A modeling approach for coupled public-private topology pattern learning models

By coupling the public-private topological model learning model, combined with graph convolution and attention mechanism, the problem of discrete feature representation of the brain emotion recognition model in the existing technology is solved, and more efficient emotion recognition ability is achieved and understanding of brain emotional mechanisms.

CN117763464BActive Publication Date: 2024-12-06TIANJIN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311826158.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-27
Publication Date
2024-12-06
Estimated Expiration
2043-12-27

AI Technical Summary

Technical Problem

Existing research still has shortcomings when combining complex brain emotional cognitive mechanisms and advanced learning strategies, and lacks effective public-private feature representation coupling. The characteristics of graph convolutional neural network training have certain discreteness, which affects emotion recognition. Effect.

Method used

A coupled public-private topological mode learning model is proposed. By obtaining the SEED data set, filtering and calculating the differential entropy and Pearson correlation coefficients, the graph convolution module and attention mechanism module are used to train the shared domain mode subnet and private domain. Pattern subnet and improve emotion recognition capabilities through fusion modules and classifiers.

Benefits of technology

It improves the representation ability of emotion recognition, explores the potential stable modes of public and private feature representations of multiple subjects, enhances the revelation of brain emotional cognitive mechanisms, and provides a more robust emotion recognition method.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117763464B_ABST
    Figure CN117763464B_ABST
Patent Text Reader

Abstract

The present invention discloses a modeling method of a coupled public-private topological pattern learning model, comprising: obtaining a SEED data set and a SEED-IV data set to form an original data set; segmenting the original data set by a preset time unit window to obtain a plurality of segmented data; filtering each segmented data to obtain a plurality of frequency bands corresponding to each data; calculating differential entropy and Pearson correlation coefficient according to the plurality of frequency bands corresponding to each data; dividing the calculation results into a plurality of subsets according to label categories; dividing each subset by a leave-one-out method, and obtaining a training set and a test set according to the division results; training the coupled public-private topological pattern learning model by using the training set to obtain a trained coupled public-private topological pattern learning model; and testing the trained coupled public-private topological pattern learning model by using the test set. The present invention can mine hidden patterns of emotions and improve the representation ability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a modeling method for a coupled public-private topology pattern learning model. Background Art

[0002] As an important branch of affective computing, emotion recognition research aims to establish a strong human-computer interaction relationship and promote the development of computer intelligence. Existing research develops affective computing models based on facial expressions, text, speech, and physiological signals. However, non-physiological signal methods are greatly affected by individual subjective performance and are easily disguised, thereby modeling incorrect affective computing models. On the contrary, physiological signals can effectively avoid this problem. Among many physiological signals, electroencephalogram (EEG) can most intuitively reflect people's true emotional state. Its high temporal resolution and difficulty in disguise make it the most popular method. Therefore, the essence of EEG-based emotion recognition is to use computers to decode the emotional state of an individual's brain.

[0003] Most existing studies focus on manually extracting features and using machine learning to decode emotional states. The connections between horizontal and vertical electrodes are learned by extracting six different features: power spectral density (PSD), differential entropy (DE), differential asymmetry (DASM), rational asymmetry (RASM), asymmetry (ASM), and differential tailness (DCAU). The graph adjacency matrix is ​​obtained by calculating the phase-locked value (PLV), and then a one-to-one maximization strategy is used to constrain the impact of individual differences on emotional representation through the Gaussian kernel support vector machine model (GKSVM). Based on these findings, researchers are committed to decoding the brain's emotional representation, but because it is still constrained by complex neural mechanisms, the decryption of emotional cognitive mechanisms needs to be improved.

[0004] From a neuroscience perspective, the effect of emotion representation is the result of the coordinated interaction of various brain regions. Current emotion-based research mainly involves ten brain regions, namely the prefrontal lobe (FP, AF); frontal lobe (F); central frontal lobe (FC); central lobe (C); central parietal lobe (CP); left and right temporal lobe (LT, RT); parietal lobe (P); parietal occipital lobe (PO) and occipital lobe (O). They focus on capturing the relationship between local and global electrode channels to represent different emotional states.

[0005] From the perspective of pattern recognition, an advanced learning strategy can efficiently and automatically identify the dynamic changes of brain regions under different emotional states. In recent years, graph neural networks have been widely used to learn the spatiotemporal structure of the brain. They effectively combine graph learning theory, neural networks and neuroscience knowledge for EEG emotion recognition. Current advanced methods include regularized graph neural networks (RGNN), graph embedding convolutional neural networks (GECNN), instance adaptive graph method (IAG), joint model of feature adaptation and graph adaptive label propagation (JAGP), dynamic graph convolutional neural network (DGCNN), sparse dynamic graph convolutional neural network (SparseDGCNN). The above work has achieved good emotion recognition results on public datasets such as SEED, SEED-IV, SEED-V, DEAP, etc., which shows that graph-based learning strategies can effectively mine the complex spatiotemporal structure of the brain and play a good role in revealing the complex cognitive mechanism of brain emotions.

[0006] In addition, the huge differences between individuals have a serious negative impact on the recognition effect of the model. To overcome this problem, researchers proposed to use transfer learning to reduce the differences between source domain features and target domain features. For example, the domain generation graph network (DGGN) essentially adopts the adversarial learning idea to generate the potential representation of EEG to improve the similarity between the source domain and the target domain. The emotion domain adversarial neural network (EDANN) learns the hidden emotion information of the source domain and the target domain. However, the method based on domain adversarial has no potential for practical application and is only suitable for offline emotion detection. In addition to domain adversarial, a plug-and-play domain adaptation method is proposed, which can quickly update the trained model with target data and maintain high accuracy. Similarly, a frame-level teacher-student framework with data privacy (FLTSDP) is proposed. First, a frame-level network weight is trained, and then inherited with new data, so that the model can be repeatedly optimized and updated. The classification effect of this method on the DEAP dataset is more than 5% better than other models. This shows that the method of updating the training weights with target data can not only effectively improve the EEG emotion recognition effect, but also has the potential to be more suitable for real-time detection than the domain adversarial method.

[0007] However, existing research still has some defects: (1) The combination of complex brain emotion cognition mechanisms and advanced learning strategies remains to be explored; (2) There is a lack of effective coupling of public feature representations and private feature representations of multiple subjects to learn the hidden emotional features of the brain; (3) The features trained by existing graph convolutional neural networks have a certain degree of discreteness, and noise needs to be added to improve their robustness. Summary of the invention

[0008] The purpose of the present invention is to provide a modeling method for coupling public-private topological pattern learning models to mine hidden patterns of emotions and improve representation capabilities.

[0009] The technical solution of the present invention to solve the above technical problems is as follows:

[0010] The present invention provides a modeling method for a coupled public-private topology pattern learning model, comprising:

[0011] S1: Get the SEED dataset and SEED-IV dataset to form the original dataset;

[0012] S2: segmenting the original data set into a preset time unit window to obtain a plurality of segmented data;

[0013] S3: Perform filtering processing on each segmented data to obtain multiple frequency bands corresponding to each data;

[0014] S4: Calculate the differential entropy and the Pearson correlation coefficient according to the multiple frequency bands corresponding to each data to obtain a calculation result;

[0015] S5: Divide the calculation results into several subsets according to label categories;

[0016] S6: Use the leave-one-out method to divide each subset, and obtain the training set and test set according to the division results;

[0017] S7: training the coupled public-private topology pattern learning model using the training set to obtain a trained coupled public-private topology pattern learning model;

[0018] S8: Using a test set to test the trained coupled public-private topology pattern learning model.

[0019] Optionally, in S3, the plurality of frequency bands include a delta band, a theta band, an alpha band, a beta band and a gamma band;

[0020] The delta frequency band is 14 Hz,

[0021] The θ frequency band is (4-8) Hz,

[0022] The α frequency band is (8-14) Hz,

[0023] The β band is (14-30) Hz,

[0024] The gamma frequency band is (30-48) Hz.

[0025] Optionally, in S4, the differential entropy is calculated as follows:

[0026]

[0027] The Pearson correlation coefficient is calculated as:

[0028]

[0029] Where DE(X) represents the calculation result of differential entropy, x represents the input sample, μ represents the mean value, σ represents the standard deviation, π represents the circumference of a circle, log represents the logarithmic function, exp represents the exponential function with e as the base, and ρ represents the average value. XY represents the correlation coefficient between X and Y, Cov() represents the covariance, X and Y represent two different samples, E represents the mathematical expectation, D(X) represents the standard deviation of X, and D(Y) represents the standard deviation of Y.

[0030] Optionally, the coupled public-private topology pattern learning model includes a public domain pattern subnet, a private domain pattern subnet, a fusion module and a classifier;

[0031] The common domain pattern subnet is used to update electrode node features using the brain combination pattern as a common structure;

[0032] The private domain pattern subnet is used to update electrode node features using the brain topology pattern as a public structure;

[0033] The fusion module is used to fuse the output results of the public domain mode subnet and the private domain mode subnet, and output the fusion result to the classifier.

[0034] Optionally, the shared domain mode subnet and / or the private domain mode subnet includes an input layer, a graph convolution module and an attention mechanism module arranged in sequence;

[0035] The graph convolution module includes a plurality of graph convolution units, each of which includes an original graph matrix and a GNGAM matrix arranged in parallel, the original graph matrix is ​​a feature matrix output by the input layer, and the GNGAM matrix is ​​a randomly generated feature matrix;

[0036] The attention mechanism module includes several attention mechanism units, the input of each attention mechanism unit is the output result of each graph convolution unit, and the output result of each attention mechanism unit is used as the output result of the common domain mode subnet and / or the private domain mode subnet.

[0037] Optionally, the S7 includes:

[0038] S71: using Pearson calculation to train the private domain mode subnet in the coupled public-private topology mode learning model to obtain a trained private domain mode subnet;

[0039] S72: training the common domain pattern subnet in the coupled public-private topology pattern learning model by using a graph noise granularity adjacency matrix method to obtain a trained common domain pattern subnet;

[0040] S73: Obtain a trained coupled public-private topology pattern learning model according to the trained private domain pattern subnet and the trained shared domain pattern subnet.

[0041] Optionally, the S72 includes:

[0042] S721: inputting the combined brain pattern feature matrix in the training set into the common domain pattern subnetwork in the coupled public-private topological pattern learning model;

[0043] S722: Calculating a graph convolution operator according to the Laplacian matrix of the combined brain pattern feature matrix;

[0044] S723: Obtaining a graph filter common pattern encoder trained for M emotional states according to the graph convolution operator;

[0045] S724: Obtaining noise granularity filter encoders for the M emotional states according to the graph filter common pattern encoders trained for the M emotional states and the GNGAM matrix;

[0046] S725: Fusing the graph filter common mode encoder trained for the M emotional states and the noise granularity filter encoder for the M emotional states to iterate and update the parameters of the common domain mode subnet and / or the parameters of the private domain mode subnet;

[0047] S726: based on a given learning rate and function loss value, obtaining an updated combined brain pattern feature;

[0048] S727: Update the filter using the updated combined brain pattern features to obtain a trained common domain pattern subnet.

[0049] Optionally, in S722, the graph convolution operator is:

[0050]

[0051] Among them, y c Represents the calculation result of the graph convolution operator, U CBP represents the orthogonal Fourier basis, CBP represents the combined brain pattern feature matrix, g() represents the Laplace filter function, Λ CBP represents the diagonal matrix of CBP, represents the orthogonal Fourier basis transpose, D represents the differential entropy calculation result, θ n represents the n-th order Chebyshev coefficient vector, T n represents a Chebyshev polynomial of order n.

[0052] Optionally, in S723, the graph filtering common pattern encoder after the M kinds of emotional state training is:

[0053] Filter c ={Filter c1 , Filter c2 , …, Filter cM}

[0054] In S724, the noise granularity filter encoders of the M emotional states are:

[0055] Filter cn ={Filter n1 , Filter n2 , …, Filter nM}

[0056] Optionally, in S726, the updated combined brain pattern feature C is:

[0057]

[0058] Among them, L r represents the learning rate, Loss represents the function loss value, CBP represents the combined brain pattern feature matrix, GNGAM represents the graph noise granularity adjacency matrix, Represents the sign of partial derivative.

[0059] Optionally, the loss function of the coupled public-private topology pattern learning model is:

[0060] L=L task +αL CP +βL couple +λ||θ||

[0061] Among them, L task indicates mission loss and m represents m samples, n represents the nth sample, y true and pred denote the true label and the predicted label respectively, log denotes the logarithmic function, α and β denote the parameters that mutually constrain the contribution to the overall loss of the model, and L CP Indicates common loss and R c represents the potential common topological pattern, R p represents the potential private topological pattern, F represents the norm, L couple represents the coupling loss and Euc(A,B) represents the Euclidean distance between the two, E(A·B) represents the transposition of A or B multiplied by their maximum difference and subtracted, RA C Represents the potential common topological pattern after enhancement, RA Prepresents the enhanced latent private topological pattern, and λ||θ|| represents the regularized update model parameters.

[0062] The present invention has the following beneficial effects:

[0063] (1) This paper proposes a coupled public-private topological pattern learning model (CCP-LS) to mine the hidden patterns of emotions and improve the representation ability;

[0064] (2) The proposed graph noise granular adjacency matrix (GNGAM) can adaptively feature-driven update learning of potential brain topological patterns;

[0065] (3) The present invention explores the potential stable mode of coupling the shared representation and private representation of multiple subjects to improve emotion recognition ability, providing a new direction for further revealing the complex emotion cognition mechanism of the brain. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A flow chart of a modeling method for coupling a public-private topology pattern learning model of the present invention;

[0067] Figure 2 It is a structural diagram of the coupled public-private topology mode learning model of the present invention. DETAILED DESCRIPTION

[0068] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0069] The present invention provides a modeling method for coupling a public-private topology pattern learning model, referring to Figure 1 As shown, including:

[0070] S1: Get the SEED dataset and SEED-IV dataset to form the original dataset;

[0071] The SEED dataset contains 15 subjects and three experiments. Each subject needs to watch 15 videos that can induce positive, neutral and negative emotions in each experiment. The experiment collected data from 62 electrode channels of each subject, and performed certain processing and feature calculations. The present invention uses all 45 sessions to evaluate the proposed learning strategy.

[0072] The SEED-IV dataset is similar to the SEED dataset, and also contains 15 subjects and three experiments. However, each subject needs to watch 24 videos that can induce four discrete emotions: happiness, sadness, fear, and neutrality. Similarly, the experiment collected data from 62 electrode channels for each subject, and performed certain processing and feature calculations. The present invention uses all 45 sessions to evaluate the proposed learning strategy.

[0073] S2: segmenting the original data set into a preset time unit window to obtain a plurality of segmented data;

[0074] It should be noted that those skilled in the art can set the preset time unit according to actual conditions, and the present invention does not impose any specific limitation. As an implementation mode, the preset time unit of the present invention is 1s.

[0075] S3: Perform filtering processing on each segmented data to obtain multiple frequency bands corresponding to each data;

[0076] The plurality of frequency bands include a delta band, a theta band, an alpha band, a beta band, and a gamma band;

[0077] The delta band is 14 Hz, the theta band is (4-8) Hz, the alpha band is (8-14) Hz, the beta band is (14-30) Hz, and the gamma band is (30-48) Hz.

[0078] S4: Calculate the differential entropy and the Pearson correlation coefficient according to the multiple frequency bands corresponding to each data to obtain a calculation result;

[0079] The calculation method of differential entropy is:

[0080]

[0081] The Pearson correlation coefficient is calculated as:

[0082]

[0083] Where DE(X) represents the calculation result of differential entropy, x represents the input sample, μ represents the mean value, σ represents the standard deviation, π represents the circumference of a circle, log represents the logarithmic function, exp represents the exponential function with e as the base, and ρ represents the average value. XY represents the correlation coefficient between X and Y, Cov() represents the covariance, X and Y represent two different samples, E represents the mathematical expectation, D(X) represents the standard deviation of X, and D(Y) represents the standard deviation of Y.

[0084] The present invention uses the Pearson correlation coefficient because the Pearson correlation coefficient can calculate the positive and negative correlation of electrode channels.

[0085] S5: Divide the calculation results into several subsets according to label categories;

[0086] Here, the label categories are different emotional states.

[0087] S6: Use the leave-one-out method to divide each subset, and obtain a training set and a test set according to the division results; since the leave-one-out method is a prior art, the present invention will not be described in detail.

[0088] S7: training the coupled public-private topology pattern learning model using the training set to obtain a trained coupled public-private topology pattern learning model;

[0089] Here, reference Figure 2 As shown, the coupled public-private topology mode learning model provided by the present invention includes a shared domain mode subnet, a private domain mode subnet, a fusion module and a classifier;

[0090] The common domain pattern subnet is used to update electrode node features using the brain combination pattern as a common structure;

[0091] The private domain pattern subnet is used to update electrode node features using the brain topology pattern as a public structure;

[0092] The fusion module is used to fuse the output results of the public domain mode subnet and the private domain mode subnet, and output the fusion result to the classifier.

[0093] The shared domain mode subnet and / or the private domain mode subnet include an input layer, a graph convolution module and an attention mechanism module arranged in sequence;

[0094] The graph convolution module includes a plurality of graph convolution units, each of which includes an original graph matrix and a GNGAM matrix arranged in parallel, the original graph matrix is ​​a feature matrix output by the input layer, and the GNGAM matrix is ​​a randomly generated feature matrix;

[0095] The attention mechanism module includes several attention mechanism units, the input of each attention mechanism unit is the output result of each graph convolution unit, and the output result of each attention mechanism unit is used as the output result of the common domain mode subnet and / or the private domain mode subnet.

[0096] Based on this, S7 specifically includes:

[0097] S71: training the private domain mode subnet in the coupled public-private topology mode learning model by using Pearson calculation to obtain a trained private domain mode subnet;

[0098] S72: training the common domain pattern subnet in the coupled public-private topology pattern learning model by using a graph noise granularity adjacency matrix method to obtain a trained common domain pattern subnet;

[0099] As a specific implementation, the specific training process provided by the present invention includes:

[0100] S721: inputting the combined brain pattern feature matrix in the training set into the common domain pattern subnetwork in the coupled public-private topological pattern learning model;

[0101] The training set is recorded as:

[0102]

[0103] Among them, the characteristics of a single subject can be expressed as:

[0104] K represents the electrode channel set.

[0105] Here, the number of subjects in each data set is recorded as S, and S-1 subjects are selected in sequence each time for combined brain pattern calculation. Specifically, the correlation coefficient of each corresponding electrode channel (K) in each frequency band (B) of S-1 subjects is calculated, and the calculation formula is as follows:

[0106]

[0107] For the combined features of the subjects, the average dimensionality reduction method is used to reduce (S-1)*(S-1) to one dimension, from which we can obtain:

[0108]

[0109] right The last two dimensions of are matrix transposed and multiplied to obtain the combined brain pattern representation:

[0110]

[0111] S722: Calculating a graph convolution operator according to the Laplacian matrix of the combined brain pattern feature matrix;

[0112] Input CBP into CDM-SubNet and update D features through filter encoding and attention decoding. For the extracted CBP and D, record L CBP is the Laplace matrix of the matrix CBP, so based on the formula y=g(L)x=g(UΛU T )x=Ug(Λ)U T x can be obtained:

[0113]

[0114] In order to reduce the computational complexity, g(Λ CBP ) is converted to an n-th order Chebyshev polynomial:

[0115]

[0116] where θ n represents the Chebyshev coefficient vector, T n (Λ CBP ) represents the Chebyshev polynomial based on the diagonal matrix CBP, so the formula can be converted to:

[0117]

[0118] Among them, y c Represents the calculation result of the graph convolution operator, U CBP represents the orthogonal Fourier basis, CBP represents the combined brain pattern feature matrix, g() represents the Laplace filter function, Λ CBP represents the diagonal matrix of CBP, represents the orthogonal Fourier basis transpose, D represents the differential entropy calculation result, θ n represents the n-th order Chebyshev coefficient vector, T n represents a Chebyshev polynomial of order n.

[0119] S723: Obtaining a graph filter common pattern encoder trained for M emotional states according to the graph convolution operator;

[0120] The graph filter common pattern encoder after training M emotional states is:

[0121] Filter c ={Filter c1 , Filter c2 , …, Filter cM}

[0122] S724: Obtaining noise granularity filter encoders for the M emotional states according to the graph filter common pattern encoders trained for the M emotional states and the GNGAM matrix;

[0123] At the same time, in order to improve the robustness of the trained common patterns, the present invention proposes a graph noise granular adjacency matrix (GNGAM) for adaptive feature-driven update learning of potential common topological patterns. For the randomly generated K*K GNGAM, the formula The noise granularity filter encoder for M emotional states can be obtained as:

[0124] Filter cn ={Filter n1 , Filter n2 , …, Filter nM}

[0125] S725: Fusing the graph filter common mode encoder trained for the M emotional states and the noise granularity filter encoder for the M emotional states to iterate and update the parameters of the common domain mode subnet;

[0126] Then through the fusion filter cand Filter cn To update the iterative model parameters:

[0127] Filter c+cn ={Filter c1+n1 , Filter c2+n2 , …, Filter cM+nM}

[0128] S726: based on a given learning rate and function loss value, obtaining an updated combined brain pattern feature;

[0129] Here we describe the process of GNGAM updating CBP. For a given learning rate L r And the function loss value Loss is obtained by simultaneously updating CBP and GNGAM through back propagation and Then you can update CBP to C:

[0130]

[0131] Among them, L r represents the learning rate, Loss represents the function loss value, CBP represents the combined brain pattern feature matrix, GNGAM represents the graph noise granularity adjacency matrix, Represents the sign of partial derivative.

[0132] S727: Update the filter using the updated combined brain pattern features to obtain a trained common domain pattern subnet.

[0133] Finally, the potential robust common topological pattern is obtained by updating the iterative filter.

[0134]

[0135] Where L C and W represent the updated structural mode and filter weight parameters respectively.

[0136] Next, the present invention uses the attention mechanism to perform decoding operations. The attention mechanism can learn the key frequency bands and brain regions reflected by various emotional states. C , the attention score of the brain area activation state corresponding to each emotion can be expressed as:

[0137]

[0138] where θ CM and θ CM 'represent the parameters of the two fully connected layers, σ soft represents the Softmax activation function. Then the attention score and R CPotential shared topological patterns after multiplication are strengthened.

[0139]

[0140] Similarly, the enhanced potential private topology pattern can be expressed as:

[0141]

[0142] After obtaining the latent shared and private topological patterns, they are concatenated and fed into a classifier to identify emotions.

[0143]

[0144]

[0145] in, represents the prediction result of the classifier, σ sig represents the Sigmoid activation function, θ CP and θ CP 'represent the parameters of the two fully connected layers respectively.

[0146] S73: Obtain a trained coupled public-private topology pattern learning model according to the trained private domain pattern subnet and the trained shared domain pattern subnet.

[0147] During the model training process, there are three loss functions. The first one is the classification task loss L task The second is the common loss L of the public-private topological pattern learned after the encoder CP , which allows the model to learn the potential common subspace. The third is the coupling loss L that strengthens the public-private topological pattern learned after the decoder. couple , its essence is to propose the idea of ​​orthogonalization of soft sub-patterns to extract the specific pattern information after decoding to couple the public-private representation. Therefore, the loss function of the coupled public-private topological pattern learning model of the present invention is:

[0148] L=L task +αL CP +βL couple +λ||θ||

[0149] Among them, L task represents the classification task loss and m represents m samples, n represents the nth sample, y true and pred denote the true label and the predicted label respectively, log denotes the logarithmic function, α and β denote the parameters that mutually constrain the contribution to the overall loss of the model, and L CP Indicates common loss and Rc represents the potential common topological pattern, R p represents the potential private topological pattern, F represents the norm, L couple represents the coupling loss and Euc(A,B) represents the Euclidean distance between the two, E(A·B) represents the transposition of A or B multiplied by their maximum difference and subtracted, RA C Represents the potential common topological pattern after enhancement, RA P represents the enhanced latent private topological pattern, and λ||θ|| represents the regularized update model parameters.

[0150] The following describes the convergence process of the coupled public-private topology pattern learning model (CCP-LS). The essence of this learning strategy is to continuously reduce the final RA through iterative updates. C and RA P Here we record RA Cn l , R.A. Pn l represents the nth sample after the lth iteration. The present invention assumes that the two samples have a large difference before the iteration, that is, RA CPn l = |RA Cn l -RA Pn l | norm approaches 1. During the iteration process, RA CPn l It keeps decreasing and tends to a robust value r∈(0,1). Therefore, the differential normalization equation can be expressed as:

[0151]

[0152] Can be transformed into:

[0153]

[0154] However, and They can be expressed as:

[0155]

[0156] From the basic inequality principle, we can get:

[0157]

[0158] therefore:

[0159]

[0160] Can get RACPn l ≤RA CPn l-1 , which means that the difference of public-private topological patterns after each iteration is smaller than that before the iteration. CPn l =RA CPn l-1 When CPn l The robust value, lim, has been reached l→∞ (RA CPn l )=r.

[0161] S8: Using a test set to test the trained coupled public-private topology pattern learning model.

[0162] For a given test set Γ test ={(PBP I+1 ,D I+1 )},with{I+1=(s),…,(S)}, and use it to update the already trained filters to extract multiple emotion state patterns.

[0163] RA (test )=Filter RA (PBP I+1 ;D I+1 )

[0164] Filter RA Represents the trained public-private topology pattern filter. The prediction result is then input through the fully connected layer.

[0165]

[0166] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A modeling method for a coupled public-private topology pattern learning model, characterized in that: include: S1: Get the SEED dataset and SEED-IV dataset to form the original dataset; S2: segmenting the original data set into a preset time unit window to obtain a plurality of segmented data; S3: Perform filtering processing on each segmented data to obtain multiple frequency bands corresponding to each data; S4: Calculate the differential entropy and the Pearson correlation coefficient according to the multiple frequency bands corresponding to each data to obtain a calculation result; S5: Divide the calculation results into several subsets according to label categories; S6: Use the leave-one-out method to divide each subset, and obtain the training set and test set according to the division results; S7: training the coupled public-private topology pattern learning model using the training set to obtain a trained coupled public-private topology pattern learning model; The coupled public-private topology pattern learning model includes a shared domain pattern subnet, a private domain pattern subnet, a fusion module and a classifier; The common domain pattern subnet is used to update electrode node features using the brain combination pattern as a common structure; The private domain pattern subnet is used to update electrode node features using the brain topology pattern as a public structure; The fusion module is used to fuse the output results of the public domain mode subnet and the private domain mode subnet, and output the fusion result to the classifier; The shared domain mode subnet and / or the private domain mode subnet include an input layer, a graph convolution module and an attention mechanism module arranged in sequence; The graph convolution module includes a plurality of graph convolution units, each of which includes an original graph matrix and a GNGAM matrix arranged in parallel, the original graph matrix is ​​a feature matrix output by the input layer, and the GNGAM matrix is ​​a randomly generated feature matrix; The attention mechanism module includes a plurality of attention mechanism units, the input of each of the attention mechanism units is the output result of each of the graph convolution units, and the output result of each of the attention mechanism units is used as the output result of the common domain mode subnet and / or the private domain mode subnet; S8: Testing the trained coupled public-private topology pattern learning model using a test set; The step S7 specifically includes the following sub-steps: S71: using Pearson calculation to train the private domain mode subnet in the coupled public-private topology mode learning model to obtain a trained private domain mode subnet; S72: training the common domain pattern subnet in the coupled public-private topology pattern learning model by using a graph noise granularity adjacency matrix method to obtain a trained common domain pattern subnet; S73: Obtaining a trained coupled public-private topology pattern learning model according to the trained private domain pattern subnet and the trained shared domain pattern subnet; The step S72 specifically includes the following sub-steps: S721: Inputting the combined brain pattern feature matrix in the training set into the common domain pattern subnet in the coupled public-private topological pattern learning model, specifically including the following formula: The training set is recorded as: Γ train ={(CBP1,CBP2,…,CBP I ),(PBP1,PBP2,…,PBP I ),(D1,D2,…,D I )}, Among them, the characteristics of a single subject can be expressed as: K represents the electrode channel set; The number of subjects in each data set is recorded as S, and S-1 subjects are selected in sequence each time for combined brain pattern calculation. Specifically, the correlation coefficient of each corresponding electrode channel in each frequency band of S-1 subjects is calculated. The calculation formula is: Among them, B represents the frequency band of each subject; For the combined features of the subjects, the average dimensionality reduction method is used to reduce (S-1)*(S-1) to one dimension, from which we can get: right The last two dimensions of are matrix transposed and multiplied to obtain a combined brain model. The combined brain model is represented by the formula: S722: Calculate a graph convolution operator according to the Laplacian matrix of the combined brain pattern feature matrix, specifically including the following formula: Input CBP into CDM-SubNet and update D features through filter encoding and attention decoding. For the extracted CBP and D, record L CBP As the Laplace matrix of matrix CBP, we can get: In order to reduce the computational complexity, g(Λ CBP ) is converted to an n-th order Chebyshev polynomial: where θ n represents the Chebyshev coefficient vector, T n (Λ CBP ) represents the Chebyshev polynomial based on the diagonal matrix CBP, so the formula can be converted to: Among them, y c Represents the calculation result of the graph convolution operator, U CBP represents the orthogonal Fourier basis, CBP represents the combined brain pattern feature matrix, g() represents the Laplace filter function, Λ CBP represents the diagonal matrix of CBP, represents the orthogonal Fourier basis transpose, and D represents the differential entropy calculation result; S723: According to the graph convolution operator, a graph filter common pattern encoder trained for M kinds of emotional states is obtained, wherein the graph filter common pattern encoder trained for M kinds of emotional states is: Filter c ={Filter c1 ,Filter c2 ,…,Filter cM }; S724: According to the graph filter common pattern encoder and GNGAM matrix trained for M emotional states, a noise granularity filter encoder for M emotional states is obtained. The noise granularity filter encoder for the M emotional states is: Filter cn ={Filter n1 ,Filter n2 ,…,Filter nM }; S725: Fusing the graph filter common mode encoder trained for the M emotional states and the noise granularity filter encoder for the M emotional states to iterate and update the parameters of the common domain mode subnet and / or the parameters of the private domain mode subnet, specifically: By fusion filter c and Filter cn Update iterative model parameters: Filter c+cn ={Filter c1+n1 ,Filter c2+n2 ,…,Filter cM+nM }; S726: Based on the given learning rate and function loss value, an updated combined brain pattern feature is obtained, which specifically includes the following formula: For a given learning rate L r And the function loss value Loss is obtained by simultaneously updating CBP and GNGAM through back propagation and Then the updated combined brain pattern feature C is: Among them, CBP represents the combined brain pattern feature matrix, GNGAM represents the graph noise granularity adjacency matrix, represents the sign of partial derivative; S727: Update the filter using the updated combined brain pattern feature to obtain a trained common domain pattern subnet, and finally obtain a potential robust common topology pattern by updating the iterative filter. The specific formula is: R C ={R C1 ,R C2 ,…,R CM }, Among them, L C and W represent the updated structural mode and filter weight parameters respectively; The loss function of the coupled public-private topology pattern learning model is: L=L task +αL CP +βL couple +λ||θ|| Among them, L task indicates mission loss and m represents m samples, n represents the nth sample, y true and pred denote the true label and the predicted label respectively, log denotes the logarithmic function, α and β denote the parameters that mutually constrain the contribution to the overall loss of the model, and L CP Indicates common loss and R c represents the potential common topological pattern, R p represents the potential private topological pattern, F represents the norm, L couple represents the coupling loss and Euc(A,B) represents the Euclidean distance between the two, E(A·B) represents the transposition of A or B multiplied by their maximum difference and subtracted, RA C Represents the potential common topological pattern after enhancement, RA P represents the enhanced latent private topological pattern, and λ||θ|| represents the regularized update model parameters.

2. The modeling method of the coupled public-private topology pattern learning model according to claim 1, characterized in that: In said S3, the plurality of frequency bands include a delta band, a theta band, an alpha band, a beta band and a gamma band; The delta frequency band is 14 Hz, The θ frequency band is (4-8) Hz, The α frequency band is (8-14) Hz, The β band is (14-30) Hz, The gamma frequency band is (30-48) Hz.

3. The modeling method of the coupled public-private topology pattern learning model according to claim 1, characterized in that: In S4, the differential entropy is calculated as follows: The Pearson correlation coefficient is calculated as: Where DE(X) represents the calculation result of differential entropy, x represents the input sample, μ represents the mean value, σ represents the standard deviation, π represents the circumference of a circle, log represents the logarithmic function, exp represents the exponential function with e as the base, and ρ represents the average value. XY represents the correlation coefficient between X and Y, Cov() represents the covariance, X and Y represent two different samples, E represents the mathematical expectation, D(X) represents the standard deviation of X, and D(Y) represents the standard deviation of Y.