A taste classification method based on electroencephalogram data

Abstract

The present application discloses a taste classification method based on electroencephalogram (EEG) data, which relates to the field of taste classification. The method comprises collecting EEG signals of a subject when receiving different basic taste stimuli; the basic tastes include sour, sweet, bitter, salty and umami; constructing a full-feature data set based on the EEG signals; determining an optimal basic classifier based on the full-feature data set, and screening electrode channels and extracting an optimal feature subset based on the optimal basic classifier; constructing a DE-Stacking ensemble model using the optimal basic classifier as a meta-model and multiple machine learning models as base models, and optimizing overall hyperparameters using a differential evolution algorithm; the multiple machine learning models include a support vector machine, a K-nearest neighbor model and a naive Bayes model; training the DE-Stacking ensemble model using the optimal feature subset, and determining the basic taste corresponding to the EEG signal to be tested based on the trained DE-Stacking ensemble model. The present application improves the accuracy of taste classification.

Description

Technical Field

[0001] The present application relates to the field of taste classification, and in particular to a taste classification method based on electroencephalogram (EEG) data. Background Art

[0002] The human brain processes a variety of sensory information, enabling individuals to perceive and respond to external stimuli. Among these sensory experiences, taste, as one of the most important, not only helps the body select and identify foods and protects against harmful substances, but also enhances the pleasure of eating and improves well-being. This has profound implications for medicine, technology, and many other areas of daily life.

[0003] Sour, sweet, bitter, salty, and umami are the five basic tastes. Sourness promotes food digestion and nutrient absorption, and also enhances taste sensitivity. Sweetness is often associated with energy-rich nutrients and brings a pleasant experience. Bitterness is generally considered unpleasant and serves as a natural defense mechanism against potentially harmful substances. Saltiness is essential for maintaining electrolyte balance and is key to transmitting nerve signals. Umami enhances the palatability of food. Given the important role of taste perception in human functioning, research on taste recognition is of great significance. However, the subjectivity of taste is influenced by individual genetics, physiology, and even psychological state, making objective analysis complicated. Furthermore, the neural mechanisms underlying taste perception are complex, involving intricate interactions between taste receptors, neural pathways, and brain regions.

[0004] Current methods for detecting various tastes, such as manual evaluation and electronic tongues, suffer from various drawbacks. Manual evaluation is highly subjective, while electronic tongues fail to capture the signal transmission and conversion between the brain and taste perception. Therefore, using brain imaging technology to identify tastes and explore the relationship between the brain and taste perception is of great significance.

[0005] Electroencephalograms (EEGs) can capture millisecond-level changes in brain electrical activity. They are also portable and low-cost. By monitoring electrical signals across various brain regions in real time, they objectively reflect a subject's actual feelings and have been widely used in various sensory stimulation research fields. By extracting brain signals generated by different taste stimuli through EEG, the relationship between the brain and taste perception can be effectively explored and analyzed.

[0006] Current research on taste EEG classification and recognition often suffers from low recognition accuracy and extensive redundant computation. EEG signals are typically acquired using multi-channel equipment, such as 32-, 64-, or 128-channel electrodes. However, not all electrodes provide useful classification information. Some electrodes may be irrelevant to taste or contain redundant information, resulting in low taste classification accuracy. Summary of the Invention

[0007] The purpose of this application is to provide a taste classification method based on EEG data to solve the problem of low taste classification accuracy.

[0008] To achieve the above objectives, this application provides the following solutions.

[0009] In a first aspect, the present application provides a taste classification method based on EEG data, comprising the following steps.

[0010] The EEG signals of the subjects were collected when they were stimulated with different basic tastes, including sour, sweet, bitter, salty and umami.

[0011] Based on the EEG signal, a full feature data set is constructed; the full feature data set includes the features of the preprocessed EEG signal; the features include linear features, nonlinear features and functional connectivity features; the linear features include mean, extreme value, variance and power spectral density; the nonlinear features include fuzzy entropy; the functional connectivity features include phase lag index.

[0012] Based on the full feature data set, an optimal basic classifier is determined, and according to the optimal basic classifier, electrode channels are screened to extract an optimal feature subset.

[0013] A DE-Stacking integration model is constructed using the optimal basic classifier as the meta-model and multiple machine learning models as base models, and using the differential evolution algorithm to tune the overall hyperparameters; the multiple machine learning models include support vector machines, K-nearest neighbor models, and naive Bayes models.

[0014] The DE-Stacking integrated model is trained using the optimal feature subset, and the basic taste corresponding to the EEG signal to be tested is determined according to the trained DE-Stacking integrated model.

[0015] According to the specific embodiments provided in this application, this application discloses the following technical effects.

[0016] This application uses a channel selection method to screen electrode channels before classification, which can reduce computing requirements and improve classification accuracy, or significantly reduce computing time while maintaining accuracy. Channel selection not only reduces the number of channels but also reduces the complexity and cost of experimental operations, thereby further expanding the advantages of EEG and improving the accuracy of taste classification.

[0017] In addition, this application selects the machine learning model with the highest classification accuracy as the optimal basic classifier, combines multiple machine learning models, constructs a DE-Stacking integration model, and then optimizes the hyperparameters of the DE-Stacking integration model. Compared with the method of individually tuning the hyperparameters of multiple machine learning models, the constructed integration model still has a certain tuning space. This application uses the Differential Evolution Algorithm (DE) to tune the hyperparameters of the DE-Stacking integration model, which can further improve the classification accuracy of the final Stacking integration learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 This is a flow chart of the taste classification method based on EEG data provided in this application.

[0020] Figure 2 A comparison chart of the basic classifiers provided in this application.

[0021] Figure 3 This is the flow chart for extracting the optimal feature subset provided by this application.

[0022] Figure 4 This is the channel selection effect diagram provided by this application.

[0023] Figure 5 This is the effect diagram of the optimal feature subset provided by this application.

[0024] Figure 6 This is the hyperparameter optimization flowchart of the DE-Stacking ensemble learning model provided in this application. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0026] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0027] The embodiment of the present application provides a taste classification method based on EEG data, which is executed by a computer device, specifically a computer device such as a terminal or a server, or a terminal and a server. In the embodiment of the present application, Figure 1 As shown, the method includes the following steps.

[0028] S1: Collecting the EEG signals of the subjects when they are stimulated with different basic tastes; the basic tastes include sour, sweet, bitter, salty and umami.

[0029] S2: Based on the EEG signal, a full feature data set is constructed; the full feature data set includes the features of the preprocessed EEG signal; the features include linear features, nonlinear features and functional connectivity features; the linear features include mean, extreme value, variance and power spectral density; the nonlinear features include fuzzy entropy; the functional connectivity features include phase lag index.

[0030] S3: Based on the full feature data set, determine the optimal basic classifier, and screen the electrode channels according to the optimal basic classifier to extract the optimal feature subset.

[0031] S4: Using the optimal basic classifier as the meta-model and multiple machine learning models as base models, and using the differential evolution algorithm to tune the overall hyperparameters, a DE-Stacking integration model is constructed; the multiple machine learning models include support vector machines, K-nearest neighbor models, and naive Bayes models.

[0032] S5: Using the optimal feature subset to train the DE-Stacking integrated model, and determining the basic taste corresponding to the EEG signal to be tested according to the trained DE-Stacking integrated model.

[0033] In an exemplary embodiment, S1 may be replaced by the following steps.

[0034] S11: A non-invasive data acquisition method was used. The acquisition equipment was a 32-lead wireless EEG acquisition device from Boricon. Reference electrodes were placed on both mastoid processes. The impedance of each lead on the electrode cap was confirmed to be less than 5kΩ. The sampling rate was 1000Hz.

[0035] S12: 31 young people with good physical and mental health and sensitive taste were selected as subjects. The task was carried out in an EEG test room that was safe, quiet, odor-free and at a suitable temperature. The subjects rinsed their mouths and closed their eyes and took deep breaths to maintain a relaxed state. Next, a taste stimulus sequence was randomly generated, and five taste stimulants were placed in order on the table. The subjects sipped 20 ml of the five tastant solutions in turn, held them in their mouths, and then pressed a button to mark them. Each tastant solution was kept in the mouth for 40 seconds, and the EEG signal was obtained for 40 seconds before the waste liquid was spit out. After obtaining a taste EEG signal, the subjects rinsed their mouths with pure water and rested for 40 seconds to eliminate the influence of the stimulus, allowing the recorded electrical signal to return to near the baseline, and then started the next taste EEG signal test. After all recordings were completed, the task ended.

[0036] In an exemplary embodiment, S2 may be replaced by the following steps.

[0037] S21: Preprocessing the EEG signal to determine a preprocessed EEG signal.

[0038] S22: Extracting features of the preprocessed EEG signals under different rhythms to construct a full feature dataset.

[0039] In an exemplary embodiment, S21 may be replaced by the following steps.

[0040] S211: Locate the electrode channels and segment the EEG signals in each electrode channel using a 1s non-overlapping time window to obtain the segmented EEG signals of each electrode channel.

[0041] S212: Downsampling and filtering are performed on the EEG signals after segmentation of each electrode channel to determine a filtered EEG signal.

[0042] In practical applications, under the premise of complying with the Nyquist theorem, in order to improve processing efficiency, the sampling frequency is reduced to 200 Hz; a finite impulse response filter (FIR) is used for 2 Hz high-pass filtering and 45 Hz low-pass filtering.

[0043] S213: Using the electrodes at the bilateral mastoid processes as reference electrodes, re-reference the filtered EEG signal.

[0044] S214: Segment processing is performed on the referenced EEG signal, and independent component analysis is performed on the EEG signal of each segment to determine the preprocessed EEG signal.

[0045] In practical applications, independent component analysis is performed on each segment of EEG signals to remove artifacts such as electrooculogram (EOG).

[0046] In practical applications, in order to increase the sample size, according to the experimental process, a 1-second non-overlapping time window is selected to segment the referenced EEG signals.

[0047] In practical applications, in an exemplary embodiment, S22 can be replaced by the following steps.

[0048] S221: Differentiate between different rhythms. Features are extracted from seven frequency bands: Delta (2-4Hz), Theta (4-8Hz), Alpha1 (8-10Hz), Alpha2 (10-13Hz), Beta1 (13-20Hz), Beta2 (20-30Hz), and Gama (30-45Hz).

[0049] S222: Feature extraction. Common EEG features can be categorized into three types: linear features, nonlinear features, and functional connectivity features. Representative features from each of these three categories are extracted, such as the mean, extreme value, variance, and power spectral density (PSD) for linear features, the fuzzy entropy (FE) for nonlinear features, and the phase lag index (PLI) for functional connectivity.

[0050] 1) Mean.

[0051] (1)

[0052] in, is the mean; N is the total number of EEG signal data points in the time interval; is the EEG signal value at the pth moment, p=1,2,3...

[0053] 2) Extreme value.

[0054] (2)

[0055] in, is an extreme value.

[0056] 3) Variance.

[0057] (3)

[0058] in, s is the variance; is the mean value of the EEG signal in the time interval.

[0059] 4) PSD.

[0060] PSD is the power of the signal within a unit frequency band. Its significance lies in converting brain waves whose amplitude changes with time into a spectrum diagram in which brain power changes with frequency, so that the distribution and changes of brain wave rhythms can be observed intuitively.

[0061] (4)

[0062] (5)

[0063] in, For EEG signals x ( n ), is an imaginary number; is the angular frequency; is the sequence number of the EEG signal data point; is the result of PSD estimation.

[0064] 5) FE.

[0065] FE is a method to measure the complexity of time series and is used to describe the uncertainty and irregularity of the system.

[0066] (6)

[0067] (7)

[0068] (8)

[0069] (9)

[0070] (10)

[0071] (11)

[0072] in, is the delay vector; For the The EEG signal value at the time, m is the pattern dimension. Since the length of each EEG data segment (i.e., EEG signal) (N=1s×200Hz=200) is less than 1000, the value of the pattern dimension m is set to 2; is the average value of the m-dimensional delay vector; ; Is the index, representing the delay vector At various points in time; For the p+k Signal value at time For the q+k Signal value at time For the pdelay vector and the q Similarity measure between delay vectors; For the p delay vector and the q The maximum distance between delay vectors; r represents the width of the fuzzy function boundary, r Usually equal to 0.2×std(x), std(x) is the standard deviation of the time series; The embedding dimension is m The average similarity when is the fuzzy entropy; The embedding dimension is m Average similarity at +1.

[0073] 6) PLI.

[0074] Among functional connectivity metrics, PLI is used to evaluate the connection strength between different channels in EEG signals.

[0075] (12)

[0076] in, is the value of PLI, which is used to evaluate the connection strength between different channels in the EEG signal; sign represents the sign function; is the phase value of the x channel signal at time t; is the phase value of the y channel signal at time t.

[0077] The PLI value range is 0-1. A PLI of 0 indicates no coupling or a coupled phase difference close to 0. The higher the phase coupling, the closer the PLI value is to 1.

[0078] In an exemplary embodiment, S3 may be replaced by the following steps.

[0079] S31: Using the full-feature dataset, determine the classification accuracy of different machine learning models; the machine learning models include a random forest (RF) model, a support vector machine, a K-nearest neighbor model, and a naive Bayes model.

[0080] In practical applications, commonly used machine learning models (random forest, support vector machine, K-nearest nearest neighbor and naive Bayes) are used to classify full-feature datasets, such as Figure 2 As shown in the figure, the random forest model with the highest classification accuracy was selected as the basic classifier, with an accuracy of 67.32%±13.97%.

[0081] S32 selects the machine learning model with the highest classification accuracy as the optimal basic classifier.

[0082] S33: Using a genetic algorithm (GA) to select electrode channels in the full feature data set to determine the screened electrode channels.

[0083] S34: Extracting an optimal feature subset based on the screened electrode channels.

[0084] In practical applications, such as Figure 3 As shown in the figure, a genetic algorithm is used to select electrode channels, and then the optimal feature subset is extracted through the contribution rate of the random forest model.

[0085] Genetic algorithm is used to select the electrode channels. From the 30 available EEG channels, the goal is to reduce the number of channels while improving the accuracy of the classification task. Due to the large number of possible channel combinations, there are 2 30 -1 combination, and calculating all possibilities is impractical. A genetic algorithm calculates the fitness function of the solution objective to automatically narrow the search direction and scale for the optimal channel combination. Furthermore, the genetic algorithm's mutation operation can, to a certain extent, help the population escape the local search range and find the global optimal solution, performing well for combinatorial optimization problems.

[0086] During channel selection, different feature types are processed differently. For linear and nonlinear features, the features of the selected channels can be directly incorporated. However, processing functional connectivity features is more complex. These functional connectivity features must be converted into feature indices based on the channel combinations selected by the genetic algorithm, and then features corresponding to all indices are selected.

[0087] like Figure 4 The results of channel selection are shown. When the 15 electrode channels FP1, FP2, F4, F7, FZ, FC1, FC5, FC6, C3, CP1, CP5, P3, P8, PO3 and OZ are selected, the classification accuracy of RF reaches 68.27%±13.85%.

[0088] In an exemplary embodiment, S33 may be replaced by the following steps.

[0089] S331: Let each chromosome represent an electrode channel, the initial population number and the maximum number of iterations; each gene on the chromosome is encoded in binary, 0 indicates that the electrode channel is not selected, and 1 indicates that the electrode channel is selected.

[0090] S332: The average classification accuracy of all subjects is used as the fitness value.

[0091] S333: Select the individuals with the highest fitness from the population to enter the next generation.

[0092] S334: Crossover operation: During each generation iteration, two crossover points are randomly selected from the gene sequences of the two parent individuals, and the partial gene sequences between the two crossover points are exchanged.

[0093] S335: Mutation operation: Based on the crossover operation, individual genes are randomly changed according to the mutation probability to introduce new genes.

[0094] S336: Repeat the crossover operation and mutation operation until the maximum number of iterations is reached, and determine the screened electrode channel.

[0095] In practical applications, the parameters of the genetic algorithm are set as follows.

[0096] 1) Initialize the population.

[0097] Each chromosome represents a channel. Each gene is encoded in binary, with 0 indicating that the channel is not selected and 1 indicating that the channel is selected. The initial population size is 100, and the maximum number of iterations is 200.

[0098] 2) Fitness calculation.

[0099] The average classification accuracy of the final classification results of all subjects is selected as the fitness value.

[0100] 3) Select an operation.

[0101] Using tournament selection, a number of individuals are randomly selected from the population to form a "small tournament" and then the individuals with the highest fitness are selected to enter the next generation. This method is better suited to maintaining diversity and preventing premature convergence.

[0102] 4) Crossover operation.

[0103] The crossover process uses a two-point crossover, randomly selecting two crossover points in the genetic sequences of the two parent individuals and exchanging the portion of the genetic sequence between these two points. The crossover rate is set to 0.8. A higher crossover rate will cause a larger number of individuals to cross over, which helps explore the search space, increases population diversity, and prevents the algorithm from falling into a local optimum.

[0104] 5) Mutation operation.

[0105] Based on the mutation probability, the genes of individuals are randomly altered to introduce new genes and maintain population diversity. The mutation rate is set to 0.05. A low mutation rate strikes a balance between introducing new genes and maintaining high-quality genes. However, a high mutation rate can destroy existing high-quality genes and may prevent the algorithm from converging, so it should be set with caution.

[0106] In an exemplary embodiment, S34 may be replaced by the following steps.

[0107] S341: Based on the screened electrode channels, and according to the feature contribution rate of the optimal basic classifier, the features of the screened electrode channels are sorted by importance to generate a feature importance sequence.

[0108] S342: Adopting a step-by-step feature addition strategy, adding features to the optimal basic classifier in order of feature importance.

[0109] S343: Based on the optimal basic classifier, calculate the average classification accuracy of all subjects.

[0110] S344: Using the average classification accuracy as an evaluation indicator, a 5-fold cross-validation method is used to extract the feature subset with the highest classification accuracy as the optimal feature subset.

[0111] In practice, the optimal feature subset is extracted using the contribution rate of the random forest. After filtering the channels, the remaining features are ranked by importance based on their contribution rate. A strategy of incremental feature addition is then adopted, where only one feature is added to the classifier at each step until all features are included.

[0112] Among them, assuming that there are multiple features in this application, the first step is to add the feature with the highest contribution rate to obtain a result, the second step is to add the top two features with the top two contribution rates to obtain a result, the third step is to add the top three features, and so on.

[0113] Given the significant differences in EEG signals among different subjects, the data of each subject were analyzed separately, and the average classification accuracy of all subjects was calculated as the main evaluation indicator.

[0114] Furthermore, a 5-fold cross-validation method was used to evenly divide the original dataset into five subsets. Each subset was used as the test set in turn, while the remaining four subsets were used as the training set. This process was repeated five times, each time using a different subset as the test set, to ensure the stability and reliability of the evaluation results. Finally, the results of the five evaluations were averaged to determine the final classification performance.

[0115] Through this method, the feature subset with the highest classification accuracy is extracted from all subjects as the optimal feature subset. Figure 5 The figure shows the result of the optimal feature subset. When the 101 features with the highest contribution rate are taken, the classification accuracy reaches its peak.

[0116] In an exemplary embodiment, S5 may be replaced by the following steps.

[0117] S51: Using the optimal feature subset to train the DE-Stacking integration model, and using the differential evolution algorithm to tune the hyperparameters of the DE-Stacking integration model to determine the trained DE-Stacking integration model.

[0118] S52: Determine the basic taste corresponding to the EEG signal to be tested according to the trained DE-Stacking integrated model.

[0119] like Figure 6 As shown in the figure, the hyperparameters of the DE-Stacking ensemble model are automatically tuned through the differential evolution algorithm, and the classification accuracy is improved using the DE-Stacking ensemble model.

[0120] In this application, the hyperparameter selection and tuning of the base classifiers were not rigorously optimized. This is primarily because, even if the optimal hyperparameters are found for individual classifier models through methods such as grid search, when these individual models are integrated, the optimal hyperparameter combination may not necessarily lead to a global optimal DE-Stacking ensemble model. Furthermore, grid search in this case consumes a large amount of computing resources and is inefficient. Therefore, the DE-Stacking ensemble model still has significant optimization potential.

[0121] To achieve global tuning of the entire DE-Stacking ensemble model, traditional grid search methods are clearly unsuitable given the model's complexity and the large number of parameters that require optimization. In contrast, optimization algorithms based on random search, such as genetic algorithms and differential evolution, are more effective in searching in high-dimensional parameter spaces. Compared to genetic algorithms, differential evolution performs better in continuous optimization problems. Therefore, differential evolution was chosen to globally optimize the hyperparameters of the DE-Stacking ensemble model, further improving final classification accuracy.

[0122] After experimental verification, Table 1 is a comparison table of the effects of this application and other methods. As shown in Table 1, compared with other methods, this application has significantly higher recognition performance of taste EEG than other methods and has better generalization performance.

[0123] Table 1

[0124]

[0125] This application uses a genetic algorithm (GA)-RF method for channel selection, which not only improves classification accuracy but also reduces the use of redundant channels, which directly affects the power consumption, data processing load, and device size of the wearable system. In wearable BCI systems, especially in taste perception applications, reducing device complexity is crucial. By reducing the number of sensors, not only can hardware costs be reduced, but the system's portability can also be improved, making it easier for users to use. This also helps promote the development of brain-computer interface technology into a wider range of application scenarios.

[0126] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A taste classification method based on EEG data, characterized in that: The taste classification method based on EEG data includes: Collecting EEG signals of the subjects when they are stimulated with different basic tastes; the basic tastes include sour, sweet, bitter, salty and umami; Constructing a full feature data set based on the EEG signal; the full feature data set includes features of the preprocessed EEG signal; the features include linear features, nonlinear features, and functional connectivity features; the linear features include mean, extreme value, variance, and power spectral density; the nonlinear features include fuzzy entropy; and the functional connectivity features include phase lag index; Based on the full feature data set, an optimal basic classifier is determined, and according to the optimal basic classifier, electrode channels are screened to extract an optimal feature subset, specifically including: Using the full-featured dataset, determine the classification accuracy of different machine learning models; the different machine learning models include random forest model, support vector machine, K nearest neighbor model, and naive Bayes model; Select the machine learning model with the highest classification accuracy as the optimal basic classifier; Selecting electrode channels in the full feature data set using a genetic algorithm to determine screened electrode channels; Based on the screened electrode channels, an optimal feature subset is extracted, specifically including: Based on the screened electrode channels, and according to the feature contribution rate of the optimal basic classifier, the features of the screened electrode channels are sorted by importance to generate a feature importance sequence; Adopting a step-by-step feature addition strategy, adding features to the optimal basic classifier in order of feature importance; Based on the optimal basic classifier, calculate the average classification accuracy of all subjects; Taking the average classification accuracy as the evaluation index, a 5-fold cross-validation method is used to extract the feature subset with the highest classification accuracy as the optimal feature subset; A DE-Stacking ensemble model is constructed using the optimal basic classifier as a meta-model and multiple machine learning models as base models, and using a differential evolution algorithm to tune overall hyperparameters; the multiple machine learning models include a support vector machine, a K-nearest neighbor model, and a naive Bayes model; The DE-Stacking integrated model is trained using the optimal feature subset, and the basic taste corresponding to the EEG signal to be tested is determined according to the trained DE-Stacking integrated model.

2. The taste classification method based on EEG data according to claim 1, characterized in that: Based on the EEG signal, a full feature dataset is constructed, specifically including: Preprocessing the EEG signal to determine a preprocessed EEG signal; The features of the preprocessed EEG signals under different rhythms are extracted to construct a full feature dataset.

3. The taste classification method based on EEG data according to claim 2, characterized in that: Preprocessing the EEG signal to determine the preprocessed EEG signal specifically includes: Locate the electrode channels and segment the EEG signals in each electrode channel using a 1s non-overlapping time window to obtain the segmented EEG signals of each electrode channel; Downsampling and filtering are performed on the EEG signals after segmentation of each electrode channel to determine the filtered EEG signals; Using electrodes at the bilateral mastoid processes as reference electrodes to re-reference the filtered EEG signal; The referenced EEG signal is segmented and the independent component analysis is performed on each segment of the EEG signal to determine the preprocessed EEG signal.

4. The taste classification method based on EEG data according to claim 1, characterized in that: The electrode channels in the full feature data set are selected using a genetic algorithm to determine the screened electrode channels, specifically including: Let each chromosome represent an electrode channel, the initial population number and the maximum number of iterations; each gene on the chromosome is encoded in binary, 0 indicates that the electrode channel is not selected, and 1 indicates that the electrode channel is selected; The average classification accuracy of all subjects is used as the fitness value; Select the individuals with the highest fitness from the population to enter the next generation; Crossover operation: In each iteration, two crossover points are randomly selected from the gene sequences of the two parent individuals, and the partial gene sequences between the two crossover points are exchanged; Mutation operation: Based on the crossover operation, the individual genes are randomly changed according to the mutation probability to introduce new genes; The crossover operation and the mutation operation are repeated until a maximum number of iterations is reached to determine the screened electrode channels.

5. The taste classification method based on EEG data according to claim 1, characterized in that: The optimal basic classifier is used as the meta-model, multiple machine learning models are used as base models, and the differential evolution algorithm is used to tune the overall hyperparameters to build a DE-Stacking integrated model, which specifically includes: A stacking ensemble model is constructed using the optimal basic classifier as a meta-model and multiple machine learning models as base models; The differential evolution algorithm is used to tune the overall hyperparameters of the Stacking ensemble model and construct the DE-Stacking ensemble model.

Images