Baijiu flavor preference degree analysis method and perception evaluation system

Abstract

The invention provides a Baijiu flavor preference degree analysis method. The Baijiu flavor preference degree analysis method comprises the following steps: (1) collecting preference degree sensory evaluation data and electroencephalogram data after a testee stimulates different Baijiu flavors; (2) preprocessing the electroencephalogram data obtained in the step (1) to obtain different electroencephalogram segment data; (3) taking the different brain wave segment data obtained in the step (2) as input features, combining the preference degree sensory evaluation data collected in the step (1), performing XGBoost model training to obtain a trained XGBoost prediction model, and predicting the preference degree of the testee based on the XGBoost prediction model to obtain a predicted value X; and (4) calculating a preference degree result Y of the testee based on the predicted value X obtained in the step (3) in combination with ratio data between different band powers. According to the method, the XGBoost model is used for carrying out feature learning on the electroencephalogram data, so that the preference degree of common consumers to different white spirit flavors can be explored, and data support is provided for development of new white spirit products.

Description

Technical Field

[0001] The present application relates to the technical field of wine detection, and in particular to a liquor flavor preference analysis method and a perception evaluation system. Background Art

[0002] The flavor of wine affects consumers' preference for the corresponding wine. At present, the evaluation of wine flavor mainly includes manual sensory evaluation and instrumental evaluation. Among them, instrumental evaluation cannot evaluate people's feelings. The current manual sensory evaluation is mainly manual sensory evaluation by professionals, and the evaluation of professionals cannot represent the evaluation of ordinary consumers. In addition, the flavor characteristics of liquor are rich and complex, and it is difficult for ordinary consumers to evaluate certain characteristic aromas. Therefore, a method that can evaluate wine flavor more comprehensively is needed. Summary of the invention

[0003] In order to solve one of the above problems, on the one hand, the present application provides a liquor flavor preference analysis method to improve the problem of low adaptability of liquor flavor preference analysis in the prior art, including the following technical solutions:

[0004] (1) Collect sensory evaluation data and EEG data of subjects’ preference for different liquor flavors;

[0005] (2) preprocessing the EEG data obtained in step (1) to obtain data of different EEG wave bands;

[0006] (3) Using the different EEG band data obtained in step (2) as input features, combined with the preference sensory evaluation data collected in step (1), XGBoost model training is performed to obtain a trained XGBoost prediction model, and the preference level of the subject is predicted based on the XGBoost prediction model to obtain a prediction value X;

[0007] (4) Based on the predicted value X obtained in step (3) and the ratio data between the powers of different bands, the subject's preference result Y is calculated.

[0008] In some embodiments, in step (1), the evaluation includes: using a digital scent player to provide aroma stimulation of different liquor aromas.

[0009] In some embodiments, in step (1), the method for obtaining the sensory evaluation data of preference includes: after each smell is played, a scoring method of 1-9 points is used, and the larger the score, the higher the preference.

[0010] In some embodiments, in step (1), subjects are selected based on the age characteristics of the liquor consumption group.

[0011] In some embodiments, in step (1), the subjects are aged between 18 and 65 years old, with a male to female ratio of 1:1, and an age distribution of 18-27 years old: 28-40 years old: 41-65 years old of 1:3:1.

[0012] In some embodiments, in step (1), the evaluation includes: different liquor aromas are divided into multiple groups for stimulation, each group is provided with multiple parallel aromas, and an interval is set between every two groups of flavor tests.

[0013] In some embodiments, in step (1), the stimulation method is: placing the olfactory tube of the digital scent player 3-5 cm in front of the subject's nose, first without stimulation, then playing the fragrance, resting for a while, and repeating.

[0014] In some embodiments, in step (1), the stimulation method is: place the olfactory tube of the digital odor player 3-5 cm in front of the subject's nose, first without stimulation for 10 seconds, then play the aroma for 20 seconds, rest for 5 minutes, and repeat 5 times.

[0015] In some embodiments, in step (2), the preprocessing includes: detrending, bandpass filtering and removing power frequency interference from the EEG data obtained in step (2); designing the bandwidth, attenuation coefficient and filter order of the digital filter according to the frequency range of different EEG bands, filtering the EEG signal after the power frequency interference removal with the corresponding digital filter, extracting the corresponding frequency band signal characteristics, and obtaining different EEG band data.

[0016] In some embodiments, in step (2), the preprocessing includes: detrending, bandpass filtering and removing power frequency interference on the EEG data obtained in step (2); constructing a digital filter according to a specific EEG rhythm; designing the bandwidth, attenuation coefficient and filter order of the digital filter according to the cutoff frequency of the frequency band; and filtering the preprocessed EEG signal with digital filters of different frequency bands to obtain different EEG rhythms.

[0017] In some embodiments, in step (3), the preprocessing includes: in step (3), the EEG data obtained in step (2) are preprocessed as follows: first, the original EEG data is detrended to remove the linear trend or baseline drift in the data to eliminate possible low-frequency interference; then, the detrended data is band-pass filtered, the frequency band range is set to 0.5-80Hz, meaningful EEG activity is retained, and high-frequency noise and extremely low-frequency components are removed; then, power frequency interference is removed to eliminate the influence of power frequency on the EEG signal, and the main loop is removed. Next, according to the frequency range of different EEG bands, where the frequency range of δ wave is 0.5-4Hz, θ wave is 4-8Hz, α wave is 8-12Hz, β wave is 12-30Hz, and γ wave is 30-100Hz, the bandwidth, attenuation coefficient and filter order of the digital filter are designed to construct digital filters suitable for each frequency band; finally, these filters are used to filter the preprocessed EEG signals, extract the corresponding frequency band signal features, and obtain different EEG band data to provide input for subsequent feature extraction and model training.

[0018] In some embodiments, in step (4), the calculation of the subject's preference result Y based on the predicted value X obtained in step (3) combined with the ratio data between the powers of different bands includes: adding the predicted value X and the corresponding ratio between the powers of different bands according to the corresponding weights to obtain the preference result Y.

[0019] In some embodiments, the ratios between the corresponding powers of different wavebands include at least one of a ratio of delta wave power to beta wave power, a ratio of theta wave power to beta wave power, and a ratio of alpha wave power to beta wave power.

[0020] In some embodiments, the preference result Y is calculated by the following formula:

[0021] Y=X×w1+DBR×w2+TBR×w3+ABR×w4,where w1=0.8~0.9, w2=0.01~0.04, w3=0.005~0.02, w4=0.05~0.3, P δ , P β , P α , P θ They are the power response values ​​of delta wave, beta wave, alpha wave and theta wave respectively.

[0022] In some embodiments, w1=0.85, w2=0.02, w3=0.01, w4=0.12.

[0023] In some embodiments, in step (4), the objective function of the XGBoost model consists of a loss function and a regularization term, and the loss function is the sum of the squares of the differences between the predicted value and the true value.

[0024] In some embodiments, in step (1), the flavors of the different liquors include at least one dimension of grassy aroma, fruity aroma, floral aroma, sweet aroma, nutty aroma, dry plant aroma, sour aroma, and other aromas.

[0025] In some embodiments, in step (1), the flavors of the different liquors include eight dimensions: grassy aroma, fruity aroma, floral aroma, sweet aroma, nutty aroma, dried plant aroma, sour aroma, and other aromas; the grassy aroma includes: grassy aroma, raw grain aroma, and bamboo aroma; the fruity aroma includes: apple aroma, pineapple, and banana; the floral aroma includes: rose, gardenia, and violet; the sweet aroma includes: vanilla aroma and toffee; the nutty aroma includes: cocoa aroma and paste aroma; the dried plant aroma includes: Chinese medicine aroma; the sour aroma includes: vinegar aroma; the other aromas include: mushroom aroma, cellar aroma, sulfur aroma, yogurt aroma, and oil aroma.

[0026] In some embodiments, the different EEG wave band data include at least one of power ratio data and signal complexity data corresponding to different wave bands.

[0027] In some embodiments, the power ratio data corresponding to different bands include at least one of: the ratio of the θ band power to the β band power, the ratio of the α band power to the β band power, the ratio of the δ band power to the β band power, and the ratio of the inverse of the α band power to the sum of the α band and θ band powers.

[0028] In some embodiments, the signal complexity includes: C0 complexity, Hjorth activity parameter, and cross-frequency phase-amplitude coupling.

[0029] In some embodiments, the calculation method of the C0 complexity includes: converting the EEG signal to the frequency domain through fast Fourier transform, calculating the average value of the frequency domain amplitude, screening out the components with amplitudes greater than the average value and then performing inverse Fourier transform to obtain a filtered time domain signal, calculating the sum of the absolute differences between the filtered time domain signal and the original time domain signal at each time point, and then dividing this sum by the sum of the absolute values ​​of all points of the original signal.

[0030] In some embodiments, the calculation method of the C0 complexity includes: FFT calculation: performing a fast Fourier transform on the signal to obtain a signal fft_data represented in the frequency domain; then calculating the average value fft_average of its frequency domain amplitude; screening high amplitude components: screening out components with amplitudes greater than the average value fft_average from fft_data, setting other components to 0, and forming a new frequency domain signal r fft; Inverse FFT: for the filtered frequency domain signal r fft Perform inverse Fourier transform to obtain the filtered time domain signal ifft_data, and calculate the C0 complexity by the following formula:

[0031]

[0032] Where: x i represents the original signal of the i-th sampling point; y i represents the filtered signal (i.e., ifft_data) of the i-th sampling point; N is the number of data points.

[0033] In some embodiments, the Hjorth activity parameter is obtained by calculating the ratio of the variance of the EEG signal to the square of the mean to analyze the change in the amplitude of the signal in the time domain.

[0034] In some embodiments, the calculation formula of the Hjorth activity parameter is:

[0035]

[0036] N is the number of data points; x i represents the EEG signal to be analyzed at the i-th data point; is the mean of the signal, and Activity is the Hjorth activity parameter.

[0037] In some embodiments, the cross-frequency phase-amplitude coupling is obtained by applying Hilbert transform to the signal in the γ frequency band to obtain instantaneous power, applying Hilbert transform to the θ signal to obtain the phase, and then calculating the phase of the γ power, and obtaining the synchronization index SI represented by the average of the complex exponentials of the phase difference.

[0038] In some embodiments, the calculation of the cross-frequency phase-amplitude coupling includes the following method: applying Hilbert transform to the signal in the γ frequency band to obtain envelope information in complex form, calculating the square of the envelope to obtain the instantaneous power in the γ frequency band; applying Hilbert transform to the θ signal to obtain the phase θ phase , apply the Hilbert transform to the γ power and obtain its phase γ power phase ; The cross-frequency phase-amplitude coupling is the synchronization index SI represented by the average of the complex exponential of the phase difference, and the calculation formula is as follows:

[0039] PAC=|SI|

[0040]

[0041] Where PAC is the cross-frequency phase-amplitude coupling, SI is the synchronization index; θ phase is the phase of the θ frequency band, γpower phase is the power phase in the γ band.

[0042] In some embodiments, the different EEG band data obtained in step (2) are used as input features to perform XGBoost model training to obtain a trained XGBoost prediction model. The XGBoost model training includes: respectively calculating the EEG signal features corresponding to no odor stimulation and odor stimulation and inputting them into the XGBoost model, setting an initial prediction value based on the preference sensory evaluation data, and gradually reducing the error between the initial prediction value and the actual target value through an iterative method until the model performance reaches a predetermined condition and then stopping the training to obtain the XGBoost prediction model.

[0043] In some embodiments, the error between the initial prediction value and the actual target value is gradually reduced by iteration until the model performance reaches a predetermined condition and then the training is stopped. This includes: gradually reducing the error between the initial prediction value and the actual target value by iteratively adding decision trees, and the prediction results of each tree are added to the prediction value of the previous round to form a new prediction value, which is used as the basis for the next round of iterations, and the training is stopped after continuous iteration until the model performance reaches a predetermined condition.

[0044] On the other hand, the present application also provides a liquor flavor perception evaluation system, which performs liquor flavor perception evaluation based on the aforementioned liquor flavor preference analysis method.

[0045] In some embodiments, the liquor flavor perception evaluation system includes:

[0046] Liquor flavor perception evaluation module: The liquor flavor perception evaluation module is used to obtain the sensory evaluation data of the subjects on different liquor flavors, and obtain the EEG signal data of the subjects in the process of sensory evaluation of different liquor flavors;

[0047] EEG signal data processing module: The EEG signal data processing module analyzes and processes the EGG signal data and sensory evaluation data collected and transmitted by the liquor flavor perception evaluation module to obtain the results of the subjects' preference for different liquor flavors;

[0048] Preference result output module: The preference result output module outputs the preference result of the subject obtained by the EEG signal data processing module.

[0049] The above-mentioned liquor flavor preference analysis method collects EEG data and preference sensory evaluation data after stimulation of liquor of different aroma types, and pre-processes the obtained EEG data, uses the XGBoost model to perform feature learning on the pre-processed EEG data to predict the preference, takes the preference sensory evaluation data as a reference standard, and optimizes the predicted preference with the band response value, so that the obtained final preference prediction value Y is closer to the actual value. Since the present application mainly uses EEG data for preference analysis, since the time series data recorded in each channel of EEG data can contain a large number of sampling points, EEG data has the characteristics of high dimension, and XGBoost uses parallel computing technology, which is suitable for processing the high-dimensional characteristics of EEG data. In addition, there are complex nonlinear relationships in the features of EEG signals. XGBoost can capture these nonlinear relationships through the combination of gradient boosting and decision trees, and can help identify which EEG features are most critical to predicting the preference through the feature importance evaluation mechanism of XGBoost; since the consumer's preference is usually continuous rather than binary or multivariate classification, the XGBoost model can predict continuous preference scores rather than simple classification results. This means that it can more finely reflect the user's actual preferences and more accurately predict the subjects' true preferences. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of the structure of a liquor flavor perception and evaluation system in one embodiment;

[0051] Figure 2 is a schematic diagram of the structure of a digital scent player in one embodiment;

[0052] Figure 3 The figure is a flow chart of scent playing of a digital scent player in one embodiment. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is further described below by specific embodiments, which do not limit the protection scope of the present invention. Some non-essential modifications and adjustments made by others based on the concept of the present invention still fall within the protection scope of the present invention.

[0054] Example 1 A method for analyzing liquor flavor preference

[0055] 1. Collect sensory evaluation data and EEG data of subjects’ preference for different types of liquor after being stimulated by the aroma:

[0056] (1) Subjects

[0057] There were 30 participants in the experiment, with a male to female ratio of 1:1; their ages ranged from 18 to 65 years old. Taking into account the liquor consumption group, the population ratio was set to 18-27 years old: 28-40 years old: 41-65 years old, which was 1:3:1.

[0058] (2) The flavor types of liquors with different aroma types evaluated are shown in Table 1:

[0059] Table 1: Liquor flavor types

[0060]

[0061]

[0062] (3) Test process

[0063] use Figure 2 The digital smell player shown is used to conduct an evaluation test on different types of liquor flavors on subjects, wherein the digital smell player comprises: a controller, an air valve, and an olfactory tube; the controller is electrically connected to the air pump and the air valve, and can control the switching of the air pump and the air valve; the air pump is connected to a plurality of air valves, each of which controls a different smell; when the air valve is opened, the corresponding smell controlled by the air valve flows into the olfactory tube.

[0064] The specific test process is as follows:

[0065] 1) Preparation before the experiment: Before the experiment, the subjects were asked not to drink alcohol for eight hours and not to eat for two hours. The laboratory was ventilated by opening windows and exhaust fans, and the central air conditioner was turned on to maintain the laboratory ambient temperature at 21±2℃.

[0066] 2) During the experiment, the subjects were asked to relax, breathe evenly, and follow the prompts on the software interface.

[0067] 3) Fill in the subject's information, including gender, age, and drinking habits (frequently consumed brands and drinking frequency).

[0068] 4) Wear the patch correctly with the assistance of the staff; during the test, the EEG data will be transmitted to the EEG analysis processor of the device in real time. Make sure that the subject always wears the patch correctly during the test.

[0069] 5) Place the olfactory tube of the digital odor player 3-5 cm in front of the subject's nose and start the test.

[0070] 6) For the 20 odors shown in Table 1, each stimulus of each odor is Figure 3The process shown in the figure is played, and the experiment begins. The test is performed for 10 seconds without odor and 20 seconds with odor, followed by a 5-minute rest. After each odor is played, the subject scores the sensory preference on the software interface, using a 1-9 score method. The larger the score, the higher the preference. The score obtained is used as the sensory evaluation data of the preference. Then click to play the next odor and repeat the above odor playback process. The interval between the two groups of odor tests is 5 minutes, which is designed to allow the subjects to fully rest and avoid olfactory fatigue. During the rest period, open the door and fan to increase air circulation. Expel the residual odor in the olfactory induction instrument to prevent the odor of subsequent experiments from being contaminated.

[0071] 7) After all eight aroma dimensions have been tested, the test is terminated.

[0072] Through the above method, we obtained the sensory evaluation data and EEG data of 30 subjects' preference for 20 odor stimuli in eight dimensions, including grassy aroma, fruity aroma, floral aroma, sweet aroma, nutty aroma, dry plant aroma, sour aroma, and other aromas.

[0073] 2. Data preprocessing:

[0074] According to the complete EEG data collected in step 1, EEG bio-indicators with obvious dominant effects are designed based on the EEG analysis model, and EEG segments with and without odor stimulation are extracted for preprocessing. The specific steps are as follows:

[0075] First, the original EEG signal is detrended to remove the linear trend or baseline drift in the data to eliminate possible low-frequency interference;

[0076] Next, the detrended data were band-pass filtered with a frequency band range of 0.5–80 Hz to retain meaningful EEG activity and remove high-frequency noise and extremely low-frequency components;

[0077] Then, the power frequency interference is removed to eliminate the influence of the power frequency on the EEG signal, and the EEG data with the main environmental noise removed is obtained;

[0078] Next, according to the frequency ranges of different specific EEG wave bands, Delta (δ) wave: 0.5-4Hz, Theta (θ) wave: 4-8Hz, Alpha (α) wave: 8-12Hz, Beta (β) wave: 12-30Hz, Gamma (γ) wave: 30-100Hz), the bandwidth, attenuation coefficient and filter order of the digital filter are designed to construct a digital filter suitable for each frequency band;

[0079] Finally, these filters are used to filter the preprocessed EEG signals, extract the corresponding frequency band signal features, and obtain different EEG band data to provide input for subsequent feature extraction and model training.

[0080] 3. Construction of Preference Model

[0081] The following EEG signal features were input into the XGBoost model:

[0082] 1. Ratio of power of different EEG rhythms

[0083] 1) TBR: the ratio of theta wave to beta wave power;

[0084] 2) ABR: ratio of alpha wave to beta wave power;

[0085] 3) DBR: the ratio of delta wave to beta wave power;

[0086] 4) ARR: the inverse of alpha wave power;

[0087] 5) BATR: the ratio of the beta wave to the sum of the alpha and theta wave powers.

[0088] 2. Signal complexity

[0089] 1) C0: C0 complexity of the signal. C0 complexity is used to describe the proportion of irregular components in the EEG signal. Its calculation method is:

[0090] FFT calculation: Perform fast Fourier transform (FFT) on the signal to obtain the frequency domain signal fft_data. Then calculate the average value of its frequency domain amplitude fft_average;

[0091] Filter high-amplitude components: Filter out components with amplitudes greater than the average value fft_average from fft_data, set other components to 0, and form a new frequency domain signal r fft

[0092] Inverse FFT: For the filtered frequency domain signal r fft Perform an inverse Fourier transform (IFFT) to obtain the filtered time domain signal ifft_data

[0093] The C0 complexity is calculated by the following formula:

[0094]

[0095] x i represents the original signal of the i-th sampling point; y i represents the filtered signal (i.e., ifft_data) of the i-th sampling point; N is the number of data points.

[0096] 2) Hjorth activity parameter (Activity): It is usually used to analyze the change of the amplitude of EEG signals in the time domain; its calculation method is:

[0097]

[0098] N is the number of data points; x i represents the EEG signal to be analyzed at the i-th data point; is the mean of the signal,

[0099] 3) PAC: Extract the cross-frequency phase-amplitude coupling (PAC) within the specified signal. PAC is used to analyze the interaction between different frequency components. The calculation method is as follows: Apply Hilbert transform to the signal in the γ band to obtain the envelope information in complex form, calculate the square of the envelope to obtain the instantaneous power in the γ band; Apply Hilbert transform to the θ signal to obtain the phase θ phase , apply the Hilbert transform to the γ power and obtain its phase γ power phase The PAC value is the synchronization index (SI) expressed by the average of the complex exponential of the phase difference, and the calculation formula is as follows:

[0100] PAC=|SI|

[0101]

[0102] The EEG signal features of multiple samples are input into the XGBoost model for training. Based on the XGBoost (Extreme Gradient Boosting) algorithm, the EEG signal features are input. XGBoost can effectively capture the complex nonlinear relationship between these features and the degree of preference, and predict the degree of preference.

[0103] The objective function of XGBoost consists of a loss function and a regularization term. The objective function of XGBoost is:

[0104]

[0105] The objective function L(θ) will be the loss function l(y i ,f(x i )) and add the regularization term Ω(f) to minimize the overall loss. Where i represents each sample in the data set, l(y i ,f(x i )) is the loss function, which represents the predicted value and the true value y i The error between the predicted values ​​of the model f(x i ) refers to the model's preference for the prediction of the i-th sample during the training process, and the true value y iis the sensory evaluation data of the subjects’ preference for the i-th sample, input x i It is an EEG signal feature constructed by using the power and relative power of the frequency band and the complexity of the signal. i The predicted value is obtained through the model Ω(f) is a regularization term used to control model complexity and prevent overfitting.

[0106] The mean squared error (MSE) is used as the loss function. MSE optimizes the model by minimizing the error between the predicted value and the actual preference level, which measures the accuracy of the model prediction and helps to adjust the model parameters to improve the prediction performance. The squared error is the predicted value. and the true value y i The sum of the squares of the differences is as follows:

[0107]

[0108] XGBoost uses a second-order Taylor expansion to approximate the objective function to update the model using gradient and Hessian information. For a given loss function We perform a second-order Taylor expansion on the loss function of each sample point i:

[0109]

[0110] Among them, f m (x i ) is the model output value of the mth round (i.e., the output of the current weak learner), g i and h i They are respectively the loss function and the predicted value The first order (Gradient) and second order derivative (Hessian) of . The gradient is used to guide the update direction of the model parameters, and the Hessian is used to determine the step size and amplitude of the model parameter update. The definition is as follows:

[0111] Gradient is the loss function with respect to the predicted value. The first-order derivative of is used to guide the update direction of the model parameters. The gradient calculation formula of the squared error loss function is as follows:

[0112]

[0113] Hessian is the loss function for the predicted value The second-order derivative of is used to determine the step size and amplitude of the model parameter update. The Hessian calculation formula of the square error loss function is as follows:

[0114]

[0115] The levels of these indexes express a degree, and the degree division is based on the collection and manual annotation of the user's EEG signal to train the model. The training process is:

[0116] Remove irrelevant data before odor stimulation, and calculate the EEG signal features corresponding to no odor stimulation and odor stimulation respectively; input the above EEG signal features into the XGBoost model, and the model processes the above EEG signal features as follows:

[0117] 1) Initial prediction value setting: At the beginning of training, the XGBoost model sets an initial prediction value based on the sensory evaluation data of preference, which is usually the mean of the sample or other simple statistical values. This initial prediction value is only used as a baseline prediction and has no direct relationship with the EEG signal input features. It reflects the basic prediction of the model when no signal features have been learned.

[0118] 2) Error fitting and adjustment: Next, the model gradually introduces decision trees to reduce the gap between the initial prediction and the actual target value (sensory evaluation data of preference). The model constructs a series of decision trees, each of which attempts to fit the error between the current prediction value and the actual target value.

[0119] 3) Accumulated prediction values: When a new decision tree is established, it will output a prediction value for each sample. The prediction results of each decision tree will be gradually accumulated to the initial prediction value. The updated prediction value will be used as the baseline prediction for the next round to continuously reduce the error between the prediction value and the target value.

[0120] 4) Training stop condition: Training is stopped when the model meets one of the following conditions: When the latest prediction value meets the condition that the number of iterations reaches the preset maximum number of rounds (the default is 100 times) or the reduction of the training error or the verification error is lower than the preset threshold (1e-4), training is stopped and the model preference prediction value X is output.

[0121] In the above training process, respectively calculating the EEG signal features corresponding to the absence of odor stimulation and the presence of odor stimulation means taking the state without odor stimulation as a benchmark, and using the model to learn the relationship between the EEG signal features corresponding to the absence of odor stimulation and the presence of odor stimulation. Through this processing process, the processed data mainly reflects the state changes before and after the stimulation, avoiding the influence of individual differences caused by different subjects' emotions and attention during the test.

[0122] 4. Optimization of the Prediction Value of Preference

[0123] Based on the GBoost model trained in step 3 above, the model preference prediction value X is obtained. Based on the obtained model preference prediction value X, the final prediction result is further optimized. During the optimization process, multi-dimensional features are integrated. In addition to the initial value output by the model, the ratio of δ wave, θ wave, α wave, and β wave is also adjusted. The optimization method is as follows:

[0124] 1) Calculate the power ratio of each band and map the data to [1, 5]:

[0125]

[0126] Among them, P δ Indicates the power response value of the delta wave, P β Indicates the power response value of β wave, P θ Represents the power response value of theta wave, P β Indicates the power response value of β wave, P α represents the power response value of α wave,

[0127] 2) Adjust the predicted value of preference according to the following formula:

[0128] Y=X×w1+DBR×w2+TBR×w3+ABR×w4

[0129] X is the model preference prediction value, and Y is the final preference prediction value. Finally, based on the real data test, adjustment and optimization, we get: w1=0.85, w2=0.02, w3=0.01, w4=0.12.

[0130] 5. Validation of the Prediction and Analysis Methods of Preference

[0131] Taking the EGG signal features of the subjects corresponding to the flavor of liquor as the input features, the consumer's preference is predicted based on the XGBoost model trained in step three and the prediction formula of the preference prediction value optimized in step four, and compared with the preference sensory evaluation data as the true value of the preference. The specific results are shown in the table below. The results show that the error between the predicted value and the true value of all aroma preference is less than 10%, indicating that this method is accurate and applicable.

[0132]

[0133]

[0134] As can be seen above, the liquor flavor preference analysis method provided by the present invention can better predict consumers' preferences for different liquor flavors. Therefore, the analysis results obtained by this method can provide data support for the development of new liquor products. At the same time, for new products or samples that have been developed, the analysis method of this method can be used as a method for predicting preference, providing data support for the marketing of subsequent products.

[0135] In summary: the present invention collects EEG data and sensory evaluation data of preference after different types of liquor aroma stimulation, preprocesses the obtained EEG data, uses the XGBoost model to perform feature learning on the preprocessed EEG data to predict the preference level, takes the sensory evaluation data of preference as the reference standard, and optimizes the predicted preference level by the ratio of the band response values, so that the final preference level prediction value Y is closer to the actual value. The ratio of different EEG rhythm power and signal complexity reflect the activity characteristics of the brain in different states. By inputting these characteristics into the preference model, the impact of odor stimulation on the subjects can be more comprehensively understood, providing stronger explanatory power. Combining the scores of multiple indicators can reduce the impact of individual differences, environmental factors or other noises; making the EEG response analysis to odor stimulation more comprehensive and in-depth.

[0136] Since this application mainly uses EEG data for preference analysis, and since the time series data recorded in each channel of EEG data can contain a large number of sampling points, EEG data has the characteristics of high dimensionality, and XGBoost uses parallel computing technology, which is suitable for processing the high dimensional characteristics of EEG data. In addition, there are complex nonlinear relationships between the features in EEG signals. XGBoost can capture these nonlinear relationships through the combination of gradient boosting and decision trees, and can help identify which EEG features are most critical for predicting preference through XGBoost's feature importance evaluation mechanism; since consumers' preferences are usually continuous rather than binary or multivariate classifications, the XGBoost model can predict continuous preference scores rather than simple classification results. This means that it can more finely reflect the user's actual preferences and help to more accurately predict the subjects' true preferences.

[0137] At the same time, using a comprehensive sensitivity score that takes into account the ratios of multiple EEG frequency bands and their changes as one of the model input features can enable the model's prediction value to more comprehensively evaluate the impact of odor on brain activity.

[0138] Example 2 A liquor flavor perception evaluation system

[0139] This embodiment provides a liquor flavor perception evaluation system based on a liquor flavor preference analysis method provided in Example 1. Figure 1, the evaluation system includes:

[0140] Liquor flavor perception evaluation module: This module includes performing sensory evaluation tests on subjects of different liquor flavors, collecting EEG signal data of subjects during the sensory evaluation test of different liquor flavors, and transmitting the collected EEG signal data to the EEG signal data module; wherein, based on the method in Example 1, sensory evaluation tests on subjects of different liquor flavors are performed, and when a digital odor player is used to perform sensory evaluation tests on subjects of different liquor flavors, the subjects need to wear a brain patch, and the EEG signal data generated during the sensory evaluation test of different liquor flavors performed by the subjects based on the digital odor player is collected through the brain patch, and the collected EEG signal data is transmitted to the EEG analysis processor;

[0141] EEG signal data processing module: using the EEG analysis processor to analyze and process the EGG signal data collected and transmitted by the liquor flavor perception evaluation module, to obtain the results of the subjects' preference for different liquor flavors, and to transmit the results of the subjects' preference obtained by the analysis and processing to the result output module;

[0142] Preference result output module: uses the software display interface to output the subject's preference result obtained by the EEG signal data processing module.

[0143] The above description is a detailed description of the preferred feasible embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. All equivalent changes or modified changes completed under the technical spirit disclosed by the present invention should fall within the patent scope covered by the present invention.

Claims

1. A method for analyzing liquor flavor preference, characterized in that: include: (1) Collect sensory evaluation data and EEG data of subjects’ preference for different liquor flavors; (2) preprocessing the EEG data obtained in step (1) to obtain data of different EEG wave bands; (3) Using the different EEG band data obtained in step (2) as input features, combined with the preference sensory evaluation data collected in step (1), XGBoost model training is performed to obtain a trained XGBoost prediction model, and the preference level of the subject is predicted based on the XGBoost prediction model to obtain a prediction value X; (4) Based on the predicted value X obtained in step (3) and the ratio data between the powers of different bands, the subject's preference result Y is calculated.

2. The method according to claim 1, characterized in that In step (2), the preprocessing includes: detrending, bandpass filtering and removing power frequency interference from the EEG data obtained in step (2); designing the bandwidth, attenuation coefficient and filter order of the digital filter according to the frequency range of different EEG bands, filtering the EEG signal after the power frequency interference removal with the corresponding digital filter, extracting the corresponding frequency band signal characteristics, and obtaining different EEG band data.

3. The method according to claim 1, characterized in that In the step (4), the calculation of the subject's preference result Y based on the predicted value X obtained in the step (3) combined with the ratio data between the powers of different bands includes: adding the predicted value X and the corresponding ratios between the powers of different bands according to corresponding weights to obtain the preference result Y.

4. The method according to claim 3, characterized in that The corresponding ratios between powers of different wavebands include at least one of a ratio of δ wave to β wave power, a ratio of θ wave to β wave power, and a ratio of α wave to β wave power.

5. The method according to claim 1, characterized in that The preference result Y is calculated by the following formula: Y=X×w1+DBR×w2+TBR×w3+ABR×w4,where w1=0.8~0.9, w2=0.01~0.04, w3=0.005~0.02, w4=0.05~0.3, P δ , P β , P α , P θ They are the power response values ​​of delta wave, beta wave, alpha wave, and theta wave respectively; Preferably, w1=0.85, w2=0.02, w3=0.01, w4=0.

12.

6. The method according to claim 1, characterized in that In step (1), the flavors of the different liquors include at least one dimension of grassy aroma, fruity aroma, floral aroma, sweet aroma, nutty aroma, dry plant aroma, sour aroma and other aromas; Preferably, the grassy aroma includes: at least one of grassy aroma, raw grain aroma, and bamboo aroma; the fruity aroma includes: at least one of apple aroma, pineapple, and banana aroma; the floral aroma includes: rose, gardenia, and violet; the sweet aroma includes: at least one of vanilla aroma and toffee aroma; the nutty aroma includes: cocoa aroma and paste aroma; the dry plant aroma includes: at least one of traditional Chinese medicine aroma; the sour aroma is vinegar aroma; the other aromas include: at least one of mushroom aroma, cellar aroma, sulfur aroma, yogurt aroma, and oil aroma.

7. The method according to claim 1, characterized in that In step (3), the different EEG wave band data include at least one of power ratio data and signal complexity data corresponding to different wave bands; Preferably, the power ratio data corresponding to different bands include at least one of: the ratio of the θ band power to the β band power, the ratio of the α band power to the β band power, the ratio of the δ band power to the β band power, and the ratio of the inverse of the α band power to the sum of the α band and θ band powers.

8. The method according to claim 7, characterized in that In step (3), the signal complexity data includes at least one of: C0 complexity, Hjorth activity parameter, and cross-frequency phase-amplitude coupling; Preferably, the C0 complexity is to convert the EEG signal into the frequency domain by fast Fourier transform, calculate the average value of the frequency domain amplitude, screen out the components with amplitude greater than the average value, and then perform inverse Fourier transform to obtain a filtered time domain signal, calculate the sum of the absolute difference between the filtered time domain signal and the original time domain signal at each time point, and then divide this sum by the sum of the absolute values ​​of all points of the original signal; The Hjorth activity parameter is obtained by calculating the ratio of the variance of the EEG signal to the square of the mean to analyze the change in the amplitude of the signal in the time domain; The cross-frequency phase-amplitude coupling is obtained by applying Hilbert transform to the signal in the γ frequency band to obtain instantaneous power, applying Hilbert transform to the θ signal to obtain phase, and then calculating the phase of the γ power, and obtaining the synchronization index represented by the average of the complex exponentials of the phase difference.

9. The method according to claim 1, characterized in that: The different EEG band data obtained in step (2) are used as input features to perform XGBoost model training to obtain a trained XGBoost prediction model, wherein the XGBoost model training comprises: respectively calculating the EEG signal features corresponding to no odor stimulation and odor stimulation and inputting them into the XGBoost model, setting an initial prediction value based on the sensory evaluation data of preference, and gradually reducing the error between the initial prediction value and the actual target value in an iterative manner until the model performance reaches a predetermined condition and then stopping the training to obtain the XGBoost prediction model; Preferably, the step of gradually reducing the error between the initial prediction value and the actual target value by iteration until the model performance reaches a predetermined condition and then stopping the training includes: gradually reducing the error between the initial prediction value and the actual target value by iteratively adding decision trees, accumulating the prediction result of each tree to the prediction value of the previous round to form a new prediction value, which is used as the basis for the next round of iteration, and continuously iterating until the model performance reaches a predetermined condition and then stopping the training.

10. A liquor flavor perception evaluation system, characterized in that: The evaluation system performs a perceptual evaluation of liquor flavor based on the method described in any one of claims 1 to 9; Preferably, the evaluation system comprises: Liquor flavor perception evaluation module: The liquor flavor perception evaluation module is used to obtain the sensory evaluation data of the subjects on different liquor flavors, and obtain the EEG signal data of the subjects in the process of sensory evaluation of different liquor flavors; EEG signal data processing module: The EEG signal data processing module analyzes and processes the EGG signal data and sensory evaluation data collected and transmitted by the liquor flavor perception evaluation module to obtain the results of the subjects' preference for different liquor flavors; Preference result output module: The preference result output module outputs the preference result of the subject obtained by the EEG signal data processing module.

Images