial Least Squares model. The coefficient of determination (R²) can be used. 2 Using the root mean square error (RMSE) and the content of flavor compounds corresponding to sensory feature descriptors as input variables and the scores of sensory feature descriptors as output variables, multiple initial prediction models were trained to obtain the trained model and its performance for each initial prediction model. The trained model with the best performance was selected as the sensory feature prediction model corresponding to that sensory feature descriptor. All sensory feature prediction models corresponding to sensory feature descriptors were combined to form a quality prediction model, successfully constructing a quantitative correlation model between chemical components and sensory experience. Based on this quality prediction model, the sensory characteristics and quality of baijiu were predicted, aiming to explore the relationship between compounds and senses. This effectively broke through the technical bottleneck of traditional sensory evaluation relying on subjective judgment and having difficulty in reproducing results, providing a data-driven scientific path for the baijiu industry to achieve standardized flavor control, precise product upgrading, and intelligent production.

[0070] In this embodiment, for each sensory feature descriptor, the content of each screened flavor compound corresponding to the sensory feature descriptor and the score of the sensory feature descriptor for each baijiu sample are used as input to train multiple initial prediction models, resulting in multiple trained models and the model performance of each trained model. The initial prediction model is a model capable of performing the prediction function. The trained model with the best performance is selected as the sensory feature prediction model corresponding to the sensory feature descriptor. All sensory feature prediction models constitute the quality prediction model. The quality prediction model is used to predict the score of each sensory feature descriptor of baijiu to determine the quality of baijiu. It should be noted that the quality of baijiu includes the score of each sensory feature descriptor and can further rely on user experience to determine the quality grade of baijiu based on the score of each sensory feature descriptor.

[0071] The initial prediction models include: random forest model, support vector machine model and partial least squares model. The performance of the models is characterized by performance evaluation metrics, including coefficient of determination and root mean square error.

[0072] The process involves using the content of each screened flavor compound corresponding to a sensory feature descriptor and the score of that descriptor as input for each baijiu sample. Multiple initial prediction models are then trained to obtain multiple trained models and the performance of each trained model. The specific steps include: (1) Divide all liquor samples into training liquor samples and test liquor samples.

[0073] All baijiu samples can be divided into training samples and test samples in a 7:3 ratio. The training samples form the training set for model training, and the test samples form the test set for model validation to obtain model performance. Of course, other division ratios can also be used.

[0074] (2) For each initial prediction model, the content of each selected flavor compound and the score of each sensory feature descriptor corresponding to each training baijiu sample are used as input to train the initial prediction model and obtain the trained model corresponding to the initial prediction model.

[0075] The training process can be as follows: taking the content of each selected flavor compound corresponding to the sensory feature descriptor of each training baijiu sample as input, the initial prediction model obtains the predicted score of the sensory feature descriptor through forward propagation, calculates the error between the predicted score of the sensory feature descriptor and the score (which is the true label) using the loss function, then propagates the error layer by layer through the backpropagation algorithm, calculates the gradient of the model parameters, and finally updates the model parameters according to the gradient. The process is repeated to minimize the error, thereby completing the model training and obtaining the trained model corresponding to the initial prediction model.

[0076] (3) For each initial prediction model, the content of each screened flavor compound corresponding to the sensory feature descriptor of each test liquor sample is used as input. The predicted score of the sensory feature descriptor of each test liquor sample is obtained by using the trained model corresponding to the initial prediction model. Based on the score and predicted score of the sensory feature descriptor of each test liquor sample, the model performance of the trained model corresponding to the initial prediction model is determined.

[0077] Based on the scores and predicted scores of the sensory feature descriptors for each test baijiu sample, the model performance of the trained model corresponding to the initial prediction model is determined. Specifically, this includes: using the scores and predicted scores of the sensory feature descriptors for each test baijiu sample as input, calculating the coefficient of determination of the trained model corresponding to the initial prediction model using the coefficient of determination formula, and calculating the root mean square error of the trained model corresponding to the initial prediction model using the root mean square error formula, thereby determining the model performance (coefficient of determination and root mean square error) of the trained model corresponding to the initial prediction model.

[0078] Specifically, the internal processing flow of the three initial prediction models follows the logic of "preprocessing, dataset partitioning, model training, parameter optimization, and prediction evaluation" step by step: first, logarithmic transformation is performed using log(content of flavor compounds after screening + 1e-10) to compress the data range, and then standardization and centering are performed to eliminate the influence of the unit and focus on the fluctuation characteristics of the variables to complete the preprocessing; then, a random seed (set.seed(123)) is set to ensure that the results are repeatable, and the training set and test set are divided in a 7:3 ratio to complete the dataset partitioning.

[0079] For different initial prediction models, the specific processes of model training, parameter optimization, and prediction evaluation are as follows: For the random forest model, based on the training set, the scores of sensory feature descriptors are used as response variables, and the contents of all screened flavor compounds corresponding to the sensory feature descriptors are used as feature variables. 5-fold cross-validation (trainControl(method="cv", number=5)) is used to control overfitting. The mtry parameter (the number of features randomly selected at each split of the random forest, which can try 5 different levels) is automatically tuned, 500 decision trees are set (number of decision trees ntree=500), and RMSE is used as the evaluation metric to complete the model training and determine the optimal mtry parameter. After parameter optimization, the trained model is obtained. Finally, the trained model is used to predict the test set. The postResample function is used to calculate the performance evaluation metric of the test set, and the prediction results and evaluation report (including the values ​​of the performance evaluation metric) are output to complete the prediction evaluation, thus completing the whole process from input data to prediction output.

[0080] For the Support Vector Machine (SVM) model, based on the training set, the scores of sensory feature descriptors are used as response variables, and the contents of all screened flavor compounds corresponding to the sensory feature descriptors are used as feature variables. The tune.svm function is used to fine-tune the model parameters. Different combinations of the penalty coefficient cost and the kernel function parameter gamma are selected for testing. The model is constructed using the ε-insensitive loss function (the tolerance band width epsilon=0.1) and the RBF kernel function. The optimal parameter combination is selected and the best SVM model is determined. The model training and parameter optimization are completed to obtain the trained model. Finally, the trained model is used to predict the test set. The postResample function is used to calculate the performance evaluation index of the test set, and the prediction results and evaluation report are output to complete the prediction evaluation, thus completing the entire process from input data to prediction output.

[0081] For the partial least squares model, based on the training set, principal components are iteratively extracted, taking into account the correlation between the content of selected flavor compounds and the scores of sensory feature descriptors. The prediction error corresponding to different numbers of principal components is calculated through 5-fold cross-validation (PRESS (Prediction Error Sum of Squares) can be used). The principal component with the smallest prediction error is selected as the optimal number of principal components to avoid overfitting. The model training and parameter optimization are completed, and the trained model is obtained. Finally, the trained model is used to predict the test set. The content of selected flavor compounds in the test set is projected onto the trained principal component space. The regression coefficients obtained from the training are used to convert the principal component scores into predicted scores of sensory feature descriptors. The performance evaluation index of the test set is obtained by calculating the error between the predicted scores. The prediction results and evaluation report are output to complete the prediction evaluation, thus completing the whole process from input data to prediction output.

[0082] After determining the model performance, the best-performing trained model is selected as the sensory feature prediction model corresponding to the sensory feature descriptor. Specifically, this involves: combining the coefficient of determination and the root mean square error, the user determines the best-performing trained model, and the best-performing trained model is selected as the sensory feature prediction model corresponding to the sensory feature descriptor.

[0083] The following section compares and analyzes the performance of the trained models corresponding to the three initial prediction models for each sensory feature descriptor. 2 The closer the RMSE is to 1, the closer it is to 0, indicating a stronger explanatory power of the model, a smaller deviation between the predicted and actual scores, and better model performance. A detailed comparison of model performance is shown in Table 4 below. Figure 4 As shown.

[0084] Table 4 Comparison of model performance on different sensory feature descriptors.

[0085] The above results demonstrate that the RF model performs well in predicting 11 sensory feature descriptors. 2 All scores were the highest, and R was achieved simultaneously in the prediction tasks of seven sensory feature descriptors, including soy sauce aroma, yeast aroma, floral aroma, sour aroma, roasted aroma, mellow sweetness, and lingering flavor. 2 The maximum value and minimum RMSE represent the best explanatory power and the lowest prediction error. PLS and SVM models only show local advantages in predicting individual sensory feature descriptors, but their overall stability is poor. For example, the SVM model has an RMSE as high as 1.23 in predicting the sensory feature descriptor "sour and fragrant," with a prediction error significantly higher than the RF model. The PLS model, however, has a lower RMSE in predicting multiple sensory feature descriptors such as "fragrant" and "sour and fragrant." 2 A value below 0.7 indicates poor fitting performance. Comprehensive comparison shows that the RF model has the best predictive performance for the sensory characteristics of baijiu, providing reliable technical support for the accurate evaluation and quality control of baijiu flavor.

[0086] Therefore, this embodiment determines that the random forest model is the optimal model for constructing quantitative predictions of the sensory characteristics and quality of baijiu. Specifically, for each sensory feature descriptor, the random forest model is selected as the optimal model (i.e., the sensory feature prediction model corresponding to that descriptor). This optimal model successfully establishes a reliable prediction method from chemical composition to macroscopic sensory characteristics. Of course, this embodiment could also determine an optimal model for each sensory feature descriptor separately.

[0087] This embodiment further introduces the analysis of key flavor compounds based on the RF model. Specifically, in order to reveal the material basis behind the sensory evaluation, this embodiment selects the best-performing random forest model and uses the mean decrease in Gini impurity as the feature importance evaluation index to quantify the contribution of each screened flavor compound to the sensory feature descriptor, thereby screening out key flavor compounds.

[0088] In this embodiment, key flavor compounds that are significantly associated with various sensory characteristic descriptors of baijiu were screened using a random forest model, as shown in Table 5 below.

[0089] Table 5 Key flavor compounds in the sensory characteristic descriptors of Baijiu (Chinese liquor)

[0090] Analysis shows that the selected key flavor compounds are highly consistent with traditional flavor chemistry theory, clearly explaining the chemical characterization corresponding to different sensory feature descriptors. Through the analysis of these key flavor compounds, the RF model not only achieved accurate prediction of sensory feature descriptor scores, but also constructed a correlation map of "flavor compounds-sensory features" at the molecular level, verifying the chemical interpretability of the model.

[0091] In this embodiment, if the sensory feature prediction model corresponding to the sensory feature descriptor is a random forest model, after obtaining the sensory feature prediction model corresponding to the sensory feature descriptor, the training method of the liquor quality prediction model in this embodiment further includes: for each screened flavor compound corresponding to the sensory feature descriptor, calculating the average Gini impurity reduction of the screened flavor compound based on the sensory feature prediction model corresponding to the sensory feature descriptor; if the average Gini impurity reduction of the screened flavor compound is greater than a third preset value, then the screened flavor compound is taken as the key flavor compound corresponding to the sensory feature descriptor.

[0092] The third preset value can be set by the user.

[0093] For the other two models, key flavor compounds can also be screened. Specifically, for the SVM model, a support vector machine based on recursive feature elimination (SVM-RFE) method can be used to screen key flavor compounds. An SVM regression model is constructed using radial basis function (RBF), and the model parameters are optimized through grid search combined with 5-fold cross-validation. The importance score of each feature (i.e., the screened flavor compound) is calculated, and key flavor compounds are screened based on the importance score of each feature. For the PLS model, latent variables are constructed between the X (screened flavor compound) and Y (sensory feature descriptor) spaces. By projecting in the direction that maximizes the covariance of X and Y, each latent variable contains a linear combination of the X variables, and the coefficient size reflects the importance. Key flavor compounds are screened based on the coefficient size.

[0094] The unique brewing process of baijiu results in a complex composition, and there is a complex relationship between the flavor compounds and the quality of the liquor. Therefore, the content of these compounds cannot be used as a single indicator of quality. This embodiment uses |r|>0.6 and VIP>1 to screen out the selected flavor compounds. Three initial prediction models are then established: random forest, support vector machine, and partial least squares model. Through model training and performance evaluation, the sensory feature prediction model corresponding to each sensory feature descriptor is determined, forming a quality prediction model. The aim is to achieve accurate prediction of the sensory characteristics and quality of baijiu. This quality prediction model is constructed based on the explicit correlation between sensory feature descriptors of baijiu and the selected flavor compounds. Each sensory feature descriptor corresponds to a specific selected flavor compound, and there is a correspondence between the input of the content of the selected flavor compounds and the output of the sensory feature descriptor scores. By inputting the content of these selected flavor compounds, the quality prediction model can directly output the scores of the corresponding sensory feature descriptors, achieving a precise mapping from "compound data to sensory scores." This successfully realizes a quantitative mapping from chemical composition to sensory experience, overcoming the limitations of traditional sensory evaluation, such as strong subjectivity and poor reproducibility. It provides a data-driven scientific solution for baijiu flavor standardization, product optimization, and intelligent manufacturing. This method can also be extended to other food flavor omics research fields, possessing significant theoretical value and application prospects.

[0095] This embodiment utilizes a gas chromatography (GC) and gas chromatography-mass spectrometry (GC-MS / MS) platform to accurately quantitatively characterize 208 flavor compounds in baijiu (Chinese liquor). Combining the unique flavor characteristics of baijiu, a customized sensory characteristic quantitative description system was constructed, encompassing eight aroma characteristics (such as soy sauce aroma, grain aroma, and floral aroma) and three taste characteristics (such as mellow sweetness and lingering aftertaste). Through a two-dimensional screening strategy of "correlation analysis (|r|>0.6) + variable importance projection (VIP>1)," differential flavor compounds were accurately identified from complex flavor compounds, further leading to the construction of sensory characteristic prediction models and quality prediction models. The sensory characteristic prediction model employs a random forest model, and validation results show that this random forest model achieves a prediction accuracy of R0 for most sensory characteristic descriptors. 2 The accuracy is >0.85, and it performs better in terms of prediction balance and reliability compared to support vector machine and partial least squares models.

[0096] This embodiment effectively solves the technical bottlenecks of traditional human sensory evaluation, which is characterized by strong subjectivity, poor result stability, and low evaluation efficiency. It successfully realizes the digital translation from the "chemical fingerprint" of baijiu to "sensory flavor perception". This method not only provides a data-driven scientific basis for the standardized control of baijiu flavor and the optimization of product quality, but also provides a scalable technical solution and practical paradigm for the standardization research in the field of food flavor omics.

[0097] Random forest models demonstrate significant advantages over traditional linear regression models in predicting the sensory flavor of baijiu. First, traditional linear regression models rely on strict statistical assumptions such as linearity, normality, and homoscedasticity. However, the formation of baijiu flavor is essentially the result of complex nonlinear interactions among hundreds of compounds, and the simplistic assumptions of linear models cannot accurately describe this complex system. In contrast, random forest models, as an ensemble tree model, do not require pre-defined functional relationships between variables and can automatically capture the nonlinear responses and interaction effects between flavor compounds and sensory feature descriptors, thus better reflecting the actual formation mechanism of baijiu flavor. Second, linear regression models are extremely sensitive to multicollinearity, and many flavor compounds in baijiu (such as homologues and derivatives) exhibit high correlations, leading to unstable linear regression coefficient estimates and difficulties in model interpretation. In contrast, random forest models, through random feature selection for each tree, naturally overcome the multicollinearity problem, ensuring the reliability of feature importance assessment. In addition, the feature importance evaluation mechanism built into the random forest model (based on the reduction of Gini impurity or the measure of out-of-bag error) can objectively and automatically screen out the key flavor compounds that contribute the most to a specific flavor, avoiding the bias of subjective pre-setting or stepwise screening in traditional methods.

[0098] Example 2 The liquor quality prediction model and application method provided in this application embodiment can be applied to, for example... Figure 1 The application environment shown illustrates this. The terminal communicates with the server via a network. A data storage system stores the data the server needs to process. This system can be set up independently, integrated into the server, or located in the cloud or on another server. The terminal can send application requests to the server. Upon receiving the request, the server, for each sensory feature descriptor, uses the content of each screened flavor compound corresponding to the sensory feature descriptor in the liquor to be predicted as input. It then uses the sensory feature prediction model corresponding to the sensory feature descriptor in the quality prediction model to determine the predicted score for that descriptor. Based on the predicted scores for all sensory feature descriptors of the liquor to be predicted, the quality of the liquor to be predicted is determined. The server can then feed back the application result—the predicted quality of the liquor to be predicted—to the terminal.

[0099] In addition, in some embodiments, the application method of the liquor quality prediction model can also be implemented by the server or the terminal alone. For example, the terminal can directly process the application request to be processed, or the server can obtain the application request to be processed from the data storage system and process it.

[0100] The terminals can be, but are not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices, while portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. Servers can be implemented using independent servers, server clusters composed of multiple servers, or cloud servers.

[0101] In one exemplary embodiment, such as Figure 5 As shown, a method for applying a liquor quality prediction model is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 The following steps, T1-T2, are used as an example to illustrate the process of using a server in the example.

[0102] Step T1: For each sensory feature descriptor, the content of each screened flavor compound corresponding to the sensory feature descriptor in the liquor to be predicted is used as input. The predicted score corresponding to the sensory feature descriptor is determined by the sensory feature prediction model corresponding to the sensory feature descriptor in the quality prediction model. The quality prediction model is a model trained using the liquor quality prediction model training method described in Example 1.

[0103] Step T2: Determine the quality of the liquor to be predicted based on the prediction scores corresponding to all the sensory feature descriptive words of the liquor to be predicted.

[0104] This application also provides an application scenario in which the above-mentioned liquor quality prediction model application method is applied. Specifically, the liquor quality prediction model application method provided in this embodiment can be applied in a liquor quality screening scenario. The liquor quality screening scenario includes a prediction stage, a display stage, and a screening stage. The prediction stage is used to determine the quality of each liquor in a batch; the display stage is used to show the user the quality of each liquor in a batch; and the screening stage is used to screen liquor based on the quality of each liquor in a batch, identifying liquors with poor quality. The liquor quality prediction model application method provided in this embodiment belongs to the prediction stage.

[0105] Example 3 In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements a method for training a liquor quality prediction model or a method for applying a liquor quality prediction model.

[0106] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0107] In one exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the liquor quality prediction model training method in Embodiment 1 or the liquor quality prediction model application method in Embodiment 2.

[0108] Example 4 In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the liquor quality prediction model training method in Embodiment 1 or the liquor quality prediction model application method in Embodiment 2.

[0109] Example 5 In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the liquor quality prediction model training method in Embodiment 1 or the liquor quality prediction model application method in Embodiment 2.

[0110] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0111] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for training a liquor quality prediction model, characterized in that, The training method for the liquor quality prediction model includes: Obtain a dataset; the dataset includes feature data and label data for each of the multiple baijiu samples. The feature data includes the content of each flavor compound among multiple flavor compounds. The label data includes the score of each sensory descriptor among multiple sensory descriptors. The sensory descriptors include aroma sensory descriptors and taste sensory descriptors. For each of the sensory characteristic descriptors, based on the content of each flavor compound in each baijiu sample and the score of the sensory characteristic descriptor, all the flavor compounds are screened to obtain the screened flavor compounds corresponding to the sensory characteristic descriptor. For each of the sensory feature descriptors, the content of each screened flavor compound corresponding to the sensory feature descriptor and the score of the sensory feature descriptor of each baijiu sample are used as inputs to train multiple initial prediction models, resulting in multiple trained models and the model performance of each trained model; the initial prediction model is a model that can perform the prediction function. For each sensory feature descriptor, the trained model with the best performance is selected as the sensory feature prediction model corresponding to the sensory feature descriptor; all sensory feature prediction models constitute a quality prediction model, which is used to predict the score of each sensory feature descriptor of baijiu and determine the quality of baijiu.

2. The method for training a liquor quality prediction model according to claim 1, characterized in that, Based on the content of each flavor compound in each baijiu sample and the score of the sensory characteristic descriptor, all the flavor compounds are screened to obtain the screened flavor compounds corresponding to the sensory characteristic descriptor, specifically including: Using the content of each flavor compound and the score of the sensory feature descriptor for each baijiu sample as input, the correlation coefficient between each flavor compound and the sensory feature descriptor is calculated using a correlation analysis algorithm, thus obtaining the correlation coefficient corresponding to each flavor compound. Using the content of each flavor compound and the score of the sensory feature descriptor for each baijiu sample as input, the VIP value between each flavor compound and the sensory feature descriptor is calculated using the variable projection importance algorithm, thus obtaining the VIP value corresponding to each flavor compound. For each flavor compound, if the absolute value of the correlation coefficient corresponding to the flavor compound is greater than a first preset value and the VIP value corresponding to the flavor compound is greater than a second preset value, then the flavor compound is selected as a screened flavor compound to screen all the flavor compounds and obtain the screened flavor compounds corresponding to the sensory feature descriptor.

3. The method for training a liquor quality prediction model according to claim 1, characterized in that, The initial prediction models include: random forest model, support vector machine model, and partial least squares model; The model performance is characterized by performance evaluation metrics, including the coefficient of determination and root mean square error.

4. The method for training a liquor quality prediction model according to claim 1, characterized in that, Using the content of each screened flavor compound corresponding to the sensory feature descriptor and the score of the sensory feature descriptor for each baijiu sample as input, multiple initial prediction models are trained to obtain multiple trained models and the model performance of each trained model, specifically including: All the liquor samples were divided into training liquor samples and test liquor samples. For each of the initial prediction models, the content of each screened flavor compound corresponding to the sensory feature descriptor and the score of the sensory feature descriptor of each training baijiu sample are used as inputs to train the initial prediction model and obtain the trained model corresponding to the initial prediction model. For each of the initial prediction models, the content of each screened flavor compound corresponding to the sensory feature descriptor of each test liquor sample is used as input. The predicted score of the sensory feature descriptor of each test liquor sample is predicted by the trained model corresponding to the initial prediction model. Based on the score of the sensory feature descriptor and the predicted score of each test liquor sample, the model performance of the trained model corresponding to the initial prediction model is determined.

5. The method for training a liquor quality prediction model according to claim 3, characterized in that, If the sensory feature prediction model corresponding to the sensory feature descriptor is a random forest model, the training method for the liquor quality prediction model after obtaining the sensory feature prediction model corresponding to the sensory feature descriptor further includes: For each of the selected flavor compounds corresponding to the sensory feature descriptor, the average reduction in Gini impurity of the selected flavor compound is calculated based on the sensory feature prediction model corresponding to the sensory feature descriptor. If the average reduction in Gini impurity of the selected flavor compound is greater than a third preset value, then the selected flavor compound is regarded as the key flavor compound corresponding to the sensory feature descriptor.

6. The method for training a liquor quality prediction model according to claim 1, characterized in that, The aroma sensory characteristics descriptions include soy sauce aroma, grain aroma, yeast aroma, fruit aroma, sweet aroma, sour aroma, floral aroma, and baking aroma; the taste sensory characteristics descriptions include mellow sweetness, mellowness, and long-lasting flavor.

7. A method for applying a liquor quality prediction model, characterized in that, The application method of the liquor quality prediction model includes: For each sensory feature descriptor, the content of each screened flavor compound corresponding to the sensory feature descriptor in the liquor to be predicted is used as input, and the predicted score corresponding to the sensory feature descriptor is determined by the sensory feature prediction model corresponding to the sensory feature descriptor in the quality prediction model; the quality prediction model is a model trained using the liquor quality prediction model training method according to any one of claims 1-6. The quality of the liquor to be predicted is determined based on the predicted scores corresponding to all the sensory feature descriptors of the liquor to be predicted.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the liquor quality prediction model training method according to any one of claims 1-6 or the liquor quality prediction model application method according to claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method for the liquor quality prediction model according to any one of claims 1-6 or the application method for the liquor quality prediction model according to claim 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method for the liquor quality prediction model according to any one of claims 1-6 or the application method for the liquor quality prediction model according to claim 7.

Images