Electroencephalogram emotion recognition method based on double-flow adaptive graph temporal interaction network

Abstract

A brainwave emotion recognition method based on a dual-stream adaptive graph temporal interaction network is proposed, comprising brainwave signal preprocessing, key feature extraction, construction of a dual-branch brainwave emotion recognition network, training of the dual-branch brainwave emotion recognition network, and testing of the dual-branch brainwave emotion recognition network. Because this invention employs a dual-branch brainwave emotion recognition network, it solves the technical problems of existing brainwave emotion recognition methods, such as the inability of static graph topology to adapt to dynamic changes in emotions, insufficient temporal modeling capabilities, insufficient fusion of frequency and spatiotemporal information, and simplistic and coarse cross-modal feature fusion. It achieves deep interaction and information purification of dual-branch features, improving the accuracy and robustness of emotion recognition. Comparative experimental results show that the average accuracy of this invention on the SEED dataset is 97.82%. This invention has advantages such as high emotion recognition accuracy and good model interpretability, and can be used in brainwave emotion recognition technology in fields such as human-computer interaction.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of brain-computer interface and emotion computing, specifically a dual-branch EEG emotion recognition solution that integrates a dual-stream adaptive graph temporal interactive brain network architecture. Technical Background

[0002] Electroencephalogram (EEG) signals possess significant advantages such as high temporal resolution and non-invasive acquisition, and can directly reflect the real activity state of brain neurons, making them a core analytical tool of great research value in the field of affective computing. However, achieving high-precision and highly stable emotion classification from raw EEG time-series data still faces many technical challenges that urgently need to be overcome.

[0003] At the feature extraction level, traditional EEG emotion recognition solutions largely rely on manually defined features, such as time-domain statistical features and frequency-domain energy distribution features. While this type of manually constructed feature can characterize the basic brain activity patterns to some extent, the overall design process is cumbersome and time-consuming. Furthermore, manually extracted features can only partially depict single-dimensional information of the signal and cannot fully capture the complex, high-dimensional, full-domain brain activity characteristics corresponding to emotional states. In addition, manually designed features heavily depend on prior human experience, making it difficult to adapt to the inherent characteristics of EEG signals—strong individual variability, non-stationarity, and strong noise interference—and easily overlooking deeply hidden emotion discrimination information. The feature representation ability has a clear upper limit, ultimately limiting the classification accuracy and stability of emotion recognition.

[0004] With the continuous evolution of deep learning technology, convolutional neural networks (CNNs) have been widely used in EEG emotion recognition tasks. However, CNNs rely on a regular grid-based convolutional computation paradigm, which is difficult to adapt to the non-Euclidean spatial distribution characteristics of EEG channels on the scalp. The induction and processing of emotions involve the coordinated operation between multiple brain regions, and dynamic functional connectivity patterns play a crucial role in this process. Convolutional structures struggle to effectively capture this type of global, non-local functional association information between brain regions.

[0005] The introduction of graph convolutional neural networks has opened up new technical avenues for processing the irregular spatial topology of EEG signals. However, most existing methods still use static adjacency matrices pre-defined based on electrode physical locations to construct brain map topology. This fixed topology ignores the continuous dynamic reorganization of brain functional networks during emotional processing; that is, the functional coupling strength between brain regions exhibits adaptive fluctuations under different emotional states, which severely limits the model's emotion discrimination ability.

[0006] Furthermore, research in affective neuroscience indicates that emotional processing involves multi-scale characteristics of brain activity. At the frequency domain level, different emotional experiences correspond to specific frequency band neural oscillation energy variation patterns; at the time domain level, emotional responses depend on the evolution of nonlinear dynamic characteristics of EEG signals. How to organically integrate frequency and time domain information to construct a deep learning model capable of simultaneously capturing spatial topology and temporal dependencies has become a core issue in improving EEG emotion recognition performance. In recent years, state-space models and their representative work, Mamba, have provided an innovative technical path for temporal modeling of EEG signals due to their linear computational complexity advantage in long sequence modeling.

[0007] To address the aforementioned issues, this invention proposes three core innovative modules: an adaptive dual-topology graph neural network in the frequency-spatial domain, which fuses a static structured brain network based on electrode location with an adaptive brain network based on multi-head self-attention, and dynamically learns emotion-related brain functional connectivity patterns using a gating mechanism; a spatially embedded temporal Mamba network in the spatiotemporal domain, which achieves synergistic enhancement of spatiotemporal information through a spatial prior channel coding network and a bidirectional Mamba layer; and an asymmetric residual cross-feature modulation module in the fusion stage, which achieves deep fusion and information purification of dual-branch features through asymmetric bidirectional modulation and a multi-head attention mechanism. These three innovative modules work together synergistically to significantly improve the classification accuracy and model robustness of the final emotion recognition task. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network with high emotion recognition accuracy and good model interpretability.

[0009] The technical solution adopted to solve the above technical problems consists of the following steps:

[0010] (1) Preprocessing of EEG signals

[0011] Raw multi-channel EEG signals were collected from the SEED public dataset. The SEED dataset has three emotion categories: sadness, neutrality, and pleasure. The raw EEG signals were segmented using a sliding window method with a window length of 400 sampling points and a step size of 200 sampling points. A 5-fold cross-validation strategy was used to divide the dataset into training and validation sets at an 8:2 ratio.

[0012] (2) Extracting key features

[0013] The EEG signal was decomposed into δ, θ, α, β, and γ frequency bands using a Butterworth bandpass filter. Differential entropy features of each frequency band were extracted using a differential entropy calculation method based on the Gaussian assumption. Power spectral density features of each frequency band were extracted using the Welch power spectral density estimation method. Power ratio features between the θ and α bands and between the α and β bands were extracted using the power spectral density ratio method. Spectral centroid features of the entire frequency band were extracted using the power spectral density weighted average method. Spectral flatness features of the entire frequency band were extracted using the ratio of the geometric mean to the arithmetic mean of the power spectral density. The above features were concatenated to obtain a 14-dimensional frequency domain feature vector.

[0014] (3) Constructing a two-branch EEG emotion recognition network

[0015] The bi-branch EEG emotion recognition network consists of an adaptive dual-topology graph neural network and a spatially embedded temporal Mamba network connected in parallel, followed by an asymmetric residual cross-feature modulation network.

[0016] (4) Training a bibranch EEG emotion recognition network

[0017] 1) Constructing the loss function

[0018] Construct the loss function L according to equation (1):

[0019] (1)

[0020] Where N represents the total number of training samples, ranging from 24 to 32, and C represents the number of emotion categories, with a value of 3. Here, represents the weighting coefficient for class c, and γ is the focusing parameter, ranging from 2.2 to 4.5. This is a soft label that has been smoothed. This represents the predicted probability that the i-th sample belongs to the c-th class.

[0021] 2) Training a bibranch EEG emotion recognition network

[0022] The training set was input into the two-branch EEG emotion recognition network for training. The training parameters were as follows: AdamW was selected as the optimizer, with a basic initial learning rate of 0.0022 and a weight decay coefficient of 0.04; the SEED standard batch size was 32 and the maximum number of iterations was 100, and training was carried out until the loss function converged.

[0023] (5) Testing the bibranch EEG emotion recognition network

[0024] The test set is input into the trained two-branch EEG emotion recognition network for testing, and the EEG emotion recognition results are output.

[0025] In step (3) of the present invention, the construction of the dual-branch EEG emotion recognition network is composed of an adaptive dual-topology graph neural network, a static structured brain network connected in parallel, an adaptive gating fusion network, a dynamic graph convolutional network, and a channel attention module connected in series.

[0026] In step (3) of the present invention, the construction of the bi-branch EEG emotion recognition network is described in which the spatially embedded temporal Mamba network is composed of a spatial prior channel coding network, a channel projection layer, a bidirectional Mamba network, and a temporal attention pooling layer connected in series.

[0027] The adaptive brain network of this invention is composed of a global average pooling layer 1, a fully connected layer 1, a fully connected layer 2, a multi-head attention layer 1, a back projection layer, a feature fusion and normalization layer, and a dynamic adjacency matrix generation layer connected in series.

[0028] The adaptive gating fusion network of the present invention is composed of a fully connected layer 3, a fully connected layer 4, a fully connected layer 5, and a weighted fusion layer 1 connected in series.

[0029] The spatial prior channel coding network of this invention is composed of a compression-excitation attention branch, a channel mixing branch, and the output of the spatial prior branch connected in parallel, followed by a weighted fusion layer 2 and a layer normalization layer connected in series. The compression-excitation attention branch is composed of a global average pooling layer 2 connected in series with a fully connected layer 6, a fully connected layer 7, and a channel feature weighting layer.

[0030] In step (3) of the present invention, the construction of the dual-branch EEG emotion recognition network is composed of an asymmetric residual cross-feature modulation network, a residual gating fusion module, a multi-head attention layer 2, and a classification module connected in series.

[0031] The asymmetric cross-modulation module of this invention is composed of fully connected layers 8, 9, 10, and 11, and a FiLM modulation layer connected together. Fully connected layers 8 and 9 are connected in series, and fully connected layers 10 and 11 are connected in series. The outputs of fully connected layers 10 and 11 are connected to the FiLM modulation layer. The residual gated fusion module is composed of fully connected layers 12 and 13 connected in series.

[0032] The method for constructing the FiLM modulation layer of the present invention is as follows:

[0033] Perform bidirectional cross-branch eigenmodulation calculation using the following formula:

[0034]

[0035]

[0036] in, To spatially embed the original temporal features of the temporal Mamba network output, The original frequency domain features output by the adaptive dual-topology graph neural network; , The scaling and offset factors are used to guide the modulation of timing features in the frequency domain branch. , The scaling and offset coefficients for frequency domain feature modulation guided by the time-series branch are all obtained through network training. For Hadamard's element-wise multiplication operation; The time-series characteristics modulated by frequency domain information, The above operations construct a FiLM modulation layer, which represents the frequency domain characteristics modulated by time-series information.

[0037] This invention employs a dual-stream adaptive graph temporal interactive brain network. An adaptive dual-topology graph neural network is constructed in the frequency-spatial domain branch, fusing a static structured brain network based on electrode location with an adaptive brain network based on multi-head self-attention. Through a gating fusion mechanism, it dynamically learns emotion-related brain function connectivity patterns. Simultaneously, a spatially embedded temporal Mamba network is constructed in the spatiotemporal domain branch. A spatial prior channel coding network models the spatial relationships between channels, and a bidirectional Mamba network is introduced to capture long-range temporal dependencies. This addresses the technical problems of existing EEG emotion recognition methods, such as the difficulty of adapting static graph topology to dynamic emotion changes, insufficient temporal modeling capabilities, and inadequate fusion of frequency and spatiotemporal information. Furthermore, through an asymmetric residual cross-feature modulation network, employing an asymmetric bidirectional feature modulation strategy and a residual gating fusion mechanism, deep interaction and information purification of the dual-branch features are achieved. This solves the technical problem of simple and coarse cross-modal feature fusion in existing methods, improving the accuracy and robustness of emotion recognition. The present invention was compared with existing EEG emotion recognition models on the SEED and SEED-IV datasets. The experimental results show that the present invention achieves an average accuracy of 97.82% on the SEED dataset and an average accuracy of 93.91% on the SEED-IV dataset. The present invention has the advantages of high emotion recognition accuracy and good model interpretability, and can be used in EEG emotion recognition technology in fields such as human-computer interaction. Attached Figure Description

[0038] Figure 1 This is a flowchart of Embodiment 1 of the present invention.

[0039] Figure 2 This is a schematic diagram of the structure of a two-branch EEG emotion recognition network.

[0040] Figure 3 yes Figure 2 A schematic diagram of the structure of the adaptive brain network.

[0041] Figure 4 yes Figure 2 A schematic diagram of the structure of the adaptive gating fusion network.

[0042] Figure 5 yes Figure 2 A schematic diagram of the structure of a mid-space prior channel coding network.

[0043] Figure 6 yes Figure 2 A schematic diagram of the structure of asymmetric residual cross-feature modulation network.

[0044] Figure 7 The curves show the comparison of the recognition accuracy of the method in Example 1 with existing comparison algorithms on the SEED dataset for each subject. Detailed Implementation

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.

[0046] Example 1

[0047] like Figure 1 As shown, the EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network in this embodiment consists of the following steps.

[0048] (1) Preprocessing of EEG signals

[0049] Raw multi-channel EEG signals were collected from the SEED public dataset. The SEED dataset has three emotion categories: sadness, neutrality, and pleasure. The raw EEG signals were segmented using a sliding window method with a window length of 400 sampling points and a step size of 200 sampling points. A 5-fold cross-validation strategy was used to divide the dataset into training and validation sets at an 8:2 ratio.

[0050] (2) Extracting key features

[0051] The EEG signal was decomposed into δ, θ, α, β, and γ frequency bands using a Butterworth bandpass filter. Differential entropy features of each frequency band were extracted using a differential entropy calculation method based on the Gaussian assumption. Power spectral density features of each frequency band were extracted using the Welch power spectral density estimation method. Power ratio features between the θ and α bands and between the α and β bands were extracted using the power spectral density ratio method. Spectral centroid features of the entire frequency band were extracted using the power spectral density weighted average method. Spectral flatness features of the entire frequency band were extracted using the ratio of the geometric mean to the arithmetic mean of the power spectral density. The above features were concatenated to obtain a 14-dimensional frequency domain feature vector.

[0052] (3) Constructing a two-branch EEG emotion recognition network

[0053] Figure 2 A schematic diagram of the structure of the dual-branch EEG emotion recognition network in this embodiment is provided. Figure 2 In this embodiment, the dual-branch EEG emotion recognition network consists of an adaptive dual-topology graph neural network and a spatially embedded temporal Mamba network connected in parallel, followed by an asymmetric residual cross-feature modulation network.

[0054] The adaptive dual-topology graph neural network in this embodiment is composed of an adaptive brain network, a static structure brain network connected in parallel, an adaptive gating fusion network, a dynamic graph convolutional network, and a channel attention module connected in series.

[0055] The spatially embedded temporal Mamba network in this embodiment is composed of a spatial prior channel coding network, a channel projection layer, a bidirectional Mamba network, and a temporal attention pooling layer connected in series.

[0056] Figure 3 Given Figure 2 A schematic diagram of the structure of a mid-adaptive brain network. Figure 3 In this embodiment, the adaptive brain network consists of a global average pooling layer 1, a fully connected layer 1, a fully connected layer 2, a multi-head attention layer 1, a back projection layer, a feature fusion and normalization layer, and a dynamic adjacency matrix generation layer, sequentially connected in series. This network is used to dynamically learn emotion-related brain function connectivity patterns from input multi-channel EEG features.

[0057] Figure 4 Given Figure 2 A schematic diagram of the structure of an adaptive gating fusion network. Figure 4 In this embodiment, the adaptive gating fusion network is composed of a fully connected layer 3, a fully connected layer 4, a fully connected layer 5, and a weighted fusion layer 1 connected in series.

[0058] Figure 5 Given Figure 2 A schematic diagram of the structure of a mid-space prior channel coding network. Figure 5 In this embodiment, the spatial prior channel coding network consists of a compression-excitation attention branch, a channel mixing branch, and the output of the spatial prior branch connected in parallel, followed by a weighted fusion layer 2 and a layer normalization layer connected in series. This network is used to model the spatial relationships between channels, enhancing the representation capability of spatiotemporal features.

[0059] In this embodiment, the compressed-excitation attention branch is composed of a global average pooling layer 2, a fully connected layer 6, a fully connected layer 7, and a channel feature weighting layer connected in series.

[0060] Figure 6 Given Figure 2 A schematic diagram of the structure of an asymmetric residual crossover feature modulation network. Figure 6In this embodiment, the asymmetric residual cross-feature modulation network is composed of an asymmetric cross-modulation module, a residual gated fusion module, a multi-head attention layer 2, and a classification module connected in series. This network is used to achieve deep interaction and adaptive fusion of frequency-spatial domain features and spatiotemporal domain features, generating a highly discriminative joint feature representation.

[0061] The asymmetric cross-modulation module in this embodiment is composed of a fully connected layer 8, a fully connected layer 9, a fully connected layer 10, a fully connected layer 11, and a FiLM modulation layer connected together. The fully connected layer 8 and the fully connected layer 9 are connected in series, the fully connected layer 10 and the fully connected layer 11 are connected in series, and the output terminals of the fully connected layers 10 and 11 are connected to the FiLM modulation layer.

[0062] The residual gated fusion module is composed of a fully connected layer 12 and a fully connected layer 13 connected in series.

[0063] The method for constructing the FiLM modulation layer is as follows:

[0064] Perform bidirectional cross-branch eigenmodulation calculation using the following formula:

[0065]

[0066]

[0067] in, To spatially embed the original temporal features of the temporal Mamba network output, The original frequency domain features output by the adaptive dual-topology graph neural network; , The scaling and offset factors are used to guide the modulation of timing features in the frequency domain branch. , The scaling and offset coefficients for frequency domain feature modulation guided by the time-series branch are all obtained through network training. For Hadamard's element-wise multiplication operation; The time-series characteristics modulated by frequency domain information, The above operations construct a FiLM modulation layer, which represents the frequency domain characteristics modulated by time-series information.

[0068] (4) Training a bibranch EEG emotion recognition network

[0069] 1) Constructing the loss function

[0070] Construct the loss function L according to equation (1):

[0071] (1)

[0072] Where N represents the total number of training samples, ranging from 24 to 32; in this embodiment, N is 32. C represents the number of emotion categories, with a value of 3. Here, γ is the weighting coefficient for class c, and γ is the focusing parameter, ranging from 2.2 to 4.5. In this embodiment, γ is set to 2.2. This is a soft label that has been smoothed. This represents the predicted probability that the i-th sample belongs to the c-th class.

[0073] 2) Training a bibranch EEG emotion recognition network

[0074] The training set was input into the two-branch EEG emotion recognition network for training. The training parameters were as follows: AdamW was selected as the optimizer, with a basic initial learning rate of 0.0022 and a weight decay coefficient of 0.04; the SEED standard batch size was 32 and the maximum number of iterations was 100, and training was carried out until the loss function converged.

[0075] (5) Testing the bibranch EEG emotion recognition network

[0076] The test set is input into the trained two-branch EEG emotion recognition network for testing, and the EEG emotion recognition results are output.

[0077] A brainwave emotion recognition method based on a dual-stream adaptive graph temporal interaction network was developed.

[0078] Example 2

[0079] The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network in this embodiment consists of the following steps.

[0080] (1) Preprocessing of EEG signals

[0081] The steps are the same as in Example 1.

[0082] (2) Extracting key features

[0083] The steps are the same as in Example 1.

[0084] (3) Constructing a two-branch EEG emotion recognition network

[0085] The steps are the same as in Example 1.

[0086] (4) Training a bibranch EEG emotion recognition network

[0087] 1) Constructing the loss function

[0088] Construct the loss function L according to equation (1):

[0089] The expression of equation (1) is the same as that in Example 1.

[0090] In equation (1), N represents the total number of training samples, ranging from 24 to 32. In this embodiment, N is 24, and C represents the number of emotion categories, with a value of 3. γ is the weighting coefficient for class c, and γ is the focusing parameter with a value range of 2.2 to 4.5. In this embodiment, γ is set to 3.3. Other parameters and variables, as well as their value ranges, are the same as in Example 1.

[0091] The other steps are the same as in Example 1. A brainwave emotion recognition method based on a dual-stream adaptive graph temporal interaction network is thus completed.

[0092] Example 3

[0093] The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network in this embodiment consists of the following steps.

[0094] (1) Preprocessing of EEG signals

[0095] The steps are the same as in Example 1.

[0096] (2) Extracting key features

[0097] The steps are the same as in Example 1.

[0098] (3) Constructing a two-branch EEG emotion recognition network

[0099] The steps are the same as in Example 1.

[0100] (4) Training a bibranch EEG emotion recognition network

[0101] 1) Constructing the loss function

[0102] Construct the loss function L according to equation (1):

[0103] The expression of equation (1) is the same as that in Example 1.

[0104] In equation (1), N represents the total number of training samples, ranging from 24 to 32. In this embodiment, N is 28, and C represents the number of emotion categories, with a value of 3. γ is the weighting coefficient for class c, and γ is the focusing parameter with a value range of 2.2 to 4.5. In this embodiment, γ is set to 4.5. Other parameters and variables, as well as their value ranges, are the same as in Example 1.

[0105] The other steps are the same as in Example 1. A brainwave emotion recognition method based on a dual-stream adaptive graph temporal interaction network is thus completed.

[0106] To verify the beneficial effects of the present invention, a computer simulation experiment and a comparative experiment with existing methods were conducted using the EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network according to Embodiment 1 of the present invention.

[0107] 1. Computer simulation experiment

[0108] The results of the computer simulation experiment are shown in Table 1.

[0109] Table 1. Results of computer simulation experiment in Example 1

[0110]

[0111] As shown in Table 1, the average recognition accuracy of this invention on the SEED dataset reached 97.82% for 15 subjects, with an average F1 score of 0.9782, an average Kappa coefficient of 0.9672, and a standard deviation of recognition accuracy of 0.67% for each subject. This indicates that the invention has high accuracy and high stability in 3-class emotion recognition.

[0112] 2. Comparative Experiment

[0113] The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network (hereinafter referred to as Example 1) of this invention was compared with the dynamic graph convolutional neural network method (hereinafter referred to as Comparison Method 1), the four-dimensional convolutional recurrent neural network method (hereinafter referred to as Comparison Method 2), and the spatiotemporal spectral hierarchical graph convolutional network method (hereinafter referred to as Comparison Method 3) in a comparative experiment. The experimental results are shown in […]. Figure 7 .

[0114] Depend on Figure 7 As can be seen, Example 1 outperformed Comparative Methods 1, 2, and 3 in terms of recognition accuracy across all 15 participants. The average recognition accuracy of Example 1 reached 97.82%, compared to 89.57% for Comparative Method 1, 93.44% for Comparative Method 2, and 95.36% for Comparative Method 3. The average recognition accuracy of Example 1 was 8.25, 4.38, and 2.46 percentage points higher than that of Comparative Methods 1, 2, and 3, respectively. The experimental results indicate that although Comparative Methods 2 and 3 approached the level of Example 1 in individual participants, the overall performance of Example 1 across the 15 participants showed less performance fluctuation and demonstrated stronger generalization ability and stability.

Claims

1. A brainwave emotion recognition method based on a two-stream adaptive graph temporal interaction network, characterized in that... It consists of the following steps: (1) Preprocessing of EEG signals Raw multi-channel EEG signals were collected from the SEED public dataset, which categorizes emotions into three types: sadness, neutrality, and pleasure. The raw EEG signals were segmented using a sliding window method with a window length of 400 sampling points and a step size of 200 sampling points. A 5-fold cross-validation strategy was used to divide the dataset into training and validation sets at an 8:2 ratio. (2) Extracting key features The EEG signal was decomposed into δ, θ, α, β, and γ frequency bands using a Butterworth bandpass filter. The differential entropy features of each frequency band were extracted using a differential entropy calculation method based on the Gaussian assumption. The power spectral density features of each frequency band were extracted using the Welch power spectral density estimation method. The power ratio features between the θ and α frequency bands and between the α and β frequency bands were extracted using the power spectral density ratio method. The spectral centroid features of the entire frequency band were extracted using the power spectral weighted average method. The spectral flatness features of the entire frequency band were extracted using the ratio of the geometric mean to the arithmetic mean of the power spectral density. The above features were concatenated to obtain a 14-dimensional frequency domain feature vector. (3) Constructing a two-branch EEG emotion recognition network The two-branch EEG emotion recognition network consists of an adaptive dual-topology graph neural network and a spatially embedded temporal Mamba network connected in parallel, followed by an asymmetric residual cross-feature modulation network. (4) Training a bibranch EEG emotion recognition network 1) Constructing the loss function Construct the loss function L according to equation (1): （1） Where N represents the total number of training samples, ranging from 24 to 32, and C represents the number of emotion categories, with a value of 3. Here, represents the weighting coefficient for class c, and γ is the focusing parameter, ranging from 2.2 to 4.

5. This is a soft label that has been smoothed. This represents the predicted probability that the i-th sample belongs to the c-th class; 2) Training a bibranch EEG emotion recognition network The training set was input into the two-branch EEG emotion recognition network for training. The training parameters were as follows: AdamW was selected as the optimizer, with a basic initial learning rate of 0.0022 and a weight decay coefficient of 0.04; the standard batch size of the SEED test subject was 32, and the maximum number of iterations was 100. Training was carried out until the loss function converged. (5) Testing the bibranch EEG emotion recognition network The test set is input into the trained two-branch EEG emotion recognition network for testing, and the EEG emotion recognition results are output.

2. The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network according to claim 1, characterized in that: In step (3), the adaptive dual-topology graph neural network is constructed by connecting an adaptive brain network, a static structure brain network in parallel, an adaptive gating fusion network, a dynamic graph convolutional network, and a channel attention module in series.

3. The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network according to claim 1, characterized in that: In step (3), the spatially embedded temporal Mamba network is constructed by sequentially connecting a spatial prior channel coding network, a channel projection layer, a bidirectional Mamba network, and a temporal attention pooling layer.

4. The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network according to claim 2, characterized in that: The adaptive brain network is composed of a global average pooling layer 1, a fully connected layer 1, a fully connected layer 2, a multi-head attention layer 1, a back projection layer, a feature fusion and normalization layer, and a dynamic adjacency matrix generation layer connected in series.

5. The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network according to claim 2, characterized in that: The adaptive gating fusion network is composed of a fully connected layer 3, a fully connected layer 4, a fully connected layer 5, and a weighted fusion layer 1 connected in series.

6. The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network according to claim 3, characterized in that: The spatial prior channel coding network is composed of a compression-excitation attention branch, a channel hybrid branch, and the output of the spatial prior branch connected in parallel, which are then connected in series with a weighted fusion layer 2 and a layer normalization layer. The compressed-excitation attention branch is composed of a global average pooling layer 2, a fully connected layer 6, a fully connected layer 7, and a channel feature weighting layer connected in series.

7. The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network according to claim 1, characterized in that: In step (3), the dual-branch EEG emotion recognition network is constructed by sequentially connecting the asymmetric residual cross-feature modulation network, the residual gating fusion module, the multi-head attention layer 2, and the classification module.

8. The EEG emotion recognition method based on a dual-stream adaptive graph temporal interaction network according to claim 7, characterized in that: The asymmetric cross-modulation module is composed of a fully connected layer 8, a fully connected layer 9, a fully connected layer 10, a fully connected layer 11, and a FiLM modulation layer connected together; the fully connected layer 8 and the fully connected layer 9 are connected in series, the fully connected layer 10 and the fully connected layer 11 are connected in series, and the output terminals of the fully connected layer 10 and the fully connected layer 11 are connected to the FiLM modulation layer. The residual gated fusion module is composed of a fully connected layer 12 and a fully connected layer 13 connected in series; The method for constructing the FiLM modulation layer is as follows: Perform bidirectional cross-branch eigenmodulation calculation using the following formula: in, To spatially embed the original temporal features of the temporal Mamba network output, The original frequency domain features output by the adaptive dual-topology graph neural network; , The scaling and offset factors are used to guide the modulation of timing features in the frequency domain branch. , The scaling and offset coefficients for frequency domain feature modulation guided by the time-series branch are all obtained through network training. For Hadamard's element-wise multiplication operation; The time-series characteristics modulated by frequency domain information, The above operations construct a FiLM modulation layer, which represents the frequency domain characteristics modulated by time-series information.

Images