Open-set cross-subject eeg emotion recognition method based on unknown sample likelihood

Abstract

The application discloses an open set cross-subject EEG emotion recognition method based on unknown sample possibility. The application proposes an adversarial domain adaptation model with an auxiliary classifier structure. The model introduces a weighting module for evaluating the probability weight of unknown classes in the target domain, assigns weights to target samples that can better represent whether they belong to known classes and unknown classes, realizes reasonable division of known classes and unknown classes, encourages the model to perform positive transfer in adversarial training, and reduces the domain gap between known classes in the source domain and the target domain. In addition, the application improves feature norm alignment, uses uncertainty entropy of prediction results of target domain samples to constrain feature norm alignment, encourages the model to learn known class features with larger norm in two domains, and avoids degradation of the target domain. The method fully considers alignment on the feature norm level between the target domain and the source domain, has high universality, and improves the generalization ability of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electroencephalogram (EEG) emotion recognition, and proposes an open set domain adaptation method based on unknown sample possibility, which is used in an emotion recognition scene where unknown classes in the source domain exist in the label space of the target domain, solves the open set domain adaptation problem of cross-subject EEG individual difference and cross-domain class imbalance. BACKGROUND

[0002] With the development of artificial intelligence (AI) and brain-computer interface systems (BCIs), the demand for machines to understand human emotional cognition is growing. By analyzing human cognitive state through electroencephalogram (EEG) signals, machines can understand human cognition and objectively feedback emotional state, and even express emotions, which has great significance in the field of brain-computer interaction. Due to the non-stationarity, individual difference and low signal-to-noise ratio of EEG signals, it is difficult to train and use artificial intelligence and brain-computer interfaces based on EEG signals on a large scale. With the introduction of domain adaptation (DA), a large number of cross-subject EEG recognition methods based on domain adaptation have emerged, and the recognition accuracy has been further improved. However, existing domain adaptation methods are limited to EEG signal recognition tasks under class balance. In practical applications, there is still a class imbalance problem between different subjects. Under class imbalance, the model recognition accuracy is greatly affected by the negative transfer phenomenon of the class space, resulting in a decrease in model recognition accuracy. Therefore, a cross-subject EEG emotion recognition method is needed that can match the feature distribution of source domain and target domain samples while avoiding the negative transfer caused by class imbalance to improve the generalization performance of EEG emotion recognition models.

[0003] In order to improve the generalization of cross-subject recognition of physiological signals such as electroencephalogram (EEG) and electrooculogram (EOG), and reduce the performance of individual differences, transfer learning (TL) is currently the most popular research direction in cross-subject emotion recognition. In the definition of transfer learning, different data distributions are referred to as domains, and the training set data is defined as the source domain, and the test set that needs to be transferred is defined as the target domain. The main goal of transfer learning is to reduce domain shift so that the model trained on the source domain data can be successfully applied to the target domain. In the definition of cross-subject emotion recognition, a subject is considered as a domain, and the individual differences between subjects are considered as a domain shift. In the transfer learning method, the domain adaptation method has successfully reduced the influence of individual differences in multiple problems in the field of BCIs. In particular, the introduction of the domain adaptation method based on deep adversarial network architecture has further improved the performance of cross-subject emotion recognition.

[0004] However, current domain adaptation methods usually assume that the class distribution of the source domain is similar to that of the target domain, i.e., closed-set domain adaptation. Or the source domain has private classes that do not exist in the target domain, i.e., partial domain adaptation, but cannot solve the case where the class space of the target domain is greater than that of the source domain. Since the known source data in the electroencephalogram signal field is much less than the unknown target domain data, this open-set domain adaptation method has more application value.

[0005] Existing emotional data sets have fewer emotional labels. When encountering target domain emotional labels that do not exist in the source data set, it can be considered as an open-set emotional recognition task with class imbalance. In this setting, the target domain includes "unknown class" labels and "known class" labels. However, only "known class" labels exist in the source domain. Therefore, the present application will construct an open-set domain adaptation model applicable to a cross-subject open-set emotional recognition algorithm with individual differences and class imbalance. SUMMARY

[0006] To solve this specific open-set domain adaptation setting, existing methods mostly use a fixed threshold to distinguish known and unknown target samples and lack the ability to handle negative transfer, and can lead to a decrease in the recognition rate of known samples. The present application proposes an open-set domain adaptation algorithm based on the possibility of unknown samples applicable to cross-subject open-set emotional recognition, aiming at the problem of negative transfer caused by the existence of unknown classes of source data in target data leading to domain adaptation. The proposed algorithm mainly has two aspects:

[0007] On the one hand, a weighted adversarial network based on the possibility of unknown samples: the open-set domain adaptation based on backpropagation lacks an indicator to quantify the possibility of known and unknown classes, but instead uses a threshold T based on experience to guide the division of known and unknown classes. Therefore, the present application proposes a weighted scheme to separate the known and unknown classes of the target domain based on the structure of the auxiliary classifier, calculates the probability of the target domain belonging to the known class through the auxiliary classifier, and obtains the probability of the unknown class of the target domain from the ordinary classifier, and combines the two to evaluate the weight of the unknown class of the target domain. These weights indicate whether they are more likely to belong to the known class or the unknown class. In the model, the maximum-minimum game target between the generator and the classifier is to align the target samples to the known classes of the source domain or classify them as unknown classes based on the weights.

[0008] On the other hand, a known class compensation strategy based on adaptive feature norm: the degradation of the model in the target domain is mainly due to its smaller feature norm relative to the source domain, and the smaller norm of the target task means that features with less information are generated, leading to domain shift. Therefore, the present application optimizes based on adaptive feature norm, introduces a known class compensation strategy based on adaptive feature norm, and uses the uncertainty entropy of the prediction results of the target domain to constrain the alignment of cross-domain feature norms, encouraging the model to learn known class features with larger norms in both domains to avoid degradation of the target domain.

[0009] In summary, the present application is directed to emotion recognition based on open set cross-subject EEG, and the core technology mainly proposes a new adversarial domain adaptation model with an auxiliary classifier structure. The model introduces a weighted value for evaluating the probability weight of unknown classes in the target domain, assigns weights to target samples that can better represent whether they belong to known classes and unknown classes, realizes reasonable division of known classes and unknown classes, encourages the model to perform positive transfer in adversarial training, and reduces the domain gap between known classes in the source domain and the target domain. In addition, the model improves the feature norm alignment method, uses the uncertainty entropy of the prediction result of the target domain sample to constrain the feature norm alignment, encourages the model to learn known class features with larger norm in the two domains, to avoid degradation of the target domain, and the method fully considers the alignment of the feature norm level of the target domain and the source domain, has high universality, and helps to improve the generalization ability of the model.

[0010] The technical scheme adopted by the present application to overcome the deficiencies of the prior art is as follows:

[0011] The open set domain adaptation method based on unknown sample possibility provided by the present application is suitable for open set cross-subject EEG emotion recognition. The emotional state of the subject is classified by analyzing the features of the EEG emotion data.

[0012] The present application takes the SEED-IV emotion data set as the premise:

[0013] Step 1: Data acquisition

[0014] The SEED-IV data set experiment induces the subject's happy, sad, neutral, and fearful emotions by combining visual and auditory video stimuli. In order to ensure that the subject can fully immerse in the experiment, the video materials are selected from the emotionally intense segments of several Chinese movies. The SEED-IV data set contains 15 subjects, and the experiment uses movie segments containing 4 emotion categories as emotion inducing materials, and each emotion state corresponds to 6 movie segments. Therefore, each subject needs to watch 24 movie materials in the experiment. The 62-channel ESI neural scanning system is used to record the electroencephalogram signal at a sampling rate of 1000Hz.

[0015] Step 2: Data processing

[0016] The electroencephalogram signal of the data set is preprocessed before inputting into the framework. Differential entropy (DE) features are extracted from 5 frequency bands of the SEED-IV data set per second: delta: 1-3Hz, theta: 4-7Hz, alpha: 8-13Hz, beta: 14-30Hz, gamma: 31-50Hz. The feature dimension is 310 (62 channels x 5 frequency bands), and the sample number of each subject is 2505.

[0017] Step 3: Construct and train the EEG emotion recognition model;

[0018] Input: source domain D with labeled data s = {X s , Y s}, number of classes C, target domain data D t = {X t}, number of classes K unknown, and maximum number of iterations T;

[0019] 3-1. Use the generator G f to extract the common features of source domain X s and target domain sample data X t , and map the extracted features to a common feature space;

[0020] 3-2. Optimize the classifier G s with the class label specific to the source domain sample data X y , and the objective function is the supervised classification loss L ce ;

[0021] 3-3. Obtain a pseudo boundary to divide known classes and unknown classes by adversarial training, in order to train the classifier to output P(y = C + 1|x t ) = w t , let the probability that the target domain sample is an unknown class be w t , train the generator to deceive the classifier to maximize the error of the classifier, the generator will choose to increase the probability (regard the sample as an unknown class) or reduce the probability (align the sample with the source domain) to maximize the error of the classifier, use the binary cross-entropy loss function, and the objective function is L adv ;

[0022] 3-4. The auxiliary classifier structure proposed in the present application includes an auxiliary source classifier an auxiliary domain classifier and a classifier G y . The structure is used to evaluate the likelihood of unknown classes in the target domain, and calculate the weight probability of each target sample

[0023] 3-5. The auxiliary source classifier uses a leaky-softmax function to predict the source class label, which can make the probability sum less than 1. Convert the features output by the generator G f to C-dimensional probability;

[0024] 3-6. Since the auxiliary source classifier is trained only by source domain data, use When classifying target domain data, lower probability values and uncertain prediction results are obtained. The probability of each target sample belonging to the known classes of the source domain can be calculated by this feature;

[0025] 3-7. To comprehensively evaluate the possibility of unknown target samples, the output of the classifier G y can be further utilized, as it contains the probability of the unknown class P(y = C+1|x). Finally, the weight containing information about the known classes of the source and the unknown classes of the target is obtained, and the lower the value of , the more likely it is to belong to the unknown class;

[0026] 3-8. Train the auxiliary source classifier using the one-versus-all binary loss function L aug_c At the same time, train the auxiliary domain classifier using the loss function L aug_adv ;

[0027] 3-9. Introduce known class compensation based on adaptive feature norm. The adaptability of the known class feature norm is compensated by the uncertainty entropy of the target domain prediction, and the influence of unknown samples on training is avoided. The classifier is used to generate target domain prediction probability, and all target samples with unknown prediction labels are filtered;

[0028] 3-10. Add a weight value to each target sample . This weight value is calculated based on the uncertainty entropy H of the target domain prediction value, which quantifies the uncertainty of the prediction;

[0029] 3-11. The uncertainty entropy H is normalized to be within [0, 1]. The smaller the entropy, the more accurate the prediction, and the larger the corresponding weight value, which is represented by γ i ;

[0030] 3.12. Obtain the weighted feature norm alignment loss function L wnfa ;

[0031] 3-13. Perform end-to-end training by a max-min optimization process to unify the unknown sample possibility evaluation and the weighted feature norm alignment. The parameters of G f , G y , are defined as θ f , θ y , In adversarial training, the gradient reversal layer GRL is used, so G f , G y can be optimized simultaneously;

[0032] 3-14. Iterate the model T times according to the above steps.

[0033] ​Output: Target domain sample prediction label.

[0034] The present application has the following advantages:

[0035] Firstly, a new problem in the field of emotion recognition is discovered and solved, that is, the open set domain adaptation problem in unsupervised domain adaptation. The present application proposes an open set cross-subject EEG emotion recognition method based on unknown sample possibility, which is suitable for open set emotion recognition. Research shows that many existing methods use a fixed threshold to divide the known classes and unknown classes of the target domain, but may not be able to cope with complex data sets, resulting in negative transfer phenomenon and reducing the recognition rate of known classes. Therefore, the present application starts from the possibility of unknown classes in the target domain, proposes an unknown class possibility weighting scheme, fully considers the information of unknown classes and known classes in the target domain, realizes the separation of known classes and unknown classes through generator and classifier adversarial learning. In addition, through the known class compensation strategy based on adaptive feature norm, the cross-domain known class alignment is promoted, and the performance decline of known class recognition is avoided. Finally, the present application effectively solves the open set emotion recognition problem based on EEG, can be applied to cognitive state recognition based on EEG under any task, has strong generalization ability, and can be well applied to clinical diagnosis and practical application. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is the model structure diagram of the present application.

[0037] Figure 2 is the auxiliary source classifier structure proposed by the present application. DETAILED DESCRIPTION

[0038] The present application will be further described below in combination with the drawings and examples.

[0039] The present application discloses an open set cross-subject EEG emotion recognition method based on unknown sample possibility. Based on the disclosed SEED-IV emotion recognition data set, the present application proposes an adversarial domain adaptation model with an auxiliary classifier structure for open set cross-subject EEG emotion recognition. The model introduces a weighting value for evaluating the probability weight of unknown classes in the target domain, assigns weights to target samples that can better represent whether they belong to known classes and unknown classes, realizes reasonable division of known classes and unknown classes, encourages the model to perform positive transfer in adversarial training, and at the same time reduces the domain gap between known classes in the source domain and the target domain. In addition, the present application improves the feature norm alignment, uses the uncertainty entropy constraint of the prediction result of the target domain sample to constrain the feature norm alignment, encourages the model to learn known class features with larger norm in two domains to avoid degradation of the target domain. The method fully considers the alignment of feature norms between the target domain and the source domain, has high universality, and improves the generalization ability of the model.

[0040] AsFigure 1 The structure diagram of the open-set cross-subject EEG emotion recognition method based on unknown sample possibility is shown, and mainly includes the following steps:

[0041] Step 1: Data preprocessing

[0042] The electroencephalogram signals of the data set are preprocessed before being input into the framework. Differential entropy (DE) features are extracted from 5 frequency bands of the SEED-IV data set every second: δ: 1-3 Hz, θ: 4-7 Hz, α: 8-13 Hz, β: 14-30 Hz, and γ: 31-50 Hz. The feature dimension is 310 (62 channels x 5 frequency bands).

[0043] For a specific length of EEG signal approximately obeying Gaussian distribution , the differential entropy is:

[0044]

[0045] which is equal to the logarithm of the energy spectrum in a specific frequency band.

[0046] Step 2: Data definition

[0047] Given N EEG data with individual differences, all subject individuals are taken as source domains and their labels are known, and the source domain is represented as D s ={X s ,Y s}, the source domain label is represented as Y s , the number of categories is C, and the target domain is represented as D t ={X t}, and the number of categories K is unknown.

[0048] Step 3: Open-set cross-subject EEG emotion recognition method based on unknown sample possibility

[0049] Input: source domain sample data D s ={X s ,Y s} with emotion classification labels, the number of categories C, target domain sample data D t ={X t} without emotion state labels (the number of categories K is unknown), and the maximum number of iterations T;

[0050] 3-1. Use the generator G f to extract the latent common features of the source domain and target domain sample data, and map the extracted features to a common feature space; the generator G f uses a neural network with 3 hidden layers, each hidden layer has 512 nodes, and uses a ReLU activation function, and the output of the network is 128 dimensions;

[0051] 3-2. With source domain sample data X s as constraints to optimize the classifier G y , the network layer number of the classifier G y is set to 3 layers. Each hidden layer of the network has 64 nodes, and ReLU activation function is used. The input of the network is 128 dimensions, and the output is C+1 dimensions. The regular cross-entropy loss function is used as the classification loss L ce :

[0052]

[0053] where C is the number of source domain categories. The present application adds an unknown class to the training. When the classifier predicts the result as C+1 class, it is considered that the sample is an unknown class.

[0054] 3-3. Divide the known class and the unknown class by obtaining a pseudo boundary through adversarial training. In order to train the classifier to output P(y=C+1|x t ) = w t , the weight probability that the target domain sample is an unknown class is w t , the generator is trained to deceive the classifier to maximize the error of the classifier. The generator will select to increase the probability (regard the sample as an unknown class) or reduce the probability (align the sample with the source domain) to maximize the error of the classifier. The binary cross-entropy function is used as the loss function, so the loss function L adv is as follows:

[0055]

[0056] where n t represents the number of target samples , represents the i-th target sample; represents the weight probability that the i-th target sample is an unknown class.

[0057] 3-4. The auxiliary classifier structure is as shown in Figure 2 . The purpose is to evaluate the possibility of the unknown class of the target domain and calculate the weight probability The structure includes an auxiliary source classifier and an auxiliary domain classifier , wherein and have the same structure, including three fully connected layers, and using ReLU activation function and Dropout function after the first two fully connected layers;

[0058] 3-5. The auxiliary source classifier The leaky-softmax function is used to predict the source class label, ensuring that the sum of probabilities is less than 1. (Auxiliary source classifier) Generator G f The output feature z is converted into C-dimensional class probabilities:

[0059]

[0060] Among them, z c Represents generator G f The output features belonging to class c;

[0061] 3-6. Since the auxiliary source classifier is trained only on source domain data, using it to classify target domain data will result in lower probability values ​​and uncertain predictions. Based on this characteristic, the calculation of each target sample... Probability of belonging to a known class:

[0062]

[0063] 3-7. To comprehensively assess the possibility that the target sample is unknown, G can be further utilized. y The output result is obtained because it includes the probability P(y=C+1|x) of the unknown class. Finally, the weight w, which includes information about the known source class and the unknown target class, is obtained. t For the target sample, w t The lower the value, the more likely it is to belong to the unknown class for each target sample. weight The calculation formula is as follows:

[0064]

[0065] in, Represents target domain samples Probability of belonging to the unknown class

[0066] 3-8. Using a one-to-many binary loss function L aug_c Train the auxiliary source classifier, and simultaneously train the auxiliary domain classifier, with a loss function of L. aug_adv ;

[0067]

[0068]

[0069] Where, n s Indicates the number of samples in the source domain. Represents source domain samples The true label, Indicates sample The probability that the target sample belongs to the c-th class.

[0070] 3-9. Known class compensation based on adaptive feature norm is introduced, the adaptability of known class feature norm is compensated by the uncertainty entropy of target domain prediction, and the influence of unknown samples on training is avoided. The target domain prediction probability is generated using the classifier, and all target samples with unknown prediction results are filtered.

[0071] 3-10. For each target sample A weight value is added. The weight value is calculated according to the uncertainty entropy H of the target domain prediction value, which quantifies the uncertainty of the prediction, and the uncertainty entropy calculation formula is as follows:

[0072]

[0073] Wherein, is the probability that the target sample belongs to the c-th class, which is calculated by the Softmax function.

[0074] 3-11. The uncertainty entropy H is limited to [0, 1] by normalization. The smaller the entropy, the more accurate the prediction, and the greater the weight value, which is represented by γ i , so the weight is defined as follows:

[0075]

[0076] 3.12. Finally, the weighted feature norm alignment loss function L wnfa is obtained, and the formula is as follows:

[0077]

[0078] Wherein, represents the number of target domain samples excluding unknown classes, represents the target domain excluding unknown classes, D s represents the source domain, and h(.) represents the function operator existing in the network, such as G f (x). The dis function is an L-2 distance function, and λ is a hyperparameter. Δr represents the residual scalar that controls the expansion of the feature norm. θ0 and θ represent the updated model parameters in the last iteration and the current iteration.

[0079] 3-13. The present application can be trained end-to-end, and the unknown sample possibility evaluation and weighted feature norm alignment are unified through a maximum and minimum optimization process, and the parameters of G f , G y , are defined as θ f , θ y , In the adversarial training, a gradient reversal layer GRL is used, so the synchronous optimization G f 、G y The specific optimization process is as follows:

[0080]

[0081]

[0082]

[0083] Output: target domain sample predicted label.

[0084] Embodiments:

[0085] In order to prove the effectiveness of the proposed application, in the experiment, leave-one-subject cross-validation is carried out to study the cross-subject emotion recognition performance. Specifically, in the validation iteration process, each iteration selects one subject as the target domain (Target), and the other 14 subjects as the source domain (Source). Since the problem solved by the application is the class imbalance cross-subject emotion recognition under the open set condition, for the SEED-IV dataset, the first two classes are set as known (Known) classes, and the remaining classes are set as unknown (Unknown) classes.

[0086] The application performs open set experiments on the SEED-IV dataset, and the comparison experiments are shown in Table 1. Each row gives the performance of each comparison method, including OS, OS* and Unknown three performance indicators. Among them, OS represents the average accuracy (%) of known and unknown classes, OS* represents the average accuracy (%) of known classes, Unknown represents the accuracy (%) of unknown classes, and H-score represents the harmonic mean of OS* and Unknown.

[0087] Table 1 Open set setting cross-subject emotion recognition accuracy results (SEED-IV)

[0088]

[0089]

[0090] In addition, the application mainly consists of two components: an adversarial loss L adv based on the possibility of unknown samples wfna and a weighted feature norm alignment loss L To verify the effectiveness of each component, an ablation experiment is performed, and the ablation experiment results are shown in Table 2. The results in the table are the average results obtained by leave-one-subject validation.

[0091] Table 2 Ablation experiment results

[0092] Comparative method w t ]]> [[ L adv ]]> [[ L wfna ]]> OS OS* Unknown H-score Source only 44.73 67.10 0.00 0.00 OSBP √ 62.22 74.00 38.66 50.79 UPOSDA (wo adv) √ 61.80 77.1 31.2 44.42 UPOSDA (wo wnfa) √ √ 67.22 80.8 40.07 53.57 UPOSDA √ √ √ 70.81 78.48 55.46 64.99

[0093] Finally, it is to be understood that the embodiments are for purposes of illustration only and that various changes and modifications can be made by those skilled in the art without departing from the scope of the application as disclosed in the specification and as defined in the appended claims. Therefore, the scope of the application is not to be limited to the embodiments disclosed herein but is to be accorded the full scope permissible by the appended claims and legal equivalents thereof.

Claims

1. An open-set cross-subject EEG emotion recognition method based on unknown sample possibility, characterized in that Comprising the following steps: Step 1: data collection; Select the SEED-IV dataset, which contains 15 subjects, and use movie clips containing 4 emotional categories as emotional induction materials in the experiment, and each emotional state corresponds to 6 movie clips; therefore, each subject needs to watch 24 movie materials in the experiment; Step 2: data preprocessing and definition; Step 3: build and train the EEG emotion recognition model; Input: source domain sample data with emotion classification labels , number of classes , for source domain sample data, for source domain sample data emotion classification labels; target domain sample data without emotion state labels , number of classes unknown, for target domain sample data, and maximum number of iterations T; 3-1, with a generator extracting source domain sample data and target domain sample data latent common features, mapping the extracted features to a common feature space; 3-2, with source domain sample data The unique class label is a constraint optimization classifier The objective function is a supervised classification loss ; 3-3, get a pseudo-boundary by adversarial training to divide known classes and unknown classes, in order to train the classifier output , the probability that the target domain sample is an unknown class is , train the generator to deceive the classifier to maximize the error of the classifier, the generator will choose to increase the probability or reduce the probability to maximize the error of the classifier, using binary cross-entropy loss function, the objective function is ; 3-4, the auxiliary classifier structure includes an auxiliary source classifier , an auxiliary domain classifier , and a classifier ; the possibility of the unknown class of the target domain is evaluated by the auxiliary classifier structure, and the weight probability of each target sample is calculated ; wherein and The structure is the same, containing three full connection layers, and using Relu activation function and Dropout function after the first two full connection layers; 3-5, auxiliary source classifier A leaky-softmax function is used to predict the source class label, which allows the probabilities to sum to less than 1; The generator The output features are converted to dimensional probabilities; 3-6, due to the auxiliary source classifier Only trained by source domain data, therefore using By this property, the probability that each target sample belongs to the known classes of the source domain is calculated, as the lower probability values and uncertain predictions are obtained when classifying the target domain data. 3-7, classifier The output result contains the probability of unknown class Finally, the weight containing the source known class information and the target unknown class information is obtained, and the lower the value of the more likely it belongs to the unknown class. 3-8, by a pair of many two-division loss function training an auxiliary source classifier ; at the same time, training an auxiliary domain classifier, the loss function is ; 3-9, introduce known class compensation based on adaptive feature norm, compensate the adaptability of known class feature norm through the uncertainty entropy of target domain prediction, and avoid the influence of unknown samples on training; use the classifier to generate target domain prediction probability, and filter all target samples whose predicted labels are unknown; 3-10, for each target sample adding a weight value, which is calculated according to the uncertainty entropy H of the target domain prediction value, and quantifies the uncertainty of the prediction; 3-11, the uncertain entropy H is normalized in [0, 1]; the smaller the entropy is, the more accurate the prediction is, and the greater the corresponding weight value is, and the weight value is represented by . 3-12, obtaining a weighted feature norm alignment loss function ; 3-13. Perform end-to-end training, align the unknown sample likelihood evaluation with the weighted feature norm through a max-min optimization process, and respectively define the parameters of , , , , ; in the adversarial training, the gradient reversal layer GRL is used, so that , ;​ 3-14, iterate the model T times according to the above steps; Output: target domain sample prediction label.

2. The open-set cross-subject EEG emotion recognition method based on unknown sample possibility according to claim 1, characterized in that Step 2 is implemented as follows: Differential entropy DE features were extracted from 5 frequency bands of SEED-IV dataset per second: : 1-3 Hz, : 4-7 Hz, : 8-13 Hz, : 14-30 Hz, : 31-50 Hz; feature dimension is 310; For a given length of an approximately Gaussian distributed EEG signal, its differential entropy ​ Equal to the logarithm of its energy spectrum in a certain frequency band.

3. The open-set cross-subject EEG emotion recognition method based on unknown sample possibility according to claim 1 or 2, characterized in that Step 2 data definition is as follows: Given N EEG data with individual differences, all subjects are taken as source domains and their labels are known, which is represented as , the source domain label is represented as , the number of categories , and the target domain is represented as , the number of categories is unknown.

4. The open-set cross-subject EEG emotion recognition method based on unknown sample possibility according to claim 3, characterized in that Cross-entropy loss function as classification loss : （2） Wherein, C is the number of source domain categories, and the unknown class is added to the training as a class, and when the classifier prediction result is C+1 class, it is considered that the sample is an unknown class; Binary cross-entropy loss function As follows: （3） in, Indicates target sample Quantity, Indicates the first One target sample; Indicates the first Each target sample represents the weighted probability of an unknown class.

5. The open-set cross-subject EEG emotion recognition method based on unknown sample possibility according to claim 4, characterized in that Auxiliary source classifier Generators Output features Convert to C-dimensional class probabilities: （4） wherein, a generator outputted belonging to a class of features; each target sample probability of belonging to a known class: （5） each target sample weight of each target sample is calculated according to the following formula: （6） wherein, representing target domain samples probability of belonging to unknown class .

6. The open-set cross-subject EEG emotion recognition method based on unknown sample possibility according to claim 5, characterized in that By a one-to-many binary loss function The auxiliary source classifier is trained, and at the same time, the auxiliary domain classifier is trained, and the loss function is ; + （7） （8） wherein, denotes the number of source domain samples, denotes the source domain sample true label, denotes the probability that a sample belongs to class c.

7. The open-set cross-subject EEG emotion recognition method based on unknown sample possibility according to claim 6, characterized in that The weight value is calculated according to the uncertainty entropy H of the target domain prediction value, and the uncertainty entropy calculation formula is as follows: （9） wherein, is the target sample is the probability of belonging to the c-th class, computed by the Softmax function; The weight values are represented by so that the weights are defined as follows: （10）。 8. The open-set cross-subject EEG emotion recognition method based on unknown sample possibility according to claim 7, characterized in that The resulting weighted feature norm alignment loss function is given by the following equation: （11） where, denotes the number of target domain samples excluding the unknown class, denotes the target domain excluding the unknown class, denotes the source domain, (.) denotes a function operator present in the network; the dis function is an L-2 distance function, is a hyper-parameter; denotes a residual scalar controlling the expansion of feature norms; and denotes the model parameters updated in the last and current iteration.

9. The open-set cross-subject EEG emotion recognition method based on unknown sample possibility according to claim 8, characterized in that Synchronization optimization ,The process of is shown below: （12） 。

Images