A method for EEG emotion recognition based on component decoupling and aliasing analysis

Abstract

The present invention provides an EEG emotion recognition method based on component decoupling and aliasing analysis. The method is as follows: 1. Multiple subjects collect EEG data under different emotion induction tasks. 2. Preprocess and extract features of EEG data. 3. Construct an emotion recognition model that introduces a drag direction matrix and a drag degree matrix. 4. Iteratively optimize the variables in the objective function to obtain a globally optimized objective function. 5. Use the emotion recognition model to identify the EEG features of the subject and obtain the emotion category of the subject. The present invention uses latent low-rank decomposition technology to effectively capture the main structural features and significant features in EEG signals. At the same time, the present invention uses label dragging technology to adjust the model boundary according to the degree of aliasing, thereby enhancing the ability to recognize subtle differences in emotional states. Therefore, the emotion decoding method proposed in the present invention combines latent low-rank decomposition and label dragging technology to optimize the continuous trial analysis of emotional states.

Description

Technical Field

[0001] The present invention belongs to the field of EEG signal processing of brain-computer interface (BCI), mainly focuses on the method of emotion recognition, and has important applications in improving the accuracy of emotion state monitoring. Background Art

[0002] Brain-computer interface (BCI) technology is an interdisciplinary subject that integrates neuroscience and computer technology. In this field, emotional BCI focuses on capturing and understanding the user's emotional state. Electroencephalogram (EEG) is widely used in emotion decoding research due to its non-invasive and cost-effective advantages. It records the brain's electrical signals through an electrode array on the scalp, thereby revealing the brain's activity patterns under different emotional states. Emotional states are closely related to the activities of multiple brain regions, which are jointly involved in the complex process of emotional experience, regulation, and expression. This study focuses on a key problem in existing technologies: the aliasing of EEG signals between different emotional tasks.

[0003] The phenomenon of emotional aliasing in emotion monitoring refers to the emotional stimulation of the next trial starting before the emotional response of the previous trial has completely decayed between consecutive trials, resulting in partial overlap of the emotional responses of the previous and subsequent trials in brain activity. Specifically, this phenomenon occurs when the physiological components of the emotional response (such as EEG activity) have not returned to the baseline state, and new emotional stimulation has intervened, causing data mixing. In order to analyze the degree of aliasing, researchers can evaluate the overlap and transition between emotional states by calculating the correlation, cross-correlation or dynamic time warping (DTW) of EEG signals between consecutive trials. Although the overlap between emotional states can be qualitatively analyzed, these methods are limited in the precise quantification of emotional aliasing. The present invention adopts a label drag model based on implicit low-rank decomposition to explicitly quantify the degree of aliasing between different emotional states, so as to perform a clearer emotional state analysis. Summary of the invention

[0004] The purpose of the present invention is to provide a method for EEG emotional state recognition and aliasing detection based on latent low-rank decomposition and label drag model. The data decomposition technology can effectively capture the main structural features and significant features in EEG signals, while separating random or non-correlated noise components, thereby enhancing the robustness of the model while obtaining "pure" data. The introduction of the label drag model makes the boundary between different emotional states in the feature space more reasonable, thereby more accurately identifying the aliasing between emotional states, and further improving the accuracy and efficiency of emotion recognition.

[0005] In a first aspect, the present invention provides an EEG emotion recognition method based on component decoupling and aliasing analysis, which comprises the following steps:

[0006] Step 1: The subject collects EEG data under different emotional states to obtain the subject's EEG data and its corresponding emotion category label.

[0007] Step 2: Preprocess and extract features of the original EEG data collected in step 1 to obtain a sample matrix X.

[0008] Step 3: Construct an emotion recognition model. The objective function of the emotion recognition model introduces the drag direction matrix B and the drag degree matrix M in the form of B⊙M for label dragging, and performs implicit low-rank decomposition on the sample matrix X; ⊙ represents the Hadamard product.

[0009] Step 4: Iteratively optimize the variables in the objective function to obtain the globally optimized objective function.

[0010] Step 5: Use the emotion recognition model to identify the EEG characteristics of the subject and obtain the subject's emotion category.

[0011] Preferably, the objective function of the emotion recognition model is as follows:

[0012]

[0013] In formula (1), X is the input sample matrix; Z is the global structural feature matrix, L is the emotion-related feature matrix; E represents noise or outliers; W and b are weights and biases, respectively; Y represents the label matrix; Y u represents the label matrix corresponding to the unlabeled samples; γ and λ are regularization parameters.

[0014] Preferably, the drag direction matrix B and the drag degree matrix M in the objective function work together to help the emotion recognition model identify and adjust the state of emotion aliasing. During the training process of step 4, for unlabeled samples, the matrix B is dynamically adjusted as the model predicts the category; at this time, the value of each row of the matrix B represents the model's confidence that each sample belongs to each category. The matrix M is updated during the training process, and for data points that are difficult to distinguish between different categories, the corresponding elements in the matrix M are increased to enlarge the interval between categories.

[0015] Preferably, the initial value of the matrix B is set according to the one-hot encoding of the label, and the unlabeled data is set using a uniform distribution; the initial value of the matrix M is usually all zero or a positive value close to zero.

[0016] Preferably, after step 4 is executed, the size of the aliasing value ρ(i) in the timing signal is calculated using the matrix E, as shown below:

[0017] ρ(i)=||e i ||2 (2)

[0018] Among them, ||·||2 represents the l2 norm of the vector, e i It represents the i-th column of the matrix E, and i represents the sequence number of the EEG samples collected in chronological order.

[0019] During the iterative optimization process in step 4, the aliasing degree of E is reduced by minimizing ρ(i).

[0020] The larger the aliasing value ρ(i), the stronger the residual effect of the previous emotional task is, and the greater the interference to the current task is. By measuring the aliasing data between adjacent trials, we can quantitatively examine the degree of mutual influence of emotional states when performing different tasks from a data level. This helps to understand the degree of influence of the previous emotional task on the current task, as well as the changing trend of this influence over time, so as to more accurately evaluate the performance and reliability of the emotion recognition system.

[0021] Preferably, the specific process of step 5 is: collect EEG signals from the subject, preprocess and extract features of the obtained EEG signals to obtain EEG features; use the obtained EEG features as the sample matrix X (including labeled and unlabeled samples). Input the sample matrix x into a semi-supervised learning framework. Under this framework, by iteratively optimizing the objective function, the label matrix Y is continuously updated and finally obtained. u , determine the emotional category of the subject.

[0022] Preferably, the emotion categories to be measured include happiness, sadness, neutrality and fear. In step 1, the subject is allowed to switch between different emotion states by visual induction.

[0023] Preferably, in step 1, the EEG acquisition uses 62 leads, and 5 frequency bands are selected for feature extraction, namely: 1-4Hz, 4-8Hz, 8-14Hz, 14-31Hz and 31-50Hz.

[0024] Preferably, in step 4, the present invention optimizes a selected single variable and fixes all variables except this one to minimize the objective function. And by optimizing each variable one by one, the optimal solution of the problem is gradually approached. The specific update rules are as follows:

[0025] 4-1. Update the drag matrix M and B, and use the objective function:

[0026]

[0027] 4-2. Update Z and L, using the objective function:

[0028]

[0029] 4-2-1 Update J, the objective function used is:

[0030]

[0031] 4-2-2 Update S, the objective function used is:

[0032]

[0033] 4-2-3 Fix other variables, update Z, and use the objective function:

[0034] Z=(I+X T X) -1 (X T (X-LX-E)+J+(X T Y1-Y2) / μ) (7)

[0035] 4-2-4 Fix other variables, update L, and use the objective function:

[0036]

[0037] 4-3. Fix other variables, update E, and use the objective function:

[0038]

[0039] 4-4. Fix other variables, update b, and use the objective function:

[0040]

[0041] 4-5. Fix other variables and update Y u , the objective function used is:

[0042]

[0043] 4-6. Fix other variables, update W, and use the objective function:

[0044]

[0045] In a second aspect, the present invention provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the aforementioned EEG emotion recognition method.

[0046] In a third aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the aforementioned EEG emotion recognition method.

[0047] The present invention has the following beneficial effects:

[0048] 1. The aliasing measurement matrix created by the present invention can quantitatively analyze the degree of aliasing between consecutive emotional tasks, quantify the influence of one emotional task on subsequent emotional tasks, and the changing trend of this influence.

[0049] 2. By introducing implicit low-rank decomposition, we can capture the structural information in the data more finely, including global structural features and significant structural features. For the aliasing analysis of emotional states, this means that we can consider both the general patterns of emotional responses and task-related patterns, providing a more comprehensive way of emotion analysis.

[0050] 3. The introduced label drag model introduces additional category information through the B⊙M term, which helps the model distinguish different categories, especially when the boundaries between some categories are not very clear. The model fine-tunes the classification boundaries of each category to better distinguish those situations with subtle differences between emotional states. This is particularly beneficial for identifying aliased emotional states because it allows the model to adjust the boundaries to distinguish emotional states that affect each other in consecutive trials.

[0051] 4. The emotion decoding method proposed in the present invention combines data denoising and decoding processes, and uses the nuclear norm to enhance the robustness of the model, which makes the decoding method more reliable when facing emotion state aliasing, and effectively improves the accuracy and efficiency of emotion state recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the overall framework of the present invention;

[0053] Figure 2 Schematic diagram of the emotion aliasing measurement of adjacent trial tasks in the present invention;

[0054] Figure 3 A diagram showing how to adjust the aliasing level by dragging the model with a label. DETAILED DESCRIPTION

[0055] The present invention will be further described below in conjunction with the accompanying drawings.

[0056] This paper introduces an EEG emotion recognition method based on component decoupling and aliasing analysis. It is used to address the low recognition accuracy of EEG emotion state recognition datasets and the important problem of aliasing between emotion states. The method is based on latent low-rank decomposition and label drag model. Through this data decomposition technology, the main structural features and significant features in EEG signals can be effectively captured, and random or non-correlated noise components can be separated. This not only obtains relatively "pure" data, but also enhances the robustness of the model. At the same time, the application of the label drag model can more accurately identify the aliasing between emotion states, enhance the ability to recognize subtle differences in emotional states, and is of great significance for improving the accuracy of emotion recognition.

[0057] like Figure 1 As shown, the EEG emotion recognition method includes the following steps:

[0058] Step 1, EEG data collection. This embodiment uses visual stimulation to induce emotional responses, specifically by watching movie clips of different emotional categories. The experiment selected four movie clips covering common basic human emotional responses, happiness, sadness, fear, and neutrality, as materials for emotional induction to ensure the effectiveness and consistency of the emotional induction process. This induction method helps to reduce the impact of external environmental variables on experimental results and maintain the consistency of experimental conditions.

[0059] Data collection was conducted using a 64-channel ESINeuroScan system in conjunction with an SMI eye tracker, designed to record brainwave and eye movement data while subjects watched emotion-inducing film clips. Each subject watched four emotion categories of film clips in a single experiment, each emotion category contained six clips, so there were a total of 24 clips, each clip lasting about 2 minutes, to induce the corresponding emotional response. Before each 2-minute film clip, the subject received a 5-second preview prompt and performed a 45-second emotional self-assessment after watching. In order to obtain reliable data, each participant participated in three days of data collection, with an interval of about one week between each experiment to ensure the recovery of emotional state and comparability of data.

[0060] Step 2: Preprocess and extract features from the raw EEG data collected in step 1 to obtain a sample matrix X; the sample matrix X contains both labeled features and unlabeled features extracted based on the measured EEG data. The preprocessing steps include filtering, artifact removal, and centering, in order to improve data quality and ensure the accuracy and reliability of subsequent analysis. The specific process is as follows:

[0061] 2-1. Downsample the original EEG data to 200Hz. This embodiment uses bandpass filtering with a frequency set at 1-75Hz to include common frequency bands of brain activity: Delta (1-4Hz), Theta (4-8Hz), Alpha (8-14Hz), Beta (14-31Hz) and Gamma (31-50Hz). Next, the independent component analysis method is used to remove signal artifacts from non-brain activity sources such as muscle movements, eye movements, or equipment interference. Finally, the EEG data after filtering and artifact removal is centered. This step involves subtracting the mean value from the signal of each channel to unify the data baseline, thereby providing a more accurate signal representation for subsequent analysis.

[0062] 2-2. After preprocessing, the EEG data of different frequency bands are subjected to short-time Fourier transform (STFT) with a time window of 3 seconds and no overlap to extract differential entropy features. Assuming that the EEG signal sample X conforms to the Gaussian distribution: X~N(μ,σ 2 ), the differential entropy feature is defined as follows:

[0063]

[0064] It can be seen that the essence of the differential entropy feature is the logarithmic form of the power spectrum density feature, that is,

[0065] Step 3: Build an emotion recognition model. The emotion recognition model is specifically a machine learning model that combines latent low-rank decomposition and label dragging for EEG emotion recognition and aliasing measurement; the objective function of the model is expressed by introducing a series of variables and regularization terms to ensure that the model can effectively distinguish different emotional states while measuring and controlling emotional aliasing. Through the above steps, the robust recognition objective function originally proposed can be obtained as follows:

[0066]

[0067] In formula (14), is the input sample matrix; is a matrix representing global structural features extracted from the sample matrix X; is the matrix of emotion-related features extracted from the sample matrix X; Represents noise or outliers; and They are the weights and biases of the machine learning model; Represents the label matrix, n = l + u, l corresponds to the number of labeled samples, and u corresponds to the number of unlabeled samples; Represents the label matrix corresponding to the unlabeled samples; and They are the drag direction matrix and drag degree matrix for dragging labels, which can measure the degree of sentiment aliasing; γ and λ are regularization parameters.

[0068] For a matrix V, its nuclear norm can be expressed as where σ i are the singular values ​​of the matrix V, and r is the rank of the matrix. ||·|| 2，1 l represents the matrix 2，1 norm, n, m represent the number of rows and columns of the matrix, v ij Represents the element in the i-th row and j-th column of the matrix V. ||·|| 1，2 l represents the matrix 1，2 norm,

[0069] In the objective function, the nuclear norm ||·|| * Used to promote low-rank structure, thus helping the model remain robust when processing data, ||·|| 1，2 and ||·|| 2，1 The norm is used to improve the robustness of the model to noise. The objective function also includes a square term of the Frobenius norm, which ensures the consistency between the model output and the actual emotional label. Considering the impact of emotional aliasing, the actual emotional label is represented by a label drag model. The overall model balances the complexity and performance of the model through regularization terms and constraints.

[0070] The drag direction matrix B and the drag degree matrix M in the objective function work together to help the emotion recognition model identify and adjust the state of emotion aliasing. In the initial stage, the initial value of the matrix B is set according to the unique hot encoding of the label, and the unlabeled data is set using a uniform distribution; the initial value of the matrix M is usually all zero or a small positive value, indicating that no assumptions are made about the intervals between categories at the beginning.

[0071] Step 4: First, u and W are initialized, and then for the objective function obtained in step 3, the variables Z, L, E, B, M, W, Y in the objective function are updated in turn by fixing some variables and updating other variables. u , b is updated to achieve the purpose of joint optimization. This process is repeated multiple times until convergence to complete the global optimization of the objective function.

[0072] During the training process, matrix B is adjusted according to the category labels and the model's predictions. Therefore, for labeled samples, matrix B is relatively stable; for unlabeled samples, matrix B will be dynamically adjusted as the model's predictions change. At this time, the value of each row of matrix B represents the model's confidence that each sample belongs to each category, rather than simply -1 or 1. Matrix M is updated during the training process to reflect the need to increase the intervals between different categories. If data points between certain categories are difficult to distinguish, the corresponding values ​​in matrix M will increase to enlarge the intervals between these categories.

[0073] The result of the calculation of B⊙M is a new matrix, in which each element combines the direction of matrix B and the amplitude of matrix M, so that the original strict binary labels are relaxed into loose labels. For example, anger and excitement may show similarities in some physiological signs, and label dragging can help the model establish a clearer dividing line between these emotions. As shown below, assuming that there is a set of four-category sample data, a sample is selected from each category, and their binary label matrix H is obtained:

[0074]

[0075] It can be calculated that the distance between different samples is fixed, which is Next, we introduce a small perturbation ∈ ij ≥0, indicating dragging of labels in feature space:

[0076]

[0077] These small perturbations ∈ ij The label becomes less strict, and the position of the sample in the feature space changes, resulting in an increase in the distance between samples of different categories. For example, the distance between the second and third samples can be calculated as follows:

[0078]

[0079] Therefore, the perturbation introduces greater flexibility, allowing the positions of different emotion categories in the feature space to be adjusted according to the specific situation, thereby better adapting to the complexity of emotional states in real data.

[0080] For samples with mixed emotions, label dragging technology can make these samples relatively central in the feature space, so that these samples are not completely close to the center of any single emotion category, thus helping to identify the mixed state of emotions. By creating a more reasonable decision boundary by dragging labels in the feature space, it is also possible to more accurately classify samples that are not obvious in emotional expression or fluctuate between several emotions.

[0081] like Figure 2 and 3 As shown in Figure 2, through the iterative process, B and M will gradually adapt to the data, helping the model to better distinguish between various emotion categories and reduce classification uncertainty caused by data aliasing. This process is applicable not only to labeled data, but also to unlabeled data, allowing the model to use all available information for more accurate emotional state classification.

[0082] In summary, by combining the analysis of data aliasing by the E matrix and the spacing adjustment function of the B⊙M label drag model, the present invention successfully achieves accurate classification of emotion categories and significantly reduces the uncertainty in the classification process. Specifically, after obtaining the matrix E, ||E|| 1，2 The term is used to measure the error or noise in the data and is obtained by applying the l2 norm to each column of the matrix E and applying the l1 norm to the result. It is effective for finding sparse outliers (i.e., error signals that stand out in some trials) because l 1，2 The norm promotes the sparsity of the rows. The size of the aliased value π(i) in the timing signal is calculated using the matrix E as follows

[0083] ρ(i)=||e i ||2 (18)

[0084] Among them, ||·||2 represents the l2 norm of the vector, e i It represents the i-th column of the matrix E, and i represents the sequence number of the EEG data of the subjects collected in chronological order.

[0085] By observing the structure of E, especially which rows have large l2 norm values, we can determine which trials contain more aliasing. If the row norm values ​​corresponding to some trials are abnormally high, it means that the data of these trials do not match the main emotional state captured by the model, which may indicate emotional aliasing.

[0086] In the process of model optimization, by minimizing ρ(i) and reducing the aliasing degree of E, B⊙M can dynamically adjust the position of the emotional state according to the reduced aliasing information, making the boundary between different emotional states in the feature space clearer, which helps to improve the accuracy of classification. Therefore, B⊙M and ||E|| 1，2 The joint optimization of can be a powerful tool for identifying and measuring emotional aliasing. By adjusting the corresponding terms in the objective function, we can improve the model's ability to distinguish emotional states and reduce the aliasing between different emotional states, thereby improving the overall performance of the model.

[0087] Step 5: Input the sample matrix X obtained in step 2 into the objective function, and obtain the corresponding predicted value labels through iterative optimization in step 4. These predicted value labels represent the emotional state type of the subjects corresponding to the samples at the time of data collection. This method uses the inherent structural characteristics of existing data to enhance the model's ability to predict new samples, thereby predicting the category of emotional EEG data of subjects with unknown labels.

[0088] The specific solution process of step 4 is as follows:

[0089] 4-1. Update M

[0090] Because B and M are usually directly related, this step can be used to adjust the class boundaries. Where Y is the category matrix and W is the weight matrix. The objective function becomes: It can be transformed into the form of sub-problems: Therefore, we can get:

[0091]

[0092] 4-2. Update Z and L

[0093] Let Z≈J, L≈S, and transform the updates of Z and L into updates of Z, L, J, and S. The original objective function can be transformed into a Lagrangian function with equations as follows:

[0094]

[0095] 4-2-1. Update J and construct the optimization function as follows

[0096]

[0097] This is a nuclear norm regularization problem, whose goal is to make the matrix close to the given data matrix while keeping the matrix low rank. This type of problem can be solved by the singular value thresholding (SVT) technique. First, initialize the parameters: select the initial J, Y2 and μ. The entire iterative process is as follows:

[0098] (1) Singular Value Thresholding (SVT): Apply SVT to update J: in, is the singular value thresholding operation, is the threshold parameter of the current iteration, which is used to determine the degree of soft thresholding applied to each singular value in SVT. Specifically, the target matrix is ​​first constructed: Thus, the current iteration Z and the corresponding Lagrange multipliers are integrated to form a new target matrix, and a J close to this matrix will help satisfy the constraints of the optimization problem.

[0099] Next, in order to update J, it is necessary to perform singular value decomposition (SVD) on the A1 formed above. Singular value decomposition decomposes A1 into three parts: the left singular vector matrix U1, the singular value diagonal matrix Σ1 and the right singular vector matrix For every singular value σ in Σ1 i , apply threshold Make adjustments: The updated singular value diagonal matrix Σ1′ is composed of these adjusted singular values ​​σ i ′. Finally, J is reconstructed using the updated singular value diagonal matrix Σ1′ and the original singular vectors U1 and V1: This reconstruction process ensures that J k+1 As low rank as possible, and as close as possible to the adjusted target matrix A1.

[0100] (2) Update the Lagrange multiplier:

[0101]

[0102] (3) Update or adjust μ: According to the descent rate of the objective function and the satisfaction of the constraints, the penalty parameter is adjusted as follows: μ k+1 = Adjustment strategy (μ k ).

[0103] (4) Repeat steps (1) to (3) multiple times until the number of iterations k reaches a preset maximum number of iterations.

[0104] 4-2-2. Update S and construct the optimization function as follows

[0105]

[0106] This embodiment adopts a similar strategy to update J to solve S. Based on singular value thresholding (SVT), problem (23) can be simplified to finding a matrix S that minimizes its nuclear norm and is close to the adjusted L (considering the influence of Lagrange multiplier Y3).

[0107] Next, the matrix Applying singular value decomposition (SVD), we get:

[0108]

[0109] In which, each singular value σ in the singular value diagonal matrix Σ2 i satisfy: Thus, the updated singular value diagonal matrix Σ′2 is obtained, and then the update of S is completed: S k+1 =UΣ′V T .

[0110] 4-2-3. Fix other variables and update Z to get the optimization formula as follows

[0111]

[0112] Let the above equation be differentiated to zero, and we can get

[0113] Z=(I+X T X) -1 (X T (X-LX-E)+J+(X T Y1-Y2) / μ) (26)

[0114] 4-2-4. Fix other variables and update L. We can get the update rule of variable L as follows

[0115]

[0116] Let the above equation be differentiated to zero, and we can get

[0117]

[0118] After further simplification, we get

[0119]

[0120] Converted to standard form A'X'+X'B'=C'. B'=μ(XX T +I)(XX T ) -1 ; A′=WW T ; C′=(μ(X-XZ-E)X T +(μS+Y1X T -Y3)-W(1 n b T -YB☉M) T X T )(XX T ) -1

[0121] 4-3. Fix other variables and update E. The update rule of variable E can be obtained as follows

[0122]

[0123] Let H = X-XZ-LX, the original objective function can be transformed into

[0124]

[0125] make Then the above formula is equivalent to the following problem

[0126]

[0127] Further broken down into:

[0128]

[0129] Among them, e j and g j are the jth columns of matrices E and G respectively. The solution to the above objective function is

[0130]

[0131] 4-4. Fix other variables and update b. The update rule of bias variable b can be obtained as follows:

[0132]

[0133] The expression of b is as follows:

[0134]

[0135] 4-5. Fix other variables and update Y u , the objective function is about the variable Y u The expression is

[0136]

[0137] For l+1≤i≤l+u, the above problems are independent, so we can decompose Y by rows u , for each y i , we have:

[0138]

[0139] Assume that m = (X T L T ) i W+1 n b T -(B☉M) i , the above problem is equivalent to the following

[0140]

[0141] It is a Euclidean projection onto a simplex and can be solved using a combination of Newton's method and the method of Lagrange multipliers.

[0142] 4-6. Fix other variables and update W. The objective function of W is as follows

[0143]

[0144] Integrate the bias term b into the objective function and get the new objective function as follows:

[0145]

[0146] Next, convert it into an equivalent trace form

[0147]

[0148] in:

[0149] After simplification, we can get

[0150]

[0151] in Q=LX(Y+B☉M),

[0152] Next, to optimize W, the generalized power iteration method can be used.

Claims

1. A method for EEG emotion recognition based on component decoupling and aliasing analysis, characterized by: The following steps are involved: Step 1: The subject collects EEG data under different emotional states to obtain the subject's EEG data and its corresponding emotion category label; Step 2: preprocess and extract features of the original EEG data collected in step 1 to obtain a sample matrix X; Step 3, construct an emotion recognition model; the objective function of the emotion recognition model introduces a drag direction matrix B and a drag degree matrix M in the form of B⊙M for dragging labels, and performs implicit low-rank decomposition on the sample matrix X; ⊙ represents the Hadamard product; the objective function of the emotion recognition model is as follows: In formula (1), X is the input sample matrix; Z is the global structural feature matrix, and L is the emotion-related feature matrix; E represents the noise matrix or outlier matrix; W and b are weights and biases respectively; Y represents the label matrix; Y u represents the label matrix corresponding to the unlabeled samples; γ and λ are regularization parameters; n = l + u, l corresponds to the number of labeled samples, and u corresponds to the number of unlabeled samples; ||·|| * is the nuclear norm; ||·|| 1，2 Represents the matrix norm;||·|| 2，1 Represents the matrix norm; Step 4: Iteratively optimize the variables in the objective function to obtain the globally optimized objective function. The iterative optimization process is: optimize the selected single variable, fix all variables except this one, to minimize the objective function; and gradually approach the optimal solution of the problem by optimizing each variable one by one; use the matrix E to calculate the size of the aliasing value ρ(i) in the timing signal, as shown below: ρ(i)=||e i ||2 (2) Among them, ||·||2 represents the l2 norm of the vector, e i represents the i-th column of the matrix E, i represents the sequence number of the EEG samples collected in chronological order; In the iterative optimization process of step 4, the aliasing degree of E is reduced by minimizing ρ(i); Step 5: Use the emotion recognition model to identify the EEG characteristics of the subject and obtain the subject's emotion category.

2. The method for EEG emotion recognition based on component decoupling and aliasing analysis according to claim 1, characterized in that: The drag direction matrix B and the drag degree matrix M in the objective function work together to help the emotion recognition model identify and adjust the state of emotion aliasing; in the training process of step 4, for unlabeled samples, the matrix B is dynamically adjusted as the model predicts the category changes; at this time, the value of each row of the matrix B represents the model's confidence that each sample belongs to each category; the matrix M is updated during the training process, and for data points that are difficult to distinguish between different categories, the corresponding elements in the matrix M are increased to enlarge the category interval.

3. The method for EEG emotion recognition based on component decoupling and aliasing analysis according to claim 1, characterized in that: The initial value of matrix B is set according to the unique hot encoding of the label, and the unlabeled data is set using a uniform distribution; the initial value of matrix M is usually all zero or a positive value close to zero.

4. The method for EEG emotion recognition based on component decoupling and aliasing analysis according to claim 1, characterized in that: The specific process of step 5 is: collect EEG signals from the subjects, preprocess and extract features of the EEG signals to obtain EEG features; use the obtained EEG features as the sample matrix X, which includes labeled and unlabeled samples; input the sample matrix X into a semi-supervised learning framework; under this framework, by iteratively optimizing the objective function, continuously update and finally obtain the label matrix Y u , determine the emotional category of the subject.

5. The method for EEG emotion recognition based on component decoupling and aliasing analysis according to claim 1, characterized in that: In step 1, the EEG acquisition uses 62 leads, and 5 frequency bands are selected for feature extraction, namely: 1-4Hz, 4-8Hz, 8-14Hz, 14-31Hz and 31-50Hz.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: The memory stores computer programs; the processor executes the EEG emotion recognition method as described in any one of claims 1-5.

7. A readable storage medium storing a computer program; characterized in that: When the computer program is executed by a processor, it is used to implement the EEG emotion recognition method as described in any one of claims 1 to 5.

Images