Cross-domain electroencephalogram emotion recognition method for progressive target tested sample mining

Abstract

The invention discloses a cross-domain electroencephalogram emotion recognition method for progressive target tested sample mining, which comprises the following steps of: firstly, collecting electroencephalogram data in different emotion inducing scenes, preprocessing, extracting characteristics of different channels and frequency bands, and generating source domain and target domain data; secondly, by aligning marginal distribution and conditional distribution of a source domain and a target domain, a target domain characteristic label matrix is determined based on similarity, and a sample confidence coefficient weight matrix is obtained; secondly, measuring an error between a sample and a characteristic label by using cross entropy, and constructing a target function by combining a sample confidence coefficient weight matrix, marginal distribution and a conditional distribution alignment error; and finally, according to the target function, carrying out joint iterative optimization, and according to the target domain characteristic tag matrix after iterative optimization, judging the emotion category of the testee during electroencephalogram data acquisition. According to the method, the overall performance of the model is effectively optimized, the global optimality of the result is ensured, and more accurate electroencephalogram emotion recognition is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram signal processing, and specifically refers to a cross-domain electroencephalogram emotion recognition method, device and storage medium for progressive target subject sample discovery. Background Technique

[0002] Emotion is a complex mental state that plays a crucial role in interpersonal communication. It is an individual's natural reaction to external stimuli or internal mental activities, usually occurring spontaneously in an unconscious manner rather than being actively controlled by the will. The occurrence of emotion involves physiological activities of the central nervous system and the peripheral nervous system, demonstrating multi-level complexity. Generally, emotional states can be presented through external behaviors such as subjective experiences, verbal expressions, facial expressions, and body movements. However, these external manifestations are easily influenced by cultural backgrounds, individual intentions, or social environments and may be deliberately concealed or disguised, resulting in uncertainties and inaccuracies in emotion recognition.

[0003] In contrast, electroencephalogram (EEG), as a non-invasive neuroimaging technology, can directly capture the neural activities of the brain and provide more subtle and accurate physiological signals for the objective assessment of emotional states. EEG signals are spontaneously generated by the brain during activities, featuring high temporal resolution, low cost, high reliability, and relatively high emotion recognition effects. Therefore, by analyzing EEG signals in emotional states, emotional changes can be effectively recognized, thus playing an important role in mental health monitoring and early warning of emotional abnormalities. In addition, in human-computer interaction systems, EEG-based emotion recognition technology enables machines to more accurately understand users' emotional needs and preferences, thereby providing a more personalized and adaptive interaction experience.

[0004] Currently, the common semi-supervised analysis methods mainly have the following five defects:

[0005] 1. Common semi-supervised learning methods may have the risk of spreading errors in the initial stage, leading the model to gradually deviate in the wrong direction.

[0006] 2. Most existing models learn domain-invariant features sequentially and estimate target domain label information. This two-stage strategy breaks the internal connection between these two processes and inevitably leads to sub-optimality.

[0007] 3. Different individuals may have significant differences in emotional responses to the same event, which reflects the individual differences in emotional experiences and expressions. This difference directly affects the effectiveness of samples in emotion classification models. However, most existing electroencephalogram emotion classification models often assign equal importance to all collected sample data and do not fully consider the differences in the quality and reliability of the collected data.

[0008] 4. In practical applications, due to the small amount of emotional EEG signal data with clear labels, the analysis results have certain limitations. And it is often challenging to obtain a large amount of emotional EEG signal data with clear labels.

[0009] 5. Many emotion classification methods rely on deep learning techniques, such as convolutional neural networks, recurrent neural networks, and transformer models. Although these models perform excellently in terms of performance, due to their complex internal mechanisms, they are usually regarded as "black boxes", and it is difficult to intuitively understand their decision-making processes. Summary of the Invention

[0010] In view of the deficiencies of the prior art, the present invention proposes a cross-domain EEG emotion recognition method for progressive target subject sample discovery, which can make full use of the emotional representation of EEG data. The model is based on the cross-domain adaptation of joint optimization of the shared subspace double projection matrix and the target label. By constructing a feature alignment constraint through the bidirectional projection matrices of the source domain and the target domain, the target domain features are mapped to the source domain subspace, and high-confidence pseudo-labels are generated based on the source domain class prototypes, so as to alleviate the domain shift problem in the early stage of training; further introduce an adaptive sample weight mechanism, dynamically quantify the sample transferability according to the confidence of the pseudo-labels in the iteration, and focus on the highly credible samples through the weighted classification loss to suppress noise interference. Finally, the model integrates the subspace alignment loss, the weighted classification loss, and the regularization term to realize the collaborative optimization of the projection matrix and the classifier. This strategy effectively combines accurate label estimation and label measurement, and can accurately evaluate and utilize the unique EEG features of the target domain to reduce the impact of error propagation in the initial stage, thereby improving the accuracy of the model.

[0011] To solve the above technical problems, the technical solution of the present invention is as follows:

[0012] A cross-domain EEG emotion recognition method for progressive target subject sample discovery, which includes the following steps:

[0013] Step 1: Under different emotion induction scenarios, collect the EEG data of multiple subjects, preprocess the collected EEG data, and extract the features of different channels and frequency bands of the preprocessed EEG data to generate source domain and target domain data.

[0014] Step 2: By aligning the marginal distribution and conditional distribution of the source domain and the target domain, calculate the distance-based similarity between each target domain sample in the subspace and the source domain subspace prototype, and determine the target domain characteristic label matrix F t , and according to F t Calculate the sample confidence weight matrix U using the left-truncated Gaussian distribution, that is, the sample transferability metric.

[0015] Step 3: Use cross-entropy to measure the error between the sample and the characteristic label, and then jointly align the sample confidence weight matrix U, the marginal distribution, and the conditional distribution alignment error to construct the objective function as follows:

[0016]

[0017] Among them, represents the source domain sample set, and n s represents the total number of source domain samples. represents the target domain sample set, and n t represents the total number of target domain samples, where n = n t + n s ; represents the estimated sample size of the c-th class, N s , N t is a diagonal matrix, and its diagonal elements are and are the projection matrices for the source domain and the target domain respectively, and k is the dimension of the projection space; C is the number of sample categories. represents the number of samples belonging to the c-th class in the source domain, represents the number of samples belonging to the c-th class in the target domain; represents the classification matrix, b = [b1, b2,... b n represents the offset vector, represents the label information of the source domain, F t is the target domain characteristic label matrix, sample confidence weight matrix where v i represents the confidence weight of the i-th sample; α, β, and λ all represent adaptive weights, and 1 represents a vector of all 1s.

[0018] Step 4: According to the constructed objective function, jointly iteratively optimize the source domain and target domain projection matrices P s , P t , the classification matrix W, the offset vector b, the target domain characteristic label matrix F t , and the sample confidence weight matrix U, so as to judge the emotional category of the subject during EEG data acquisition according to the iteratively optimized target domain characteristic label matrix F t .

[0019] Preferably, since EEG data is multi-rhythm and multi-channel. The features extracted from different frequency bands are concatenated to form a feature vector. Suppose an EEG signal dataset with 64 channels is used, and the Fz and Cz electrodes are selected as reference electrodes. All signals are collected in accordance with the international 10-20 system for standardization.

[0020] Preferably, the specific process of preprocessing and feature extraction of EEG data in step 1 is as follows: The EEG data is downsampled and then the noise and electrooculogram artifacts are filtered out by a band-pass filter. The filtered data is divided into 5 common EEG frequency bands, namely Delta (1 - 4 Hz), Theta (4 - 8 Hz), Alpha (8 - 14 Hz), Beta (14 - 31 Hz) and Gamma (31 - 50 Hz). On the basis of this data, the differential entropy features of different channels and different frequency bands are extracted.

[0021] Preferably, the specific process of constructing the objective function in step 3 is as follows: The Euclidean distance is used to measure the alignment error of the subspaces respectively mapped by the source domain and the target domain. Considering the marginal distribution and conditional distribution for each class, a construction model can be obtained

[0022]

[0023] The characteristic label matrix and the sample confidence matrix are obtained through distribution alignment, and finally the objective function formula (1) is obtained by combining the cross-entropy loss of the credible samples and the labels.

[0024] Preferably, the method of joint iterative optimization in step 4 is as follows:

[0025] Step 4.1: Initialize the projection matrices P s 、P t , the classification matrix W, and the offset vector b.

[0026] Step 4.2: Update the projection matrices P s 、P t by the least squares method: Fix the classification matrix W, the offset vector b, and the target domain characteristic label matrix F t 、the sample confidence weight matrix U, and the objective functions for obtaining P s 、P t are

[0027]

[0028] Solving equation (3) can update the projection matrices P s 、P t .

[0029] Step 4.3: Update the target domain characteristic label matrix F t : Update the target domain characteristic label matrix F according to step 2 t .

[0030] Step 4.4: Update the sample confidence weight matrix U: Update the sample confidence weight matrix U according to step 2.

[0031] Step 4.5: Update the classification matrix W by least squares: Fix the projection matrices P s and P t , the offset vector b, and the target domain feature label matrix F t , and the sample confidence weight matrix U, and the objective function of W is obtained as

[0032]

[0033] Solving Equation (4) can update the classification matrix W.

[0034] Step 4.6: Update the offset vector b by least squares: Fix the projection matrices P s and P t , the classification matrix W, the target domain feature label matrix F t , and the sample confidence weight diagonal matrix U, and the objective function of b is obtained as

[0035]

[0036] Solving Equation (5) can update the offset vector b.

[0037] Step 4.7: Repeat steps 4.2, 4.3, 4.4, 4.5, and 4.6 until the objective function converges, and finally complete the joint iterative optimization of the projection matrices P s and P t , the classification matrix W, the offset vector b, the target domain feature label matrix F t , and the sample confidence weight matrix U.

[0038] Preferably, the sample confidence weight matrix U is a diagonal matrix, and v i represents the confidence weight of the i-th sample.

[0039] The present invention has the following characteristics and beneficial effects:

[0040] 1. The present invention proposes a cross-domain EEG emotion recognition method for progressive target subject sample discovery. This method realizes the global improvement of model performance by jointly optimizing the shared subspace projection matrix and the target pseudo-label. The loss of the target label is calculated by combining the in-domain and inter-domain estimations, thereby effectively optimizing the overall performance of the model and ensuring the global optimality of the results.

[0041] 2. The present invention provides a mechanism for accurately evaluating target domain labels based on aligned subspace class prototypes. Specifically, by analyzing the similarity between target domain subspace samples and source domain subspace class prototypes, more accurate pseudo-labels are generated. On this basis, interactive learning is combined with the introduction of a pseudo-label reliability matrix to improve the overall performance and robustness of the model. On the one hand, a more correct reliability matrix can be obtained after accurately estimating the pseudo-labels. On the other hand, obtaining a more correct reliability matrix can promote the alignment of subspace class prototypes, thereby obtaining a more accurate pseudo-label estimate. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0043] Figure 1 It is a framework diagram of a cross-domain EEG emotion recognition method for progressive target subject sample discovery in this embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0045] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings.

[0046] On the contrary, the present invention covers any alternatives, modifications, equivalent methods and solutions made within the spirit and scope of the present invention defined by the claims. Further, in order to enable the public to have a better understanding of the present invention, in the following detailed description of the present invention, some specific details are described in detail. Those skilled in the art can fully understand the present invention without the description of these details.

[0047] Embodiment 1

[0048] This embodiment provides a cross-domain EEG emotion recognition method for progressive target subject sample discovery, as Figure 1 shown, specifically including the following steps:

[0049] Step 1: Collection and preprocessing of emotional EEG data;

[0050] During the experiment, the subjects need to wear an EEG cap to record the EEG signals generated under emotional stimuli and store them as raw data. After the EEG signal acquisition is completed, the subjects need to subjectively evaluate their own emotional experiences and label the corresponding EEG signals according to their self-perceived emotional states. Subsequently, preprocessing and feature extraction are performed on the emotional EEG data. In this embodiment, first, the obtained EEG data is downsampled, and the sampling rate is adjusted to 200 Hz. Subsequently, the data is processed using a band-pass filter with a set frequency range of 1 Hz to 50 Hz to reduce the interference of noise and eye movement artifacts. After filtering, the signal is divided into five different frequency bands, including Delta (1 - 4 Hz), Theta (4 - 8 Hz), Alpha (8 - 14 Hz), Beta (14 - 31 Hz), and Gamma (31 - 50 Hz). Based on these processed data, the differential entropy features of each channel in different frequency bands are further calculated.

[0051] Step 2: Calculate the distance-based similarity by aligning the marginal and conditional distributions of the source domain and the target domain to determine the target domain characteristic label matrix F t , and based on F t Calculate the sample confidence weight matrix U using the left-truncated Gaussian distribution.

[0052] Step 2.1: Obtain the feature projections of the source domain and the target domain respectively to obtain their representations in the common subspace and On this basis, calculate the class center values of the source domain subspace features Z s , and these class centers represent the prototypes of each category in the subspace. Next, for each sample X (j) in the target domain, calculate its subspace feature and the distances to each source domain prototype. To quantify this distance relationship and convert it into a similarity score, take the reciprocal of each distance value, so that a smaller distance corresponds to a higher similarity score. Based on the above similarity scores, a characteristic label matrix F t is constructed, and this matrix is used to assign pseudo-labels to each sample in the target domain. Specifically, for each target domain sample X (j) , determine the membership probability of the j-th sample to the c-th class according to its similarity score to the c-th class prototype

[0053]

[0054] where is an indicator function that returns 1 if and returns 0 otherwise.

[0055] Step 2.2: To further evaluate the confidence of each sample, a method based on the left-truncated Gaussian distribution is introduced. Specifically, for each sample According to the feature label matrix F t Obtain its probability for the most likely class And use the left-truncated Gaussian distribution F to calculate the confidence v of this sample j , that is, the j-th element in the diagonal matrix U.

[0056]

[0057] Among them, erf(x) represents the error function, μ is the average of the probabilities of the most likely classes of all samples, and σ is the variance.

[0058] Step 3: Use cross-entropy to measure the error between the sample and the feature label, and then jointly align the error of the sample confidence weight matrix U, the marginal distribution, and the conditional distribution to construct the objective function of the cross-domain EEG emotion recognition model for progressive target subject discovery.

[0059] Use the method based on the maximum mean discrepancy criterion to align the source data X s And the target data X t The marginal distributions in the latent subspace, as shown in Equation (4)

[0060]

[0061] Among them, P s Represents the latent subspace of the original data, P t Represents the latent subspace of the target data, X s，i Represents the i-th source domain sample, X t，i Represents the i-th target domain sample.

[0062] Is equivalent to Equation (5):

[0063]

[0064] Among them, Is a full single-column vector of length n s (n t ).

[0065] Similarly, the conditional distributions of the aligned source data and target data can also be achieved by minimization:

[0066]

[0067] After simplification, it can be obtained:

[0068]

[0069] Among them is an unknown label metric matrix in the target domain, which is obtained through distribution alignment and satisfies non-negativity and row normalization constraints. For example, if F t is [0.11, 0.15, 0.39, 0.35], the probabilities that the i-th target sample belongs to four emotional states are 0.11, 0.15, 0.39, and 0.35 respectively.

[0070] Since the target samples are unlabeled, the true sample size of each class cannot be obtained. Therefore, we use to estimate the true sample size. N s (N t ) is a diagonal matrix, whose diagonal elements are

[0071] By jointly considering the marginal distribution and the conditional distribution, we obtain

[0072]

[0073] where the second term restricts the two projection matrices from deviating too much.

[0074] The data represented in the subspace is combined with the label information obtained through the subspace information by the classification matrix W, and the regularization of W is added to prevent the model from overfitting to obtain

[0075]

[0076] On this basis, to reduce the risk of propagation errors in the initial stage, the confidence weights of each sample are jointly considered; thus, the model of the cross-domain EEG emotion recognition method for progressive target subject sample discovery is obtained

[0077]

[0078] where, represents the source domain sample set, n s represents the total number of source domain samples, represents the target domain sample set, n t represents the total number of target domain samples, n = n t + n s ; and are the projection matrices for the source domain and the target domain respectively, k is the dimension of the projection space; c is the number of sample categories, represents the number of samples belonging to the c-th category in the source domain, represents the number of samples belonging to the c-th category in the target domain; represents the label information of the source domain, represents the classification matrix, b = [b1, b2,... b nDenote the offset vector; v i Denote the confidence weight of the i-th sample; β represents the adaptive weight, and 1 represents a vector of all 1s.

[0079] Generally speaking, by combining Equation (8) and Equation (10), the objective function (11) of the cross-domain EEG emotion recognition model for progressive target subject sample discovery is finally obtained:

[0080]

[0081] Step 4. According to the constructed objective function, for the source domain and target domain projection matrices P s , P t , the classification matrix W, the offset vector b, and the target domain feature label matrix F t , and the sample confidence weight matrix U, perform joint iterative optimization;

[0082] Step 4.1. Before optimizing the objective function, all these data need to be initialized. Among them, all elements of the sample confidence weight matrix U are initialized to 1 when initialized; the target domain feature label matrix F t is initialized with each element being The source domain and target domain projection matrices P s , P t , the classification matrix W are all initialized according to principal component analysis; the offset vector b is randomly initialized.

[0083] Step 4.2. Update the source domain and target domain projection matrices P s , P t : Fix the classification matrix W, the offset vector b, the target domain feature label matrix F t , and the sample confidence weight matrix U, and obtain the objective function of P s , P t as

[0084] For the convenience of solution, let M = [I d , -I d ; -I d , I d , where I d is a d×d identity matrix, and N s (N t ) is a diagonal matrix, and the c-th diagonal element is Then denote is rewritten as λtr(P T TP) + αtr(P T MP), where The same is true in the following discussion.

[0085] Let \(G = UX\) T 、\(H = U1\), \(J = UY\), and by solving, the objective function of \(P\) is:[[]]

[0086]

[0087] Take the partial derivative of \(P\) and set it to 0, we get:[[]]

[0088]

[0089] After simplification, we get:[[]]

[0090] \((G T G)^{-1}(\lambda T+\alpha M)P + PWW T =(G T G) -1 (G T JW T -G T Hb T W T ). (15)

[0091] It is equivalent to the form of \(AP + PB = C\), where \(A=(G T G) -1 (\lambda T+\alpha M)\), \(B = WW T \), \(C=(G T G) -1 (G T JW T -G T Hb T W T ). Obviously, formula (15) is a standard Sylvester equation, where \(P\) is the variable, and then introduce \(\varepsilon I T \) on the diagonal of \((G d G)\) to avoid possible singularities.

[0092] Step 4.3, Update the target domain feature label matrix \(F t : Update the target domain feature label matrix \(F\) according to Step 2.1 t .

[0093] Step 4.4, Update the sample confidence weight matrix \(U\): Update the sample confidence weight matrix \(U\) according to Step 2.2.

[0094] Step 4.5, Update the classification matrix \(W\): Fix the projection matrices \(P s \), \(P t \), the offset vector \(b\), the target domain feature label matrix \(F t \), and the sample confidence weight matrix \(U\), and the objective function of \(W\) is:[[]]

[0095]

[0096] Let \(G = UX\) T \(, H = U1, J = UY\), the formula (16) can be rewritten as:

[0097]

[0098] Equivalent to:

[0099] \(\min(\text{tr}(W T P T G T + bH T - J T )(GPW + Hb T - J)+ \beta\text{tr}(W T W)). (18)

[0100] Take the partial derivative of \(W\) and set it to 0, we get:

[0101]

[0102] The solution is:

[0103] \(W = (\beta1 + P T G T GP) -1 P T G T (J - Hb T ). (20)

[0104] Step 4.6. Update the offset vector \(b\): Fix the projection matrices \(P\) of the source domain and the target domain s \(, P t \), the classification matrix \(W\), the target domain feature label matrix \(F t \), the sample confidence weight diagonal matrix \(U\), and the objective function of \(b\) is:

[0105]

[0106] The formula (21) can be rewritten as:

[0107]

[0108] Equivalent to:

[0109]

[0110] Take the partial derivative of \(b\) and set it to 0, we get:

[0111]

[0112] The solution is:

[0113] \(b = (H T H)-1 (J-W T P T G T )H. (25)

[0114] Step 4.7: Repeat and iterate Steps 4.2, 4.3, 4.4, 4.5, and 4.6 until the objective function converges, and finally complete the projection matrices P for the fixed source domain and the target domain. s 、P t , the offset vector B, and the target domain feature label matrix F t , the sample confidence weight matrix U, and the joint iterative optimization of the classification matrix W, so as to obtain the corresponding predicted value labels.

[0115] In this experiment, the benchmark emotion EEG dataset SEED-IV was used. SEED-IV is an EEG dataset collected by Shanghai Jiao Tong University and is dedicated to emotion recognition research. This dataset recruited 15 subjects, and each subject participated in 3 experiments. In each experiment, the subjects were required to watch 24 movie clips, with 6 clips corresponding to each emotional state (happy, angry, sad, and neutral) to stimulate specific emotions. During the viewing process, 62-channel EEG signals were recorded using an EEG device with a sampling rate of 1000 Hz, and data preprocessing was performed to remove noise. SEED-IV is widely used in emotion computing, brain-computer interface (BCI), and machine learning research, providing high-quality data support for multi-modal emotion recognition.

[0116] Table 1 Cross-subject emotion recognition results of the one-to-one transfer paradigm on the SEED-IV dataset

[0117]

[0118]

[0119] Through comparative experiments, it can be seen that the results of this embodiment have higher recognition accuracy in three scenarios compared with the semi-supervised method of the currently popular joint EEG feature transfer and semi-supervised cross-subject emotion recognition model JTSR. The average improvement in the first scenario is 7.02%, the average improvement in the second scenario is 2.29%, and the average improvement in the third scenario is 3.97%.

[0120] In this embodiment, by using the projection matrices P s 、P t , the offset vector B, and the target domain feature label matrix F t, the joint iterative optimization of the sample confidence weight matrix U and the classification matrix W enables the final model to effectively determine the labels of unlabeled samples; and obtain the weight coefficients that quantitatively describe the sample confidence. In this embodiment, on the one hand, the model combines the domain subspace features and the target domain label information, avoiding the suboptimal solution problem caused by separating the two processes and destroying their internal connection; on the other hand, the model gradually labels the confident target samples with the aligned subspace class prototypes, fusing accurate label estimation and label confidence measurement, and effectively reducing the risk of error propagation in the initial stage.

[0121] The above has described the embodiments of the present invention in detail in conjunction with the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, without departing from the principles and spirit of the present invention, various changes, modifications, substitutions, and variations to these embodiments including components still fall within the protection scope of the present invention.

Claims

1. A cross-domain EEG emotion recognition method for progressive target subject sample mining, characterized in that: It includes the following steps: Step 1: Collect EEG data of multiple subjects in different emotion-inducing scenarios, extract the features of different channels and frequency bands of EEG data after preprocessing, and generate source domain and target domain data; Step 2: By aligning the marginal distribution and conditional distribution of the source domain and the target domain, determine the target domain feature label matrix F based on similarity t , and according to F t Obtain the sample confidence weight matrix U; Step 3: Use cross entropy to measure the error between the sample and the feature label, combine the sample confidence weight matrix and the marginal distribution and conditional distribution alignment error to construct the objective function; Step 4: According to the objective function, the source domain and target domain projection matrices, classification matrices and offset vectors, target domain feature label matrix, and sample confidence weight matrix are jointly iteratively optimized. Based on the iteratively optimized target domain feature label matrix, the emotion category of the subject during EEG data collection is determined.

2. The cross-domain EEG emotion recognition method for progressive target subject sample mining according to claim 1 is characterized in that: The specific process of preprocessing and feature extraction of EEG data in step 1 is as follows: after downsampling the EEG data, noise and electrooculogram artifacts are filtered out through a bandpass filter; the filtered data is divided into 5 EEG frequency bands, namely Delta: 1-4Hz, Theta: 4-8Hz, Alpha: 8-14Hz, Beta: 14-31Hz, Gamma: 31-50Hz, and differential entropy features of different channels and frequency bands are extracted based on this data.

3. The cross-domain EEG emotion recognition method for progressive target subject sample mining according to claim 1 is characterized in that: The specific implementation process of step 2 is as follows: Step 2.1: Obtain the source domain and target domain for feature projection respectively to obtain their representation in the common subspace and Calculate the source domain subspace feature Z s The class center value; for each sample X in the target domain (j) , calculate its subspace features The distance between each source domain prototype, take the inverse of each distance value to get the similarity score; based on the similarity score, construct a feature label matrix F t , used to assign pseudo labels to each sample in the target domain. For each target domain sample X (j) , determine the probability of the jth sample belonging to the cth class according to its similarity score with the cth class prototype in, is an indicator function, which means that if If yes, it returns 1, otherwise it returns 0; Step 2.2: For each sample According to F t Get the probability of its most likely class And use the left truncated Gaussian distribution F to calculate the sample confidence v j , the sample confidence weight matrix U is a diagonal matrix, v j That is, the jth element in the diagonal matrix U: erf(x) represents the error function, μ is the average probability of the largest possible class of all samples, and σ is the variance.

4. The cross-domain EEG emotion recognition method for progressive target subject sample mining according to claim 3 is characterized in that: The objective function is as follows: in, represents the source domain sample set, n s represents the total number of source domain samples, represents the target domain sample set, n t Represents the total number of samples in the target domain, n = n t +n s ; Indicates the estimated sample size of the cth class, N s 、N t is a diagonal matrix, The diagonal elements are and are the projection matrices for the source domain and the target domain respectively, k is the dimension of the projection space; C is the number of sample categories, represents the number of samples belonging to the cth class in the source domain, Represents the number of samples belonging to the cth class in the target domain; Represents the classification matrix, b=[b1,b2,…b n ] represents the offset vector, Indicates the label information of the source domain, F t is the target domain feature label matrix, Sample confidence weight matrix where v i Represents the confidence weight of the i-th sample; α, β, and λ are all expressed as adaptive weights, and 1 represents a vector of all 1s.

5. The cross-domain EEG emotion recognition method for progressive target subject sample mining according to claim 4 is characterized in that: The specific process of constructing the objective function is: using the Euclidean distance to measure the alignment error of the subspaces mapped from the source domain and the target domain, and considering the marginal distribution and conditional distribution for each class to obtain the constructed model: The feature label matrix and sample confidence matrix are obtained through distribution alignment, and the objective function is finally obtained by combining the cross entropy loss of the reliable samples and labels.

6. The cross-domain EEG emotion recognition method for progressive target subject sample mining according to claim 4 is characterized in that: The method of joint iterative optimization in step 4 is: Step 4.1: Initialize the projection matrix P of the source domain and the target domain s and P t , classification matrix W, offset vector b; Step 4.2: Update the projection matrix P of the source domain and target domain by least squares s , P t :Fixed classification matrix W, offset vector b, target domain feature label matrix F t , sample confidence weight matrix U, get P s , P t The objective function is: Solve the above equation to update the projection matrix P of the source domain and the target domain s , P t ; Step 4.3: Update the target domain feature label matrix F according to step 2 t ; Step 4.4, update the sample confidence weight matrix U according to step 2; Step 4.5: Update the classification matrix W by least squares: Fix the projection matrix P of the source domain and the target domain s and P t , offset vector b, target domain feature label matrix F t , the sample confidence weight matrix U, and the objective function of W is: The above formula is used to update the classification matrix W; Step 4.6: Update the offset vector b by least squares: Fix the projection matrix P of the source domain and the target domain s and P t , classification matrix W, target domain feature label matrix F t , the sample confidence weight diagonal matrix U, and the objective function of b is: Calculate the above formula to update the offset vector b; Step 4.7, repeat steps 4.2, 4.3, 4.4, 4.5, and 4.6 until the objective function converges, and finally complete the projection matrix P of the source domain and the target domain s and P t , classification matrix W, offset vector b, target domain feature label matrix F t , joint iterative optimization of the sample confidence weight matrix U.

Images