Rapid pre-discrimination method for outputting sensory values according to flavor characteristics of cow milk

Abstract

The invention provides a rapid pre-discrimination method for outputting sensory values according to flavor characteristics of cow milk. The rapid pre-discrimination method comprises the following steps: performing program optimization on a raw milk sample by adopting HS-SPME-GC-MS to determine volatile matters; heating the raw milk and then carrying out sensory evaluation; measuring by an electronic nose to obtain a sensor response value; establishing a sensory value prediction model based on volatile flavor factor data or fusion of volatile flavor factors and sensor signal data, and evaluating a model effect; and introducing a new sample, and pre-judging the flavor attribute according to the two established models. According to the method, a multi-data fusion method is creatively adopted, that is, the accuracy is higher by combining the temperament data and the electronic nose data and establishing the model with the sensory values, the model accuracy can be improved by increasing the sample number in a gradient mode, and the sensory values are more comprehensively and accurately output. The reason for establishing the model for the two types of data is that the sensory value information can be predicted more accurately according to which data source, and the subjective error and time cost of sensory evaluation personnel are effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of detection technology, and in particular to a rapid pre-judgment method for outputting sensory values ​​based on the flavor characteristics of milk. Background Technology

[0002] Flavor is a crucial factor determining consumer acceptance and purchasing power, even surpassing product price in importance. The flavor of dairy products is influenced at all four key stages of the supply chain: pasture, raw milk, processing, and packaging. Raw milk, in particular, varies significantly in flavor between different dairy farms due to differences in dairy cow feed and management methods. Developing effective raw milk testing models is essential for identifying ideal flavor attributes early in production to improve the overall quality of dairy products for consumers.

[0003] The China Raw Milk Quality, Safety and Nutrition Database employs a scientific classification system, categorizing raw milk based on four key indicators: milk fat content, milk protein content, total bacterial count, and somatic cell count. For a long time, stakeholders in the dairy industry, including dairy farmers, buyers, and experts, have focused on maximizing milk yield, increasing nutritional content, and enhancing milk activity, while flavor—a crucial factor influencing the sensory quality of food—has often been overlooked. This is partly due to the challenges of scientifically investigating flavor perception, particularly when objectively measuring sensory experiences, where traditional assessment methods are limited by subjectivity. While population-based sensory panels can provide more authentic data, they are often costly, time-consuming, and difficult to implement on a large scale, especially for raw milk. Therefore, developing inexpensive and scalable sensory flavor assessment models for raw milk is crucial for predicting sensory attributes in advance. Summary of the Invention

[0004] In view of this, the technical problem to be solved by the present invention is to provide a rapid predictive method for outputting sensory values ​​based on the flavor characteristics of milk. The method provided by the present invention can predict sensory attributes based on key flavor components and can be further applied to the preliminary grading of dairy product flavor quality. This method can replace manual sensory evaluation, reduce the cost of sensory evaluation, and at the same time has a certain degree of stability.

[0005] This invention provides a rapid pre-judgment method for outputting sensory values ​​based on milk flavor characteristics, comprising the following steps:

[0006] A) The volatile flavor factors of the raw milk samples were determined by HS-SPME-GC-MS to obtain the volatile components of the raw milk samples;

[0007] B) Heat the raw milk, perform sensory evaluation, and obtain sensory evaluation data;

[0008] C) The raw milk was measured using an electronic nose to obtain the response value of the electronic nose sensor;

[0009] D) Establish a Python algorithm to build a sensory value prediction model based on the volatile components and sensory evaluation data of raw milk samples, and based on the volatile components of raw milk fused with the response value of the electronic nose sensor and sensory evaluation data, and evaluate the results.

[0010] E) Based on the above sensory value prediction model, introduce new raw milk samples and establish machine learning models based on their volatile flavor factor data or the combination of volatile flavor factors and sensor signal data to predict their flavor attributes.

[0011] In some specific embodiments, the method for determining volatile flavor factors in step A) includes:

[0012] 8 mL of raw milk and 1.0000±0.0005 g of NaCl were placed in a headspace sample vial. After equilibration in a 40 °C water bath for 20 min, SPME fiber was inserted into the headspace vial for adsorption for 30 min. 2-Methyl-3-heptanone was used as the internal standard for the volatile flavor experiment. The headspace gas was collected, and then the SPME extraction fiber enriched with volatile flavor was inserted into the GC inlet for desorption for 5 min. After that, a gradient temperature program was run.

[0013] In some specific implementations, step A) GC-MS parameters include: gas chromatography; the chromatographic column is a DB-WAX highly polar column (column size 30m*0.25mm*0.25μm).

[0014] Helium was used as the carrier gas at a flow rate of 1 mL / min. Electron impact ionization (EI) mode was used at 70 eV; the mass spectrometer ion source temperature was 230 °C, the quadrupole temperature was 150 °C, and the injection port temperature was 250 °C.

[0015] The volatile aroma components were identified using the NIST 20 spectral library.

[0016] In some specific embodiments, step C) includes:

[0017] Signals were acquired using a PEN3 portable electronic sensor. The settling time for each sample was between 115 and 120 seconds, and the response value for each sensor was calculated as G / G0, where G and G0 represent the sensor's response to the sample and air, respectively.

[0018] In some specific implementations, the evaluation described in step D) includes:

[0019] Based on the volatile components and sensory evaluation data of raw milk samples, 19 models were established, and the regression results were evaluated according to the mean absolute error (MAE) and accuracy (ACC).

[0020] In some specific implementations, the evaluation in step D) includes: evaluating the regression results based on mean absolute error and accuracy.

[0021] In some specific implementations, the model optimization in step D) includes: increasing the dataset in increments of 10, from 30 samples to 80 samples, to train the model and improve the accuracy of the model in predicting sensory information. This is also a prominent innovation of this patent.

[0022] In some specific implementations, the prediction model is a prediction model for creamy flavor, milky flavor, salty flavor, and overall preference.

[0023] In some specific implementations, the model described in step D) includes:

[0024] (Random Forest model, bagging method (bootstrapping aggregation method), adaptive boosting algorithm, gradient boosting machine, support vector machine (radial basis kernel function), extreme gradient boosting model (XGBoost model), Bayesian model (usually referring to Naive Bayes classifier), support vector machine (multinomial kernel function), lasso regression (L1 regularized regression), decision tree, support vector machine (linear kernel function), ridge regression (L2 regularized regression), elastic network regression (L1+L2 regularized), linear regression, neural network, artificial neural network, K-nearest neighbor algorithm, logistic regression, multilayer perceptron, a total of 19 machine learning models).

[0025] In some specific implementations, the raw milk samples are sourced from typical pastures in Northwest China.

[0026] Compared with existing technologies, this invention provides a rapid predictive method for sensory values ​​based on the flavor characteristics of milk, comprising the following steps: A) Optimizing the GC-MS heating procedure for raw milk and using HS-SPME-GC-MS to determine volatile flavor factors, obtaining the volatile components of the raw milk sample; B) Heating the raw milk and performing sensory evaluation to obtain sensory evaluation data; C) Measuring the raw milk using an electronic nose to obtain the electronic nose sensor response value; D) Using a Python script, establishing a multi-data fusion sensory value prediction model based on the volatile components and sensory values ​​of the raw milk sample, and based on the volatile components combined with the electronic nose sensor response value and sensory evaluation value, and evaluating the model accuracy and model error results; E) To examine the impact of sample size on the accuracy of the modeling results, this invention innovatively increases the sample size on the overall model accuracy; F) Based on the above sensory value prediction model, introducing new milk samples and predicting their flavor attributes based on their volatile flavor factor data or volatile flavor factors and sensor signal data. This invention finds that gradually increasing the sample size can improve model accuracy, and a large amount of data is beneficial to the overall representativeness of the model. This invention creatively discovers that employing a multi-data fusion method—combining gas chromatography-mass spectrometry (GC-MS) data with electronic nose data and sensory evaluation to build a model—results in higher accuracy and a more comprehensive prediction of milk's sensory information. The GC-MS + electronic nose data fusion method can more comprehensively and detailedly represent the overall flavor profile of milk, making it closer to the sensory evaluation model. Furthermore, as the gradient increases the sample size, the accuracy of the model based on volatile components combined with E-nose and sensory evaluation increases, which is crucial for improving model accuracy. By more comprehensively capturing data distribution and reducing random errors, the generalization ability and reliability of the model are fundamentally improved. Attached Figure Description

[0027] Figure 1 Technical roadmap for constructing a raw milk flavor quality evaluation model;

[0028] Figure 2 Total ion chromatogram of GC-MS data for a single sample in Comparative Example 1 (GC-MS optimization program);

[0029] Figure 3 Total ion chromatogram of GC-MS data for a single sample in Example 1 (GC-MS raw program);

[0030] Figure 4 E-nose spectra of all samples in Example 2;

[0031] Figure 5 Example 2: Accuracy analysis of predicting milk flavor, cream flavor, and salty flavor based on GC-MS volatile component data gradient with increased sample size and preference.

[0032] Figure 6 Based on GC-MS volatile component data and electronic nose E-Nose data, the sample size was increased in a gradient to predict the milky aroma, creamy aroma, and salty aroma, and the accuracy of preference analysis was performed. Detailed Implementation

[0033] This invention provides a rapid pre-judgment method for outputting sensory values ​​based on the flavor characteristics of milk. Those skilled in the art can refer to the content of this document and appropriately modify the process parameters to achieve this. It should be particularly noted that all similar substitutions and modifications are obvious to those skilled in the art and fall within the scope of protection of this invention. The method and application of this invention have been described through preferred embodiments. Those skilled in the art can obviously modify or appropriately change and combine the method and application described herein without departing from the content, spirit, and scope of this invention to realize and apply the technology of this invention.

[0034] This invention employs multimodal data fusion to establish a machine learning model for predicting the sensory characteristics of milk and consumer preferences. It addresses the critical need for efficient flavor phenotypic analysis in the dairy industry, providing technical support for flavor enhancement research. The method can be used directly on raw milk without requiring any processing equipment or a large number of sensory evaluation personnel. It enables rapid online flavor prediction of raw milk. It is green and economical, reduces energy consumption, and is easy to promote. This technology has considerable market potential.

[0035] This invention provides a rapid pre-judgment method for outputting sensory values ​​based on milk flavor characteristics, comprising the following steps:

[0036] A) The volatile flavor factors of the raw milk samples were determined by HS-SPME-GC-MS to obtain the volatile components of the raw milk samples;

[0037] B) Heat the raw milk, perform sensory evaluation, and obtain sensory evaluation data;

[0038] C) The raw milk was measured using an electronic nose to obtain the response value of the electronic nose sensor;

[0039] D) Based on the volatile components and sensory evaluation data of raw milk samples, the volatile components are fused with the electronic nose sensor response value and sensory evaluation data to establish a sensory value prediction model and evaluate the results.

[0040] E) Based on the above sensory value prediction model, introduce new raw milk samples and predict their flavor attributes based on their volatile flavor factor data or volatile flavor factor fusion sensor signal data.

[0041] F) Innovatively, gradient increases the number of prediction set samples, increasing the accuracy and breadth of modeling.

[0042] This invention provides a rapid pre-judgment method for outputting sensory values ​​based on the flavor characteristics of milk. First, the volatile flavor factors of the raw milk sample are determined by HS-SPME-GC-MS to obtain the volatile components of the raw milk sample.

[0043] The raw milk used in this invention comes from typical pastoral areas in Northwest China. Fresh raw milk is collected from the pastures and transported to the laboratory under refrigeration for 3 hours. SPME adsorption of flavor compounds is performed on the raw milk, and four parallel data points are obtained for each sample. The volatile flavor components are qualitatively and quantitatively determined using NIST 20, retention index, and standards.

[0044] This invention collects raw milk samples from typical Chinese pastures, with four parallel groups for each differential sample. Qualitative and quantitative analysis is performed using HS-SPME-GC-MS. Sodium chloride is added to each raw milk (RM) sample (the purpose of sodium chloride is to break the emulsion and promote the release of volatile aromas), and the samples are placed in headspace vials. The internal standard is 2-methyl-3-heptanone (containing 0.03 ppm of acetone).

[0045] Helium was used as the carrier gas at a flow rate of 1 mL / min. Mass spectrometry was performed at an electron voltage of 70 eV, with the mass source, quadrupole, and interface temperatures set at 230 °C, 150 °C, and 250 °C, respectively. Volatile organic compounds (VOCs) were identified by comparing the mass spectra with those in the National Institute of Standards and Technology (NIST) 20.0 database. Qualitative analysis was conducted using standards, and VOCs were analyzed using an internal standard semi-quantitative method to provide a comprehensive overview of volatile flavor components.

[0046] In some specific embodiments, the method for determining volatile flavor factors includes:

[0047] The raw milk was placed in a headspace vial, and the heating temperature was 40°C. After a 20-minute water bath, SPME fibers were inserted into the headspace vial and equilibrated for 30 minutes. The headspace gas release was collected, and sensor signal values ​​were acquired over 115–120 seconds.

[0048] GC-MS parameters include: Gas Chromatography 8890GC System 5977GC / MSD; Column: DB-WAX high polarity column (30m*0.25mm*0.25μm). Gradient temperature program: 30℃ for 1 min, then increase to 90℃ at 3℃ / min, hold for 0 min, and finally increase to 220℃ at 6℃ / min, hold for 5 min (this program is the optimized program of this patent); This invention achieves the following by controlling the above parameters as a whole: ① Solvent focusing locks low-boiling-point substances; ② Cold trap focusing confines high-boiling-point substances; ③ Protects the chromatographic column and MS system; ④ Compresses peak width to improve sensitivity, ultimately achieving efficient and accurate analysis of trace components in complex mixtures.

[0049] The inventors have discovered that a low initial temperature heating program is beneficial for the detection of more compounds such as acids and ketones.

[0050] The volatile aroma components were identified using the NIST 20 spectral library.

[0051] The raw milk was heated, and sensory evaluation was conducted to obtain sensory evaluation data.

[0052] The present invention, after being optimized, is subjected to sensory evaluation using the following method:

[0053] A sensory team of 12 members with over 13 years of sensory skills training was recruited from the Molecular Sensory Science Laboratory of Beijing Technology and Business University (Beijing, China). The team consisted of 6 men and 6 women, with an average age of 22-30. The team members then underwent a week of routine training to ensure accuracy in identifying the supplied samples.

[0054] The sensory evaluation room was equipped with the following conditions: clean, odorless, noise-free, well-lit, 25°C, and 65% relative humidity. Separate cubicles were set up to ensure each group member was undisturbed. Group members selected four attributes: creamy, milky, milky-salty, and overall preference. A 7-point sensory evaluation scale was used, starting from 1 and ending at 7, with the following score distribution: 1 (no odor), 2-4 (slightly weak), 5-6 (moderate), and 7 (very strong). 8 mL of sample was accurately weighed and placed in a clean, odorless brown headspace vial labeled with a random three-digit number corresponding to the number on the sensory evaluation scale. After standing for 20 minutes, the sample was provided to the sensory group for evaluation. Before evaluating the next sample, group members received coffee beans to refresh their sense of smell. Each sample was evaluated three times by each group member.

[0055] The raw milk was tested using an electronic nose to obtain the response value of the electronic nose sensor.

[0056] Some specific implementations include:

[0057] Signals were acquired using a PEN3 portable electronic sensor. The settling time for each sample was between 115 and 120 seconds, and the response value for each sensor was calculated as G / G0, where G and G0 represent the sensor's response to the sample and air, respectively.

[0058] Based on the volatile components of raw milk samples, the response values ​​of electronic nose sensors, and sensory evaluation data, a sensory value prediction model was established, and the results were evaluated.

[0059] The inventors established a model based on volatile flavor components from gas chromatography-mass spectrometry (GC-MS) measurements and sensory data. They found that predicting sensory values ​​solely based on GC-MS volatile data had low accuracy. Therefore, they employed a multi-data fusion method, combining GC-MS data with electronic nose data and sensory values ​​to build a model that achieved higher accuracy and a more comprehensive prediction of milk's sensory properties. In some specific embodiments, the evaluation includes:

[0060] Based on the volatile components and sensory evaluation data of raw milk samples, 19 models were established, and the regression results were evaluated according to the mean absolute error (MAE) and accuracy (ACC).

[0061] In some specific implementations, the evaluation includes: evaluating the regression results based on mean absolute error and accuracy.

[0062] In some specific implementations, the prediction model is a prediction model for creamy flavor, salty flavor, milky flavor, and overall preference.

[0063] In some specific embodiments, the model includes:

[0064] Random Forest, Bag-of-the-Domain (BOP) clustering, Adaptive Boosting, Gradient Boosting Machine, Support Vector Machine (Radial Basis Kernel), Extreme Gradient Boosting (XGBoost), Bayesian Model (usually referring to Naive Bayes classifier), Support Vector Machine (Multinomial Kernel), Lasso Regression (L1 Regularized Regression), Decision Tree, Support Vector Machine (Linear Kernel), Ridge Regression (L2 Regularized Regression), Elastic Network Regression (L1+L2 Regularized), Linear Regression, Neural Network, Artificial Neural Network, K-Nearest Neighbors Algorithm, Logistic Regression, and Multilayer Perceptron—a total of 19 machine learning models.

[0065] In some specific implementations, to further evaluate the impact of the flavor holographic dataset on the model, based on the aforementioned modeling method, the inventors analyzed 30 randomly selected samples as a test set. Then, the dataset was increased in increments of 10, from 30 samples to 80 samples, to train the model and improve its accuracy in predicting sensory responses. Gradient techniques were used to progressively expand the test dataset from 30 instances to 80 instances in increments of 10 samples; the new samples were subsequently used for external validation.

[0066] This invention utilizes GC-MS data and sensory data for modeling, combining GC-MS data with E-nose and sensory data. Nineteen machine learning models were established using these two methods, outputting sensory values ​​based on volatile compounds or volatile compounds combined with electronic nose data. Compared to the traditional OPLS-DA model, our machine learning models break the linear assumption, enhance predictive capabilities (high-precision classification / regression), expand application scenarios (real-time monitoring, multimodal data fusion), and improve computational efficiency. We found that different models are suitable for sensory prediction of different flavors, providing a good paradigm for predicting diverse sensory preferences. Overall, among the 19 models based on volatile flavors and sensory data, eight models achieved an average prediction accuracy of over 85% for overall sensory attributes, particularly the Random Forest model, bagging method, gradient boosting machine, support vector machine (radial kernel), Bayesian regression, Lasso regression, support vector machine (linear sum function), and K-nearest neighbor model. Of the 19 models built based on volatile flavor component data, electronic nose data, and sensory values, 10 models achieved a prediction rate of over 85% for overall sensory attributes. These models are: Random Forest, Gradient Boosting Machine, Support Vector Machine (Meridian Kernel), Extreme Gradient Boosting (XGBoost), Bayesian regression, Lasso regression, Decision Tree, Support Vector Machine (Linear), Ridge regression, and K-Nearest Neighbor. Among the 19 models built based on GC-MS data, 12, 15, and 14 models, respectively, achieved accuracy rates above 0.7 in predicting milky aroma, milky-salty flavor, and creamy flavor. Furthermore, 13, 19, and 19 models, respectively, based on fusion data from GC-MS and electronic nose, achieved accuracy rates above 0.7 in predicting milky aroma, milky-salty flavor, and creamy flavor. In conclusion, models built based on volatile flavor data, electronic nose data, and sensory values ​​showed higher accuracy in outputting sensory values, indicating that data fusion improved the accuracy of the modeling. Furthermore, this patent discovers that different models are built based on two data sources. The inventors found that different sensory value predictions can be predicted better using different models, which is related to the overall characteristics of the data and the generalization ability of the model.

[0067] The model of this invention has a low mean absolute error (MAE) and a high accuracy (ACC).

[0068] The sensory attribute pre-judgment model established by the method of this invention avoids the problem of relying on a large number of consumers in traditional sensory perception quantification methods, and overcomes the limitations of traditional sensory evaluation being highly subjective and time-consuming. It is suitable for high-throughput detection of large numbers of samples, and this high-throughput flavor phenotypic analysis method is expected to significantly improve the efficiency of flavor analysis. By implementing a specialized dairy product processing strategy based on the milk source sensory pre-judgment model, disputes between purchasers, suppliers, and consumers caused by inconsistencies in milk source quality can be effectively reduced. This method is crucial for assessing and improving the quality of finished dairy products.

[0069] It should be understood that the expression “one or more of…” individually includes each of the objects described after the expression, as well as various different combinations of two or more of the described objects, unless otherwise understood from the context and usage. The expression “and / or” combined with three or more described objects should be understood to have the same meaning, unless otherwise understood from the context.

[0070] The terms “including,” “having,” or “containing,” including the use of their grammatical synonyms, should generally be understood as open-ended and non-restrictive, for example, not excluding other unstated elements or steps, unless otherwise specifically stated or understood from the context.

[0071] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural.

[0072] In this application, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions mean any combination of these items, including any combination of single or multiple items.

[0073] It should be understood that the order of the steps or the order in which certain actions are performed is not important as long as the invention remains operational. Furthermore, two or more steps or actions can be performed simultaneously.

[0074] The use of any and all instances or exemplary language such as “e.g.” or “including” in this document is merely intended to better illustrate the invention and is not intended to limit the scope of the invention unless the claims are made. No language in this specification should be construed as indicating that any unclaimed element is essential to the practice of the invention.

[0075] Furthermore, the numerical ranges and parameters used to define the present invention are approximate values, and the relevant values ​​in the specific embodiments have been presented as precisely as possible. However, any value inevitably contains standard deviations due to individual test methods. Therefore, unless explicitly stated otherwise, it should be understood that all ranges, quantities, values, and percentages used in this disclosure are modified with the word "approximately." Here, "approximately" generally means an actual value within plus or minus 10%, 5%, 1%, or 0.5% of a particular value or range.

[0076] It should be understood that in the various embodiments of this application, the order of the above processes does not imply the order of execution. Some or all steps may be executed in parallel or sequentially. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0077] The embodiments and comparative examples of this invention describe some examples, in which the embodiments illustrate certain implementations of the invention. However, this does not mean that the effects of the invention can only be achieved in these examples.

[0078] To further illustrate the present invention, the following describes in detail, with reference to embodiments, a rapid pre-judgment method for outputting sensory values ​​based on milk flavor characteristics provided by the present invention.

[0079] Comparative Example 1:

[0080] Regarding the HS-SPME-GC-MS parameters, a DB-WAX highly polar column was used. The temperature program consisted of a 35°C hold for 1 min, followed by a 3°C / min ramp to 90°C hold for 0 min, and finally a 6°C / min ramp to 210°C hold for 2 min. This temperature program is the standard procedure for milk analysis. The total ion chromatogram of a single sample in the comparative example is shown below. Figure 2 .

[0081] Example 1:

[0082] Regarding the HS-SPME-GC-MS parameters, a DB-WAX highly polar column was used. The temperature program employed was a newly optimized program for raw milk: 30℃ for 1 min, then increased to 90℃ at a rate of 3℃ / min, held for 0 min, and finally increased to 220℃ at a rate of 6℃ / min, held for 5 min. This temperature program is the latest optimized program for raw milk, and compared with related programs reported in the literature, it uses a lower initial temperature. The optimized GC-MS spectrum is shown in Figure (…). Figure 3 As shown in the image:

[0083] Samples prepared using a milk-based temperature program with an initial temperature of 30°C showed high resolution and yielded high levels of volatile compounds. A lower initial temperature (30°C) detected some hydrocarbons, such as eicosane, styrene, and 4-methyloctane; and high-boiling-point acids, such as pterin-6-carboxylic acid and octanoic acid / hexanoic acid. The GC-MS temperature program, starting at a low temperature, essentially utilizes temperature control to achieve physical focusing of components at the column head, addressing the core challenge of separating samples with a wide boiling range. This strategy achieves efficient and accurate analysis of trace components in complex mixtures through: ① solvent focusing to lock in low-boiling-point compounds; ② cold trap focusing to constrain high-boiling-point compounds; ③ protection of the column and MS system; and ④ peak width compression to improve sensitivity. Experimental results show that a lower initial temperature program produces better peak shapes, facilitating the detection of more acids and ketones. In comparison, fewer volatile flavor compounds were detected in the comparative example.

[0084] Example 2

[0085] like Figure 1 The diagram illustrates a rapid predictive method for outputting sensory values ​​based on milk flavor characteristics, specifically including:

[0086] Step 1: Sensory evaluation of preheated raw milk samples. The sensory analysis was conducted by 20 trained reviewers (aged 20-35) with a history of daily milk consumption. Reviewers were in good health, possessed excellent verbal communication skills, and had no history of smoking or alcohol consumption. After recruitment, reviewers received training to familiarize themselves with flavor terminology and relevant standard compounds, and scored the milk aroma. The intensity of each identified aroma and flavor was evaluated, and a sensory evaluation of the flavor was performed.

[0087] 2. Raw milk samples were collected from typical Chinese pastures, with four parallel groups for each differential sample. Qualitative and quantitative analysis was performed using HS-SPME-GC-MS. 1.0000±0.0005 g of sodium chloride was added to each raw milk (RM) sample and placed in a headspace vial. 2-Methyl-3-heptanone (containing 0.03 ppm of acetone) was used as the internal standard. Helium was used as the carrier gas at a flow rate of 1 mL / min. Mass spectrometry was performed at an electron voltage of 70 eV, with the mass source, quadrupole, and interface temperatures set to 230℃, 150℃, and 250℃, respectively. Volatile organic compounds were identified by comparing the mass spectra with those in the National Institute of Standards and Technology (NIST 20.0) database. The samples were then placed in headspace vials, incubated in a water bath for 20 min, and then the SPME fiber was inserted into the headspace vial for equilibration for 30 min. Helium was used as the carrier gas at a flow rate of 1 mL / min. Mass spectrometry was recorded at an electronic voltage of 70 eV, with the mass source, quadrupole, and interface temperatures set at 230 °C, 150 °C, and 250 °C, respectively. Compound analysis was performed by comparing the total ion chromatogram with mass spectra from the National Institute of Standards and Technology (NIST) 20.0 database and by qualitative analysis using standards. Volatile organic compounds were analyzed using an internal standard semi-quantitative method, providing a comprehensive analysis of volatile flavor components.

[0088] 3. Volatile aroma release analysis. Signals were acquired using a PEN3 portable electronic sensor. The settling time for each sample was between 115 and 120 seconds, and the response value for each sensor was calculated as G / G0, where G and G0 represent the sensor's response to the sample and air, respectively.

[0089] 4. Select the volatile component data from step (2) and the sensory data from step (1) for machine learning modeling. The 19 models include: Random Forest model, bagging method (bootstrap aggregation method), adaptive boosting algorithm, gradient boosting machine, support vector machine (radial basis kernel function), extreme gradient boosting model (XGBoost model), Bayesian model (usually referring to Naive Bayes classifier), support vector machine (multinomial kernel function), lasso regression (L1 regularized regression), decision tree, support vector machine (linear kernel function), ridge regression (L2 regularized regression), elastic network regression (L1+L2 regularization), linear regression, neural network, artificial neural network, K-nearest neighbor algorithm, logistic regression, and multilayer perceptron.

[0090] Of the 19 models built based on volatile flavor component data and sensory values, 8 models had a prediction rate of over 85% for overall sensory attributes: Random Forest, Bagged Method, Gradient Boosting Machine, Support Vector Machine (Meridian Kernel), Bayesian Method, Lasso Regression, Linear Support Vector Machine, and K-Nearest Neighbor (see Table 1). Of the 19 models built based on volatile flavor component data, electronic nose data, and sensory values, 10 models had a prediction rate of over 85% for overall sensory attributes: Random Forest, Gradient Boosting Machine, Support Vector Machine (Meridian Kernel), XG Boost, Bayesian Method, Lasso Regression, Decision Tree, Linear Support Vector Machine, Ridge Regression, and K-Nearest Neighbor (see Table 2). The regression results were evaluated based on Mean Absolute Error (MAE) and Accuracy (ACC).

[0091] 5. To further evaluate the impact of the flavor holographic dataset on the model, based on the modeling method in (4), we analyzed 30 randomly selected samples as the test set. Then, we increased the dataset by 10 increments, from 30 samples to 80 samples, to train the model and improve the accuracy of the model in predicting sensory information.

[0092] 6. In order to improve the accuracy of model (4), we combined flavor omics data with electronic nose gas sensor data to build a more accurate prediction model for milky, creamy, salty, and preference flavors, and evaluated the predictive ability of the model.

[0093] 7. Modeling method based on (6). Gradient technique is used to gradually expand the test dataset from 30 instances to 80 instances in increments of 10 samples. The new samples are then used for external validation.

[0094] 8. In one embodiment, the specific steps include:

[0095] (1) In one embodiment, the raw milk is heated to the same temperature and then subjected to sensory evaluation (milk oil flavor, saltiness, overall preference);

[0096] (2) The volatile aroma components were analyzed using the volatile flavor collection method in step (2) of Example 2;

[0097] (3) The volatile aroma release characteristics were analyzed using the volatile aroma signal acquisition method in step (3) of Example 2;

[0098] (4) In the regression models built based on temperament data and overall sensory values, the absolute error (MAE) of all eight models, namely Random Forest, Bag Method, K-Nearest Neighbor, Gradient Boosting Machine, Bayesian Model, Lasso Regression (L1 Regularized Regression), Support Vector Machine (Linear Kernel Function), and Support Vector Regression-Radial Basis Kernel Function, is <1, and the accuracy is above 85%.

[0099] (5) Among the regression models built based on temperament data combined with electronic nose and overall sensory values, random forest, gradient boosting machine, K-nearest neighbor, extreme gradient boosting, support vector regression with radial basis function, Bayesian model, lasso model, decision tree, support vector machine (linear kernel function), and ridge regression were used. The mean absolute error (MAE) of all 10 models was <1, and the accuracy was above 85%. The mean absolute error (MAE) of the association model based on temperament combined with electronic nose and sensory values ​​was even lower.

[0100] (6) Nineteen models were established based on GC-MS data to predict different sensory attributes (milky aroma, milky saltiness, and creamy flavor) in milk. Twelve models achieved a prediction accuracy of over 70% for milk aroma, fourteen for milky saltiness (over 90%), and fourteen for creamy flavor (over 70%) (Table 3). Nineteen machine learning models were established based on GC-MS combined with electronic nose sensors and sensory evaluation attributes. Thirteen models achieved a prediction accuracy of over 70% for milk aroma, nineteen for milky saltiness (over 90%), and nineteen for creamy flavor (over 70%) (Table 4). Overall, the models established based on GC-MS data fused with electronic nose data showed high accuracy in predicting sensory values. However, the accuracy of different models varied for predicting different sensory attributes (milky aroma, milky saltiness, and creamy flavor). Models with higher accuracy could be selected for specific sensory attributes.

[0101] (7) Using the methods mentioned in steps (2) and (4), establish prediction models for milk flavor, cream flavor, saltiness, and overall preference, and optimize the machine learning model by increasing the number of prediction sets by 10 samples. In the prediction of overall preference for cream, when the number of samples gradually increases from 30 to 80 with a gradient of 10, the linear model, elastic network, and ridge regression model achieve a prediction accuracy of over 80% for the cream flavor in raw milk. When the sample size reaches 60 or more, the linear regression, elastic network regression, and ridge regression achieve a prediction accuracy of over 80% for the saltiness attribute in milk. With the increase in the number of samples, the accuracy of K-nearest neighbor, decision tree, and radial basis function support vector machine models for milk flavor has improved. When the sample size increases from 70 to 80, the bagging method, random forest model, and artificial neural network model increase the prediction accuracy for milk preference. Figure 5 );

[0102] (8) Using the method mentioned in steps (2)(3)(4)(5) based on volatile component data and E-Nose multimodal data, prediction models for milk flavor, cream flavor, salty flavor, and overall preference were established. For the milk flavor attribute, when the number of prediction set samples reached more than 60, the accuracy of predictions for milk flavor was improved by increasing the number of samples, using K-nearest neighbors, and employing radial basis function kernel support vector machines and ridge regression models. When the number of prediction set models exceeded 40, the prediction accuracy for salty flavor increased with the increase in the number of prediction set samples; when the number of prediction set samples exceeded 70, the prediction accuracy for milk flavor increased with the increase in the number of prediction set samples; and for the prediction of overall preference, when the number of prediction set samples exceeded 40, the prediction accuracy of overall preference increased with the increase in the number of prediction set samples using radial basis function kernel support vector machines, gradient boosting machines, and extreme gradient boosting machines. Figure 6 ).

[0103] The above conclusions show that the method based on the fusion of two flavor data (GC-MS combined with electronic nose signal) is better at predicting sensory characteristics, and increasing the number of samples in the prediction set significantly increases the accuracy of the model prediction.

[0104] Table 1. Analysis of the accuracy and MAE value prediction of overall sensory value based on GC-MS volatile component data modeling (19 models) in Example 2.

[0105] Table 2. Accuracy and MAE value analysis of overall sensory value prediction based on GC-MS volatile component data and electronic nose E-Nose data (19 models) in Example 2.

[0106] Table 3. Accuracy analysis of predictions for milky, creamy, and salty flavors based on GC-MS volatile component data modeling (19 models) in Example 2.

[0107] Table 4. Accuracy analysis of predictions for milky, creamy, and salty-milky flavors based on GC-MS volatile component data fusion with electronic nose data (19 models) in Example 2.

[0108] Table 1

[0109]

[0110]

[0111] Table 2

[0112]

[0113] Table 3

[0114]

[0115]

[0116] Table 4

[0117]

[0118] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A rapid predictive method for outputting sensory values ​​based on milk flavor characteristics, characterized in that, Includes the following steps: A) The volatile flavor factors of the raw milk samples were determined by HS-SPME-GC-MS to obtain the volatile components of the raw milk samples; B) Heat the raw milk, perform sensory evaluation, and obtain sensory evaluation data; C) The raw milk was measured using an electronic nose to obtain the response value of the electronic nose sensor; D) Establish a sensory value prediction model based on the volatile components and sensory properties of raw milk samples, based on volatile flavor components, electronic nose, and sensory properties, and evaluate the model results; E) Based on the models built from the above two types of data, the number of sample prediction sets is gradually increased to optimize the model accuracy; F) Based on the above sensory value prediction model, introduce new milk samples and predict their sensory attributes based on their volatile flavor factor data or volatile flavor factor and sensor signal data.

2. The method according to claim 1, characterized in that, Step A) The method for determining volatile flavor factors includes: The raw milk was placed in a headspace sample vial. After heating in a 40°C water bath for 20 minutes, the SPME fiber was inserted into the headspace vial and equilibrated for 30 minutes. The headspace gas was collected and released, and sensor signal values ​​were acquired for 115–120 seconds.

3. The method according to claim 1, characterized in that, Step A) GC-MS parameters include: The gas chromatograph is model 8890GC System / 5977C GC / MSD; the column is a DB-WAX highly polar column; and the ion source is an EI ion source. The carrier gas was helium, with a flow rate of 1 mL / min; the standard electron energy of the EI ionization source was 70 eV; the mass spectrometer source was 230℃, the quadrupole source was 150℃; and the interface temperature was 250℃. Gradient temperature program: Hold at 30℃ for 1 min, then increase to 90℃ at 3℃ / min, hold for 0 min, and finally increase to 220℃ at 6℃ / min, hold for 5 min; The volatile aroma components were identified using the NIST 20 spectral library.

4. The method according to claim 1, characterized in that, Step C) includes: Signals were acquired using a PEN3 portable electronic sensor; the settling time for each sample was between 115 and 120 seconds, and the response value for each sensor was calculated as G / G0, where G and G0 represent the sensor’s response to the sample and air, respectively.

5. The method according to claim 1, characterized in that, The assessment described in step D) includes: Nineteen models were established based on the volatile components and sensory evaluation data of raw milk samples, respectively. Nineteen models were also established based on the volatile components of raw milk samples, combined with electronic nose data and sensory values.

6. The method according to claim 1, characterized in that, The evaluation in step D) includes: evaluating the regression results based on mean absolute error and accuracy.

7. The method according to claim 1, characterized in that, Step E) The model's samples include: increasing the dataset in increments of 10, from 30 samples to 80 samples, to train the model and improve the model's accuracy in predicting sensory information.

8. The method according to claim 1, characterized in that, The prediction model described in steps D and E) is a prediction model based on four sensory attribute values: milky flavor, creamy flavor, salty flavor, and overall preference.

9. The method according to claim 6, characterized in that, The model described in step D) includes: (19 machine learning models in total, including Random Forest, Bag-of-the-Domain (BOT) clustering, Adaptive Boosting, Gradient Boosting Machine, Support Vector Machine (Radial Basis Kernel), Extreme Gradient Boosting (XGBoost), Bayesian model (usually Naive Bayes classifier), Support Vector Machine (Multinomial Kernel), Lasso Regression (L1 Regularized Regression), Decision Tree, Support Vector Machine (Linear Kernel), Ridge Regression (L2 Regularized Regression), Elastic Network Regression (L1+L2 Regularized), Linear Regression, Neural Network, Artificial Neural Network, K-Nearest Neighbors, Logistic Regression, and Multilayer Perception. All machine learning models are built using Python algorithms.) 10. The method according to claim 1, characterized in that, The raw milk samples came from typical pastures in Northwest China.

Images