An Emotion EEG Network Recognition System Based on Bayesian Spectral Regression Strategy

Abstract

This invention discloses an emotion EEG network recognition system based on a Bayesian spectral regression strategy, belonging to the field of EEG signal processing. This system can robustly mine discriminative feature information of the emotion EEG network, achieving effective characterization of the emotion EEG network. To verify the robustness of the proposed emotion recognition system, simulation and real-world data experiments based on the DEAP and MAHNOB datasets were designed. Experimental results show that the robust Bayesian spectral regression model based on Student's T and Laplace prior distributions proposed in this application can effectively improve the recognition accuracy of emotion EEG signals, achieving robust recognition of individual emotional states. This application provides a potential solution for the design and implementation of robust emotion recognition and emotion-based brain-computer interface systems based on EEG.

Description

Technical Field

[0001] This invention relates to techniques for decoding and classifying electroencephalogram (EEG) networks in the field of cognitive neuroscience research, and particularly to an EEG network classification method based on Bayesian spectral regression. Background Technology

[0002] Electroencephalogram (EEG) signals, as a primary manifestation of brain activity, play a crucial role in decoding human cognitive activities and human-computer interaction. Emotion, as one of the brain's higher cognitive functions, makes network analysis of emotional EEG signals significant for further revealing emotional cognitive processing and brain cognitive patterns. In recent years, with the continuous development of computer and brain-computer interface (BCI) technologies, researchers have proposed the concept of affective brain-computer interfaces (BCIs) based on traditional motor decoding BCIs, giving rise to a new research field—affective computing. Affective computing aims to capture the differences in brain activity patterns under different emotional cognitive states through data processing and information mining of emotional EEG signals, achieving effective identification of individual emotional states. Benefiting from the high temporal resolution, ease of operation, and low cost of EEG signals, research on decoding individual emotional states based on EEG signals has always been a research hotspot in the field of affective computing. However, due to the noise-sensitive nature of EEG, research on decoding individual emotional states based on EEG signals still faces significant challenges.

[0003] Electroencephalogram (EEG) network analysis, as an effective cognitive neural information processing technique, is widely used in research on higher cognitive processes in the brain. Currently, EEG network analysis, as a highly efficient cognitive neuroscience research tool, offers the possibility of further revealing higher cognitive processes in the brain. Essentially, EEG network analysis characterizes the functional connectivity between different brain regions by evaluating the correlation or consistency between EEG signals in different leads, thereby representing the interaction between various brain regions during cognitive activities. Compared to traditional EEG signal spectral feature analysis methods, which can only utilize information from signals within a single lead, EEG networks can reflect the collaborative processing and information interaction patterns of cognitive information in various brain regions during cognitive neural activities. This not only suppresses the influence of EEG noise on brain signal analysis to a certain extent but also better reflects the global information transmission characteristics of the brain during emotional processing, which is of great significance for further exploring key brain regions and functional connectivity patterns related to cognitive tasks. However, as a random signal with a low signal-to-noise ratio, EEG signals cannot completely eliminate the influence of noise on the results in EEG network analysis. The EEG network estimated from EEG signals contaminated by noise often contains pseudo-functional connections and functional connection deviations caused by noise interference, which poses certain challenges to subsequent EEG network analysis and individual state recognition.

[0004] Among numerous methods for identifying EEG networks, graph theory treats EEG networks as generalized graph signals, considering brain regions as points in an image and connection strengths as edges, and using graph theory to analyze the characteristics of EEG networks. Graph learning often involves Rayleigh entropy eigenvalue problems, such as Locality Preserving Projection (LPP), Neighborhood Preserving Embedding (NPE), and Linear Discriminant Analysis (LDA). To address computational space and model optimization, Cai et al. proposed Spectral Regression Discriminant Analysis (SRDA) based on regularized least squares. This method captures the feature signals that map images to lower dimensions, clustering high-dimensional image signals by mapping them to lower-dimensional information. However, due to the square limitation inherent in the square method, existing spectral regression methods struggle to handle the impact of noise on feature extraction. During the solution of the mapping matrix, they are highly susceptible to outliers, leading to biased clustering results. Therefore, developing more efficient methods for noise reduction in EEG network analysis and data dimensionality reduction is of great significance for further improving research on individual emotion recognition based on EEG networks. Summary of the Invention

[0005] This invention solves the following technical problems: To address the deficiencies in the background technology, this invention proposes a spectral regression method based on a Bayesian framework. It assumes that the likelihood function follows a Laplace distribution and that both the mapping matrix and the likelihood function follow a Student's t distribution, respectively, to adapt to the actual signal distribution affected by noise. It then uses a hierarchical Bayesian method combined with variational Bayesian methods to solve the problem, thereby minimizing the impact of outliers.

[0006] This application designs and implements an emotion EEG network recognition system based on a Bayesian spectral regression strategy. The system includes: a system input module, a system decoding module, and a system output module.

[0007] (I) System Input Module:

[0008] After wearing the EEG data collection device, the participants watched an emotion-evoking video while their EEG signals were collected simultaneously. Throughout the data collection process, the participants were asked to maintain their posture, watch the corresponding emotion video quietly, and remain silent after the video finished playing. They were also asked to complete a questionnaire while recovering their mood to avoid the previous video interfering with their emotions.

[0009] (II) System Decoding Module:

[0010] As a key module of the system, the system decoding module preprocesses the acquired EEG signals and calculates the corresponding EEG network. Then, it feeds the data into a robust Bayesian spectral regression model for training to obtain the corresponding mapping matrix and class center matrix. The robust Bayesian spectral regression model is an optimization of the spectral regression method by introducing Bayesian structure, making the method more robust than the original recognition method and adaptable to noisy EEG signals in real-world situations.

[0011] The method for calculating the mapping matrix using a robust Bayesian spectral regression algorithm is as follows:

[0012] Graph regression transforms the traditional Rayleigh entropy-based optimization problem into a problem of solving eigenvectors and regression. In a supervised manner, the eigenvectors are transformed into prior constraints related to label information. The regression process can be optimized by introducing a hierarchical Bayesian structure. The specific implementation process is as follows:

[0013] First, under supervised learning, an affinity matrix is ​​constructed, and the corresponding feature matrix is ​​also constructed.

[0014] In a supervised study, the affinity matrix can be viewed as a block matrix containing c blocks of shared classes; where sample points of the same class belong to the same block in the affinity matrix. Therefore, the problem of finding the eigenvalues ​​of the affinity matrix is ​​transformed into finding the eigenvalues ​​of the block matrix, where the eigenvector of each block is represented as:

[0015] y c =[0…c…c…0]

[0016] Where c represents the category. Combining the characteristics of the category blocks in the affinity matrix, the feature vector of the affinity matrix is ​​determined by the category under supervised conditions. As long as two samples belong to the same category, the feature vectors in the corresponding blocks are the same.

[0017] In the Bayesian spectral regression process, the feature matrix is ​​fitted using an electroencephalogram (EEG) network, and its matrix representation is as follows:

[0018] Y = XG + Z

[0019] Where X = [x1, x2, ..., x n ],X∈R N×(M×T) This represents the EEG network signal, where N represents the number of channels, M represents the network connections, T represents the training set length, and Y = [y1, y2, ..., y]. t ],Y∈R N×(c-1) Let Z represent the feature matrix constructed from the label signals, where Z ∈ R. T N(0,Σ) represents the Gaussian white noise that appears in the fitting problem. v ), Σ vLet G ∈ R be the covariance matrix corresponding to the Gaussian distribution. (M×T)×(c-1) This represents the corresponding mapping matrix;

[0020] In the process of Bayesian spectral regression, the matrix form of the feature matrix is ​​transformed into a vector expression as: y = Xg + z; where y, g, and z are the vector representations of the feature matrix, the mapping matrix, and the noise, respectively.

[0021] The prior distribution of the mapping matrix is ​​set to a Gaussian distribution g ~ N(0, Σ). g ), Σ g Let the covariance matrix of the corresponding Gaussian distribution be denoted by: and the corresponding likelihood function be:

[0022]

[0023] Therefore, according to Bayes' theorem, the posterior inference can be obtained as follows:

[0024] p(g|y,Θ)∝p(y|X,g,Θ y )×p(g|Θ g )

[0025] Among them, Θ, Θ y Θ g This represents all the hyperparameters needed in the posterior solution process. The entire posterior formula expresses that the probability p(g|y,Θ) is proportional to the probability p(y|X,g,Θ). y ) and probability p(g|Θ g The product of ) . The posterior distribution is solved using the EM method:

[0026] Step 1: Calculate the expected value and the posterior distribution of the latent variables;

[0027] ln(p(g|y,Θ))=E g lnp(y|X,g,Σ v )+E g lnp(g|Σ g )

[0028] Among them, E g This indicates the expectation of the parameter g in the formula;

[0029] Utilizing during the solution process and Let these represent the expectation and covariance of parameter g, respectively:

[0030]

[0031]

[0032] Step 2: Maximize, take the partial derivative of the log-likelihood function from Step 1 and maximize it. The parameters updated in the kth step are:

[0033]

[0034]

[0035] Decoding test: The dataset is mapped using the test set portion of the dataset divided by 10-fold cross-validation. Then, the mapped dimensionality reduction matrix is ​​compared with the class centers, and the closest class center is used as the predicted label. Finally, the prediction accuracy of the entire process is calculated by comparing it with the true label.

[0036] (III) System Output Module: The system displays the predicted labels and outputs the prediction accuracy to determine the overall performance of the system.

[0037] Furthermore, the calculation process for Bayesian spectral regression based on the Laplace distribution prior is as follows:

[0038] Represent the Laplace distribution using Gaussian and exponential distributions:

[0039]

[0040] in, τ≥0, meaning τ is a parameter with σ 2 A random variable with an exponential distribution.

[0041] Assuming the prior distribution of the likelihood function is a Laplace distribution, the corresponding posterior distribution is solved using variational inference within the hierarchical Bayesian framework. The q-distribution, which is easier to derive through variational inference, is obtained because the exponential distribution is also represented as another form of the Gamma distribution. Therefore, based on the Gaussian conjugate relationship between the Gaussian and Gamma distributions, we obtain:

[0042]

[0043]

[0044] Where q(·) represents the q-distribution, This represents a more easily expressible Gaussian distribution for the corresponding q-distribution, where and For its corresponding mean and variance, similarly... That's also true.

[0045] The updated parameter results are obtained:

[0046]

[0047]

[0048]

[0049] Furthermore, the specific calculation process for Bayesian spectral regression based on the Student's T-distribution prior is as follows:

[0050] Represent the student's T-distribution using Gaussian and Gamma distributions:

[0051]

[0052] Assuming that both the noise distribution and the likelihood function are Student's T-distributions, within the framework of hierarchical Bayesian methods, a more easily solvable q-distribution can be derived using variational Bayesian methods. Based on the conjugate relationship between the Gaussian and Gamma distributions, we can obtain:

[0053]

[0054]

[0055] Therefore, the parameter update process can be further derived as follows:

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062] In this application, the emotional EEG network is treated as a generalized graph feature. Based on the graph regression method, a robust Bayesian spectral regression model based on Student's T and Laplace distributions is proposed. This model optimizes the original model by designing the prior distribution of noise or likelihood function of the emotional EEG network, thereby further suppressing the influence of various EEG artifacts on the connection errors caused by EEG network analysis, thus achieving effective dimensionality reduction and identification of the emotional EEG network. Specifically, this invention designs two robust Bayesian models: a Bayesian spectral regression model based on a prior distribution of both noise and likelihood function following Student's T, and a Bayesian spectral regression model based on a prior distribution of noise following Laplace. The contributions of different prior distributions and placement regions to the robustness of the models are discussed. Furthermore, based on this model, an emotion recognition system based on EEG signals is designed and implemented. This system can robustly mine discriminative feature information of the emotional EEG network, achieving effective characterization of the emotional EEG network. To verify the robustness of the proposed emotion recognition system, simulations and real-world experiments based on the DEAP and MAHNOB datasets are designed. Experimental results show that the robust Bayesian spectral regression model based on Student's T and Laplace prior distributions proposed in this application can effectively improve the recognition accuracy of emotional EEG signals and achieve robust recognition of individual emotional states. This application provides a potential solution for the design and implementation of robust emotion recognition and emotion-based brain-computer interface systems based on EEG. Attached Figure Description

[0063] Figure 1 This is a flowchart of a system testing experiment based on simulation data.

[0064] Figure 2 This is a flowchart of a system testing experiment based on real EEG data.

[0065] Figure 3 Experimental procedures for the DEAP and MAHNOB datasets.

[0066] Figure 4 This is the flowchart of the Bayesian spectral regression algorithm.

[0067] Figure 5 The flowchart is for the Bayesian spectral regression algorithm based on the student's T prior distribution.

[0068] Figure 6 This is the flowchart of the Bayesian spectral regression algorithm based on the Laplace prior distribution. Detailed Implementation

[0069] See Figure 1 The example shows a system test experiment flowchart based on simulation data from the system of this invention. The simulation system is specifically (e.g.) Figure 1(As shown) includes: system data generation module, system decoding module, and system decoding module.

[0070] (I) System Data Generation Module:

[0071] First, simulation data of two classes were generated. The multivariate Gaussian distribution function built into MATLAB was used to generate two classes with different means and variances. Since the optimal hyperplane for distinguishing the two classes of signals is in the 135-degree direction, this application designed one multivariate Gaussian distribution with a mean of (1.85, 1.85) and a variance of (0.50, 0.50), and the other multivariate Gaussian distribution with a mean of (3.00, 3.00) and a variance of (0.50, 0.50). This experiment constructed 200 samples, with 100 samples in each class.

[0072] Secondly, to better test the model's noise adaptability and simulate real data under the influence of external noise, this application sequentially adds noise signals with different signal-to-noise ratios (SNR) to two types of simulation signals (snr = 15, 10, 5, 0), respectively. By modifying the SNR, the resistance of different methods to noise under different intensities is reflected. Adding outliers causes deviations in the hyperplane, altering the data classification performance; the Bayesian method is better able to adapt to data changes.

[0073] (II) The system decoding module includes:

[0074] The system decoding module constructs a signal by taking the upper triangle of 200 samples and uses 10-fold cross-validation to divide the data into training and test sets.

[0075] For the training set data, this invention incorporates it, along with the feature matrix constructed under supervised conditions, into the spectral regression and robust Bayesian spectral regression models in the system of this application (e.g., Figure 4 (The model solution process in sections 5 and 6) involves solving for the corresponding mapping matrix and class center matrix. In the robust Bayesian spectral regression model, the mapping matrix is ​​determined by the iteration parameters. The class centers are determined by the cluster centers of the training set.

[0076] For the test set data, this system reduces the dimensionality of the data using a mapping matrix, compares it with the class centers of different classes, and uses the class closest to the class center as the predicted label.

[0077] (III) The system output module includes:

[0078] Finally, the predicted label values ​​output by the system are used, and the predicted labels are compared with the original test set labels to further calculate the label prediction accuracy of the system.

[0079] In summary, simulation experiments verified the reliability and stability of the proposed system, demonstrating that the proposed robust Bayesian spectral regression method for noise reduction is an effective feature extraction method. To test whether the system performs equally well on real-world data, further experiments were conducted on two datasets to verify the overall reliability of the system.

[0080] See Figure 2 The example shown is a flowchart of the system test experiment using real EEG data from the system of this invention. The system specifically includes (e.g.) Figure 2 (As shown): System input module, system decoding module, and system output module.

[0081] (a) The system input module includes:

[0082] First, input the emotional EEG data from the DEAP and MAHNOB datasets, along with the corresponding scale information;

[0083] The DEAP dataset collected EEG signals from 32 participants, including 16 men and 16 women. Participants wore a 32-lead EEG acquisition system with the assistance of the experimenter, and their EEG signals were collected at a sampling frequency of 512 Hz. Finally, participants followed the system prompts (such as…). Figure 3 (As shown) Participants watched evoked videos and completed related scales. Throughout the process, participants were asked to rate each 40-minute music video based on arousal, valence, liking or disliking, dominance, and familiarity.

[0084] The MAHNOB dataset collected EEG signals from 30 young, healthy adult participants, including 17 women and 13 men, aged 19 to 40. Participants wore a 32-lead EEG acquisition system with the assistance of the experimenter and followed the system prompts (such as...). Figure 3 (As shown) Participants watched evoked videos and completed related scales. The entire process lasted approximately 40 minutes, and participants labeled their emotional state based on valence and level of arousal after watching each emotional video.

[0085] The self-report component was assessed after each video based on five questions: 1. Emotional label, 2. Arousal, 3. Valence, 4. Dominance, and 5. Predictability. The emotional labels were: 1. Sadness, 2. Happiness / Joy, 3. Disgust, 4. Neutrality, 5. Amusement, 6. Anger, 7. Fear, 8. Surprise, and 9. Anxiety. Participants answered each question using a key. Note that questions 2 through 5 were scored out of 9.

[0086] (II) System Decoding Module:

[0087] As the core part of this system, the system decoding module includes a data processing module and a label prediction module.

[0088] (1) Data processing module: preprocesses the acquired emotional EEG data. The processing flow includes baseline correction, average reference and bandpass filtering of different frequency bands 4-8Hz, 8-12Hz, 12-30Hz, 30-75Hz, 30-49Hz and 51-80Hz.

[0089] The signals were categorized into four classes: Class 1 (Low Arousal, Low Valence (LALV)) – arousal less than 5 and valence greater than 5 – Class 2 (High Arousal, Low Valence (LAHV) – arousal greater than 5 and valence less than 5 – Class 3 (High Arousal, Low Valence (HALV) – arousal greater than 5 and valence greater than 5 – and Class 4 (High Arousal, High Valence (HAHV)) – arousal greater than 5 and valence greater than 5. The EEG signals corresponding to Class 1 and Class 2 labels were then used in the following experiments.

[0090] Constructing a brain network from emotional EEG data: A brain network was constructed from the preprocessed EEG data using the PLV method. Hilbert transform was applied to the time series x1 and x2 corresponding to two leads to extract the instantaneous phase and amplitude information of the signals.

[0091]

[0092]

[0093] Using the phase difference of the calculated phase values, calculate the corresponding phase-locked value (PLV) for that instant:

[0094]

[0095] Finally, using electrode locations as points in the brain network and phase-locked values ​​as corresponding functional connections, an emotional brain network is constructed.

[0096] The brain network data was divided into training and testing sets using ten-fold cross-validation. Since the brain network is a symmetric matrix, this application incorporates the upper triangular shape of the brain network from each trial of the training data, along with the feature matrix constructed under supervised conditions, into the spectral regression and robust Bayesian spectral regression models of this invention (e.g., ...). Figure 4 (The model solution process in sections 5 and 6) involves solving for the mapping matrix and class centers in the prediction model. Similarly, the mapping matrix of the robust Bayesian spectral regression model is determined by the iteration parameters. The class centers are determined by the cluster centers of the training set.

[0097] (2) Label prediction module: After the test set is reduced in dimensionality by the mapping matrix, the predicted label is determined by calculating the distance to the class center. It is assumed that the label of the class center with the closest Euclidean distance after dimensionality reduction is the same.

[0098] (III) System Output Module:

[0099] The system displays the predicted tag signal and compares the test tag with the predicted tag to calculate the corresponding accuracy. The performance of this application is then judged based on the accuracy of the predicted tag.

[0100] After completing the simulation and real experimental systems, this application further discusses the advantages and disadvantages of different feature extraction methods in the system decoding module. By using the predicted labels corresponding to different model methods obtained in the system decoding module and the prediction accuracy results of the simulation system and the real data system obtained in the system output module, the overall system situation can be obtained.

[0101] The examples provide simulation results including:

[0102] The simulation system in this application uses two types of simulation data. After repeating 100 times of 10-fold cross-validation, it compares the classification performance of the spectral regression method (SR), the 1-norm-based spectral regression method (L1SR), the Bayesian spectral regression method based on Gaussian distribution (BVSR), the Bayesian spectral regression method based on Laplace distribution (LASR), and the Bayesian spectral regression method based on Student's T distribution (STSR) under different signal-to-noise ratios, i.e., the noise suppression performance of different methods.

[0103] Simulation results show that the accuracy of all five methods decreases significantly with increasing signal-to-noise ratio, indicating that noise does indeed have a significant impact on the prediction process of spectral regression methods. Furthermore, a horizontal comparison reveals that the STSR method significantly suppresses noise better than the LASR and BVSR methods, and is also significantly better than the SR method. This demonstrates that the Bayesian spectral regression method proposed in this application can effectively suppress the influence of outliers and better capture the characteristic information of the data.

[0104] See Table 1 for the experimental results of two publicly available datasets provided in the examples, including:

[0105] The real-world data system in this application uses two publicly available datasets. With 10-fold cross-validation, the system classification output performance of the spectral regression (SR), L1-norm-based spectral regression (L1SR), Gaussian-based Bayesian spectral regression (BVSR), Laplace-based Bayesian spectral regression (LASR), and Student's T-distribution-based Bayesian spectral regression (STSR) methods on different datasets was compared.

[0106] The results in Table 1 show that the spectral regression method in the system has an accuracy similar to that of the spectral regression based on the Gaussian distribution. This is because BVSR is equivalent to the SR method from the perspective of maximum likelihood; it solves a regularized least squares problem. Under the constraint of regularization, overfitting is prevented during the problem-solving process. Therefore, compared with the general least squares method, this method has a certain degree of robustness and can still predict the label information of the data relatively well in noisy network data. At the same time, a horizontal comparison shows that the LASR method has better predictive ability for data labels than BVSR. This is because the LASR method changes the prior distribution to the Laplace distribution, which is essentially equivalent to solving a least squares problem with a norm-regularized term. As a result, this method not only performs well in solving the overfitting problem but also has a certain effect on noise suppression. Similarly, the STSR method, due to its unique heavy-tailed characteristics, performs well in noisy brain network data and shows good robustness in the label prediction process.

[0107] Tables 1 and 2 both demonstrate that the system proposed in this application, based on robust Bayesian spectral regression, is more reliable, especially for noisy EEG signals with low signal-to-noise ratios, where the system exhibits higher robustness. In summary, the entire emotion recognition system, aided by the robust Bayesian spectral regression model, can suppress the influence of artifacts on the results, improve the ability to predict subject emotion labels, and enhance the system's robustness, making the entire system more suitable for real-world noisy environments.

[0108] Advantages of the present invention

[0109] This application's system fully considers the impact of artifacts on EEG signals, thereby utilizing brain network signals with global brain information interaction connotations for dimensionality reduction. Simultaneously, considering the functional connectivity bias in brain networks caused by artifacts, this application proposes a Bayesian spectral regression system based on priors of Student's T-distribution and Laplace distribution, further suppressing the influence of noise from the perspective of data distribution fitting, improving the stability of the entire emotion recognition system, and providing a more stable recognition method for predicting emotion categories using online high-noise EEG signals.

[0110] In summary, the method proposed in this application is stable and robust, and it improves the prediction accuracy on both the DEAP and MAHNOB datasets compared to traditional spectral regression methods. Specifically, it improves the accuracy by about 11% on the binary classification MAHNOB dataset.

[0111] Experiments have demonstrated that the system of this invention performs well in online emotion decoding tasks within the field of emotion brain-computer interfaces. Through this invention, those skilled in the art will realize that the embodiments described are intended to help readers understand the principles of the invention and should be understood as not limiting the scope of protection of the invention to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the essence of the invention, and these modifications and combinations are still within the scope of protection of this invention.

[0112] Table 1 shows the experimental results of the open dataset.

[0113]

[0114] Table 2 shows the comparison of the effects of the present invention and the prior art.

[0115]

[0116]

[0117]

Claims

1. An emotion EEG network recognition system based on a Bayesian spectral regression strategy, the system comprising: System input module, system decoding module, and system output module; (I) System Input Module: After wearing the EEG data collection device, the participants watched an emotion-evoking video while their EEG signals were collected simultaneously. Throughout the data collection process, the participants were required to maintain their posture, watch the corresponding emotion video quietly, and remain silent after the video finished playing. They were also asked to complete a questionnaire while recovering their mood to avoid the previous video interfering with their emotions. (II) System Decoding Module: As a key module of the system, the system decoding module preprocesses the acquired EEG signals and calculates the corresponding EEG network. Then, it feeds the data into a robust Bayesian spectral regression model for training to obtain the corresponding mapping matrix and class center matrix. The robust Bayesian spectral regression model is an optimization of the spectral regression method by introducing Bayesian structure, making the method more robust than the original recognition method and adaptable to noisy EEG signals in real-world situations. The method for calculating the mapping matrix using a robust Bayesian spectral regression algorithm is as follows: The graphical regression method transforms the traditional Rayleigh entropy-based optimization problem into a problem of solving eigenvectors and regression. In the supervised learning process, the feature vectors are transformed into prior constraints related to the label information. The regression process can then be optimized by introducing a hierarchical Bayesian structure, and the specific implementation process is as follows: First, under supervised learning, an affinity matrix is ​​constructed, and the corresponding feature matrix is ​​also constructed. In a supervised study, the affinity matrix can be viewed as a block matrix containing c blocks of shared classes; where sample points of the same class belong to the same block in the affinity matrix. Therefore, the problem of finding the eigenvalues ​​of the affinity matrix is ​​transformed into finding the eigenvalues ​​of the block matrix, where the eigenvector of each block is represented as: y c =[0…c…c…0] Where c represents the category. Combining the characteristics of the category blocks in the affinity matrix, the feature vector of the affinity matrix is ​​determined by the category under supervised conditions. As long as two samples belong to the same category, the feature vectors in the corresponding blocks are the same. In the Bayesian spectral regression process, the feature matrix is ​​fitted using an electroencephalogram (EEG) network, and its matrix representation is as follows: Y = XG + Z Where X = [x1, x2, ..., x n ],X∈R N×(M×T) This represents the EEG network signal, where N represents the number of channels, M represents the network connections, T represents the training set length, and Y = [y1, y2, ..., y]. t ],Y∈R N×(c-1) Let Z represent the feature matrix constructed from the label signals, where Z ∈ R. T N(0,Σ) represents the Gaussian white noise that appears in the fitting problem. v ), Σ v Let G ∈ R be the covariance matrix corresponding to the Gaussian distribution. (M×T)×(c-1) This represents the corresponding mapping matrix; In the process of Bayesian spectral regression, the matrix form of the feature matrix is ​​transformed into a vector expression as: y = Xg + z; where y, g, and z are the vector representations of the feature matrix, the mapping matrix, and the noise, respectively. The prior distribution of the mapping matrix is ​​set to a Gaussian distribution g ~ N(0, Σ). g ), Σ g Let the covariance matrix of the corresponding Gaussian distribution be denoted by: and the corresponding likelihood function be: Therefore, according to Bayes' theorem, the posterior inference can be obtained as follows: p(g|y,Θ)∝p(y|X,g,Θ y )×p(g|Θ g ) Among them, Θ, Θ y Θ g This represents all the hyperparameters needed in the posterior solution process. The entire posterior formula expresses that the probability p(g|y,Θ) is proportional to the probability p(y|X,g,Θ). y ) and probability p(g|Θ g The product of ); the posterior distribution is solved using the EM method: Step 1: Calculate the expected value and the posterior distribution of the latent variables; ln(p(g|y,Θ))=E g lnp(y|X,g,Σ v )+E g lnp(g|Σ g ) Among them, E g This indicates the expectation of the parameter g in the formula; Utilizing during the solution process and Let the expected value and covariance of parameter g be represented respectively: Step 2: Maximize, take the partial derivative of the log-likelihood function from Step 1 and maximize it. The parameters updated in the kth step are: Decoding test: The dataset is mapped using the test set portion of the dataset divided by 10-fold cross-validation. Then, the mapped dimensionality reduction matrix is ​​compared with the class centers, and the closest class center is used as the predicted label. Finally, the prediction accuracy of the entire process is calculated by comparing it with the true label. (III) System Output Module: The system displays the predicted labels and outputs the prediction accuracy to determine the overall performance of the system.

2. The emotion EEG network recognition system based on Bayesian spectral regression strategy as described in claim 1, characterized in that, The calculation process of Bayesian spectral regression based on Laplace distribution prior is as follows: Represent the Laplace distribution using Gaussian and exponential distributions: in, τ≥0, meaning τ is a parameter with σ 2 A random variable with an exponential distribution; Assuming the prior distribution of the likelihood function is a Laplace distribution, the corresponding posterior distribution is solved using variational inference within the hierarchical Bayesian framework. The q-distribution, which is easier to derive through variational inference, is obtained because the exponential distribution is also represented as another form of the Gamma distribution. Therefore, based on the Gaussian conjugate relationship between the Gaussian and Gamma distributions, we obtain: Where q(·) represents the q-distribution, This represents a more easily expressible Gaussian distribution for the corresponding q-distribution, where and For its corresponding mean and variance, similarly... That is also true; The updated parameter results are obtained:

3. The emotion EEG network recognition system based on Bayesian spectral regression strategy as described in claim 1, characterized in that, The specific calculation process of Bayesian spectral regression based on the Student's T-distribution prior is as follows: Represent the student's T-distribution using Gaussian and Gamma distributions: Assuming that both the noise distribution and the likelihood function are Student's T-distributions, within the framework of hierarchical Bayesian methods, a more easily solvable q-distribution can be derived using variational Bayesian methods. Based on the conjugate relationship between the Gaussian and Gamma distributions, we can obtain: Therefore, the parameter update process can be further derived as follows:

Images