Emotion recognition method and system based on dynamic clustering and variational bayesian meta-learning

Abstract

The application discloses a kind of based on dynamic clustering and variational bayes meta-learning emotional recognition method and system, specifically use restaurant process to dynamically determine the number of crowd categories, and in view of the difference existing in the mode of emotional expression by physiological signal for different crowd categories, the idea of bayes meta-learning is introduced to learn the commonality of crowd category emotional expression mode, and the characteristics of crowd individual emotional expression mode are learned by the sub-process of meta-learning.There are four steps: data arrangement and preprocessing, initialization of model crowd dynamic clustering process parameters and meta-parameters, online training and parameter updating of algorithm model, model performance test and put into use;The application can establish different emotional classification models with different parameter distributions for different crowd categories, thereby realizing high-precision emotion recognition, and simultaneously having strong generalization ability of knowledge transfer between different crowds.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of information technology, and particularly relates to an emotion recognition method and system based on dynamic clustering and variational Bayesian meta-learning using electroencephalogram signals. BACKGROUND

[0002] In recent years, with the increasing attention of the whole society to people's mental health diseases, the corresponding emotion recognition technology has also been continuously developed. As an important physiological characteristic different from traditional facial expressions, sounds and other easy-to-fake features, electroencephalogram (EEG) has been widely valued and studied in the field of emotion recognition. However, the existing emotion recognition methods often have difficulty in balancing the classification accuracy of emotions and the generalization ability of the model in dealing with physiological signal characteristics of different populations when processing emotion classification of different populations.

[0003] Traditional emotion recognition methods usually assume that the emotion classification patterns of all individuals are the same (for example, ordinary people, psychologically sensitive people or people with special mental diseases), which leads to the fact that traditional emotion recognition methods are often not ideal in practical applications. A simple example is that different populations will produce different emotions when experiencing the same event. If an emotion classification model is trained separately for a certain population, two difficulties will arise, one is that there is no unified measurement standard to divide the dimension of emotion expression patterns of the population, and the other is that it is often difficult to obtain a large amount of experimental data in the corresponding population to train the model. Therefore, it is of great significance in the field of emotion research and application to study a method that can dynamically adapt to the emotion classification patterns of different populations and obtain high emotion classification performance with a small amount of sample support. SUMMARY

[0004] The purpose of the present application is to solve the problems that the existing emotion recognition methods often have difficulty in balancing the classification accuracy of emotions and the generalization ability of the model in dealing with physiological signal characteristics of different populations when processing emotion classification of different populations, and the traditional emotion recognition methods usually assume that the emotion classification patterns of all individuals are the same, which leads to the fact that the traditional emotion recognition methods are often not ideal in practical applications.

[0005] In order to solve the above problems, the present application provides an emotion recognition method based on dynamic clustering and variational Bayesian meta-learning, comprising the following steps:

[0006] Step S1: data arrangement and preprocessing;

[0007] Step S1-1: arrange the existing subject electroencephalogram signals, uniformly downsample the electroencephalogram signals to 128HZ, filter the electroencephalogram signals using a 5 to 45HZ band-pass filter, and obtain physiological signal characteristic data called by the method;

[0008] Step S1-2: The subject's sequence number is t, and the set of physiological signal characteristic data sequences obtained from the subject is denoted as:

[0009]

[0010] The set of subjects' true emotional labels is denoted as:

[0011]

[0012] In the formula, This refers to the multi-channel physiological signals obtained from subjects in a single emotional stimulation experiment. The number of characteristic channels representing physiological signals. The product of the duration and frequency of the physiological signal represents the number of time points at which the physiological signal is ultimately obtained. Represents the set of real numbers. This represents the number of physiological signal samples with emotion labels obtained from the t-th subject, and the corresponding emotion labels. Using one-hot encoding, the liking level dimension of the emotion annotation based on the DEAP dataset is given, and the subjects' emotions are divided into two categories: liking and disliking, which are represented by vectors [0,1] and [1,0] respectively;

[0013] Steps S1-3: Collect the sample set of the subjects. The support set samples were split into training samples of the subjects' own emotion classifier. and a sample query set for testing The splitting was based on randomly shuffling the data and then dividing it according to the specified proportions.

[0014] Step S2: Initialization of parameters and meta-parameters in the dynamic clustering process of the population in the model;

[0015] Step S2-1: Initialize the dynamic clustering parameters of the population based on the Chinese restaurant process CRP, set a hyperparameter α to determine the population category of the newly entered i-th subject, and use... The formula for the label representing its category is as follows:

[0016]

[0017] In the formula, Indicates the first The number of subjects expected to be assigned to a population, α>0 is used as a concentration hyperparameter to control the number of population types;

[0018] Step S2-2: The probability of a certain subject being classified in a certain population category needs to be determined by constructing a stick-breaking representation for each population category. According to the Dirichlet process principle, the initialization process is as follows:

[0019]

[0020] wherein, represents a probability value vector, wherein each element represents the probability of classifying the subject individual into different population categories, and the values of the elements in the vector need to be non-negative and the sum of all values is 1. The probability of the kth category is represented as

[0021] Step S2-3: The traditional meta-learning assumes that there is a meta-model with a parameter . A corresponding emotion classifier is generated using the meta-parameters in the meta-model, and the generation process is as follows:

[0022]

[0023] wherein, represents a Gaussian mixture distribution, represents sampling from the probability value vector according to the corresponding probability values to obtain a certain result of , i.e., the current subject t is explicitly classified into the corresponding population category, and the meta-parameter distribution of the subject's population category is determined. represents the population category to which the current subject belongs, represents sampling a corresponding meta-parameter framework from the classifier parameter distribution of the corresponding population, represents a mean vector subject to Gaussian mixture distribution, represents a semi-positive covariance matrix;

[0024] generates an emotion classifier sub-model with initialized parameters for the current subject, and the specific process is as follows:

[0025]

[0026] Step S3: Online training and parameter updating of the algorithm model;

[0027] Step S3-1: When the sample data of the tth subject is input, the distribution parameters of the four parameters are output by the posterior method through structured inference. The parameters of the Beta distribution subject to in formula 2 are denoted as , and the subject to ​The parameters of the distribution are denoted as The task-specific parameters in equation 5 are denoted as The parameters of the Gaussian mixture distribution are denoted as The variational parameters of equation 6 are denoted as The variational parameters of equation 6 are denoted as The model parameters can be updated or used online, and the specific formula is:

[0028]

[0029] When the tth subject arrives, the observed lower bound of evidence ELOB is further described according to equation 8:

[0030]

[0031] In the formula, represents the mathematical expectation, represents the Kullback-Leibler divergence, and the optimization model parameters can be completed by maximizing the loss function, q represents the distribution predicted by the model, and p represents the true distribution;

[0032] Step S3-2: Generation and optimization process of four distribution parameters:

[0033] (1) The variational distribution of is only considered in the item related to in ELOB, and the optimal variational distribution is:

[0034]

[0035] (2) The variational distribution of is also only considered in the item related to in ELOB, and the optimal variational distribution is:

[0036]

[0037] That is, a parameterized classifier distribution:

[0038]

[0039] In the formula, is the variational distribution of

[0040] (3) The variational distribution of is when the tth subject arrives, the kth population category to which it belongs is inferred After that, the meta-learning parameters of the corresponding population distribution are denoted as: ​​​

[0041]

[0042] Assuming that each population category is subject to a Gaussian distribution with a parameter of

[0043]

[0044] Without loss of generality, the new prior knowledge of the tth subject is given by the posterior knowledge of the last step, i.e.

[0045]

[0046] After the current subject's data training is completed, the new posterior knowledge is used to accept the new subject emotion classification task, and the parameters are updated by the stochastic gradient descent method;

[0047] (4) The variational distribution of the generated emotion classifier with the initialized parameters, formula is:

[0048]

[0049] In the formula, indicates the emotion classifier neural network parameters generated by the method for the current subject, represents the sampled variational parameters, and the variational parameters of the population category to which the current subject belongs are used for training data with the current subject Perform several steps of gradient descent, i.e.

[0050]

[0051] In the formula, is the step of gradient descent, is the learning rate, and the gradient descent strategy uses stochastic gradient descent;

[0052] Step S4: performance test and put into use;

[0053] When a subject physiological signal sample set with only a small number of sample labels ​​After the method is called, the neural network trained in step S3 is called to accurately divide the subject into the corresponding population, and the meta-learning framework trained in the corresponding population is called to generate an emotion classifier with initial parameters for the subject. Through an end-to-end manner, the physiological signal is input as an input, and an emotion classification result is output. A small amount of samples are used to quickly train the emotion classifier for use or testing. At this time, the test steps are as follows: when one or a batch of physiological signal samples without emotion labels are input, the corresponding physiological signal samples are input into the corresponding emotion classifier, and the emotion prediction output of the meta-parameter model for the physiological signal samples can be obtained.

[0054] In a preferred manner, the emotion classifier sub-model belongs to a replaceable component, and a multi-layer convolutional neural network module is selected in the method.

[0055] In a preferred manner, the step S1-3 splitting ratio is 60% and 40%.

[0056] An emotion recognition system based on dynamic clustering and variational Bayesian meta-learning includes a data arrangement and data preprocessing module, a dynamic clustering parameter module based on Chinese restaurant process CRP, a meta-parameter module, a model parameter updating module, and a model use module:

[0057] Step M1: The data arrangement and preprocessing module includes the following specific implementations:

[0058] Step M1-1: The existing subject electroencephalogram signal is arranged, and the electroencephalogram signal is uniformly down-sampled to 128 Hz. A 5-45 Hz band-pass filter is used to filter the electroencephalogram signal to obtain physiological signal feature data called by the method.

[0059] Step M1-2: The subject sequence number is t, and the physiological signal feature data sequence set obtained from the subject is denoted as:

[0060]

[0061] The real emotion label set of the subject is denoted as:

[0062]

[0063] In the formula, represents the multi-channel physiological signal obtained from the subject in a single emotion stimulation experiment, represents the number of feature channels of the physiological signal, represents the product of the time length and frequency of the physiological signal, that is, the number of time points of the finally obtained physiological signal, represents a real number set, This represents the number of physiological signal samples with emotion labels obtained from the t-th subject, and the corresponding emotion labels. Using one-hot encoding, the liking level dimension of the emotion annotation based on the DEAP dataset is given, and the subjects' emotions are divided into two categories: liking and disliking, which are represented by vectors [0,1] and [1,0] respectively;

[0064] Step M1-3: Collect the sample set of the subjects. The support set samples were split into training samples of the subjects' own emotion classifier. and a sample query set for testing The splitting was based on a random shuffling, with the components divided into 60% and 40% portions respectively.

[0065] Step M2: Dynamic clustering parameter module based on Chinese restaurant process CRP;

[0066] Step M2-1: The Chinese restaurant process controls the category determination for the newly entered i-th subject by setting a hyperparameter α, and uses... The formula for the label representing its category is as follows:

[0067]

[0068] In the formula, Indicates the first The number of subjects expected to be assigned to a population, α>0 is used as a concentration hyperparameter to control the number of population types;

[0069] Step M2-2: Determining the probability of a subject being classified into a specific population category requires constructing a broken bar representation for each category. Based on the Dirichlet process principle, the initialization formula is as follows:

[0070]

[0071] In the formula, Represents a probability value vector. Each element in the vector represents the probability of an individual subject being assigned to a different population category. It is necessary to ensure that all elements in the vector are non-negative and that the sum of all values ​​is 1. The probability of the k-th category is expressed as... ;

[0072] Step M3: The specific implementation method of the meta-parameter module is as follows:

[0073] The traditional meta-learning assumption has a parameter of The meta-model is used to generate a corresponding sentiment classifier using the meta-parameters in the meta-model. The generation process is as follows:

[0074]

[0075] In the formula, Indicates a mixture Gaussian distribution. Represents the probability value vector Each component is sampled according to its corresponding probability value to obtain a definite value. The result is that the current subject t is clearly classified into the corresponding population category, and the distribution of meta-parameters of the population category to which the subject belongs is determined; This indicates the population category to which the current subject belongs. This indicates that a corresponding meta-parameter framework is sampled from the classifier parameter distribution of the corresponding population. This represents the mean vector of a Gaussian mixture distribution. Represents the positive semi-covariance matrix;

[0076] Generate a corresponding emotion classifier sub-model with initialization parameters for the current subject. The specific process is as follows:

[0077]

[0078] Step M4: Model parameter update module;

[0079] Step M4-1: Using structured inference, input the sample data of the t-th subject, output the distribution parameters of the four parameters, and apply the formula in equation 2. The parameters of the Beta distribution are denoted as . Formula 4 Obedience The parameters of the distribution are denoted as In formula 5 The parameters that follow a Gaussian mixture distribution are denoted as . The task-specific parameters in Formula 6 The variational parameter is denoted as This allows for continuous online updates to model parameters or deployment. The specific formula is as follows:

[0080]

[0081] When the t-th subject arrives, the observed lower bound of evidence (ELOB) is further described according to Formula 8 as follows:

[0082]

[0083] In the formula, Represents the mathematical expectation. The Kullback-Leibler divergence is represented, and the optimization of model parameters can be achieved by maximizing this loss function. q represents the distribution predicted by the model, and p represents the true distribution.

[0084] Step M4-2: Generation and optimization process of four distribution parameters:

[0085] (1) The variational distribution of is only related to the term in the ELOB, and the optimal variational distribution is:

[0086]

[0087] (2) The variational distribution of is also only related to the term in the ELOB, and the optimal variational distribution is:

[0088] That is, a parameterized classifier distribution:

[0089]

[0090]

[0091] In the formula, is the variational distribution of

[0092] (3) The variational distribution of is when the tth subject arrives, the kth population category to which it belongs is inferred After that, the meta-learning parameters of the corresponding population Distribution is represented as:

[0093]

[0094] Assume that each population category is subject to a Gaussian distribution with a parameter

[0095]

[0096] Without loss of generality, when the tth subject arrives, the prior knowledge of the new is given by the posterior knowledge of the existing , that is:

[0097]

[0098] And after the data of the current subject is trained, the new posterior knowledge is replaced to accept the new subject emotion classification task, and the parameters are updated by the method of stochastic gradient descent.​​​

[0099] (4) the variational distribution, i.e., the generated emotion classifier with initialized parameters, is given by:

[0100]

[0101] where, denotes the emotion classifier neural network parameters generated by the present method for the current subject, represents the sampled variational parameters, the variational parameters of the subject class to which the current subject has been determined to belong for training with the training data of the current subject performing gradient descent in several steps, i.e.,

[0102]

[0103] where, is the step of gradient descent, is the learning rate, and the gradient descent strategy uses stochastic gradient descent;

[0104] Step M5: Model deployment module

[0105] When a subject physiological signal sample set with only a small number of sample labels After being input into the present method, the neural network trained in step S3 is called to accurately divide the subject into the corresponding population, and a trained meta-learning framework in the corresponding population is called to generate an emotion classifier with initialized parameters for the subject. Through an end-to-end manner, i.e., taking physiological signals as input and outputting emotion classification results, the emotion classifier can be quickly trained using the corresponding small number of samples to be put into use or tested. At this time, the testing step is as follows: when a physiological signal sample without emotion annotation is input, the corresponding physiological signal sample is input into the corresponding emotion classifier, and the emotion prediction output of the meta-parameter model for the physiological signal sample can be obtained.

[0106] The Monte Carlo method includes: randomly generating a large number of parameter samples and applying them to the model to observe the classification performance.

[0107] The electroencephalogram signal in step S1 is information recorded by existing devices or information obtained by other means. The present application does not focus on signal acquisition, but on subsequent processing of the existing signal.

[0108] Similarly, the present application relates to the following:

[0109] A set of physiological signal feature data sequences obtained from a subject denotes the multi-channel physiological signals obtained from the subjects in a single emotional stimulation experiment, denotes the number of feature channels of the physiological signals, denotes the product of the time length and the frequency of the physiological signals, i.e., the number of time points at which the physiological signals are finally obtained, denotes the real number set, denotes the number of emotional-labeled physiological signal samples obtained from the tth subject, and the corresponding emotional label is given in the form of one-hot encoding. The process of obtaining the above data is not involved in the present application, and the present application only takes and uses the existing data for subsequent steps.

[0110] To be more specific, the present application does not involve the process of directly acting on the human body to obtain signals. The concept of "obtaining signals" or "collecting signals" related to the acquisition of signals in the present application only takes and uses existing signal data, and does not collect signals by acting on the human body.

[0111] The present application has the following beneficial effects: in terms of the performance of emotion classification, the accuracy rate ACC commonly used in the field of emotion classification is used, and compared with other traditional machine learning methods which do not distinguish between populations, a classic machine learning model such as a support vector machine can achieve 69.85%; the accuracy rate of a traditional CNN convolutional network can reach 76.45%; the performance of a latest neural network based on deep learning such as a Tsception framework can reach about 82.37%; and the classification accuracy rate of the present application can reach about 85.23% under the same experimental conditions and using the same experimental environment, which is superior to the existing methods in the performance of emotion classification. Compared with other traditional researches, the present application supports model testing on individuals of a population with a small number of samples, for example: in the DEAP data set, the present application can be used in application after being trained by using a small number of individual samples (40 emotional samples of 16 subjects) of a population, and the accuracy rate of emotion classification can reach about 75%, which effectively proves the effectiveness of the present application in the small sample scenario and the effectiveness of the innovative idea of distinguishing between different populations. BRIEF DESCRIPTION OF DRAWINGS

[0112] Figure 1 is a method flowchart of the present application;

[0113] Figure 2 is a diagram showing the difference in parameter setting and parameter generation between the present application and traditional meta-learning. DETAILED DESCRIPTION

[0114] In combination Figure 1 shows the method flow of the present application, which includes the following steps:

[0115] Step S1: data sorting and preprocessing;

[0116] Step S1-1: Organize the existing subject's EEG signals, uniformly downsample the EEG signals to 128HZ, filter the EEG signals using a 5 to 45 HZ band-pass filter, and obtain the physiological signal feature data called by the method;

[0117] Step S1-2: The subject sequence number is t, and the physiological signal feature data sequence set obtained from the subject is denoted as:

[0118]

[0119] The set of real emotion labels of the subject is denoted as:

[0120]

[0121] In the formula, represents the multi-channel physiological signal obtained from the subject in a single emotion stimulation experiment, represents the number of feature channels of the physiological signal, represents the product of the time length and frequency of the physiological signal, i.e., the number of time points of the finally obtained physiological signal, represents a real number set, represents the number of emotion-labeled physiological signal samples obtained from the tth subject, and the corresponding emotion label is given in the form of one-hot encoding. The like dimension of the DEAP data set is annotated with emotion, and the subject's emotion is divided into two categories: like and dislike, represented by vectors [0, 1] and [1, 0], respectively.

[0122] Step S1-3: Split the sample set of the subject into support set samples for training the subject's own emotion classifier and query set samples for testing , and split according to the proportion after random shuffling;

[0123] Step S2: Initialization of model population dynamic clustering process parameters and meta-parameters;

[0124] Step S2-1: Population dynamic clustering parameter initialization based on Chinese restaurant process CRP, set a hyperparameter a to determine the population category of the newly entered ith subject, and use to represent the label of its category, and the formula is as follows:

[0125]

[0126] In the formula, represents the ith subject, and The number of subjects in the personal group expected to be divided, and a concentration hyperparameter controls the number of types of the population;

[0127] Step S2-2: The probability of a subject being divided into a certain personal group category needs to be determined by constructing a stick-breaking representation for each personal group category. According to the principle of Dirichlet process, the formula for the initialization process is:

[0128]

[0129] In the formula, represents a probability value vector, where each element represents the probability of dividing a subject into different personal group categories. It is necessary to ensure that the values of the elements in the vector are non-negative and the sum of all values is 1. The probability of the kth category is represented as

[0130] Step S2-3: The traditional meta-learning assumes that there is a meta-model with a parameter , and a corresponding emotion classifier is generated using the meta-parameters in the meta-model. The generation process is:

[0131]

[0132] In the formula, represents a Gaussian mixture distribution, represents sampling from the probability value vector according to the corresponding probability value to obtain a certain result of , that is, the current subject t is explicitly divided into the corresponding personal group category, and the meta-parameter distribution of the personal group category to which the subject belongs is determined. represents the personal group category to which the current subject belongs, represents sampling a corresponding meta-parameter framework from the classifier parameter distribution of the corresponding personal group, represents a mean vector subject to Gaussian mixture distribution, represents a semi-positive covariance matrix;

[0133] Generate an emotion classifier sub-model with initialized parameters for the current subject. The specific process is:

[0134]

[0135] Step S3: Online training and parameter updating of the algorithm model;

[0136] Step S3-1: Through structured inference, when the sample data of the tth subject is input, the distribution parameters of the four parameters are output through the posterior method. The formula 2 ​The parameters of the Beta distribution are denoted as . Formula 4 Obedience The parameters of the distribution are denoted as In formula 5 The parameters that follow a Gaussian mixture distribution are denoted as . The task-specific parameters in Formula 6 The variational parameter is denoted as This allows for continuous online updates to model parameters or deployment. The specific formula is as follows:

[0137]

[0138] When the t-th subject arrives, the observed lower bound of evidence (ELOB) is further described according to Formula 8 as follows:

[0139]

[0140] In the formula, Represents the mathematical expectation. represents the Kullback-Leibler divergence. Maximizing this loss function achieves the goal of optimizing the model parameters. q represents the distribution predicted by the model, and p represents the true distribution.

[0141] Step S3-2: The generation and optimization process of the four distribution parameters:

[0142] (1) The variational distribution of ELOB is considered only in relation to... The optimal variational distribution of the relevant terms is:

[0143]

[0144] (2) The variational distribution of ELOB is also considered only in relation to... The relevant terms are used to derive their optimal variational distribution as follows:

[0145]

[0146] That is, a parameterized classifier distribution:

[0147]

[0148] In the formula, yes Variational distribution;

[0149] (3) Given the variational distribution, when the t-th subject arrives, we can infer the k-th population category to which the subject belongs. After that, the meta-learning parameters of the corresponding crowd The distribution is expressed as:

[0150]

[0151] Assuming that each crowd category is subject to a Gaussian distribution with a parameter of

[0152]

[0153] Without loss of generality, the new prior knowledge of the tth subject is given by the existing posterior knowledge in the last step, that is:

[0154]

[0155] After the current subject's data training is completed, the new posterior knowledge is replaced to accept the new subject emotion classification task, and the parameters are updated by the stochastic gradient descent method.

[0156] (4) The variational distribution of the generated emotion classifier with the initialized parameters, the formula is:

[0157]

[0158] In the formula, represents the emotion classifier neural network parameters generated by the current subject, represents the sampled variational parameters, and the variational parameters of the crowd category to which the current subject belongs are used to train the data of the current subject Perform several steps of gradient descent, that is:

[0159]

[0160] In the formula, is the step of gradient descent, is the learning rate, and the gradient descent strategy uses stochastic gradient descent.

[0161] Step S4: performance test and put into use.

[0162] When a subject physiological signal sample set with only a small number of sample labels ​​After the method is transmitted, the neural network trained in step S3 is called to accurately divide the subject into the corresponding population, the meta-learning framework trained in the corresponding population is called to generate an emotion classifier with initial parameters for the subject, and the emotion classification result is output through an end-to-end manner, i.e., the physiological signal is taken as an input, the emotion classifier is quickly trained using a corresponding small amount of samples, and the emotion classifier can be put into use or tested; at this time, the test step is that when one or a batch of physiological signal samples without emotion labels are transmitted, the corresponding physiological signal samples are input into the corresponding emotion classifier, and the emotion prediction output of the meta-parameter model for the physiological signal samples can be obtained.

[0163] The above is the preferred embodiment of the present application, but the present application is not limited to the above examples, and it should be noted that other improvements and changes directly derived or imagined by those skilled in the art without departing from the spirit and concept of the present application should be considered to be included in the protection scope of the present application.

Claims

1. An emotion recognition method based on dynamic clustering and variational Bayesian meta-learning, characterized in that, Comprising the following steps: Step S1: data arrangement and preprocessing; Step S1-1: arrange the existing subject electroencephalogram signals, uniformly downsample the electroencephalogram signals to 128HZ, filter the electroencephalogram signals using a 5 to 45HZ band-pass filter, and obtain physiological signal feature data called by the method; Step S1-2: the subject sequence number is t, and the physiological signal feature data sequence set obtained from the subject is denoted as: The real emotion label set of the subject is denoted as: wherein, denotes the multi-channel physiological signals obtained from the subjects in the single emotional stimulus experiment, denotes the number of feature channels of the physiological signals, denotes the product of the length and frequency of the physiological signals, i.e. the number of time points at which the physiological signals are finally obtained, denotes the set of real numbers, denotes the number of emotional-labeled physiological signal samples obtained from the t-th subject, and the corresponding emotional label is given in the form of one-hot encoding, and the dimension of the like degree annotated by the DEAP dataset divides the subjects' emotions into two categories: like and dislike, which are represented by the vectors [0, 1] and [1, 0], respectively. Step S1-3: splitting the sample set of the subject Splitting the support set samples for training the subject's own emotion classifier and the query set samples for testing The splitting is according to the proportion after random shuffling respectively; Step S2: initialization of the population dynamic clustering process parameters and meta-parameters in the model; Step S2-1: Population dynamic clustering parameter initialization based on Chinese restaurant process CRP, set a hyper parameter a to determine the population category of the new i-th subject, and use to represent the label of its category, the formula is as follows: wherein, denotes the number of clusters the number of subjects in the population that are expected to be assigned to the cluster, and a > 0 controls the number of clusters as a concentration hyperparameter. Step S2-2: the probability of a certain subject being divided into a certain population category needs to be completed by constructing a broken stick representation for each population category, and according to the principle of Dirichlet process, the formula for the initialization process is: In the formula, represents a probability value vector, wherein the value of each element represents the probability of the individual subject being classified into different population categories, and it is required that the values of the elements in the vector are non-negative and the sum of all values is 1, and the probability of the kth category is represented as ; Step S2-3: The conventional meta-learning assumes that there is a meta-model with a parameter , and a corresponding emotion classifier is generated using the meta-parameter in the meta-model, and the generation process is: wherein, denotes a Gaussian mixture distribution, denotes sampling from the probability value vector to obtain a certain result, i.e. the current subject t is explicitly classified into the corresponding population class, and the meta-parameter distribution of the population class to which the subject belongs is determined; denotes the population class to which the current subject belongs, denotes sampling a corresponding meta-parameter framework from the classifier parameter distribution of the corresponding population, denotes a mean vector subject to a Gaussian mixture distribution, denotes a semi-positive definite covariance matrix. Generate an emotion classifier sub-model corresponding to the current subject with initialized parameters, and the specific process is: Step S3: online training and parameter updating of the algorithm model; Step S3-1: By means of structured inference, when the sample data of the tth subject is input, the distribution parameters of the four parameters are output in a posterior manner, and the parameters of the Beta distribution in formula 2 are updated as follows: The parameters of the Beta distribution are denoted as The parameters of the Beta distribution in formula 4 are updated as follows: The parameters of the Beta distribution are denoted as The parameters of the Beta distribution are denoted as The parameters of the Beta distribution in formula 5 are updated as follows: The parameters of the Gaussian mixture distribution are denoted as The variational parameters of the task-specific parameters in formula 6 are denoted as That is, the model parameters can be continuously updated or put into use in an online manner, and the specific formula is as follows: When the tthsubject arrives, the observed lower bound of evidence ELOB is further described according to formula 8: wherein denotes the mathematical expectation, denotes the Kullback-Leibler divergence, which is maximized to optimize the model parameters, q denotes the distribution predicted by the model, and p denotes the true distribution. Step S3-2: Generation and optimization process of four distribution parameters: (1) The variational distribution of the ELOB is given by The optimal variational distribution of the ELOB is given by (2) The variational distribution of ELOB, only considering the terms related to The optimal variational distribution of ELOB is derived as: That is, a classifier distribution with parameters: wherein is the variational distribution; (3) the variational distribution, when the t-th subject arrives, infer its k-th population class After that, the meta-learning parameters of the corresponding population are represented as: Assuming that each population category is subject to a parameter: Gaussian distribution, without loss of generality, the new prior knowledge of the tth subject is given by the stored posterior knowledge of the tth subject, i.e. And after the data training of the current subject is completed, replace the new posterior knowledge to accept the new subject emotion classification task, and update the parameters by the way of stochastic gradient descent; (4) the variational distribution, i.e., the generated emotion classifier with the initialization parameters, is given by: wherein, denotes the emotion classifier neural network parameters generated by the method for the current subject, denotes the sampled variational parameters, the variational parameters of the subject class to which the current subject has been determined to belong for training with the training data of the current subject performing gradient descent over several steps, namely: In the formula, is the step of gradient descent, is the learning rate, and the gradient descent strategy adopts a stochastic gradient descent. Step S4: performance test and put into use; When a subject physiological signal sample set with only a small amount of sample labels After the method is passed, the neural network trained in step S3 is called to accurately divide the subject into the corresponding population, and the meta-learning framework trained in the corresponding population is called to generate an emotion classifier with initial parameters for the subject. Through an end-to-end manner, i.e., taking physiological signals as input and outputting emotion classification results, the emotion classifier can be quickly trained using a small amount of samples for use or testing; At this time, the test step is when one or a batch of physiological signal samples without emotion labels are passed in, the corresponding physiological signal samples are input into the corresponding emotion classifier, and the emotion prediction output of the meta-parameter model for the physiological signal samples can be obtained. 2.The emotion recognition method based on dynamic clustering and variational Bayesian meta-learning according to claim 1, characterized in that, The emotion classifier sub-model belongs to a replaceable component, and a module based on a multi-layer convolutional neural network is selected in the method. 3.The emotion recognition method based on dynamic clustering and variational Bayesian meta-learning according to claim 1, characterized in that, The splitting ratio is 60% and 40%.

4. An emotion recognition system based on dynamic clustering and variational Bayesian meta-learning, characterized in that, Comprising: Data arrangement and data preprocessing module, dynamic clustering parameter module based on Chinese restaurant process CRP, meta-parameter module, model parameter updating module, model use module: Step M1: the data arrangement and preprocessing module specifically comprises: Step M1-1: arrange the existing subject electroencephalogram signals, uniformly downsample the electroencephalogram signals to 128HZ, filter the electroencephalogram signals using a 5 to 45HZ band-pass filter, and obtain physiological signal feature data called by the method; Step M1-2: the subject sequence number is t, and the physiological signal feature data sequence set obtained from the subject is denoted as: The real emotion label set of the subject is denoted as: wherein, denotes the multi-channel physiological signals obtained from the subject in a single emotional stimulus experiment, denotes the number of feature channels of the physiological signals, denotes the product of the length and frequency of the physiological signals, i.e. the number of time points at which the physiological signals are finally obtained, denotes the set of real numbers, denotes the number of emotional-labeled physiological signal samples obtained from the t-th subject, and the corresponding emotional label is given in the form of one-hot encoding, and the dimension of the like degree of the emotional annotation in the DEAP dataset divides the emotions of the subjects into two categories: like and dislike, which are represented by the vectors [0, 1] and [1, 0], respectively. Step M1-3: Splitting the set of samples of the subject Splitting the support set samples of the training subject's own emotion classifier and the query set samples for testing The split is according to a random shuffle and a 60% and 40% split, respectively; Step M2: dynamic clustering parameter module based on Chinese restaurant process CRP; Step M2-1: Chinese restaurant process controls the class determination for the new incoming ith subject by setting a hyperparameter a and uses to represent the label of its class, which is given by the following formula: wherein represents the number of clusters the number of subjects in the population expected to be assigned to each cluster, and a > 0 controls the number of clusters as a concentration hyperparameter; Step M2-2: the probability of a certain subject being divided into a certain population category needs to be completed by constructing a broken stick representation for each population category, and according to the principle of Dirichlet process, the formula for the initialization process is: In the formula, represents a probability value vector, wherein the value of each element represents the probability of the individual subject being classified into different population categories, and it is required that the values of the elements in the vector are non-negative and the sum of all values is 1, and the probability of the kth category is represented as ; Step M3: the specific implementation of the meta-parameter module is: Traditional meta-learning assumes that there is a meta-model with a parameter and generates a corresponding emotion classifier using the meta-parameters in the meta-model, and the generation process is: In the formula, represents a Gaussian mixture distribution, represents sampling from the probability value vector to obtain a certain result, i.e., the current subject t is definitely divided into the corresponding population category, and the meta-parameter distribution of the population category to which the subject belongs is determined; represents the population category to which the current subject belongs, represents sampling a corresponding meta-parameter framework from the classifier parameter distribution of the corresponding population, represents a mean vector subject to a Gaussian mixture distribution, represents a semi-positive covariance matrix; Generate an emotion classifier sub-model corresponding to the current subject with initialized parameters, and the specific process is: Step M4: model parameter updating module; Step M4-1: Input the sample data of the t-th subject in a structured inference manner, output the distribution parameters of the four parameters, and the formula 2 is The parameters of the Beta distribution are denoted as The formula 4 is The parameters of the Beta distribution are denoted as The parameters of the Beta distribution are denoted as The formula 5 is The parameters of the Gaussian mixture distribution are denoted as The variational parameters of the task-specific parameters in the formula 6 are denoted as That is, the model parameters can be continuously updated or put into use in an online manner, and the specific formula is: When the tthsubject arrives, the observed lower bound of evidence ELOB is further described according to formula 8: wherein denotes the mathematical expectation, denotes the Kullback-Leibler divergence, which is maximized to optimize the model parameters, q denotes the distribution predicted by the model, and p denotes the true distribution; Step M4-2: Generation and optimization of four distribution parameters: (1) The variational distribution of the ELOB is given by The optimal variational distribution of the ELOB is given by (2) The variational distribution of ELOB, only considering the terms related to The optimal variational distribution of ELOB is derived as: That is, a classifier distribution with parameters: wherein is the variational distribution; (3) the variational distribution of the k-th population class when the t-th subject arrives After that, the meta-learning parameters of the corresponding population are represented as: Assuming that each population category is subject to a parameter: Gaussian distribution, without loss of generality, the new prior knowledge of the tth subject is given by the stored posterior knowledge of the tth subject, i.e. And after the current subject data training is completed, new posterior knowledge is replaced to accept a new subject emotion classification task, and the parameters are updated by means of stochastic gradient descent; (4) the variational distribution, i.e., the generated emotion classifier with the initialization parameters, is given by: wherein, denotes the emotion classifier neural network parameters generated by the method for the current subject, denotes the sampled variational parameters, the variational parameters of the class of the population to which the current subject has been determined to belong for the current subject performing gradient descent over several steps, namely: In the formula, is the step of gradient descent, is the learning rate, and the gradient descent strategy adopts a stochastic gradient descent. Step M5: model putting into use module; When a subject physiological signal sample set with only a small amount of sample labels After the method is passed, the neural network trained in step S3 is called to accurately divide the subject into the corresponding population, and the meta-learning framework trained in the corresponding population is called to generate an emotion classifier with initial parameters for the subject. Through an end-to-end manner, the physiological signal is taken as an input, and an emotion classification result is output. The corresponding small amount of samples are used to quickly train the emotion classifier for use or testing. At this time, the testing step is that when one or a batch of physiological signal samples without emotion labels are passed in, the corresponding physiological signal samples are input into the corresponding emotion classifier, and the emotion prediction output of the meta-parameter model for the physiological signal samples is obtained.

Images